-
-
Notifications
You must be signed in to change notification settings - Fork 575
/
llvm_backend_expr.cpp
5464 lines (4622 loc) · 171 KB
/
llvm_backend_expr.cpp
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
gb_internal lbValue lb_emit_arith_matrix(lbProcedure *p, TokenKind op, lbValue lhs, lbValue rhs, Type *type, bool component_wise);
gb_internal lbValue lb_emit_logical_binary_expr(lbProcedure *p, TokenKind op, Ast *left, Ast *right, Type *final_type) {
lbModule *m = p->module;
lbBlock *rhs = lb_create_block(p, "logical.cmp.rhs");
lbBlock *done = lb_create_block(p, "logical.cmp.done");
lbValue short_circuit = {};
if (op == Token_CmpAnd) {
lb_build_cond(p, left, rhs, done);
short_circuit = lb_const_bool(m, t_llvm_bool, false);
} else if (op == Token_CmpOr) {
lb_build_cond(p, left, done, rhs);
short_circuit = lb_const_bool(m, t_llvm_bool, true);
}
if (rhs->preds.count == 0) {
lb_start_block(p, done);
return short_circuit;
}
if (done->preds.count == 0) {
lb_start_block(p, rhs);
if (lb_is_expr_untyped_const(right)) {
return lb_expr_untyped_const_to_typed(m, right, default_type(final_type));
}
return lb_build_expr(p, right);
}
Array<LLVMValueRef> incoming_values = {};
Array<LLVMBasicBlockRef> incoming_blocks = {};
array_init(&incoming_values, heap_allocator(), done->preds.count+1);
array_init(&incoming_blocks, heap_allocator(), done->preds.count+1);
for_array(i, done->preds) {
incoming_values[i] = short_circuit.value;
incoming_blocks[i] = done->preds[i]->block;
}
lb_start_block(p, rhs);
lbValue edge = {};
if (lb_is_expr_untyped_const(right)) {
edge = lb_expr_untyped_const_to_typed(m, right, t_llvm_bool);
} else {
edge = lb_emit_conv(p, lb_build_expr(p, right), t_llvm_bool);
}
GB_ASSERT(edge.type == t_llvm_bool);
incoming_values[done->preds.count] = edge.value;
incoming_blocks[done->preds.count] = p->curr_block->block;
lb_emit_jump(p, done);
lb_start_block(p, done);
LLVMTypeRef dst_type = lb_type(m, t_llvm_bool);
LLVMValueRef phi = nullptr;
GB_ASSERT(incoming_values.count == incoming_blocks.count);
GB_ASSERT(incoming_values.count > 0);
LLVMTypeRef phi_type = nullptr;
for (LLVMValueRef incoming_value : incoming_values) {
if (!LLVMIsConstant(incoming_value)) {
phi_type = LLVMTypeOf(incoming_value);
break;
}
}
lbValue res = {};
if (phi_type == nullptr) {
phi = LLVMBuildPhi(p->builder, dst_type, "");
LLVMAddIncoming(phi, incoming_values.data, incoming_blocks.data, cast(unsigned)incoming_values.count);
res.value = phi;
res.type = t_llvm_bool;
} else {
for_array(i, incoming_values) {
LLVMValueRef incoming_value = incoming_values[i];
LLVMTypeRef incoming_type = LLVMTypeOf(incoming_value);
if (phi_type != incoming_type) {
GB_ASSERT_MSG(LLVMIsConstant(incoming_value), "%s vs %s", LLVMPrintTypeToString(phi_type), LLVMPrintTypeToString(incoming_type));
bool ok = !!LLVMConstIntGetZExtValue(incoming_value);
incoming_values[i] = LLVMConstInt(phi_type, ok, false);
}
}
// NOTE(bill): this now only uses i1 for the logic to prevent issues with corrupted booleans which are not of value 0 or 1 (e.g. 2)
// Doing this may produce slightly worse code as a result but it will be correct behaviour
phi = LLVMBuildPhi(p->builder, phi_type, "");
LLVMAddIncoming(phi, incoming_values.data, incoming_blocks.data, cast(unsigned)incoming_values.count);
res.value = phi;
res.type = t_llvm_bool;
}
return lb_emit_conv(p, res, default_type(final_type));
}
gb_internal lbValue lb_emit_unary_arith(lbProcedure *p, TokenKind op, lbValue x, Type *type) {
switch (op) {
case Token_Add:
return x;
case Token_Not: // Boolean not
case Token_Xor: // Bitwise not
case Token_Sub: // Number negation
break;
case Token_Pointer:
GB_PANIC("This should be handled elsewhere");
break;
}
if (is_type_array_like(x.type)) {
// IMPORTANT TODO(bill): This is very wasteful with regards to stack memory
Type *tl = base_type(x.type);
