Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 1 | // Copyright 2014 The Go Authors. All rights reserved. |
| 2 | // Use of this source code is governed by a BSD-style |
| 3 | // license that can be found in the LICENSE file. |
| 4 | |
Anthony Martin | 8303a13 | 2014-03-12 18:12:25 -0700 | [diff] [blame] | 5 | // +build !plan9 |
| 6 | |
Russ Cox | cb040d5 | 2014-09-04 23:05:18 -0400 | [diff] [blame] | 7 | #include "textflag.h" |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 8 | |
Russ Cox | ee6c6d9 | 2014-09-09 17:12:05 -0400 | [diff] [blame] | 9 | // NOTE: Windows externalthreadhandler expects memclr to preserve DX. |
| 10 | |
Austin Clements | 87e48c5 | 2016-10-17 18:41:56 -0400 | [diff] [blame] | 11 | // void runtime·memclrNoHeapPointers(void*, uintptr) |
| 12 | TEXT runtime·memclrNoHeapPointers(SB), NOSPLIT, $0-8 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 13 | MOVL ptr+0(FP), DI |
| 14 | MOVL n+4(FP), BX |
| 15 | XORL AX, AX |
| 16 | |
| 17 | // MOVOU seems always faster than REP STOSL. |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 18 | tail: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 19 | TESTL BX, BX |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 20 | JEQ _0 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 21 | CMPL BX, $2 |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 22 | JBE _1or2 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 23 | CMPL BX, $4 |
Keith Randall | 4b7d5f0 | 2015-11-05 12:39:56 -0800 | [diff] [blame] | 24 | JB _3 |
| 25 | JE _4 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 26 | CMPL BX, $8 |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 27 | JBE _5through8 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 28 | CMPL BX, $16 |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 29 | JBE _9through16 |
Martin Möhrmann | 5a6c580 | 2017-04-27 08:30:27 +0200 | [diff] [blame] | 30 | CMPB runtime·support_sse2(SB), $1 |
| 31 | JNE nosse2 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 32 | PXOR X0, X0 |
| 33 | CMPL BX, $32 |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 34 | JBE _17through32 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 35 | CMPL BX, $64 |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 36 | JBE _33through64 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 37 | CMPL BX, $128 |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 38 | JBE _65through128 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 39 | CMPL BX, $256 |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 40 | JBE _129through256 |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 41 | // TODO: use branch table and BSR to make this just a single dispatch |
| 42 | |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 43 | loop: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 44 | MOVOU X0, 0(DI) |
| 45 | MOVOU X0, 16(DI) |
| 46 | MOVOU X0, 32(DI) |
| 47 | MOVOU X0, 48(DI) |
| 48 | MOVOU X0, 64(DI) |
| 49 | MOVOU X0, 80(DI) |
| 50 | MOVOU X0, 96(DI) |
| 51 | MOVOU X0, 112(DI) |
| 52 | MOVOU X0, 128(DI) |
| 53 | MOVOU X0, 144(DI) |
| 54 | MOVOU X0, 160(DI) |
| 55 | MOVOU X0, 176(DI) |
| 56 | MOVOU X0, 192(DI) |
| 57 | MOVOU X0, 208(DI) |
| 58 | MOVOU X0, 224(DI) |
| 59 | MOVOU X0, 240(DI) |
| 60 | SUBL $256, BX |
| 61 | ADDL $256, DI |
| 62 | CMPL BX, $256 |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 63 | JAE loop |
| 64 | JMP tail |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 65 | |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 66 | _1or2: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 67 | MOVB AX, (DI) |
| 68 | MOVB AX, -1(DI)(BX*1) |
Josh Bleecher Snyder | bc9e160 | 2014-08-01 06:21:08 -0700 | [diff] [blame] | 69 | RET |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 70 | _0: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 71 | RET |
Keith Randall | 4b7d5f0 | 2015-11-05 12:39:56 -0800 | [diff] [blame] | 72 | _3: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 73 | MOVW AX, (DI) |
Keith Randall | 4b7d5f0 | 2015-11-05 12:39:56 -0800 | [diff] [blame] | 74 | MOVB AX, 2(DI) |
| 75 | RET |
| 76 | _4: |
| 77 | // We need a separate case for 4 to make sure we clear pointers atomically. |
| 78 | MOVL AX, (DI) |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 79 | RET |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 80 | _5through8: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 81 | MOVL AX, (DI) |
| 82 | MOVL AX, -4(DI)(BX*1) |
| 83 | RET |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 84 | _9through16: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 85 | MOVL AX, (DI) |
| 86 | MOVL AX, 4(DI) |
| 87 | MOVL AX, -8(DI)(BX*1) |
| 88 | MOVL AX, -4(DI)(BX*1) |
| 89 | RET |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 90 | _17through32: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 91 | MOVOU X0, (DI) |
| 92 | MOVOU X0, -16(DI)(BX*1) |
| 93 | RET |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 94 | _33through64: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 95 | MOVOU X0, (DI) |
| 96 | MOVOU X0, 16(DI) |
| 97 | MOVOU X0, -32(DI)(BX*1) |
| 98 | MOVOU X0, -16(DI)(BX*1) |
| 99 | RET |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 100 | _65through128: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 101 | MOVOU X0, (DI) |
| 102 | MOVOU X0, 16(DI) |
| 103 | MOVOU X0, 32(DI) |
| 104 | MOVOU X0, 48(DI) |
| 105 | MOVOU X0, -64(DI)(BX*1) |
| 106 | MOVOU X0, -48(DI)(BX*1) |
| 107 | MOVOU X0, -32(DI)(BX*1) |
| 108 | MOVOU X0, -16(DI)(BX*1) |
| 109 | RET |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 110 | _129through256: |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 111 | MOVOU X0, (DI) |
| 112 | MOVOU X0, 16(DI) |
| 113 | MOVOU X0, 32(DI) |
| 114 | MOVOU X0, 48(DI) |
| 115 | MOVOU X0, 64(DI) |
| 116 | MOVOU X0, 80(DI) |
| 117 | MOVOU X0, 96(DI) |
| 118 | MOVOU X0, 112(DI) |
| 119 | MOVOU X0, -128(DI)(BX*1) |
| 120 | MOVOU X0, -112(DI)(BX*1) |
| 121 | MOVOU X0, -96(DI)(BX*1) |
| 122 | MOVOU X0, -80(DI)(BX*1) |
| 123 | MOVOU X0, -64(DI)(BX*1) |
| 124 | MOVOU X0, -48(DI)(BX*1) |
| 125 | MOVOU X0, -32(DI)(BX*1) |
| 126 | MOVOU X0, -16(DI)(BX*1) |
| 127 | RET |
| 128 | nosse2: |
| 129 | MOVL BX, CX |
| 130 | SHRL $2, CX |
| 131 | REP |
| 132 | STOSL |
| 133 | ANDL $3, BX |
Russ Cox | b55791e | 2014-10-28 21:50:16 -0400 | [diff] [blame] | 134 | JNE tail |
Keith Randall | da7cf0b | 2014-02-06 17:43:22 -0800 | [diff] [blame] | 135 | RET |