GeezerVison

zipimporter.get_filename() (and therefore find_spec()) read, decompressed and unmarshalled the whole .pyc just to determine the module's file name, and exec_module() then did it all again. Only the 16-byte header matters for picking the file: a bad or stale header falls back to the sibling .py, while errors in the marshalled body never do.

This splits the header checks out of _unmarshal_code() and has get_filename() run only those, reading (and for deflate/zstd members, decompressing) just enough of the member to cover the header.

Observable change: a .pyc with a valid header and a corrupt body now fails in exec_module() instead of find_spec(), with the same exception. GH-153799 made the equivalent change for a SyntaxError in source modules.

Benchmarks

Release build, Linux, warm page cache, timestamp-based pycs.

Importing 27 stdlib packages (148 modules) from a zip of ~350 stdlib modules with .pyc next to .py, min of 40 runs:

Per module, find_spec() + get_code() for a single .pyc (in-process medians):

find_spec() itself no longer scales with pyc size: ~0.1 ms regardless of size, versus 43 ms for a 5 MB pyc on main.

zstd gains less because ZstdDecompressor construction dominates small reads and zstd decodes a whole block before producing output.

• Issue: zipimport reads and unmarshals pyc's twice per import #161644