Message ID | 20180514153713.GA4320@redhat.com |
---|---|
State | New |
Headers | show |
Series | Add __attribute__((malloc) to allocator and remove unused code | expand |
On Mon, 14 May 2018, Jonathan Wakely wrote: > As discussed at https://gcc.gnu.org/ml/libstdc++/2018-01/msg00073.html > we can simplify the allocator function for valarray memory. I also > noticed that the _Array(size_t) constructor is never used. > > * include/bits/valarray_array.h (__valarray_get_memory): Remove. > (__valarray_get_storage): Call operator new directly. Remove ignored > top-level restrict qualifier and add malloc attribute instead. I am trying to understand the point of adding this attribute. The function is just { return static_cast<_Tp*>(operator new(__n * sizeof(_Tp))); } The idea is that it isn't safe (? see PR 23383) to mark operator new with the attribute, but it is safe for this particular use? When optimizing, I certainly hope this trivial function gets inlined, and then the attribute is lost (should the inliner add 'restrict' when inlining a function with attribute malloc?) and all that matters is operator new.
On 17/05/18 12:54 +0200, Marc Glisse wrote: >On Mon, 14 May 2018, Jonathan Wakely wrote: > >>As discussed at https://gcc.gnu.org/ml/libstdc++/2018-01/msg00073.html >>we can simplify the allocator function for valarray memory. I also >>noticed that the _Array(size_t) constructor is never used. >> >> * include/bits/valarray_array.h (__valarray_get_memory): Remove. >> (__valarray_get_storage): Call operator new directly. Remove ignored >> top-level restrict qualifier and add malloc attribute instead. > >I am trying to understand the point of adding this attribute. The >function is just > >{ return static_cast<_Tp*>(operator new(__n * sizeof(_Tp))); } > >The idea is that it isn't safe (? see PR 23383) to mark operator new >with the attribute, but it is safe for this particular use? I'd forgotten about that (I was assuming the compiler doesn't need to be told about the properties of operator new, because they're defined by the language). We can remove the attribute. >When optimizing, I certainly hope this trivial function gets inlined, >and then the attribute is lost (should the inliner add 'restrict' when >inlining a function with attribute malloc?) and all that matters is >operator new.
On Thu, 17 May 2018, Jonathan Wakely wrote: > On 17/05/18 12:54 +0200, Marc Glisse wrote: >> On Mon, 14 May 2018, Jonathan Wakely wrote: >> >>> As discussed at https://gcc.gnu.org/ml/libstdc++/2018-01/msg00073.html >>> we can simplify the allocator function for valarray memory. I also >>> noticed that the _Array(size_t) constructor is never used. >>> >>> * include/bits/valarray_array.h (__valarray_get_memory): Remove. >>> (__valarray_get_storage): Call operator new directly. Remove ignored >>> top-level restrict qualifier and add malloc attribute instead. >> >> I am trying to understand the point of adding this attribute. The function >> is just >> >> { return static_cast<_Tp*>(operator new(__n * sizeof(_Tp))); } >> >> The idea is that it isn't safe (? see PR 23383) to mark operator new with >> the attribute, but it is safe for this particular use? > > I'd forgotten about that (I was assuming the compiler doesn't need to > be told about the properties of operator new, because they're defined > by the language). We can remove the attribute. I am not necessarily asking to remove it. I don't have a good understanding of what would break if we marked operator new with the attribute, so I have no idea if those reasons also apply for this use in valarray. >> When optimizing, I certainly hope this trivial function gets inlined, and >> then the attribute is lost (should the inliner add 'restrict' when inlining >> a function with attribute malloc?) and all that matters is operator new. If we determine that using the attribute here but not on operator new is the right choice, then I believe we need some middle-end tweaks so it isn't ignored.
