2 * Copyright (C) 2002 Frederic 'dilb' Boulay
4 * Author: Frederic Boulay <dilb@handhelds.org>
6 * The function defined in this file is derived from the simple_idct function
7 * from the libavcodec library part of the Libav project.
9 * This file is part of Libav.
11 * Libav is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public
13 * License as published by the Free Software Foundation; either
14 * version 2.1 of the License, or (at your option) any later version.
16 * Libav is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with Libav; if not, write to the Free Software
23 * 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 function ff_simple_idct_arm, export=1
59 @@ void simple_idct_arm(int16_t *block)
60 @@ save stack for reg needed (take all of them),
61 @@ R0-R3 are scratch regs, so no need to save them, but R0 contains the pointer to block
62 @@ so it must not be overwritten, if it is not saved!!
63 @@ R12 is another scratch register, so it should not be saved too
65 stmfd sp!, {r4-r11, r14} @ R14 is also called LR
66 @@ at this point, R0=block, other registers are free.
67 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.
68 adr r12, __constant_ptr__ @ R12=__constant_ptr__, the vector containing the constants, probably not necessary to reserve a register for it
69 @@ add 2 temporary variables in the stack: R0 and R14
70 sub sp, sp, #8 @ allow 2 local variables
71 str r0, [sp, #0] @ save block in sp[0]
77 @@ at this point, R0=block, R14=&block[56], R12=__const_ptr_, R1-R11 free
81 @@ 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 :)
82 ldr r1, [r14, #0] @ R1=(int32)(R12)[0]=ROWr32[0] (relative row cast to a 32b pointer)
83 ldr r2, [r14, #4] @ R2=(int32)(R12)[1]=ROWr32[1]
84 ldr r3, [r14, #8] @ R3=ROWr32[2]
85 ldr r4, [r14, #12] @ R4=ROWr32[3]
86 @@ check if the words are null, if all of them are null, then proceed with next row (branch __end_row_loop),
87 @@ if ROWr16[0] is the only one not null, then proceed with this special case (branch __almost_empty_row)
88 @@ else follow the complete algorithm.
89 @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
90 @@ R3=ROWr32[2], R4=ROWr32[3], R5-R11 free
91 orr r5, r4, r3 @ R5=R4 | R3
92 orr r5, r5, r2 @ R5=R4 | R3 | R2
93 orrs r6, r5, r1 @ Test R5 | R1 (the aim is to check if everything is null)
95 mov r7, r1, asr #16 @ R7=R1>>16=ROWr16[1] (evaluate it now, as it could be useful later)
96 ldrsh r6, [r14, #0] @ R6=ROWr16[0]
97 orrs r5, r5, r7 @ R5=R4 | R3 | R2 | R7
98 beq __almost_empty_row
101 @@ at this point, R0=block (temp), R1(free), R2=ROWr32[1], R3=ROWr32[2], R4=ROWr32[3],
102 @@ R5=(temp), R6=ROWr16[0], R7=ROWr16[1], R8-R11 free,
103 @@ R12=__const_ptr_, R14=&block[n]
104 @@ to save some registers/calls, proceed with b0-b3 first, followed by a0-a3
106 @@ MUL16(b0, W1, row[1]);
107 @@ MUL16(b1, W3, row[1]);
108 @@ MUL16(b2, W5, row[1]);
109 @@ MUL16(b3, W7, row[1]);
110 @@ MAC16(b0, W3, row[3]);
111 @@ MAC16(b1, -W7, row[3]);
112 @@ MAC16(b2, -W1, row[3]);
113 @@ MAC16(b3, -W5, row[3]);
114 ldr r8, [r12, #offW1] @ R8=W1
115 mov r2, r2, asr #16 @ R2=ROWr16[3]
116 mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
117 ldr r9, [r12, #offW3] @ R9=W3
118 ldr r10, [r12, #offW5] @ R10=W5
119 mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
120 ldr r11, [r12, #offW7] @ R11=W7
121 mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
122 mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
123 teq r2, #0 @ if null avoid muls
125 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)
126 rsbne r2, r2, #0 @ R2=-ROWr16[3]
127 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)
128 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
154 mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5]=b0
155 mov r4, r4, asr #16 @ R4=ROWr16[7]
157 mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5]=b2
158 mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5]=b3
159 rsbne r3, r3, #0 @ R3=-ROWr16[5]
160 mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5]=b1
162 teq r4, #0 @ if null avoid muls
164 mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7]=b0