lbValue val = lb_address_from_load_or_generate_local(p, x);
GB_ASSERT(is_type_array_like(type));
Type *elem_type = base_array_type(type);
// NOTE(bill): Doesn't need to be zero because it will be initialized in the loops
lbAddr res_addr = lb_add_local(p, type, nullptr, false, true);
lbValue res = lb_addr_get_ptr(p, res_addr);
bool inline_array_arith = lb_can_try_to_inline_array_arith(type);
i32 count = cast(i32)get_array_type_count(tl);
LLVMTypeRef vector_type = nullptr;
if (op != Token_Not && lb_try_vector_cast(p->module, val, &vector_type)) {
LLVMValueRef vp = LLVMBuildPointerCast(p->builder, val.value, LLVMPointerType(vector_type, 0), "");
LLVMValueRef v = LLVMBuildLoad2(p->builder, vector_type, vp, "");
LLVMValueRef opv = nullptr;
switch (op) {
case Token_Xor:
opv = LLVMBuildNot(p->builder, v, "");
break;
case Token_Sub:
if (is_type_float(elem_type)) {
opv = LLVMBuildFNeg(p->builder, v, "");
} else {
opv = LLVMBuildNeg(p->builder, v, "");
}
break;
}
if (opv != nullptr) {
LLVMSetAlignment(res.value, cast(unsigned)lb_alignof(vector_type));
LLVMValueRef res_ptr = LLVMBuildPointerCast(p->builder, res.value, LLVMPointerType(vector_type, 0), "");
LLVMBuildStore(p->builder, opv, res_ptr);
return lb_emit_conv(p, lb_emit_load(p, res), type);
}
}
if (inline_array_arith) {
// inline
for (i32 i = 0; i < count; i++) {
lbValue e = lb_emit_load(p, lb_emit_array_epi(p, val, i));
lbValue z = lb_emit_unary_arith(p, op, e, elem_type);
lb_emit_store(p, lb_emit_array_epi(p, res, i), z);
}
} else {
auto loop_data = lb_loop_start(p, count, t_i32);
lbValue e = lb_emit_load(p, lb_emit_array_ep(p, val, loop_data.idx));
lbValue z = lb_emit_unary_arith(p, op, e, elem_type);
lb_emit_store(p, lb_emit_array_ep(p, res, loop_data.idx), z);
lb_loop_end(p, loop_data);
}
return lb_emit_load(p, res);
}
if (op == Token_Xor) {
lbValue cmp = {};
cmp.value = LLVMBuildNot(p->builder, x.value, "");
cmp.type = x.type;
return lb_emit_conv(p, cmp, type);
}
if (op == Token_Not) {
lbValue cmp = {};
LLVMValueRef zero = LLVMConstInt(lb_type(p->module, x.type), 0, false);
cmp.value = LLVMBuildICmp(p->builder, LLVMIntEQ, x.value, zero, "");
cmp.type = t_llvm_bool;
return lb_emit_conv(p, cmp, type);
}
if (op == Token_Sub && is_type_integer(type) && is_type_different_to_arch_endianness(type)) {
Type *platform_type = integer_endian_type_to_platform_type(type);
lbValue v = lb_emit_byte_swap(p, x, platform_type);
lbValue res = {};
res.value = LLVMBuildNeg(p->builder, v.value, "");
res.type = platform_type;
return lb_emit_byte_swap(p, res, type);
}
if (op == Token_Sub && is_type_float(type) && is_type_different_to_arch_endianness(type)) {
Type *platform_type = integer_endian_type_to_platform_type(type);
lbValue v = lb_emit_byte_swap(p, x, platform_type);
lbValue res = {};
res.value = LLVMBuildFNeg(p->builder, v.value, "");
res.type = platform_type;
return lb_emit_byte_swap(p, res, type);
}
lbValue res = {};
switch (op) {
case Token_Not: // Boolean not
case Token_Xor: // Bitwise not
res.value = LLVMBuildNot(p->builder, x.value, "");
res.type = x.type;
return res;
case Token_Sub: // Number negation
if (is_type_integer(x.type)) {
res.value = LLVMBuildNeg(p->builder, x.value, "");
} else if (is_type_float(x.type)) {
res.value = LLVMBuildFNeg(p->builder, x.value, "");
} else if (is_type_complex(x.type)) {
LLVMValueRef v0 = LLVMBuildFNeg(p->builder, LLVMBuildExtractValue(p->builder, x.value, 0, ""), "");
LLVMValueRef v1 = LLVMBuildFNeg(p->builder, LLVMBuildExtractValue(p->builder, x.value, 1, ""), "");
lbAddr addr = lb_add_local_generated(p, x.type, false);
LLVMTypeRef type = llvm_addr_type(p->module, addr.addr);
LLVMBuildStore(p->builder, v0, LLVMBuildStructGEP2(p->builder, type, addr.addr.value, 0, ""));
LLVMBuildStore(p->builder, v1, LLVMBuildStructGEP2(p->builder, type, addr.addr.value, 1, ""));
return lb_addr_load(p, addr);
} else if (is_type_quaternion(x.type)) {
