1 | ; $Id: ASMMemFirstMismatchingU8.asm 93115 2022-01-01 11:31:46Z vboxsync $
|
---|
2 | ;; @file
|
---|
3 | ; IPRT - ASMMemFirstMismatchingU8().
|
---|
4 | ;
|
---|
5 |
|
---|
6 | ;
|
---|
7 | ; Copyright (C) 2006-2022 Oracle Corporation
|
---|
8 | ;
|
---|
9 | ; This file is part of VirtualBox Open Source Edition (OSE), as
|
---|
10 | ; available from http://www.virtualbox.org. This file is free software;
|
---|
11 | ; you can redistribute it and/or modify it under the terms of the GNU
|
---|
12 | ; General Public License (GPL) as published by the Free Software
|
---|
13 | ; Foundation, in version 2 as it comes in the "COPYING" file of the
|
---|
14 | ; VirtualBox OSE distribution. VirtualBox OSE is distributed in the
|
---|
15 | ; hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
|
---|
16 | ;
|
---|
17 | ; The contents of this file may alternatively be used under the terms
|
---|
18 | ; of the Common Development and Distribution License Version 1.0
|
---|
19 | ; (CDDL) only, as it comes in the "COPYING.CDDL" file of the
|
---|
20 | ; VirtualBox OSE distribution, in which case the provisions of the
|
---|
21 | ; CDDL are applicable instead of those of the GPL.
|
---|
22 | ;
|
---|
23 | ; You may elect to license modified versions of this file under the
|
---|
24 | ; terms and conditions of either the GPL or the CDDL or both.
|
---|
25 | ;
|
---|
26 |
|
---|
27 |
|
---|
28 | ;*******************************************************************************
|
---|
29 | ;* Header Files *
|
---|
30 | ;*******************************************************************************
|
---|
31 | %define RT_ASM_WITH_SEH64
|
---|
32 | %include "iprt/asmdefs.mac"
|
---|
33 |
|
---|
34 |
|
---|
35 | BEGINCODE
|
---|
36 |
|
---|
37 | ;;
|
---|
38 | ; Variant of ASMMemFirstMismatchingU8 with a fixed @a u8 value.
|
---|
39 | ; We repeat the prolog and join the generic function.
|
---|
40 | ;
|
---|
41 | RT_BEGINPROC ASMMemFirstNonZero
|
---|
42 | ;
|
---|
43 | ; Prologue.
|
---|
44 | ;
|
---|
45 | %if ARCH_BITS != 64
|
---|
46 | push xBP
|
---|
47 | mov xBP, xSP
|
---|
48 | push xDI
|
---|
49 | %if ARCH_BITS == 16
|
---|
50 | push es
|
---|
51 | %endif
|
---|
52 | %elifdef ASM_CALL64_MSC
|
---|
53 | mov r9, rdi ; save rdi in r9
|
---|
54 | %endif
|
---|
55 | SEH64_END_PROLOGUE
|
---|
56 |
|
---|
57 | ;
|
---|
58 | ; Normalize input; rdi=pv, rcx=cb, rax=0
|
---|
59 | ;
|
---|
60 | %if ARCH_BITS == 64
|
---|
61 | %ifdef ASM_CALL64_MSC
|
---|
62 | mov rdi, rcx
|
---|
63 | mov rcx, rdx
|
---|
64 | jrcxz RT_CONCAT(NAME(ASMMemFirstMismatchingU8),.return_all_same)
|
---|
65 | xor eax, eax
|
---|
66 | %else
|
---|
67 | mov rcx, rsi
|
---|
68 | jrcxz RT_CONCAT(NAME(ASMMemFirstMismatchingU8),.return_all_same)
|
---|
69 | xor eax, eax
|
---|
70 | %endif
|
---|
71 |
|
---|
72 | %elif ARCH_BITS == 32
|
---|
73 | mov ecx, [ebp + 0ch]
|
---|
74 | jecxz RT_CONCAT(NAME(ASMMemFirstMismatchingU8),.return_all_same)
|
---|
75 | mov edi, [ebp + 08h]
|
---|
76 | xor eax, eax
|
---|
77 |
|
---|
78 | %elif ARCH_BITS == 16
|
---|
79 | mov cx, [bp + 08h] ; cb
|
---|
80 | jcxz RT_CONCAT(NAME(ASMMemFirstMismatchingU8),.return16_all_same)
|
---|
81 | les di, [bp + 04h] ; pv (far)
|
---|
82 | xor ax, ax
|
---|
83 |
|
---|
84 | %else
|
---|
85 | %error "Invalid ARCH_BITS value"
|
---|
86 | %endif
|
---|
87 |
|
---|
88 | ;
|
---|
89 | ; Join ASMMemFirstMismatchingU8
|
---|
90 | ;
|
---|
91 | jmp RT_CONCAT(NAME(ASMMemFirstMismatchingU8),.is_all_zero_joining)
|
---|
92 | ENDPROC ASMMemFirstNonZero
|
---|
93 |
|
---|
94 |
|
---|
95 | ;;
|
---|
96 | ; Inverted memchr.