On Thu, May 17, 2018 at 1:14 PM Marc Glisse <marc.glisse@inria.fr> wrote: > On Thu, 17 May 2018, Jonathan Wakely wrote: > > On 17/05/18 12:54 +0200, Marc Glisse wrote: > >> On Mon, 14 May 2018, Jonathan Wakely wrote: > >> > >>> As discussed at https://gcc.gnu.org/ml/libstdc++/2018-01/msg00073.html > >>> we can simplify the allocator function for valarray memory. I also > >>> noticed that the _Array(size_t) constructor is never used. > >>> > >>> * include/bits/valarray_array.h (__valarray_get_memory): Remove. > >>> (__valarray_get_storage): Call operator new directly. Remove ignored > >>> top-level restrict qualifier and add malloc attribute instead. > >> > >> I am trying to understand the point of adding this attribute. The function > >> is just > >> > >> { return static_cast<_Tp*>(operator new(__n * sizeof(_Tp))); } > >> > >> The idea is that it isn't safe (? see PR 23383) to mark operator new with > >> the attribute, but it is safe for this particular use? > > > > I'd forgotten about that (I was assuming the compiler doesn't need to > > be told about the properties of operator new, because they're defined > > by the language). We can remove the attribute. > I am not necessarily asking to remove it. I don't have a good > understanding of what would break if we marked operator new with the > attribute, so I have no idea if those reasons also apply for this use in > valarray. > >> When optimizing, I certainly hope this trivial function gets inlined, and > >> then the attribute is lost (should the inliner add 'restrict' when inlining > >> a function with attribute malloc?) and all that matters is operator new. > If we determine that using the attribute here but not on operator new is > the right choice, then I believe we need some middle-end tweaks so it > isn't ignored. We don't have a good way to do this. Your suggestion of adding restrict would work if it were not that only function-scope restrict uses are later handled... Richard. > -- > Marc Glisse
On Thu, 17 May 2018, Richard Biener wrote: > On Thu, May 17, 2018 at 1:14 PM Marc Glisse <marc.glisse@inria.fr> wrote: > >> On Thu, 17 May 2018, Jonathan Wakely wrote: > >>> On 17/05/18 12:54 +0200, Marc Glisse wrote: >>>> On Mon, 14 May 2018, Jonathan Wakely wrote: >>>> >>>>> As discussed at https://gcc.gnu.org/ml/libstdc++/2018-01/msg00073.html >>>>> we can simplify the allocator function for valarray memory. I also >>>>> noticed that the _Array(size_t) constructor is never used. >>>>> >>>>> * include/bits/valarray_array.h (__valarray_get_memory): Remove. >>>>> (__valarray_get_storage): Call operator new directly. Remove > ignored >>>>> top-level restrict qualifier and add malloc attribute instead. >>>> >>>> I am trying to understand the point of adding this attribute. The > function >>>> is just >>>> >>>> { return static_cast<_Tp*>(operator new(__n * sizeof(_Tp))); } >>>> >>>> The idea is that it isn't safe (? see PR 23383) to mark operator new > with >>>> the attribute, but it is safe for this particular use? >>> >>> I'd forgotten about that (I was assuming the compiler doesn't need to >>> be told about the properties of operator new, because they're defined >>> by the language). We can remove the attribute. > >> I am not necessarily asking to remove it. I don't have a good >> understanding of what would break if we marked operator new with the >> attribute, so I have no idea if those reasons also apply for this use in >> valarray. > >>>> When optimizing, I certainly hope this trivial function gets inlined, > and >>>> then the attribute is lost (should the inliner add 'restrict' when > inlining >>>> a function with attribute malloc?) and all that matters is operator > new. > >> If we determine that using the attribute here but not on operator new is >> the right choice, then I believe we need some middle-end tweaks so it >> isn't ignored. > > We don't have a good way to do this. Your suggestion of adding restrict > would work if it were not that only function-scope restrict uses are later > handled... This seems extremely similar to the issue of inlining functions with restrict arguments. I have written a PR, but it is probably not worth submitting.
