Resolve glob matches in one path space (fixes #191)

[typeless]

expand_glob_pattern merges two sources of matches: a filesystem scan, whose paths are source-relative, and a scan of Generated graph nodes, whose paths carry the build-root prefix. Out-of-tree those are different path spaces, and everything downstream compared them as plain strings — the sort, the dedup, and apply_exclusions.

Three symptoms, one cause

%f order depended on the name of the build directory. "../AAA/a/gen.dat" sorts before "keep.dat"; "../zz/a/gen.dat" sorts after:

putup -B AAA  ->  cat ../AAA/a/gen.dat keep.dat     all.txt: GEN\nKEEP
putup -B zz   ->  cat keep.dat ../zz/a/gen.dat      all.txt: KEEP\nGEN

A generated file shadowing a same-named source appeared twice — the halves spelled it a/x.dat and build/a/x.dat, so the dedup kept both and the checked-in source was never consumed:

cat ../build/a/x.dat ../build/a/x.dat > ../build/a/all.txt

Exclusions silently stopped applying. Normalized source-relative, compared against a build-prefixed path, so !skip.gen held in-tree and was ignored out-of-tree.

The fix

The loop already computed the build-root-stripped path to run the glob against, then discarded it in favour of the physical one. It now pushes the stripped path, so both halves are source-relative and the existing sort, dedup and literal exclusion comparison are correct as written.

resolve_input_node checks BuildRoot before SourceRoot at source-relative paths, so the stripped spelling still names the generated node — and the node is known to exist, since these matches come from iterating the Generated nodes themselves.

PR #180 introduced this merge and a follow-up normalized the ../ component lexically. That was one axis of a two-axis problem; this is the other.

Tests

Three e2e scenarios, each the shape that was never covered — a glob matching a generated and a source file, out-of-tree. Each observed failing first; the shadowing case showed FROMRULE twice.

Note the test-suite gap was narrower than I first wrote in the issue: out-of-tree coverage is extensive (265 scenarios use -B build). What was missing was specifically a glob spanning both node kinds, which is the only shape where the two spaces meet.

Also checked by hand, because the previous fix in this area got a related case half-right: a ../ pattern crossing directories out-of-tree resolves to distinct files in source-relative order, with correct physical paths rendered.

cat ../build/b/gen.q ../b/src.q > ../build/m/out.txt

Verification

make test: 142404 assertions, 637 cases, 32 e2e shards. make tidy, make iwyu clean. GCC BSP parses (24 Tupfiles, 1853 commands).

🤖 Generated with Claude Code

https://claude.ai/code/session_01TLob4Ef3qyMd3Dnp9q3XmA

---

Review corrections

An adversarial review found the claims here were broader than the fix, and one test asserted the wrong thing.

The exclusion axis is only half closed. A glob exclusion (!s*.gen) is expanded against the filesystem, so it never sees generated files and is still ignored out-of-tree; in-tree it is unstable, rendering echo keep.gen skip.gen on build 1 and echo keep.gen on build 2. Bin {group} members are still pushed build-prefixed into the same list. Both pre-existing, both verified on main too, now filed as #195. The comments in this PR no longer claim more than it does.

The order test had a blind spot. Asserting only that two build directories agree would still pass if the filesystem half stopped contributing entirely — both would be GEN and both equal. It now pins "GEN\nKEEP\n", covering source half, generated half and canonical order together.

The shadowing test canonicalized a divergence. Real tup rejects that project outright:

tup error: Attempting to insert 'x.dat' as a generated node when it already exists as a
different type (normal file).

and putup in-tree overwrites the checked-in source file on disk (FROMSRC → FROMRULE). Filed as #194. The scenario is now tagged [deviation], states that it asserts only the dedup this PR is about, and points at the issue. Consequence worth noting: #191's second symptom is only reachable in a project #194 would reject, so that half of this fix is defensive rather than user-facing.

Behaviour change for existing projects. A Tupfile that worked around #191 by spelling an exclusion build-prefixed (!build/skip.gen) silently stops excluding. The source-relative spelling is the tup-conformant one, so the break is intended, but it is a break.

The review also verified independently, by execution, what I had argued: in-tree output is byte-identical to main (including a ../ cross-directory glob), the 3-tree GCC BSP dry run is byte-identical over 3530 commands, every consumer of these strings resolves BuildRoot-first before any exists() probe, an index written by main upgrades cleanly in one re-run, and all three new tests fail when the one-line change is reverted.

[github-actions]

PR metrics

Performance (gcc example, Linux)

Workload	Instructions	CPU time	Page faults	D1 miss	LL miss	Wall	Peak RSS
parse	1076 M	0.51 s	14.2 k	0.8%	0.1%	0.505 s	31.1 MB (-0.1MB)
dry-run	1881 M	0.62 s	16.6 k	0.9%	0.1%	0.619 s	40.2 MB

Deterministic signals: instructions (cachegrind-simulated instruction reads — exact across runs, no PMU needed), page faults, peak RSS, and the cachegrind D1/LL miss rates. CPU time is user+sys from time(1).

Internal statistics (gcc example, up-to-date dry run)

Metric	Value
Tupfiles parsed	24
Commands	3545
Commands scheduled	0
Files checked	5818
Files changed	0
Files in index	6087
Graph edges	367913
Index size (bytes)	7378282
Implicit deps	330624
Hash computations	0
Hashes skipped (stat cache)	5818
Stat calls	5869
Parse time (ms)	485.5
Total time (ms)	601.1
Runner CPU	AMD EPYC 9V74 80-Core Processor

Counters from putup -n --stat on the fully-built gcc example (up-to-date dry run): deterministic work measures — a jump in commands scheduled, hash computations, or stat calls is a real behavior change, not noise. Timings are the minimum over repeated runs, compared only against a baseline from the same CPU model; the counters are the regression signal.
Timing deltas suppressed: baseline ran on different hardware (AMD EPYC 7763 64-Core Processor).

Binary size (Linux)

Binary	.text	.data	.bss	File
putup	517.5 KB	1.6 KB	98.7 KB	615.2 KB

Test coverage (lines)

Overall	Median file	Min file	Max file
86.7%	95%	0.0% src/platform/path-posix.cpp	100.0% include/pup/core/arena.hpp

103 files · 15164/17494 lines covered

Deltas vs main@fc14f89de.

Updated for e1f4f60