Two more doloop patches

Two more doloop patches

[Bernd Schmidt]

[-- Attachment #1: Type: text/plain, Size: 857 bytes --]

These two patches strengthen functions in loop-iv.c so that we allow 
frame and stack pointers in the expressions we're examining.  For loops 
that walk an array on the stack, the exit test often is a pointer 
comparison involving addresses based on fp or sp, and the compiler can 
make useful simplifications if these substitutions are allowed.

Bootstrapped and regression tested on i686-linux (irrelevant extra 
failures in libmudflap due to expect truncating compilre output). 
Equivalent patches have been in our Blackfin 4.3 tree for a while now 
and had a fair amount of testing.


Bernd
-- 
This footer brought to you by insane German lawmakers.
Analog Devices GmbH      Wilhelm-Wagenfeld-Str. 6      80807 Muenchen
Sitz der Gesellschaft Muenchen, Registergericht Muenchen HRB 40368
Geschaeftsfuehrer Thomas Wessel, William A. Martin, Margaret Seif

[-- Attachment #2: fn-invariant.diff --]
[-- Type: text/plain, Size: 1638 bytes --]

	* loop-iv.c (replace_single_def_regs): Follow chains of
	single-definition regs.  Use reg notes where possible.  Allow not just
	constants, but anything that's function_invariant_p.

Index: loop-iv.c
===================================================================
--- loop-iv.c	(revision 145357)
+++ loop-iv.c	(working copy)
@@ -1376,23 +1376,43 @@ replace_single_def_regs (rtx *reg, void 
 {
   unsigned regno;
   df_ref adef;
-  rtx set;
+  rtx set, src;
   rtx *expr = (rtx *)expr1;
 
   if (!REG_P (*reg))
     return 0;
 
   regno = REGNO (*reg);
-  adef = DF_REG_DEF_CHAIN (regno);
-  if (adef == NULL || DF_REF_NEXT_REG (adef) != NULL
-      || DF_REF_IS_ARTIFICIAL (adef))
-    return -1;
+  for (;;)
+    {
+      rtx note;
+      adef = DF_REG_DEF_CHAIN (regno);
+      if (adef == NULL || DF_REF_NEXT_REG (adef) != NULL
+	    || DF_REF_IS_ARTIFICIAL (adef))
+	return -1;
 
-  set = single_set (DF_REF_INSN (adef));
-  if (set == NULL || SET_DEST (set) != *reg || !CONSTANT_P (SET_SRC (set)))
+      set = single_set (DF_REF_INSN (adef));
+      if (set == NULL || !REG_P (SET_DEST (set))
+	  || REGNO (SET_DEST (set)) != regno)
+	return -1;
+
+      note = find_reg_equal_equiv_note (DF_REF_INSN (adef));
+      if (note && function_invariant_p (XEXP (note, 0)))
+	src = XEXP (note, 0);
+      else
+	src = SET_SRC (set);
+
+      if (REG_P (SET_SRC (set)))
+	{
+	  regno = REGNO (SET_SRC (set));
+	  continue;
+	}
+      break;
+    }
+  if (!function_invariant_p (src))
     return -1;
 
-  *expr = simplify_replace_rtx (*expr, *reg, SET_SRC (set));
+  *expr = simplify_replace_rtx (*expr, *reg, src);
   return 1;
 }
 

[-- Attachment #3: rhs-invariant.diff --]
[-- Type: text/plain, Size: 880 bytes --]

	* loop-iv.c (simple_rhs_p): Allow regs that are function_invariant_p.

