3 * Copyright (C) 2002 Frederic 'dilb' Boulay.
6 * Author: Frederic Boulay <dilb@handhelds.org>
8 * The function defined in this file is derived from the simple_idct function
9 * from the libavcodec library part of the FFmpeg project.
11 * This file is part of FFmpeg.
13 * FFmpeg is free software; you can redistribute it and/or
14 * modify it under the terms of the GNU Lesser General Public
15 * License as published by the Free Software Foundation; either
16 * version 2.1 of the License, or (at your option) any later version.
18 * FFmpeg is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
21 * Lesser General Public License for more details.
23 * You should have received a copy of the GNU Lesser General Public
24 * License along with FFmpeg; if not, write to the Free Software
25 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
28 /* useful constants for the algorithm, they are save in __constant_ptr__ at */
29 /* the end of the source code.*/
37 #define MASK_MSHW 0xFFFF0000
39 /* offsets of the constants in the vector */
47 #define offMASK_MSHW 28
50 #define ROW_SHIFT2MSHW (16-11)
52 #define ROW_SHIFTED_1 1024 /* 1<< (ROW_SHIFT-1) */
53 #define COL_SHIFTED_1 524288 /* 1<< (COL_SHIFT-1) */
58 .global simple_idct_ARM
61 @@ void simple_idct_ARM(int16_t *block)
62 @@ save stack for reg needed (take all of them),
63 @@ R0-R3 are scratch regs, so no need to save them, but R0 contains the pointer to block
64 @@ so it must not be overwritten, if it is not saved!!
65 @@ R12 is another scratch register, so it should not be saved too
67 stmfd sp!, {r4-r11, r14} @ R14 is also called LR
68 @@ at this point, R0=block, other registers are free.
69 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.
70 add r12, pc, #(__constant_ptr__-.-8) @ R12=__constant_ptr__, the vector containing the constants, probably not necessary to reserve a register for it
71 @@ add 2 temporary variables in the stack: R0 and R14
72 sub sp, sp, #8 @ allow 2 local variables
73 str r0, [sp, #0] @ save block in sp[0]
79 @@ at this point, R0=block, R14=&block[56], R12=__const_ptr_, R1-R11 free
83 @@ read the row and check if it is null, almost null, or not, according to strongarm specs, it is not necessary to optimise ldr accesses (i.e. split 32bits in 2 16bits words), at least it gives more usable registers :)
84 ldr r1, [r14, #0] @ R1=(int32)(R12)[0]=ROWr32[0] (relative row cast to a 32b pointer)
85 ldr r2, [r14, #4] @ R2=(int32)(R12)[1]=ROWr32[1]
86 ldr r3, [r14, #8] @ R3=ROWr32[2]
87 ldr r4, [r14, #12] @ R4=ROWr32[3]
88 @@ check if the words are null, if all of them are null, then proceed with next row (branch __end_row_loop),
89 @@ if ROWr16[0] is the only one not null, then proceed with this special case (branch __almost_empty_row)
90 @@ else follow the complete algorithm.
91 @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
92 @@ R3=ROWr32[2], R4=ROWr32[3], R5-R11 free
93 orr r5, r4, r3 @ R5=R4 | R3
94 orr r5, r5, r2 @ R5=R4 | R3 | R2
95 orrs r6, r5, r1 @ Test R5 | R1 (the aim is to check if everything is null)
97 mov r7, r1, asr #16 @ R7=R1>>16=ROWr16[1] (evaluate it now, as it could be useful later)
98 ldrsh r6, [r14, #0] @ R6=ROWr16[0]
99 orrs r5, r5, r7 @ R5=R4 | R3 | R2 | R7
100 beq __almost_empty_row
103 @@ at this point, R0=block (temp), R1(free), R2=ROWr32[1], R3=ROWr32[2], R4=ROWr32[3],
104 @@ R5=(temp), R6=ROWr16[0], R7=ROWr16[1], R8-R11 free,
105 @@ R12=__const_ptr_, R14=&block[n]
106 @@ to save some registers/calls, proceed with b0-b3 first, followed by a0-a3
108 @@ MUL16(b0, W1, row[1]);
109 @@ MUL16(b1, W3, row[1]);
110 @@ MUL16(b2, W5, row[1]);
111 @@ MUL16(b3, W7, row[1]);
112 @@ MAC16(b0, W3, row[3]);
113 @@ MAC16(b1, -W7, row[3]);
114 @@ MAC16(b2, -W1, row[3]);
115 @@ MAC16(b3, -W5, row[3]);
116 ldr r8, [r12, #offW1] @ R8=W1
117 mov r2, r2, asr #16 @ R2=ROWr16[3]
118 mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
119 ldr r9, [r12, #offW3] @ R9=W3
120 ldr r10, [r12, #offW5] @ R10=W5
121 mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
122 ldr r11, [r12, #offW7] @ R11=W7
123 mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
124 mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
125 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)
