FYI, your "8‿64 •bit._cast -⟜@ raw" in bqn/util.bqn results in passing to •bit._cast integers in the range 0..255, whereas it expects integers ¯128..127 (and does arbitrary platform-specific things on integers outside that for the int16→int8 narrow); removing the "-⟜@" makes it work properly, and 2.5-5x faster while at it from not going to intermediate int16 (allowing the bitcast to be O(1) by itself). (•FBytes → •file.MapBytes also probably improves things, at the cost of the obvious mmap problem of issues if the file changes, as BQNs arrays are immutable). This is also what caused the problem at https://github.com/ClickHouse/ClickBench/pull/939/changes/40... (there's been some discussion about what to do about invalid •bit._cast before, considering it's rather perf-sensitive; it's the only place in CBQN with such quiet very-wrong results on invalid inputs)

The default CBQN "make o3" on x86-64 also results in it only using SSE2 (utilizing function multiversioning is on the ever-infinite TODO list, though somewhat-low on it considering it's strictly-unnecessary in any specific situation; there's also AVX-512 usage on a branch, but mostly only AVX2 on mainline; and no arm SVE)

That all said, CBQN doesn't currently do any loop fusion, so being significantly-slower for sequences of operations over larger-than-cache arrays would kinda just be expected. BQN also just isn't particularly intended for database work anyway.

(didn't look much at the specific query impls, though "Pair" in utils.bqn is at least an overlong version of "Pair ← ⋈¨"; and some if not all of those Pairs would be better as "≍˘" to avoid nested arrays and ensuing pointer chasing; and, of course, if some of the columns are bools/int8/int16/int32, it'd be beneficial to store & load them as such instead of float64)