[PATCH] m68k: fix builtins/memset.c

[PATCH] m68k: fix builtins/memset.c

[Roman Zippel]

Hi,

This patch fixes builtins/memset.c, which currently fails because 
"memset(ptr, val, 1)" isn't optimized to a simple set operation. The 
current definition of MOVE_BY_PIECES_P prevents byte moves, theoretically 
it's possible to fix this also by defining a new STORE_BY_PIECES_P, but 
the easier solution is to just remove MOVE_BY_PIECES_P. The reason it was 
added is long gone, as bytes pushes on the stack are handled via the 
pushqi1 pattern now. To make sure this stays this way I replaced the code 
in output_move_qimode() which used to generate this code with an abort().

The patch is tested on linux-m68k.

bye, Roman

2004-07-13  Roman Zippel  <zippel@linux-m68k.org>

	* config/m68k/m68k.c (output_move_qimode): Abort on an attempt to 
	generate code which is generated by pushqi1 now
	* config/m68k/m68k.h (MOVE_BY_PIECES_P): Remove.


Index: gcc/config/m68k/m68k.c
===================================================================
RCS file: /home/roman/src/gcc-cvs/gcc/gcc/config/m68k/m68k.c,v
retrieving revision 1.134
diff -u -r1.134 m68k.c
--- gcc/config/m68k/m68k.c	16 Jun 2004 21:53:46 -0000	1.134
+++ gcc/config/m68k/m68k.c	13 Jul 2004 09:29:21 -0000
@@ -1780,10 +1780,6 @@
 const char *
 output_move_qimode (rtx *operands)
 {
-  rtx xoperands[4];
-
-  /* This is probably useless, since it loses for pushing a struct
-     of several bytes a byte at a time.	 */
   /* 68k family always modifies the stack pointer by at least 2, even for
      byte pushes.  The 5200 (ColdFire) does not do this.  */
   if (GET_CODE (operands[0]) == MEM
@@ -1791,22 +1787,8 @@
       && XEXP (XEXP (operands[0], 0), 0) == stack_pointer_rtx
       && ! ADDRESS_REG_P (operands[1])
       && ! TARGET_COLDFIRE)
-    {
-      xoperands[1] = operands[1];
-      xoperands[2]
-	= gen_rtx_MEM (QImode,
-		       gen_rtx_PLUS (VOIDmode, stack_pointer_rtx, const1_rtx));
-      /* Just pushing a byte puts it in the high byte of the halfword.	*/
-      /* We must put it in the low-order, high-numbered byte.  */
-      if (!reg_mentioned_p (stack_pointer_rtx, operands[1]))
-	{
-	  xoperands[3] = stack_pointer_rtx;
-	  output_asm_insn ("subq%.l #2,%3\n\tmove%.b %1,%2", xoperands);
-	}
-      else
-	output_asm_insn ("move%.b %1,%-\n\tmove%.b %@,%2", xoperands);
-      return "";
-    }
+    /* generated by pushqi1 pattern now */
+    abort ();
 
   /* clr and st insns on 68000 read before writing.
      This isn't so on the 68010, but we have no TARGET_68010.  */
Index: gcc/config/m68k/m68k.h
===================================================================
RCS file: /home/roman/src/gcc-cvs/gcc/gcc/config/m68k/m68k.h,v
retrieving revision 1.116
diff -u -r1.116 m68k.h
--- gcc/config/m68k/m68k.h	25 Jun 2004 05:05:16 -0000	1.116
+++ gcc/config/m68k/m68k.h	13 Jul 2004 09:29:22 -0000
@@ -675,11 +675,6 @@
    On the ColdFire, sp@- in a byte insn pushes just a byte.  */
 #define PUSH_ROUNDING(BYTES) (TARGET_COLDFIRE ? BYTES : ((BYTES) + 1) & ~1)
 
-/* We want to avoid trying to push bytes.  */
-#define MOVE_BY_PIECES_P(SIZE, ALIGN) \
-  (move_by_pieces_ninsns (SIZE, ALIGN) < MOVE_RATIO \
-    && (((SIZE) >=16 && (ALIGN) >= 16) || (TARGET_COLDFIRE)))
-
 #define FIRST_PARM_OFFSET(FNDECL) 8
 
 /* On the 68000, the RTS insn cannot pop anything.

Re: [PATCH] m68k: fix builtins/memset.c

[Richard Henderson]

On Tue, Jul 13, 2004 at 07:55:56PM +0200, Roman Zippel wrote:
> 	* config/m68k/m68k.c (output_move_qimode): Abort on an attempt to 
> 	generate code which is generated by pushqi1 now
> 	* config/m68k/m68k.h (MOVE_BY_PIECES_P): Remove.

Ok.


r~