LLVMValueRef v0 = LLVMBuildFNeg(p->builder, LLVMBuildExtractValue(p->builder, x.value, 0, ""), "");
LLVMValueRef v1 = LLVMBuildFNeg(p->builder, LLVMBuildExtractValue(p->builder, x.value, 1, ""), "");
LLVMValueRef v2 = LLVMBuildFNeg(p->builder, LLVMBuildExtractValue(p->builder, x.value, 2, ""), "");
LLVMValueRef v3 = LLVMBuildFNeg(p->builder, LLVMBuildExtractValue(p->builder, x.value, 3, ""), "");
lbAddr addr = lb_add_local_generated(p, x.type, false);
LLVMTypeRef type = llvm_addr_type(p->module, addr.addr);
LLVMBuildStore(p->builder, v0, LLVMBuildStructGEP2(p->builder, type, addr.addr.value, 0, ""));
LLVMBuildStore(p->builder, v1, LLVMBuildStructGEP2(p->builder, type, addr.addr.value, 1, ""));
LLVMBuildStore(p->builder, v2, LLVMBuildStructGEP2(p->builder, type, addr.addr.value, 2, ""));
LLVMBuildStore(p->builder, v3, LLVMBuildStructGEP2(p->builder, type, addr.addr.value, 3, ""));
return lb_addr_load(p, addr);
} else if (is_type_simd_vector(x.type)) {
Type *elem = base_array_type(x.type);
if (is_type_float(elem)) {
res.value = LLVMBuildFNeg(p->builder, x.value, "");
} else {
res.value = LLVMBuildNeg(p->builder, x.value, "");
}
} else if (is_type_matrix(x.type)) {
lbValue zero = {};
zero.value = LLVMConstNull(lb_type(p->module, type));
zero.type = type;
return lb_emit_arith_matrix(p, Token_Sub, zero, x, type, true);
} else {
GB_PANIC("Unhandled type %s", type_to_string(x.type));
}
res.type = x.type;
return res;
}
return res;
}
gb_internal bool lb_try_direct_vector_arith(lbProcedure *p, TokenKind op, lbValue lhs, lbValue rhs, Type *type, lbValue *res_) {
GB_ASSERT(is_type_array_like(type));
Type *elem_type = base_array_type(type);
// NOTE(bill): Shift operations cannot be easily dealt with due to Odin's semantics
if (op == Token_Shl || op == Token_Shr) {
return false;
}
if (!LLVMIsALoadInst(lhs.value) || !LLVMIsALoadInst(rhs.value)) {
return false;
}
lbValue lhs_ptr = {};
lbValue rhs_ptr = {};
lhs_ptr.value = LLVMGetOperand(lhs.value, 0);
lhs_ptr.type = alloc_type_pointer(lhs.type);
rhs_ptr.value = LLVMGetOperand(rhs.value, 0);
rhs_ptr.type = alloc_type_pointer(rhs.type);
LLVMTypeRef vector_type0 = nullptr;
LLVMTypeRef vector_type1 = nullptr;
if (lb_try_vector_cast(p->module, lhs_ptr, &vector_type0) &&
lb_try_vector_cast(p->module, rhs_ptr, &vector_type1)) {
GB_ASSERT(vector_type0 == vector_type1);
LLVMTypeRef vector_type = vector_type0;
Type *integral_type = base_type(elem_type);
if (is_type_simd_vector(integral_type)) {
integral_type = core_array_type(integral_type);
}
if (is_type_bit_set(integral_type)) {
switch (op) {
case Token_Add: op = Token_Or; break;
case Token_Sub: op = Token_AndNot; break;
}
Type *u = bit_set_to_int(type);
if (is_type_array(u)) {
return false;
}
}
LLVMValueRef lhs_vp = LLVMBuildPointerCast(p->builder, lhs_ptr.value, LLVMPointerType(vector_type, 0), "");
LLVMValueRef rhs_vp = LLVMBuildPointerCast(p->builder, rhs_ptr.value, LLVMPointerType(vector_type, 0), "");
LLVMValueRef x = LLVMBuildLoad2(p->builder, vector_type, lhs_vp, "");
LLVMValueRef y = LLVMBuildLoad2(p->builder, vector_type, rhs_vp, "");
LLVMValueRef z = nullptr;
if (is_type_float(integral_type)) {
switch (op) {
case Token_Add:
z = LLVMBuildFAdd(p->builder, x, y, "");
break;
case Token_Sub:
z = LLVMBuildFSub(p->builder, x, y, "");
break;
case Token_Mul:
z = LLVMBuildFMul(p->builder, x, y, "");
break;
case Token_Quo:
z = LLVMBuildFDiv(p->builder, x, y, "");
break;
case Token_Mod:
z = LLVMBuildFRem(p->builder, x, y, "");
break;
default:
GB_PANIC("Unsupported vector operation %.*s", LIT(token_strings[op]));
break;
}
} else {
switch (op) {
case Token_Add:
z = LLVMBuildAdd(p->builder, x, y, "");
break;
case Token_Sub:
z = LLVMBuildSub(p->builder, x, y, "");
break;
case Token_Mul:
z = LLVMBuildMul(p->builder, x, y, "");
break;
case Token_Quo:
if (is_type_unsigned(integral_type)) {
z = LLVMBuildUDiv(p->builder, x, y, "");
