On Tue, Nov 30, 2021 at 3:18 PM Andrew Pinski <pins...@gmail.com> wrote: > > On Tue, Nov 30, 2021 at 3:08 PM Navid Rahimi <navidrah...@microsoft.com> > wrote: > > > > Hi Andrew, > > > > Thanks for your detailed comment. There are two problem I wanted to discuss > > with you about: > > > > a) The optimization I have sent patch, does optimize variable length "<<" > > too(for example B0 << x, where x is variable). This [1] link shows the > > actual optimization and a link for the proof is included in the editor. > > > > b) I am unable to prove the optimization you are describing for > > non-constant length shift. You can take a look at the code example [2] and > > proof [3]. I am getting "Transformation doesn't verify!" when I do > > implement the optimization you mentioned for non-constant shift. > > > > The optimization you are describing only works for "(take: (t << 1) != 0) > > -> ((t & 0x7fffffff) != 0)" which only is provable and works for > > INTEGER_CST. > > No it works with non constants too: > t << y != 0 -> t & (-1u>>y) != 0 > > When y == 0, you have t != 0. > Which is exactly what you think it should be. > Which can be further reduced to t != 0 as y >= sizeof(t)*BITS_PER_UNIT > is undefined.
I filed PR 103509 for the above issue. Note it only works when comparing against 0. I also notice that LLVM, ICC and MSVC do not do the optimization either. But they (except for MSVC) do handle (-1u>>y) != 0 to be always true. Thanks, Andrew Pinski > > Thanks, > Andrew Pinski > > > > > > My understanding might be incorrect here, please don't hesitate to correct > > me. > > > > 1) https://compiler-explorer.com/z/r46znh4Tj > > 2) https://compiler-explorer.com/z/K1so39dbK > > 3) https://alive2.llvm.org/ce/z/-54zZv > > > > Best wishes, > > Navid. > > > > ________________________________________ > > From: Andrew Pinski <pins...@gmail.com> > > Sent: Tuesday, November 30, 2021 14:03 > > To: Navid Rahimi > > Cc: Navid Rahimi via Gcc-patches > > Subject: [EXTERNAL] Re: [PATCH] tree-optimization/98956 Optimizing out > > boolean left shift > > > > On Tue, Nov 30, 2021 at 8:35 AM Navid Rahimi via Gcc-patches > > <gcc-patches@gcc.gnu.org> wrote: > > > > > > Hi GCC community, > > > > > > This patch will add the missed pattern described in bug 98956 [1] to the > > > match.pd. The codegen and correctness proof for this pattern is here > > > [2,3] in case anyone is curious. Tested on x86_64 Linux. > > > > > > > A better way to optimize this is the following (which I describe in PR > > 64992): > > take: (t << 1) != 0; > > > > This should be transformed into: > > (t & 0x7fffffff) != 0 > > > > The rest will just fall out really. That is there is no reason to > > special case bool here. > > I have most of the patch except for creating the mask part which > > should be simple, I just did not want to look up the wi:: functions at > > the time I was writing it into the bug report. > > > > Thanks, > > Andrew Pinski > > > > > > > > > Tree-optimization/98956: > > > > > > Adding new optimization to match.pd: > > > * match.pd ((B0 << x) cmp 0) -> B0 cmp 0 : New > > > optimization. > > > * gcc.dg/tree-ssa/pr98956.c: testcase for this > > > optimization. > > > * gcc.dg/tree-ssa/pr98956-2.c: testcase for node with > > > side-effect. > > > > > > 1) > > > https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgcc.gnu.org%2Fbugzilla%2Fshow_bug.cgi%3Fid%3D98956&data=04%7C01%7Cnavidrahimi%40microsoft.com%7Cd83f36080fd94b563ab608d9b44d4d1f%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739066369079450%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=EO7zAIa9sux4JklTDeALImoX3Kcjqeug%2BssU0E%2Fp6mY%3D&reserved=0 > > > 2) > > > https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fcompiler-explorer.com%2Fz%2Fnj4PTrecW&data=04%7C01%7Cnavidrahimi%40microsoft.com%7Cd83f36080fd94b563ab608d9b44d4d1f%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739066369079450%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=GyivNuda31%2FPXJQQ4Z9tK2cFtj3N9YcvRdtM7rVkhHg%3D&reserved=0 > > > 3) > > > https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Falive2.llvm.org%2Fce%2Fz%2FjyJAoS&data=04%7C01%7Cnavidrahimi%40microsoft.com%7Cd83f36080fd94b563ab608d9b44d4d1f%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637739066369079450%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=esqOKjKS5JZDbNBmAi0Bwwk0JTTHzInQ2Lgeq%2BPHJ9w%3D&reserved=0 > > > > > > Best wishes, > > > Navid.
