Arturo Pesaro / lwip-sys

lwip-sys with changes

Fork of lwip-sys by mbed official

arch/checksum.c@10:09b0951b1899, 2013-09-10 (annotated)

Committer:
bogdanm
Date:
Tue Sep 10 15:14:42 2013 +0300
Revision:
10:09b0951b1899
Sync with git revision 171dda705c947bf910926a0b73d6a4797802554d

Who changed what in which revision?

UserRevisionLine numberNew contents of line
bogdanm 10:09b0951b1899 1 /* Copyright (C) 2013 - Adam Green (https://github.com/adamgreen)
bogdanm 10:09b0951b1899 2
bogdanm 10:09b0951b1899 3 Licensed under the Apache License, Version 2.0 (the "License");
bogdanm 10:09b0951b1899 4 you may not use this file except in compliance with the License.
bogdanm 10:09b0951b1899 5 You may obtain a copy of the License at
bogdanm 10:09b0951b1899 6
bogdanm 10:09b0951b1899 7 http://www.apache.org/licenses/LICENSE-2.0
bogdanm 10:09b0951b1899 8
bogdanm 10:09b0951b1899 9 Unless required by applicable law or agreed to in writing, software
bogdanm 10:09b0951b1899 10 distributed under the License is distributed on an "AS IS" BASIS,
bogdanm 10:09b0951b1899 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
bogdanm 10:09b0951b1899 12 See the License for the specific language governing permissions and
bogdanm 10:09b0951b1899 13 limitations under the License.
bogdanm 10:09b0951b1899 14 */
bogdanm 10:09b0951b1899 15 #if defined(TOOLCHAIN_GCC) && defined(__thumb2__)
bogdanm 10:09b0951b1899 16
bogdanm 10:09b0951b1899 17
bogdanm 10:09b0951b1899 18 /* This is a hand written Thumb-2 assembly language version of the
bogdanm 10:09b0951b1899 19 algorithm 3 version of lwip_standard_chksum in lwIP's inet_chksum.c. It
bogdanm 10:09b0951b1899 20 performs the checksumming 32-bits at a time and even unrolls the loop to
bogdanm 10:09b0951b1899 21 perform two of these 32-bit adds per loop iteration.
bogdanm 10:09b0951b1899 22
bogdanm 10:09b0951b1899 23 Returns:
bogdanm 10:09b0951b1899 24 16-bit 1's complement summation (not inversed).
bogdanm 10:09b0951b1899 25
bogdanm 10:09b0951b1899 26 NOTE: This function does return a uint16_t from the assembly language code
bogdanm 10:09b0951b1899 27 but is marked as void so that GCC doesn't issue warning because it
bogdanm 10:09b0951b1899 28 doesn't know about this low level return.
bogdanm 10:09b0951b1899 29 */
bogdanm 10:09b0951b1899 30 __attribute__((naked)) void /*uint16_t*/ thumb2_checksum(const void* pData, int length)
bogdanm 10:09b0951b1899 31 {
bogdanm 10:09b0951b1899 32 __asm (
bogdanm 10:09b0951b1899 33 ".syntax unified\n"
bogdanm 10:09b0951b1899 34 ".thumb\n"
bogdanm 10:09b0951b1899 35
bogdanm 10:09b0951b1899 36 // Push non-volatile registers we use on stack. Push link register too to
bogdanm 10:09b0951b1899 37 // keep stack 8-byte aligned and allow single pop to restore and return.
bogdanm 10:09b0951b1899 38 " push {r4, lr}\n"
bogdanm 10:09b0951b1899 39 // Initialize sum, r2, to 0.
bogdanm 10:09b0951b1899 40 " movs r2, #0\n"
bogdanm 10:09b0951b1899 41 // Remember whether pData was at odd address in r3. This is used later to
bogdanm 10:09b0951b1899 42 // know if it needs to swap the result since the summation will be done at
bogdanm 10:09b0951b1899 43 // an offset of 1, rather than 0.
bogdanm 10:09b0951b1899 44 " ands r3, r0, #1\n"
bogdanm 10:09b0951b1899 45 // Need to 2-byte align? If not skip ahead.
bogdanm 10:09b0951b1899 46 " beq 1$\n"
bogdanm 10:09b0951b1899 47 // We can return if there are no bytes to sum.
bogdanm 10:09b0951b1899 48 " cbz r1, 9$\n"
bogdanm 10:09b0951b1899 49
bogdanm 10:09b0951b1899 50 // 2-byte align.
bogdanm 10:09b0951b1899 51 // Place the first data byte in odd summation location since it needs to be
