| agl@chromium.org | aab4090 | 2010-06-04 14:47:38 +0000 | [diff] [blame] | 1 | /*************************************************************************** |
| epoger@google.com | fd03db0 | 2011-07-28 14:24:55 +0000 | [diff] [blame] | 2 | * Copyright (c) 2009,2010, Code Aurora Forum. All rights reserved. |
| 3 | * |
| 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 | aab4090 | 2010-06-04 14:47:38 +0000 | [diff] [blame] | 6 | ***************************************************************************/ |
| 7 | |
| 8 | .code 32 |
| 9 | .fpu neon |
| 10 | .align 4 |
| 11 | .globl memset32_neon |
| agl@chromium.org | aab4090 | 2010-06-04 14:47:38 +0000 | [diff] [blame] | 12 | |
| 13 | /* r0 = buffer, r1 = value, r2 = times to write */ |
| 14 | memset32_neon: |
| 15 | cmp r2, #1 |
| 16 | streq r1, [r0], #4 |
| 17 | bxeq lr |
| 18 | |
| 19 | cmp r2, #4 |
| 20 | bgt memset32_neon_start |
| 21 | cmp r2, #0 |
| 22 | bxeq lr |
| 23 | memset32_neon_small: |
| 24 | str r1, [r0], #4 |
| 25 | subs r2, r2, #1 |
| 26 | bne memset32_neon_small |
| 27 | bx lr |
| 28 | memset32_neon_start: |
| 29 | cmp r2, #16 |
| 30 | blt memset32_dropthru |
| 31 | vdup.32 q0, r1 |
| 32 | vmov q1, q0 |
| 33 | cmp r2, #32 |
| 34 | blt memset32_16 |
| 35 | cmp r2, #64 |
| 36 | blt memset32_32 |
| 37 | cmp r2, #128 |
| 38 | blt memset32_64 |
| 39 | memset32_128: |
| 40 | movs r12, r2, lsr #7 |
| 41 | memset32_loop128: |
| 42 | subs r12, r12, #1 |
| 43 | vst1.64 {q0, q1}, [r0]! |
| 44 | vst1.64 {q0, q1}, [r0]! |
| 45 | vst1.64 {q0, q1}, [r0]! |
| 46 | vst1.64 {q0, q1}, [r0]! |
| 47 | vst1.64 {q0, q1}, [r0]! |
| 48 | vst1.64 {q0, q1}, [r0]! |
| 49 | vst1.64 {q0, q1}, [r0]! |
| 50 | vst1.64 {q0, q1}, [r0]! |
| 51 | vst1.64 {q0, q1}, [r0]! |
| 52 | vst1.64 {q0, q1}, [r0]! |
| 53 | vst1.64 {q0, q1}, [r0]! |
| 54 | vst1.64 {q0, q1}, [r0]! |
| 55 | vst1.64 {q0, q1}, [r0]! |
| 56 | vst1.64 {q0, q1}, [r0]! |
| 57 | vst1.64 {q0, q1}, [r0]! |
| 58 | vst1.64 {q0, q1}, [r0]! |
| 59 | bne memset32_loop128 |
| 60 | ands r2, r2, #0x7f |
| 61 | bxeq lr |
| 62 | memset32_64: |
| 63 | movs r12, r2, lsr #6 |
| 64 | beq memset32_32 |
| 65 | vst1.64 {q0, q1}, [r0]! |
| 66 | vst1.64 {q0, q1}, [r0]! |
| 67 | vst1.64 {q0, q1}, [r0]! |
| 68 | vst1.64 {q0, q1}, [r0]! |
| 69 | vst1.64 {q0, q1}, [r0]! |
| 70 | vst1.64 {q0, q1}, [r0]! |
| 71 | vst1.64 {q0, q1}, [r0]! |
| 72 | vst1.64 {q0, q1}, [r0]! |
| 73 | ands r2, r2, #0x3f |
| 74 | bxeq lr |
| 75 | memset32_32: |
| 76 | movs r12, r2, lsr #5 |
| 77 | beq memset32_16 |
| 78 | vst1.64 {q0, q1}, [r0]! |
| 79 | vst1.64 {q0, q1}, [r0]! |
| 80 | vst1.64 {q0, q1}, [r0]! |
| 81 | vst1.64 {q0, q1}, [r0]! |
| 82 | ands r2, r2, #0x1f |
| 83 | bxeq lr |
| 84 | memset32_16: |
| 85 | movs r12, r2, lsr #4 |
| 86 | beq memset32_dropthru |
| 87 | and r2, r2, #0xf |
| 88 | vst1.64 {q0, q1}, [r0]! |
| 89 | vst1.64 {q0, q1}, [r0]! |
| 90 | memset32_dropthru: |
| 91 | rsb r2, r2, #15 |
| 92 | add pc, pc, r2, lsl #2 |
| 93 | nop |
| 94 | str r1, [r0, #56] |
| 95 | str r1, [r0, #52] |
| 96 | str r1, [r0, #48] |
| 97 | str r1, [r0, #44] |
| 98 | str r1, [r0, #40] |
| 99 | str r1, [r0, #36] |
| 100 | str r1, [r0, #32] |
| 101 | str r1, [r0, #28] |
| 102 | str r1, [r0, #24] |
| 103 | str r1, [r0, #20] |
| 104 | str r1, [r0, #16] |
| 105 | str r1, [r0, #12] |
| 106 | str r1, [r0, #8] |
| 107 | str r1, [r0, #4] |
| 108 | str r1, [r0, #0] |
| 109 | bx lr |
| 110 | |
| agl@chromium.org | aab4090 | 2010-06-04 14:47:38 +0000 | [diff] [blame] | 111 | .end |