Patrick Williams | 975a06f | 2022-10-21 14:42:47 -0500 | [diff] [blame^] | 1 | From 03d97c104f2d68cffd1bfc48cd62727e13a64712 Mon Sep 17 00:00:00 2001 |
| 2 | From: Rui Miguel Silva <rui.silva@linaro.org> |
| 3 | Date: Fri, 14 Oct 2022 17:42:52 +0100 |
| 4 | Subject: [PATCH] newlib: memcpy: remove optimized version |
| 5 | |
| 6 | When creating messages packed to send over openamp we may need |
| 7 | to do some copy in unaligned address, because of that we may |
| 8 | not always use the assembler optimized version, which will |
| 9 | trough a data-abort on aligned address exception. |
| 10 | |
| 11 | So, we may just use the version in string.h (the same used in |
| 12 | optee-os) that will take care to check and use different |
| 13 | optimization based on given source or destination address's. |
| 14 | |
| 15 | Upstream-Status: Pending |
| 16 | Signed-off-by: Rui Miguel Silva <rui.silva@linaro.org> |
| 17 | --- |
| 18 | newlib/libc/machine/aarch64/memcpy-stub.c | 2 +- |
| 19 | newlib/libc/machine/aarch64/memcpy.S | 166 ---------------------- |
| 20 | 2 files changed, 1 insertion(+), 167 deletions(-) |
| 21 | |
| 22 | diff --git a/newlib/libc/machine/aarch64/memcpy-stub.c b/newlib/libc/machine/aarch64/memcpy-stub.c |
| 23 | index cd6d72a8b8af..5f2b7968c7fc 100644 |
| 24 | --- a/newlib/libc/machine/aarch64/memcpy-stub.c |
| 25 | +++ b/newlib/libc/machine/aarch64/memcpy-stub.c |
| 26 | @@ -27,5 +27,5 @@ |
| 27 | #if (defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED)) |
| 28 | # include "../../string/memcpy.c" |
| 29 | #else |
| 30 | -/* See memcpy.S */ |
| 31 | +# include "../../string/memcpy.c" |
| 32 | #endif |
| 33 | diff --git a/newlib/libc/machine/aarch64/memcpy.S b/newlib/libc/machine/aarch64/memcpy.S |
| 34 | index 463bad0a1816..2a1460546374 100644 |
| 35 | --- a/newlib/libc/machine/aarch64/memcpy.S |
| 36 | +++ b/newlib/libc/machine/aarch64/memcpy.S |
| 37 | @@ -61,170 +61,4 @@ |
| 38 | #if (defined (__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED)) |
| 39 | /* See memcpy-stub.c */ |
| 40 | #else |
| 41 | - |
| 42 | -#define dstin x0 |
| 43 | -#define src x1 |
| 44 | -#define count x2 |
| 45 | -#define dst x3 |
| 46 | -#define srcend x4 |
| 47 | -#define dstend x5 |
| 48 | -#define A_l x6 |
| 49 | -#define A_lw w6 |
| 50 | -#define A_h x7 |
| 51 | -#define A_hw w7 |
| 52 | -#define B_l x8 |
| 53 | -#define B_lw w8 |
| 54 | -#define B_h x9 |
| 55 | -#define C_l x10 |
| 56 | -#define C_h x11 |
| 57 | -#define D_l x12 |
| 58 | -#define D_h x13 |
| 59 | -#define E_l src |
| 60 | -#define E_h count |
| 61 | -#define F_l srcend |
| 62 | -#define F_h dst |
| 63 | -#define tmp1 x9 |
| 64 | - |
| 65 | -#define L(l) .L ## l |
| 66 | - |
| 67 | - .macro def_fn f p2align=0 |
| 68 | - .text |
| 69 | - .p2align \p2align |
| 70 | - .global \f |
| 71 | - .type \f, %function |
| 72 | -\f: |
| 73 | - .endm |
| 74 | - |
| 75 | -/* Copies are split into 3 main cases: small copies of up to 16 bytes, |
| 76 | - medium copies of 17..96 bytes which are fully unrolled. Large copies |
| 77 | - of more than 96 bytes align the destination and use an unrolled loop |
| 78 | - processing 64 bytes per iteration. |
| 79 | - Small and medium copies read all data before writing, allowing any |
| 80 | - kind of overlap, and memmove tailcalls memcpy for these cases as |
| 81 | - well as non-overlapping copies. |
| 82 | -*/ |
| 83 | - |
| 84 | -def_fn memcpy p2align=6 |
| 85 | - prfm PLDL1KEEP, [src] |
| 86 | - add srcend, src, count |
| 87 | - add dstend, dstin, count |
| 88 | - cmp count, 16 |
| 89 | - b.ls L(copy16) |
| 90 | - cmp count, 96 |
| 91 | - b.hi L(copy_long) |
| 92 | - |
| 93 | - /* Medium copies: 17..96 bytes. */ |
| 94 | - sub tmp1, count, 1 |
| 95 | - ldp A_l, A_h, [src] |
| 96 | - tbnz tmp1, 6, L(copy96) |
| 97 | - ldp D_l, D_h, [srcend, -16] |
| 98 | - tbz tmp1, 5, 1f |
