[PATCH] D18164: [tsan] Do not instrument reads/writes to instruction profile counters.

[Xinliang David Li via llvm-commits]

On Wed, Mar 16, 2016 at 1:21 AM, Dmitry Vyukov <dvyukov at google.com> wrote:

> dvyukov added a comment.
>
> Anna reported that there is in fact some degradation if we use monotonic
> atomics. I am fine if we ignore the counters during tsan instrumentation,



Fine with that too if that is a must, but what is wrong with test
configuration  modified to avoid this known limitation?  According to Anna,
the following configs are tested:

- coverage + tsan
- coverage + asan
..

If the motivation is to save build time (by merging intrumentations),

How about just

- coverage + asan
- tsan
-...

This is also faster.  There does not seem to be any value to repeatedly
enable coverage.



> but also file bugs on optimizer (there must be no performance difference).


DSE won't happen with monotonic used (as below) -- I am not sure if that is
a bug. The LangRef has a paragraph about Optimization guidance.

define void @foo() #0 {
  store atomic i32 1, i32* @g monotonic, align 4
  store atomic i32 2, i32* @g monotonic, align 4
  ret void
}

thanks,

David



> This whole discussion has little to do with profiler, we are discussing
> that llvm does not permit users to have fast and correct code at the same
> time.
>
> I don't that switch to atomic RMWs under tsan is a good idea: having
> precise counters only in tsan build does not make sense. That could be a
> separate option (if user wants precise counters), but it is orthogonal to
> this discussion.
>
>
> http://reviews.llvm.org/D18164
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160316/2c57eff3/attachment.html>