Hi,
On 2018-03-17 18:55:11 +0100, Tomas Vondra wrote:
> Not sure, but the backbranches seem to be working fine, and the commit
> that triggers the issue is from December 31.
Well, that added the test. Are you saying that if you execute similar
code on an older branch it doesn't fail?
> Anyway, I can confirm that the fix suggested by Tom does the trick
Could you try the attached patch, too?
I'm a bit afraid that we'll have to go a lot further if we can't make
icc do safe signed overflow in all cases. This case is easy enough to
fix, but we were lucky in a way to find it. Given that apparently we
know that ICC also optimizes based on overflow sematics it seems we
should really work towards making all of the backend safe with that :(.
- Andres