Error reading bolt data input file: line 1, column 43: expected single char for mispred bit about bolt HOT 13 CLOSED

It could be the new format of perf output. What is the perf version that your are running and on what hardware? Ideally I'd like to see the output of perf script -F pid,brstack ..., or at least at a part of it that was causing the trouble.

Also, our diagnostics was broken. Could you please pull the latest changes and re-run the command?

from bolt.

./perf --version
perf version 3.10.0-693.5.2.el7.x86_64.debug
The hardware is Intel Xeon Phi 7250. The hardware i am running perf2bolt, though is Intel(R) Xeon(R) CPU E5-2660 v2 @ 2.20GHz, same kernel and OS version, Centos 7.

Will respond your other questions later.
Thanks.

from bolt.

With the fixed diagnostics it looks like this:

Error reading bolt data input file: line 1, column 43: expected single char for mispred bit
Found: -
PERF2BOLT: Failed to parse samples

from bolt.

sudo perf script -f -F pid,brstack
First few lines of output:

 7466  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0 
 7466  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0 
 7466  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0 
 7466  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0  0xffffffff816ac32c/0x7f2c9cf100fe/-/-/-/0

from bolt.

For the perf issue it's an easy fix, assuming that otherwise the output format is the same. If you can upgrade to perf version 4.5 or later it might be easier to guarantee that. The more serious issue I see is with __intel_avx_rep_memcpy and __intel_new_memcpy functions referenced in the log. We'll need to add a support for jump tables / data in code, otherwise we cannot process them in relocation mode. The workaround is to use -relocs=0, and it is not good for performance.

from bolt.

Grabbed perf from kernel 4.5:
by git checkout v4.5 and compiled it.
$ perf --version
perf version 4.5.gb562e4

Collected traces :
./perf record -e cycles:u -j any,u -o perf.data -- (...)

Then on a different machine after copying the perf.data:
sudo perf script -F pid,brstack -f

172359  0xffffffff816ac32c/0x7fc9b763c0fe/-/-/-/0 
172359  0xffffffff816ac32c/0x7fc9b763c0fe/-/-/-/0  0xffffffff816ac32c/0x7fc9b763c0fe/-/-/-/0 
172359  0xffffffff816ac32c/0x7fc9b763c0fe/-/-/-/0  0xffffffff816ac32c/0x7fc9b763c0fe/-/-/-/0  0xffffffff816ac32c/0x7fc9b763c0fe/-/-/-/0 

I have also tried perf version 4.14.rc5.g7c584a with the same results.

from bolt.

I guess the problem here is not in perf version but possibly perf on KNL not supporting the correct hw event.
From:
https://gitlab.imag.fr/kaunetem/linux-kaunetem/commit/dc323ce8e72d6d1beb9af9bbd29c4d55ce3d7fb0
M/P/-: M=branch target mispredicted or branch direction was mispredicted, P=target predicted or direction predicted, -=not supported

I will have a look how perf gets it.

from bolt.

Yep this one is a KNL issue. On other CPU i got:
0xffffffff816ac32c/0x7f7b322365a0/P/-/-/0

from bolt.

Interesting. We only use misprediction bit for ICP optimization, so it should be safe to ignore it. I'll add a patch. Thank you for the investigation.

from bolt.

Great, thanks! So how about the intel memcpy (and others) replacements?
Is there anything i can do to help? Should a new issue be raised?

from bolt.

Is there any way to force intel compiler to use libc functions? It justifies the new issue in any case.

from bolt.

Looks good, closing:

PERF2BOLT: Waiting for perf tasks collection to finish...
PERF2BOLT: Parsing perf-script tasks output
PERF2BOLT: Input binary is associated with 1 PID(s)
PERF2BOLT: Waiting for perf events collection to finish...
PERF2BOLT: Aggregating branch events...
PERF2BOLT-WARNING: misprediction bit is missing in profile

from bolt.

Is there any way to force intel compiler to use libc functions?

-ffreestanding seems to cut it.
https://software.intel.com/en-us/forums/intel-c-compiler/topic/306063

from bolt.