> -----Original Message----- > From: Segher Boessenkool [mailto:seg...@kernel.crashing.org] > Sent: Friday, March 13, 2020 7:50 AM > To: Yangfei (Felix) <felix.y...@huawei.com> > Cc: gcc-patches@gcc.gnu.org; Zhanghaijian (A) <z.zhanghaij...@huawei.com> > Subject: Re: [PATCH PR94026] combine missed opportunity to simplify > comparisons with zero > > Hi! > > Please Cc: me on combine patches; you sent it nine days ago, and I didn't see > it > until now.
OK. > On Wed, Mar 04, 2020 at 08:39:36AM +0000, Yangfei (Felix) wrote: > > This is a simple fix for PR94026. > > With this fix, combine will try make an extraction if we are in a equality > comparison and this is an AND > > with a constant which is power of two minus one. Shift here should be an > constant. For example, combine > > will transform (compare (and (lshiftrt x 8) 6) 0) to (compare > > (zero_extract > (x 2 9)) 0). > > Why is that a good thing? The reported test case is reduced from spec2017 541.leela_r. I have pasted original code snippet on the bugzilla. We found other compilers like aocc/llvm can catch this pattern and simplify it. > (There should be thorough tests on many archs, showing it helps on average, > and it doesn't regress anything. I can do that for you, but not right now). I only have aarch64 & x86_64 linux available and have tested this patch with spec17 on both platforms. No obvious improvement & regression witnessed. This is expected as only one instruction is reduced here. It's appreciated if this can be tested on other archs. > The code needs more comments, and the commit message should say what is > done and why you made those choices. > In general, we should have *fewer* zero_extract, not more. OK. I can add more comments & commit message if finally we are inclined to go with this patch. Thanks, Felix
