diff options
author | Marcus Shawcroft <marcus.shawcroft@linaro.org> | 2013-01-16 13:28:39 +0000 |
---|---|---|
committer | Marcus Shawcroft <marcus.shawcroft@linaro.org> | 2013-01-17 10:56:49 +0000 |
commit | c2b6221ed3be421c2f7a8ad32e70ad501340b55b (patch) | |
tree | 892be23e2afe0ea27879d2122457a554c3baf00d /ports/sysdeps/aarch64/memmove.S | |
parent | 857c8d222813f26e5178d3de1f272d1c3d9e24ac (diff) | |
download | glibc-c2b6221ed3be421c2f7a8ad32e70ad501340b55b.tar.gz glibc-c2b6221ed3be421c2f7a8ad32e70ad501340b55b.tar.xz glibc-c2b6221ed3be421c2f7a8ad32e70ad501340b55b.zip |
AArch64: Implement optimized memmove.
Diffstat (limited to 'ports/sysdeps/aarch64/memmove.S')
-rw-r--r-- | ports/sysdeps/aarch64/memmove.S | 312 |
1 files changed, 312 insertions, 0 deletions
diff --git a/ports/sysdeps/aarch64/memmove.S b/ports/sysdeps/aarch64/memmove.S new file mode 100644 index 0000000000..c42eb1c13a --- /dev/null +++ b/ports/sysdeps/aarch64/memmove.S @@ -0,0 +1,312 @@ +/* Copyright (C) 2012-2013 Free Software Foundation, Inc. + + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library. If not, see + <http://www.gnu.org/licenses/>. */ + +#include <sysdep.h> + +/* Assumptions: + * + * ARMv8-a, AArch64 + * Unaligned accesses + */ + +/* Parameters and result. */ +#define dstin x0 +#define src x1 +#define count x2 +#define tmp1 x3 +#define tmp1w w3 +#define tmp2 x4 +#define tmp2w w4 +#define tmp3 x5 +#define tmp3w w5 +#define dst x6 + +#define A_l x7 +#define A_h x8 +#define B_l x9 +#define B_h x10 +#define C_l x11 +#define C_h x12 +#define D_l x13 +#define D_h x14 + +ENTRY_ALIGN (memmove, 6) + + cmp dstin, src + b.lo L(downwards) + add tmp1, src, count + cmp dstin, tmp1 + b.hs memcpy /* No overlap. */ + + /* Upwards move with potential overlap. + * Need to move from the tail backwards. SRC and DST point one + * byte beyond the remaining data to move. */ + add dst, dstin, count + add src, src, count + cmp count, #64 + b.ge L(mov_not_short_up) + + /* Deal with small moves quickly by dropping straight into the + * exit block. */ +L(tail63up): + /* Move up to 48 bytes of data. At this point we only need the + * bottom 6 bits of count to be accurate. */ + ands tmp1, count, #0x30 + b.eq L(tail15up) + sub dst, dst, tmp1 + sub src, src, tmp1 + cmp tmp1w, #0x20 + b.eq 1f + b.lt 2f + ldp A_l, A_h, [src, #32] + stp A_l, A_h, [dst, #32] +1: + ldp A_l, A_h, [src, #16] + stp A_l, A_h, [dst, #16] +2: + ldp A_l, A_h, [src] + stp A_l, A_h, [dst] +L(tail15up): + /* Move up to 15 bytes of data. Does not assume additional data + * being moved. */ + tbz count, #3, 1f + ldr tmp1, [src, #-8]! + str tmp1, [dst, #-8]! +1: + tbz count, #2, 1f + ldr tmp1w, [src, #-4]! + str tmp1w, [dst, #-4]! +1: + tbz count, #1, 1f + ldrh tmp1w, [src, #-2]! + strh tmp1w, [dst, #-2]! +1: + tbz count, #0, 1f + ldrb tmp1w, [src, #-1] + strb tmp1w, [dst, #-1] +1: + RET + +L(mov_not_short_up): + /* We don't much care about the alignment of DST, but we want SRC + * to be 128-bit (16 byte) aligned so that we don't cross cache line + * boundaries on both loads and stores. */ + ands tmp2, src, #15 /* Bytes to reach alignment. */ + b.eq 2f + sub count, count, tmp2 + /* Move enough data to reach alignment; unlike memcpy, we have to + * be aware of the overlap, which means we can't move data twice. */ + tbz tmp2, #3, 1f + ldr tmp1, [src, #-8]! + str tmp1, [dst, #-8]! +1: + tbz tmp2, #2, 1f + ldr tmp1w, [src, #-4]! + str tmp1w, [dst, #-4]! +1: + tbz tmp2, #1, 1f + ldrh tmp1w, [src, #-2]! + strh tmp1w, [dst, #-2]! +1: + tbz tmp2, #0, 1f + ldrb tmp1w, [src, #-1]! + strb tmp1w, [dst, #-1]! +1: + + /* There may be less than 63 bytes to go now. */ + cmp count, #63 + b.le L(tail63up) +2: + subs count, count, #128 + b.ge L(mov_body_large_up) + /* Less than 128 bytes to move, so handle 64 here and then jump + * to the tail. */ + ldp A_l, A_h, [src, #-64]! + ldp B_l, B_h, [src, #16] + ldp C_l, C_h, [src, #32] + ldp D_l, D_h, [src, #48] + stp A_l, A_h, [dst, #-64]! + stp B_l, B_h, [dst, #16] + stp C_l, C_h, [dst, #32] + stp D_l, D_h, [dst, #48] + tst count, #0x3f + b.ne L(tail63up) + RET + + /* Critical loop. Start