summaryrefslogtreecommitdiff
path: root/src/gallium/drivers/llvmpipe/lp_bld_arit.c
blob: 83ca06acf86541398b0aa357656b297bc7f91f53 (plain)
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
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
/**************************************************************************
 *
 * Copyright 2009 VMware, Inc.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the
 * "Software"), to deal in the Software without restriction, including
 * without limitation the rights to use, copy, modify, merge, publish,
 * distribute, sub license, and/or sell copies of the Software, and to
 * permit persons to whom the Software is furnished to do so, subject to
 * the following conditions:
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 **************************************************************************/


/**
 * @file
 * Helper
 *
 * LLVM IR doesn't support all basic arithmetic operations we care about (most
 * notably min/max and saturated operations), and it is often necessary to
 * resort machine-specific intrinsics directly. The functions here hide all
 * these implementation details from the other modules.
 *
 * We also do simple expressions simplification here. Reasons are:
 * - it is very easy given we have all necessary information readily available
 * - LLVM optimization passes fail to simplify several vector expressions
 * - We often know value constraints which the optimization passes have no way
 *   of knowing, such as when source arguments are known to be in [0, 1] range.
 *
 * @author Jose Fonseca <jfonseca@vmware.com>
 */


#include "util/u_memory.h"
#include "util/u_debug.h"
#include "util/u_string.h"
#include "util/u_cpu_detect.h"

#include "lp_bld_type.h"
#include "lp_bld_const.h"
#include "lp_bld_intr.h"
#include "lp_bld_logic.h"
#include "lp_bld_pack.h"
#include "lp_bld_debug.h"
#include "lp_bld_arit.h"


/**
 * Generate min(a, b)
 * No checks for special case values of a or b = 1 or 0 are done.
 */
static LLVMValueRef
lp_build_min_simple(struct lp_build_context *bld,
                    LLVMValueRef a,
                    LLVMValueRef b)
{
   const struct lp_type type = bld->type;
   const char *intrinsic = NULL;
   LLVMValueRef cond;

   /* TODO: optimize the constant case */

   if(type.width * type.length == 128) {
      if(type.floating) {
         if(type.width == 32 && util_cpu_caps.has_sse)
            intrinsic = "llvm.x86.sse.min.ps";
         if(type.width == 64 && util_cpu_caps.has_sse2)
            intrinsic = "llvm.x86.sse2.min.pd";
      }
      else {
         if(type.width == 8 && !type.sign && util_cpu_caps.has_sse2)
            intrinsic = "llvm.x86.sse2.pminu.b";
         if(type.width == 8 && type.sign && util_cpu_caps.has_sse4_1)
            intrinsic = "llvm.x86.sse41.pminsb";
         if(type.width == 16 && !type.sign && util_cpu_caps.has_sse4_1)
            intrinsic = "llvm.x86.sse41.pminuw";
         if(type.width == 16 && type.sign && util_cpu_caps.has_sse2)
            intrinsic = "llvm.x86.sse2.pmins.w";
         if(type.width == 32 && !type.sign && util_cpu_caps.has_sse4_1)
            intrinsic = "llvm.x86.sse41.pminud";
         if(type.width == 32 && type.sign && util_cpu_caps.has_sse4_1)
            intrinsic = "llvm.x86.sse41.pminsd";
      }
   }

   if(intrinsic)
      return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);

   cond = lp_build_cmp(bld, PIPE_FUNC_LESS, a, b);
   return lp_build_select(bld, cond, a, b);
}


/**
 * Generate max(a, b)
 * No checks for special case values of a or b = 1 or 0 are done.
 */
static LLVMValueRef
lp_build_max_simple(struct lp_build_context *bld,
                    LLVMValueRef a,
                    LLVMValueRef b)
{
   const struct lp_type type = bld->type;
   const char *intrinsic = NULL;
   LLVMValueRef cond;

