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 FFmpeg project.
9 * This file is part of FFmpeg.
11 * FFmpeg 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 * FFmpeg 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 FFmpeg; if not, write to the Free Software
23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 #include "libavutil/arm/asm.S"
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) */
56 function ff_simple_idct_arm, export=1
57 @@ void simple_idct_arm(int16_t *block)
58 @@ save stack for reg needed (take all of them),
59 @@ R0-R3 are scratch regs, so no need to save them, but R0 contains the pointer to block
60 @@ so it must not be overwritten, if it is not saved!!
61 @@ R12 is another scratch register, so it should not be saved too
63 stmfd sp!, {r4-r11, r14} @ R14 is also called LR
64 @@ at this point, R0=block, other registers are free.
65 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.
66 adr r12, __constant_ptr__ @ R12=__constant_ptr__, the vector containing the constants, probably not necessary to reserve a register for it
67 @@ add 2 temporary variables in the stack: R0 and R14
68 sub sp, sp, #8 @ allow 2 local variables
69 str r0, [sp, #0] @ save block in sp[0]
75 @@ at this point, R0=block, R14=&block[56], R12=__const_ptr_, R1-R11 free
79 @@ 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 :)
80 ldr r1, [r14, #0] @ R1=(int32)(R12)[0]=ROWr32[0] (relative row cast to a 32b pointer)
81 ldr r2, [r14, #4] @ R2=(int32)(R12)[1]=ROWr32[1]
82 ldr r3, [r14, #8] @ R3=ROWr32[2]
83 ldr r4, [r14, #12] @ R4=ROWr32[3]
84 @@ check if the words are null, if all of them are null, then proceed with next row (branch __end_row_loop),
85 @@ if ROWr16[0] is the only one not null, then proceed with this special case (branch __almost_empty_row)
86 @@ else follow the complete algorithm.
87 @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
88 @@ R3=ROWr32[2], R4=ROWr32[3], R5-R11 free
89 orr r5, r4, r3 @ R5=R4 | R3
90 orr r5, r5, r2 @ R5=R4 | R3 | R2
91 orrs r6, r5, r1 @ Test R5 | R1 (the aim is to check if everything is null)
93 mov r7, r1, asr #16 @ R7=R1>>16=ROWr16[1] (evaluate it now, as it could be useful later)
94 ldrsh r6, [r14, #0] @ R6=ROWr16[0]
95 orrs r5, r5, r7 @ R5=R4 | R3 | R2 | R7
96 beq __almost_empty_row
99 @@ at this point, R0=block (temp), R1(free), R2=ROWr32[1], R3=ROWr32[2], R4=ROWr32[3],
100 @@ R5=(temp), R6=ROWr16[0], R7=ROWr16[1], R8-R11 free,
101 @@ R12=__const_ptr_, R14=&block[n]
102 @@ to save some registers/calls, proceed with b0-b3 first, followed by a0-a3
104 @@ MUL16(b0, W1, row[1]);
105 @@ MUL16(b1, W3, row[1]);
106 @@ MUL16(b2, W5, row[1]);
107 @@ MUL16(b3, W7, row[1]);
108 @@ MAC16(b0, W3, row[3]);
109 @@ MAC16(b1, -W7, row[3]);
110 @@ MAC16(b2, -W1, row[3]);
111 @@ MAC16(b3, -W5, row[3]);
112 ldr r8, [r12, #offW1] @ R8=W1
113 mov r2, r2, asr #16 @ R2=ROWr16[3]
114 mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
115 ldr r9, [r12, #offW3] @ R9=W3
116 ldr r10, [r12, #offW5] @ R10=W5
117 mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
118 ldr r11, [r12, #offW7] @ R11=W7
119 mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
120 mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
121 teq r2, #0 @ if null avoid muls
123 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)
124 rsbne r2, r2, #0 @ R2=-ROWr16[3]
125 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)
126 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)
128 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)
130 @@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3],
131 @@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
132 @@ R12=__const_ptr_, R14=&block[n]
133 @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
