patches/gcc/3.4.6/140-arm-ldm-peephole.patch
author "Yann E. MORIN" <yann.morin.1998@anciens.enib.fr>
Wed Dec 30 15:36:22 2009 +0100 (2009-12-30)
changeset 1669 61edd9d19e3c
parent 746 b150d6f590fc
permissions -rw-r--r--
scripts/functions: add aria2, a powerfull downloader

aria2 is a powerfull downloader that is capable of chunking and
parallel retrieval.

Due to li;itations in crosstool-NG retrieval facilities, it's not possible
to take fully advantage of aria2. It might happen that, in the future,
those limitations get lifted away, so we can take use features such as
parallel downloading from more than one server at the same time. For now,
it should still speed up downloads thanks to parallel downloading of chunks.
     1 diff -durN gcc-3.4.6.orig/gcc/config/arm/arm.c gcc-3.4.6/gcc/config/arm/arm.c
     2 --- gcc-3.4.6.orig/gcc/config/arm/arm.c	2005-10-01 15:31:38.000000000 +0200
     3 +++ gcc-3.4.6/gcc/config/arm/arm.c	2007-08-15 22:54:59.000000000 +0200
     4 @@ -4857,6 +4857,11 @@
     5        *load_offset = unsorted_offsets[order[0]];
     6      }
     7  
     8 +  /* For XScale a two-word LDM is a performance loss, so only do this if
     9 +     size is more important.  See comments in arm_gen_load_multiple.  */
    10 +  if (nops == 2 && arm_tune_xscale && !optimize_size)
    11 +    return 0;
    12 +
    13    if (unsorted_offsets[order[0]] == 0)
    14      return 1; /* ldmia */
    15  
    16 @@ -5083,6 +5088,11 @@
    17        *load_offset = unsorted_offsets[order[0]];
    18      }
    19  
    20 +  /* For XScale a two-word LDM is a performance loss, so only do this if
    21 +     size is more important.  See comments in arm_gen_load_multiple.  */
    22 +  if (nops == 2 && arm_tune_xscale && !optimize_size)
    23 +    return 0;
    24 +
    25    if (unsorted_offsets[order[0]] == 0)
    26      return 1; /* stmia */
    27  
    28 diff -durN gcc-3.4.6.orig/gcc/config/arm/arm.md gcc-3.4.6/gcc/config/arm/arm.md
    29 --- gcc-3.4.6.orig/gcc/config/arm/arm.md	2005-10-01 15:31:38.000000000 +0200
    30 +++ gcc-3.4.6/gcc/config/arm/arm.md	2007-08-15 22:54:59.000000000 +0200
    31 @@ -8811,13 +8811,16 @@
    32     (set_attr "length" "4,8,8")]
    33  )
    34  
    35 +; Try to convert LDR+LDR+arith into [add+]LDM+arith
    36 +; On XScale, LDM is always slower than two LDRs, so only do this if
    37 +; optimising for size.
    38  (define_insn "*arith_adjacentmem"
    39    [(set (match_operand:SI 0 "s_register_operand" "=r")
    40  	(match_operator:SI 1 "shiftable_operator"
    41  	 [(match_operand:SI 2 "memory_operand" "m")
    42  	  (match_operand:SI 3 "memory_operand" "m")]))
    43     (clobber (match_scratch:SI 4 "=r"))]
    44 -  "TARGET_ARM && adjacent_mem_locations (operands[2], operands[3])"
    45 +  "TARGET_ARM && (!arm_tune_xscale || optimize_size) && adjacent_mem_locations (operands[2], operands[3])"
    46    "*
    47    {
    48      rtx ldm[3];
    49 @@ -8852,6 +8855,8 @@
    50        }
    51     if (val1 && val2)
    52        {
    53 +	/* This would be a loss on a Harvard core, but adjacent_mem_locations()
    54 +	   will prevent it from happening.  */
    55  	rtx ops[3];
    56  	ldm[0] = ops[0] = operands[4];
    57  	ops[1] = XEXP (XEXP (operands[2], 0), 0);
    58 diff -durN gcc-3.4.6.orig/gcc/genpeep.c gcc-3.4.6/gcc/genpeep.c
    59 --- gcc-3.4.6.orig/gcc/genpeep.c	2003-07-05 07:27:22.000000000 +0200
    60 +++ gcc-3.4.6/gcc/genpeep.c	2007-08-15 22:54:59.000000000 +0200
    61 @@ -381,6 +381,7 @@
    62    printf ("#include \"recog.h\"\n");
    63    printf ("#include \"except.h\"\n\n");
    64    printf ("#include \"function.h\"\n\n");
    65 +  printf ("#include \"flags.h\"\n\n");
    66  
    67    printf ("#ifdef HAVE_peephole\n");
    68    printf ("extern rtx peep_operand[];\n\n");