165 mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7]=b2
166 rsbne r4, r4, #0 @ R4=-ROWr16[7]
167 mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7]=b3
169 mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7]=b1
172 @@ at this point, R0=b0, R1=b1, R2=ROWr32[2] | ROWr32[3] (tmp), R3 (free), R4 (free),
173 @@ R5=b2, R6=ROWr16[0], R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
174 @@ R12=__const_ptr_, R14=&block[n]
177 @@ a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1));
178 @@ a1 = a0 + W6 * row[2];
179 @@ a2 = a0 - W6 * row[2];
180 @@ a3 = a0 - W2 * row[2];
181 @@ a0 = a0 + W2 * row[2];
182 ldr r9, [r12, #offW4] @ R9=W4
183 mul r6, r9, r6 @ R6=W4*ROWr16[0]
184 ldr r10, [r12, #offW6] @ R10=W6
185 ldrsh r4, [r14, #4] @ R4=ROWr16[2] (a3 not defined yet)
186 add r6, r6, #ROW_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(ROW_SHIFT-1) (a0)
188 mul r11, r10, r4 @ R11=W6*ROWr16[2]
189 ldr r8, [r12, #offW2] @ R8=W2
190 sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2)
191 @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
194 beq __end_bef_a_evaluation
196 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
197 mul r11, r8, r4 @ R11=W2*ROWr16[2]
198 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
199 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
202 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
203 @@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
204 @@ R12=__const_ptr_, R14=&block[n]
211 ldrsh r11, [r14, #8] @ R11=ROWr16[4]
212 teq r11, #0 @ if null avoid muls
214 mulne r11, r9, r11 @ R11=W4*ROWr16[4]
216 ldrsh r9, [r14, #12] @ R9=ROWr16[6]
218 addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0)
219 subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1)
220 subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2)
221 addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3)
222 @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
223 teq r9, #0 @ if null avoid muls
225 mulne r11, r10, r9 @ R11=W6*ROWr16[6]
226 addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0)
227 mulne r10, r8, r9 @ R10=W2*ROWr16[6]
232 subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3)
234 subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1)
235 addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2)
238 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
239 @@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
240 @@ R12=__const_ptr_, R14=&block[n]
241 @@ row[0] = (a0 + b0) >> ROW_SHIFT;
242 @@ row[1] = (a1 + b1) >> ROW_SHIFT;
243 @@ row[2] = (a2 + b2) >> ROW_SHIFT;
244 @@ row[3] = (a3 + b3) >> ROW_SHIFT;
245 @@ row[4] = (a3 - b3) >> ROW_SHIFT;
246 @@ row[5] = (a2 - b2) >> ROW_SHIFT;
247 @@ row[6] = (a1 - b1) >> ROW_SHIFT;
248 @@ row[7] = (a0 - b0) >> ROW_SHIFT;
249 add r8, r6, r0 @ R8=a0+b0
250 add r9, r2, r1 @ R9=a1+b1
251 @@ put 2 16 bits half-words in a 32bits word
252 @@ ROWr32[0]=ROWr16[0] | (ROWr16[1]<<16) (only Little Endian compliant then!!!)
253 ldr r10, [r12, #offMASK_MSHW] @ R10=0xFFFF0000
254 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a1+b1)<<5)
255 mvn r11, r10 @ R11= NOT R10= 0x0000FFFF
256 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a0+b0)>>11)
260 add r8, r3, r5 @ R8=a2+b2
261 add r9, r4, r7 @ R9=a3+b3
262 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a3+b3)<<5)
263 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a2+b2)>>11)
267 sub r8, r4, r7 @ R8=a3-b3
268 sub r9, r3, r5 @ R9=a2-b2
269 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a2-b2)<<5)
270 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a3-b3)>>11)
274 sub r8, r2, r1 @ R8=a1-b1
275 sub r9, r6, r0 @ R9=a0-b0
276 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a0-b0)<<5)
277 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a1-b1)>>11)
284 @@ the row was empty, except ROWr16[0], now, management of this special case
285 @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
286 @@ R3=ROWr32[2], R4=ROWr32[3], R5=(temp), R6=ROWr16[0], R7=ROWr16[1],
287 @@ R8=0xFFFF (temp), R9-R11 free
288 mov r8, #0x10000 @ R8=0xFFFF (2 steps needed!) it saves a ldr call (because of delay run).
