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