} else {
z = LLVMBuildSDiv(p->builder, x, y, "");
}
break;
case Token_Mod:
if (is_type_unsigned(integral_type)) {
z = LLVMBuildURem(p->builder, x, y, "");
} else {
z = LLVMBuildSRem(p->builder, x, y, "");
}
break;
case Token_ModMod:
if (is_type_unsigned(integral_type)) {
z = LLVMBuildURem(p->builder, x, y, "");
} else {
LLVMValueRef a = LLVMBuildSRem(p->builder, x, y, "");
LLVMValueRef b = LLVMBuildAdd(p->builder, a, y, "");
z = LLVMBuildSRem(p->builder, b, y, "");
}
break;
case Token_And:
z = LLVMBuildAnd(p->builder, x, y, "");
break;
case Token_AndNot:
z = LLVMBuildAnd(p->builder, x, LLVMBuildNot(p->builder, y, ""), "");
break;
case Token_Or:
z = LLVMBuildOr(p->builder, x, y, "");
break;
case Token_Xor:
z = LLVMBuildXor(p->builder, x, y, "");
break;
default:
GB_PANIC("Unsupported vector operation");
break;
}
}
if (z != nullptr) {
lbAddr res = lb_add_local_generated_temp(p, type, lb_alignof(vector_type));
LLVMValueRef vp = LLVMBuildPointerCast(p->builder, res.addr.value, LLVMPointerType(vector_type, 0), "");
LLVMBuildStore(p->builder, z, vp);
lbValue v = lb_addr_load(p, res);
if (res_) *res_ = v;
return true;
}
}
return false;
}
gb_internal lbValue lb_emit_arith_array(lbProcedure *p, TokenKind op, lbValue lhs, lbValue rhs, Type *type) {
GB_ASSERT(is_type_array_like(lhs.type) || is_type_array_like(rhs.type));
lhs = lb_emit_conv(p, lhs, type);
rhs = lb_emit_conv(p, rhs, type);
GB_ASSERT(is_type_array_like(type));
Type *elem_type = base_array_type(type);
i64 count = get_array_type_count(type);
unsigned n = cast(unsigned)count;
// NOTE(bill, 2021-06-12): Try to do a direct operation as a vector, if possible
lbValue direct_vector_res = {};
if (lb_try_direct_vector_arith(p, op, lhs, rhs, type, &direct_vector_res)) {
return direct_vector_res;
}
bool inline_array_arith = lb_can_try_to_inline_array_arith(type);
if (inline_array_arith) {
auto dst_ptrs = slice_make<lbValue>(temporary_allocator(), n);
auto a_loads = slice_make<lbValue>(temporary_allocator(), n);
auto b_loads = slice_make<lbValue>(temporary_allocator(), n);
auto c_ops = slice_make<lbValue>(temporary_allocator(), n);
for (unsigned i = 0; i < n; i++) {
a_loads[i].value = LLVMBuildExtractValue(p->builder, lhs.value, i, "");
a_loads[i].type = elem_type;
}
for (unsigned i = 0; i < n; i++) {
b_loads[i].value = LLVMBuildExtractValue(p->builder, rhs.value, i, "");
b_loads[i].type = elem_type;
}
for (unsigned i = 0; i < n; i++) {
c_ops[i] = lb_emit_arith(p, op, a_loads[i], b_loads[i], elem_type);
}
lbAddr res = lb_add_local_generated(p, type, false);
for (unsigned i = 0; i < n; i++) {
dst_ptrs[i] = lb_emit_array_epi(p, res.addr, i);
}
for (unsigned i = 0; i < n; i++) {
lb_emit_store(p, dst_ptrs[i], c_ops[i]);
}
return lb_addr_load(p, res);
} else {
lbValue x = lb_address_from_load_or_generate_local(p, lhs);
lbValue y = lb_address_from_load_or_generate_local(p, rhs);
lbAddr res = lb_add_local_generated(p, type, false);
auto loop_data = lb_loop_start(p, cast(isize)count, t_i32);
lbValue a_ptr = lb_emit_array_ep(p, x, loop_data.idx);
lbValue b_ptr = lb_emit_array_ep(p, y, loop_data.idx);
lbValue dst_ptr = lb_emit_array_ep(p, res.addr, loop_data.idx);
lbValue a = lb_emit_load(p, a_ptr);
lbValue b = lb_emit_load(p, b_ptr);
lbValue c = lb_emit_arith(p, op, a, b, elem_type);
lb_emit_store(p, dst_ptr, c);
lb_loop_end(p, loop_data);
return lb_addr_load(p, res);
}
}
gb_internal bool lb_is_matrix_simdable(Type *t) {
Type *mt = base_type(t);
GB_ASSERT(mt->kind == Type_Matrix);
Type *elem = core_type(mt->Matrix.elem);
if (is_type_complex(elem)) {
return false;
}
if (is_type_different_to_arch_endianness(elem)) {
return false;
}
switch (build_context.metrics.arch) {
default:
return false;
case TargetArch_amd64:
case TargetArch_arm64:
break;
}
if (type_align_of(t) < 16) {
// it's not aligned well enough to use the vector instructions
