openwrt/toolchain/gcc/3.4.2/600-gcc34-arm-ldm-peephole.patch

   1 --- gcc-3.4.0/gcc/config/arm/arm.md.arm-ldm-peephole    2004-01-13 08:24:37.000000000 -0500
   2 +++ gcc-3.4.0/gcc/config/arm/arm.md     2004-04-24 18:18:04.000000000 -0400
   3 @@ -8810,13 +8810,16 @@
   4     (set_attr "length" "4,8,8")]
   5  )
   6
   7 +; Try to convert LDR+LDR+arith into [add+]LDM+arith
   8 +; On XScale, LDM is always slower than two LDRs, so only do this if
   9 +; optimising for size.
  10  (define_insn "*arith_adjacentmem"
  11    [(set (match_operand:SI 0 "s_register_operand" "=r")
  12         (match_operator:SI 1 "shiftable_operator"
  13          [(match_operand:SI 2 "memory_operand" "m")
  14           (match_operand:SI 3 "memory_operand" "m")]))
  15     (clobber (match_scratch:SI 4 "=r"))]
  16 -  "TARGET_ARM && adjacent_mem_locations (operands[2], operands[3])"
  17 +  "TARGET_ARM && (!arm_tune_xscale || optimize_size) && adjacent_mem_locations (operands[2], operands[3])"
  18    "*
  19    {
  20      rtx ldm[3];
  21 @@ -8851,6 +8854,8 @@
  22        }
  23     if (val1 && val2)
  24        {
  25 +       /* This would be a loss on a Harvard core, but adjacent_mem_locations()
  26 +          will prevent it from happening.  */
  27         rtx ops[3];
  28         ldm[0] = ops[0] = operands[4];
  29         ops[1] = XEXP (XEXP (operands[2], 0), 0);
  30 --- gcc-3.4.0/gcc/config/arm/arm.c.arm-ldm-peephole     2004-04-24 18:16:25.000000000 -0400
  31 +++ gcc-3.4.0/gcc/config/arm/arm.c      2004-04-24 18:18:04.000000000 -0400
  32 @@ -4593,8 +4593,11 @@
  33          arith_adjacentmem pattern to output an overlong sequence.  */
  34        if (!const_ok_for_op (PLUS, val0) || !const_ok_for_op (PLUS, val1))
  35         return 0;
  36 -
  37 -      return (reg0 == reg1) && ((val1 - val0) == 4 || (val0 - val1) == 4);
  38 +
  39 +      /* For Harvard cores, only accept pairs where one offset is zero.
  40 +         See comment in load_multiple_sequence.  */
  41 +      return (reg0 == reg1) && ((val1 - val0) == 4 || (val0 - val1) == 4)
  42 +       && (!arm_ld_sched || val0 == 0 || val1 == 0);
  43      }
  44    return 0;
  45  }
  46 @@ -4838,6 +4841,11 @@
  47        *load_offset = unsorted_offsets[order[0]];
  48      }
  49
  50 +  /* For XScale a two-word LDM is a performance loss, so only do this if
  51 +     size is more important.  See comments in arm_gen_load_multiple.  */
  52 +  if (nops == 2 && arm_tune_xscale && !optimize_size)
  53 +    return 0;
  54 +
  55    if (unsorted_offsets[order[0]] == 0)
  56      return 1; /* ldmia */
  57
  58 @@ -5064,6 +5072,11 @@
  59        *load_offset = unsorted_offsets[order[0]];
  60      }
  61
  62 +  /* For XScale a two-word LDM is a performance loss, so only do this if
  63 +     size is more important.  See comments in arm_gen_load_multiple.  */
  64 +  if (nops == 2 && arm_tune_xscale && !optimize_size)
  65 +    return 0;
  66 +
  67    if (unsorted_offsets[order[0]] == 0)
  68      return 1; /* stmia */
  69
  70 --- gcc-3.4.0/gcc/genpeep.c.arm-ldm-peephole    2003-07-05 01:27:22.000000000 -0400
  71 +++ gcc-3.4.0/gcc/genpeep.c     2004-04-24 18:18:04.000000000 -0400
  72 @@ -381,6 +381,7 @@
  73    printf ("#include \"recog.h\"\n");
  74    printf ("#include \"except.h\"\n\n");
  75    printf ("#include \"function.h\"\n\n");
  76 +  printf ("#include \"flags.h\"\n\n");
  77
  78    printf ("#ifdef HAVE_peephole\n");
  79    printf ("extern rtx peep_operand[];\n\n");