[mips] std::regex_traits considers '-' to be a member of the 'w' class

[llvmbot]

Bugzilla Link	26476
Resolution	FIXED
Resolved on	Feb 17, 2016 09:04
Version	3.8
OS	Linux
Blocks	#26433
Reporter	LLVM Bugzilla Contributor
CC	@mclow

Extended Description

It currently fails with this assertion:
isctype.pass.cpp:30: int main(): Assertion `!t.isctype('-', t.lookup_classname(s.begin(), s.end()))' failed.
this test passed in the previous release.

[mclow]

Is this the only failure in that test?
(i.e, if you comment out that assert and re-run it, does it pass)

I suspect that ctype_base (defined in __locale) is not getting built correctly for your system; there's a set of #ifdefs for different systems there, and none for mips.

[llvmbot]

The 'assert(!t.isctype('@', t.lookup_classname(s.begin(), s.end())));' on line 31 and then the wchar_t version of the same two on lines 159 and 160 also fail.

I've only had a quick look at ctype_base in __locale so far but my machine should be covered by the GLIBC. I'll keep digging.

[llvmbot]

It looks like the result of lookup_classname() is wrong. If I understand this correctly, it should return a bitmask to test against the values in classic_table.

I think this class should be digit|alpha|upper|lower but some how 'graph' has crept in too. This then matches the punct|graph|print used for the '-'.

[llvmbot]

I'm fairly certain the problem is __regex_word. It's trying to use an unoccupied bit (0x80) but this bit is only unoccupied on little endian machines.

On big-endian the bytes are reversed in the enum containing _ISgraph/_ISprint/etc. (by _ISbit() in /usr/include/ctype.h) so __regex_word and _ISgraph happen to have the same value. It seems that __regex_word needs to be 0x8000 on a big-endian machine.

It's late so I'm leaving a build running overnight to confirm. I should know if this is the problem in the morning.

[llvmbot]

I can confirm that this was the problem. I've posted a patch at http://reviews.llvm.org/D17132

[mclow]

In r260527, I added some tests to catch this if it happens again.
Can you please update to that version and run the tests, with and without your patch?

[llvmbot]

The tests you added in r260527 pass even without my patch. This is because the table is shared between C and C++ and we're colliding with the _ISgraph bit used by C rather than any of the C++ related bits.

[llvmbot]

Committed an '#ifdef mips' version in r261088 and merged it in r261097