summaryrefslogtreecommitdiffstats
blob: 092f140924b7870fb42d79a5853455ed0ab58af4 (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
/* libs/pixelflinger/codeflinger/blending.cpp
**
** Copyright 2006, The Android Open Source Project
**
** Licensed under the Apache License, Version 2.0 (the "License");
** you may not use this file except in compliance with the License.
** You may obtain a copy of the License at
**
**     http://www.apache.org/licenses/LICENSE-2.0
**
** Unless required by applicable law or agreed to in writing, software
** distributed under the License is distributed on an "AS IS" BASIS,
** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
** See the License for the specific language governing permissions and
** limitations under the License.
*/

#define LOG_TAG "pixelflinger-code"

#include <assert.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>

#include <android/log.h>

#include "GGLAssembler.h"

namespace android {

void GGLAssembler::build_fog(
                        component_t& temp,      // incomming fragment / output
                        int component,
                        Scratch& regs)
{
   if (mInfo[component].fog) {
        Scratch scratches(registerFile());
        comment("fog");

        integer_t fragment(temp.reg, temp.h, temp.flags);
        if (!(temp.flags & CORRUPTIBLE)) {
            temp.reg = regs.obtain();
            temp.flags |= CORRUPTIBLE;
        }

        integer_t fogColor(scratches.obtain(), 8, CORRUPTIBLE); 
        LDRB(AL, fogColor.reg, mBuilderContext.Rctx,
                immed12_pre(GGL_OFFSETOF(state.fog.color[component])));

        integer_t factor(scratches.obtain(), 16, CORRUPTIBLE);
        CONTEXT_LOAD(factor.reg, generated_vars.f);

        // clamp fog factor (TODO: see if there is a way to guarantee
        // we won't overflow, when setting the iterators)
        BIC(AL, 0, factor.reg, factor.reg, reg_imm(factor.reg, ASR, 31));
        CMP(AL, factor.reg, imm( 0x10000 ));
        MOV(HS, 0, factor.reg, imm( 0x10000 ));

        build_blendFOneMinusF(temp, factor, fragment, fogColor);
    }
}

void GGLAssembler::build_blending(
                        component_t& temp,      // incomming fragment / output
                        const pixel_t& pixel,   // framebuffer
                        int component,
                        Scratch& regs)
{
   if (!mInfo[component].blend)
        return;
        
    int fs = component==GGLFormat::ALPHA ? mBlendSrcA : mBlendSrc;
    int fd = component==GGLFormat::ALPHA ? mBlendDstA : mBlendDst;
    if (fs==GGL_SRC_ALPHA_SATURATE && component==GGLFormat::ALPHA)
        fs = GGL_ONE;
    const int blending = blending_codes(fs, fd);
    if (!temp.size()) {
        // here, blending will produce something which doesn't depend on
        // that component (eg: GL_ZERO:GL_*), so the register has not been
        // allocated yet. Will never be used as a source.
        temp = component_t(regs.obtain(), CORRUPTIBLE);
    }

    // we are doing real blending...
    // fb:          extracted dst
    // fragment:    extracted src
    // temp:        component_t(fragment) and result

    // scoped register allocator
    Scratch scratches(registerFile());
    comment("blending");

    // we can optimize these cases a bit...
    // (1) saturation is not needed
    // (2) we can use only one multiply instead of 2
    // (3) we can reduce the register pressure
    //      R = S*f + D*(1-f) = (S-D)*f + D
    //      R = S*(1-f) + D*f = (D-S)*f + S

    const bool same_factor_opt1 =
        (fs==GGL_DST_COLOR && fd==GGL_ONE_MINUS_DST_COLOR) ||
        (fs==GGL_SRC_COLOR && fd==GGL_ONE_MINUS_SRC_COLOR) ||
        (fs==GGL_DST_ALPHA && fd==GGL_ONE_MINUS_DST_ALPHA) ||
        (fs==GGL_SRC_ALPHA && fd==GGL_ONE_MINUS_SRC_ALPHA);

    const bool same_factor_opt2 =
        (fs==GGL_ONE_MINUS_DST_COLOR && fd==GGL_DST_COLOR) ||
        (fs==GGL_ONE_MINUS_SRC_COLOR && fd==GGL_SRC_COLOR) || 
        (fs==GGL_ONE_MINUS_DST_ALPHA && fd==GGL_DST_ALPHA) ||
        (fs==GGL_ONE_MINUS_SRC_ALPHA && fd==GGL_SRC_ALPHA);


