[llvm-dev] RFC: Consider changing the semantics of 'fast' flag implying all fast-math-flags

Thu Nov 17 00:51:16 PST 2016

Those are all good points.  Your reassociation point in the context of inlining is particularly interesting.

FWIW, we also have a case where a customer wants '-fno-associative-math' to suppress reassociation under '-ffastmath'.  It would take me a while to find the specifics of the issue, but it was (if my memory is right) more of a real use-case.  (That is to say, the code that was "failing" due to reassociation didn't have an obvious fix like the reciprocal situation, here, other than to turn off fast-math.)  In fact, the request to suppress reassociation was the motivation for creating PR27372 in the first place (which eventually fed into this thread).  I have to say that on the reassociation point, my concern is that to really suppress that, we will have to suppress so much, that there will hardly be any point in using -ffast-math.

I'd say your comments here are very similar to what Nicolai said in another subthread of this discussion:

>> I'd be really curious to know if there is anybody who really needs arcp
>> without fp-contract=fast or vice versa, or who needs both of these but
>> not the X*log2(0.5*Y) transform you mentioned, and so on.[1]
>> ...
>> [1] One case I _can_ think of (and which may have been a reason for the
>> proliferation of flags in the first place) is somebody who enables fast
>> math, but then doesn't want their results to change when they update the
>> compiler and get a new set of optimizations. But IMO that's a use case
>> that should be explicitly rejected.

I think those are all really good points, and an argument can be made that when -ffast-math gives you results you don't want, then you just have to turn it off.  Essentially, the user can't "have his cake and eat it too".

All that said, I think we (the company I work for, Sony) will have to implement support for these switches.  It comes down to GCC has these switches (e.g., -fno-reciprocal-math and -fno-associative-math), and they do suppress the transformations for our customers.  They switch to Clang/LLVM, they use the same switches, and it doesn't "work".  So as a practical matter, I think we will support them.  Whether the LLVM community in general feels that that's required, is another question.  Until for your recent comments here, and Nicolai's comments above, I would have thought the answer was clearly yes.  But maybe that's not the case.

In summary, irrespective of any (subjective?) assessment of how legitimate a particular use-case is, do we want switches like:

    -ffast-math -fno-reciprocal-math
     -ffast-math -fno-associative-math

to work?

For me, the answer is yes, because I have multiple customers that tell me they really want to leave -ffast-math on, but they want to be able to disable these sub-categories.  I've been approaching this under the assumption that the answer is yes for the Clang/LLVM community in general.

Thanks,
-Warren

From: mehdi.amini at apple.com [mailto:mehdi.amini at apple.com]
Sent: Wednesday, November 16, 2016 10:46 PM
To: Ristow, Warren <warren.ristow at sony.com>
Cc: Kaylor, Andrew <andrew.kaylor at intel.com>; llvm-dev at lists.llvm.org; Nicolai Hähnle <nhaehnle at gmail.com>
Subject: Re: [llvm-dev] RFC: Consider changing the semantics of 'fast' flag implying all fast-math-flags

On Nov 16, 2016, at 10:04 PM, Ristow, Warren <warren.ristow at sony.com<mailto:warren.ristow at sony.com>> wrote:

> Can you elaborate what kind of runtime failure is the reciprocal transformation triggering?

Yes.  It was along the lines of:

    {
      float x = a / c;
      float y = b / c;

      if (y == 1.0f) {
        // do some processing for when 'b' and 'c' are equal
      } else {
        // do other processing
      }

      use(x, y);
    }

Of course they understood they could easily change this code once they understood the issue.

But the fact that it "failed" for non-edge-case values of 'c', they were worried.  As an example of the non-edge-case aspect, when 'c' is 41.0f (and so of course 'b' is 41.0f), intuitively they felt that this “would work precisely”, even with fast-math.  Once they understood more, they agreed this was reasonable with fast-math, but they had the underlying concern that if they encountered one case where 'num' and 'den' were equal (and non-edge-case), yet 'num / den' wasn't precisely 1.0f, then even if they fixed this situation where they encountered it, it might be lurking elsewhere in their code, and so they wanted to disable that transformation.

