[BOLT][perf2bolt] Perf2bolt does not support perf.data data sampled by SimplePerf record

[qianwindz]

In ubuntu20.04 ,i use simpleperf record -e cs-etm:u -- ./etm_test_loop.But the perf.data can't use，and the following error occurred

vboxuser@autofdo:~/Desktop/build/bin$ ./perf2bolt -p perf.data -o perf.fdata --itrace=l64i1us ./Matrix
BOLT-INFO: shared object or position-independent executable detected
PERF2BOLT: Starting data aggregation job for perf.data
PERF2BOLT: No perf executable found!

[llvmbot]

@llvm/issue-subscribers-bolt

Author: None (qianwindz)

In ubuntu20.04 ,i use simpleperf record -e cs-etm:u -- ./etm_test_loop.But the perf.data can't use，and the following error occurred

vboxuser@autofdo:~/Desktop/build/bin$ ./perf2bolt -p perf.data -o perf.fdata --itrace=l64i1us ./Matrix
BOLT-INFO: shared object or position-independent executable detected
PERF2BOLT: Starting data aggregation job for perf.data
PERF2BOLT: No perf executable found!

[aaupov]

perf2bolt invokes perf script when processing perf.data. Please try installing perf tools (linux-tools-generic package on Ubuntu).

[aaupov]

Or try following simpleperf instructions for converting etm data:
https://android.googlesource.com/platform/system/extras/+/master/simpleperf/doc/collect_etm_data_for_autofdo.md#convert-etm-data-for-llvm_bolt-experiment

[paschalis-mpeis]

perf2bolt invokes perf script when processing perf.data. Please try installing perf tools (linux-tools-generic package on Ubuntu).

I believe installing perf tools won't help as it looks like the perf.data output of simpleperf is not compatible with perf.

But using pre-aggregation for ETM data should work (eg --pa flag to get a perf.fdatafile).

[qianwindz]

I tried the method you mentioned and found that simpleperf Inject - output bolt and perf2bolt - pa did work, but later on, the error in the picture still appeared, indicating a segmentation fault . What caused this error? Is there any solution? Looking forward to your reply.

[paschalis-mpeis]

Code refactoring has recently introduced AArch64MCSymbolizer:

• [BOLT][AArch64] Add symbolizer for AArch64 disassembler. NFCI #127969

along with some other PRs for related improvements. Unsure if this crash was already there though.

(cc @maksfb)

[qianwindz]

I found that this error may have been caused by me not splitting the data of a single binary library but using the entire file. After I split the data of libhwui, the previous error disappeared and another error appeared. Why is this error?The libhwui.data is the second picture.

[aaupov]

Regarding pre-aggregation: to my knowledge, there are no tools to convert perf.data into pre-aggregated format (out of the box). But if you can synthesize and extract brstacks, you should be able to aggregate them in a script and output the profile in a required format (B+F or better yet T format).

[qianwindz]

I tried to obtain the bolt data for etm_test-loop according to Google's instructions. When using etm data, perf2bolt seemed to work, but the same operation did not work on the so library. So what should I do if I need to optimize some SO libraries? We look forward to your reply.

[paschalis-mpeis]

The screenshots appear to be from your successful attempt on the binary.
Can you share what was the issue when trying with libraries?

[qianwindz]

@paschalis-mpeis Of course, when I optimize the SO library, there will be an error in perf2bolt as shown in the image below.

[paschalis-mpeis]

Great, thanks! Can you enable line numbers and report back the exact line that caused this crash?
ie reconfigure cmake amending the below flags, build bolt again, repeat the perf2bolt step, and paste here the relevant text.

-DCMAKE_C_FLAGS=-gline-tables-only
-DCMAKE_CXX_FLAGS=-gline-tables-only

BTW you may also copy-paste your output in textual format, wrapping it in triple backticks.

[qianwindz]

@paschalis-mpeis Hello, I'm not sure if this format will satisfy you.If it doesn't work, I will add those options and recompile.

[paschalis-mpeis]

Oh, apologies for the confusion. I'm interested in the actual source code line where the crash occurred in AArch64MCSymbolizer::adjustRelocation. Without the provided flags we only get a hex number in the stack trace, and there are manual steps on our end to figure that out.

If you could recompile with the provided flags then we'll see the exact line it crashed. Also by pasting out just the text, we can easily copy-paste relevant bits on our end.