Re: PATCH to hoist loads/stores out of loops

[Richard Henderson <rth at cygnus dot com>]

On Mon, Jul 20, 1998 at 05:19:59PM -0700, Mark Mitchell wrote:
> Actually, the code I submitted does do this.  What, you say, it didn't
> when you tried it?  Oops, I forgot to tell you that I haven't yet
> implemented C's `restrict'.

Following up on my own message (which hasn't shown up yet) looking
at the actual code rather than the patch brings to my attention that
load_mems is run before movables are found and moved.  Duh.

Just moving the call of load_mem to the end of scan_loop as pictured
below, was enough to get Toon's test case to work.  Also cleaned is
the REG_USERVAR_P nonsense, since we now happen after moveables.

I don't know if calling record_base_value is helpful or not.  The only
time it would matter is if your induced value were used in an address,
and I can't immediately come up with a way that could occur directly.


r~


--- loop.c.mm	Tue Jul 21 00:32:36 1998
+++ loop.c	Tue Jul 21 00:55:16 1998
@@ -781,35 +781,6 @@ scan_loop (loop_start, end, unroll_p)
     may_not_optimize[i] = 1, n_times_set[i] = 1;
   bcopy ((char *) n_times_set, (char *) n_times_used, nregs * sizeof (int));
 
-  /* Load MEMs into REGs for the duration of the loop, if
-     appropriate.  Note that calculate n_times_set, etc., before this
-     point so that we can tell what things are loop-invariant in
-     load_mems.  */
-  load_mems (scan_start, end, loop_top, loop_start);
-
-  /* Recalculate n_times_set and friends since load_mems may have
-     created new registers.  */
-  if (max_reg_num () > nregs - loop_mems_idx)
-    {
-      bzero ((char *) n_times_set, nregs * sizeof (int));
-      bzero (may_not_optimize, nregs);
-      if (loop_has_call)
-	bzero ((char *) reg_single_usage, nregs * sizeof (rtx));
-
-      count_loop_regs_set (loop_top ? loop_top : loop_start, end,
-			   may_not_optimize, reg_single_usage,
-			   &insn_count, nregs); 
-
-      for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
-	may_not_optimize[i] = 1, n_times_set[i] = 1;
-      bcopy ((char *) n_times_set, (char *) n_times_used, 
-	     nregs * sizeof (int)); 
-
-      /* Set nregs to reflect the real number of registers.  There's no
-	 need to include any extra slop.  */
-      nregs = max_reg_num ();
-    }
-
   if (loop_dump_stream)
     {
       fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
@@ -1182,6 +1153,11 @@ scan_loop (loop_start, end, unroll_p)
     if (n_times_set[i] < 0)
       n_times_set[i] = n_times_used[i];
 
+  /* Load MEMs into REGs for the duration of the loop, if appropriate. 
+     Note that the adjust of n_times_set is before this point so that
+     we can tell what things are loop-invariant in load_mems.  */
+  load_mems (scan_start, end, loop_top, loop_start);
+
   if (flag_strength_reduce)
     {
       the_movables = movables;
@@ -8534,13 +8510,12 @@ load_mems (scan_start, end, loop_top, st
 	       loop, but later discovered that we could not.  */
 	    continue;
 
-	  /* Allocate a pseudo for this MEM.  We set REG_USERVAR_P in
-	     order to keep scan_loop from moving stores to this MEM
-	     out of the loop just because this REG is neither a
-	     user-variable nor used in the loop test.  */
+	  /* Allocate a pseudo for this MEM.  Initialize n_times_set for
+	     this reg; note that it is zero by design -- our one set is
+	     outside the loop under consideration.  */
 	  reg = gen_reg_rtx (GET_MODE (mem));
-	  REG_USERVAR_P (reg) = 1;
 	  loop_mems[i].reg = reg;
+	  n_times_set[REGNO (reg)] = 0;
 
 	  /* Now, replace all references to the MEM with the
 	     corresponding pesudos.  */
@@ -8580,6 +8555,14 @@ load_mems (scan_start, end, loop_top, st
 		   the NOTE_LOOP_END.  */
 		  set = gen_rtx_SET (GET_MODE (reg), mem, reg); 
 		  emit_insn_after (set, label);
+		}
+
+	      if (loop_dump_stream)
+		{
+		  fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
+			   REGNO (reg), (written ? "r/w" : "r/o"));
+	  	  print_rtl (loop_dump_stream, mem);
+		  fputc ('\n', loop_dump_stream);
 		}
 	    }
 	}