135 orrs r2, r3, r4 @ R2=ROWr32[2] | ROWr32[3]
136 beq __end_b_evaluation
138 @@ at this point, R0=b0, R1=b1, R2 (free), R3=ROWr32[2], R4=ROWr32[3],
139 @@ R5=b2, R6=ROWr16[0], R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
140 @@ R12=__const_ptr_, R14=&block[n]
141 @@ MAC16(b0, W5, row[5]);
142 @@ MAC16(b2, W7, row[5]);
143 @@ MAC16(b3, W3, row[5]);
144 @@ MAC16(b1, -W1, row[5]);
145 @@ MAC16(b0, W7, row[7]);
146 @@ MAC16(b2, W3, row[7]);
147 @@ MAC16(b3, -W1, row[7]);
148 @@ MAC16(b1, -W5, row[7]);
149 mov r3, r3, asr #16 @ R3=ROWr16[5]
150 teq r3, #0 @ if null avoid muls
152 mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5]=b0
153 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
162 mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7]=b0
163 mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7]=b2
164 rsbne r4, r4, #0 @ R4=-ROWr16[7]
165 mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7]=b3
167 mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7]=b1
170 @@ at this point, R0=b0, R1=b1, R2=ROWr32[2] | ROWr32[3] (tmp), R3 (free), R4 (free),
171 @@ R5=b2, R6=ROWr16[0], R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
172 @@ R12=__const_ptr_, R14=&block[n]
175 @@ a0 = (W4 * row[0]) + (1 << (ROW_SHIFT - 1));
176 @@ a1 = a0 + W6 * row[2];
177 @@ a2 = a0 - W6 * row[2];
178 @@ a3 = a0 - W2 * row[2];
179 @@ a0 = a0 + W2 * row[2];
180 ldr r9, [r12, #offW4] @ R9=W4
181 mul r6, r9, r6 @ R6=W4*ROWr16[0]
182 ldr r10, [r12, #offW6] @ R10=W6
183 ldrsh r4, [r14, #4] @ R4=ROWr16[2] (a3 not defined yet)
184 add r6, r6, #ROW_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(ROW_SHIFT-1) (a0)
186 mul r11, r10, r4 @ R11=W6*ROWr16[2]
187 ldr r8, [r12, #offW2] @ R8=W2
188 sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2)
189 @@ temp = ((uint32_t*)row)[2] | ((uint32_t*)row)[3];
192 beq __end_bef_a_evaluation
194 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
195 mul r11, r8, r4 @ R11=W2*ROWr16[2]
196 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
197 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
200 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
201 @@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
202 @@ R12=__const_ptr_, R14=&block[n]
209 ldrsh r11, [r14, #8] @ R11=ROWr16[4]
210 teq r11, #0 @ if null avoid muls
212 mulne r11, r9, r11 @ R11=W4*ROWr16[4]
214 ldrsh r9, [r14, #12] @ R9=ROWr16[6]
216 addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0)
217 subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1)
218 subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2)
219 addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3)
220 @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
221 teq r9, #0 @ if null avoid muls
223 mulne r11, r10, r9 @ R11=W6*ROWr16[6]
224 addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0)
225 mulne r10, r8, r9 @ R10=W2*ROWr16[6]
230 subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3)
232 subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1)
233 addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2)
236 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
237 @@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
238 @@ R12=__const_ptr_, R14=&block[n]
239 @@ row[0] = (a0 + b0) >> ROW_SHIFT;
240 @@ row[1] = (a1 + b1) >> ROW_SHIFT;
241 @@ row[2] = (a2 + b2) >> ROW_SHIFT;
242 @@ row[3] = (a3 + b3) >> ROW_SHIFT;
243 @@ row[4] = (a3 - b3) >> ROW_SHIFT;
244 @@ row[5] = (a2 - b2) >> ROW_SHIFT;
245 @@ row[6] = (a1 - b1) >> ROW_SHIFT;
246 @@ row[7] = (a0 - b0) >> ROW_SHIFT;
247 add r8, r6, r0 @ R8=a0+b0
248 add r9, r2, r1 @ R9=a1+b1
249 @@ put 2 16 bits half-words in a 32bits word
250 @@ ROWr32[0]=ROWr16[0] | (ROWr16[1]<<16) (only Little Endian compliant then!!!)
