configure result in wrong hash namespace/header with gcc 4.1.2

[GoogleCodeExporter]

What steps will reproduce the problem?
1. configure with gcc 4.1.2
2. configure script using hash in __gnu_cxx, but it's
   really in tr1

What is the expected output? What do you see instead?
using hash in tr1

What version of the product are you using? On what operating system?
1.9

Please provide any additional information below.

i think gcc 4.1 with patch level >= 1 provide hash function in
tr1 namespace, but m4/stl_hash.m4 only check for __GNUC_MINOR__ < 2.
change the test to this fix the problem:

   AC_TRY_COMPILE([#if defined(__GNUC__) && (__GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 1 && __GNUC_PATCHLEVEL__ < 1))

Original issue reported on code.google.com by james.fa...@gmail.com on 6 Dec 2010 at 10:56

[GoogleCodeExporter]

As the comment in m4/stl_hash.m4 says, "I've seen unexplainable unordered_map 
bugs with -O2 on older gcc's."  So I purposefully still use the __gnu_cxx hash, 
rather than the tr1 hash, with gcc 4.1.2.

Both hashes are available in 4.1.2, so I'm not sure exactly what the problem is 
that you're reporting.  Do you have code that doesn't work?

Original comment by csilv...@gmail.com on 6 Dec 2010 at 8:24

• Added labels: Priority-Medium, Type-Defect

[GoogleCodeExporter]

sorry, i did not give any attentions to the comment.

when i migrate to 1.9, it complains about __gnu_cxx. it seems our code's old 
version
sparsehash is using the tr1 namespace.
maybe someone manually edit the sparseconfig.h to force using tr1. :)

Original comment by james.fa...@gmail.com on 7 Dec 2010 at 3:29

[GoogleCodeExporter]

Hmm, what is the exact complaint you're seeing?  Can you give a specific 
example of a compile command you try to run, and the errors you get in response?

Original comment by csilv...@gmail.com on 7 Dec 2010 at 4:24

[GoogleCodeExporter]

here it is:
/usr/local/include/google/sparsehash/hashtable-common.h:65: error: 
‘operator()’ is not a member of ‘__gnu_cxx::hash<std::basic_string<char, 
std::char_traits<char>, std::allocator<char> > >’

our code does not provide a hash function for std::string in __gnu_cxx,
it just force sparsehash to use tr1 instead.

Original comment by james.fa...@gmail.com on 7 Dec 2010 at 6:37

[GoogleCodeExporter]

Ah, I see what the problem is.  __gnu_cxx::hash<> does not provide a default 
hash for string, though it does for char*.  tr1::hash<> is the other way 
around: it does provide hash<string>, but not hash<char*>.  The problem isn't 
the fact it's in __gnu_cxx, the problem is you're depending on the system to 
provide a hash function for you which it's not.

I'm closing this as NotABug since it's really not sparsehash's responsibility 
to deal with hash functions at all.  The approach you're using now, of forcing 
tr1::hash<>, is fine.  Another approach would be to just define your own hash 
function, so you don't have to worry about what the system provides.

Original comment by csilv...@gmail.com on 7 Dec 2010 at 6:56

• Changed state: Invalid

[GoogleCodeExporter]

/home/zhangchi/local/include/sparsehash/sparse_hash_map:249:75:   instantiated 
from 'google::sparse_hash_map<Key, T, HashFcn, EqualKey, Alloc>::iterator 
google::sparse_hash_map<Key, T, HashFcn, EqualKey, 
Alloc>::find(google::sparse_hash_map<Key, T, HashFcn, EqualKey, 
Alloc>::key_type&) [with Key = long long unsigned int, T = bool, HashFcn = 
__gnu_cxx::hash<long long unsigned int>, EqualKey = std::equal_to<long long 
unsigned int>, Alloc = google::libc_allocator_with_realloc<std::pair<const long 
long unsigned int, bool> >, google::sparse_hash_map<Key, T, HashFcn, EqualKey, 
Alloc>::iterator = google::sparse_hashtable_iterator<std::pair<const long long 
unsigned int, bool>, long long unsigned int, __gnu_cxx::hash<long long unsigned 
int>, google::sparse_hash_map<long long unsigned int, bool>::SelectKey, 
google::sparse_hash_map<long long unsigned int, bool>::SetKey, 
std::equal_to<long long unsigned int>, 
google::libc_allocator_with_realloc<std::pair<const long long unsigned int, 
bool> > >, google::sparse_hash_map<Key, T, HashFcn, EqualKey, Alloc>::key_type 
= long long unsigned int]'
GenericNodeMemWrapper.cpp:960:55:   instantiated from here
/home/zhangchi/local/include/sparsehash/internal/hashtable-common.h:250:62: 
error: 'operator()' is not a member of 
'google::sparsehash_internal::sh_hashtable_settings<long long unsigned int, 
__gnu_cxx::hash<long long unsigned int>, long unsigned int, 4>::hasher'

Original comment by hewm2...@gmail.com on 23 Jul 2013 at 5:59

[GoogleCodeExporter]

Sorry I am stumbling across this old issue but on our old servers we found that 
you can use boost to fix this by changing the config file to use boost headers. 
Change some of the values in src/config.h to be using something similar to this:

/* src/config.h.  Generated by configure.  */
/* src/config.h.in.  Generated from configure.ac by autoheader.  */
/* Namespace for Google classes */
#define GOOGLE_NAMESPACE ::google
/* the location of the header defining hash functions */
#define HASH_FUN_H <boost/functional/hash/hash.hpp>
/* the location of <unordered_map> or <hash_map> */
#define HASH_MAP_H <boost/tr1/unordered_map.hpp>
/* the namespace of the hash<> function */
#define HASH_NAMESPACE boost
/* the location of <unordered_set> or <hash_set> */
#define HASH_SET_H <boost/tr1/unordered_set.hpp>
/* Define to 1 if you have the <google/malloc_extension.h> header file. */
/* #undef HAVE_GOOGLE_MALLOC_EXTENSION_H */
/* define if the compiler has hash_map */
#define HAVE_HASH_MAP 0
/* define if the compiler has hash_set */
#define HAVE_HASH_SET 0
/* define if the compiler supports unordered_{map,set} */
#define HAVE_UNORDERED_MAP 1

Original comment by colin.di...@gmail.com on 2 Apr 2015 at 7:44