3 * Copyright (C) 2002 Frederic 'dilb' Boulay
5 * Author: Frederic Boulay <dilb@handhelds.org>
7 * The function defined in this file is derived from the simple_idct function
8 * from the libavcodec library part of the FFmpeg project.
10 * This file is part of FFmpeg.
12 * FFmpeg is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public
14 * License as published by the Free Software Foundation; either
15 * version 2.1 of the License, or (at your option) any later version.
17 * FFmpeg is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with FFmpeg; if not, write to the Free Software
24 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
29 /* useful constants for the algorithm, they are save in __constant_ptr__ at */
30 /* the end of the source code.*/
38 #define MASK_MSHW 0xFFFF0000
40 /* offsets of the constants in the vector */
48 #define offMASK_MSHW 28
51 #define ROW_SHIFT2MSHW (16-11)
53 #define ROW_SHIFTED_1 1024 /* 1<< (ROW_SHIFT-1) */
54 #define COL_SHIFTED_1 524288 /* 1<< (COL_SHIFT-1) */
59 function ff_simple_idct_arm, export=1
60 @@ void simple_idct_arm(int16_t *block)
61 @@ save stack for reg needed (take all of them),
62 @@ R0-R3 are scratch regs, so no need to save them, but R0 contains the pointer to block
63 @@ so it must not be overwritten, if it is not saved!!
64 @@ R12 is another scratch register, so it should not be saved too
66 stmfd sp!, {r4-r11, r14} @ R14 is also called LR
67 @@ at this point, R0=block, other registers are free.
68 add r14, r0, #112 @ R14=&block[8*7], better start from the last row, and decrease the value until row=0, i.e. R12=block.
69 adr r12, __constant_ptr__ @ R12=__constant_ptr__, the vector containing the constants, probably not necessary to reserve a register for it
70 @@ add 2 temporary variables in the stack: R0 and R14
71 sub sp, sp, #8 @ allow 2 local variables
72 str r0, [sp, #0] @ save block in sp[0]
78 @@ at this point, R0=block, R14=&block[56], R12=__const_ptr_, R1-R11 free
82 @@ read the row and check if it is null, almost null, or not, according to strongarm specs, it is not necessary to optimize ldr accesses (i.e. split 32bits in 2 16bits words), at least it gives more usable registers :)
83 ldr r1, [r14, #0] @ R1=(int32)(R12)[0]=ROWr32[0] (relative row cast to a 32b pointer)
84 ldr r2, [r14, #4] @ R2=(int32)(R12)[1]=ROWr32[1]
85 ldr r3, [r14, #8] @ R3=ROWr32[2]
86 ldr r4, [r14, #12] @ R4=ROWr32[3]
87 @@ check if the words are null, if all of them are null, then proceed with next row (branch __end_row_loop),
88 @@ if ROWr16[0] is the only one not null, then proceed with this special case (branch __almost_empty_row)
89 @@ else follow the complete algorithm.
90 @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
91 @@ R3=ROWr32[2], R4=ROWr32[3], R5-R11 free
92 orr r5, r4, r3 @ R5=R4 | R3
93 orr r5, r5, r2 @ R5=R4 | R3 | R2
94 orrs r6, r5, r1 @ Test R5 | R1 (the aim is to check if everything is null)
96 mov r7, r1, asr #16 @ R7=R1>>16=ROWr16[1] (evaluate it now, as it could be useful later)
97 ldrsh r6, [r14, #0] @ R6=ROWr16[0]
98 orrs r5, r5, r7 @ R5=R4 | R3 | R2 | R7
99 beq __almost_empty_row
102 @@ at this point, R0=block (temp), R1(free), R2=ROWr32[1], R3=ROWr32[2], R4=ROWr32[3],
103 @@ R5=(temp), R6=ROWr16[0], R7=ROWr16[1], R8-R11 free,
104 @@ R12=__const_ptr_, R14=&block[n]
105 @@ to save some registers/calls, proceed with b0-b3 first, followed by a0-a3
107 @@ MUL16(b0, W1, row[1]);
108 @@ MUL16(b1, W3, row[1]);
109 @@ MUL16(b2, W5, row[1]);
110 @@ MUL16(b3, W7, row[1]);
111 @@ MAC16(b0, W3, row[3]);
112 @@ MAC16(b1, -W7, row[3]);
113 @@ MAC16(b2, -W1, row[3]);
114 @@ MAC16(b3, -W5, row[3]);
115 ldr r8, [r12, #offW1] @ R8=W1
116 mov r2, r2, asr #16 @ R2=ROWr16[3]
117 mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
118 ldr r9, [r12, #offW3] @ R9=W3
119 ldr r10, [r12, #offW5] @ R10=W5
120 mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
121 ldr r11, [r12, #offW7] @ R11=W7
122 mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
123 mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
124 teq r2, #0 @ if null avoid muls
126 mlane r0, r9, r2, r0 @ R0+=W3*ROWr16[3]=b0 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
127 rsbne r2, r2, #0 @ R2=-ROWr16[3]
128 mlane r1, r11, r2, r1 @ R1-=W7*ROWr16[3]=b1 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