|
---|
97 | ;
|
---|
98 | ; @returns Pointer to the byte which doesn't equal u8.
|
---|
99 | ; @returns NULL if all equal to u8.
|
---|
100 | ;
|
---|
101 | ; @param msc:rcx gcc:rdi pv Pointer to the memory block.
|
---|
102 | ; @param msc:rdx gcc:rsi cb Number of bytes in the block. This MUST be aligned on 32-bit!
|
---|
103 | ; @param msc:r8b gcc:dl u8 The value it's supposed to be filled with.
|
---|
104 | ;
|
---|
105 | ; @cproto DECLINLINE(void *) ASMMemFirstMismatchingU8(void const *pv, size_t cb, uint8_t u8)
|
---|
106 | ;
|
---|
107 | RT_BEGINPROC ASMMemFirstMismatchingU8
|
---|
108 | ;
|
---|
109 | ; Prologue.
|
---|
110 | ;
|
---|
111 | %if ARCH_BITS != 64
|
---|
112 | push xBP
|
---|
113 | mov xBP, xSP
|
---|
114 | push xDI
|
---|
115 | %if ARCH_BITS == 16
|
---|
116 | push es
|
---|
117 | %endif
|
---|
118 | %elifdef ASM_CALL64_MSC
|
---|
119 | mov r9, rdi ; save rdi in r9
|
---|
120 | %endif
|
---|
121 | SEH64_END_PROLOGUE
|
---|
122 |
|
---|
123 | %if ARCH_BITS != 16
|
---|
124 | ;
|
---|
125 | ; The 32-bit and 64-bit variant of the code.
|
---|
126 | ;
|
---|
127 |
|
---|
128 | ; Normalize input; rdi=pv, rcx=cb, rax=eight-times-u8
|
---|
129 | %if ARCH_BITS == 64
|
---|
130 | %ifdef ASM_CALL64_MSC
|
---|
131 | mov rdi, rcx
|
---|
132 | mov rcx, rdx
|
---|
133 | jrcxz .return_all_same
|
---|
134 | movzx r8d, r8b
|
---|
135 | mov rax, qword 0101010101010101h
|
---|
136 | imul rax, r8
|
---|
137 | %else
|
---|
138 | mov rcx, rsi
|
---|
139 | jrcxz .return_all_same
|
---|
140 | movzx edx, dl
|
---|
141 | mov rax, qword 0101010101010101h
|
---|
142 | imul rax, rdx
|
---|
143 | %endif
|
---|
144 |
|
---|
145 | %elif ARCH_BITS == 32
|
---|
146 | mov ecx, [ebp + 0ch]
|
---|
147 | jecxz .return_all_same
|
---|
148 | mov edi, [ebp + 08h]
|
---|
149 | movzx eax, byte [ebp + 10h]
|
---|
150 | mov ah, al
|
---|
151 | movzx edx, ax
|
---|
152 | shl eax, 16
|
---|
153 | or eax, edx
|
---|
154 | %else
|
---|
155 | %error "Invalid ARCH_BITS value"
|
---|
156 | %endif
|
---|
157 |
|
---|
158 | .is_all_zero_joining:
|
---|
159 | cld
|
---|
160 |
|
---|
161 | ; Unaligned pointer? Align it (elsewhere).
