On Fri, Sep 6, 2013 at 12:20 PM, Joern Rennecke wrote: > We found that > > std::basic_string<char, std::char_traits<char>, std::allocator<char> > ::copy(char*, unsigned long, unsigned long) const > > got miscompiled for ARC because reorg thought that all call-clobbered > registers were dead after a conditional call.
Hmm, interesting to have a target with both predication and delay slots... > I can't reproduce the test case with current trunk + ARC port, but I reckon > this is just due to the fickleness of the register allocator. Chances are, > every other version, some other obscure function is miscompiled without > this patch. > > bootstrapped on i686-pc-linux-gnu. That's not a very good test, i686 doesn't use this code ;-) > OK to apply? Yes, the patch is OK. However, I wouldn't be surprised if using COND_EXECs together with delay slots has more hidden problems. Browsing through resource.c, the "if (jump_insn)" block at lines 1095:1119 jumps out as another place you should probably look into. Ciao! Steven
