On Mon, 26 Jan 2015, Jakub Jelinek wrote: > On Mon, Jan 26, 2015 at 04:06:11PM +0100, Richard Biener wrote: > > > > This is the 2nd thing I came up with after looking at PR64277. > > VRP does a poor job computing value-ranges of unrolled loop IVs > > thus a very simple thing to do is to factor in previous VRP results > > by intersecting what VRP2 computes with recorded SSA name range infos > > (that also makes errors in those more likely to pop up... :/). > > > > This reduces the number of array bound warnings I get from the testcase > > but doesn't fix it completely. It also ends up with saner value-ranges. > > > > Bootstrapped and tested on x86_64-unknown-linux-gnu (with one > > libstdc++ runtime failure which I am now checking if caused by the patch, > > and thus likely an existing latent issue with SSA name range info). > > > > Ok for trunk? Or should I delay this to GCC 6? > > Does this work even without the other patch?
Yes, I've actually developed 2/2 first. The other patch only ever emits more warnings... > What do you think about Ilya's patch to set TREE_NO_WARNING in the unrolled > iterations where we had to keep the exit check? I don't like it too much - it papers over the real issue and prevents valid warnings from being emitted. Maybe we can set TREE_NO_WARNING on the last iteration (that is the only one we usually end up warning on - like after this patch). There are also many dups - regressions in 4.8 I belive where we warn for the last iteration in an unrolled loop. Richard.