251 ldr r10, [r12, #offMASK_MSHW] @ R10=0xFFFF0000
252 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a1+b1)<<5)
253 mvn r11, r10 @ R11= NOT R10= 0x0000FFFF
254 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a0+b0)>>11)
258 add r8, r3, r5 @ R8=a2+b2
259 add r9, r4, r7 @ R9=a3+b3
260 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a3+b3)<<5)
261 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a2+b2)>>11)
265 sub r8, r4, r7 @ R8=a3-b3
266 sub r9, r3, r5 @ R9=a2-b2
267 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a2-b2)<<5)
268 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a3-b3)>>11)
272 sub r8, r2, r1 @ R8=a1-b1
273 sub r9, r6, r0 @ R9=a0-b0
274 and r9, r10, r9, lsl #ROW_SHIFT2MSHW @ R9=0xFFFF0000 & ((a0-b0)<<5)
275 and r8, r11, r8, asr #ROW_SHIFT @ R8=0x0000FFFF & ((a1-b1)>>11)
282 @@ the row was empty, except ROWr16[0], now, management of this special case
283 @@ at this point, R0=block, R14=&block[n], R12=__const_ptr_, R1=ROWr32[0], R2=ROWr32[1],
284 @@ R3=ROWr32[2], R4=ROWr32[3], R5=(temp), R6=ROWr16[0], R7=ROWr16[1],
285 @@ R8=0xFFFF (temp), R9-R11 free
286 mov r8, #0x10000 @ R8=0xFFFF (2 steps needed!) it saves a ldr call (because of delay run).
287 sub r8, r8, #1 @ R8 is now ready.
288 and r5, r8, r6, lsl #3 @ R5=R8 & (R6<<3)= (ROWr16[0]<<3) & 0xFFFF
289 orr r5, r5, r5, lsl #16 @ R5=R5 | (R5<<16)
290 str r5, [r14, #0] @ R14[0]=ROWr32[0]=R5
291 str r5, [r14, #4] @ R14[4]=ROWr32[1]=R5
292 str r5, [r14, #8] @ R14[8]=ROWr32[2]=R5
293 str r5, [r14, #12] @ R14[12]=ROWr32[3]=R5
296 @@ at this point, R0-R11 (free)
297 @@ R12=__const_ptr_, R14=&block[n]
298 ldr r0, [sp, #0] @ R0=block
299 teq r0, r14 @ compare current &block[8*n] to block, when block is reached, the loop is finished.
305 @@ at this point, R0=block, R1-R11 (free)
306 @@ R12=__const_ptr_, R14=&block[n]
307 add r14, r0, #14 @ R14=&block[7], better start from the last col, and decrease the value until col=0, i.e. R14=block.