   /* TODO: optimize the constant case */

   if(type.width * type.length == 128) {
      if(type.floating) {
         if(type.width == 32 && util_cpu_caps.has_sse)
            intrinsic = "llvm.x86.sse.max.ps";
         if(type.width == 64 && util_cpu_caps.has_sse2)
            intrinsic = "llvm.x86.sse2.max.pd";
      }
      else {
         if(type.width == 8 && !type.sign && util_cpu_caps.has_sse2)
            intrinsic = "llvm.x86.sse2.pmaxu.b";
         if(type.width == 8 && type.sign && util_cpu_caps.has_sse4_1)
            intrinsic = "llvm.x86.sse41.pmaxsb";
         if(type.width == 16 && !type.sign && util_cpu_caps.has_sse4_1)
            intrinsic = "llvm.x86.sse41.pmaxuw";
         if(type.width == 16 && type.sign && util_cpu_caps.has_sse2)
            intrinsic = "llvm.x86.sse2.pmaxs.w";
         if(type.width == 32 && !type.sign && util_cpu_caps.has_sse4_1)
            intrinsic = "llvm.x86.sse41.pmaxud";
         if(type.width == 32 && type.sign && util_cpu_caps.has_sse4_1)
            intrinsic = "llvm.x86.sse41.pmaxsd";
      }
   }

   if(intrinsic)
      return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);

   cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, b);
   return lp_build_select(bld, cond, a, b);
}


/**
 * Generate 1 - a, or ~a depending on bld->type.
 */
LLVMValueRef
lp_build_comp(struct lp_build_context *bld,
              LLVMValueRef a)
{
   const struct lp_type type = bld->type;

   if(a == bld->one)
      return bld->zero;
   if(a == bld->zero)
      return bld->one;

   if(type.norm && !type.floating && !type.fixed && !type.sign) {
      if(LLVMIsConstant(a))
         return LLVMConstNot(a);
      else
         return LLVMBuildNot(bld->builder, a, "");
   }

   if(LLVMIsConstant(a))
      return LLVMConstSub(bld->one, a);
   else
      return LLVMBuildSub(bld->builder, bld->one, a, "");
}


/**
 * Generate a + b
 */
LLVMValueRef
lp_build_add(struct lp_build_context *bld,
             LLVMValueRef a,
             LLVMValueRef b)
{
   const struct lp_type type = bld->type;
   LLVMValueRef res;

   if(a == bld->zero)
      return b;
   if(b == bld->zero)
      return a;
   if(a == bld->undef || b == bld->undef)
      return bld->undef;

   if(bld->type.norm) {
      const char *intrinsic = NULL;

      if(a == bld->one || b == bld->one)
        return bld->one;

      if(util_cpu_caps.has_sse2 &&
         type.width * type.length == 128 &&
         !type.floating && !type.fixed) {
         if(type.width == 8)
            intrinsic = type.sign ? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
         if(type.width == 16)
            intrinsic = type.sign ? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
      }
   
      if(intrinsic)
         return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
   }

   if(LLVMIsConstant(a) && LLVMIsConstant(b))
      res = LLVMConstAdd(a, b);
   else
      res = LLVMBuildAdd(bld->builder, a, b, "");

   /* clamp to ceiling of 1.0 */
   if(bld->type.norm && (bld->type.floating || bld->type.fixed))
      res = lp_build_min_simple(bld, res, bld->one);

   /* XXX clamp to floor of -1 or 0??? */

   return res;
}


/**
 * Generate a - b
 */
LLVMValueRef
lp_build_sub(struct lp_build_context *bld,
             LLVMValueRef a,
             LLVMValueRef b)
{
   const struct lp_type type = bld->type;
   LLVMValueRef res;

   if(b == bld->zero)
      return a;
   if(a == bld->undef || b == bld->undef)
      return bld->undef;
   if(a == b)
      return bld->zero;

   if(bld->type.norm) {
      const char *intrinsic = NULL;

      if(b == bld->one)
        return bld->zero;

      if(util_cpu_caps.has_sse2 &&
         type.width * type.length == 128 &&
         !type.floating && !type.fixed) {
         if(type.width == 8)
            intrinsic = type.sign ? "llvm.x86.sse2.psubs.b" : "llvm.x86.sse2.psubus.b";
         if(type.width == 16)
            intrinsic = type.sign ? "llvm.x86.sse2.psubs.w" : "llvm.x86.sse2.psubus.w";
      }
   
      if(intrinsic)
         return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
   }

   if(LLVMIsConstant(a) && LLVMIsConstant(b))
      res = LLVMConstSub(a, b);
   else
      res = LLVMBuildSub(bld->builder, a, b, "");

   if(bld->type.norm && (bld->type.floating || bld->type.fixed))
      res = lp_build_max_simple(bld, res, bld->zero);

