http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52890
--- Comment #10 from Pat Haugen <pthaugen at gcc dot gnu.org> 2012-09-12 23:04:55 UTC --- Created attachment 28181 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28181 Reduced testcase Martin, Have you done any more digging on this? I just discovered that cpu2006 benchmark 471.omnetpp suffers the same problem (8 byte loads turned in to sequence of byte loads/shifts/ors). It causes a 12% degradation on PowerPC, and also goes away when -fno-ipa-sra is added. I'm attaching a reduced testcase that can be compiled with g++ -S -O3 -m64 -mcpu=power7.
