update

docs/pre12-batch-02.md

@@ -0,0 +1,31 @@
+# Pre12 Batch 02
+
+This batch replaces exactly three potentially unaligned typed loads with the
+FreeBSD little-endian byte decoders planned for Pre12.
+
+- `xattr.c`: `hdrbuf` contains two raw metadata bytes returned by
+  `erofs_xattr_read_backing(..., sizeof(raw_len), ...)`; `le16dec(hdrbuf)`
+  decodes the prefix-record length.
+- `inode.c`: `buf` contains at least one complete 32-byte compact on-disk inode
+  returned by `erofs_read_metadata`; its first two bytes are the little-endian
+  `i_format` field decoded by `le16dec(buf)`.
+- `data.c`: the non-indexes branch sets `entry_size` to
+  `EROFS_BLOCK_MAP_ENTRY_SIZE == sizeof(__le32) == 4`; `idx` points to those
+  four raw block-map bytes and `le32dec(idx)` decodes the block address.
+
+For bytes `b0..b3`, both old expressions and the new helpers produce
+`b0 | b1 << 8` or `b0 | b1 << 8 | b2 << 16 | b3 << 24`.  On a little-endian
+host the typed load already has that value; on a big-endian host `leXXtoh`
+byte-swaps the typed value to that same formula.  The helpers preserve this
+result while avoiding any alignment requirement.
+
+`xattr.c` and `data.c` receive the helpers through `internal.h` to
+`erofs_fs.h` to `<sys/endian.h>`; `inode.c` also includes `<sys/endian.h>`
+directly.  Existing `le16dec` and `le32dec` consumers confirm the established
+include convention.  Error checks, branches, value ranges, and buffer release
+ordering are unchanged.  No point was skipped and no other endian read was
+changed.
+
+Validation used exact-expression counts, buffer-source and length inspection,
+symbolic endian equivalence, source-diff inspection, and `git diff --check`.
+No build, QEMU, smoke, feature, or other dynamic test was run.