- Boost - lists.boost.org

Re: [boost] [review] hash functions
by brangdon＠cix.compulink.co.uk 16 Mar '05

16 Mar '05

In-Reply-To: <989aceac05031508512fb3bd43(a)mail.gmail.com> (reply-to:to:subject:in-reply-to:mime-version:content-type:content-transfe r-encoding:references; b=R/dgT5RcmFTM4l6wo1FsmJV0WmZRxxS4UwfDvrZfxiBXu3i07NjqdqXmY0k93EfeA2rzuv0U dyun6iZo4Mn86mrySznULAY97Xn/zgqN3tU49cmIB/EtvwNWQzyHhnrH/iKu8pfDQoo6swyqX4 AJ8ih0MJXLu4umRrhBw7YSD2E=) wrote (abridged): > > However, it yields hash values for int * different to the values for > > char *, so with the current design aims we can't use it. > > It also has the undesirable side-effect of making the hash value > depend on the machine endianness. The same set of integers would hash > to different values on (say) Sparc and Intel machines. True - another variant of the stability issue. -- Dave Harris, Nottingham, UK

1 0

Re: [boost] Re: [review] hash functions
by brangdon＠cix.compulink.co.uk 16 Mar '05

16 Mar '05

In-Reply-To: <d17nlh$8o3$1(a)sea.gmane.org> daniel(a)calamity.org.uk (Daniel James) wrote (abridged): > I'm glad that I'm not the only person who's worried about alignment. > But this is a problems for a cross-platform library. Either I'll have > to introduce a load of macros to tweak it on various platforms (I'll > need help there) or just accept it. We can probably find something which gives reasonable results across a wide range of platforms. Shuffling the bits in a hash_combine-like manner would help. Eg something like: template <typename T> size_t hash_value( const T *p ) { size_t value = reinterpret_cast<size_t>(p); return value + value / sizeof(T); } If there's an easy way to turn sizeof(T) into a power of 2, use it. In many cases it will already be a power of 2, and I suspect the performance difference isn't crucial anyway. Alternatively we could just use: return value + (value >> 3); which ought to be reasonable for most small objects. It may also be reasonable for large objects that are heap-allocated, depending on how the heap is implemented. Anyway, I think it is better to do something like this than to just return value. -- Dave Harris, Nottingham, UK

1 0

Re: [boost] Re: Re: [review] hash functions
by brangdon＠cix.compulink.co.uk 16 Mar '05

16 Mar '05

In-Reply-To: <d174ak$vp3$1(a)sea.gmane.org> nesotto(a)cs.auc.dk (Thorsten Ottosen) wrote (abridged): > are you saying we don't want to hash a Base* as a void*, but use > object properties instead? Sometimes. It's a reasonable thing to do. It's what my own app actually does. It shouldn't be the default, of course. Mainly I'm saying that equality tests sometimes need to be polymorphic, and when they are the corresponding hash_value needs to be as well. So sometimes we will be comparing hash values from different types. (Other people here have claimed that never happens.) -- Dave Harris, Nottingham, UK

1 0

No tests on como?
by Peter Dimov 16 Mar '05

16 Mar '05

We don't seem to be running any tests on Comeau C++, why? Not using the "reference compiler" for our testing doesn't make much sense. Has the "reference compiler" changed without me noticing? -- Peter Dimov http://www.pdimov.com

8 23

How to add examples ot the regression testing?
by Gennadiy Rozental 16 Mar '05

16 Mar '05

I want to make sure that all examples are successfully compilable. What is the best way to do this? Gennadiy

3 3

Re: Sockets: proposal for a library design
by Don G 16 Mar '05

16 Mar '05

Hi Scott, > Networking is inherently asynchronous. Initiation and acceptance > of connections and I/O may be disguised as synchronous but other > activity (e.g. "network down") cannot. Network APIs will often > defer such errors until the next reporting opportunity (e.g. next > I/O request) but this is a band-aid that is sometimes enough and > at other times downright misleading. I assume that something > going by the name of Boost.Net would be targeting a higher > sophistication. Amen. It took me some time to get used to the work around that one should always try to have a pending read in order to detect problems. In some cases that works, but it depends on the protocol whether or not you can read at the right time. > Proposing that all network communication is asynchronous does > not preclude any synchronous coding. As discussed in other > threads (i.e. RPC) a "low-level", asynchronous library may be > the "core" of a Boost.net while synchronous services may be > built on top. While this will work, in my experience there are efficiencies to be had if the synchronous methods aren't forced to create/store callbacks that simply signal condition variables. Of course, the gains will depend on the cost of the async mechanism in question. > I think an asynchronous, core Boost.Net should expose fundamental > network services. In the case of TCP, that would be the reliable, > async 2-way transfer of data blocks. Achieving a clean and > portable library at this level would be achievement enough. What, no UDP? :) > Dragging I/O streams into the "network discussion" at this > point is almost premature; it moves the focus up the network > stack. Again this doesn't preclude those who need to from > integrating with I/O streams at a later point. I agree. > $0.01 Hey! That was at least $0.03! Best, Don __________________________________ Do you Yahoo!? Make Yahoo! your home page http://www.yahoo.com/r/hs

