Skip to content

Apple's open-source XNU kernel on Android phones: its own loader and drivers, Metal on the Adreno. Runs from RAM; a reboot gives the phone back to Android.

C

docs/air: two ways a comparison answers about your own tree instead

vcvkkbcc807db5534
diff --git a/docs/air.md b/docs/air.md
index 60ae047..0d01b3f 100644
--- a/docs/air.md
+++ b/docs/air.md
@@ -86,6 +86,20 @@ functions and not entry points at all.
The row "run on the phone's GPU | 7" in the first table above is unaffected and still reads 7:
it counts modules of the 456-module corpus, and these are generated modules that are not in the
corpus. The two measurements are deliberately not added together.
+- **Compare against someone else's tree from a pristine extract, never from your own build.**
+ Checking whether Mesa's `main` still had the rounding fault meant comparing one file against
+ ours — and ours has our own patch in it, 202 487 bytes against 202 144 clean. Taken from this
+ tree the two would have differed by exactly our own change, and the search would have been for a
+ difference in `main` that is not there. Unpacking the tarball again costs a second and removes
+ the working copy from the question.
+
+ The same mistake has a twin in how a check is run: `patch -s --dry-run` exited 0 against `main`'s
+ file and that was written down as "applies cleanly", when without `-s` it says
+ "Hunk #1 succeeded at 410 (offset 2 lines)". The context matched and the line number had moved,
+ which is a different and more useful statement. **Quiet output is not clean output** — a flag
+ that hides a success's details makes the success less informative than it appears, which is the
+ same shape as the skips below.
+
- **A `continue` that adds nothing to a total is where this check has gone wrong three times.**
Not a maxim but a thing to grep for. Three times on 20 October an absent check was dressed as a
passing one, and each time the shape was identical: a loop skipped a case and incremented no