    // XXX: we could also optimize these cases:
    // R = S*f + D*f = (S+D)*f
    // R = S*(1-f) + D*(1-f) = (S+D)*(1-f)
    // R = S*D + D*S = 2*S*D


    // see if we need to extract 'component' from the destination (fb)
    integer_t fb;
    if (blending & (BLEND_DST|FACTOR_DST)) { 
        fb.setTo(scratches.obtain(), 32); 
        extract(fb, pixel, component);
        if (mDithering) {
            // XXX: maybe what we should do instead, is simply
            // expand fb -or- fragment to the larger of the two
            if (fb.size() < temp.size()) {
                // for now we expand 'fb' to min(fragment, 8)
                int new_size = temp.size() < 8 ? temp.size() : 8;
                expand(fb, fb, new_size);
            }
        }
    }


    // convert input fragment to integer_t
    if (temp.l && (temp.flags & CORRUPTIBLE)) {
        MOV(AL, 0, temp.reg, reg_imm(temp.reg, LSR, temp.l));
        temp.h -= temp.l;
        temp.l = 0;
    }
    integer_t fragment(temp.reg, temp.size(), temp.flags);

    // if not done yet, convert input fragment to integer_t
    if (temp.l) {
        // here we know temp is not CORRUPTIBLE
        fragment.reg = scratches.obtain();
        MOV(AL, 0, fragment.reg, reg_imm(temp.reg, LSR, temp.l));
        fragment.flags |= CORRUPTIBLE;
    }

    if (!(temp.flags & CORRUPTIBLE)) {
        // temp is not corruptible, but since it's the destination it
        // will be modified, so we need to allocate a new register.
        temp.reg = regs.obtain();
        temp.flags &= ~CORRUPTIBLE;
        fragment.flags &= ~CORRUPTIBLE;
    }

    if ((blending & BLEND_SRC) && !same_factor_opt1) {
        // source (fragment) is needed for the blending stage
        // so it's not CORRUPTIBLE (unless we're doing same_factor_opt1)
        fragment.flags &= ~CORRUPTIBLE;
    }


    if (same_factor_opt1) {
        //  R = S*f + D*(1-f) = (S-D)*f + D
        integer_t factor;
        build_blend_factor(factor, fs, 
                component, pixel, fragment, fb, scratches);
        // fb is always corruptible from this point
        fb.flags |= CORRUPTIBLE;
        build_blendFOneMinusF(temp, factor, fragment, fb);
    } else if (same_factor_opt2) {
        //  R = S*(1-f) + D*f = (D-S)*f + S
        integer_t factor;
        // fb is always corrruptible here
        fb.flags |= CORRUPTIBLE;
        build_blend_factor(factor, fd,
                component, pixel, fragment, fb, scratches);
        build_blendOneMinusFF(temp, factor, fragment, fb);
    } else {
        integer_t src_factor;
        integer_t dst_factor;

        // if destination (fb) is not needed for the blending stage, 
        // then it can be marked as CORRUPTIBLE
        if (!(blending & BLEND_DST)) {
            fb.flags |= CORRUPTIBLE;
        }

        // XXX: try to mark some registers as CORRUPTIBLE
        // in most case we could make those corruptible
        // when we're processing the last component
        // but not always, for instance
        //    when fragment is constant and not reloaded
        //    when fb is needed for logic-ops or masking
        //    when a register is aliased (for instance with mAlphaSource)