130 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)
132 @@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3],
133 @@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
134 @@ R12=__const_ptr_, R14=&block[n]
135 @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
137 orrs r2, r3, r4 @ R2=ROWr32[2] | ROWr32[3]
138 beq __end_b_evaluation
140 @@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3],
141 @@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
142 @@ R12=__const_ptr_, R14=&block[n]
143 @@ MAC16(b0, W5, row[5]);
144 @@ MAC16(b2, W7, row[5]);
145 @@ MAC16(b3, W3, row[5]);
146 @@ MAC16(b1, -W1, row[5]);
147 @@ MAC16(b0, W7, row[7]);
148 @@ MAC16(b2, W3, row[7]);
149 @@ MAC16(b3, -W1, row[7]);
150 @@ MAC16(b1, -W5, row[7]);
151 mov r3, r3, asr #16 @ R3=ROWr16[5]
152 teq r3, #0 @ if null avoid muls
153 mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5]=b0
154 mov r4, r4, asr #16 @ R4=ROWr16[7]
155 mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5]=b2
156 mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5]=b3
157 rsbne r3, r3, #0 @ R3=-ROWr16[5]
158 mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5]=b1
160 teq r4, #0 @ if null avoid muls
161 mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7]=b0
162 mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7]=b2
163 rsbne r4, r4, #0 @ R4=-ROWr16[7]
164 mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7]=b3
165 mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7]=b1
168 @@ at this point, R0=b0, R1=b1, R2=ROWr32[2] | ROWr32[3] (tmp), R3 (free), R4 (free),
169 @@ R5=b2, R6=ROWr16[0], R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
170 @@ R12=__const_ptr_, R14=&block[n]
173 @@ a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1));
174 @@ a1 = a0 + W6 * row[2];
175 @@ a2 = a0 - W6 * row[2];
176 @@ a3 = a0 - W2 * row[2];
177 @@ a0 = a0 + W2 * row[2];
178 ldr r9, [r12, #offW4] @ R9=W4
179 mul r6, r9, r6 @ R6=W4*ROWr16[0]
180 ldr r10, [r12, #offW6] @ R10=W6
181 ldrsh r4, [r14, #4] @ R4=ROWr16[2] (a3 not defined yet)
182 add r6, r6, #ROW_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(ROW_SHIFT-1) (a0)
184 mul r11, r10, r4 @ R11=W6*ROWr16[2]
185 ldr r8, [r12, #offW2] @ R8=W2
186 sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2)
187 @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
190 beq __end_bef_a_evaluation
192 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
193 mul r11, r8, r4 @ R11=W2*ROWr16[2]
194 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
195 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
198 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
199 @@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
200 @@ R12=__const_ptr_, R14=&block[n]
207 ldrsh r11, [r14, #8] @ R11=ROWr16[4]
208 teq r11, #0 @ if null avoid muls
209 mulne r11, r9, r11 @ R11=W4*ROWr16[4]
211 ldrsh r9, [r14, #12] @ R9=ROWr16[6]
212 addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0)
213 subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1)
214 subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2)
215 addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3)
216 @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
217 teq r9, #0 @ if null avoid muls
218 mulne r11, r10, r9 @ R11=W6*ROWr16[6]
219 addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0)
220 mulne r10, r8, r9 @ R10=W2*ROWr16[6]
225 subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3)
226 subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1)
227 addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2)
230 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
231 @@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
232 @@ R12=__const_ptr_, R14=&block[n]
233 @@ row[0] = (a0 + b0) >> ROW_SHIFT;
234 @@ row[1] = (a1 + b1) >> ROW_SHIFT;
235 @@ row[2] = (a2 + b2) >> ROW_SHIFT;
236 @@ row[3] = (a3 + b3) >> ROW_SHIFT;
237 @@ row[4] = (a3 - b3) >> ROW_SHIFT;
238 @@ row[5] = (a2 - b2) >> ROW_SHIFT;
239 @@ row[6] = (a1 - b1) >> ROW_SHIFT;
240 @@ row[7] = (a0 - b0) >> ROW_SHIFT;
241 add r8, r6, r0 @ R8=a0+b0
242 add r9, r2, r1 @ R9=a1+b1
243 @@ put 2 16 bits half-words in a 32bits word
244 @@ ROWr32[0]=ROWr16[0] | (ROWr16[1]<<16) (only Little Endian compliant then!!!)
