| agl@chromium.org | 21417a7 | 2010-08-13 17:05:28 +0000 | [diff] [blame] | 1 | /* |
| epoger@google.com | fd03db0 | 2011-07-28 14:24:55 +0000 | [diff] [blame] | 2 | * Copyright 2010 The Android Open Source Project |
| agl@chromium.org | 21417a7 | 2010-08-13 17:05:28 +0000 | [diff] [blame] | 3 | * |
| epoger@google.com | fd03db0 | 2011-07-28 14:24:55 +0000 | [diff] [blame] | 4 | * Use of this source code is governed by a BSD-style license that can be |
| 5 | * found in the LICENSE file. |
| agl@chromium.org | 21417a7 | 2010-08-13 17:05:28 +0000 | [diff] [blame] | 6 | */ |
| 7 | |
| 8 | /* Changes: |
| 9 | * 2010-08-11 Steve McIntyre <steve.mcintyre@arm.com> |
| 10 | * Added small changes to the two functions to make them work on the |
| 11 | * specified number of 16- or 32-bit values rather than the original |
| 12 | * code which was specified as a count of bytes. More verbose comments |
| 13 | * to aid future maintenance. |
| 14 | */ |
| 15 | |
| 16 | .text |
| 17 | .align |
| 18 | |
| 19 | .global arm_memset32 |
| 20 | .type arm_memset32, %function |
| 21 | .global arm_memset16 |
| 22 | .type arm_memset16, %function |
| 23 | |
| 24 | /* |
| 25 | * Optimized memset functions for ARM. |
| 26 | * |
| 27 | * void arm_memset16(uint16_t* dst, uint16_t value, int count); |
| 28 | * void arm_memset32(uint32_t* dst, uint32_t value, int count); |
| 29 | * |
| 30 | */ |
| 31 | arm_memset16: |
| 32 | .fnstart |
| 33 | push {lr} |
| 34 | |
| djsollen@google.com | a44e6c6 | 2012-01-09 14:38:25 +0000 | [diff] [blame^] | 35 | /* Multiply count by 2 - go from the number of 16-bit shorts |
| 36 | * to the number of bytes desired. */ |
| 37 | mov r2, r2, lsl #1 |
| 38 | |
| agl@chromium.org | 21417a7 | 2010-08-13 17:05:28 +0000 | [diff] [blame] | 39 | /* expand the data to 32 bits */ |
| 40 | orr r1, r1, lsl #16 |
| 41 | |
| 42 | /* align to 32 bits */ |
| 43 | tst r0, #2 |
| 44 | strneh r1, [r0], #2 |
| 45 | subne r2, r2, #2 |
| 46 | |
| agl@chromium.org | 21417a7 | 2010-08-13 17:05:28 +0000 | [diff] [blame] | 47 | /* Now jump into the main loop below. */ |
| 48 | b .Lwork_32 |
| 49 | .fnend |
| 50 | |
| 51 | arm_memset32: |
| 52 | .fnstart |
| 53 | push {lr} |
| 54 | |
| 55 | /* Multiply count by 4 - go from the number of 32-bit words to |
| 56 | * the number of bytes desired. */ |
| 57 | mov r2, r2, lsl #2 |
| 58 | |
| 59 | .Lwork_32: |
| 60 | /* Set up registers ready for writing them out. */ |
| 61 | mov ip, r1 |
| 62 | mov lr, r1 |
| 63 | |
| 64 | /* Try to align the destination to a cache line. Assume 32 |
| 65 | * byte (8 word) cache lines, it's the common case. */ |
| 66 | rsb r3, r0, #0 |
| 67 | ands r3, r3, #0x1C |
| 68 | beq .Laligned32 |
| 69 | cmp r3, r2 |
| 70 | andhi r3, r2, #0x1C |
| 71 | sub r2, r2, r3 |
| 72 | |
| 73 | /* (Optionally) write any unaligned leading bytes. |
| 74 | * (0-28 bytes, length in r3) */ |
| 75 | movs r3, r3, lsl #28 |
| 76 | stmcsia r0!, {r1, lr} |
| 77 | stmcsia r0!, {r1, lr} |
| 78 | stmmiia r0!, {r1, lr} |
| 79 | movs r3, r3, lsl #2 |
| 80 | strcs r1, [r0], #4 |
| 81 | |
| 82 | /* Now quickly loop through the cache-aligned data. */ |
| 83 | .Laligned32: |
| 84 | mov r3, r1 |
| 85 | 1: subs r2, r2, #32 |
| 86 | stmhsia r0!, {r1,r3,ip,lr} |
| 87 | stmhsia r0!, {r1,r3,ip,lr} |
| 88 | bhs 1b |
| 89 | add r2, r2, #32 |
| 90 | |
| 91 | /* (Optionally) store any remaining trailing bytes. |
| 92 | * (0-30 bytes, length in r2) */ |
| 93 | movs r2, r2, lsl #28 |
| 94 | stmcsia r0!, {r1,r3,ip,lr} |
| 95 | stmmiia r0!, {r1,lr} |
| 96 | movs r2, r2, lsl #2 |
| 97 | strcs r1, [r0], #4 |
| 98 | strmih lr, [r0], #2 |
| 99 | |
| 100 | pop {pc} |
| 101 | .fnend |