On Thu, Apr 12, 2012 at 9:22 AM, Alan Modra <amo...@gmail.com> wrote:
> I tried that. It doesn't work without something else in the insn to
> stop rtl-dce deleting it, so you may as well use SETs. But thanks for
> the prod in the right direction. We do get slightly better results
> when the regs are not hidden away in an UNSPEC, for instance
> non-stack writes/reads are seen by the alias oracle to not conflict
> with the epilogue frame deallocation.
> Bootstrapped etc. powerpc-linux. OK to apply, David?
> PR target/52828
> * config/rs6000/rs6000.c (rs6000_emit_stack_tie): Rewrite with
> tie regs on destination of sets. Delete forward declaration.
> (rs6000_emit_stack_reset): Update rs6000_emit_stack_tie calls.
> (rs6000_emit_prologue): Likewise.
> (rs6000_emit_epilogue): Likewise. Use in place of gen_frame_tie
> and gen_stack_tie.
> (is_mem_ref): Use tie_operand to recognise stack ties.
> * config/rs6000/predicates.md (tie_operand): New.
> * config/rs6000/rs6000.md (UNSPEC_TIE): Delete.
> (restore_stack_block): Generate new stack tie rtl.
> (restore_stack_nonlocal): Likewise.
> (stack_tie): Update.
> (frame_tie): Delete.
This probably is getting close to the best we can do with GCC's RTL
alias analysis infrastructure.
This version is okay, but I also want to give Richi and Olivier an
opportunity to comment if they still have any concerns.