|
---|
162 | test edi, xCB - 1
|
---|
163 | jnz .unaligned_pv
|
---|
164 | .aligned_pv:
|
---|
165 |
|
---|
166 | ; Do the dword/qword scan.
|
---|
167 | mov edx, xCB - 1
|
---|
168 | and edx, ecx ; Remaining bytes for tail scan
|
---|
169 | %if ARCH_BITS == 64
|
---|
170 | shr xCX, 3
|
---|
171 | repe scasq
|
---|
172 | %else
|
---|
173 | shr xCX, 2
|
---|
174 | repe scasd
|
---|
175 | %endif
|
---|
176 | jne .multibyte_mismatch
|
---|
177 |
|
---|
178 | ; Prep for tail scan.
|
---|
179 | mov ecx, edx
|
---|
180 |
|
---|
181 | ;
|
---|
182 | ; Byte by byte scan.
|
---|
183 | ;
|
---|
184 | .byte_by_byte:
|
---|
185 | repe scasb
|
---|
186 | jne .return_xDI
|
---|
187 |
|
---|
188 | .return_all_same:
|
---|
189 | xor eax, eax
|
---|
190 | %ifdef ASM_CALL64_MSC
|
---|
191 | mov rdi, r9 ; restore rdi
|
---|
192 | %elif ARCH_BITS == 32
|
---|
193 | pop edi
|
---|
194 | leave
|
---|
195 | %endif
|
---|
196 | ret
|
---|
197 |
|
---|
198 | ; Return after byte scan mismatch.
|
---|
199 | .return_xDI:
|
---|
200 | lea xAX, [xDI - 1]
|
---|
201 | %ifdef ASM_CALL64_MSC
|
---|
202 | mov rdi, r9 ; restore rdi
|
---|
203 | %elif ARCH_BITS == 32
|
---|
204 | pop edi
|
---|
205 | leave
|
---|
206 | %endif
|
---|
207 | ret
|
---|
208 |
|
---|
209 | ;
|
---|
210 | ; Multibyte mismatch. We rewind and do a byte scan of the remainder.
|
---|
211 | ; (can't just search the qword as the buffer must be considered volatile).
|
---|
212 | ;
|
---|
213 | .multibyte_mismatch:
|
---|
214 | lea xDI, [xDI - xCB]
|
---|
215 | lea xCX, [xCX * xCB + xCB]
|
---|
216 | or ecx, edx
|
---|
217 | jmp .byte_by_byte
|
---|
218 |
|
---|
219 | ;
|
---|
220 | ; Unaligned pointer. If it's worth it, align the pointer, but if the
|
---|
221 | ; memory block is too small do the byte scan variant.
|
---|
222 | ;
|
---|
223 | .unaligned_pv:
|
---|
224 | cmp xCX, 4*xCB ; 4 steps seems reasonable.
|
---|
225 | jbe .byte_by_byte
|
---|
226 |
|
---|
227 | ; Unrolled buffer realignment.