311 @@ at this point, R0=block (temp), R1-R11 (free)
312 @@ R12=__const_ptr_, R14=&block[n]
313 @@ proceed with b0-b3 first, followed by a0-a3
314 @@ MUL16(b0, W1, col[8x1]);
315 @@ MUL16(b1, W3, col[8x1]);
316 @@ MUL16(b2, W5, col[8x1]);
317 @@ MUL16(b3, W7, col[8x1]);
318 @@ MAC16(b0, W3, col[8x3]);
319 @@ MAC16(b1, -W7, col[8x3]);
320 @@ MAC16(b2, -W1, col[8x3]);
321 @@ MAC16(b3, -W5, col[8x3]);
322 ldr r8, [r12, #offW1] @ R8=W1
324 mul r0, r8, r7 @ R0=W1*ROWr16[1]=b0 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
325 ldr r9, [r12, #offW3] @ R9=W3
326 ldr r10, [r12, #offW5] @ R10=W5
327 mul r1, r9, r7 @ R1=W3*ROWr16[1]=b1 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
328 ldr r11, [r12, #offW7] @ R11=W7
329 mul r5, r10, r7 @ R5=W5*ROWr16[1]=b2 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
331 mul r7, r11, r7 @ R7=W7*ROWr16[1]=b3 (ROWr16[1] must be the second arg, to have the possibility to save 1 cycle)
332 teq r2, #0 @ if 0, then avoid muls
334 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)
335 rsbne r2, r2, #0 @ R2=-ROWr16[3]
336 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)
337 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)
339 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)
341 @@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free),
342 @@ R5=b2, R6 (free), R7=b3, R8=W1, R9=W3, R10=W5, R11=W7,
343 @@ R12=__const_ptr_, R14=&block[n]
344 @@ MAC16(b0, W5, col[5x8]);
345 @@ MAC16(b2, W7, col[5x8]);
346 @@ MAC16(b3, W3, col[5x8]);
347 @@ MAC16(b1, -W1, col[5x8]);
348 @@ MAC16(b0, W7, col[7x8]);
349 @@ MAC16(b2, W3, col[7x8]);
350 @@ MAC16(b3, -W1, col[7x8]);
351 @@ MAC16(b1, -W5, col[7x8]);
352 ldrsh r3, [r14, #80] @ R3=COLr16[5x8]
353 teq r3, #0 @ if 0 then avoid muls
355 mlane r0, r10, r3, r0 @ R0+=W5*ROWr16[5x8]=b0
356 mlane r5, r11, r3, r5 @ R5+=W7*ROWr16[5x8]=b2
357 mlane r7, r9, r3, r7 @ R7+=W3*ROWr16[5x8]=b3
358 rsbne r3, r3, #0 @ R3=-ROWr16[5x8]
359 ldrsh r4, [r14, #112] @ R4=COLr16[7x8]
361 mlane r1, r8, r3, r1 @ R7-=W1*ROWr16[5x8]=b1
363 teq r4, #0 @ if 0 then avoid muls
365 mlane r0, r11, r4, r0 @ R0+=W7*ROWr16[7x8]=b0
366 mlane r5, r9, r4, r5 @ R5+=W3*ROWr16[7x8]=b2
367 rsbne r4, r4, #0 @ R4=-ROWr16[7x8]
368 mlane r7, r8, r4, r7 @ R7-=W1*ROWr16[7x8]=b3
370 mlane r1, r10, r4, r1 @ R1-=W5*ROWr16[7x8]=b1
373 @@ at this point, R0=b0, R1=b1, R2 (free), R3 (free), R4 (free),
374 @@ R5=b2, R6 (free), R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
375 @@ R12=__const_ptr_, R14=&block[n]
378 @@ a0 = (W4 * col[8x0]) + (1 << (COL_SHIFT - 1));
379 @@ a1 = a0 + W6 * row[2];
380 @@ a2 = a0 - W6 * row[2];
381 @@ a3 = a0 - W2 * row[2];
382 @@ a0 = a0 + W2 * row[2];
384 ldr r9, [r12, #offW4] @ R9=W4
385 mul r6, r9, r6 @ R6=W4*ROWr16[0]
386 ldr r10, [r12, #offW6] @ R10=W6
387 ldrsh r4, [r14, #32] @ R4=ROWr16[2] (a3 not defined yet)
388 add r6, r6, #COL_SHIFTED_1 @ R6=W4*ROWr16[0] + 1<<(COL_SHIFT-1) (a0)
389 mul r11, r10, r4 @ R11=W6*ROWr16[2]