Thanks for elaborating.

I’d be reluctant to call this situation a real use-case though.
Is the the distinction on reciprocal really make sense here? This user can have the same “surprising" anywhere in their code-base with reassociation as well:

void foo (float a, float b) {
  float x = a - b;
  if (x == 0)
     … // only if a == b
}

That would sound totally reasonable, unless foo is inlined and reassociation would lead to a non-zero value for x even when a and b passed in to foo "if it wasn’t inlined" would be identical!

(Reminds me somehow of a client that was bitten by nnan: their assumption was that as long as they didn’t introduce NaN in the program everything was fine. However with fast-math some transformations were introducing NaN where there wasn’t before and propagating to other computation that were transformed under the assumption that no NaN would show up, it also turns out that making the code safe against NaN and efficient at the same time is hard, especially when the code itself it compiled with fast-math).

—
Mehdi

From: mehdi.amini at apple.com<mailto:mehdi.amini at apple.com> [mailto:mehdi.amini at apple.com]
Sent: Wednesday, November 16, 2016 7:11 PM
To: Ristow, Warren <warren.ristow at sony.com<mailto:warren.ristow at sony.com>>
Cc: Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>>; llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>; Nicolai Hähnle <nhaehnle at gmail.com<mailto:nhaehnle at gmail.com>>
Subject: Re: [llvm-dev] RFC: Consider changing the semantics of 'fast' flag implying all fast-math-flags

On Nov 16, 2016, at 6:22 PM, Ristow, Warren <warren.ristow at sony.com<mailto:warren.ristow at sony.com>> wrote:

> ... except that Warren’s proposal that started this discussion seems to imply that he
> has a use case that requires reciprocals to be turned off separately.

Just to close this loose end, yes I have a use case.

Specifically, we have a customer that turns on '‑ffast‑math', but was getting a runtime failure due to the reciprocal-transformation being done.

Can you elaborate what kind of runtime failure is the reciprocal transformation triggering?

—
Mehdi

They don't want turn off fast‑math because they like the performance improvement, and can live with the imprecision in most cases.  So they wanted to suppress just the reciprocal-transformation.  I intended to tell them the solution was simple: use '‑ffast‑math ‑fno‑reciprocal‑math'.  But on trying it myself, I ran into the issue here.

Thanks,
-Warren

From: Kaylor, Andrew [mailto:andrew.kaylor at intel.com]
Sent: Wednesday, November 16, 2016 4:13 PM
To: Mehdi Amini <mehdi.amini at apple.com<mailto:mehdi.amini at apple.com>>; Ristow, Warren <warren.ristow at sony.com<mailto:warren.ristow at sony.com>>; llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>; Nicolai Hähnle <nhaehnle at gmail.com<mailto:nhaehnle at gmail.com>>
Subject: RE: [llvm-dev] RFC: Consider changing the semantics of 'fast' flag implying all fast-math-flags

I don’t really like the idea of updating checks of UnsafeAlgebra() to depend on all of the other flags.  It seems like it would be preferable to look at each optimization and figure out which flags it actually requires.  I suspect that in many cases the “new” flag (i.e. allowing reassociation, etc.) will be what is actually needed anyway.

I would be inclined to agree with Niolai’s suggestion of combining all the flags related to value safety, except that Warren’s proposal that started this discussion seems to imply that he has a use case that requires reciprocals to be turned off separately.

-Andy

From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of Mehdi Amini via llvm-dev
Sent: Wednesday, November 16, 2016 8:55 AM
To: Ristow, Warren <warren.ristow at sony.com<mailto:warren.ristow at sony.com>>
Cc: llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] RFC: Consider changing the semantics of 'fast' flag implying all fast-math-flags

On Nov 15, 2016, at 11:59 PM, Ristow, Warren <warren.ristow at sony.com<mailto:warren.ristow at sony.com>> wrote:

Hi,

Thanks for the quick feedback.  I see your points, but I have a few questions/comments.  I'll start at the end of the previous post:

> ...
> I think these are valuable problems to solve, but you should tackle them piece by piece:
>
> 1) the clang part of overriding the individual FMF and emitting the right IR is the first thing to fix.
> 2) the backend is still using the global UnsafeFPMath and it should be killed.

