yann@1: --- gcc-3.4.0/gcc/config/arm/arm.md.arm-ldm-peephole	2004-01-13 08:24:37.000000000 -0500
yann@1: +++ gcc-3.4.0/gcc/config/arm/arm.md	2004-04-24 18:18:04.000000000 -0400
yann@1: @@ -8810,13 +8810,16 @@
yann@1:     (set_attr "length" "4,8,8")]
yann@1:  )
yann@1:  
yann@1: +; Try to convert LDR+LDR+arith into [add+]LDM+arith
yann@1: +; On XScale, LDM is always slower than two LDRs, so only do this if
yann@1: +; optimising for size.
yann@1:  (define_insn "*arith_adjacentmem"
yann@1:    [(set (match_operand:SI 0 "s_register_operand" "=r")
yann@1:  	(match_operator:SI 1 "shiftable_operator"
yann@1:  	 [(match_operand:SI 2 "memory_operand" "m")
yann@1:  	  (match_operand:SI 3 "memory_operand" "m")]))
yann@1:     (clobber (match_scratch:SI 4 "=r"))]
yann@1: -  "TARGET_ARM && adjacent_mem_locations (operands[2], operands[3])"
yann@1: +  "TARGET_ARM && (!arm_tune_xscale || optimize_size) && adjacent_mem_locations (operands[2], operands[3])"
yann@1:    "*
yann@1:    {
yann@1:      rtx ldm[3];
yann@1: @@ -8851,6 +8854,8 @@
yann@1:        }
yann@1:     if (val1 && val2)
yann@1:        {
yann@1: +	/* This would be a loss on a Harvard core, but adjacent_mem_locations()
yann@1: +	   will prevent it from happening.  */
yann@1:  	rtx ops[3];
yann@1:  	ldm[0] = ops[0] = operands[4];
yann@1:  	ops[1] = XEXP (XEXP (operands[2], 0), 0);
yann@1: --- gcc-3.4.0/gcc/config/arm/arm.c.arm-ldm-peephole	2004-04-24 18:16:25.000000000 -0400
yann@1: +++ gcc-3.4.0/gcc/config/arm/arm.c	2004-04-24 18:18:04.000000000 -0400
yann@1: @@ -4838,6 +4841,11 @@
yann@1:        *load_offset = unsorted_offsets[order[0]];
yann@1:      }
yann@1:  
yann@1: +  /* For XScale a two-word LDM is a performance loss, so only do this if
yann@1: +     size is more important.  See comments in arm_gen_load_multiple.  */
yann@1: +  if (nops == 2 && arm_tune_xscale && !optimize_size)
yann@1: +    return 0;
yann@1: +
yann@1:    if (unsorted_offsets[order[0]] == 0)
yann@1:      return 1; /* ldmia */
yann@1:  
yann@1: @@ -5064,6 +5072,11 @@
yann@1:        *load_offset = unsorted_offsets[order[0]];
yann@1:      }
yann@1:  
yann@1: +  /* For XScale a two-word LDM is a performance loss, so only do this if
yann@1: +     size is more important.  See comments in arm_gen_load_multiple.  */
yann@1: +  if (nops == 2 && arm_tune_xscale && !optimize_size)
yann@1: +    return 0;
yann@1: +
yann@1:    if (unsorted_offsets[order[0]] == 0)
yann@1:      return 1; /* stmia */
yann@1:  
yann@1: --- gcc-3.4.0/gcc/genpeep.c.arm-ldm-peephole	2003-07-05 01:27:22.000000000 -0400
yann@1: +++ gcc-3.4.0/gcc/genpeep.c	2004-04-24 18:18:04.000000000 -0400
yann@1: @@ -381,6 +381,7 @@
yann@1:    printf ("#include \"recog.h\"\n");
yann@1:    printf ("#include \"except.h\"\n\n");
yann@1:    printf ("#include \"function.h\"\n\n");
yann@1: +  printf ("#include \"flags.h\"\n\n");
yann@1:  
yann@1:    printf ("#ifdef HAVE_peephole\n");
yann@1:    printf ("extern rtx peep_operand[];\n\n");