[Henglie]

Fixes #54.

Root cause

StringLocator assumed two things that the DEX format does not require:

1. the string_data_items are written back to back, so the first and last string_ids entry bound the region;
2. they are written in string_ids order, so the item behind a match carries index + 1 — which let the locator map a match through the following offset and subtract one.

d8/R8 satisfy both, but a layout pass such as dexlayout permutes the items, and vendor tools repackage system APKs that way. A synthetic DEX whose string_data is permuted (offsets, signature and checksum rewritten, code unchanged) shows both failure modes from the report:

layout	before	after
index order	correct	correct
full reversal	region collapses to start > end, every query returns nothing	correct
middle permutation	first → {1} instead of {3}	correct

Fix

• The region is taken from the smallest and largest offset instead of the first and last entry.
• A match ends at the first terminator behind its start, since a MUTF-8 payload never holds 0x00 on its own. A match that reaches past that terminator runs into the following item and names no string.
• When the offsets are already in index order, the old single dict lookup still resolves the match, so d8/R8 output keeps its cost. A permuted layout searches the ascending offsets with bisect instead, and the layout is decided once while building the map.
• Building the map unpacks the offset array once instead of 68k unpack_from calls, and fills the lookup with update(zip(...)).

Tests

tests/test_string_locator_order.py covers the ordered, reversed and middle-permutation layouts, a substring that must map to the item that holds it, a match that spans two items, and the findrefs output for the reversed layout. tests/dex_fixture.py gains reorder_string_data(), which moves the items inside their region and recomputes the SHA-1 signature and Adler-32 checksum, so the result stays a valid DEX — the test asserts that every index still resolves to the same string before using it.

Verification

• Reverting string_locator.py to 9216fe2 makes the new file fail 13 subtests; with the fix all 7 pass.
• tests/run_tests.py --suite unit: 72 tests pass (65 existing + 7 new).
• tests/run_tests.py --suite integration: 5 tests pass.
• tests/benchmark_compare.py --baseline <base> --samples 15 --cases unit: no gated regression. unit/string_locator −35%, unit/type_locator −4%, unit/method_locator +2.5% (p=0.30), unit/field_locator +2.0% (p=0.30).
• tests/benchmark_reference.py --samples 3: all 16 count metrics match reference-baseline.json.

Not verified here: the GUI suite (this machine is Windows with no Tk display), and benchmark_reference.py's 0.0880 s decompilation budget, which the base commit misses on this machine as well (0.28 s vs 0.33 s) — the recorded baseline is a Linux CI number and the decompiler path is untouched.

Note

The end of an item is still taken as the start of the next one: string_data_item has no alignment requirement and both d8 and dexlayout write the items back to back. A layout that pads between items falls back to the search path, which bounds the item by the following offset.

[MG1937]

貌似有性能劣化，两个测试集没过，要不优化一下？