yann@339: diff -durN gcc-3.4.6.orig/gcc/config/arm/arm.c gcc-3.4.6/gcc/config/arm/arm.c
yann@339: --- gcc-3.4.6.orig/gcc/config/arm/arm.c	2005-10-01 15:31:38.000000000 +0200
yann@339: +++ gcc-3.4.6/gcc/config/arm/arm.c	2007-08-15 22:54:59.000000000 +0200
yann@339: @@ -4857,6 +4857,11 @@
yann@339:        *load_offset = unsorted_offsets[order[0]];
yann@339:      }
yann@339:  
yann@339: +  /* For XScale a two-word LDM is a performance loss, so only do this if
yann@339: +     size is more important.  See comments in arm_gen_load_multiple.  */
yann@339: +  if (nops == 2 && arm_tune_xscale && !optimize_size)
yann@339: +    return 0;
yann@339: +
yann@339:    if (unsorted_offsets[order[0]] == 0)
yann@339:      return 1; /* ldmia */
yann@339:  
yann@339: @@ -5083,6 +5088,11 @@
yann@339:        *load_offset = unsorted_offsets[order[0]];
yann@339:      }
yann@339:  
yann@339: +  /* For XScale a two-word LDM is a performance loss, so only do this if
yann@339: +     size is more important.  See comments in arm_gen_load_multiple.  */
yann@339: +  if (nops == 2 && arm_tune_xscale && !optimize_size)
yann@339: +    return 0;
yann@339: +
yann@339:    if (unsorted_offsets[order[0]] == 0)
yann@339:      return 1; /* stmia */
yann@339:  
yann@339: diff -durN gcc-3.4.6.orig/gcc/config/arm/arm.md gcc-3.4.6/gcc/config/arm/arm.md
yann@339: --- gcc-3.4.6.orig/gcc/config/arm/arm.md	2005-10-01 15:31:38.000000000 +0200
yann@339: +++ gcc-3.4.6/gcc/config/arm/arm.md	2007-08-15 22:54:59.000000000 +0200
yann@339: @@ -8811,13 +8811,16 @@
yann@339:     (set_attr "length" "4,8,8")]
yann@339:  )
yann@339:  
yann@339: +; Try to convert LDR+LDR+arith into [add+]LDM+arith
yann@339: +; On XScale, LDM is always slower than two LDRs, so only do this if
yann@339: +; optimising for size.
yann@339:  (define_insn "*arith_adjacentmem"
yann@339:    [(set (match_operand:SI 0 "s_register_operand" "=r")
yann@339:  	(match_operator:SI 1 "shiftable_operator"
yann@339:  	 [(match_operand:SI 2 "memory_operand" "m")
yann@339:  	  (match_operand:SI 3 "memory_operand" "m")]))
yann@339:     (clobber (match_scratch:SI 4 "=r"))]
yann@339: -  "TARGET_ARM && adjacent_mem_locations (operands[2], operands[3])"
yann@339: +  "TARGET_ARM && (!arm_tune_xscale || optimize_size) && adjacent_mem_locations (operands[2], operands[3])"
yann@339:    "*
yann@339:    {
yann@339:      rtx ldm[3];
yann@339: @@ -8852,6 +8855,8 @@
yann@339:        }
yann@339:     if (val1 && val2)
yann@339:        {
yann@339: +	/* This would be a loss on a Harvard core, but adjacent_mem_locations()
yann@339: +	   will prevent it from happening.  */
yann@339:  	rtx ops[3];
yann@339:  	ldm[0] = ops[0] = operands[4];
yann@339:  	ops[1] = XEXP (XEXP (operands[2], 0), 0);
yann@339: diff -durN gcc-3.4.6.orig/gcc/genpeep.c gcc-3.4.6/gcc/genpeep.c
yann@339: --- gcc-3.4.6.orig/gcc/genpeep.c	2003-07-05 07:27:22.000000000 +0200
yann@339: +++ gcc-3.4.6/gcc/genpeep.c	2007-08-15 22:54:59.000000000 +0200
yann@339: @@ -381,6 +381,7 @@
yann@339:    printf ("#include \"recog.h\"\n");
yann@339:    printf ("#include \"except.h\"\n\n");
yann@339:    printf ("#include \"function.h\"\n\n");
yann@339: +  printf ("#include \"flags.h\"\n\n");
yann@339:  
yann@339:    printf ("#ifdef HAVE_peephole\n");
yann@339:    printf ("extern rtx peep_operand[];\n\n");