return false;
}
if ((mt->Matrix.row_count & 1) ^ (mt->Matrix.column_count & 1)) {
return false;
}
if (mt->Matrix.is_row_major) {
// TODO(bill): make #row_major matrices work with SIMD
return false;
}
if (elem->kind == Type_Basic) {
switch (elem->Basic.kind) {
case Basic_f16:
case Basic_f16le:
case Basic_f16be:
switch (build_context.metrics.arch) {
case TargetArch_amd64:
return false;
case TargetArch_arm64:
// TODO(bill): determine when this is fine
return true;
case TargetArch_i386:
case TargetArch_wasm32:
case TargetArch_wasm64p32:
return false;
}
}
}
return true;
}
gb_internal LLVMValueRef lb_matrix_to_vector(lbProcedure *p, lbValue matrix) {
Type *mt = base_type(matrix.type);
GB_ASSERT(mt->kind == Type_Matrix);
LLVMTypeRef elem_type = lb_type(p->module, mt->Matrix.elem);
unsigned total_count = cast(unsigned)matrix_type_total_internal_elems(mt);
LLVMTypeRef total_matrix_type = LLVMVectorType(elem_type, total_count);
#if 1
LLVMValueRef ptr = lb_address_from_load_or_generate_local(p, matrix).value;
LLVMValueRef matrix_vector_ptr = LLVMBuildPointerCast(p->builder, ptr, LLVMPointerType(total_matrix_type, 0), "");
LLVMValueRef matrix_vector = LLVMBuildLoad2(p->builder, total_matrix_type, matrix_vector_ptr, "");
LLVMSetAlignment(matrix_vector, cast(unsigned)type_align_of(mt));
return matrix_vector;
#else
LLVMValueRef matrix_vector = LLVMBuildBitCast(p->builder, matrix.value, total_matrix_type, "");
return matrix_vector;
#endif
}
gb_internal LLVMValueRef lb_matrix_trimmed_vector_mask(lbProcedure *p, Type *mt) {
mt = base_type(mt);
GB_ASSERT(mt->kind == Type_Matrix);
unsigned stride = cast(unsigned)matrix_type_stride_in_elems(mt);
unsigned row_count = cast(unsigned)mt->Matrix.row_count;
unsigned column_count = cast(unsigned)mt->Matrix.column_count;
unsigned mask_elems_index = 0;
auto mask_elems = slice_make<LLVMValueRef>(permanent_allocator(), row_count*column_count);
for (unsigned j = 0; j < column_count; j++) {
for (unsigned i = 0; i < row_count; i++) {
unsigned offset = stride*j + i;
mask_elems[mask_elems_index++] = lb_const_int(p->module, t_u32, offset).value;
}
}
LLVMValueRef mask = LLVMConstVector(mask_elems.data, cast(unsigned)mask_elems.count);
return mask;
}
gb_internal LLVMValueRef lb_matrix_to_trimmed_vector(lbProcedure *p, lbValue m) {
LLVMValueRef vector = lb_matrix_to_vector(p, m);
Type *mt = base_type(m.type);
GB_ASSERT(mt->kind == Type_Matrix);
unsigned stride = cast(unsigned)matrix_type_stride_in_elems(mt);
unsigned row_count = cast(unsigned)mt->Matrix.row_count;
if (stride == row_count) {
return vector;
}
LLVMValueRef mask = lb_matrix_trimmed_vector_mask(p, mt);
LLVMValueRef trimmed_vector = llvm_basic_shuffle(p, vector, mask);
return trimmed_vector;
}
gb_internal lbValue lb_emit_matrix_tranpose(lbProcedure *p, lbValue m, Type *type) {
if (is_type_array(m.type)) {
i32 rank = type_math_rank(m.type);
if (rank == 2) {
lbAddr addr = lb_add_local_generated(p, type, false);
lbValue dst = addr.addr;
lbValue src = m;
i32 n = cast(i32)get_array_type_count(m.type);
i32 m = cast(i32)get_array_type_count(type);
// m.type == [n][m]T
// type == [m][n]T
for (i32 j = 0; j < m; j++) {
lbValue dst_col = lb_emit_struct_ep(p, dst, j);
for (i32 i = 0; i < n; i++) {
lbValue dst_row = lb_emit_struct_ep(p, dst_col, i);
lbValue src_col = lb_emit_struct_ev(p, src, i);
lbValue src_row = lb_emit_struct_ev(p, src_col, j);
lb_emit_store(p, dst_row, src_row);
}
}
return lb_addr_load(p, addr);
}
// no-op
m.type = type;
return m;
}
Type *mt = base_type(m.type);
GB_ASSERT(mt->kind == Type_Matrix);
if (lb_is_matrix_simdable(mt)) {
unsigned stride = cast(unsigned)matrix_type_stride_in_elems(mt);
unsigned row_count = cast(unsigned)mt->Matrix.row_count;
unsigned column_count = cast(unsigned)mt->Matrix.column_count;
auto rows = slice_make<LLVMValueRef>(permanent_allocator(), row_count);