at a new Icache line boundary. Assuming + * 64 bytes per line this ensures the entire loop is in one line. */ + .p2align 6 +L(mov_body_large_up): + /* There are at least 128 bytes to move. */ + ldp A_l, A_h, [src, #-16] + ldp B_l, B_h, [src, #-32] + ldp C_l, C_h, [src, #-48] + ldp D_l, D_h, [src, #-64]! +1: + stp A_l, A_h, [dst, #-16] + ldp A_l, A_h, [src, #-16] + stp B_l, B_h, [dst, #-32] + ldp B_l, B_h, [src, #-32] + stp C_l, C_h, [dst, #-48] + ldp C_l, C_h, [src, #-48] + stp D_l, D_h, [dst, #-64]! + ldp D_l, D_h, [src, #-64]! + subs count, count, #64 + b.ge 1b + stp A_l, A_h, [dst, #-16] + stp B_l, B_h, [dst, #-32] + stp C_l, C_h, [dst, #-48] + stp D_l, D_h, [dst, #-64]! + tst count, #0x3f + b.ne L(tail63up) + RET + +L(downwards): + /* For a downwards move we can safely use memcpy provided that + * DST is more than 16 bytes away from SRC. */ + sub tmp1, src, #16 + cmp dstin, tmp1 + b.ls memcpy /* May overlap, but not critically. */ + + mov dst, dstin /* Preserve DSTIN for return value. */ + cmp count, #64 + b.ge L(mov_not_short_down) + + /* Deal with small moves quickly by dropping straight into the + * exit block. */ +L(tail63down): + /* Move up to 48 bytes of data. At this point we only need the + * bottom 6 bits of count to be accurate. */ + ands tmp1, count, #0x30 + b.eq L(tail15down) + add dst, dst, tmp1 + add src, src, tmp1 + cmp tmp1w, #0x20 + b.eq 1f + b.lt 2f + ldp A_l, A_h, [src, #-48] + stp A_l, A_h, [dst, #-48] +1: + ldp A_l, A_h, [src, #-32] + stp A_l, A_h, [dst, #-32] +2: + ldp A_l, A_h, [src, #-16] + stp A_l, A_h, [dst, #-16] +L(tail15down): + /* Move up to 15 bytes of data. Does not assume additional data + being moved. */ + tbz count, #3, 1f + ldr tmp1, [src], #8 + str tmp1, [dst], #8 +1: + tbz count, #2, 1f + ldr tmp1w, [src], #4 + str tmp1w, [dst], #4 +1: + tbz count, #1, 1f + ldrh tmp1w, [src], #2 + strh tmp1w, [dst], #2 +1: + tbz count, #0, 1f + ldrb tmp1w, [src] + strb tmp1w, [dst] +1: + RET + +L(mov_not_short_down): + /* We don't much care about the alignment of DST, but we want SRC + * to be 128-bit (16 byte) aligned so that we don't cross cache line + * boundaries on both loads and stores. */ + neg tmp2, src + ands tmp2, tmp2, #15 /* Bytes to reach alignment. */ + b.eq 2f + sub count, count, tmp2 + /* Move enough data to reach alignment; unlike memcpy, we have to + * be aware of the overlap, which means we can't move data twice. */ + tbz tmp2, #3, 1f + ldr tmp1, [src], #8 + str tmp1, [dst], #8 +1: + tbz tmp2, #2, 1f + ldr tmp1w, [src], #4 + str tmp1w, [dst], #4 +1: + tbz tmp2, #1, 1f + ldrh tmp1w, [src], #2 + strh tmp1w, [dst], #2 +1: + tbz tmp2, #0, 1f + ldrb tmp1w, [src], #1 + strb tmp1w, [dst], #1 +1: + + /* There may be less than 63 bytes to go now. */ + cmp count, #63 + b.le L(tail63down) +2: + subs count, count, #128 + b.ge L(mov_body_large_down) + /* Less than 128 bytes to move, so handle 64 here and then jump + * to the tail. */ + ldp A_l, A_h, [src] + ldp B_l, B_h, [src, #16] + ldp C_l, C_h, [src, #32] + ldp D_l, D_h, [src, #48] + stp A_l, A_h, [dst] + stp B_l, B_h, [dst, #16] + stp C_l, C_h, [dst, #32] + stp D_l, D_h, [dst, #48] + tst count, #0x3f + add src, src, #64 + add dst, dst, #64 + b.ne L(tail63down) + RET + + /* Critical loop. Start at a new cache line boundary. Assuming + * 64 bytes per line this ensures the entire loop is in one line. */ + .p2align 6 +L(mov_body_large_down): + /* There are at least 128 bytes to move. */ + ldp A_l, A_h, [src, #0] + sub dst, dst, #16 /* Pre-bias. */ + ldp B_l, B_h, [src, #16] + ldp C_l, C_h, [src, #32] + ldp D_l, D_h, [src, #48]! /* src += 64 - Pre-bias. */ +1: + stp A_l, A_h, [dst, #16] + ldp A_l, A_h, [src, #16] + stp B_l, B_h, [dst, #32] + ldp B_l, B_h, [src, #32] + stp C_l, C_h, [dst, #48] + ldp C_l, C_h, [src, #48] + stp D_l, D_h, [dst, #64]! + ldp D_l, D_h, [src, #64]! + subs count, count, #64 + b.ge 1b + stp A_l, A_h, [dst, #16] + stp B_l, B_h, [dst, #32] + stp C_l, C_h, [dst, #48] + stp D_l, D_h, [dst, #64] + add src, src, #16 + add dst, dst, #64 + 16 + tst count, #0x3f + b.ne L(tail63down) + RET +END (memmove) + +libc_hidden_builtin_def (memmove) |