On 12/19/2016 11:56 AM, Jakub Jelinek wrote:
On Mon, Dec 19, 2016 at 11:46:24AM -0700, Jeff Law wrote:
But I don't see that as inherently blocking this patch. It's pointing out a
bad API interface. It's no different than when I added teh NULL pointer
dereference warnings a while ago -- we had the exact same kinds of problems.
The question is how many of them are there. We *know* this kind of thing is
going to happen. Again, at this point I don't see 78859 as inherently
meaning Martin's patch should be reverted.
profiledbootstrap is meant to be supported without --disable-werror, has
been working that way for at least last 10 years.
But we also have to twiddle things to deal with profiledbootstrap
selecting different jump threads to realize and as a result we get
different Wuninitialized warnings. This happens all the time
And so is normal
bootstrap on powerpc* or hppa*.
Agreed, but we need to investigate those at a level deep enough to
understand the real problem.
So broken bootstrap is a
strong reason for having to do something.
Certainly we have to do something. But that doesn't mean flat out
reversion and rejection of the patch. It means careful analysis of the
costs & benefits. Jumping to reversion & rejection because it triggers
some warnings in cases we don't like is too hasty.
By moving the warning earlier, we'll still warn for the most cases, but
won't warn in the more convoluted cases. We can perhaps work on it further
in GCC 8. If we keep it as is, I think most users will just -Wno-nonnull
as soon as they run into some warning that will be hard to figure out what
is going on. At that point they will not get warnings even for the obvious
cases that we used to warn. Look at how the Linux kernel folks disable most
of warnings even for smaller reasons.
But again, the user case is no different than other warnings that are
sensitive to optimizations.
My sense is that we should revert and table until gcc-8 where we can
evaluate the space more thoroughly.
Jeff