auto mask_elems = slice_make<LLVMValueRef>(permanent_allocator(), column_count);
LLVMValueRef vector = lb_matrix_to_vector(p, m);
for (unsigned i = 0; i < row_count; i++) {
for (unsigned j = 0; j < column_count; j++) {
unsigned offset = stride*j + i;
mask_elems[j] = lb_const_int(p->module, t_u32, offset).value;
}
// transpose mask
LLVMValueRef mask = LLVMConstVector(mask_elems.data, column_count);
LLVMValueRef row = llvm_basic_shuffle(p, vector, mask);
rows[i] = row;
}
lbAddr res = lb_add_local_generated(p, type, true);
for_array(i, rows) {
LLVMValueRef row = rows[i];
lbValue dst_row_ptr = lb_emit_matrix_epi(p, res.addr, 0, i);
LLVMValueRef ptr = dst_row_ptr.value;
ptr = LLVMBuildPointerCast(p->builder, ptr, LLVMPointerType(LLVMTypeOf(row), 0), "");
LLVMBuildStore(p->builder, row, ptr);
}
return lb_addr_load(p, res);
}
lbAddr res = lb_add_local_generated(p, type, true);
i64 row_count = mt->Matrix.row_count;
i64 column_count = mt->Matrix.column_count;
for (i64 j = 0; j < column_count; j++) {
for (i64 i = 0; i < row_count; i++) {
lbValue src = lb_emit_matrix_ev(p, m, i, j);
lbValue dst = lb_emit_matrix_epi(p, res.addr, j, i);
lb_emit_store(p, dst, src);
}
}
return lb_addr_load(p, res);
}
gb_internal lbValue lb_matrix_cast_vector_to_type(lbProcedure *p, LLVMValueRef vector, Type *type) {
lbAddr res = lb_add_local_generated(p, type, true);
LLVMValueRef res_ptr = res.addr.value;
unsigned alignment = cast(unsigned)gb_max(type_align_of(type), lb_alignof(LLVMTypeOf(vector)));
LLVMSetAlignment(res_ptr, alignment);
res_ptr = LLVMBuildPointerCast(p->builder, res_ptr, LLVMPointerType(LLVMTypeOf(vector), 0), "");
LLVMBuildStore(p->builder, vector, res_ptr);
return lb_addr_load(p, res);
}
gb_internal lbValue lb_emit_matrix_flatten(lbProcedure *p, lbValue m, Type *type) {
if (is_type_array(m.type)) {
// no-op
m.type = type;
return m;
}
Type *mt = base_type(m.type);
GB_ASSERT(mt->kind == Type_Matrix);
lbAddr res = lb_add_local_generated(p, type, true);
i64 row_count = mt->Matrix.row_count;
i64 column_count = mt->Matrix.column_count;
TEMPORARY_ALLOCATOR_GUARD();
auto srcs = array_make<lbValue>(temporary_allocator(), 0, row_count*column_count);
auto dsts = array_make<lbValue>(temporary_allocator(), 0, row_count*column_count);
for (i64 j = 0; j < column_count; j++) {
for (i64 i = 0; i < row_count; i++) {
lbValue src = lb_emit_matrix_ev(p, m, i, j);
array_add(&srcs, src);
}
}
for (i64 j = 0; j < column_count; j++) {
for (i64 i = 0; i < row_count; i++) {
lbValue dst = lb_emit_array_epi(p, res.addr, i + j*row_count);
array_add(&dsts, dst);
}
}
GB_ASSERT(srcs.count == dsts.count);
for_array(i, srcs) {
lb_emit_store(p, dsts[i], srcs[i]);
}
return lb_addr_load(p, res);
}
gb_internal lbValue lb_emit_outer_product(lbProcedure *p, lbValue a, lbValue b, Type *type) {
Type *mt = base_type(type);
Type *at = base_type(a.type);
Type *bt = base_type(b.type);
GB_ASSERT(mt->kind == Type_Matrix);
GB_ASSERT(at->kind == Type_Array);
GB_ASSERT(bt->kind == Type_Array);
i64 row_count = mt->Matrix.row_count;
i64 column_count = mt->Matrix.column_count;
GB_ASSERT(row_count == at->Array.count);
GB_ASSERT(column_count == bt->Array.count);
lbAddr res = lb_add_local_generated(p, type, true);
for (i64 j = 0; j < column_count; j++) {
for (i64 i = 0; i < row_count; i++) {
lbValue x = lb_emit_struct_ev(p, a, cast(i32)i);
lbValue y = lb_emit_struct_ev(p, b, cast(i32)j);
lbValue src = lb_emit_arith(p, Token_Mul, x, y, mt->Matrix.elem);
lbValue dst = lb_emit_matrix_epi(p, res.addr, i, j);
lb_emit_store(p, dst, src);
}
}
return lb_addr_load(p, res);
}
gb_internal lbValue lb_emit_matrix_mul(lbProcedure *p, lbValue lhs, lbValue rhs, Type *type) {
// TODO(bill): Handle edge case for f16 types on x86(-64) platforms
Type *xt = base_type(lhs.type);
Type *yt = base_type(rhs.type);
GB_ASSERT(is_type_matrix(type));
GB_ASSERT(is_type_matrix(xt));
GB_ASSERT(is_type_matrix(yt));
