Re: [patch] PR debug/53948

[Richard Biener <richard dot guenther at gmail dot com>]

On Wed, Jul 18, 2012 at 7:46 PM, Steven Bosscher <stevenb.gcc@gmail.com> wrote:
> Hello,
>
> This is my proposed fix for PR53948. We don't want to put user
> variables in callee-clobbered registers, but obviously function
> arguments are OK there. REG_USERVAR_P is set on PARM_DECLs and on user
> variables, so it can't be used to distinguish between the two.
>
> As it turns out, I can hi-jack a bit for that: 'unchanging' (currently
> incorrectly documented as used on REG) for a new macro
> REG_FUNCTION_PARM_P. I found one obvious place where this bit can be
> used instead of REG_USERVAR_P, and probably there are a few more
> places where this is useful (TBD, I'm going to look at all places
> where RTL code looks at tree's PARM_DECL later).
>
> Bootstrapped&tested on powerpc64-unknown-linux-gnu.
> OK for trunk?

Just being pointed back to this patch ... I wonder if simply looking
at REG_EXPR of a REG_USERVAR_P reg and checking whether
it's a PARM_DECL isn't good enough (and simpler)?

I suppose the optabs.c change was the "real" fix, thus sth like

Index: gcc/optabs.c
===================================================================
--- gcc/optabs.c        (revision 194552)
+++ gcc/optabs.c        (working copy)
@@ -3848,9 +3848,13 @@ emit_libcall_block_1 (rtx insns, rtx tar
   rtx final_dest = target;
   rtx next, last, insn;

-  /* If this is a reg with REG_USERVAR_P set, then it could possibly turn
-     into a MEM later.  Protect the libcall block from this change.  */
-  if (! REG_P (target) || REG_USERVAR_P (target))
+  /* If this is a parameter with REG_USERVAR_P set, then it could possibly turn
+     into a MEM later (e.g. if a REG_EQUIV note is attached to the insns that
+     sets the reg).  Protect the libcall block from this change.  */
+  if (! REG_P (target)
+      || (REG_USERVAR_P (target)
+         && REG_EXPR (target) != NULL_TREE
+         && TREE_CODE (REG_EXPR (target)) == PARM_DECL))
     target = gen_reg_rtx (GET_MODE (target));

   /* If we're using non-call exceptions, a libcall corresponding to an

(untested)

Thanks,
Richard.

> Ciao!
> Steven
>
>         PR debug/53948
>         * rtl.h (REG_FUNCTION_PARM_P): New flag on a REG.  Re-use 'unchaning'.
>         * emit-rtl.c (mark_function_parm_reg): New function.
>         * function.c (assign_parm_setup_reg): Use mark_function_parm_reg
>         instead of mark_user_reg.
>         * combine.c (can_change_dest_mode): Preserve REG_FUNCTION_PARM_P.
>         * web.c (entry_register): Likewise.
>         * reload1.c (reload): Likewise.
>         * ira-emit.c (ira_create_new_reg): Likewise.
>         * reginfo.c (reg_scan_mark_refs): Likewise.
>         * optabs.c (emit_libcall_block_1): Use REG_FUNCTION_PARM_P instead
>         of REG_USERVAR_P.
>         * regstat.c (dump_reg_info): Print REG_FUNCTION_PARM_P.
>         * doc/rtl.texi (REG_FUNCTION_PARM_P): Document it.
>         ('unchanging' flag): Fix documentation.