[PATCH] Jan's fix to PR target/7434

[PATCH] Jan's fix to PR target/7434

[Gwenole Beauchesne]

Hi,

1) Just in case this gets lost, the following is what Jan appended to the
PR only.

On Sat, 10 Aug 2002, Jan Hubicka wrote:

> Hi,
> the attached patch fixes the problem.  Crash was caused by confussion of
> modes in 64bit splitting.  In the 32bit mode all values involved are
> SImode, but for 64bit we need to change them.
> 
> Bootstrapped/regtested i386. OK for mainline and 3.2.x once opened?
> 
> Sat Aug 10 22:14:41 CEST 2002  Jan Hubicka  <jh@suse.cz>
> 
> 	* i386.c (ix86_split_long_move): Fix modes in the precomputed address
> 	references.
> 
> *** i386.c.old	Sat Aug 10 18:58:35 2002
> --- i386.c	Sat Aug 10 22:13:44 2002
> *************** ix86_split_long_move (operands)
> *** 9561,9575 ****
>   	 Do an lea to the last part and use only one colliding move.  */
>         else if (collisions > 1)
>   	{
>   	  collisions = 1;
> ! 	  emit_insn (gen_rtx_SET (VOIDmode, part[0][nparts - 1],
>   				  XEXP (part[1][0], 0)));
> ! 	  part[1][0] = change_address (part[1][0],
> ! 				       TARGET_64BIT ? DImode : SImode,
> ! 				       part[0][nparts - 1]);
> ! 	  part[1][1] = adjust_address (part[1][0], VOIDmode, UNITS_PER_WORD);
>   	  if (nparts == 3)
> ! 	    part[1][2] = adjust_address (part[1][0], VOIDmode, 8);
>   	}
>       }
>   
> --- 9561,9575 ----
>   	 Do an lea to the last part and use only one colliding move.  */
>         else if (collisions > 1)
>   	{
> + 	  rtx addrreg = gen_lowpart (Pmode, part[0][nparts - 1]);
>   	  collisions = 1;
> ! 	  emit_insn (gen_rtx_SET (VOIDmode, addrreg,
>   				  XEXP (part[1][0], 0)));
> ! 	  part[1][0] = change_address (part[1][0], GET_MODE (part[0][0]),
> ! 				       addrreg);
> ! 	  part[1][1] = adjust_address (part[1][0], GET_MODE (part[0][1]), UNITS_PER_WORD);
>   	  if (nparts == 3)
> ! 	    part[1][2] = adjust_address (part[1][0], GET_MODE (part[0][2]), 8);
>   	}
>       }

2) The testcase:

2002-07-29  Gwenole Beauchesne  <gbeauchesne@mandrakesoft.com>

	PR target/7434
	* testsuite/gcc.dg/20020729-2.c: New test.

--- gcc-3.2/gcc/testsuite/gcc.dg/20020729-2.c.x86_64-split-long-moves	2002-08-13 11:19:27.000000000 +0200
+++ gcc-3.2/gcc/testsuite/gcc.dg/20020729-2.c	2002-08-13 11:19:08.000000000 +0200
@@ -0,0 +1,14 @@
+/* PR target/7434
+   This testcase ICE'd gcc on x86-64, because of confusion of modes in
+   64-bit splitting.  */
+/* { dg-do compile } */
+
+int main (void)
+{
+  static const int align_g[] = { 1, 2, 4, 8, 16 };
+  char * buf;
+  int i = 0;
+  volatile long double val = 0;
+  val = *((long double *)(buf + align_g[i]));
+  return 0;
+}

Bye,
Gwenole.