This is the mail archive of the gcc-bugs@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[Bug tree-optimization/69702] [4.9/5/6 Regression] excessive stack usage with -fprofile-arcs, LIM store motion lacks a register pressure aware cost model


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69702

--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 10 Feb 2016, arnd at linaro dot org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69702
> 
> --- Comment #2 from Arnd Bergmann <arnd at linaro dot org> ---
> Thanks, I have now added -fno-tree-loop-im to the kernel gcov cflags, so files
> we profile will be built with that. I can confirm that it fixes all stack size
> warnings that show up with -fprofile-arcs, I found around a dozen of them in
> various parts of the kernel.
> 
> Are there any downsides to doing this for all compiler versions? I don't think
> we care much about a missed optimization when CGOV is used, and nobody ships
> that on production systems. I guess for older gcc releases it's not necessary
> to disable tree-loop-im, but it's not easy to make this conditional on the gcc
> version without major surgery in the kernel build system.

Downsides are only missed optimizations.  Esp. for profiling it was
requested that the couter updates were combined which only will happen
(possibly) when we apply store motion.

> I'm also adding -Wno-maybe-uninitialized now, because -fprofile-arcs also
> introduces countless "error: '...' may be used uninitialized in this function
> [-Werror=maybe-uninitialized]" warnings that I don't see without
> -fprofile-arcs, and those are all false positives. Let me know if I should open
> another bug report for those (I assume there isn't really much to do, based on
> my reading of https://gcc.gnu.org/wiki/Better_Uninitialized_Warnings).

I don't remember seeing such bug so yes please.

> I also found two other files in the kernel that doen't build correctly once
> -fprofile-arcs is enabled (with or without tree-loop-im), I'll report those as
> separate bugs.

Thanks.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]