1 0

Re: Sockets: proposal for a library design
by Don G 16 Mar '05

16 Mar '05

Hi Boris, My responses are below... >> I would suggest that a proposed Boost.Net library should focus >> on networking in a more abstract manner, and not be merely a >> socket encapsulation. IMHO, sockets are a terrible API and >> should be hidden as fast as possible. A socket wrapper class >> is appropriate, but I would not go to any trouble to change >> the socket API. Simply wrap the resources and provide portable >> access to the underlying system. The real end-user part of the >> Net library should not talk about sockets (unless one provides >> a method like get_native to return the socket handle, and that >> method is defined as non-portable). > > This reminds me of the layers requirement from > http://www.crystalclearsoftware.com/cgi-bin/boost_wiki/wiki.pl? > BoostSocket/SocketRequirements. > I agree that it would be nice to have some easy-to-use high-level > classes for concepts that are used in network applications again > and again. Yes; sockets are just a means to that end and not the end in themselves. As such, they need not receive much special design consideration because that is not where you want users to interact with the library. A robust socket layer may only encourage undesirable use. The socket layer that has involved in my own code consists of a single Socket class with: a destructor for cleanup, methods that handle the mapping of non-portable EFOOBAR codes to portable exception types or return values, const-correctness, default arguments in a few places, etc.. The network layer simply uses sockets to implement the abstraction, and is the only direct user of sockets. > I agree again. There shouldn't be any of the four I/O models > omitted. We will see as we proceed if supporting various I/O > models is easy to implement and to use at runtime (especially > switching between I/O models). The I/O model that troubles me is the select/poll model. That is the mechanism I have used to implement the other three and it is not readily apparent how to expose it to the user as well. I may be surprised, but I think most folks would not mind seeing that model go away. The async I/O model with callbacks (essentially made from inside the select/poll/whatever loop) would provide a robust replacement for that tedious body of code. >> As others have also suggested, I would agree that the net >> facilities not be tied up into I/O streams. It would be >> better to provide the network facilities and then layer >> on I/O streams for those that want it. > > Several people dislike the idea of a network library built > on I/O streams. What are the reasons for this? I can only speak for myself, but I just don't like I/O streams. They seem to have a lot of clunk and little value, IMHO. They may work for text protocols, but in my use of network communications, they don't feel right. For example, the output would probably need consistent flushing do get data out of the buffer and on to the wire. > Is this a request for a more low-level access to socket > functions like read(), write(), send() and recv()? Again, I don't view socket access as the goal, but essentially yes: access to the "primitives" is what I would want. For example, I would want to be able to differentiate EOF from ECONNRESET(?) from ETIMEOUT, etc.. > The only reason I see why I/O streams should not be used is > the lack of support for an asynchronous I/O model. The > network library would have to add functions for asynchronous > support Yes, the network layer has to define what asynchronous I/O looks like. Again, I think a function<> with an appropriate signature is the answer. If the users needs those calls to be queued/marshalled, that is not the problem of the network layer. > until one day I/O streams support an asynchronous > I/O model by default (btw, anyone working on this?). I don't see how I/O streams could be asynchronous because async file I/O is not portable. For example Windows 95/98/Me don't have it. I suspect that many other systems are similar. Even though Windows NT (and 2000, XP, 2K Server) have async file I/O, it is not something that select() can handle as it can on *nix. >> Usage examples: > > Thanks again! When the socket class hierarchy is more > stable I hope to move upwards to high-level classes > representing concepts from the network world you used > in your examples. Most welcome. If I recall, the actual number of calls to socket API's made in my entire code base was so few (probably around 10-20) that even the little time I had spent on the socket layer was probably overkill. >> One thing I think is essential when using a blocking model >> is cancellation. For example, I have a reset() method on >> Acceptor, Address, Datagram and Stream that can be called >> at any time to cause blocking methods to terminate with a >> cancel exception (obviously this must be from another >> thread). Ironically, these reset() methods meant I could >> not use blocking socket calls to implement my blocking >> methods - recv() cannot be canceled once it blocks (at >> least not portably)! > > I added this to the Wiki but I am not sure if a network > library should support it. If you have two threads in your > application and one thread is halted because of a blocking > call it might have been better to use a > nonblocking/asynchronous I/O model? Perhaps, but some produces are just best designed as synchronous. They can go along merrily producing data and waiting for it to be consumed before producing more. Eventually they will block, and if they cannot be canceled (forced to unblock), you have a problem. IMO, any mechanism that can block indefinitely really needs some way to cancel. This should not be confused with thread cancellation. In this situation, once a stream is canceled/reset, it is almost completely unusable (most methods will throw the cancel exception). Best, Don __________________________________ Do you Yahoo!? Make Yahoo! your home page http://www.yahoo.com/r/hs