bogdanm 10:09b0951b1899 52 // swapped later. It's ok to overwrite r2 here as it only had a value of 0
bogdanm 10:09b0951b1899 53 // up until now. Advance r0 pointer and decrement r1 length as we go.
bogdanm 10:09b0951b1899 54 " ldrb r2, [r0], #1\n"
bogdanm 10:09b0951b1899 55 " lsls r2, r2, #8\n"
bogdanm 10:09b0951b1899 56 " subs r1, r1, #1\n"
bogdanm 10:09b0951b1899 57
bogdanm 10:09b0951b1899 58 // Need to 4-byte align? If not skip ahead.
bogdanm 10:09b0951b1899 59 "1$:\n"
bogdanm 10:09b0951b1899 60 " ands r4, r0, #3\n"
bogdanm 10:09b0951b1899 61 " beq 2$\n"
bogdanm 10:09b0951b1899 62 // Have more than 1 byte left to align? If not skip ahead to take care of
bogdanm 10:09b0951b1899 63 // trailing byte.
bogdanm 10:09b0951b1899 64 " cmp r1, #2\n"
bogdanm 10:09b0951b1899 65 " blt 7$\n"
bogdanm 10:09b0951b1899 66
bogdanm 10:09b0951b1899 67 // 4-byte align.
bogdanm 10:09b0951b1899 68 " ldrh r4, [r0], #2\n"
bogdanm 10:09b0951b1899 69 " adds r2, r2, r4\n"
bogdanm 10:09b0951b1899 70 " subs r1, r1, #2\n"
bogdanm 10:09b0951b1899 71
bogdanm 10:09b0951b1899 72 // Main summing loop which sums up data 2 words at a time.
bogdanm 10:09b0951b1899 73 // Make sure that we have more than 7 bytes left to sum.
bogdanm 10:09b0951b1899 74 "2$:\n"
bogdanm 10:09b0951b1899 75 " cmp r1, #8\n"
bogdanm 10:09b0951b1899 76 " blt 3$\n"
bogdanm 10:09b0951b1899 77 // Sum next two words. Applying previous upper 16-bit carry to
bogdanm 10:09b0951b1899 78 // lower 16-bits.
bogdanm 10:09b0951b1899 79 " ldr r4, [r0], #4\n"
bogdanm 10:09b0951b1899 80 " adds r2, r4\n"
bogdanm 10:09b0951b1899 81 " adc r2, r2, #0\n"
bogdanm 10:09b0951b1899 82 " ldr r4, [r0], #4\n"
bogdanm 10:09b0951b1899 83 " adds r2, r4\n"
bogdanm 10:09b0951b1899 84 " adc r2, r2, #0\n"
bogdanm 10:09b0951b1899 85 " subs r1, r1, #8\n"
bogdanm 10:09b0951b1899 86 " b 2$\n"
bogdanm 10:09b0951b1899 87
bogdanm 10:09b0951b1899 88 // Sum up any remaining half-words.
bogdanm 10:09b0951b1899 89 "3$:\n"
bogdanm 10:09b0951b1899 90 // Make sure that we have more than 1 byte left to sum.
bogdanm 10:09b0951b1899 91 " cmp r1, #2\n"
bogdanm 10:09b0951b1899 92 " blt 7$\n"
bogdanm 10:09b0951b1899 93 // Sum up next half word, continue to apply carry.
bogdanm 10:09b0951b1899 94 " ldrh r4, [r0], #2\n"
bogdanm 10:09b0951b1899 95 " adds r2, r4\n"
bogdanm 10:09b0951b1899 96 " adc r2, r2, #0\n"
bogdanm 10:09b0951b1899 97 " subs r1, r1, #2\n"
bogdanm 10:09b0951b1899 98 " b 3$\n"
bogdanm 10:09b0951b1899 99
bogdanm 10:09b0951b1899 100 // Handle trailing byte, if it exists
bogdanm 10:09b0951b1899 101 "7$:\n"
bogdanm 10:09b0951b1899 102 " cbz r1, 8$\n"
bogdanm 10:09b0951b1899 103 " ldrb r4, [r0]\n"
bogdanm 10:09b0951b1899 104 " adds r2, r4\n"
bogdanm 10:09b0951b1899 105 " adc r2, r2, #0\n"
bogdanm 10:09b0951b1899 106
bogdanm 10:09b0951b1899 107 // Fold 32-bit checksum into 16-bit checksum.
bogdanm 10:09b0951b1899 108 "8$:\n"
bogdanm 10:09b0951b1899 109 " ubfx r4, r2, #16, #16\n"
bogdanm 10:09b0951b1899 110 " ubfx r2, r2, #0, #16\n"
bogdanm 10:09b0951b1899 111 " adds r2, r4\n"
bogdanm 10:09b0951b1899 112 " ubfx r4, r2, #16, #16\n"
bogdanm 10:09b0951b1899 113 " ubfx r2, r2, #0, #16\n"
bogdanm 10:09b0951b1899 114 " adds r2, r4\n"
bogdanm 10:09b0951b1899 115
bogdanm 10:09b0951b1899 116 // Swap bytes if started at odd address
bogdanm 10:09b0951b1899 117 " cbz r3, 9$\n"
bogdanm 10:09b0951b1899 118 " rev16 r2, r2\n"
bogdanm 10:09b0951b1899 119
bogdanm 10:09b0951b1899 120 // Return final sum.
bogdanm 10:09b0951b1899 121 "9$: mov r0, r2\n"
bogdanm 10:09b0951b1899 122 " pop {r4, pc}\n"
bogdanm 10:09b0951b1899 123 );
bogdanm 10:09b0951b1899 124 }
bogdanm 10:09b0951b1899 125
bogdanm 10:09b0951b1899 126 #endif