http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49338
--- Comment #4 from Hans Meier <hsmeier at arcor dot de> 2011-06-09 08:42:31 UTC --- (In reply to comment #2) > is an obvious aliasing violation. yes, but why does this lead to read from other objects than it is intended? - The "[1..1]" comes from m_ppSelect[i] when reading m_ppUpperLimit[i] and m_ppLowerLimit[i], if we change Select to 0 the range also changes to [0..0] etc. btw: this code sequence in our software is called about 10 million times per second from more than 1000 occurrences in 100 classes and only in this singe constellation it doesn't work. Considering these figures it is THE hotspot for performance optimization and this is the fastest implementation I could find within weeks of experimenting.