289 sub r8, r8, #1 @ R8 is now ready.
290 and r5, r8, r6, lsl #3 @ R5=R8 & (R6<<3)= (ROWr16[0]<<3) & 0xFFFF
291 orr r5, r5, r5, lsl #16 @ R5=R5 | (R5<<16)
292 str r5, [r14, #0] @ R14[0]=ROWr32[0]=R5
293 str r5, [r14, #4] @ R14[4]=ROWr32[1]=R5
294 str r5, [r14, #8] @ R14[8]=ROWr32[2]=R5
295 str r5, [r14, #12] @ R14[12]=ROWr32[3]=R5
298 @@ at this point, R0-R11 (free)
299 @@ R12=__const_ptr_, R14=&block[n]
300 ldr r0, [sp, #0] @ R0=block
301 teq r0, r14 @ compare current &block[8*n] to block, when block is reached, the loop is finished.
307 @@ at this point, R0=block, R1-R11 (free)
308 @@ R12=__const_ptr_, R14=&block[n]
309 add r14, r0, #14 @ R14=&block[7], better start from the last col, and decrease the value until col=0, i.e. R14=block.
313 @@ at this point, R0=block (temp), R1-R11 (free)
314 @@ R12=__const_ptr_, R14=&block[n]
315 @@ proceed with b0-b3 first, followed by a0-a3
316 @@ MUL16(b0, W1, col[8x1]);
317 @@ MUL16(b1, W3, col[8x1]);
318 @@ MUL16(b2, W5, col[8x1]);
319 @@ MUL16(b3, W7, col[8x1]);
320 @@ MAC16(b0, W3, col[8x3]);
321 @@ MAC16(b1, -W7, col[8x3]);
322 @@ MAC16(b2, -W1, col[8x3]);
323 @@ MAC16(b3, -W5, col[8x3]);
324 ldr r8, [r12, #offW1] @ R8=W1
326 mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
327 ldr r9, [r12, #offW3] @ R9=W3
328 ldr r10, [r12, #offW5] @ R10=W5
329 mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
330 ldr r11, [r12, #offW7] @ R11=W7
331 mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
333 mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
334 teq r2, #0 @ if 0, then avoid muls
336 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)
337 rsbne r2, r2, #0 @ R2=-ROWr16[3]
338 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)
339 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)
341 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)
343 @@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free),
344 @@ R5=b2, R6 (free), R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
345 @@ R12=__const_ptr_, R14=&block[n]
346 @@ MAC16(b0, W5, col[5x8]);
347 @@ MAC16(b2, W7, col[5x8]);
348 @@ MAC16(b3, W3, col[5x8]);
349 @@ MAC16(b1, -W1, col[5x8]);
350 @@ MAC16(b0, W7, col[7x8]);
351 @@ MAC16(b2, W3, col[7x8]);
352 @@ MAC16(b3, -W1, col[7x8]);
353 @@ MAC16(b1, -W5, col[7x8]);
354 ldrsh r3, [r14, #80] @ R3=COLr16[5x8]
355 teq r3, #0 @ if 0 then avoid muls
357 mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5x8]=b0
358 mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5x8]=b2
359 mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5x8]=b3
360 rsbne r3, r3, #0 @ R3=-ROWr16[5x8]
361 ldrsh r4, [r14, #112] @ R4=COLr16[7x8]
363 mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5x8]=b1
365 teq r4, #0 @ if 0 then avoid muls
367 mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7x8]=b0
368 mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7x8]=b2
369 rsbne r4, r4, #0 @ R4=-ROWr16[7x8]
370 mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7x8]=b3
372 mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7x8]=b1
375 @@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free),
376 @@ R5=b2, R6 (free), R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
377 @@ R12=__const_ptr_, R14=&block[n]
380 @@ a0 = (W4 * col[8x0]) + (1 << (COL_SHIFT - 1));
381 @@ a1 = a0 + W6 * row[2];
382 @@ a2 = a0 - W6 * row[2];
383 @@ a3 = a0 - W2 * row[2];
384 @@ a0 = a0 + W2 * row[2];