129 mlane r5, r8, r2, r5 @ R5-=W1*ROWr16[3]=b2 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
131 mlane r7, r10, r2, r7 @ R7-=W5*ROWr16[3]=b3 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
133 @@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3],
134 @@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
135 @@ R12=__const_ptr_, R14=&block[n]
136 @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
138 orrs r2, r3, r4 @ R2=ROWr32[2] | ROWr32[3]
139 beq __end_b_evaluation
141 @@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3],
142 @@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
143 @@ R12=__const_ptr_, R14=&block[n]
144 @@ MAC16(b0, W5, row[5]);
145 @@ MAC16(b2, W7, row[5]);
146 @@ MAC16(b3, W3, row[5]);
147 @@ MAC16(b1, -W1, row[5]);
148 @@ MAC16(b0, W7, row[7]);
149 @@ MAC16(b2, W3, row[7]);
150 @@ MAC16(b3, -W1, row[7]);
151 @@ MAC16(b1, -W5, row[7]);
152 mov r3, r3, asr #16 @ R3=ROWr16[5]
153 teq r3, #0 @ if null avoid muls
155 mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5]=b0
156 mov r4, r4, asr #16 @ R4=ROWr16[7]
158 mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5]=b2
159 mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5]=b3
160 rsbne r3, r3, #0 @ R3=-ROWr16[5]
161 mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5]=b1
163 teq r4, #0 @ if null avoid muls
165 mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7]=b0
166 mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7]=b2
167 rsbne r4, r4, #0 @ R4=-ROWr16[7]
168 mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7]=b3
170 mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7]=b1
173 @@ at this point, R0=b0, R1=b1, R2=ROWr32[2] | ROWr32[3] (tmp), R3 (free), R4 (free),
174 @@ R5=b2, R6=ROWr16[0], R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
175 @@ R12=__const_ptr_, R14=&block[n]
178 @@ a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1));
179 @@ a1 = a0 + W6 * row[2];
180 @@ a2 = a0 - W6 * row[2];
181 @@ a3 = a0 - W2 * row[2];
182 @@ a0 = a0 + W2 * row[2];
183 ldr r9, [r12, #offW4] @ R9=W4
184 mul r6, r9, r6 @ R6=W4*ROWr16[0]
185 ldr r10, [r12, #offW6] @ R10=W6
186 ldrsh r4, [r14, #4] @ R4=ROWr16[2] (a3 not defined yet)
187 add r6, r6, #ROW_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(ROW_SHIFT-1) (a0)
189 mul r11, r10, r4 @ R11=W6*ROWr16[2]
190 ldr r8, [r12, #offW2] @ R8=W2
191 sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2)
192 @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
195 beq __end_bef_a_evaluation
197 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
198 mul r11, r8, r4 @ R11=W2*ROWr16[2]
199 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
200 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
203 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
204 @@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
205 @@ R12=__const_ptr_, R14=&block[n]
212 ldrsh r11, [r14, #8] @ R11=ROWr16[4]
213 teq r11, #0 @ if null avoid muls
215 mulne r11, r9, r11 @ R11=W4*ROWr16[4]
217 ldrsh r9, [r14, #12] @ R9=ROWr16[6]
219 addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0)
220 subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1)
221 subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2)
222 addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3)
223 @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
224 teq r9, #0 @ if null avoid muls
226 mulne r11, r10, r9 @ R11=W6*ROWr16[6]
227 addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0)
228 mulne r10, r8, r9 @ R10=W2*ROWr16[6]
233 subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3)
235 subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1)
236 addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2)
239 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
240 @@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
241 @@ R12=__const_ptr_, R14=&block[n]
242 @@ row[0] = (a0 + b0) >> ROW_SHIFT;
243 @@ row[1] = (a1 + b1) >> ROW_SHIFT;
244 @@ row[2] = (a2 + b2) >> ROW_SHIFT;
245 @@ row[3] = (a3 + b3) >> ROW_SHIFT;
246 @@ row[4] = (a3 - b3) >> ROW_SHIFT;
247 @@ row[5] = (a2 - b2) >> ROW_SHIFT;
248 @@ row[6] = (a1 - b1) >> ROW_SHIFT;
249 @@ row[7] = (a0 - b0) >> ROW_SHIFT;
250 add r8, r6, r0 @ R8=a0+b0
251 add r9, r2, r1 @ R9=a1+b1
252 @@ put 2 16 bits half-words in a 32bits word
253 @@ ROWr32[0]=ROWr16[0] | (ROWr16[1]<<16) (only Little Endian compliant then!!!)
