I've attached the code. The code does not make direct use of unsafePerformIO. It uses QuickCheck, but I don't think, this is a QuickCheck bug. The used Eq-instance is the one for Float. I've only managed to reproduce this bug on 32-bit-linux with ghc-7.4.2 when compiling with -O2. (The code might seem a bit odd, but this is the most boiled down version I could come up with. Even removing the "module Main where" line changes the behaviour.) Cheers, Sönke On 07/10/2012 12:51 PM, malcolm.wallace wrote:

Also, it is more likely to be a buggy instance of Eq, than a real loss of referential transparency.

Regards, Malcolm

On Jul 10, 2012, at 11:49 AM, Christopher Done wrote:

...

Depends what the real offending code is. For example, if it contains unsafePerformIO then it's not a bug.

On 10 July 2012 12:42, Sönke Hahn mailto:shahn@cs.tu-berlin.de> wrote:

...

Hi!

I've discovered a strange bug that violates simple equational reasoning. Basically, something similar to this:

let a = f x in a == f x

evaluates to False.

I'd like to report this on ghc-trac, but I realised, that I don't know a good name for behaviour like this. Is there one? "Broken referential transparency", perhaps?

Thanks, Sönke

_______________________________________________ Glasgow-haskell-users mailing list Glasgow-haskell-users@haskell.org mailto:Glasgow-haskell-users@haskell.org http://www.haskell.org/mailman/listinfo/glasgow-haskell-users

_______________________________________________ Glasgow-haskell-users mailing list Glasgow-haskell-users@haskell.org mailto:Glasgow-haskell-users@haskell.org http://www.haskell.org/mailman/listinfo/glasgow-haskell-users

_______________________________________________ Glasgow-haskell-users mailing list Glasgow-haskell-users@haskell.org http://www.haskell.org/mailman/listinfo/glasgow-haskell-users

On 10/07/2012 12:21, Aleksey Khudyakov wrote:

On Tue, Jul 10, 2012 at 3:06 PM, Sönke Hahn wrote:

...

I've attached the code. The code does not make direct use of unsafePerformIO. It uses QuickCheck, but I don't think, this is a QuickCheck bug. The used Eq-instance is the one for Float.

I've only managed to reproduce this bug on 32-bit-linux with ghc-7.4.2 when compiling with -O2.

It's expected behaviour with floats. Calculations in FPU are done in maximul precision available. If one evaluation result is kept in registers and another has been moved to memory and rounded and move back to registers number will be not the same indeed.

In short. Never compare floating point number for equality unless you really know what are you doing.

I consider it a bug, because as the original poster pointed out it is a violation of referential transparency. What's more, it is *not* an inherent property of floating point arithmetic, because if the compiler is careful to do all the operations at the correct precision then you can get determinstic results. This is why GHC has the -fexcess-precision flag: you have to explicitly ask to break referential transparency. The bug is that the x86 native code generator behaves as if -fexcess-precision is always on. I seriously doubt that we'll ever fix this "bug": you can get correct behaviour by enabling -msse2, or using a 64-bit machine. I don't off-hand know what the LLVM backend does here, but I would guess that it has the same bug. Cheers, Simon

On 11/07/2012 08:36, Christian Maeder wrote:

Hi,

I think this bug is serious and should be turned into a ticket on http://hackage.haskell.org/trac/ghc/ Would you do so Sönke?

The abstraction of floats (Float or Double) is broken if equality considers (random and invisible) excess bits that are not part of the ordinary sign, exponent and fraction representation.

It should also hold: show f1 == show f2 => f1 == f2 and: read (show f) == f (apart from NaN)

Why do you "doubt that we'll ever fix this", Simon?

Several reasons: - the fix hurts performance badly, because you have to store floats into memory after every operation. (c.f. gcc's -ffloat-store option) - the fix is complicated - good workarounds exist (-msse2) - it is rarely a problem

What is the problem to disable -fexcess-precision or enable -msse2 (on most machines) by default?

-fexcess-precision cannot be disabled on x86 (that is the bug). -msse2 is not supported on all processors, so we can't enable it by default. Cheers, Simon

Cheers Christian

Am 10.07.2012 14:33, schrieb Simon Marlow:

...

On 10/07/2012 12:21, Aleksey Khudyakov wrote:

...

