[imlvts]

Grafting nothing at a branch must neither create nor destroy a location: an existing branch is emptied and survives as a dangling path, an absent one stays absent, as after graft of an empty source. graft_child_maps' slow path left the old branch in place for an empty map or one without a root value, its fast path grafted an empty node (tripping a debug assertion) and created the focus, and graft_masked_branches created the focus when every masked branch was absent from the source.

[Adam-Vandervorst]

Can we get a specialized micro benchmark for them in the progress so we can compare fixes?

[imlvts]

Change from master to fix/graft-nothing (worktree: aac0433 + faster graft_child_maps), with the 95% confidence interval in brackets. Positive means slower.
30 alternating runs per version, pinned to one CPU core; intervals from resampling the per-run medians.
Bold: the interval excludes 0 and the change is at least 5%.

bench	1	2	3	8	30	100	200	at 200: master → fix	at 200, committed aac0433
graft_masked_branches
src_full	+1 [+1, +2]	+4 [+4, +5]	+1 [+1, +2]	+1 [+0, +2]	+2 [+1, +3]	+2 [+1, +3]	+3 [+2, +4]	2.38 µs → 2.45 µs	+3%
src_half	+0 [−0, +1]	+6 [+5, +7]	+2 [+1, +2]	+2 [+2, +3]	+6 [+5, +7]	+8 [+7, +9]	+11 [+10, +12]	1.85 µs → 2.05 µs	+11%
src_half_remove_unset	−2 [−4, +1]	−8 [−9, −7]	+0 [−3, +4]	+2 [+2, +3]	+0 [−1, +1]	−1 [−2, +0]	+0 [−0, +1]	1.04 µs → 1.04 µs	−0%
src_dangling	+0 [−0, +1]	+2 [+2, +3]	+0 [−0, +0]	+1 [+0, +1]	+2 [+1, +3]	+1 [+0, +2]	+4 [+3, +4]	1.77 µs → 1.84 µs	+4%
src_none_masked	+0 [−0, +1]	+2 [+1, +2]	+2 [+1, +2]	+5 [+4, +6]	+10 [+9, +11]	+17 [+16, +18]	+40 [+39, +41]	1.29 µs → 1.81 µs	+40%
src_bare_value	−0 [−1, +0]	+1 [+0, +2]	−4 [−5, −4]	−1 [−2, −1]	+7 [+6, +8]	+15 [+14, +16]	+37 [+36, +38]	1.32 µs → 1.81 µs	+39%
missing_focus_src_full	−0 [−1, +1]	+0 [−1, +1]	−3 [−4, −2]	−1 [−2, +0]	−0 [−1, +0]	+2 [+1, +3]	+1 [+0, +1]	3.59 µs → 3.62 µs	+1%
missing_focus_src_none_masked	−4 [−5, −4]	−6 [−6, −5]	−30 [−31, −30]	−26 [−26, −25]	−14 [−14, −13]	−6 [−7, −5]	−6 [−7, −6]	2.53 µs → 2.37 µs	−7%
graft_child_maps
keep_full	+12 [+10, +13]	+9 [+8, +9]	−5 [−6, −4]	−22 [−23, −22]	−37 [−37, −36]	−49 [−50, −49]	−64 [−64, −64]	12.88 µs → 4.60 µs	−20%
keep_half_empty	+11 [+10, +12]	+24 [+24, +25]	+1 [+0, +2]	−2 [−2, −1]	−12 [−12, −11]	−26 [−27, −26]	−54 [−54, −53]	7.11 µs → 3.29 µs	+28%
remove_unset_full	−2 [−4, −1]	+1 [+0, +2]	−6 [−7, −4]	+5 [+1, +8]	−6 [−7, −6]	−4 [−5, −4]	−2 [−3, −2]	1.42 µs → 1.39 µs	+5%
remove_unset_half_empty	−3 [−4, −2]	−0 [−1, +1]	−7 [−8, −6]	+9 [+7, +11]	−4 [−5, −3]	−5 [−5, −4]	−4 [−4, −3]	953 ns → 917 ns	+2%

[imlvts]

According to claude, the 40% perf loss src_none_masked and src_bare_value is inherent to the fix itself. on master, not doing that work is the bug.

[Adam-Vandervorst]

Hmm that's quite a swing for only serving the purpose of giving a faster zipper algebra -- I can't recall how much this specialization saved. Especially keep half empty, I guess. Do we have benchmarks on zipper algebra @marcin-rzeznicki ?

[imlvts]

keep_half_empty is slower on small maps but faster on bigger maps (vs master)