GB_ASSERT(xt->Matrix.column_count == yt->Matrix.row_count);
GB_ASSERT(are_types_identical(xt->Matrix.elem, yt->Matrix.elem));
GB_ASSERT(xt->Matrix.is_row_major == yt->Matrix.is_row_major);
Type *elem = xt->Matrix.elem;
unsigned outer_rows = cast(unsigned)xt->Matrix.row_count;
unsigned inner = cast(unsigned)xt->Matrix.column_count;
unsigned outer_columns = cast(unsigned)yt->Matrix.column_count;
if (!xt->Matrix.is_row_major && lb_is_matrix_simdable(xt)) {
unsigned x_stride = cast(unsigned)matrix_type_stride_in_elems(xt);
unsigned y_stride = cast(unsigned)matrix_type_stride_in_elems(yt);
auto x_rows = slice_make<LLVMValueRef>(permanent_allocator(), outer_rows);
auto y_columns = slice_make<LLVMValueRef>(permanent_allocator(), outer_columns);
LLVMValueRef x_vector = lb_matrix_to_vector(p, lhs);
LLVMValueRef y_vector = lb_matrix_to_vector(p, rhs);
auto mask_elems = slice_make<LLVMValueRef>(permanent_allocator(), inner);
for (unsigned i = 0; i < outer_rows; i++) {
for (unsigned j = 0; j < inner; j++) {
unsigned offset = x_stride*j + i;
mask_elems[j] = lb_const_int(p->module, t_u32, offset).value;
}
// transpose mask
LLVMValueRef mask = LLVMConstVector(mask_elems.data, inner);
LLVMValueRef row = llvm_basic_shuffle(p, x_vector, mask);
x_rows[i] = row;
}
for (unsigned i = 0; i < outer_columns; i++) {
LLVMValueRef mask = llvm_mask_iota(p->module, y_stride*i, inner);
LLVMValueRef column = llvm_basic_shuffle(p, y_vector, mask);
y_columns[i] = column;
}
lbAddr res = lb_add_local_generated(p, type, true);
for_array(i, x_rows) {
LLVMValueRef x_row = x_rows[i];
for_array(j, y_columns) {
LLVMValueRef y_column = y_columns[j];
LLVMValueRef elem = llvm_vector_dot(p, x_row, y_column);
lbValue dst = lb_emit_matrix_epi(p, res.addr, i, j);
LLVMBuildStore(p->builder, elem, dst.value);
}
}
return lb_addr_load(p, res);
}
if (!xt->Matrix.is_row_major) {
lbAddr res = lb_add_local_generated(p, type, true);
auto inners = slice_make<lbValue[2]>(permanent_allocator(), inner);
for (unsigned j = 0; j < outer_columns; j++) {
for (unsigned i = 0; i < outer_rows; i++) {
lbValue dst = lb_emit_matrix_epi(p, res.addr, i, j);
for (unsigned k = 0; k < inner; k++) {
inners[k][0] = lb_emit_matrix_ev(p, lhs, i, k);
inners[k][1] = lb_emit_matrix_ev(p, rhs, k, j);
}
lbValue sum = lb_const_nil(p->module, elem);
for (unsigned k = 0; k < inner; k++) {
lbValue a = inners[k][0];
lbValue b = inners[k][1];
sum = lb_emit_mul_add(p, a, b, sum, elem);
}
lb_emit_store(p, dst, sum);
}
}
return lb_addr_load(p, res);
} else {
lbAddr res = lb_add_local_generated(p, type, true);
auto inners = slice_make<lbValue[2]>(permanent_allocator(), inner);
for (unsigned i = 0; i < outer_rows; i++) {
for (unsigned j = 0; j < outer_columns; j++) {
lbValue dst = lb_emit_matrix_epi(p, res.addr, i, j);
for (unsigned k = 0; k < inner; k++) {
inners[k][0] = lb_emit_matrix_ev(p, lhs, i, k);
inners[k][1] = lb_emit_matrix_ev(p, rhs, k, j);
}
lbValue sum = lb_const_nil(p->module, elem);
for (unsigned k = 0; k < inner; k++) {
lbValue a = inners[k][0];
lbValue b = inners[k][1];
sum = lb_emit_mul_add(p, a, b, sum, elem);
}
lb_emit_store(p, dst, sum);
}
}
return lb_addr_load(p, res);
}
}
gb_internal lbValue lb_emit_matrix_mul_vector(lbProcedure *p, lbValue lhs, lbValue rhs, Type *type) {
// TODO(bill): Handle edge case for f16 types on x86(-64) platforms
Type *mt = base_type(lhs.type);
Type *vt = base_type(rhs.type);
GB_ASSERT(is_type_matrix(mt));
GB_ASSERT(is_type_array_like(vt));
i64 vector_count = get_array_type_count(vt);
GB_ASSERT(mt->Matrix.column_count == vector_count);
GB_ASSERT(are_types_identical(mt->Matrix.elem, base_array_type(vt)));
Type *elem = mt->Matrix.elem;
if (!mt->Matrix.is_row_major && lb_is_matrix_simdable(mt)) {
unsigned stride = cast(unsigned)matrix_type_stride_in_elems(mt);
unsigned row_count = cast(unsigned)mt->Matrix.row_count;
