Warnings and compile-time failure on 458.sjeng

Wed May 25 13:39:32 PDT 2016

> On May 25, 2016, at 1:20 PM, Xinliang David Li <davidxl at google.com> wrote:
> 
> 
> On Wed, May 25, 2016 at 1:17 PM, Steven Wu <stevenwu at apple.com> wrote:
> It is -O3 PGO for arm64 with training input.
> 
> I am looking as well. Seems PData->Values is NULL but I don't know how yet.

Steven sent me a core dump. I'll take a look at this to double-check the value of PData.

> 
> With static allocation, this should not happen -- the field is statically initialized.

But, we know static allocation is enabled because we get a warning from allocateOneNode().

If enough warnings occur, is it possible for CurrentVNode to wrap around s.t:

  COMPILER_RT_PTR_FETCH_ADD(ValueProfNode, CurrentVNode, 1) < EndVNode ?

If this happens, allocateOneNode() can eventually return an incorrect VNode which will then be stored into PData->Values.

vedant

> 
> David
>  
> 
> 
> Steven
> 
>> On May 25, 2016, at 1:06 PM, Xinliang David Li <davidxl at google.com> wrote:
>> 
>> I could not reproduce the issue (only saw out of nodes warning). What additional option is used (I used O3)? What is the input (train, ref, test)?
>> 
>> David
>> 
>> On Wed, May 25, 2016 at 11:24 AM, Vedant Kumar <vsk at apple.com> wrote:
>> Hi David,
>> 
>> We're seeing another issue that we think is related to the recent static VP node allocation changes. Instrumented programs are hitting a segfault:
>> 
>> ```
>> 458.sjeng is fixed but 445.gobmk is broken.
>> LLVM Profile Warning: Unable to track new values: Running out of static counters.  Consider using option -mllvm -vp-counters-per-site=<n> to allocate more value profile counters at compile time.
>> /Users/buildslave/jenkins/workspace/Performance_ARM64_SPEC2006_INT-O3_LTO_PGO-master/spec2006/cur_run/nt/build/LNTBased/speccpu2006/int/445.gobmk/tools/timeit-target: error: child terminated by signal 11
>> ```
>> 
>> The backtrace is:
>> 
>> ```
>> * frame #0: 0x0000000100112f10 445.gobmk.simple`__llvm_profile_instrument_target + 132 at InstrProfilingValue.c:137
>>     frame #1: 0x000000010007d528 445.gobmk.simple`shapes_callback + 2352
>>     frame #2: 0x0000000100035b0c 445.gobmk.simple`matchpat_loop + 2088
>>     frame #3: 0x0000000100034b38 445.gobmk.simple`matchpat_goal_anchor + 1124
>>     frame #4: 0x000000010007cb0c 445.gobmk.simple`shapes + 384
>>     frame #5: 0x0000000100029e4c 445.gobmk.simple`do_genmove + 2344
>>     frame #6: 0x00000001000a5a48 445.gobmk.simple`gtp_gg_genmove + 216
>>     frame #7: 0x0000000100099a6c 445.gobmk.simple`gtp_main_loop + 660
>>     frame #8: 0x000000010009bd00 445.gobmk.simple`main + 7188
>> ```
>> 
>> In __llvm_profile_instrument_target(), it looks like we're failing the condition: CounterIndex < NumVSites.
>> 
>> Do you mind taking a look?
>> 
>> thanks,
>> vedant
>> 
>> 
>> > On May 23, 2016, at 12:29 PM, Xinliang David Li <davidxl at google.com> wrote:
>> >
>> > Fix is on the way.
>> >
>> > David
>> >
>> > On Mon, May 23, 2016 at 12:27 PM, Vedant Kumar <vsk at apple.com> wrote:
>> >
>> > > On May 23, 2016, at 12:19 PM, Xinliang David Li <davidxl at google.com> wrote:
>> > >
>> > >
>> > >
>> > > On Mon, May 23, 2016 at 12:15 PM, Vedant Kumar <vsk at apple.com> wrote:
>> > > Hi David,
>> > >
>> > > I think one of the SPEC2006 tests doesn't have enough statically-allocated VP nodes per site. We're seeing:
>> > >
>> > > > "Child terminated by signal 25" (SIGXFSZ) after:
>> > >
>> > > Is this related?
>> >
>> > I don't think so. I suspect that's happening because the device is overloaded with logging info.
>> >
>> >
>> > > >
>> > > > LLVM Profile Warning: Running out of nodes: site_0 at func_12822962448227433604, value=4295054468
>> > > > ...
>> > > > LLVM Profile Warning: Running out of nodes: site_0 at func_12822962448227433604, value=4295052980
>> > > > LLVM Profile Warning: Running out of nodes: site_0 at func_12822962/Users/buildslave/jenkins/workspace/Performance_ARM64_SPEC2006_INT-O3_LTO_PGO-master/spec2006/cur_run/nt/build/LNTBased/speccpu2006/int/458.sjeng
>> > >
>> > > It seems like the fix for now is to either tweak vp-counters-per-site for the test or to set -vp-static-alloc=false.
>> > >
>> > > In the long term, do you think it's worth adjusting vp-counters-per-site s.t we can run SPEC without modifications? If so, is SPEC the right testbed?
>> > >
>> > > Rong had helped collecting SPEC related statistics, but looks like there is something missing.  I will investigate. In the meantime, can you try the workaround?
>> >
>> > Yes, we'll do that.
>> >
>> > Here's all the logging info from our bot:
>> >
>> >
>> >
>> >
>> >
>> > thanks,
>> > vedant
>> >
>> >
>> > >
>> > > thanks,
>> > >
>> > > David
>> > >
>> > >
>> > > thanks,
>> > > vedant
>> > >
>> >
>> >
>> >
>> 
>> 
> 
> 