245 ldr r10, [r12, #offMASK_MSHW] @ R10=0xFFFF0000
246 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a1+b1)<<5)
247 mvn r11, r10 @ R11= NOT R10= 0x0000FFFF
248 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a0+b0)>>11)
252 add r8, r3, r5 @ R8=a2+b2
253 add r9, r4, r7 @ R9=a3+b3
254 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a3+b3)<<5)
255 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a2+b2)>>11)
259 sub r8, r4, r7 @ R8=a3-b3
260 sub r9, r3, r5 @ R9=a2-b2
261 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a2-b2)<<5)
262 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a3-b3)>>11)
266 sub r8, r2, r1 @ R8=a1-b1
267 sub r9, r6, r0 @ R9=a0-b0
268 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a0-b0)<<5)
269 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a1-b1)>>11)
276 @@ the row was empty, except ROWr16[0], now, management of this special case
277 @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
278 @@ R3=ROWr32[2], R4=ROWr32[3], R5=(temp), R6=ROWr16[0], R7=ROWr16[1],
279 @@ R8=0xFFFF (temp), R9-R11 free
280 mov r8, #0x10000 @ R8=0xFFFF (2 steps needed!) it saves a ldr call (because of delay run).
281 sub r8, r8, #1 @ R8 is now ready.
282 and r5, r8, r6, lsl #3 @ R5=R8 & (R6<<3)= (ROWr16[0]<<3) & 0xFFFF
283 orr r5, r5, r5, lsl #16 @ R5=R5 | (R5<<16)
284 str r5, [r14, #0] @ R14[0]=ROWr32[0]=R5
285 str r5, [r14, #4] @ R14[4]=ROWr32[1]=R5
286 str r5, [r14, #8] @ R14[8]=ROWr32[2]=R5
287 str r5, [r14, #12] @ R14[12]=ROWr32[3]=R5
290 @@ at this point, R0-R11 (free)
291 @@ R12=__const_ptr_, R14=&block[n]
292 ldr r0, [sp, #0] @ R0=block
293 teq r0, r14 @ compare current &block[8*n] to block, when block is reached, the loop is finished.
299 @@ at this point, R0=block, R1-R11 (free)
300 @@ R12=__const_ptr_, R14=&block[n]
301 add r14, r0, #14 @ R14=&block[7], better start from the last col, and decrease the value until col=0, i.e. R14=block.
305 @@ at this point, R0=block (temp), R1-R11 (free)
306 @@ R12=__const_ptr_, R14=&block[n]
307 @@ proceed with b0-b3 first, followed by a0-a3
308 @@ MUL16(b0, W1, col[8x1]);
309 @@ MUL16(b1, W3, col[8x1]);
310 @@ MUL16(b2, W5, col[8x1]);
311 @@ MUL16(b3, W7, col[8x1]);
312 @@ MAC16(b0, W3, col[8x3]);
313 @@ MAC16(b1, -W7, col[8x3]);
314 @@ MAC16(b2, -W1, col[8x3]);
315 @@ MAC16(b3, -W5, col[8x3]);
316 ldr r8, [r12, #offW1] @ R8=W1
318 mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
319 ldr r9, [r12, #offW3] @ R9=W3
320 ldr r10, [r12, #offW5] @ R10=W5
321 mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
322 ldr r11, [r12, #offW7] @ R11=W7
323 mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
325 mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
326 teq r2, #0 @ if 0, then avoid muls
327 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)
328 rsbne r2, r2, #0 @ R2=-ROWr16[3]
329 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)
330 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)
331 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)
333 @@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free),
334 @@ R5=b2, R6 (free), R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
335 @@ R12=__const_ptr_, R14=&block[n]
336 @@ MAC16(b0, W5, col[5x8]);
337 @@ MAC16(b2, W7, col[5x8]);
338 @@ MAC16(b3, W3, col[5x8]);
339 @@ MAC16(b1, -W1, col[5x8]);
340 @@ MAC16(b0, W7, col[7x8]);
341 @@ MAC16(b2, W3, col[7x8]);
342 @@ MAC16(b3, -W1, col[7x8]);
343 @@ MAC16(b1, -W5, col[7x8]);
344 ldrsh r3, [r14, #80] @ R3=COLr16[5x8]
345 teq r3, #0 @ if 0 then avoid muls
346 mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5x8]=b0
347 mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5x8]=b2
348 mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5x8]=b3
349 rsbne r3, r3, #0 @ R3=-ROWr16[5x8]
350 ldrsh r4, [r14, #112] @ R4=COLr16[7x8]
351 mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5x8]=b1
353 teq r4, #0 @ if 0 then avoid muls
354 mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7x8]=b0