390 ldr r8, [r12, #offW2] @ R8=W2
391 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
392 sub r3, r6, r11 @ R3=a0-W6*ROWr16[2] (a2)
393 mul r11, r8, r4 @ R11=W2*ROWr16[2]
394 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
395 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
397 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
398 @@ R5=b2, R6=a0, R7=b3, R8=W2, R9=W4, R10=W6, R11 (free),
399 @@ R12=__const_ptr_, R14=&block[n]
404 ldrsh r11, [r14, #64] @ R11=ROWr16[4]
405 teq r11, #0 @ if null avoid muls
407 mulne r11, r9, r11 @ R11=W4*ROWr16[4]
409 addne r6, r6, r11 @ R6+=W4*ROWr16[4] (a0)
410 subne r2, r2, r11 @ R2-=W4*ROWr16[4] (a1)
411 subne r3, r3, r11 @ R3-=W4*ROWr16[4] (a2)
412 ldrsh r9, [r14, #96] @ R9=ROWr16[6]
414 addne r4, r4, r11 @ R4+=W4*ROWr16[4] (a3)
415 @@ W6 alone is no more useful, save W2*ROWr16[6] in it instead
416 teq r9, #0 @ if null avoid muls
418 mulne r11, r10, r9 @ R11=W6*ROWr16[6]
419 addne r6, r6, r11 @ R6+=W6*ROWr16[6] (a0)
420 mulne r10, r8, r9 @ R10=W2*ROWr16[6]
425 subne r4, r4, r11 @ R4-=W6*ROWr16[6] (a3)
427 subne r2, r2, r10 @ R2-=W2*ROWr16[6] (a1)
428 addne r3, r3, r10 @ R3+=W2*ROWr16[6] (a2)
430 @@ at this point, R0=b0, R1=b1, R2=a1, R3=a2, R4=a3,
431 @@ R5=b2, R6=a0, R7=b3, R8 (free), R9 (free), R10 (free), R11 (free),
432 @@ R12=__const_ptr_, R14=&block[n]
433 @@ col[0 ] = ((a0 + b0) >> COL_SHIFT);
434 @@ col[8 ] = ((a1 + b1) >> COL_SHIFT);
435 @@ col[16] = ((a2 + b2) >> COL_SHIFT);
436 @@ col[24] = ((a3 + b3) >> COL_SHIFT);
437 @@ col[32] = ((a3 - b3) >> COL_SHIFT);
438 @@ col[40] = ((a2 - b2) >> COL_SHIFT);
439 @@ col[48] = ((a1 - b1) >> COL_SHIFT);
440 @@ col[56] = ((a0 - b0) >> COL_SHIFT);
441 @@@@@ no optimization here @@@@@
442 add r8, r6, r0 @ R8=a0+b0
443 add r9, r2, r1 @ R9=a1+b1
444 mov r8, r8, asr #COL_SHIFT
445 mov r9, r9, asr #COL_SHIFT
448 add r8, r3, r5 @ R8=a2+b2
449 add r9, r4, r7 @ R9=a3+b3
450 mov r8, r8, asr #COL_SHIFT
451 mov r9, r9, asr #COL_SHIFT
454 sub r8, r4, r7 @ R8=a3-b3
455 sub r9, r3, r5 @ R9=a2-b2
456 mov r8, r8, asr #COL_SHIFT
457 mov r9, r9, asr #COL_SHIFT
460 sub r8, r2, r1 @ R8=a1-b1
461 sub r9, r6, r0 @ R9=a0-b0
462 mov r8, r8, asr #COL_SHIFT
463 mov r9, r9, asr #COL_SHIFT
468 @@ at this point, R0-R11 (free)
469 @@ R12=__const_ptr_, R14=&block[n]
470 ldr r0, [sp, #0] @ R0=block
471 teq r0, r14 @ compare current &block[n] to block, when block is reached, the loop is finished.
478 __end_simple_idct_arm:
479 @@ restore registers to previous status!
480 add sp, sp, #8 @@ the local variables!
481 ldmfd sp!, {r4-r11, r15} @@ update PC with LR content.
485 @@ kind of sub-function, here not to overload the common case.
486 __end_bef_a_evaluation:
487 add r2, r6, r11 @ R2=a0+W6*ROWr16[2] (a1)
488 mul r11, r8, r4 @ R11=W2*ROWr16[2]
489 sub r4, r6, r11 @ R4=a0-W2*ROWr16[2] (a3)
490 add r6, r6, r11 @ R6=a0+W2*ROWr16[2] (a0)
491 bal __end_a_evaluation
495 __constant_ptr__: @@ see #defines at the beginning of the source code for values.