HWPMC: Fix clang stack frame parsing
ClosedPublic
Actions

Authored by fabient on Aug 2 2016, 4:14 PM.

Details

Reviewers

emaste
andrew
imp
gnn
• ian

Commits

rS315270: Fix arm stack frame walking support:

Summary

clang generate stack frame differently than gcc.
The patch handle both compiler and enable stack frame when HWPMC is enabled.
If some userland program is using gcc the stack frame will not be captured correctly. A future possible evolution is to add a heuristic on failure to change from one mode to the other.

Test Plan

Without patch "pmcstat -n1 -S clock.hard -T" will not show any level above captured location.
With the patch chain is captured correctly.

Diff Detail

Repository

rS FreeBSD src repository - subversion

Lint

Lint Skipped

Unit

Tests Skipped

Event Timeline

fabient updated this revision to Diff 18961.Aug 2 2016, 4:14 PM

fabient retitled this revision from to HWPMC: Fix clang stack frame parsing.

fabient updated this object.

fabient edited the test plan for this revision. (Show Details)

fabient set the repository for this revision to rS FreeBSD src repository - subversion.

Herald added subscribers: gnn, andrew, imp. · View Herald TranscriptAug 2 2016, 4:14 PM

fabient added reviewers: gnn, imp, andrew, • ian.Aug 2 2016, 4:16 PM

We should be using the EABI unwinding function. The location of any given register on the EABI stack is compiler specific so we shouldn't hard code it.

sys/conf/Makefile.arm
54	Your source tree seems to be out of date. `WITHOUT_ARM_EABI` was removed from this file over 2 years ago.

In D7396#153730, @andrew wrote:

We should be using the EABI unwinding function. The location of any given register on the EABI stack is compiler specific so we shouldn't hard code it.

Yes but the EABI unwinding cost is not the same.
At the moment every platform use basic stack walk.
This patch only repair something broken since clang as default compiler.
https://llvm.org/bugs/show_bug.cgi?id=18505

As a second step it will be interesting to benchmark.