   return res;
}


/**
 * Normalized 8bit multiplication.
 *
 * - alpha plus one
 *
 *     makes the following approximation to the division (Sree)
 *    
 *       a*b/255 ~= (a*(b + 1)) >> 256
 *    
 *     which is the fastest method that satisfies the following OpenGL criteria
 *    
 *       0*0 = 0 and 255*255 = 255
 *
 * - geometric series
 *
 *     takes the geometric series approximation to the division
 *
 *       t/255 = (t >> 8) + (t >> 16) + (t >> 24) ..
 *
 *     in this case just the first two terms to fit in 16bit arithmetic
 *
 *       t/255 ~= (t + (t >> 8)) >> 8
 *
 *     note that just by itself it doesn't satisfies the OpenGL criteria, as
 *     255*255 = 254, so the special case b = 255 must be accounted or roundoff
 *     must be used
 *
 * - geometric series plus rounding
 *
 *     when using a geometric series division instead of truncating the result
 *     use roundoff in the approximation (Jim Blinn)
 *
 *       t/255 ~= (t + (t >> 8) + 0x80) >> 8
 *
 *     achieving the exact results
 *
 * @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995, 
 *     ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
 * @sa Michael Herf, The "double blend trick", May 2000, 
 *     http://www.stereopsis.com/doubleblend.html
 */
static LLVMValueRef
lp_build_mul_u8n(LLVMBuilderRef builder,
                 struct lp_type i16_type,
                 LLVMValueRef a, LLVMValueRef b)
{
   LLVMValueRef c8;
   LLVMValueRef ab;

   c8 = lp_build_int_const_scalar(i16_type, 8);
   
#if 0
   
   /* a*b/255 ~= (a*(b + 1)) >> 256 */
   b = LLVMBuildAdd(builder, b, lp_build_int_const_scalar(i16_type, 1), "");
   ab = LLVMBuildMul(builder, a, b, "");

#else
   
   /* ab/255 ~= (ab + (ab >> 8) + 0x80) >> 8 */
   ab = LLVMBuildMul(builder, a, b, "");
   ab = LLVMBuildAdd(builder, ab, LLVMBuildLShr(builder, ab, c8, ""), "");
   ab = LLVMBuildAdd(builder, ab, lp_build_int_const_scalar(i16_type, 0x80), "");

#endif
   
   ab = LLVMBuildLShr(builder, ab, c8, "");

   return ab;
}


/**
 * Generate a * b
 */
LLVMValueRef
lp_build_mul(struct lp_build_context *bld,
             LLVMValueRef a,
             LLVMValueRef b)
{
   const struct lp_type type = bld->type;

   if(a == bld->zero)
      return bld->zero;
   if(a == bld->one)
      return b;
   if(b == bld->zero)
      return bld->zero;
   if(b == bld->one)
      return a;
   if(a == bld->undef || b == bld->undef)
      return bld->undef;

   if(!type.floating && !type.fixed && type.norm) {
      if(type.width == 8) {
         struct lp_type i16_type = lp_wider_type(type);
         LLVMValueRef al, ah, bl, bh, abl, abh, ab;

         lp_build_unpack2(bld->builder, type, i16_type, a, &al, &ah);
         lp_build_unpack2(bld->builder, type, i16_type, b, &bl, &bh);

         /* PMULLW, PSRLW, PADDW */
         abl = lp_build_mul_u8n(bld->builder, i16_type, al, bl);
         abh = lp_build_mul_u8n(bld->builder, i16_type, ah, bh);

         ab = lp_build_pack2(bld->builder, i16_type, type, abl, abh);
         
         return ab;
      }

      /* FIXME */
      assert(0);
   }

   if(LLVMIsConstant(a) && LLVMIsConstant(b))
      return LLVMConstMul(a, b);

   return LLVMBuildMul(bld->builder, a, b, "");
}


/**
 * Generate a / b
 */
LLVMValueRef
lp_build_div(struct lp_build_context *bld,
             LLVMValueRef a,
             LLVMValueRef b)
{
   const struct lp_type type = bld->type;

   if(a == bld->zero)
      return bld->zero;
   if(a == bld->one)
      return lp_build_rcp(bld, b);
   if(b == bld->zero)
      return bld->undef;
   if(b == bld->one)
      return a;
   if(a == bld->undef || b == bld->undef)
      return bld->undef;

   if(LLVMIsConstant(a) && LLVMIsConstant(b))
      return LLVMConstFDiv(a, b);

   if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4)
      return lp_build_mul(bld, a, lp_build_rcp(bld, b));

   return LLVMBuildFDiv(bld->builder, a, b, "");
}


LLVMValueRef
lp_build_lerp(struct lp_build_context *bld,
              LLVMValueRef x,
              LLVMValueRef v0,
              LLVMValueRef v1)
{
   return lp_build_add(bld, v0, lp_build_mul(bld, x, lp_build_sub(bld, v1, v0)));
}