        // blend away...
        if (fs==GGL_ZERO) {
            if (fd==GGL_ZERO) {         // R = 0
                // already taken care of
            } else if (fd==GGL_ONE) {   // R = D
                // already taken care of
            } else {                    // R = D*fd
                // compute fd
                build_blend_factor(dst_factor, fd,
                        component, pixel, fragment, fb, scratches);
                mul_factor(temp, fb, dst_factor);
            }
        } else if (fs==GGL_ONE) {
            if (fd==GGL_ZERO) {         // R = S
                // NOP, taken care of
            } else if (fd==GGL_ONE) {   // R = S + D
                component_add(temp, fb, fragment); // args order matters
                component_sat(temp);
            } else {                    // R = S + D*fd
                // compute fd
                build_blend_factor(dst_factor, fd,
                        component, pixel, fragment, fb, scratches);
                mul_factor_add(temp, fb, dst_factor, component_t(fragment));
                component_sat(temp);
            }
        } else {
            // compute fs
            build_blend_factor(src_factor, fs, 
                    component, pixel, fragment, fb, scratches);
            if (fd==GGL_ZERO) {         // R = S*fs
                mul_factor(temp, fragment, src_factor);
            } else if (fd==GGL_ONE) {   // R = S*fs + D
                mul_factor_add(temp, fragment, src_factor, component_t(fb));
                component_sat(temp);
            } else {                    // R = S*fs + D*fd
                mul_factor(temp, fragment, src_factor);
                if (scratches.isUsed(src_factor.reg))
                    scratches.recycle(src_factor.reg);
                // compute fd
                build_blend_factor(dst_factor, fd,
                        component, pixel, fragment, fb, scratches);
                mul_factor_add(temp, fb, dst_factor, temp);
                if (!same_factor_opt1 && !same_factor_opt2) {
                    component_sat(temp);
                }
            }
        }
    }

    // now we can be corrupted (it's the dest)
    temp.flags |= CORRUPTIBLE;
}

void GGLAssembler::build_blend_factor(
        integer_t& factor, int f, int component,
        const pixel_t& dst_pixel,
        integer_t& fragment,
        integer_t& fb,
        Scratch& scratches)
{
    integer_t src_alpha(fragment);

    // src_factor/dst_factor won't be used after blending,
    // so it's fine to mark them as CORRUPTIBLE (if not aliased)
    factor.flags |= CORRUPTIBLE;

    switch(f) {
    case GGL_ONE_MINUS_SRC_ALPHA:
    case GGL_SRC_ALPHA:
        if (component==GGLFormat::ALPHA && !isAlphaSourceNeeded()) {
            // we're processing alpha, so we already have
            // src-alpha in fragment, and we need src-alpha just this time.
        } else {
           // alpha-src will be needed for other components
            if (!mBlendFactorCached || mBlendFactorCached==f) {
                src_alpha = mAlphaSource;
                factor = mAlphaSource;
                factor.flags &= ~CORRUPTIBLE;           
                // we already computed the blend factor before, nothing to do.
                if (mBlendFactorCached)
                    return;
                // this is the first time, make sure to compute the blend
                // factor properly.
                mBlendFactorCached = f;
                break;
            } else {
                // we have a cached alpha blend factor, but we want another one,
                // this should really not happen because by construction,
                // we cannot have BOTH source and destination
                // blend factors use ALPHA *and* ONE_MINUS_ALPHA (because
                // the blending stage uses the f/(1-f) optimization
                
                // for completeness, we handle this case though. Since there
                // are only 2 choices, this meens we want "the other one"
                // (1-factor)
                factor = mAlphaSource;
                factor.flags &= ~CORRUPTIBLE;           
                RSB(AL, 0, factor.reg, factor.reg, imm((1<<factor.s)));
                mBlendFactorCached = f;
                return;
            }                
        }
        // fall-through...
    case GGL_ONE_MINUS_DST_COLOR:
    case GGL_DST_COLOR:
    case GGL_ONE_MINUS_SRC_COLOR:
    case GGL_SRC_COLOR:
    case GGL_ONE_MINUS_DST_ALPHA:
    case GGL_DST_ALPHA:
    case GGL_SRC_ALPHA_SATURATE:
        // help us find out what register we can use for the blend-factor
        // CORRUPTIBLE registers are chosen first, or a new one is allocated.
        if (fragment.flags & CORRUPTIBLE) {
            factor.setTo(fragment.reg, 32, CORRUPTIBLE);
            fragment.flags &= ~CORRUPTIBLE;
        } else if (fb.flags & CORRUPTIBLE) {
            factor.setTo(fb.reg, 32, CORRUPTIBLE);
            fb.flags &= ~CORRUPTIBLE;
        } else {
            factor.setTo(scratches.obtain(), 32, CORRUPTIBLE);
        } 
        break;
    }

