Load static archives with a /SYM64/ symbol table

[zardus]

THIS MESSAGE WAS GENERATED BY AN AUTOMATED PROCESS

cle.Loader cannot open a static archive whose index is the 64-bit variant. GNU ar names that member /SYM64/; arpy classifies it as a member whose real name lives in the long filename table and then raises a bare ValueError parsing SYM64 as the offset into that table. Every mips64 archive is written this way, as is any archive too large to index with 32-bit offsets.

arpy discards the symbol table for either spelling, so the AR backend now skips a leading /SYM64/ member itself, and only when the member header fits inside the file. It also wraps arpy's parse failures in CLEInvalidBinaryError. The classification really belongs in arpy, but both the pinned 1.1.1 and the current release get it wrong.

The regression loads a real mips64 /SYM64/ library added by angr/binaries#176, so the checks here stay red until that merges.

Validation: #736 (comment)

[zardus]

THIS MESSAGE WAS GENERATED BY AN AUTOMATED PROCESS

Validation record for head 1ffb9f17daeb4049fc4dc5f5ba3f1255963eeebd against baseline b58ea02a446106647cdaae32bdf91b7062404cc1.

• Focused: pytest tests/test_static_archive.py — 4 passed
• Regression: with cle/backends/static_archive.py reverted to the baseline and the tests left alone, three of the four fail. test_sym64_symbol_table and test_sym64_size_past_end both raise ValueError: invalid literal for int() with base 10: b'SYM64' from arpy.py:277, escaping cle.Loader uncaught, and test_bad_header_magic raises arpy.ArchiveFormatError: file header magic doesn't match. The 32-bit control test_symbol_table passes either way, which is what it is there for.
• Full suite: pytest tests — 206 passed, 9 skipped, on Python 3.12.13 with arpy 1.1.1
• Lint/type against the merge base: pylint cle/backends/static_archive.py 9.67 -> 9.81 and tests/test_static_archive.py 10.00 as a new file; pyright badness 0.20 -> 0.11 and 0.00
• Pre-commit: the 24 configured hooks over --all-files, all passed or skipped, no rewrites

The regression loads tests/mips64/sym64_archive.a, added by angr/binaries#176: a 3,104-byte MIPS64 big-endian library indexed by a /SYM64/ table, holding one member x11_xcb.o. The 32-bit control loads tests_src/i2c_master_read-nucleol152re/mbed/TARGET_NUCLEO_L152RE/TOOLCHAIN_GCC_ARM/libmbed.a, already in angr/binaries, which has an ordinary / symbol table and a // long filename table. The two malformed cases copy one of those fixtures and overwrite a single member header field — the terminating magic and the symbol table size — so each input differs from a real archive only in the field under test. cle's CI checks out binaries master unconditionally, so the checks here stay red until angr/binaries#176 merges; the results above come from a local checkout of that branch. GNU binutils 2.46 reads both fixtures: ar t lists every member, including the ones named through the long filename table, and nm --print-armap prints the index for the /SYM64/ and / spellings alike.

The failure came out of a sweep of real-world static archives run on cle 638301bbf02865e53394e6643e1eb6e66083ee94. All 118 mips64 archives in it use /SYM64/ and none of them loaded, every one raising ValueError: invalid literal for int() with base 10: b'SYM64'. The new fixture is one of those archives, sha256 992dec1ee6f8400f30759f7998d7f3699261a39f7d7c03d175f1d5f7269e5782, taken from a public OS distribution.

Inputs probed by hand beyond the regression: even- and odd-length symbol tables, an empty index, a size field padded with spaces, a negative size, a non-numeric size, a size running past the end of the file, a /nnn long filename member that must not be taken for an index, a BSD member, and a /SYM64/ archive nested inside another. Each now either loads or raises CLEInvalidBinaryError; on the baseline every /SYM64/ case raised a bare ValueError. Removing the file-length bound on its own makes test_sym64_size_past_end fail with ValueError: No arch is assigned yet.

Caveats: the first Windows CI attempt failed before collecting anything, when two pytest-xdist workers raced to install the pyvex FFI parser cache and os.replace hit PermissionError: [WinError 5]; the rerun is green and an unrelated cle PR started at the same minute failed identically. An archive holding no members at all still fails with ValueError: No arch is assigned yet out of Backend.arch, which is unchanged from the baseline and independent of the symbol table spelling. Only cle and binaries are checked out for this branch, so the other workspace suites did not run.

[angr-bot]

Corpus decompilation diffs can be found at angr/dec-snapshots@master...angr/cle_736