I addressed this point (2) for the reciprocal aspect in the patch, but of course that wasn't useful without doing something about (1).

Regarding (1), over at https://reviews.llvm.org/D26708#596610, David made the same point that it should be done in Clang.  I can understand that, but I wonder whether having the concept of the 'fast' flag in the IR that implies all the other FMF makes sense?  I'm not seeing a good reason for it, but since this is very new to me, I can easily imagine I'm missing the big picture.

For example, in the LLVM IR (http://llvm.org/docs/LangRef.html#fast-math-flags) the fast-math flags 'nnan', 'ninf', 'nsz', 'arcp' and 'fast’ are defined.  Except for 'fast', each of these has a fairly specific definition of what they mean.  For example, for 'arcp':

    arcp => "Allow optimizations to use the reciprocal of an argument rather
             than perform division."

'fast' is unusual, in that it describes a fairly generic set of aggressive floating-point optimizations:

    fast => "Allow algebraically equivalent transformations that may dramatically
            change results in floating point (e.g. reassociate). This flag implies
            all the others."

Very loosely, 'fast' means "all the aggressive FP-transformations that are not controlled by one of the other 4, plus it implies all the other 4".  If for terminology, we call those additional aggressive optimizations 'aggr', then we have:

    'fast' == 'aggr' + 'nnan' + 'ninf' + 'nsz' + 'arcp'

So as I see it, if we want to disable only one of the other ones (like 'arcp', in my case), there isn't any way to express that with these IR flags defined this way.  In short, we cannot turn on all the flags besides 'arcp'.  To do that, what we want is that somehow for the Clang switches:

  '-ffast-math -fno-reciprocal-math'

to ultimately result in LLVM IR that has the following flags on in appropriate FP ops:

  'aggr' + 'nnan' + 'ninf' + ‘nsz'

Make sense, I missed that we can’t *subtract* from fast at the IR level.

I wouldn’t be opposed to have something along the line of “aggr”, but there is a tradeoff: some transformation may be harder to guard with this model.

Maybe that could be a starting point: changing the “UnsafeAlgebra” bit in the FMF to be “aggr” you mention and replace all the query to FastMathFlags::UnsafeAlgebra() to return true if all the bits are set in the Flags. This alone should be nothing more than a mechanical change I believe.
The important part is then auditing all the users of UnsafeAlgebra() in the middle end and check if they can be “downgraded” to aggr safely: i.e. if they don’t need aggr *and* another flag.

—
Mehdi

But I don't see a way to express 'aggr' in the IR.  We could do this, if we change the definition of the IR 'fast' flag to remove that sentence about implying all the others:

    fast => "Allow algebraically equivalent transformations that may dramatically
            change results in floating point (e.g. reassociate).

(If we do something like that, we may want to change the name from 'fast' to something else (like 'aggr'), to avoid tying it too closely to the concept of the '-ffast-math' switch.)

As an aside, I don't know if the "reassociate" example is the only other transformation that's allowed by 'fast' (I presume it isn't), but I think reassociation would be better expressed by a separate flag, which could then be controlled independently via '-f[no]-associative-math' switch.  Not having that flag exist separately in the FMF is the origin of PR27372.  But creating that flag and using it in the appropriate places would still run into these problems of 'fast' implying all the others, which would make it impossible to disable reassociation while leaving all the other FMF transformations enabled.

To ask a concrete question using the current definition of 'fast' (which includes enabling reassociation, as the LLVM IR documentation of FMF says), how can we express in the IR that reciprocal-transformations are not allowed, but reassociation is allowed?

So the bottom line is that I do see there are issues in Clang that are relevant.  But as long as 'fast' means "'aggr' plus all the other FMF transformations", I don't see how we can effectively disable a subset of those other FMF transformations (while leaving 'aggr' transformations, such as reassociation, enabled).  With that in mind, my patch took one step in having 'fast' no longer imply all the others.

Thanks,
-Warren

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161117/8500827b/attachment-0001.html>