This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [tree-ssa] replace sibcall.c take 3
> On Tue, Nov 18, 2003 at 03:13:54AM +0100, Jan Hubicka wrote:
> > * Makefile.in (sibcall.o): Kill.
> > (tree-tailcall.o): Add except.h dependency
> > * sibcall.c: Kill.
> > (purge_reg_equiv_notes, purge_mem_unchanging_flag): Move to ...
> > * calls.c (purge_reg_equiv_notes, purge_mem_unchanging_flag) ... here.
> > (expand_call): Do not produce placeholders; do not deal with tail
> > recursion; update equivalencies after sibcall production.
> > * toplev.c (rest_of_handle_sibling_calls): Kill.
> > (rest_of_compialtion): Do not use rest_of_handle_sibling_calls.
> > * tree-dump.c (dump_files): Add tail2
> > * tree-flow.h (tree_optimize_tail_calls): Update prototype.
> > * tree-optimize.c (optimize_function_tree): Do tail optimization twice.
> > * tree-tailcall.c: Inlucde except.h
> > (suitable_for_tail_call_opt_p): New.
> > (optimize_tail_call): Add opt_tailcalls argument; optimize tailcalls.
> > (tree_optimize_tail_calls): Add opt_tailcalls/pass arguments.
> > * tree.h (CALL_EXPR_TAILCALL): New.
> > (tree_dump_index): Add tail2
> >
> > * gcc.dg/tree-ssa/tailcall-1.c: New.
> > * gcc.dg/tree-ssa/tailrecursion-?.c: Rename dump
>
> Ok for now.
>
> > ! /* If tail call production suceeded, we need to remove REG_EQUIV notes on
> > ! arguments too, as argument area is now clobbered by the call. */
> > ! if (tail_call_insns)
> > {
> ...
> > ! purge_reg_equiv_notes ();
> ...
> > ! /* A sibling call sequence also may invalidate RTX_UNCHANGING_P
> > ! flag of some incoming arguments MEM RTLs, because it can write into
> > ! those slots. We clear all those bits now.
> > !
> > ! This is (slight) overkill, we could keep track of which arguments
> > ! we actually write into. */
> > ! for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
> > ! {
> > ! if (INSN_P (insn))
> > ! purge_mem_unchanging_flag (PATTERN (insn));
> > ! }
>
> One of the first things I'd like you to clean up is to do this
> only once per function, rather than once per sibcall.
Yes, I want to do that next. But in fact I originally wanted to ask
here and forgot about that...
>
> Better yet, notice if any sibcalls exist in the statement stream
> (either via .tail2 or front-end setting the bit) and don't emit
> the notes and /u flag in the first place.
It seems to me that the notes really needs to be removed only for
sibcalls actually using memory argument.
So perhaps I can collect the information about how large portion
of argument area is clobbered and remove the flags more curefully?
Honza
>
>
> r~