LLVMValueRef
lp_build_lerp_2d(struct lp_build_context *bld,
                 LLVMValueRef x,
                 LLVMValueRef y,
                 LLVMValueRef v00,
                 LLVMValueRef v01,
                 LLVMValueRef v10,
                 LLVMValueRef v11)
{
   LLVMValueRef v0 = lp_build_lerp(bld, x, v00, v01);
   LLVMValueRef v1 = lp_build_lerp(bld, x, v10, v11);
   return lp_build_lerp(bld, y, v0, v1);
}


/**
 * Generate min(a, b)
 * Do checks for special cases.
 */
LLVMValueRef
lp_build_min(struct lp_build_context *bld,
             LLVMValueRef a,
             LLVMValueRef b)
{
   if(a == bld->undef || b == bld->undef)
      return bld->undef;

   if(a == b)
      return a;

   if(bld->type.norm) {
      if(a == bld->zero || b == bld->zero)
         return bld->zero;
      if(a == bld->one)
         return b;
      if(b == bld->one)
         return a;
   }

   return lp_build_min_simple(bld, a, b);
}


/**
 * Generate max(a, b)
 * Do checks for special cases.
 */
LLVMValueRef
lp_build_max(struct lp_build_context *bld,
             LLVMValueRef a,
             LLVMValueRef b)
{
   if(a == bld->undef || b == bld->undef)
      return bld->undef;

   if(a == b)
      return a;

   if(bld->type.norm) {
      if(a == bld->one || b == bld->one)
         return bld->one;
      if(a == bld->zero)
         return b;
      if(b == bld->zero)
         return a;
   }

   return lp_build_max_simple(bld, a, b);
}


/**
 * Generate abs(a)
 */
LLVMValueRef
lp_build_abs(struct lp_build_context *bld,
             LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef vec_type = lp_build_vec_type(type);

   if(!type.sign)
      return a;

   if(type.floating) {
      /* Mask out the sign bit */
      LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
      LLVMValueRef mask = lp_build_int_const_scalar(type, ((unsigned long long)1 << type.width) - 1);
      a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
      a = LLVMBuildAnd(bld->builder, a, mask, "");
      a = LLVMBuildBitCast(bld->builder, a, vec_type, "");
      return a;
   }

   if(type.width*type.length == 128 && util_cpu_caps.has_ssse3) {
      switch(type.width) {
      case 8:
         return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.b.128", vec_type, a);
      case 16:
         return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.w.128", vec_type, a);
      case 32:
         return lp_build_intrinsic_unary(bld->builder, "llvm.x86.ssse3.pabs.d.128", vec_type, a);
      }
   }

   return lp_build_max(bld, a, LLVMBuildNeg(bld->builder, a, ""));
}


LLVMValueRef
lp_build_sgn(struct lp_build_context *bld,
             LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef vec_type = lp_build_vec_type(type);
   LLVMValueRef cond;
   LLVMValueRef res;

   /* Handle non-zero case */
   if(!type.sign) {
      /* if not zero then sign must be positive */
      res = bld->one;
   }
   else if(type.floating) {
      /* Take the sign bit and add it to 1 constant */
      LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
      LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
      LLVMValueRef sign;
      LLVMValueRef one;
      sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
      sign = LLVMBuildAnd(bld->builder, sign, mask, "");
      one = LLVMConstBitCast(bld->one, int_vec_type);
      res = LLVMBuildOr(bld->builder, sign, one, "");
      res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
   }
   else
   {
      LLVMValueRef minus_one = lp_build_const_scalar(type, -1.0);
      cond = lp_build_cmp(bld, PIPE_FUNC_GREATER, a, bld->zero);
      res = lp_build_select(bld, cond, bld->one, minus_one);
   }

