Currently the integer tests report an error on 64 bit platforms, as can be seen on http://tinyurl.com/6whtg The following code in integer_test.cpp causes the problem: PRIVATE_FIT_TESTS( int_t, least, long, LONG_MAX ); PRIVATE_FIT_TESTS( int_t, fast, long, LONG_MAX ); PRIVATE_FIT_TESTS( uint_t, least, unsigned long, ULONG_MAX ); PRIVATE_FIT_TESTS( uint_t, fast, unsigned long, ULONG_MAX ); The first line expands to: do { long v = LONG_MAX ; PRIVATE_FIT_TEST(int_t, 32, least, v); v >>= 1; PRIVATE_FIT_TEST(int_t, 31, least, v); v >>= 1; ... } which becomes: do { long v = LONG_MAX ; BOOST_TEST( int_t < 32 > :: least ( v ) == v ); v >>= 1; BOOST_TEST( int_t < 31 > :: least ( v ) == v ); v >>= 1; ... } which of course alway fails on 64 bit platforms because LONG_MAX is a 64 bit number. A possible fix would be to start the test with 64 bit and not 32 on 64 bit platforms. But how can 64 bit platforms be detected? Probably with #if (LONG_MAX > INT_MAX) ... #endif or something like this. Another fix (which ignores 64 bit values alltogether) would be to change the lines above to not use long and LONG_MAX but int and INT_MAX. Opinions, anyone? Markus
On Wed, Sep 01, 2004 at 12:50:32PM +0200, Markus Sch?pflin wrote:
Currently the integer tests report an error on 64 bit platforms, as can be seen on http://tinyurl.com/6whtg
The following code in integer_test.cpp causes the problem:
PRIVATE_FIT_TESTS( int_t, least, long, LONG_MAX ); PRIVATE_FIT_TESTS( int_t, fast, long, LONG_MAX ); PRIVATE_FIT_TESTS( uint_t, least, unsigned long, ULONG_MAX ); PRIVATE_FIT_TESTS( uint_t, fast, unsigned long, ULONG_MAX );
The first line expands to:
do { long v = LONG_MAX ; PRIVATE_FIT_TEST(int_t, 32, least, v); v >>= 1; PRIVATE_FIT_TEST(int_t, 31, least, v); v >>= 1; ... }
which becomes:
do { long v = LONG_MAX ; BOOST_TEST( int_t < 32 > :: least ( v ) == v ); v >>= 1; BOOST_TEST( int_t < 31 > :: least ( v ) == v ); v >>= 1; ... }
which of course alway fails on 64 bit platforms because LONG_MAX is a 64 bit number.
A possible fix would be to start the test with 64 bit and not 32 on 64 bit platforms. But how can 64 bit platforms be detected? Probably with #if (LONG_MAX > INT_MAX) ... #endif or something like this.
This is not _guaranteed_ to work, but would be OK in practice. An ILP64 platform would have sizeof(int) == sizeof(long) == 8 and an LLP64 platform would have sizeof(int) == sizeof(long) == 4 In both cases LONG_MAX == INT_MAX. I think all 64-bit unices use the LP64 model, where sizeof(int) == 4 and sizeof(long) == 8, so I think your suggestion would work.
Another fix (which ignores 64 bit values alltogether) would be to change the lines above to not use long and LONG_MAX but int and INT_MAX.
As I understand it that just changes the bad assumption that long is 32-bits to assuming that int is 32 (or fewer) bits, which is still an unportable assumption, but likely to work for the foreseeable future (famous last words :-) jon -- I don't pretend to be an expert on intellectual property law, but I do know one thing. If a music industry executive claims I should agree with their agenda because it will make me more money, I put my hand on my wallet ... and check it after they leave, just to make sure nothing's missing. - Janis Ian <http://www.janisian.com/article-internet_debacle.html>
Jonathan Wakely wrote:
On Wed, Sep 01, 2004 at 12:50:32PM +0200, Markus Schöpflin wrote:
A possible fix would be to start the test with 64 bit and not 32 on 64 bit platforms. But how can 64 bit platforms be detected? Probably with #if (LONG_MAX > INT_MAX) ... #endif or something like this.
This is not _guaranteed_ to work, but would be OK in practice.
An ILP64 platform would have sizeof(int) == sizeof(long) == 8 and an LLP64 platform would have sizeof(int) == sizeof(long) == 4
In both cases LONG_MAX == INT_MAX.
I think all 64-bit unices use the LP64 model, where sizeof(int) == 4 and sizeof(long) == 8, so I think your suggestion would work.
Do you know of a way which would work portably?
Another fix (which ignores 64 bit values alltogether) would be to change the lines above to not use long and LONG_MAX but int and INT_MAX.
As I understand it that just changes the bad assumption that long is 32-bits to assuming that int is 32 (or fewer) bits, which is still an unportable assumption, but likely to work for the foreseeable future (famous last words :-)
Yes, right. But it would just be a fix to make the tests succeed, it wouldn't really count for much on a 64 bit platform. Maybe one of the authors of this test could perhaps voice an opinion on what to do? Markus
participants (3)
-
Jonathan Wakely -
Markus Schöpflin -
Markus Schöpflin