Hi, On Wed, 20 Jun 2007, Paolo Bonzini wrote:
> > This is one of the places where i slavishly copied what flow did. if > > you want to change this, go test it on at least 7 platforms and fix all > > of the problems that it causes. > > I see. Can one of you recap how it relates to the m68k problem, though? :-) I have basically this sequence of instructions: (call_insn 1 (set (%d0) (call ...))) (insn 2 (set (x) (%d0))) (call_insn 3 (set (%a0) (call ...))) ... (call_insn/j 5 (set (%a0) (call ...))) The immediate problem is that there is no REG_DEAD note generated for insn 2. reload calls the code in caller-save.c, because some other call clobbered register is associated with a pseudo and needs to be preserved across a call. It builds live information based on the REG_DEAD notes and is confused that %d0 doesn't die in insn 3. There is one problem that the live out information in this block is incorrect, as it includes the return register, but the function is exited via sibcall and there is no code to set the return register. This is made a little more complex, as on m68k the return value here is returned in %d0/%a0, so %d0 is live at the end of the block. One possible fix is to clear the uses at the edge to the exit (what the small patch previously posted does). The second problem (and the one in discussion right now) is that I think %d0 should have been clobbered already by any of the later calls, which is at least what reload and many other passes assume, they had to manually add this information, if they wanted to use the dataflow information. bye, Roman