254 ldr r10, [r12, #offMASK_MSHW] @ R10=0xFFFF0000
255 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a1+b1)<<5)
256 mvn r11, r10 @ R11= NOT R10= 0x0000FFFF
257 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a0+b0)>>11)
261 add r8, r3, r5 @ R8=a2+b2
262 add r9, r4, r7 @ R9=a3+b3
263 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a3+b3)<<5)
264 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a2+b2)>>11)
268 sub r8, r4, r7 @ R8=a3-b3
269 sub r9, r3, r5 @ R9=a2-b2
270 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a2-b2)<<5)
271 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a3-b3)>>11)
275 sub r8, r2, r1 @ R8=a1-b1
276 sub r9, r6, r0 @ R9=a0-b0
277 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a0-b0)<<5)
278 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a1-b1)>>11)
285 @@ the row was empty, except ROWr16[0], now, management of this special case
286 @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
287 @@ R3=ROWr32[2], R4=ROWr32[3], R5=(temp), R6=ROWr16[0], R7=ROWr16[1],
288 @@ R8=0xFFFF (temp), R9-R11 free
289 mov r8, #0x10000 @ R8=0xFFFF (2 steps needed!) it saves a ldr call (because of delay run).
290 sub r8, r8, #1 @ R8 is now ready.
291 and r5, r8, r6, lsl #3 @ R5=R8 & (R6<<3)= (ROWr16[0]<<3) & 0xFFFF
292 orr r5, r5, r5, lsl #16 @ R5=R5 | (R5<<16)
293 str r5, [r14, #0] @ R14[0]=ROWr32[0]=R5
294 str r5, [r14, #4] @ R14[4]=ROWr32[1]=R5
295 str r5, [r14, #8] @ R14[8]=ROWr32[2]=R5
296 str r5, [r14, #12] @ R14[12]=ROWr32[3]=R5
299 @@ at this point, R0-R11 (free)
300 @@ R12=__const_ptr_, R14=&block[n]
301 ldr r0, [sp, #0] @ R0=block
302 teq r0, r14 @ compare current &block[8*n] to block, when block is reached, the loop is finished.
308 @@ at this point, R0=block, R1-R11 (free)
309 @@ R12=__const_ptr_, R14=&block[n]
310 add r14, r0, #14 @ R14=&block[7], better start from the last col, and decrease the value until col=0, i.e. R14=block.
314 @@ at this point, R0=block (temp), R1-R11 (free)
315 @@ R12=__const_ptr_, R14=&block[n]
316 @@ proceed with b0-b3 first, followed by a0-a3
317 @@ MUL16(b0, W1, col[8x1]);
318 @@ MUL16(b1, W3, col[8x1]);
319 @@ MUL16(b2, W5, col[8x1]);
320 @@ MUL16(b3, W7, col[8x1]);
321 @@ MAC16(b0, W3, col[8x3]);
322 @@ MAC16(b1, -W7, col[8x3]);
323 @@ MAC16(b2, -W1, col[8x3]);
324 @@ MAC16(b3, -W5, col[8x3]);
325 ldr r8, [r12, #offW1] @ R8=W1
327 mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
328 ldr r9, [r12, #offW3] @ R9=W3
329 ldr r10, [r12, #offW5] @ R10=W5
330 mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
331 ldr r11, [r12, #offW7] @ R11=W7
332 mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
334 mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
335 teq r2, #0 @ if 0, then avoid muls
337 mlane r0, r9, r2, r0 @ R0+=W3*ROWr16[3]=b0 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
338 rsbne r2, r2, #0 @ R2=-ROWr16[3]
339 mlane r1, r11, r2, r1 @ R1-=W7*ROWr16[3]=b1 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
340 mlane r5, r8, r2, r5 @ R5-=W1*ROWr16[3]=b2 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
342 mlane r7, r10, r2, r7 @ R7-=W5*ROWr16[3]=b3 (ROWr16[3] must be the second arg, to have the possibility to save 1 cycle)
344 @@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free),
345 @@ R5=b2, R6 (free), R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
346 @@ R12=__const_ptr_, R14=&block[n]
347 @@ MAC16(b0, W5, col[5x8]);
348 @@ MAC16(b2, W7, col[5x8]);
349 @@ MAC16(b3, W3, col[5x8]);
350 @@ MAC16(b1, -W1, col[5x8]);
351 @@ MAC16(b0, W7, col[7x8]);
352 @@ MAC16(b2, W3, col[7x8]);
353 @@ MAC16(b3, -W1, col[7x8]);
354 @@ MAC16(b1, -W5, col[7x8]);
355 ldrsh r3, [r14, #80] @ R3=COLr16[5x8]
356 teq r3, #0 @ if 0 then avoid muls
358 mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5x8]=b0
359 mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5x8]=b2
360 mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5x8]=b3
361 rsbne r3, r3, #0 @ R3=-ROWr16[5x8]
362 ldrsh r4, [r14, #112] @ R4=COLr16[7x8]
364 mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5x8]=b1
366 teq r4, #0 @ if 0 then avoid muls
