Comment 1:

Here are the functions that cause problems on race builders:
runtime: unknown argument frame size for runtime.asmcgocall
runtime: unknown argument frame size for runtime.sigreturn
runtime: unknown argument frame size for runtime.sigtramp
runtime: unknown argument frame size for runtime.raceread
I've also seen this one *without* -race when stackalloc crashed and wanted to unwind the
stack:
runtime: unknown argument frame size for runtime.morestack
Can preemptive scheduler make things even worse? C functions do have args info, right?
Part of the solution may be to not calculate args size when we are not doing GC, because
SIGPROF causes stack unwinding at unpredictable points.

Comment 2:

Can we just make pprof refuse to enable itself in race builds? Profiling won't give
useful data in that case anyway.

Comment 3:

I agree that profiling with -race is generally useless (it's useful for optimizing the
race detector itself, though).
But this is not race-specific bug. This is due to SIGPROF arriving before/after stack
frame allocation (SUB/ADD SP).

Comment 4:

Labels changed: added priority-later, removed priority-triage.

Comment 6:

I am happy that the problems with the race builders around the time of this issue being
raised have been resolved and that this ticket should be closed.

Status changed to Fixed.