unsigned column_count = cast(unsigned)mt->Matrix.column_count;
auto m_columns = slice_make<LLVMValueRef>(permanent_allocator(), column_count);
auto v_rows = slice_make<LLVMValueRef>(permanent_allocator(), column_count);
LLVMValueRef matrix_vector = lb_matrix_to_vector(p, lhs);
for (unsigned column_index = 0; column_index < column_count; column_index++) {
LLVMValueRef mask = llvm_mask_iota(p->module, stride*column_index, row_count);
LLVMValueRef column = llvm_basic_shuffle(p, matrix_vector, mask);
m_columns[column_index] = column;
}
for (unsigned row_index = 0; row_index < column_count; row_index++) {
LLVMValueRef value = lb_emit_struct_ev(p, rhs, row_index).value;
LLVMValueRef row = llvm_vector_broadcast(p, value, row_count);
v_rows[row_index] = row;
}
GB_ASSERT(column_count > 0);
LLVMValueRef vector = nullptr;
for (i64 i = 0; i < column_count; i++) {
if (i == 0) {
vector = llvm_vector_mul(p, m_columns[i], v_rows[i]);
} else {
vector = llvm_vector_mul_add(p, m_columns[i], v_rows[i], vector);
}
}
return lb_matrix_cast_vector_to_type(p, vector, type);
}
lbAddr res = lb_add_local_generated(p, type, true);
for (i64 i = 0; i < mt->Matrix.row_count; i++) {
for (i64 j = 0; j < mt->Matrix.column_count; j++) {
lbValue dst = lb_emit_matrix_epi(p, res.addr, i, 0);
lbValue d0 = lb_emit_load(p, dst);
lbValue a = lb_emit_matrix_ev(p, lhs, i, j);
lbValue b = lb_emit_struct_ev(p, rhs, cast(i32)j);
lbValue c = lb_emit_mul_add(p, a, b, d0, elem);
lb_emit_store(p, dst, c);
}
}
return lb_addr_load(p, res);
}
gb_internal lbValue lb_emit_vector_mul_matrix(lbProcedure *p, lbValue lhs, lbValue rhs, Type *type) {
// TODO(bill): Handle edge case for f16 types on x86(-64) platforms
Type *mt = base_type(rhs.type);
Type *vt = base_type(lhs.type);
GB_ASSERT(is_type_matrix(mt));
GB_ASSERT(is_type_array_like(vt));
i64 vector_count = get_array_type_count(vt);
GB_ASSERT(vector_count == mt->Matrix.row_count);
GB_ASSERT(are_types_identical(mt->Matrix.elem, base_array_type(vt)));
Type *elem = mt->Matrix.elem;
if (!mt->Matrix.is_row_major && lb_is_matrix_simdable(mt)) {
unsigned stride = cast(unsigned)matrix_type_stride_in_elems(mt);
unsigned row_count = cast(unsigned)mt->Matrix.row_count;
unsigned column_count = cast(unsigned)mt->Matrix.column_count; gb_unused(column_count);
auto m_columns = slice_make<LLVMValueRef>(permanent_allocator(), row_count);
auto v_rows = slice_make<LLVMValueRef>(permanent_allocator(), row_count);
LLVMValueRef matrix_vector = lb_matrix_to_vector(p, rhs);
auto mask_elems = slice_make<LLVMValueRef>(permanent_allocator(), column_count);
for (unsigned row_index = 0; row_index < row_count; row_index++) {
for (unsigned column_index = 0; column_index < column_count; column_index++) {
unsigned offset = row_index + column_index*stride;
mask_elems[column_index] = lb_const_int(p->module, t_u32, offset).value;
}
// transpose mask
LLVMValueRef mask = LLVMConstVector(mask_elems.data, column_count);
LLVMValueRef column = llvm_basic_shuffle(p, matrix_vector, mask);
m_columns[row_index] = column;
}
for (unsigned column_index = 0; column_index < row_count; column_index++) {
LLVMValueRef value = lb_emit_struct_ev(p, lhs, column_index).value;
LLVMValueRef row = llvm_vector_broadcast(p, value, column_count);
v_rows[column_index] = row;
}
GB_ASSERT(row_count > 0);
LLVMValueRef vector = nullptr;
for (i64 i = 0; i < row_count; i++) {
if (i == 0) {
vector = llvm_vector_mul(p, v_rows[i], m_columns[i]);
} else {
vector = llvm_vector_mul_add(p, v_rows[i], m_columns[i], vector);
}
}
lbAddr res = lb_add_local_generated(p, type, true);
LLVMValueRef res_ptr = res.addr.value;
unsigned alignment = cast(unsigned)gb_max(type_align_of(type), lb_alignof(LLVMTypeOf(vector)));
LLVMSetAlignment(res_ptr, alignment);
res_ptr = LLVMBuildPointerCast(p->builder, res_ptr, LLVMPointerType(LLVMTypeOf(vector), 0), "");
LLVMBuildStore(p->builder, vector, res_ptr);