BTCrashFilter writes `[crash] ... addr=0x... (btl4+0xNNNN)` plus an EBP-chain
walk into the tester's day log, and its own comment claims "we hold the PDB".
We did not. Two independent gaps, both fatal to the point of the thing:
1. No Release PDB was produced AT ALL -- only build/Debug/btl4.pdb, and the
shipped exe embedded no PDB path. btl4+0xNNNN from a tester could not be
turned into a function name by anyone.
2. Release flags were /O2 /Ob2 /DNDEBUG with no /Oy- anywhere, and /O2 implies
/Oy on x86. The walker follows EBP, so the stack it printed was unreliable
even when the faulting address was not.
Net effect: when the Owens crash finally lands we would have received an address
nobody could resolve, and a call chain we could not trust. Cheaper to fix before
tonight's session than to wait for the crash to happen twice.
/Zi emit debug info -> a PDB. Does not change codegen.
/Oy- keep EBP as a frame pointer so the walk is trustworthy.
/DEBUG + /OPT:REF + /OPT:ICF -- /DEBUG turns the last two OFF by default,
which would have quietly bloated the shipped exe with unreferenced
code. The exe grew 512 bytes, the debug directory entry, and nothing else.
Also emitting /MAP, which turned out to matter: this machine has no cdb, and a
tester's operator may not have one either. The .map is plain text, so an offset
can be resolved with a text editor. tools/symcrash.py does it properly --
nearest-preceding-symbol against the archived map, module-external addresses
(ntdll, kernel32) passed through untouched since they are not ours.
Verified end to end rather than assumed, using the BT_CRASHTEST=1 hook that
exists for exactly this:
[crash] addr=0xb8260 (btl4+0x8260) access=1 target=0x0
[crash] stack: btl4+0x8260 btl4+0xfe7ed 0x763ffcc9 ...
btl4+0x8260 -> _WinMain@16 +0x7e0
btl4+0xfe7ed -> __scrt_common_main_seh +0xf8
which is exactly right: BT_CRASHTEST does *(volatile int *)0 = 0 inside WinMain,
called from the CRT entry. Correct stack AND correct names.
The PDB and map are archived per build and never shipped: .gitignore already
covers *.pdb and dist/, and mkdist only picks up .dll next to the exe, so
neither can reach a zip by accident. The session header stamps
build=4.11.<n> (<hash>), so an archived pair matches a tester's log exactly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
