error C2995: 'void fst::Prune(fst::MutableFst<A> *,const fst::PruneOptions<Arc,ArcFilter> &)': function template has already been defined

[jtrmal]

When compiling fstscript I'll get a bunch of errors saying:

1>C:\Users\jtrmal\Documents\openfst\src\include\fst/prune.h(216): error C2064: term does not evaluate to a function taking 0 arguments
1>C:\Users\jtrmal\Documents\openfst\src\include\fst/prune.h(305): error C2064: term does not evaluate to a function taking 0 arguments
1>C:\Users\jtrmal\Documents\openfst\src\include\fst/prune.h(311): error C2995: 'void fst::Prune(const fst::Fst<Arc> &,fst::MutableFst<A> *,const fst::PruneOptions<Arc,ArcFilter> &)': function template has already been defined

I was playing with that and seems it's because of the enable_if magic.
After changing

typename std::enable_if<
              (Arc::Weight::Properties() & kPath) == kPath>::type * = nullptr

to

typename = std::enable_if<
              (Arc::Weight::Properties() & kPath) == kPath>::type

(as http://en.cppreference.com/w/cpp/types/enable_if describes)
I end up with the error

error C2995: 'void fst::Prune(const fst::Fst<Arc> &,fst::MutableFst<A> *,const fst::PruneOptions<Arc,ArcFilter> &)': function template has already been defined

which I think relates to "Notes" in here:
http://en.cppreference.com/w/cpp/types/enable_if

but from there, I have no idea if there is an easy way how to resolve it

[kkm000]

Let me try to repro. What is the cc file you are compiling? And VS 2015 or 2017?

[jtrmal]

arciterator-class.cc and I'm using MSVC 17 (Community Edition) I have found discussion here: https://stackoverflow.com/questions/26753393/what-is-wrong-with-this-use-of-stdenable-if that seems to make sense y.

…

On Mon, Nov 6, 2017 at 3:54 PM, Kirill Katsnelson ***@***.***> wrote: Let me try to repro. What is the cc file you are compiling? And VS 2015 or 2017? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AKisXyRgFxBpWdPJuvor9uOisDw6lehfks5sz3IigaJpZM4QT2jT> .

[kkm000]

Yep, I am reproducing it by simply compiling this file! Let's see if I may have any ideas too.

[kkm000]

This is the weird part:

error C2039: 'type': is not a member of 'std::enable_if<false,_Ty>'

It's like BUT OF COURSE it's not, that's the whole point of this template... Very strange.

[jtrmal]

changing the functions like this: ``` template <class Arc, class ArcFilter> typename std::enable_if<(Arc::Weight::Properties() & kPath) != kPath>::type Prune(const Fst<Arc> &, MutableFst<Arc> *ofst, const PruneOptions<Arc, ArcFilter> &) { FSTERROR() << "Prune: Weight needs to have the path property: " << Arc::Weight::Type(); ofst->SetProperties(kError, kError); } ``` worked partially -- while for some function the issue got resolved, I still have issues with ``` 1>C:\Users\jtrmal\Documents\openfst\src\include\fst/prune.h(308): error C2995: 'std::enable_if<,void>::type fst::Prune(const fst::Fst<Arc> &,fst::MutableFst<A> *,const fst::PruneOptions<Arc,ArcFilter> &)': function template has already been defined 1>C:\Users\jtrmal\Documents\openfst\src\include\fst/prune.h(215): note: see declaration of 'fst::Prune' ``` which is fairly confusing as it seems there that both the condition and the negation of the condition are fullfiled and with WeightCompare which I don't know how to deal with, as it returns bool ``` 1>C:\Users\jtrmal\Documents\openfst\src\include\fst/isomorphic.h(48): error C2995: 'bool fst::internal::WeightCompare(Weight,Weight,float,bool *)': function template has already been defined 1>C:\Users\jtrmal\Documents\openfst\src\include\fst/isomorphic.h(28): note: see declaration of 'fst::internal::WeightCompare' ```

…

On Mon, Nov 6, 2017 at 4:12 PM, Kirill Katsnelson ***@***.***> wrote: This is the weird part: error C2039: 'type': is not a member of 'std::enable_if<false,_Ty>' It's like BUT OF COURSE it's not, that's the whole point of this template... Very strange. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AKisXyNduNdLthhpUQ361B6_KLmi2z2nks5sz3YngaJpZM4QT2jT> .

[kkm000]

It looks like MSVC cannot digest that. Other compilers can.

But the deeper issue is why are they using enable_if; in this case it has no point.