368 mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7x8]=b0
369 mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7x8]=b2
370 rsbne r4, r4, #0 @ R4=-ROWr16[7x8]
371 mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7x8]=b3
373 mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7x8]=b1
376 @@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free),
377 @@ R5=b2, R6 (free), R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
378 @@ R12=__const_ptr_, R14=&block[n]
381 @@ a0 = (W4 * col[8x0]) + (1 << (COL_SHIFT - 1));
382 @@ a1 = a0 + W6 * row[2];
383 @@ a2 = a0 - W6 * row[2];
384 @@ a3 = a0 - W2 * row[2];
385 @@ a0 = a0 + W2 * row[2];
387 ldr r9, [r12, #offW4] @ R9=W4
388 mul r6, r9, r6 @ R6=W4*ROWr16[0]
389 ldr r10, [r12, #offW6] @ R10=W6
390 ldrsh r4, [r14, #32] @ R4=ROWr16[2] (a3 not defined yet)
391 add r6, r6, #COL_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(COL_SHIFT-1) (a0)
392 mul r11, r10, r4 @ R11=W6*ROWr16[2]
393 ldr r8, [r12, #offW2] @ R8=W2
394 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
395 sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2)
396 mul r11, r8, r4 @ R11=W2*ROWr16[2]
397 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
398 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
400 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
401 @@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
402 @@ R12=__const_ptr_, R14=&block[n]
407 ldrsh r11, [r14, #64] @ R11=ROWr16[4]
408 teq r11, #0 @ if null avoid muls
410 mulne r11, r9, r11 @ R11=W4*ROWr16[4]
412 addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0)
413 subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1)
414 subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2)
415 ldrsh r9, [r14, #96] @ R9=ROWr16[6]
417 addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3)
418 @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
419 teq r9, #0 @ if null avoid muls
421 mulne r11, r10, r9 @ R11=W6*ROWr16[6]
422 addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0)
423 mulne r10, r8, r9 @ R10=W2*ROWr16[6]
428 subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3)
430 subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1)
431 addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2)
433 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
434 @@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
435 @@ R12=__const_ptr_, R14=&block[n]
436 @@ col[0 ] = ((a0 + b0) >> COL_SHIFT);
437 @@ col[8 ] = ((a1 + b1) >> COL_SHIFT);
438 @@ col[16] = ((a2 + b2) >> COL_SHIFT);
439 @@ col[24] = ((a3 + b3) >> COL_SHIFT);
440 @@ col[32] = ((a3 - b3) >> COL_SHIFT);
441 @@ col[40] = ((a2 - b2) >> COL_SHIFT);
442 @@ col[48] = ((a1 - b1) >> COL_SHIFT);
443 @@ col[56] = ((a0 - b0) >> COL_SHIFT);
444 @@@@@ no optimization here @@@@@
445 add r8, r6, r0 @ R8=a0+b0
446 add r9, r2, r1 @ R9=a1+b1
447 mov r8, r8, asr #COL_SHIFT
448 mov r9, r9, asr #COL_SHIFT
451 add r8, r3, r5 @ R8=a2+b2
452 add r9, r4, r7 @ R9=a3+b3
453 mov r8, r8, asr #COL_SHIFT
454 mov r9, r9, asr #COL_SHIFT
457 sub r8, r4, r7 @ R8=a3-b3
458 sub r9, r3, r5 @ R9=a2-b2
459 mov r8, r8, asr #COL_SHIFT
460 mov r9, r9, asr #COL_SHIFT
463 sub r8, r2, r1 @ R8=a1-b1
464 sub r9, r6, r0 @ R9=a0-b0
465 mov r8, r8, asr #COL_SHIFT
466 mov r9, r9, asr #COL_SHIFT
471 @@ at this point, R0-R11 (free)
472 @@ R12=__const_ptr_, R14=&block[n]
473 ldr r0, [sp, #0] @ R0=block
474 teq r0, r14 @ compare current &block[n] to block, when block is reached, the loop is finished.
481 __end_simple_idct_arm:
482 @@ restore registers to previous status!
483 add sp, sp, #8 @@ the local variables!
484 ldmfd sp!, {r4-r11, r15} @@ update PC with LR content.
488 @@ kind of sub-function, here not to overload the common case.
489 __end_bef_a_evaluation:
490 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
491 mul r11, r8, r4 @ R11=W2*ROWr16[2]
492 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
493 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
494 bal __end_a_evaluation
497 __constant_ptr__: @@ see #defines at the beginning of the source code for values.