diff -r bd5e0a849352 -r b150d6f590fc patches/gcc/3.4.6/600-arm-ldm-peephole.patch
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/patches/gcc/3.4.6/600-arm-ldm-peephole.patch	Mon Jul 28 21:08:01 2008 +0000
@@ -0,0 +1,68 @@
+diff -durN gcc-3.4.6.orig/gcc/config/arm/arm.c gcc-3.4.6/gcc/config/arm/arm.c
+--- gcc-3.4.6.orig/gcc/config/arm/arm.c	2005-10-01 15:31:38.000000000 +0200
++++ gcc-3.4.6/gcc/config/arm/arm.c	2007-08-15 22:54:59.000000000 +0200
+@@ -4857,6 +4857,11 @@
+       *load_offset = unsorted_offsets[order[0]];
+     }
+ 
++  /* For XScale a two-word LDM is a performance loss, so only do this if
++     size is more important.  See comments in arm_gen_load_multiple.  */
++  if (nops == 2 && arm_tune_xscale && !optimize_size)
++    return 0;
++
+   if (unsorted_offsets[order[0]] == 0)
+     return 1; /* ldmia */
+ 
+@@ -5083,6 +5088,11 @@
+       *load_offset = unsorted_offsets[order[0]];
+     }
+ 
++  /* For XScale a two-word LDM is a performance loss, so only do this if
++     size is more important.  See comments in arm_gen_load_multiple.  */
++  if (nops == 2 && arm_tune_xscale && !optimize_size)
++    return 0;
++
+   if (unsorted_offsets[order[0]] == 0)
+     return 1; /* stmia */
+ 
+diff -durN gcc-3.4.6.orig/gcc/config/arm/arm.md gcc-3.4.6/gcc/config/arm/arm.md
+--- gcc-3.4.6.orig/gcc/config/arm/arm.md	2005-10-01 15:31:38.000000000 +0200
++++ gcc-3.4.6/gcc/config/arm/arm.md	2007-08-15 22:54:59.000000000 +0200
+@@ -8811,13 +8811,16 @@
+    (set_attr "length" "4,8,8")]
+ )
+ 
++; Try to convert LDR+LDR+arith into [add+]LDM+arith
++; On XScale, LDM is always slower than two LDRs, so only do this if
++; optimising for size.
+ (define_insn "*arith_adjacentmem"
+   [(set (match_operand:SI 0 "s_register_operand" "=r")
+ 	(match_operator:SI 1 "shiftable_operator"
+ 	 [(match_operand:SI 2 "memory_operand" "m")
+ 	  (match_operand:SI 3 "memory_operand" "m")]))
+    (clobber (match_scratch:SI 4 "=r"))]
+-  "TARGET_ARM && adjacent_mem_locations (operands[2], operands[3])"
++  "TARGET_ARM && (!arm_tune_xscale || optimize_size) && adjacent_mem_locations (operands[2], operands[3])"
+   "*
+   {
+     rtx ldm[3];
+@@ -8852,6 +8855,8 @@
+       }
+    if (val1 && val2)
+       {
++	/* This would be a loss on a Harvard core, but adjacent_mem_locations()
++	   will prevent it from happening.  */
+ 	rtx ops[3];
+ 	ldm[0] = ops[0] = operands[4];
+ 	ops[1] = XEXP (XEXP (operands[2], 0), 0);
+diff -durN gcc-3.4.6.orig/gcc/genpeep.c gcc-3.4.6/gcc/genpeep.c
+--- gcc-3.4.6.orig/gcc/genpeep.c	2003-07-05 07:27:22.000000000 +0200
++++ gcc-3.4.6/gcc/genpeep.c	2007-08-15 22:54:59.000000000 +0200
+@@ -381,6 +381,7 @@
+   printf ("#include \"recog.h\"\n");
+   printf ("#include \"except.h\"\n\n");
+   printf ("#include \"function.h\"\n\n");
++  printf ("#include \"flags.h\"\n\n");
+ 
+   printf ("#ifdef HAVE_peephole\n");
+   printf ("extern rtx peep_operand[];\n\n");