   /* Handle zero */
   cond = lp_build_cmp(bld, PIPE_FUNC_EQUAL, a, bld->zero);
   res = lp_build_select(bld, cond, bld->zero, bld->one);

   return res;
}


enum lp_build_round_sse41_mode
{
   LP_BUILD_ROUND_SSE41_NEAREST = 0,
   LP_BUILD_ROUND_SSE41_FLOOR = 1,
   LP_BUILD_ROUND_SSE41_CEIL = 2,
   LP_BUILD_ROUND_SSE41_TRUNCATE = 3
};


static INLINE LLVMValueRef
lp_build_round_sse41(struct lp_build_context *bld,
                     LLVMValueRef a,
                     enum lp_build_round_sse41_mode mode)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef vec_type = lp_build_vec_type(type);
   const char *intrinsic;

   assert(type.floating);
   assert(type.width*type.length == 128);
   assert(lp_check_value(type, a));
   assert(util_cpu_caps.has_sse4_1);

   switch(type.width) {
   case 32:
      intrinsic = "llvm.x86.sse41.round.ps";
      break;
   case 64:
      intrinsic = "llvm.x86.sse41.round.pd";
      break;
   default:
      assert(0);
      return bld->undef;
   }

   return lp_build_intrinsic_binary(bld->builder, intrinsic, vec_type, a,
                                    LLVMConstInt(LLVMInt32Type(), mode, 0));
}


LLVMValueRef
lp_build_trunc(struct lp_build_context *bld,
               LLVMValueRef a)
{
   const struct lp_type type = bld->type;

   assert(type.floating);
   assert(lp_check_value(type, a));

   if(util_cpu_caps.has_sse4_1)
      return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_TRUNCATE);
   else {
      LLVMTypeRef vec_type = lp_build_vec_type(type);
      LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
      LLVMValueRef res;
      res = LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
      res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
      return res;
   }
}


LLVMValueRef
lp_build_round(struct lp_build_context *bld,
               LLVMValueRef a)
{
   const struct lp_type type = bld->type;

   assert(type.floating);
   assert(lp_check_value(type, a));

   if(util_cpu_caps.has_sse4_1)
      return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
   else {
      LLVMTypeRef vec_type = lp_build_vec_type(type);
      LLVMValueRef res;
      res = lp_build_iround(bld, a);
      res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
      return res;
   }
}


LLVMValueRef
lp_build_floor(struct lp_build_context *bld,
               LLVMValueRef a)
{
   const struct lp_type type = bld->type;

   assert(type.floating);

   if(util_cpu_caps.has_sse4_1)
      return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
   else {
      LLVMTypeRef vec_type = lp_build_vec_type(type);
      LLVMValueRef res;
      res = lp_build_ifloor(bld, a);
      res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
      return res;
   }
}


LLVMValueRef
lp_build_ceil(struct lp_build_context *bld,
              LLVMValueRef a)
{
   const struct lp_type type = bld->type;

   assert(type.floating);
   assert(lp_check_value(type, a));

   if(util_cpu_caps.has_sse4_1)
      return lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
   else {
      LLVMTypeRef vec_type = lp_build_vec_type(type);
      LLVMValueRef res;
      res = lp_build_iceil(bld, a);
      res = LLVMBuildSIToFP(bld->builder, res, vec_type, "");
      return res;
   }
}


/**
 * Convert to integer, through whichever rounding method that's fastest,
 * typically truncating to zero.
 */
LLVMValueRef
lp_build_itrunc(struct lp_build_context *bld,
                LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);

   assert(type.floating);
   assert(lp_check_value(type, a));

   return LLVMBuildFPToSI(bld->builder, a, int_vec_type, "");
}


LLVMValueRef
lp_build_iround(struct lp_build_context *bld,
                LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
   LLVMValueRef res;

   assert(type.floating);
   assert(lp_check_value(type, a));

   if(util_cpu_caps.has_sse4_1) {
      res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_NEAREST);
   }
   else {
      LLVMTypeRef vec_type = lp_build_vec_type(type);
      LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
      LLVMValueRef sign;
      LLVMValueRef half;

      /* get sign bit */
      sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
      sign = LLVMBuildAnd(bld->builder, sign, mask, "");

      /* sign * 0.5 */
      half = lp_build_const_scalar(type, 0.5);
      half = LLVMBuildBitCast(bld->builder, half, int_vec_type, "");
      half = LLVMBuildOr(bld->builder, sign, half, "");
      half = LLVMBuildBitCast(bld->builder, half, vec_type, "");

      res = LLVMBuildAdd(bld->builder, a, half, "");
   }

   res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");

   return res;
}


LLVMValueRef
lp_build_ifloor(struct lp_build_context *bld,
                LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
   LLVMValueRef res;

   assert(type.floating);
   assert(lp_check_value(type, a));

   if(util_cpu_caps.has_sse4_1) {
      res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
   }
   else {
      /* Take the sign bit and add it to 1 constant */
      LLVMTypeRef vec_type = lp_build_vec_type(type);
      unsigned mantissa = lp_mantissa(type);
      LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
      LLVMValueRef sign;
      LLVMValueRef offset;

