https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64205
--- Comment #4 from Michael Meissner <meissner at gcc dot gnu.org> --- Is -mcpu=power7 -mno-hard-dfp really a reasonable option? I built a toolchain using --with-cpu=power5 from subversion id 219607 on a powerpc64 linux system, and the system bootstrapped. If I compile the code that Peter mentioned using -m32 -O2 -fPIC -S foo01.c -mlong-double-128 -mno-minimal-toc -g -fbuilding-libgcc -fno-stack-protector -g -mcpu=power5, it compiles fine. As seger mentions, -mlra fixes this (but there are still problems with -mlra). It does fail if you enable the stfiwx instruction (added in power7), but disable DFP. How many people do this? I'm not sure this is a representation of the original bug. I suspect if you changed TARGET_NO_SDMODE_STACK in rs6000.h to eliminate the TARGET_DFP test it should work. #define TARGET_NO_SDMODE_STACK (TARGET_LFIWZX && TARGET_STFIWX)
