>> Or, if ARM supports unaligned loads/stores using special instructions, >> perhaps you should also benchmark the alternative of not realigning, but >> instead making sure those unaligned instructions are used for the shadow >> memory loads/stores in the asan prologue/epilogue. > I have tried to use -fno-peephole2 to shutdown instructions > combinations,but that makes me feel uncomfortable.
That should not be required. I suspect what you need is a movmisalignsi expander generating unaligned_store/loadsi patterns. Notice that these put out ldr / str instructions but with UNSPECs which means that the peepholers for ldrd / ldm will not catch them. As Jakub says later in this thread ARM is a STRICT_ALIGNMENT target. If the compiler is peepholing unaligned ldr's and str's, that would be a bug and the compiler needs to be fixed to avoid this. regards Ramana >> >> Please next time post the patch from a sane MUA that doesn't eat tabs or >> at least as an attachment. Comments should start with a capital letter, and >> with a full stop followed by two spaces, in the ChangeLog also all entries >> should end with a full stop. > Sorry about that... > BTW,I didn't modify the ChangeLog because this patch should be discussed > carefully. >> >> Jakub > >