      /* sign = a < 0 ? ~0 : 0 */
      sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
      sign = LLVMBuildAnd(bld->builder, sign, mask, "");
      sign = LLVMBuildAShr(bld->builder, sign, lp_build_int_const_scalar(type, type.width - 1), "");

      /* offset = -0.99999(9)f */
      offset = lp_build_const_scalar(type, -(double)(((unsigned long long)1 << mantissa) - 1)/((unsigned long long)1 << mantissa));
      offset = LLVMConstBitCast(offset, int_vec_type);

      /* offset = a < 0 ? -0.99999(9)f : 0.0f */
      offset = LLVMBuildAnd(bld->builder, offset, sign, "");
      offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "");

      res = LLVMBuildAdd(bld->builder, a, offset, "");
   }

   res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");

   return res;
}


LLVMValueRef
lp_build_iceil(struct lp_build_context *bld,
               LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
   LLVMValueRef res;

   assert(type.floating);
   assert(lp_check_value(type, a));

   if(util_cpu_caps.has_sse4_1) {
      res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_CEIL);
   }
   else {
      assert(0);
      res = bld->undef;
   }

   res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");

   return res;
}


LLVMValueRef
lp_build_sqrt(struct lp_build_context *bld,
              LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef vec_type = lp_build_vec_type(type);
   char intrinsic[32];

   /* TODO: optimize the constant case */
   /* TODO: optimize the constant case */

   assert(type.floating);
   util_snprintf(intrinsic, sizeof intrinsic, "llvm.sqrt.v%uf%u", type.length, type.width);

   return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
}


LLVMValueRef
lp_build_rcp(struct lp_build_context *bld,
             LLVMValueRef a)
{
   const struct lp_type type = bld->type;

   if(a == bld->zero)
      return bld->undef;
   if(a == bld->one)
      return bld->one;
   if(a == bld->undef)
      return bld->undef;

   assert(type.floating);

   if(LLVMIsConstant(a))
      return LLVMConstFDiv(bld->one, a);

   if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4)
      /* FIXME: improve precision */
      return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type), a);

   return LLVMBuildFDiv(bld->builder, bld->one, a, "");
}


/**
 * Generate 1/sqrt(a)
 */
LLVMValueRef
lp_build_rsqrt(struct lp_build_context *bld,
               LLVMValueRef a)
{
   const struct lp_type type = bld->type;

   assert(type.floating);

   if(util_cpu_caps.has_sse && type.width == 32 && type.length == 4)
      return lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rsqrt.ps", lp_build_vec_type(type), a);

   return lp_build_rcp(bld, lp_build_sqrt(bld, a));
}


/**
 * Generate cos(a)
 */
LLVMValueRef
lp_build_cos(struct lp_build_context *bld,
              LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef vec_type = lp_build_vec_type(type);
   char intrinsic[32];

   /* TODO: optimize the constant case */

   assert(type.floating);
   util_snprintf(intrinsic, sizeof intrinsic, "llvm.cos.v%uf%u", type.length, type.width);

   return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
}


/**
 * Generate sin(a)
 */
LLVMValueRef
lp_build_sin(struct lp_build_context *bld,
              LLVMValueRef a)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef vec_type = lp_build_vec_type(type);
   char intrinsic[32];

   /* TODO: optimize the constant case */

   assert(type.floating);
   util_snprintf(intrinsic, sizeof intrinsic, "llvm.sin.v%uf%u", type.length, type.width);

   return lp_build_intrinsic_unary(bld->builder, intrinsic, vec_type, a);
}


/**
 * Generate pow(x, y)
 */
LLVMValueRef
lp_build_pow(struct lp_build_context *bld,
             LLVMValueRef x,
             LLVMValueRef y)
{
   /* TODO: optimize the constant case */
   if(LLVMIsConstant(x) && LLVMIsConstant(y))
      debug_printf("%s: inefficient/imprecise constant arithmetic\n",
                   __FUNCTION__);

   return lp_build_exp2(bld, lp_build_mul(bld, lp_build_log2(bld, x), y));
}


/**
 * Generate exp(x)
 */
LLVMValueRef
lp_build_exp(struct lp_build_context *bld,
             LLVMValueRef x)
{
   /* log2(e) = 1/log(2) */
   LLVMValueRef log2e = lp_build_const_scalar(bld->type, 1.4426950408889634);

