1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/patches/gcc/3.4.4/600-gcc34-arm-ldm-peephole.patch Sun May 27 20:22:06 2007 +0000
1.3 @@ -0,0 +1,65 @@
1.4 +--- gcc-3.4.0/gcc/config/arm/arm.md.arm-ldm-peephole 2004-01-13 08:24:37.000000000 -0500
1.5 ++++ gcc-3.4.0/gcc/config/arm/arm.md 2004-04-24 18:18:04.000000000 -0400
1.6 +@@ -8810,13 +8810,16 @@
1.7 + (set_attr "length" "4,8,8")]
1.8 + )
1.9 +
1.10 ++; Try to convert LDR+LDR+arith into [add+]LDM+arith
1.11 ++; On XScale, LDM is always slower than two LDRs, so only do this if
1.12 ++; optimising for size.
1.13 + (define_insn "*arith_adjacentmem"
1.14 + [(set (match_operand:SI 0 "s_register_operand" "=r")
1.15 + (match_operator:SI 1 "shiftable_operator"
1.16 + [(match_operand:SI 2 "memory_operand" "m")
1.17 + (match_operand:SI 3 "memory_operand" "m")]))
1.18 + (clobber (match_scratch:SI 4 "=r"))]
1.19 +- "TARGET_ARM && adjacent_mem_locations (operands[2], operands[3])"
1.20 ++ "TARGET_ARM && (!arm_tune_xscale || optimize_size) && adjacent_mem_locations (operands[2], operands[3])"
1.21 + "*
1.22 + {
1.23 + rtx ldm[3];
1.24 +@@ -8851,6 +8854,8 @@
1.25 + }
1.26 + if (val1 && val2)
1.27 + {
1.28 ++ /* This would be a loss on a Harvard core, but adjacent_mem_locations()
1.29 ++ will prevent it from happening. */
1.30 + rtx ops[3];
1.31 + ldm[0] = ops[0] = operands[4];
1.32 + ops[1] = XEXP (XEXP (operands[2], 0), 0);
1.33 +--- gcc-3.4.0/gcc/config/arm/arm.c.arm-ldm-peephole 2004-04-24 18:16:25.000000000 -0400
1.34 ++++ gcc-3.4.0/gcc/config/arm/arm.c 2004-04-24 18:18:04.000000000 -0400
1.35 +@@ -4838,6 +4841,11 @@
1.36 + *load_offset = unsorted_offsets[order[0]];
1.37 + }
1.38 +
1.39 ++ /* For XScale a two-word LDM is a performance loss, so only do this if
1.40 ++ size is more important. See comments in arm_gen_load_multiple. */
1.41 ++ if (nops == 2 && arm_tune_xscale && !optimize_size)
1.42 ++ return 0;
1.43 ++
1.44 + if (unsorted_offsets[order[0]] == 0)
1.45 + return 1; /* ldmia */
1.46 +
1.47 +@@ -5064,6 +5072,11 @@
1.48 + *load_offset = unsorted_offsets[order[0]];
1.49 + }
1.50 +
1.51 ++ /* For XScale a two-word LDM is a performance loss, so only do this if
1.52 ++ size is more important. See comments in arm_gen_load_multiple. */
1.53 ++ if (nops == 2 && arm_tune_xscale && !optimize_size)
1.54 ++ return 0;
1.55 ++
1.56 + if (unsorted_offsets[order[0]] == 0)
1.57 + return 1; /* stmia */
1.58 +
1.59 +--- gcc-3.4.0/gcc/genpeep.c.arm-ldm-peephole 2003-07-05 01:27:22.000000000 -0400
1.60 ++++ gcc-3.4.0/gcc/genpeep.c 2004-04-24 18:18:04.000000000 -0400
1.61 +@@ -381,6 +381,7 @@
1.62 + printf ("#include \"recog.h\"\n");
1.63 + printf ("#include \"except.h\"\n\n");
1.64 + printf ("#include \"function.h\"\n\n");
1.65 ++ printf ("#include \"flags.h\"\n\n");
1.66 +
1.67 + printf ("#ifdef HAVE_peephole\n");
1.68 + printf ("extern rtx peep_operand[];\n\n");