| 99 | - ldp B_l, B_h, [src, 16] |
| 100 | - ldp C_l, C_h, [srcend, -32] |
| 101 | - stp B_l, B_h, [dstin, 16] |
| 102 | - stp C_l, C_h, [dstend, -32] |
| 103 | -1: |
| 104 | - stp A_l, A_h, [dstin] |
| 105 | - stp D_l, D_h, [dstend, -16] |
| 106 | - ret |
| 107 | - |
| 108 | - .p2align 4 |
| 109 | - /* Small copies: 0..16 bytes. */ |
| 110 | -L(copy16): |
| 111 | - cmp count, 8 |
| 112 | - b.lo 1f |
| 113 | - ldr A_l, [src] |
| 114 | - ldr A_h, [srcend, -8] |
| 115 | - str A_l, [dstin] |
| 116 | - str A_h, [dstend, -8] |
| 117 | - ret |
| 118 | - .p2align 4 |
| 119 | -1: |
| 120 | - tbz count, 2, 1f |
| 121 | - ldr A_lw, [src] |
| 122 | - ldr A_hw, [srcend, -4] |
| 123 | - str A_lw, [dstin] |
| 124 | - str A_hw, [dstend, -4] |
| 125 | - ret |
| 126 | - |
| 127 | - /* Copy 0..3 bytes. Use a branchless sequence that copies the same |
| 128 | - byte 3 times if count==1, or the 2nd byte twice if count==2. */ |
| 129 | -1: |
| 130 | - cbz count, 2f |
| 131 | - lsr tmp1, count, 1 |
| 132 | - ldrb A_lw, [src] |
| 133 | - ldrb A_hw, [srcend, -1] |
| 134 | - ldrb B_lw, [src, tmp1] |
| 135 | - strb A_lw, [dstin] |
| 136 | - strb B_lw, [dstin, tmp1] |
| 137 | - strb A_hw, [dstend, -1] |
| 138 | -2: ret |
| 139 | - |
| 140 | - .p2align 4 |
| 141 | - /* Copy 64..96 bytes. Copy 64 bytes from the start and |
| 142 | - 32 bytes from the end. */ |
| 143 | -L(copy96): |
| 144 | - ldp B_l, B_h, [src, 16] |
| 145 | - ldp C_l, C_h, [src, 32] |
| 146 | - ldp D_l, D_h, [src, 48] |
| 147 | - ldp E_l, E_h, [srcend, -32] |
| 148 | - ldp F_l, F_h, [srcend, -16] |
| 149 | - stp A_l, A_h, [dstin] |
| 150 | - stp B_l, B_h, [dstin, 16] |
| 151 | - stp C_l, C_h, [dstin, 32] |
| 152 | - stp D_l, D_h, [dstin, 48] |
| 153 | - stp E_l, E_h, [dstend, -32] |
| 154 | - stp F_l, F_h, [dstend, -16] |
| 155 | - ret |
| 156 | - |
| 157 | - /* Align DST to 16 byte alignment so that we don't cross cache line |
| 158 | - boundaries on both loads and stores. There are at least 96 bytes |
| 159 | - to copy, so copy 16 bytes unaligned and then align. The loop |
| 160 | - copies 64 bytes per iteration and prefetches one iteration ahead. */ |
| 161 | - |
| 162 | - .p2align 4 |
| 163 | -L(copy_long): |
| 164 | - and tmp1, dstin, 15 |
| 165 | - bic dst, dstin, 15 |
| 166 | - ldp D_l, D_h, [src] |
| 167 | - sub src, src, tmp1 |
| 168 | - add count, count, tmp1 /* Count is now 16 too large. */ |
| 169 | - ldp A_l, A_h, [src, 16] |
| 170 | - stp D_l, D_h, [dstin] |
| 171 | - ldp B_l, B_h, [src, 32] |
| 172 | - ldp C_l, C_h, [src, 48] |
| 173 | - ldp D_l, D_h, [src, 64]! |
| 174 | - subs count, count, 128 + 16 /* Test and readjust count. */ |
| 175 | - b.ls 2f |
| 176 | -1: |
| 177 | - stp A_l, A_h, [dst, 16] |
| 178 | - ldp A_l, A_h, [src, 16] |
| 179 | - stp B_l, B_h, [dst, 32] |
| 180 | - ldp B_l, B_h, [src, 32] |
| 181 | - stp C_l, C_h, [dst, 48] |
| 182 | - ldp C_l, C_h, [src, 48] |
| 183 | - stp D_l, D_h, [dst, 64]! |
| 184 | - ldp D_l, D_h, [src, 64]! |
| 185 | - subs count, count, 64 |
| 186 | - b.hi 1b |
| 187 | - |
| 188 | - /* Write the last full set of 64 bytes. The remainder is at most 64 |
| 189 | - bytes, so it is safe to always copy 64 bytes from the end even if |
| 190 | - there is just 1 byte left. */ |
| 191 | -2: |
| 192 | - ldp E_l, E_h, [srcend, -64] |
| 193 | - stp A_l, A_h, [dst, 16] |
| 194 | - ldp A_l, A_h, [srcend, -48] |
| 195 | - stp B_l, B_h, [dst, 32] |
| 196 | - ldp B_l, B_h, [srcend, -32] |
| 197 | - stp C_l, C_h, [dst, 48] |
| 198 | - ldp C_l, C_h, [srcend, -16] |
| 199 | - stp D_l, D_h, [dst, 64] |
| 200 | - stp E_l, E_h, [dstend, -64] |
| 201 | - stp A_l, A_h, [dstend, -48] |
| 202 | - stp B_l, B_h, [dstend, -32] |
| 203 | - stp C_l, C_h, [dstend, -16] |
| 204 | - ret |
| 205 | - |
| 206 | - .size memcpy, . - memcpy |
| 207 | #endif |
| 208 | -- |
| 209 | 2.38.0 |
| 210 | |