The PowerPC backend has a pretty complete MC support, so the fix should be trivial, but it is probably better if someone familiar with the ISA looks at it.
My guess from the existing code is that is most cpus msync is as alias for "sync 0", but on "book E" versions there is not "sync 0" and msync is a proper instruction.
Extended Description
llc -march=ppc32 -mcpu=440 -filetype=obj -o test.o
crashes on
define void @f() { fence acquire ret void }
The PowerPC backend has a pretty complete MC support, so the fix should be trivial, but it is probably better if someone familiar with the ISA looks at it.
My guess from the existing code is that is most cpus msync is as alias for "sync 0", but on "book E" versions there is not "sync 0" and msync is a proper instruction.