   return lp_build_mul(bld, log2e, lp_build_exp2(bld, x));
}


/**
 * Generate log(x)
 */
LLVMValueRef
lp_build_log(struct lp_build_context *bld,
             LLVMValueRef x)
{
   /* log(2) */
   LLVMValueRef log2 = lp_build_const_scalar(bld->type, 1.4426950408889634);

   return lp_build_mul(bld, log2, lp_build_exp2(bld, x));
}


#define EXP_POLY_DEGREE 3
#define LOG_POLY_DEGREE 5


/**
 * Generate polynomial.
 * Ex:  x^2 * coeffs[0] + x * coeffs[1] + coeffs[2].
 */
static LLVMValueRef
lp_build_polynomial(struct lp_build_context *bld,
                    LLVMValueRef x,
                    const double *coeffs,
                    unsigned num_coeffs)
{
   const struct lp_type type = bld->type;
   LLVMValueRef res = NULL;
   unsigned i;

   /* TODO: optimize the constant case */
   if(LLVMIsConstant(x))
      debug_printf("%s: inefficient/imprecise constant arithmetic\n",
                   __FUNCTION__);

   for (i = num_coeffs; i--; ) {
      LLVMValueRef coeff = lp_build_const_scalar(type, coeffs[i]);
      if(res)
         res = lp_build_add(bld, coeff, lp_build_mul(bld, x, res));
      else
         res = coeff;
   }

   if(res)
      return res;
   else
      return bld->undef;
}


/**
 * Minimax polynomial fit of 2**x, in range [-0.5, 0.5[
 */
const double lp_build_exp2_polynomial[] = {
#if EXP_POLY_DEGREE == 5
   9.9999994e-1, 6.9315308e-1, 2.4015361e-1, 5.5826318e-2, 8.9893397e-3, 1.8775767e-3
#elif EXP_POLY_DEGREE == 4
   1.0000026, 6.9300383e-1, 2.4144275e-1, 5.2011464e-2, 1.3534167e-2
#elif EXP_POLY_DEGREE == 3
   9.9992520e-1, 6.9583356e-1, 2.2606716e-1, 7.8024521e-2
#elif EXP_POLY_DEGREE == 2
   1.0017247, 6.5763628e-1, 3.3718944e-1
#else
#error
#endif
};


void
lp_build_exp2_approx(struct lp_build_context *bld,
                     LLVMValueRef x,
                     LLVMValueRef *p_exp2_int_part,
                     LLVMValueRef *p_frac_part,
                     LLVMValueRef *p_exp2)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef vec_type = lp_build_vec_type(type);
   LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
   LLVMValueRef ipart = NULL;
   LLVMValueRef fpart = NULL;
   LLVMValueRef expipart = NULL;
   LLVMValueRef expfpart = NULL;
   LLVMValueRef res = NULL;

   if(p_exp2_int_part || p_frac_part || p_exp2) {
      /* TODO: optimize the constant case */
      if(LLVMIsConstant(x))
         debug_printf("%s: inefficient/imprecise constant arithmetic\n",
                      __FUNCTION__);

      assert(type.floating && type.width == 32);

      x = lp_build_min(bld, x, lp_build_const_scalar(type,  129.0));
      x = lp_build_max(bld, x, lp_build_const_scalar(type, -126.99999));

      /* ipart = int(x - 0.5) */
      ipart = LLVMBuildSub(bld->builder, x, lp_build_const_scalar(type, 0.5f), "");
      ipart = LLVMBuildFPToSI(bld->builder, ipart, int_vec_type, "");

      /* fpart = x - ipart */
      fpart = LLVMBuildSIToFP(bld->builder, ipart, vec_type, "");
      fpart = LLVMBuildSub(bld->builder, x, fpart, "");
   }

   if(p_exp2_int_part || p_exp2) {
      /* expipart = (float) (1 << ipart) */
      expipart = LLVMBuildAdd(bld->builder, ipart, lp_build_int_const_scalar(type, 127), "");
      expipart = LLVMBuildShl(bld->builder, expipart, lp_build_int_const_scalar(type, 23), "");
      expipart = LLVMBuildBitCast(bld->builder, expipart, vec_type, "");
   }

   if(p_exp2) {
      expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
                                     Elements(lp_build_exp2_polynomial));

      res = LLVMBuildMul(bld->builder, expipart, expfpart, "");
   }

   if(p_exp2_int_part)
      *p_exp2_int_part = expipart;

   if(p_frac_part)
      *p_frac_part = fpart;

   if(p_exp2)
      *p_exp2 = res;
}


LLVMValueRef
lp_build_exp2(struct lp_build_context *bld,
              LLVMValueRef x)
{
   LLVMValueRef res;
   lp_build_exp2_approx(bld, x, NULL, NULL, &res);
   return res;
}


