On 03/05/2018 12:58 PM, Matthew Miller wrote:
On Mon, Mar 05, 2018 at 12:07:40PM +0100, Florian Weimer wrote:
Should we start compiling Fedora with auto-vectorization, either
using -O3 or -O2 -ftree-loop-vectorize?
What's the performance impact?
You mean, on build time? Difficult to tell. Simple things like Lua do
not show build time change with -O3.
We didn't observe a substantial reduction in overall build time when we
switched Fedora glibc from -O3 to -O2 because build time is mostly
dominated by testing anyway (and discounting testing, a lot of time is
spent in make).
For run-time impact, the performance changes you'd observe depend on
whether you pick a benchmark which happens to be vectorized or not.
What are the risks?
The usual. Different compiler bugs. Additional warnings (both true
positives and false positives) which break -Werror builds.
Furthermore, -O3 slightly changes ABI because it enables
-finline-functions, which can break symbol interposition. That's why
-O2 -ftree-loop-vectorize -ftree-slp-vectorize might be the better
choice, perhaps with other bits from -O3.
Thanks,
Florian
_______________________________________________
devel mailing list -- devel@xxxxxxxxxxxxxxxxxxxxxxx
To unsubscribe send an email to devel-leave@xxxxxxxxxxxxxxxxxxxxxxx