1 0

boost::bind now with limited operator support
by Peter Dimov 15 Mar '05

15 Mar '05

boost::bind now supports operator! and the relational operators ==, !=, <, <=, >, >=. For the latter, the left operand must be a bind expression, the right operand can be a value, a placeholder, or another bind expression. Not exactly a Lambda, but good enough for many common tasks. -- Peter Dimov http://www.pdimov.com

1 0

Re: [boost] Re: [review] hash functions
by brangdon＠cix.compulink.co.uk 15 Mar '05

15 Mar '05

In-Reply-To: <d11prg$uvt$1(a)sea.gmane.org> daniel(a)calamity.org.uk (Daniel James) wrote (abridged): > > Firstly, I think we have some agreement that the signature: > > > > size_t hash_range( It first, It last ); > > > > is wrong > > I didn't get the impression of any agreement about that. > > > and there should be some way to provide an initial seed. > > Although, I agree with this. That's all I meant. I knew there wasn't much agreement over whether side effects are bad :-) My review was a lot more conservative than my comments in discussion. > > We have some agreement that using reinterpret_cast would be better. > > Yes, and this will be fixed. When I wrote it seemed there were still unresolved issues. In fact, I don't believe anyone has yet posted an acceptable hash_value for pointers. Presumably we want to shift out bits which are always 0 for alignment reasons. > It is proposed that the standard explicitly specifies the hash > functions (and I've been criticized for not doing that) but part > of the purpose of this library is to get real world experience > on what the specification should be. And change it, if necessary. OK. > I do intend to spend more time looking into alternatives in the future > (and will look into adding a constant to hash_combine very soon). OK. Although I suggested it, I'm not sure it's the best fix, as compared with using non-zero initial seeds. const size_t default_seed = 0x9e3779b9; // Or whatever. size_t hash_range( It first, It last, seed=default_seed ); A signature like that provides a hook to introduce variation between similar types if the author desires, and also allows ranges to be hashed as subranges with the output of one subrange feeding into the next. The default value allows the issue to be safely ignored by naive users. So it's a good idea anyway, and since it also fixes the problem with zeros maybe we don't need to fix that twice. This does leave the issue of other places where hash_combine is called. It would be sufficient if we could be sure that all used a non-zero initial seed, eg: size_t hash_value( std::pair<int,int> p ) { size_t hash = default_seed; combine_hash( hash, p.first ); combine_hash( hash, p.second ); return hash; } and the documentation would encourage that, but we can't really be sure of it. Another approach is to change all the primitive functions. For example: size_t hash_value( int i ) { return static_cast<unsigned int>(i) + default_seed; } Introducing the non-zero constant here is surely as efficient as in combine_hash, if not more so. Admittedly it doesn't eliminate the zero, it just moves it around, but it should massively reduce its dynamic frequency. (Actually it will eliminate the zero where it matters most, ie for types like bool, char and short which are smaller than size_t.) > Did you see Peter's original proposal? Yes, I had followed most of the links. > > For example, the proposal has no specialisation for chars at all. I > > can't tell if that is an oversight, or whether it is a design > > aim [...] > I don't think it's either. Peter? Peter has now said it is a design aim. That even you weren't sure illustrates my point :-) > When hash_range is called with a seed argument, it would be equivalent > to calling hash_combine for pair<int, int>. If I understand the principles being advocated here, pair< pair<int,int>, pair<int,int> > should have the same hash_value as int[4]. How do you achieve that without giving hash_value( pair ) an initial seed? ----- I find myself increasingly drawn to the idea of a hasher object. The main signature should be like: template <typename T> Hasher &hash( Hasher &hasher, const T &t ); with implementations looking like, eg: template <typename First, typename Second> Hasher &hash( Hasher &hasher, const pair<First,Second> &p ) { hash( hasher, first ); hash( hasher, second ); return hash; } Then hash_value, if we still need it, would be like: template <typename T> size_t hash_value( T &t ) { Hasher hasher; hash( hasher, t ); hasher.finalise(); return hasher.value(); } but would rarely be called directly. This solves the problem of the default value, allows us to store intermediate results bigger than size_t (which may be important for cryptographic-strength hashing), and the finalise() step provides a hook for the final mixing/avalanching which some algorithms use (eg Paul Hsieh's). > Perhaps it would be better spelt as: > > template <class It> > void hash_combine(std::size_t& seed, It begin, It end); Yes, I considered that. I wasn't sure whether overloading the name would help or hinder - hash_range and hash_combine are not substitutable for each other. > Well, we did discus the type of review before it started, and nobody > raised any objections. I don't object to quick reviews in principle (although if someone can only contribute, eg, on Sunday mornings, a 5-day period is going to leave them out completely). > > I am not an expert on hashing, but I have implemented my own hashing > > library. I have encountered practical problems due to different > > objects having the same hash values. > > Is this in C++? Yes. The main problem was zeros. My hash primitives produced zeros and my hash_combine turned zeros into more zeros. Just as with the initial boost proposal. > I would have thought that type safety would have helped avoid > such problems. This was specifically an issue with optional or alternate types. For example, I had something similar to: size_t MyType::hash_value() const { return use_a ? hash_value(a) : hash_value(b); } where a and b could be different types. We can avoid that mistake in library classes like boost:variant, but that doesn't prevent it recurring in user code. When I am dealing with types without much intrinsic variation, eg bools, I found it important to exploit such variation as I did have as much as possible. Hashing false and 0 to different values helped. > Thank you for reviewing the library and bringing up several good points. My pleasure. -- Dave Harris, Nottingham, UK

