mehdi_amini added a comment.
In https://reviews.llvm.org/D28867#649760, @vsk wrote:
> Ah ok, so it sounds like a better approach would be to split the missing
> record message into a separate off-by-default warning.
I'm in favor of this.
https://reviews.llvm.org/D28867
vsk added a comment.
Ah ok, so it sounds like a better approach would be to split the missing record
message into a separate off-by-default warning. I don't have the time to update
this diff this week, but will shoot for the next.
https://reviews.llvm.org/D28867
xur added a comment.
This change gonna hide the missing profile for a function permanently if there
is no mismatch. Like, if a user adds a new function without changing existing
functions, he will never get a warning if using the old profiles.
In LLVM IR level PGO, we also find the missing
vsk created this revision.
Clang warns that a profile is out-of-date if it can't find a profile
record for any function in a TU. This warning is now noisy because llvm
can dead-strip functions with profiling instrumentation.
Only emit the out-of-date warning if there is an actual record