    // XXX: doesn't work if size==1

    switch(f) {
    case GGL_ONE_MINUS_DST_COLOR:
    case GGL_DST_COLOR:
        factor.s = fb.s;
        ADD(AL, 0, factor.reg, fb.reg, reg_imm(fb.reg, LSR, fb.s-1));
        break;
    case GGL_ONE_MINUS_SRC_COLOR:
    case GGL_SRC_COLOR:
        factor.s = fragment.s;
        ADD(AL, 0, factor.reg, fragment.reg,
            reg_imm(fragment.reg, LSR, fragment.s-1));
        break;
    case GGL_ONE_MINUS_SRC_ALPHA:
    case GGL_SRC_ALPHA:
        factor.s = src_alpha.s;
        ADD(AL, 0, factor.reg, src_alpha.reg,
                reg_imm(src_alpha.reg, LSR, src_alpha.s-1));
        break;
    case GGL_ONE_MINUS_DST_ALPHA:
    case GGL_DST_ALPHA:
        // XXX: should be precomputed
        extract(factor, dst_pixel, GGLFormat::ALPHA);
        ADD(AL, 0, factor.reg, factor.reg,
                reg_imm(factor.reg, LSR, factor.s-1));
        break;
    case GGL_SRC_ALPHA_SATURATE:
        // XXX: should be precomputed
        // XXX: f = min(As, 1-Ad)
        // btw, we're guaranteed that Ad's size is <= 8, because
        // it's extracted from the framebuffer
        break;
    }

    switch(f) {
    case GGL_ONE_MINUS_DST_COLOR:
    case GGL_ONE_MINUS_SRC_COLOR:
    case GGL_ONE_MINUS_DST_ALPHA:
    case GGL_ONE_MINUS_SRC_ALPHA:
        RSB(AL, 0, factor.reg, factor.reg, imm((1<<factor.s)));
    }
    
    // don't need more than 8-bits for the blend factor
    // and this will prevent overflows in the multiplies later
    if (factor.s > 8) {
        MOV(AL, 0, factor.reg, reg_imm(factor.reg, LSR, factor.s-8));
        factor.s = 8;
    }
}

int GGLAssembler::blending_codes(int fs, int fd)
{
    int blending = 0;
    switch(fs) {
    case GGL_ONE:
        blending |= BLEND_SRC;
        break;

    case GGL_ONE_MINUS_DST_COLOR:
    case GGL_DST_COLOR:
        blending |= FACTOR_DST|BLEND_SRC;
        break;
    case GGL_ONE_MINUS_DST_ALPHA:
    case GGL_DST_ALPHA:
        // no need to extract 'component' from the destination
        // for the blend factor, because we need ALPHA only.
        blending |= BLEND_SRC;
        break;

    case GGL_ONE_MINUS_SRC_COLOR:
    case GGL_SRC_COLOR:    
        blending |= FACTOR_SRC|BLEND_SRC;
        break;
    case GGL_ONE_MINUS_SRC_ALPHA:
    case GGL_SRC_ALPHA:
    case GGL_SRC_ALPHA_SATURATE:
        blending |= FACTOR_SRC|BLEND_SRC;
        break;
    }
    switch(fd) {
    case GGL_ONE:
        blending |= BLEND_DST;
        break;

    case GGL_ONE_MINUS_DST_COLOR:
    case GGL_DST_COLOR:
        blending |= FACTOR_DST|BLEND_DST;
        break;
    case GGL_ONE_MINUS_DST_ALPHA:
    case GGL_DST_ALPHA:
        blending |= FACTOR_DST|BLEND_DST;
        break;

    case GGL_ONE_MINUS_SRC_COLOR:
    case GGL_SRC_COLOR:    
        blending |= FACTOR_SRC|BLEND_DST;
        break;
    case GGL_ONE_MINUS_SRC_ALPHA:
    case GGL_SRC_ALPHA:
        // no need to extract 'component' from the source
        // for the blend factor, because we need ALPHA only.
        blending |= BLEND_DST;
        break;
    }
    return blending;
}