I believe the best way to handle this is merge the two templates like

//   Plus(weight, Weight::One()) == Weight::One()
template <class Arc, class ArcFilter>
void Prune(MutableFst<Arc> *fst, const PruneOptions<Arc, ArcFilter> &opts) {
  using StateId = typename Arc::StateId;
  using Weight = typename Arc::Weight;
  using StateHeap = Heap<StateId, internal::PruneCompare<StateId, Weight>>;

  if ((Weight::Properties() & kPath) != kPath) {
    FSTERROR() << "Prune: Weight needs to have the path property: "
               << Arc::Weight::Type();
    fst->SetProperties(kError, kError);
    return;
  }

  auto ns = fst->NumStates();
  if (ns < 1) return;
  .  .  .

I do not understand why the code with enable_if is trying to help the compiler to compile one of the two code paths, but since the expression in the if() is pretty much constexpr (it should be, or it would not be allowed in enable_if anyway), the compiler must be totally insane to generate any code for the false branch, effectively producing the same template expansion.

Does my reasoning make sense, or am I missing something deeper?

[kkm000]

As for this error message from your second attempt. this is certainly a bug in the compiler:

C2995: 'std::enable_if<,void>::type fst::Prune(const fst::Fst<Arc> &,fst::MutableFst<A> *,const fst::PruneOptions<Arc,ArcFilter> &)': function template has already been defined

std::enable_if<,void> cannot happen after expansion. Especially given the expression is constexpr bool.

[jtrmal]

yes, IIRC that was something they had before but for some reason, they changed to enable_if semantics (perhaps saving a couple of instructions?. I wonder if that kPath test is really constexpr because it seems that it's both the condition and it's negation is evaluated as true, for which I don't see any than some deep compiler reasoning reasons. And additionally, why GCC doesn't mind? :) I'll dig for the previous version of that file and try to get the preivous version working... y.

…

On Mon, Nov 6, 2017 at 4:42 PM, Kirill Katsnelson ***@***.***> wrote: It looks like MSVC cannot digest that. Other compilers can. But the deeper issue is *why* are they using enable_if; in this case it has no point. I believe the best way to handle this is merge the two templates like // Plus(weight, Weight::One()) == Weight::One()template <class Arc, class ArcFilter>void Prune(MutableFst<Arc> *fst, const PruneOptions<Arc, ArcFilter> &opts) { using StateId = typename Arc::StateId; using Weight = typename Arc::Weight; using StateHeap = Heap<StateId, internal::PruneCompare<StateId, Weight>>; if ((Weight::Properties() & kPath) != kPath) { FSTERROR() << "Prune: Weight needs to have the path property: " << Arc::Weight::Type(); fst->SetProperties(kError, kError); return; } auto ns = fst->NumStates(); if (ns < 1) return; . . . I do not understand why the code with enable_if is trying to help the compiler to compile one of the two code paths, but since the expression in the if() is pretty much constexpr (it should be, or it would not be allowed in enable_if anyway), the compiler must be totally insane to generate any code for the false branch, effectively producing the same template expansion. Does my reasoning make sense, or am I missing something deeper? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AKisXwSMsNTwPsS1du3otOw5ae9BV9R-ks5sz30ogaJpZM4QT2jT> .

[kkm000]

I wonder if that kPath test is really constexpr because it seems that it's both the condition and it's negation is evaluated as true

If you look closer at the error message, they both are evaluated as empty! enable_if<,void>. Complete nonsense.

[jtrmal]

oh, I wasn't able to parse that function prototype before (the "emptiness" of the expression). Now I see. y.

…

On Mon, Nov 6, 2017 at 4:53 PM, Kirill Katsnelson ***@***.***> wrote: I wonder if that kPath test is really constexpr because it seems that it's both the condition and it's negation is evaluated as true If you look closer at the error message, they both are evaluated as empty! enable_if<,void>. Complete nonsense. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AKisX-ePYnHJHQ72bfojJvWueMUB7ssoks5sz3_jgaJpZM4QT2jT> .

[kkm000]

IIRC that was something they had before but for some reason, they changed to enable_if semantics (perhaps saving a couple of instructions?.

That cannot save any this way, if anything goes as it should. Essentially, a compiler always compiles if (true) expr1; else expr2 as just expr1;. And enable_if can be used only with constexpr, as it is evaluated at the template expansion time.

The only reason I could think is they worked around some bug in a certain compiler?

[kkm000]

I am getting this error in 3 headers from compiling this one file. Should not be a big patch. When you compile full fstscript, are there more than this?

[jtrmal]

I gave up when I had 5k errors, becuase the prune.h is included everywhere -- will keep working on it.

…

On Mon, Nov 6, 2017 at 5:01 PM, Kirill Katsnelson ***@***.***> wrote: I am getting this error in 3 headers from compiling this one file. Should not be a big patch. When you compile full fstscript, are there more than this? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AKisXxA0adPMFyksFytzjEYUCseKehF6ks5sz4GtgaJpZM4QT2jT> .

[kkm000]

Let me come up with a "minimal" patch, so that it would be easy to maintain.

[kkm000]

Looks like enable_if is used only in these 3 includes (the 4th use in util.h is rather valid).