On Tue, Jul 10, 2012 at 3:06 PM, Sönke Hahn wrote:

...

I've attached the code. The code does not make direct use of unsafePerformIO. It uses QuickCheck, but I don't think, this is a QuickCheck bug. The used Eq-instance is the one for Float.

I've only managed to reproduce this bug on 32-bit-linux with ghc-7.4.2 when compiling with -O2.

It's expected behaviour with floats. Calculations in FPU are done in maximul precision available. If one evaluation result is kept in registers and another has been moved to memory and rounded and move back to registers number will be not the same indeed.

In short. Never compare floating point number for equality unless you really know what are you doing.

I consider it a bug, because as the original poster pointed out it is a violation of referential transparency. What's more, it is *not* an inherent property of floating point arithmetic, because if the compiler is careful to do all the operations at the correct precision then you can get determinstic results. This is why GHC has the -fexcess-precision flag: you have to explicitly ask to break referential transparency.

The bug is that the x86 native code generator behaves as if -fexcess-precision is always on. I seriously doubt that we'll ever fix this "bug": you can get correct behaviour by enabling -msse2, or using a 64-bit machine. I don't off-hand know what the LLVM backend does here, but I would guess that it has the same bug.

Cheers, Simon

_______________________________________________ Glasgow-haskell-users mailing list Glasgow-haskell-users@haskell.org http://www.haskell.org/mailman/listinfo/glasgow-haskell-users

On 11/07/2012 09:51, Christian Maeder wrote:

Am 11.07.2012 10:25, schrieb Simon Marlow:

...

On 11/07/2012 08:36, Christian Maeder wrote:

...

Hi,

I think this bug is serious and should be turned into a ticket on http://hackage.haskell.org/trac/ghc/ Would you do so Sönke?

The abstraction of floats (Float or Double) is broken if equality considers (random and invisible) excess bits that are not part of the ordinary sign, exponent and fraction representation.

It should also hold: show f1 == show f2 => f1 == f2 and: read (show f) == f (apart from NaN)

Why do you "doubt that we'll ever fix this", Simon?

Several reasons:

- the fix hurts performance badly, because you have to store floats into memory after every operation. (c.f. gcc's -ffloat-store option)

If we sacrifice correctness for performance then we should clearly document this!

I will document it in the User's Guide along with the other known bugs.

...

...

What is the problem to disable -fexcess-precision or enable -msse2 (on most machines) by default?

-fexcess-precision cannot be disabled on x86 (that is the bug).

-msse2 is not supported on all processors, so we can't enable it by default.

Can't "configure" find this out?

Configure will detect whether the machine you're building on supports -msse2, but not whether the machine that you will eventually *run* the code on does. For instance, when building GHC for distribution we have to assume that the target machine does not support SSE2, so all the libraries must be built without -msse2. Cheers, Simon

C.

...

Cheers, Simon

...

Cheers Christian

Am 10.07.2012 14:33, schrieb Simon Marlow:

...

On 10/07/2012 12:21, Aleksey Khudyakov wrote:

...

On Tue, Jul 10, 2012 at 3:06 PM, Sönke Hahn wrote:

...

I've attached the code. The code does not make direct use of unsafePerformIO. It uses QuickCheck, but I don't think, this is a QuickCheck bug. The used Eq-instance is the one for Float.

I've only managed to reproduce this bug on 32-bit-linux with ghc-7.4.2 when compiling with -O2.

It's expected behaviour with floats. Calculations in FPU are done in maximul precision available. If one evaluation result is kept in registers and another has been moved to memory and rounded and move back to registers number will be not the same indeed.

In short. Never compare floating point number for equality unless you really know what are you doing.

I consider it a bug, because as the original poster pointed out it is a violation of referential transparency. What's more, it is *not* an inherent property of floating point arithmetic, because if the compiler is careful to do all the operations at the correct precision then you can get determinstic results. This is why GHC has the -fexcess-precision flag: you have to explicitly ask to break referential transparency.

The bug is that the x86 native code generator behaves as if -fexcess-precision is always on. I seriously doubt that we'll ever fix this "bug": you can get correct behaviour by enabling -msse2, or using a 64-bit machine. I don't off-hand know what the LLVM backend does here, but I would guess that it has the same bug.

