[Bug go/64999] s390x libgo test failure in TestMemoryProfiler

["boger at us dot ibm.com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64999

--- Comment #46 from boger at us dot ibm.com ---
(In reply to Ian Lance Taylor from comment #45)
> If we change the PC returned by backtrace_full, and then use that changed PC
> to look up file/line information, we might get different results.  That
> seems clear.  My next question is: when does this matter?
> 
> There are two ways that we use the result of runtime_callers: either we use
> the file/line information, or we use the PC.  If we use the file/line
> information, all is well, and there is nothing to discuss.  If we use the
> PC, as in runtime.Callers, then it's true that if we pass that PC back to
> libbacktrace we may get different file/line information.  But, so what? 
> We've already lost the original file/line information anyhow.
> 
> The idea is that the changed PC will be the same as the PC returned by the
> gc library.  Therefore, we should get approximately the same file/line
> information as the gc library gets.  That is why runtime/pprof/pprof.go in
> the gc library backs up the PC to the call instruction: because it knows
> that it has the PC after the call instruction.

Let me just respond to this last paragraph for now.  If the pc values in gc are
always the instruction following the call (or the excepting instruction), and
runtime_callers expects it to be that way, and we want gc and gccgo to return
the same results, then I don't understand why backtrace_full and
backtrace_simple should ever decrement the pc for any platform.  There is a
brief mention of this in comment 9 -- that we can make use the fact that the pc
had been decremented by inspecting it and making use of the fact that is an
abnormal pc value; however I don't see anywhere in the libgo code where it does
this, and besides, for some platforms after the pc is decremented it is still a
normal looking pc so there is no way to know if it had been decremented. 
Getting the right pc and the right instruction seems more important than
preserving some information that is currently not being used and can't be used
on all platforms.

That means, if the decrement was removed completely from backtrace_full and
backtrace_simple then the pc values would be correct for each platform and the
line numbers would be determined based on the correct pc.  And then we would
still need the BackupPC function for those cases where the call instruction was
needed as is done in pprof.go and some extra stuff for s390 and s390x for their
various call instructions to get the right number of bytes to back up.