|
@Jonathan_Blow | |||||
|
So I am using ThreadSanitizer, and I find myself cleaning up data races that don't matter, in order to keep the output clean. (For example, an atomic set of a flag, where another thread reads that flag word, but doesn't care about that flag). On the one hand,
|
||||||
|
||||||
|
Jonathan Blow
@Jonathan_Blow
|
2. velj |
|
this means the diagnostics will be clean from spurious data races, which helps us spot real races. But on the other hand, it can have negative performance impacts on the code, for example by increasing memory use as I introduce more flag words.
|
||
|
|
||
|
Jonathan Blow
@Jonathan_Blow
|
2. velj |
|
It reminds me a little bit of Warning Cleanup Theatre from the 1990s and 2000s (and probably today), where people insert a bunch of casts into their code and now there are no more warnings, so The Code Must Be Better Now With All These Casts. Except data races are more serious.
|
||
|
|
||
|
Jonathan Blow
@Jonathan_Blow
|
2. velj |
|
So I am doing it, but, it doesn't feel totally good to me.
P.S. What is the best practice for getting tsan to shut up if there is an actual data race that you intend to be there and is fine (e.g. a thread polls a location to see if a value shows up there), without
|
||
|
|
||
|
Jonathan Blow
@Jonathan_Blow
|
2. velj |
|
doing something that might cause you to miss a real problem later?
|
||
|
|
||
|
tbodt
@tblodt
|
2. velj |
|
you might actually be doing that atomic flag set wrong, I just tried a test program that atomically sets/clears different bits of a global variable and didn't get any noise from tsan. here's my test code which you can flip between atomic and non-atomic: gist.github.com/tbodt/d390c205…
|
||
|
|
||
|
Jonathan Blow
@Jonathan_Blow
|
2. velj |
|
Weird. It is just an atomic or, not a whole lot seems like it could happen incorrectly...
|
||
|
|
||
|
Matt Lim
@pencilflip
|
2. velj |
|
Afaik you can tell tsan to ignore entire functions/files (github.com/google/sanitiz…), but can’t distinguish between types of races
|
||
|
|
||