355 mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7x8]=b2
356 rsbne r4, r4, #0 @ R4=-ROWr16[7x8]
357 mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7x8]=b3
358 mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7x8]=b1
361 @@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free),
362 @@ R5=b2, R6 (free), R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
363 @@ R12=__const_ptr_, R14=&block[n]
366 @@ a0 = (W4 * col[8x0]) + (1 << (COL_SHIFT - 1));
367 @@ a1 = a0 + W6 * row[2];
368 @@ a2 = a0 - W6 * row[2];
369 @@ a3 = a0 - W2 * row[2];
370 @@ a0 = a0 + W2 * row[2];
372 ldr r9, [r12, #offW4] @ R9=W4
373 mul r6, r9, r6 @ R6=W4*ROWr16[0]
374 ldr r10, [r12, #offW6] @ R10=W6
375 ldrsh r4, [r14, #32] @ R4=ROWr16[2] (a3 not defined yet)
376 add r6, r6, #COL_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(COL_SHIFT-1) (a0)
377 mul r11, r10, r4 @ R11=W6*ROWr16[2]
378 ldr r8, [r12, #offW2] @ R8=W2
379 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
380 sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2)
381 mul r11, r8, r4 @ R11=W2*ROWr16[2]
382 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
383 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
385 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
386 @@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
387 @@ R12=__const_ptr_, R14=&block[n]
392 ldrsh r11, [r14, #64] @ R11=ROWr16[4]
393 teq r11, #0 @ if null avoid muls
394 mulne r11, r9, r11 @ R11=W4*ROWr16[4]
396 addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0)
397 subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1)
398 subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2)
399 ldrsh r9, [r14, #96] @ R9=ROWr16[6]
400 addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3)
401 @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
402 teq r9, #0 @ if null avoid muls
403 mulne r11, r10, r9 @ R11=W6*ROWr16[6]
404 addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0)
405 mulne r10, r8, r9 @ R10=W2*ROWr16[6]
410 subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3)
411 subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1)
412 addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2)
414 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
415 @@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
416 @@ R12=__const_ptr_, R14=&block[n]
417 @@ col[0 ] = ((a0 + b0) >> COL_SHIFT);
418 @@ col[8 ] = ((a1 + b1) >> COL_SHIFT);
419 @@ col[16] = ((a2 + b2) >> COL_SHIFT);
420 @@ col[24] = ((a3 + b3) >> COL_SHIFT);
421 @@ col[32] = ((a3 - b3) >> COL_SHIFT);
422 @@ col[40] = ((a2 - b2) >> COL_SHIFT);
423 @@ col[48] = ((a1 - b1) >> COL_SHIFT);
424 @@ col[56] = ((a0 - b0) >> COL_SHIFT);
425 @@@@@ no optimisation here @@@@@
426 add r8, r6, r0 @ R8=a0+b0
427 add r9, r2, r1 @ R9=a1+b1
428 mov r8, r8, asr #COL_SHIFT
429 mov r9, r9, asr #COL_SHIFT
432 add r8, r3, r5 @ R8=a2+b2
433 add r9, r4, r7 @ R9=a3+b3
434 mov r8, r8, asr #COL_SHIFT
435 mov r9, r9, asr #COL_SHIFT
438 sub r8, r4, r7 @ R8=a3-b3
439 sub r9, r3, r5 @ R9=a2-b2
440 mov r8, r8, asr #COL_SHIFT
441 mov r9, r9, asr #COL_SHIFT
444 sub r8, r2, r1 @ R8=a1-b1
445 sub r9, r6, r0 @ R9=a0-b0
446 mov r8, r8, asr #COL_SHIFT
447 mov r9, r9, asr #COL_SHIFT
452 @@ at this point, R0-R11 (free)
453 @@ R12=__const_ptr_, R14=&block[n]
454 ldr r0, [sp, #0] @ R0=block
455 teq r0, r14 @ compare current &block[n] to block, when block is reached, the loop is finished.
462 __end_simple_idct_ARM:
463 @@ restore registers to previous status!
464 add sp, sp, #8 @@ the local variables!
465 ldmfd sp!, {r4-r11, r15} @@ update PC with LR content.
469 @@ kind of sub-function, here not to overload the common case.
470 __end_bef_a_evaluation:
471 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
472 mul r11, r8, r4 @ R11=W2*ROWr16[2]
473 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
474 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
475 bal __end_a_evaluation
478 __constant_ptr__: @@ see #defines at the beginning of the source code for values.