3 3

boost::filesystem in Windows 95
by Jorge Lodos 15 Mar '05

15 Mar '05

Hi We have a code base for Windows that compiles with common settings for all Windows platforms. When using boost::filesystem, this is no longer possible, due to the fact that Windows implementation in operations_posix_windows.cpp uses the API function GetFileAttributesExA, not available in Windows 95. GetFileAttributesExA could be replaced by FindFirstFile, which gives the same information and is available in all Windows platforms. This API is used in two different places, BOOST_FILESYSTEM_DECL bool _is_empty( const path & ph ) implementation and BOOST_FILESYSTEM_DECL boost::intmax_t file_size( const path & ph ) implementation. Using FindFirstFile instead of GetFileAttributesExA we would get: BOOST_FILESYSTEM_DECL bool _is_empty( const path & ph ) { # ifdef BOOST_POSIX ........ # else WIN32_FIND_DATA ; HANDLE hf = ::FindFirstFileA(ph.string().c_str(), &ffd); if (hf == INVALID_HANDLE_VALUE) boost::throw_exception( filesystem_error( "boost::filesystem::is_empty", ph, fs::detail::system_error_code() ) ); CloseHandle(hf); return ( ffd.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY ) ? is_empty_directory( ph ) :( !ffd.nFileSizeHigh && !ffd.nFileSizeLow ); # endif } BOOST_FILESYSTEM_DECL boost::intmax_t file_size( const path & ph ) { # ifdef BOOST_POSIX ....... # else // by now, intmax_t is 64-bits on all Windows compilers WIN32_FIND_DATA ; HANDLE hf = ::FindFirstFileA(ph.string().c_str(), &ffd); if (hf == INVALID_HANDLE_VALUE) boost::throw_exception( filesystem_error( "boost::filesystem::file_size", ph, fs::detail::system_error_code() ) ); CloseHandle(hf); if ( (ffd.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY) !=0 ) boost::throw_exception( filesystem_error( "boost::filesystem::file_size", ph, "invalid: is a directory", is_directory_error ) ); return (static_cast<boost::intmax_t>(ffd.nFileSizeHigh) << (sizeof(ffd.nFileSizeLow)*8)) + ffd.nFileSizeLow; # endif } Is there any chance this could be solved in boost::filesystem? Thank you for boost :-) Best regards Jorge

2 1