On Thu, May 17, 2018 at 1:42 PM Marc Glisse <marc.glisse@inria.fr> wrote: > On Thu, 17 May 2018, Richard Biener wrote: > > On Thu, May 17, 2018 at 1:14 PM Marc Glisse <marc.glisse@inria.fr> wrote: > > > >> On Thu, 17 May 2018, Jonathan Wakely wrote: > > > >>> On 17/05/18 12:54 +0200, Marc Glisse wrote: > >>>> On Mon, 14 May 2018, Jonathan Wakely wrote: > >>>> > >>>>> As discussed at https://gcc.gnu.org/ml/libstdc++/2018-01/msg00073.html > >>>>> we can simplify the allocator function for valarray memory. I also > >>>>> noticed that the _Array(size_t) constructor is never used. > >>>>> > >>>>> * include/bits/valarray_array.h (__valarray_get_memory): Remove. > >>>>> (__valarray_get_storage): Call operator new directly. Remove > > ignored > >>>>> top-level restrict qualifier and add malloc attribute instead. > >>>> > >>>> I am trying to understand the point of adding this attribute. The > > function > >>>> is just > >>>> > >>>> { return static_cast<_Tp*>(operator new(__n * sizeof(_Tp))); } > >>>> > >>>> The idea is that it isn't safe (? see PR 23383) to mark operator new > > with > >>>> the attribute, but it is safe for this particular use? > >>> > >>> I'd forgotten about that (I was assuming the compiler doesn't need to > >>> be told about the properties of operator new, because they're defined > >>> by the language). We can remove the attribute. > > > >> I am not necessarily asking to remove it. I don't have a good > >> understanding of what would break if we marked operator new with the > >> attribute, so I have no idea if those reasons also apply for this use in > >> valarray. > > > >>>> When optimizing, I certainly hope this trivial function gets inlined, > > and > >>>> then the attribute is lost (should the inliner add 'restrict' when > > inlining > >>>> a function with attribute malloc?) and all that matters is operator > > new. > > > >> If we determine that using the attribute here but not on operator new is > >> the right choice, then I believe we need some middle-end tweaks so it > >> isn't ignored. > > > > We don't have a good way to do this. Your suggestion of adding restrict > > would work if it were not that only function-scope restrict uses are later > > handled... > This seems extremely similar to the issue of inlining functions with > restrict arguments. That works to the extent that the effect of restrict is reflected in the memory references in the IL by PTA. But for our case there is no restrict qualified arguments but only return values. Richard. > I have written a PR, but it is probably not worth submitting. > -- > Marc Glisse
diff --git a/libstdc++-v3/include/bits/valarray_array.h b/libstdc++-v3/include/bits/valarray_array.h index 07f38ed03ed..6759d6003e9 100644 --- a/libstdc++-v3/include/bits/valarray_array.h +++ b/libstdc++-v3/include/bits/valarray_array.h @@ -47,18 +47,15 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION // Helper functions on raw pointers // - // We get memory by the old fashion way - inline void* - __valarray_get_memory(size_t __n) - { return operator new(__n); } + // We get memory the old fashioned way + template<typename _Tp> + _Tp* + __valarray_get_storage(size_t) __attribute__((__malloc__)); template<typename _Tp> - inline _Tp*__restrict__ + inline _Tp* __valarray_get_storage(size_t __n) - { - return static_cast<_Tp*__restrict__> - (std::__valarray_get_memory(__n * sizeof(_Tp))); - } + { return static_cast<_Tp*>(operator new(__n * sizeof(_Tp))); } // Return memory to the system inline void @@ -410,7 +407,6 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION template<typename _Tp> struct _Array { - explicit _Array(size_t); explicit _Array(_Tp* const __restrict__); explicit _Array(const valarray<_Tp>&); _Array(const _Tp* __restrict__, size_t); @@ -503,12 +499,6 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION __dst._M_data, __j._M_data); } - template<typename _Tp> - inline - _Array<_Tp>::_Array(size_t __n) - : _M_data(__valarray_get_storage<_Tp>(__n)) - { std::__valarray_default_construct(_M_data, _M_data + __n); } - template<typename _Tp> inline _Array<_Tp>::_Array(_Tp* const __restrict__ __p)