similar to: [LLVMdev] Surprising SCEV / Constant folding result

Hi, +CC Max, Serguei This looks like a textbook case of why caching is hard. We first call getZeroExtendExpr on %dec, and this call does end up returning an add rec. However, in the process of simplifying the zext, it also calls into isLoopBackedgeGuardedByCond which in turn calls getZeroExtendExpr(%dec) again. However, this second (recursive) time, we don't simplify the zext and cache a

Hi Sanjoy, SCEV is behaving inconsistently when forming SCEV for this zext instruction in the attached test case- %conv5 = zext i32 %dec to i64 If we request a SCEV for the instruction, it returns- (zext i32 {{-1,+,1}<nw><%for.body>,+,-1}<nw><%for.body7> to i64) This can be seen by invoking- $ opt -analyze -scalar-evolution inconsistent-scev-zext.ll But when computing

Hi Sanjoy, Thanks for investigating the issue! I am more interested in getting the underlying problem fixed rather than making this particular test case work. I think I have more test cases where this problem crops up. I would any day prefer consistent results over compile time savings which can lead to inconsistencies. These inconsistencies require significant developer time to analyze and fix

Hi Pankaj, On Sun, Feb 11, 2018 at 2:32 PM, Chawla, Pankaj <pankaj.chawla at intel.com> wrote: > Thanks for investigating the issue! > > I am more interested in getting the underlying problem fixed rather than making this particular test case work. I think I have more test cases where this problem crops up. I would any day prefer consistent results over compile time savings which

Hi, On Tue, Aug 8, 2017 at 12:58 PM, Friedman, Eli <efriedma at codeaurora.org> wrote: > Oh, I see... yes, we do stupid things involving mutating NoWrap flags after > a SCEV is created. (grep for setNoWrapFlags in ScalarEvolution.cpp.) That's really a compile time hack -- we defer some expensive tricks to prove nsw/nuw on an add recurrences to when we've been asked to

Be careful about oversimplifying signed integer comparisons -- integer arithmetic can easily overflow, so you cannot transform A > B to A - B > 0. The compare instructions in most processors do not simply subtract and test the most significant bit; they compute what the sign of the difference would be in extended precision. On Apr 17, 2010, at 1:00 PM, llvmdev-request at cs.uiuc.edu wrote:

Hi, i am playing the ScalarEvolution these days. i found the the ScalarEvolution will simply return a SCEVUnknow for a ICmpInst, so i think maybe great to add a new kind of SCEV to the ScalarEvolution framework. for example, if i run ScalarEvolution on the bc file generate from the following C source file: int f(int a, int b, int c, int d) { return (2 * a + 5 * c + 2) > (4 * d - 3*b

Hello all, I’m pretty new to LLVM. I'm writing a pass for loop optimization. I clone and rearrange loops, setting the cloned loop as the original loop’s parent. This can be done multiple times, until there is no more work to do. The trouble is, after the first time I do this, the cloned loop's SCEVs become unknown types when they should be AddRecExpr. If I re-run the whole pass on the

Hi Sanjoy, >> I'm a bit apprehensive of adding more caching to solve problems created by caching; but if there is no way out of adding another cache, how about adding a cache that maps SCEV expressions to their simplified versions? Then we could do something like: I may be wrong but I think caching is not an issue in itself, but caching in the presence of self-recursion is. >>

Hi Sanjoy, So what is the verdict on this issue? Thanks, Pankaj -----Original Message----- From: Chawla, Pankaj Sent: Monday, February 26, 2018 11:12 AM To: Sanjoy Das <sanjoy at playingwithpointers.com> Cc: Maxim Kazantsev <max.kazantsev at azul.com>; Serguei Katkov <serguei.katkov at azul.com>; llvm-dev at lists.llvm.org Subject: RE: [SCEV] Inconsistent SCEV formation for

On 8/8/2017 1:37 PM, Friedman, Eli wrote: > On 8/8/2017 10:22 AM, Geoff Berry via llvm-dev wrote: >> Hi all, >> >> I'm looking into resolving a FIXME in the LoopDataPrefetch (and FalkorMarkStridedAccesses) pass by marking both of these passes as preserving the ScalarEvolution analysis. Unfortunately, when this change is made, LSR will generate different code. One of the

Hi all, I'm looking into resolving a FIXME in the LoopDataPrefetch (and FalkorMarkStridedAccesses) pass by marking both of these passes as preserving the ScalarEvolution analysis. Unfortunately, when this change is made, LSR will generate different code. One of the root causes seems to be that SCEV will return different nsw/nuw flags for the same Value, depending on what order the

This sounds fine to me (and sorry for the delay!). -- Sanjoy On Mon, Mar 12, 2018 at 1:09 PM, Chawla, Pankaj <pankaj.chawla at intel.com> wrote: > Hi Sanjoy, > > So what is the verdict on this issue? > > Thanks, > Pankaj > > > -----Original Message----- > From: Chawla, Pankaj > Sent: Monday, February 26, 2018 11:12 AM > To: Sanjoy Das <sanjoy at

Hi Sanjoy, Thanks for the reply! Would it be possible for you to implement this? You know the codebase better than I do. Thanks, Pankaj -----Original Message----- From: Sanjoy Das [mailto:sanjoy at playingwithpointers.com] Sent: Tuesday, March 13, 2018 1:34 PM To: Chawla, Pankaj <pankaj.chawla at intel.com> Cc: Maxim Kazantsev <max.kazantsev at azul.com>; Serguei Katkov

LLVM 2.7's ScalarEvolution.cpp has this scary comment: // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can // lead to pointer expressions which cannot safely be expanded to GEPs, // because ScalarEvolution doesn't respect the GEP aliasing rules when // simplifying integer expressions. I think I understand what the comment is saying. If a GEP has

Thanks for the helpful reply! I see that we are trying to keep ScalarEvolution stable around instruction ordering. My suggestion would be to not restrict the fix by only recursing on the first operand. By "dominator logic" I meant that if all other 'cheap' checks fail, we should decide by walking the dominator tree to see which instruction's basic block is encountered

Hi Pankaj, On Tue, Mar 13, 2018 at 1:55 PM, Chawla, Pankaj <pankaj.chawla at intel.com> wrote: > Thanks for the reply! > Would it be possible for you to implement this? I don't have cycles for this right now, but if you file a bug I can give this a shot when I have time later. Even in the best case this will have to at least wait until end of April because I'm leaving for a

On Sep 9, 2010, at 1:48 PM, David Greene wrote: > LLVM 2.7's ScalarEvolution.cpp has this scary comment: > > // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can > // lead to pointer expressions which cannot safely be expanded to GEPs, > // because ScalarEvolution doesn't respect the GEP aliasing rules when > // simplifying integer

Hi Geoff, On Wed, Aug 9, 2017 at 8:58 AM, Geoff Berry <gberry at codeaurora.org> wrote: > On 8/8/2017 8:38 PM, Sanjoy Das wrote: >> >> Hi, >> >> On Tue, Aug 8, 2017 at 12:58 PM, Friedman, Eli <efriedma at codeaurora.org> >> wrote: >>> >>> Oh, I see... yes, we do stupid things involving mutating NoWrap flags >>> after

Thanks for the feedback, Sanjoy. > SCEV is fairly conservative around PHI nodes that aren't recurrences and aren't obviously equivalent to a min-max branch-phi idiom. Is that the limitation you're running into here? Yes, that's exactly the problem. The problematic PHI nodes (%bc.resume.val and %bc.resume.val1) aren't either recurrences or related to min-max idioms. I