Cheers, Simon

_______________________________________________ Glasgow-haskell-users mailing list Glasgow-haskell-users@haskell.org http://www.haskell.org/mailman/listinfo/glasgow-haskell-users

On 11.07.2012 20:05, Tyson Whitehead wrote:

On July 11, 2012 04:51:50 Christian Maeder wrote:

...

Is it not enough to store floats into memory just before equality tests (or add rounding in the instance definitions of Float and Double in Eq and Ord)?

You have to be 100% consistent in how you do every operations in all cases otherwise different levels of rounding errors will creep into the results.

It isn't too hard to imagine a floating point expression getting inlined somewhere, and the compiler generating code to evalulate it all in registers. Intermediate operations will then be done to 80 bit precision.

Elsewhere, it doesn't get inlined and the compiler generates code to store intermediate results in memory. Intermediate operations will then be done to 32 bit precision. Different results will occur on the rounding boundaires.

Always storing and reloading after every operations will give you consistent results. I guess the other option is to disable inlining etc. or somehow ensure they are applied consistently in all use cases of an expression.

There are more possibilities. With optimizations turned on compiler may be able to squeeze everything into registers and do store/load operations otherwise. However I don't consider it a problem. If your result depends on rounding you are in deep trouble and it's quite likely result is garbage. Equality is a bit more complicated. It's rarely meaningful and if you do need exact equality you'd better to ensure that number is stored because exact behaviour depends on hardware, compiler and compiler flags.

On July 10, 2012 09:28:27 Christian Maeder wrote:

Am 10.07.2012 13:06, schrieb Sönke Hahn:

...

I've attached the code. The code does not make direct use of unsafePerformIO. It uses QuickCheck, but I don't think, this is a QuickCheck bug. The used Eq-instance is the one for Float.

The Eq-instance for floats is broken wrt NaN

Prelude> (0/0 :: Float) == 0/0 False

I do not know if you create NaN in your tests, though.

Would that really be broken though? NaN can arrise from many contexts (e.g., sqrt(-1)), so it would also not make much sense to return True. The IEEE standard actually defines a mutually exclusive fourth "unordered" state wrt to NaNs for comparisons (in addition to lesser, greater, and equal). I would like to suggest native floating point might be better modelled as "Maybe Float", with NaN being the builtin "Nothing", but leaves out Inf. Cheers! -Tyson

On 07/10/2012 12:51 PM, malcolm.wallace wrote:

Also, it is more likely to be a buggy instance of Eq, than a real loss of referential transparency.

Speaking of which... would it be remiss of me to mention the elephant in the room, namely the Eq instance for Float? AFAICT there is no possible way for a Float value to fulfill the Eq type class requirements, so why is it an instance? (I'm thinking of the "weird" Nan/Infinity behvaior primarily). Is there *really* such a huge amount of code out there that relies on an Eq (or Ord for that matter!) instance for Float? If so... shouldn't that code be fixed rather than being subtly buggy? Regards,

On 13.07.2012 14:18, Sönke Hahn wrote:

On 07/13/2012 03:12 AM, Bardur Arantsson wrote:

...

Speaking of which... would it be remiss of me to mention the elephant in the room, namely the Eq instance for Float?

AFAICT there is no possible way for a Float value to fulfill the Eq type class requirements, so why is it an instance? (I'm thinking of the "weird" Nan/Infinity behvaior primarily). Is there *really* such a huge amount of code out there that relies on an Eq (or Ord for that matter!) instance for Float? If so... shouldn't that code be fixed rather than being subtly buggy?

Num inherits from Eq, so Float couldn't have an instance for Num if we didn't have that Eq instance.

No more since GHC 7.4. But Eq is indeed superclass of Ord and it Ord is used a lot.

On 13.07.2012 19:27, Brandon Allbery wrote:

On Fri, Jul 13, 2012 at 7:29 AM, Aleksey Khudyakov mailto:alexey.skladnoy@gmail.com> wrote:

Num inherits from Eq, so Float couldn't have an instance for Num if we didn't have that Eq instance.

No more since GHC 7.4. But Eq is indeed superclass of Ord and it Ord is used a lot.

...but Float's Ord instance is even more dubious than its Eq.

What's wrong with it except for NaN? NaN surely breaks every law but otherwise it's OK.