So that we can get some idea, for nightlies, whether Breakpad is having success with CFI unwinds, FP unwinds, etc.

I think this would be a useful addition for troubleshooting unwinding in nightlies. It generates lines like this: Profiler: BPUnw frame stats: \ TOTAL 4000 CTX 224 CFI 540 FP 3236 SCAN 0 NONE 0 TOTAL is the total number of frames in this sample interval (changed to 5000 in the patch) CTX is # frames from direct context (iow direct-from-CPU program counter) CFI is # frames from CFI FP is # frames from FP chasing SCAN is # frames from stack scan NONE is # frames with some other trustworthyness metric (should always be zero) Also, because there is always exactly one CTX frame per unwind, the average unwind length is TOTAL / CTX.

(In reply to Julian Seward from comment #1) > Also, because there is always exactly one CTX frame per unwind That's neat. That tell you the number of samples.

Comment on attachment 739576 [details] [diff] [review] Patch Review of attachment 739576 [details] [diff] [review]: ----------------------------------------------------------------- I'm happy with the patch as-is but I'd like to discuss my second comment before landing. ::: tools/profiler/UnwinderThread2.cpp @@ +1550,5 @@ > + nf_CFI_SCAN++; break; > + case google_breakpad::StackFrame::FRAME_TRUST_FP: nf_FP++; break; > + case google_breakpad::StackFrame::FRAME_TRUST_CFI: nf_CFI++; break; > + case google_breakpad::StackFrame::FRAME_TRUST_CONTEXT: nf_CONTEXT++; break; > + default: break; Perhaps default should count towards NONE. Not a big deal since we can always do TOTAL - SUM(...). @@ +1712,5 @@ > if (n_frames_dubious > (unsigned int)sUnwindStackScan) > break; > > + if (LOGLEVEL >= 2) > + stats_notify_frame(frame->trust); Perhaps we should do this before we discard stack scan? This way we can watch SCAN/NONE for incomplete unwinds.