/**
 * Minimax polynomial fit of log2(x)/(x - 1), for x in range [1, 2[
 * These coefficients can be generate with
 * http://www.boost.org/doc/libs/1_36_0/libs/math/doc/sf_and_dist/html/math_toolkit/toolkit/internals2/minimax.html
 */
const double lp_build_log2_polynomial[] = {
#if LOG_POLY_DEGREE == 6
   3.11578814719469302614, -3.32419399085241980044, 2.59883907202499966007, -1.23152682416275988241, 0.318212422185251071475, -0.0344359067839062357313
#elif LOG_POLY_DEGREE == 5
   2.8882704548164776201, -2.52074962577807006663, 1.48116647521213171641, -0.465725644288844778798, 0.0596515482674574969533
#elif LOG_POLY_DEGREE == 4
   2.61761038894603480148, -1.75647175389045657003, 0.688243882994381274313, -0.107254423828329604454
#elif LOG_POLY_DEGREE == 3
   2.28330284476918490682, -1.04913055217340124191, 0.204446009836232697516
#else
#error
#endif
};


/**
 * See http://www.devmaster.net/forums/showthread.php?p=43580
 */
void
lp_build_log2_approx(struct lp_build_context *bld,
                     LLVMValueRef x,
                     LLVMValueRef *p_exp,
                     LLVMValueRef *p_floor_log2,
                     LLVMValueRef *p_log2)
{
   const struct lp_type type = bld->type;
   LLVMTypeRef vec_type = lp_build_vec_type(type);
   LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);

   LLVMValueRef expmask = lp_build_int_const_scalar(type, 0x7f800000);
   LLVMValueRef mantmask = lp_build_int_const_scalar(type, 0x007fffff);
   LLVMValueRef one = LLVMConstBitCast(bld->one, int_vec_type);

   LLVMValueRef i = NULL;
   LLVMValueRef exp = NULL;
   LLVMValueRef mant = NULL;
   LLVMValueRef logexp = NULL;
   LLVMValueRef logmant = NULL;
   LLVMValueRef res = NULL;

   if(p_exp || p_floor_log2 || p_log2) {
      /* TODO: optimize the constant case */
      if(LLVMIsConstant(x))
         debug_printf("%s: inefficient/imprecise constant arithmetic\n",
                      __FUNCTION__);

      assert(type.floating && type.width == 32);

      i = LLVMBuildBitCast(bld->builder, x, int_vec_type, "");

      /* exp = (float) exponent(x) */
      exp = LLVMBuildAnd(bld->builder, i, expmask, "");
   }

   if(p_floor_log2 || p_log2) {
      logexp = LLVMBuildLShr(bld->builder, exp, lp_build_int_const_scalar(type, 23), "");
      logexp = LLVMBuildSub(bld->builder, logexp, lp_build_int_const_scalar(type, 127), "");
      logexp = LLVMBuildSIToFP(bld->builder, logexp, vec_type, "");
   }

   if(p_log2) {
      /* mant = (float) mantissa(x) */
      mant = LLVMBuildAnd(bld->builder, i, mantmask, "");
      mant = LLVMBuildOr(bld->builder, mant, one, "");
      mant = LLVMBuildSIToFP(bld->builder, mant, vec_type, "");

      logmant = lp_build_polynomial(bld, mant, lp_build_log2_polynomial,
                                    Elements(lp_build_log2_polynomial));

      /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
      logmant = LLVMBuildMul(bld->builder, logmant, LLVMBuildMul(bld->builder, mant, bld->one, ""), "");

      res = LLVMBuildAdd(bld->builder, logmant, logexp, "");
   }

   if(p_exp)
      *p_exp = exp;

   if(p_floor_log2)
      *p_floor_log2 = logexp;

   if(p_log2)
      *p_log2 = res;
}


LLVMValueRef
lp_build_log2(struct lp_build_context *bld,
              LLVMValueRef x)
{
   LLVMValueRef res;
   lp_build_log2_approx(bld, x, NULL, NULL, &res);
   return res;
}