De-pessimize ScalarEvolution's handling of nsw/nuw?

[Andrew Trick]

> On Dec 4, 2014, at 1:39 PM, Jingyue Wu <jingyue at google.com> wrote:
> 
> I was chasing down a performance regression on a CUDA benchmark compiled for the NVPTX64 backend, and found loop strength reduction is ineffective in the presence of sign extension. Here's a reduced test case:
> 
> void foo(float *input, int n) {
>   for (int i = -n; i != n; ++i) {                                                     
>     baz(input[i + 5]);                                                            
>   }
> }
> 
> I expect &input[i + 5] to be promoted to an indvar but it's not. 
> 
> The root cause of this misoptimization is that ScalarEvolution is pessimistic about tagging nsw/nuw to a SCEVAddExpr. This pessimization was introduced in http://llvm.org/viewvc/llvm-project?view=revision&revision=145367 <http://llvm.org/viewvc/llvm-project?view=revision&revision=145367>. According to the comments there (http://llvm.org/docs/doxygen/html/ScalarEvolution_8cpp_source.html#l04087 <http://llvm.org/docs/doxygen/html/ScalarEvolution_8cpp_source.html#l04087>), ScalarEvolution does not apply an instruction's nsw/nuw flags to the corresponding SCEV expression. In the above example, &input[i + 5] corresponds to SCEV expression input + 4 * sext(i + 5). In order to promote &input[i + 5] to an indvar, we need to at least prove (i + 5) does not sign overflow so that we can reassociate the expression to (input + 5) + 4 * sext(i) which can be represented as a SCEVAddRecExpr. However, because ScalarEvolution doesn't apply sext to (i + 5), it cannot distribute sext(i + 5) to sext(i) + 5, and is thus unable to identify &input[i + 5] as a potential indvar. 
> 
> Side note: this issue kicked in after my recent recent change that disables induction variable widening for the NVPTX64 backend. This issue used to be alleviated (if any) by induction variable widening because there wouldn't be any sext if index i is already 64-bit. 
> 
> I wonder if the fix which disables applying nsw/nuw is too conservative. The comments in the source code say that ScalarEvolution does not apply an instruction's nsw/nuw flags to the corresponding SCEV expression because another non-control-equivalent instruction without nsw/nuw can be mapped to the same expression. If that's the only case we worried about, is a better fix to be not mapping instructions only differ in nsw/nuw to the same SCEV expression? That can be done by adding the wrapping flag of a SCEVAddExpr expression to the folding set that serves as the index of this expression for SCEV look-up. 
> 
> I followed this idea, and tried a preliminary change (http://reviews.llvm.org/differential/diff/16942/ <http://reviews.llvm.org/differential/diff/16942/>). It works fine so far: no transformation tests failed; some analysis tests failed but the new results seem better instead of incorrect. I wonder if I was just lucky on not breaking tests or it is the right way to go. 

SCEV is a system for computing algabreic equivalence. Control dependence has no place there. SCEV not an IR. The basic premise is that equivalent subexpressions will cancel. I don't understand how SCEV can fundamentally work if expressions are tagged with any kind of infomation derived from the IR.

NSW/NUW flags currently don't change the way SCEV works because they are lazilly and conservativelly attached to the expression without affecting its identity. Even so, they still cause problems because they introduce nondeterminism w.r.t the order SCEVs are computed. They were just an expedient hack to support a corner case in C -- which turns out to be a case people care about. But we can't do much more with them within SCEV. Instead, the user of SCEV will need to reason about NSW/NUW. There has been a lot of discussion about that but I don't think anyone wants to take it on.

I don't think LSR has a problem with your example. The incoming IR will have a sign-extend in the loop because that's what you asked for. It will treat the sign-extend as a use of the induction variable. The expression looks like this:

  LSR Use: Kind=Basic, Offsets={0}, widest fixup type: i32
    reg({(5 + (-1 * %n)),+,1}<nw><%for.body>)

I take it that you want LSR to convert this from an array index into a pointer increment, like this:

  LSR Use: Kind=Address of float, Offsets={0}, widest fixup type: float*
    reg({(20 + (4 * (sext i32 (-1 * %n) to i64)) + %input),+,4}<%for.body>)

To do that, you'll need to perform sign-extend elimination in indvars! I did think the NVPTX fix to disable IV widening was heavy-handed, but it's not my platform. It would certainly be easy to allow widening of simple cases of array[i+c]. I offered some suggestions in PR21148 I think.

-Andy


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20141215/bc187d22/attachment.html>