// ---------------------------------------------------------------------------

void GGLAssembler::build_blendFOneMinusF(
        component_t& temp,
        const integer_t& factor, 
        const integer_t& fragment,
        const integer_t& fb)
{
    //  R = S*f + D*(1-f) = (S-D)*f + D
    Scratch scratches(registerFile());
    // compute S-D
    integer_t diff(fragment.flags & CORRUPTIBLE ?
            fragment.reg : scratches.obtain(), fb.size(), CORRUPTIBLE);
    const int shift = fragment.size() - fb.size();
    if (shift>0)        RSB(AL, 0, diff.reg, fb.reg, reg_imm(fragment.reg, LSR, shift));
    else if (shift<0)   RSB(AL, 0, diff.reg, fb.reg, reg_imm(fragment.reg, LSL,-shift));
    else                RSB(AL, 0, diff.reg, fb.reg, fragment.reg);
    mul_factor_add(temp, diff, factor, component_t(fb));
}

void GGLAssembler::build_blendOneMinusFF(
        component_t& temp,
        const integer_t& factor, 
        const integer_t& fragment,
        const integer_t& fb)
{
    //  R = S*f + D*(1-f) = (S-D)*f + D
    Scratch scratches(registerFile());
    // compute D-S
    integer_t diff(fb.flags & CORRUPTIBLE ?
            fb.reg : scratches.obtain(), fb.size(), CORRUPTIBLE);
    const int shift = fragment.size() - fb.size();
    if (shift>0)        SUB(AL, 0, diff.reg, fb.reg, reg_imm(fragment.reg, LSR, shift));
    else if (shift<0)   SUB(AL, 0, diff.reg, fb.reg, reg_imm(fragment.reg, LSL,-shift));
    else                SUB(AL, 0, diff.reg, fb.reg, fragment.reg);
    mul_factor_add(temp, diff, factor, component_t(fragment));
}

// ---------------------------------------------------------------------------

void GGLAssembler::mul_factor(  component_t& d,
                                const integer_t& v,
                                const integer_t& f)
{
    int vs = v.size();
    int fs = f.size();
    int ms = vs+fs;

    // XXX: we could have special cases for 1 bit mul

    // all this code below to use the best multiply instruction
    // wrt the parameters size. We take advantage of the fact
    // that the 16-bits multiplies allow a 16-bit shift
    // The trick is that we just make sure that we have at least 8-bits
    // per component (which is enough for a 8 bits display).

    int xy;
    int vshift = 0;
    int fshift = 0;
    int smulw = 0;

    if (vs<16) {
        if (fs<16) {
            xy = xyBB;
        } else if (GGL_BETWEEN(fs, 24, 31)) {
            ms -= 16;
            xy = xyTB;
        } else {
            // eg: 15 * 18  ->  15 * 15
            fshift = fs - 15;
            ms -= fshift;
            xy = xyBB;
        }
    } else if (GGL_BETWEEN(vs, 24, 31)) {
        if (fs<16) {
            ms -= 16;
            xy = xyTB;
        } else if (GGL_BETWEEN(fs, 24, 31)) {
            ms -= 32;
            xy = xyTT;
        } else {
            // eg: 24 * 18  ->  8 * 18
            fshift = fs - 15;
            ms -= 16 + fshift;
            xy = xyTB;
        }
    } else {
        if (fs<16) {
            // eg: 18 * 15  ->  15 * 15
            vshift = vs - 15;
            ms -= vshift;
            xy = xyBB;
        } else if (GGL_BETWEEN(fs, 24, 31)) {
            // eg: 18 * 24  ->  15 * 8
            vshift = vs - 15;
            ms -= 16 + vshift;
            xy = xyBT;
        } else {
            // eg: 18 * 18  ->  (15 * 18)>>16
            fshift = fs - 15;
            ms -= 16 + fshift;
            xy = yB;    //XXX SMULWB
            smulw = 1;
        }
    }

