On Sat, Jul 23, 2011 at 6:29 PM, DJ Delorie <d...@redhat.com> wrote: > >> Yeah, the testcase is invalid - that we lost the volatile was a bug, but >> you really have to fix the kernel. > > Sadly, that's not a helpful suggestion. How else should the kernel > force a word-sized read? I thought volatile was the way to tell gcc > "do exactly what I tell you to do" ?
Huh? No. Volatile means "preserve this load/store". If you want to force it to assume the memory is aligned then you have to do that during RTL expansion or in the target. > And the kernel does produce the right code at that particular spot > now, but for some reason the usb module still doesn't work. We'll > have to debug it some more. The same code works just fine with gcc > 4.4. Fact is that GCC knows that memory is not properly aligned. Iff arm is truly a strict-align target then an aligned load/store from such memory will fault. And I seriously doubt you get the int member magically aligned by "luck" (thus, misaligning the first char member by exactly 3 bytes). So either arm should not lie and claim it's strict-align or you need to fix the kernel anyway. Richard.