386 ldr r9, [r12, #offW4] @ R9=W4
387 mul r6, r9, r6 @ R6=W4*ROWr16[0]
388 ldr r10, [r12, #offW6] @ R10=W6
389 ldrsh r4, [r14, #32] @ R4=ROWr16[2] (a3 not defined yet)
390 add r6, r6, #COL_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(COL_SHIFT-1) (a0)
391 mul r11, r10, r4 @ R11=W6*ROWr16[2]
392 ldr r8, [r12, #offW2] @ R8=W2
393 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
394 sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2)
395 mul r11, r8, r4 @ R11=W2*ROWr16[2]
396 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
397 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
399 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
400 @@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
401 @@ R12=__const_ptr_, R14=&block[n]
406 ldrsh r11, [r14, #64] @ R11=ROWr16[4]
407 teq r11, #0 @ if null avoid muls
409 mulne r11, r9, r11 @ R11=W4*ROWr16[4]
411 addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0)
412 subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1)
413 subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2)
414 ldrsh r9, [r14, #96] @ R9=ROWr16[6]
416 addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3)
417 @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
418 teq r9, #0 @ if null avoid muls
420 mulne r11, r10, r9 @ R11=W6*ROWr16[6]
421 addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0)
422 mulne r10, r8, r9 @ R10=W2*ROWr16[6]
427 subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3)
429 subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1)
430 addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2)
432 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
433 @@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
434 @@ R12=__const_ptr_, R14=&block[n]
435 @@ col[0 ] = ((a0 + b0) >> COL_SHIFT);
436 @@ col[8 ] = ((a1 + b1) >> COL_SHIFT);
437 @@ col[16] = ((a2 + b2) >> COL_SHIFT);
438 @@ col[24] = ((a3 + b3) >> COL_SHIFT);
439 @@ col[32] = ((a3 - b3) >> COL_SHIFT);
440 @@ col[40] = ((a2 - b2) >> COL_SHIFT);
441 @@ col[48] = ((a1 - b1) >> COL_SHIFT);
442 @@ col[56] = ((a0 - b0) >> COL_SHIFT);
443 @@@@@ no optimization here @@@@@
444 add r8, r6, r0 @ R8=a0+b0
445 add r9, r2, r1 @ R9=a1+b1
446 mov r8, r8, asr #COL_SHIFT
447 mov r9, r9, asr #COL_SHIFT
450 add r8, r3, r5 @ R8=a2+b2
451 add r9, r4, r7 @ R9=a3+b3
452 mov r8, r8, asr #COL_SHIFT
453 mov r9, r9, asr #COL_SHIFT
456 sub r8, r4, r7 @ R8=a3-b3
457 sub r9, r3, r5 @ R9=a2-b2
458 mov r8, r8, asr #COL_SHIFT
459 mov r9, r9, asr #COL_SHIFT
462 sub r8, r2, r1 @ R8=a1-b1
463 sub r9, r6, r0 @ R9=a0-b0
464 mov r8, r8, asr #COL_SHIFT
465 mov r9, r9, asr #COL_SHIFT
470 @@ at this point, R0-R11 (free)
471 @@ R12=__const_ptr_, R14=&block[n]
472 ldr r0, [sp, #0] @ R0=block
473 teq r0, r14 @ compare current &block[n] to block, when block is reached, the loop is finished.
480 __end_simple_idct_arm:
481 @@ restore registers to previous status!
482 add sp, sp, #8 @@ the local variables!
483 ldmfd sp!, {r4-r11, r15} @@ update PC with LR content.
487 @@ kind of sub-function, here not to overload the common case.
488 __end_bef_a_evaluation:
489 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
490 mul r11, r8, r4 @ R11=W2*ROWr16[2]
491 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
492 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
493 bal __end_a_evaluation
496 __constant_ptr__: @@ see #defines at the beginning of the source code for values.