Index: loop-iv.c
===================================================================
--- loop-iv.c	(revision 145357)
+++ loop-iv.c	(working copy)
@@ -1341,10 +1341,10 @@ simple_rhs_p (rtx rhs)
       op1 = XEXP (rhs, 1);
       /* Allow reg OP const and reg OP reg.  */
       if (!(REG_P (op0) && !HARD_REGISTER_P (op0))
-	  && !CONSTANT_P (op0))
+	  && !function_invariant_p (op0))
 	return false;
       if (!(REG_P (op1) && !HARD_REGISTER_P (op1))
-	  && !CONSTANT_P (op1))
+	  && !function_invariant_p (op1))
 	return false;
 
       return true;
@@ -1358,7 +1358,7 @@ simple_rhs_p (rtx rhs)
       /* Allow reg OP const.  */
       if (!(REG_P (op0) && !HARD_REGISTER_P (op0)))
 	return false;
-      if (!CONSTANT_P (op1))
+      if (!function_invariant_p (op1))
 	return false;
 
       return true;

Re: Two more doloop patches

[Zdenek Dvorak]

Hi,

> These two patches strengthen functions in loop-iv.c so that we allow 
> frame and stack pointers in the expressions we're examining.  For loops 
> that walk an array on the stack, the exit test often is a pointer 
> comparison involving addresses based on fp or sp, and the compiler can 
> make useful simplifications if these substitutions are allowed.
> 
> Bootstrapped and regression tested on i686-linux (irrelevant extra 
> failures in libmudflap due to expect truncating compilre output). 
> Equivalent patches have been in our Blackfin 4.3 tree for a while now 
> and had a fair amount of testing.

> +      note = find_reg_equal_equiv_note (DF_REF_INSN (adef));
> +      if (note && function_invariant_p (XEXP (note, 0)))
> +	src = XEXP (note, 0);
> +      else
> +	src = SET_SRC (set);
> +
> +      if (REG_P (SET_SRC (set)))
> +	{
> +	  regno = REGNO (SET_SRC (set));
> +	  continue;
> +	}
> +      break;

maybe

if (note && function_invariant_p (XEXP (note, 0)))
  {
    src = XEXP (note, 0);
    break;
  }
src = SET_SRC (set);

if (REG_P (src))
  ...

?

Otherwise OK,

Zdenek

Re: Two more doloop patches

[Bernd Schmidt]

Zdenek Dvorak wrote:
> Hi,
> 
>> These two patches strengthen functions in loop-iv.c so that we allow 
>> frame and stack pointers in the expressions we're examining.  For loops 
>> that walk an array on the stack, the exit test often is a pointer 
>> comparison involving addresses based on fp or sp, and the compiler can 
>> make useful simplifications if these substitutions are allowed.
>>
>> Bootstrapped and regression tested on i686-linux (irrelevant extra 
>> failures in libmudflap due to expect truncating compilre output). 
>> Equivalent patches have been in our Blackfin 4.3 tree for a while now 
>> and had a fair amount of testing.
> 
>> +      note = find_reg_equal_equiv_note (DF_REF_INSN (adef));
>> +      if (note && function_invariant_p (XEXP (note, 0)))
>> +	src = XEXP (note, 0);
>> +      else
>> +	src = SET_SRC (set);
>> +
>> +      if (REG_P (SET_SRC (set)))
>> +	{
>> +	  regno = REGNO (SET_SRC (set));
>> +	  continue;
>> +	}
>> +      break;
> 
> maybe
> 
> if (note && function_invariant_p (XEXP (note, 0)))
>   {
>     src = XEXP (note, 0);
>     break;
>   }
> src = SET_SRC (set);
> 
> if (REG_P (src))
>   ...
> 
> ?
> 
> Otherwise OK,

Changed, retested to make sure that bfin-elf produces identical output, 
and bootstrapped on i686-linux.  Committed.


Bernd
-- 
This footer brought to you by insane German lawmakers.
Analog Devices GmbH      Wilhelm-Wagenfeld-Str. 6      80807 Muenchen
Sitz der Gesellschaft Muenchen, Registergericht Muenchen HRB 40368
Geschaeftsfuehrer Thomas Wessel, William A. Martin, Margaret Seif