    ALOGE_IF(ms>=32, "mul_factor overflow vs=%d, fs=%d", vs, fs);

    int vreg = v.reg;
    int freg = f.reg;
    if (vshift) {
        MOV(AL, 0, d.reg, reg_imm(vreg, LSR, vshift));
        vreg = d.reg;
    }
    if (fshift) {
        MOV(AL, 0, d.reg, reg_imm(vreg, LSR, fshift));
        freg = d.reg;
    }
    if (smulw)  SMULW(AL, xy, d.reg, vreg, freg);
    else        SMUL(AL, xy, d.reg, vreg, freg);


    d.h = ms;
    if (mDithering) {
        d.l = 0; 
    } else {
        d.l = fs; 
        d.flags |= CLEAR_LO;
    }
}

void GGLAssembler::mul_factor_add(  component_t& d,
                                    const integer_t& v,
                                    const integer_t& f,
                                    const component_t& a)
{
    // XXX: we could have special cases for 1 bit mul
    Scratch scratches(registerFile());

    int vs = v.size();
    int fs = f.size();
    int as = a.h;
    int ms = vs+fs;

    ALOGE_IF(ms>=32, "mul_factor_add overflow vs=%d, fs=%d, as=%d", vs, fs, as);

    integer_t add(a.reg, a.h, a.flags);

    // 'a' is a component_t but it is guaranteed to have
    // its high bits set to 0. However in the dithering case,
    // we can't get away with truncating the potentially bad bits
    // so extraction is needed.

   if ((mDithering) && (a.size() < ms)) {
        // we need to expand a
        if (!(a.flags & CORRUPTIBLE)) {
            // ... but it's not corruptible, so we need to pick a
            // temporary register.
            // Try to uses the destination register first (it's likely
            // to be usable, unless it aliases an input).
            if (d.reg!=a.reg && d.reg!=v.reg && d.reg!=f.reg) {
                add.reg = d.reg;
            } else {
                add.reg = scratches.obtain();
            }
        }
        expand(add, a, ms); // extracts and expands
        as = ms;
    }

    if (ms == as) {
        if (vs<16 && fs<16) SMLABB(AL, d.reg, v.reg, f.reg, add.reg);
        else                MLA(AL, 0, d.reg, v.reg, f.reg, add.reg);
    } else {
        int temp = d.reg;
        if (temp == add.reg) {
            // the mul will modify add.reg, we need an intermediary reg
            if (v.flags & CORRUPTIBLE)      temp = v.reg;
            else if (f.flags & CORRUPTIBLE) temp = f.reg;
            else                            temp = scratches.obtain();
        }

        if (vs<16 && fs<16) SMULBB(AL, temp, v.reg, f.reg);
        else                MUL(AL, 0, temp, v.reg, f.reg);

        if (ms>as) {
            ADD(AL, 0, d.reg, temp, reg_imm(add.reg, LSL, ms-as));
        } else if (ms<as) {
            // not sure if we should expand the mul instead?
            ADD(AL, 0, d.reg, temp, reg_imm(add.reg, LSR, as-ms));
        }
    }

    d.h = ms;
    if (mDithering) {
        d.l = a.l; 
    } else {
        d.l = fs>a.l ? fs : a.l;
        d.flags |= CLEAR_LO;
    }
}

void GGLAssembler::component_add(component_t& d,
        const integer_t& dst, const integer_t& src)
{
    // here we're guaranteed that fragment.size() >= fb.size()
    const int shift = src.size() - dst.size();
    if (!shift) {
        ADD(AL, 0, d.reg, src.reg, dst.reg);
    } else {
        ADD(AL, 0, d.reg, src.reg, reg_imm(dst.reg, LSL, shift));
    }

    d.h = src.size();
    if (mDithering) {
        d.l = 0;
    } else {
        d.l = shift;
        d.flags |= CLEAR_LO;
    }
}

void GGLAssembler::component_sat(const component_t& v)
{
    const int one = ((1<<v.size())-1)<<v.l;
    CMP(AL, v.reg, imm( 1<<v.h ));
    if (isValidImmediate(one)) {
        MOV(HS, 0, v.reg, imm( one ));
    } else if (isValidImmediate(~one)) {
        MVN(HS, 0, v.reg, imm( ~one ));
    } else {
        MOV(HS, 0, v.reg, imm( 1<<v.h ));
        SUB(HS, 0, v.reg, v.reg, imm( 1<<v.l ));
    }
}

// ----------------------------------------------------------------------------

}; // namespace android