On Mon, Feb 20, 2012 at 7:18 PM, Jakub Jelinek <ja...@redhat.com> wrote: > On Mon, Feb 20, 2012 at 07:10:26PM +0100, Uros Bizjak wrote: >> > So the above is right and needed, though perhaps we might want >> > a combine pattern or peephole to turn the >> > movl $-1, %eax >> > xbegin .+6 >> > cmpl %eax, $-1 >> > jne 1f >> >> The compiler can reverse the condition and exchange arms of if >> expression. This will void the assumption that false arm (and abort >> sequence) is at pc+6. > > The xbegin .+6 instruction doesn't (from the compiler's POV) as any > JUMP_INSN, it just returns either -1 or some other value in %eax > (when coupled with setting %eax to -1 before it). > That is the semantics of the _xbegin () intrinsic. > We can (and at least in the combiner should) model it as > a JUMP_INSN with UNSPEC_VOLATILE in it, which will work similarly > to asm goto. The question is how happy cfgrtl.c etc. will be on > a conditional JUMP_INSN with UNSPEC_VOLATILE in it.
IIUC the documentation, the fallback label is a parameter to xbegin insn, but the insn itself doesn't jump anywhere - it just records the parameter as a fallback address. However, there is no guarantee that the fallback code is exactly at (pc)+6, so we have to use asm labels here. As written in your example in this thread, the label is at the top of the false branch. Instead of ".+6", we have to pass this label to xbegin insn somehow. Uros.
