[llvm-dev] Working on FP SCEV Analysis

Fri May 20 10:04:50 PDT 2016

To the best of my experience, handling case B (secondary induction) is must-have, and if I’m not mistaken,
people aren’t opposed to that.

For me, handling case A (primary induction) is “why not?”, but I certainly admit that that can be very naïve
thinking coming from lack of good understanding on SCEV and their proper usages. Now, let’s assume we
can postpone discussion about  case A. What is the best approach to handle case B? Let me summarize
the discussion so far. Hope I didn’t miss anything.

1)
Extend SCEV was the initial approach taken by Elena.
                Elena thinks this solution ”looks very structured”.
                If I’m not mistaken, some people think this is overkill and overly complicates already complicated SCEV.
                Anyone care to look at the patch Elena came up with?
2)
IndVarSimplify::handleFloatingPointIV  (mentioned by Andy)
                This transforms integer-valued FP (primary) IV into integer IV and convert.
                Chandler says most of Graphics shading language use case mentioned by Owen
                should be handled here.
                It certainly has the logic of detecting FP induction, but Andy punted discussions
                on non-integer valued IV issues to MichaelZ and Adam.
3)
extend InductionDescriptor::isInductionPHI in the vectorizer to directly analyze the PHIs without SCEV support
(mentioned by Adam)
                If this is the standard way to deal with all secondary inductions, it certainly looks attractive.
                Elena, would you try doing this and compare with 1)?

Thanks,
Hideki

----------------------
I looked at what our FORTRAN compiler (non-LLVM) does for the following (x, y, z, and f are float by default typing,
based on names).

       subroutine foo(x, y, z, A, N)
       double precision A(N)
       do f = x, y, z
         A(f) = f
       enddo
       end

The frontend computes the trip count outside of the loop (per language rule) and uses an integer primary
induction variable (compile-generated) for loop control. So, (future) FORTRAN usage doesn’t seem to be
a good example for promoting support for case A (= FP primary induction variable).

From: Chandler Carruth [mailto:chandlerc at google.com]
Sent: Thursday, May 19, 2016 7:03 PM
To: Demikhovsky, Elena <elena.demikhovsky at intel.com>; anemet at apple.com; Sanjoy Das <sanjoy at playingwithpointers.com>
Cc: Saito, Hideki <hideki.saito at intel.com>; llvm-dev <llvm-dev at lists.llvm.org>; Andrew Trick <atrick at apple.com>
Subject: Re: [llvm-dev] Working on FP SCEV Analysis

On Thu, May 19, 2016 at 7:03 AM Demikhovsky, Elena <elena.demikhovsky at intel.com<mailto:elena.demikhovsky at intel.com>> wrote:
> One option would be to extend InductionDescriptor::isInductionPHI in the vectorizer to directly analyze the PHIs without SCEV support as Sanjoy suggested.  I *think* that that could be sufficient to handle case B.

I implemented this with FP SCEV and the code looks very structured, including SCEVExpander. Extending the existing structures without implementing FP SCEV will be problematic.
And my end goal is to handle case *A*.
Ok, but there have been *numerous* requests for an explanation of why this is important, and that hasn't emerged yet.

We really need to have a clear understanding of the relative importance of solving these problems in order to understand the best approach.

-           Elena

From: anemet at apple.com<mailto:anemet at apple.com> [mailto:anemet at apple.com<mailto:anemet at apple.com>]
Sent: Thursday, May 19, 2016 07:43
To: Sanjoy Das <sanjoy at playingwithpointers.com<mailto:sanjoy at playingwithpointers.com>>
Cc: Demikhovsky, Elena <elena.demikhovsky at intel.com<mailto:elena.demikhovsky at intel.com>>; Saito, Hideki <hideki.saito at intel.com<mailto:hideki.saito at intel.com>>; llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>>; Andrew Trick <atrick at apple.com<mailto:atrick at apple.com>>

Subject: Re: [llvm-dev] Working on FP SCEV Analysis

On May 18, 2016, at 12:17 PM, Sanjoy Das via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:

Demikhovsky, Elena wrote:
> Even then, I'd personally want to see further evidence of why the
correct solution is to model the floating point IV in SCEV rather than
find a more powerful way of converting the IV to an integer that models
> the non-integer values taken on by the IV. As an example, if the use
case is the following code with appropriate flags to relax IEEE
semantics so this looks like normal algabra etc:

> for (float f = 0.01f; f < 1.0f; f += 0.01f) ç **A**

...

> I'd rather see us cleverly turn it into:

> float f = 0.01f;

> for (int i = 1; i < 100; i += 1, f += 0.01f) ç **B**

I can later try to enhance IndVarSimplify::handleFloatingPointIV() in
order to convert**A** to **B**.

But **B** is exactly the case I’m starting from. The main IV “i” is
integer. The variable “f” is also considered as IV in this loop.

And this loop is not vectorized because “f” is floating point.

I don’t think that the case **B** is uncommon.

If B is the case we actually care about, I'd say changing SCEV to work with floating points is an overkill.  How would you expect an SCEVFAddExpr to help vectorize B, other than tell you what the initial value and the increment is (and these can be found with a simple value analysis)?

One option would be to extend InductionDescriptor::isInductionPHI in the vectorizer to directly analyze the PHIs without SCEV support as Sanjoy suggested.  I *think* that that could be sufficient to handle case B.

Then if we find other pressing cases to handle we can rethink the strategy.

Also the current diagnostics is pretty bad for Hideki’s testcase with TTT as float.  This is what we currently report with -Rpass-analysis=loop-vectorize:

/tmp/sss.c:3:6: remark: loop not vectorized: value that could not be
      identified as reduction is used outside the loop
      [-Rpass-analysis=loop-vectorize]

I have no clue why we say that the value is used outside the loop.  I think this should just say that we have a loop-variant value that we couldn’t identify either as an induction or as a reduction.

Adam

If we're interested in handling complex variants of A directly: computing trip counts, proving away predicates etc. without translating the loops to use integer IVs (perhaps because we can't legally do so), then I can see FP-SCEV as a reasonable implementation strategy, but it looks like the general consensus is that such cases are rare and generally not worth optimizing?

-- Sanjoy

-*/Elena/*

---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160520/5ba72094/attachment.html>