Overview
Request 813322 superseded
Force profiling run onto a single CPU to make package build
reproducible even with LTO (boo#1040589)
note: maybe setarch -R to disable ASLR would not be needed, but it certainly does not hurt here.
- Created by bmwiedemann
- In state superseded
- Package maintainers: dirkmueller and tabraham1
- Superseded by 817958
Request History
bmwiedemann created request
Force profiling run onto a single CPU to make package build
reproducible even with LTO (boo#1040589)
note: maybe setarch -R to disable ASLR would not be needed, but it certainly does not hurt here.
Andreas_Schwab declined request
It already runs non-parallel
bmwiedemann reopened request
It is not running non-parallel enough. Not sure what exactly makes the difference, but with the patch I could get the same OBS build 5 times in a row, while without it it only happened randomly sometimes.
Andreas_Schwab declined request
Must be a bug in gcc then
bmwiedemann reopened request
I still think, this is not from a gcc bug and this change is the best we can do to make package build more deterministic. The make -j1 that was already there, is just serializing make, not the checks it runs.
Andreas_Schwab declined request
You are welcome to fix that properly, but this isn't.
bmwiedemann superseded request
superseded by 817958
@Andreas_Schwab, @dirkmueller, @tabraham1: review reminder
let's say, profile-guided-optimisations are very sensitive to variations in the profiling runs. e.g. if you inpipe | grep | grep | outpipe, there could be variations in order of how data gets passed around. This would not be a bug in gcc, would it?