← board

{*xs} skips the set's deduplication

Repro

print(len({*[1, 1, 2]}))   # CPython 2, pxx 3
print(len({1, 1, 2}))      # CPython 2, pxx 2  <- the non-star path is right
print(len({*"aab"}))       # CPython 2, pxx 3

Reproduced on pin v292 and earlier. Silent: the result is still a set-shaped value and still prints like one, just with an extra element.

Cause

PyParseListLiteralT(opener, closer, dedup) builds both list and set displays into a TPyList. When dedup is set it routes each ordinary element through TPyList.add — added by bug-nilpy-set-literal-does-not-deduplicate, which fixed exactly this defect for the non-star path.

The star-spread arm in the same procedure calls TPyList.extend unconditionally, and extend appends. So the fix for the sibling case landed on one arm of the same procedure and not the other — the shape devdocs/dev/normalise-dont-special-case.md describes, and the second time this literal parser has shown it.

Fix

Make the star arm honour dedup: either a TPyList.extend_unique in compiler/builtin/pylib.pas (a pylib change, so it needs a pin), or emit a loop of add calls, or extend-then-dedup in place. Prefer whichever leaves ONE mechanism deciding "does this display deduplicate" — the current split is what produced the bug.

Check the dict-display and comprehension paths for the same split while there: {k: v for ...} and {x for x in ...} also build a marked TPyList.

Gate

make compiler/pascal26 + the repro + tools/gate.sh quick; a pin only if pylib changes. Add the repro to test/test_nilpy_str_as_an_iterable_argument.npy, which carries a sorted({*s}) assertion and deliberately does NOT assert the length today — that omission is this ticket.

2026-08-14 — FIXED, and pylib needed nothing

PyParseListLiteralT's star arm now picks its bulk operation from the same dedup flag that already picks the per-element one:

per element spread
list display append_self extend
set display add setupdate (was extend)

TPyList.setupdate — for i := 0 to other.count - 1 do Self.add(other.at(i)) — already existed for set.update, so there was nothing to add to pylib and no pin is needed. That is the outcome worth noting: the ticket proposed writing an extend_unique, and the deduplicating bulk add was already there under the name Python gives it. Checking for the existing mechanism before adding one is what kept this from becoming a second way to do the same thing.

One thing still decides whether a display deduplicates.

The siblings the ticket asked about are already right

Measured rather than assumed, since the whole ticket is about a fix that landed on one arm and not another:

{x for x in [1,1,2,2,3]}   -> 3 elements     set comprehension: correct
{x % 2 for x in xs}        -> [0, 1]         correct
{k: 1 for k in "aab"}      -> 2 keys         dict comprehension: correct
set([1,1,2]), set("aab")   -> correct

Only the display's star arm was wrong.

Gate

test/test_nilpy_set_star_spread_dedups.npy + .expected from CPython. It carries the LIST rows deliberately — [*[1,1,2]] must still be [1, 1, 2], which is the failure a "make it dedup" fix invites. make compiler/pascal26 fixedpoint + tools/gate.sh quick.

Log