|
---|
228 | %if ARCH_BITS == 64
|
---|
229 | dec xCX
|
---|
230 | scasb
|
---|
231 | jne .return_xDI
|
---|
232 | test edi, xCB - 1
|
---|
233 | jz .aligned_pv
|
---|
234 |
|
---|
235 | dec xCX
|
---|
236 | scasb
|
---|
237 | jne .return_xDI
|
---|
238 | test edi, xCB - 1
|
---|
239 | jz .aligned_pv
|
---|
240 |
|
---|
241 | dec xCX
|
---|
242 | scasb
|
---|
243 | jne .return_xDI
|
---|
244 | test edi, xCB - 1
|
---|
245 | jz .aligned_pv
|
---|
246 |
|
---|
247 | dec xCX
|
---|
248 | scasb
|
---|
249 | jne .return_xDI
|
---|
250 | test edi, xCB - 1
|
---|
251 | jz .aligned_pv
|
---|
252 | %endif
|
---|
253 |
|
---|
254 | dec xCX
|
---|
255 | scasb
|
---|
256 | jne .return_xDI
|
---|
257 | test edi, xCB - 1
|
---|
258 | jz .aligned_pv
|
---|
259 |
|
---|
260 | dec xCX
|
---|
261 | scasb
|
---|
262 | jne .return_xDI
|
---|
263 | test edi, xCB - 1
|
---|
264 | jz .aligned_pv
|
---|
265 |
|
---|
266 | dec xCX
|
---|
267 | scasb
|
---|
268 | jne .return_xDI
|
---|
269 | jmp .aligned_pv
|
---|
270 |
|
---|
271 |
|
---|
272 | %else ; ARCH_BITS == 16
|
---|
273 |
|
---|
274 | ;
|
---|
275 | ; The 16-bit variant of the code is a little simpler since we're
|
---|
276 | ; working with two byte words in the 'fast' scan. We also keep
|
---|
277 | ; this separate from the 32-bit/64-bit code because that allows
|
---|
278 | ; avoid a few rex prefixes here and there by using extended
|
---|
279 | ; registers (e??) where we don't care about the whole register.
|
---|
280 | ;
|
---|
281 | CPU 8086
|
---|
282 |
|
---|
283 | ; Load input parameters.
|
---|
284 | mov cx, [bp + 08h] ; cb
|
---|
285 | jcxz .return16_all_same
|
---|
286 | les di, [bp + 04h] ; pv (far)
|
---|
287 | mov al, [bp + 0ah] ; u8
|
---|
288 | mov ah, al
|
---|
289 |
|
---|
290 | .is_all_zero_joining:
|
---|
291 | cld
|
---|
292 |
|
---|
293 | ; Align the pointer.
|
---|
294 | test di, 1
|
---|
295 | jz .word_scan
|
---|
296 |
|
---|
297 | dec cx
|
---|
298 | scasb
|
---|
299 | jne .return16_di
|
---|
300 | jcxz .return16_all_same
|
---|
301 |
|
---|
302 | ; Scan word-by-word.
|
---|
303 | .word_scan:
|
---|
304 | mov dx, cx
|
---|
305 | shr cx, 1
|
---|
306 | repe scasw
|
---|
307 | jne .word_mismatch
|
---|
308 |
|
---|
309 | ; do we have a tail byte?
|
---|
310 | test dl, 1
|
---|
311 | jz .return16_all_same
|
---|
312 | scasb
|
---|
313 | jne .return16_di
|
---|
314 |
|
---|
315 | .return16_all_same:
|
---|
316 | xor ax, ax
|
---|
317 | xor dx, dx
|
---|
318 | .return16:
|
---|
319 | pop es
|
---|
320 | pop di
|
---|
321 | pop bp
|
---|
322 | ret
|
---|
323 |
|
---|
324 | .word_mismatch:
|
---|
325 | ; back up a word.
|
---|
326 | inc cx
|
---|
327 | sub di, 2
|
---|
328 |
|
---|
329 | ; Do byte-by-byte scanning of the rest of the buffer.
|
---|
330 | shl cx, 1
|
---|
331 | mov dl, 1
|
---|
332 | and dl, [bp + 08h] ; cb
|
---|
333 | or cl, dl
|
---|
334 | repe scasb
|
---|
335 | je .return16_all_same
|
---|
336 |
|
---|
337 | .return16_di:
|
---|
338 | mov ax, di
|
---|
339 | dec ax
|
---|
340 | mov dx, es
|
---|
341 | jmp .return16
|
---|
342 |
|
---|
343 | %endif ; ARCH_BITS == 16
|
---|
344 | ENDPROC ASMMemFirstMismatchingU8
|
---|
345 |
|
---|