forked from Rust-GPU/rust-gpu
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbuilder_methods.rs
3265 lines (3038 loc) · 132 KB
/
builder_methods.rs
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
use super::Builder;
use crate::abi::ConvSpirvType;
use crate::builder_spirv::{BuilderCursor, SpirvConst, SpirvValue, SpirvValueExt, SpirvValueKind};
use crate::custom_insts::{CustomInst, CustomOp};
use crate::rustc_codegen_ssa::traits::BaseTypeMethods;
use crate::spirv_type::SpirvType;
use itertools::Itertools;
use rspirv::dr::{InsertPoint, Instruction, Operand};
use rspirv::spirv::{Capability, MemoryModel, MemorySemantics, Op, Scope, StorageClass, Word};
use rustc_apfloat::{ieee, Float, Round, Status};
use rustc_codegen_ssa::common::{
AtomicOrdering, AtomicRmwBinOp, IntPredicate, RealPredicate, SynchronizationScope, TypeKind,
};
use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
use rustc_codegen_ssa::mir::place::PlaceRef;
use rustc_codegen_ssa::traits::{
BackendTypes, BuilderMethods, ConstMethods, LayoutTypeMethods, OverflowOp,
};
use rustc_codegen_ssa::MemFlags;
use rustc_data_structures::fx::FxHashSet;
use rustc_middle::bug;
use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
use rustc_middle::ty::layout::LayoutOf;
use rustc_middle::ty::{self, Ty};
use rustc_span::Span;
use rustc_target::abi::call::FnAbi;
use rustc_target::abi::{Abi, Align, Scalar, Size, WrappingRange};
use smallvec::SmallVec;
use std::borrow::Cow;
use std::cell::Cell;
use std::iter::{self, empty};
use std::ops::RangeInclusive;
macro_rules! simple_op {
(
$func_name:ident, $inst_name:ident
$(, fold_const {
$(int($fold_int_lhs:ident, $fold_int_rhs:ident) => $fold_int:expr)?
})?
) => {
fn $func_name(&mut self, lhs: Self::Value, rhs: Self::Value) -> Self::Value {
assert_ty_eq!(self, lhs.ty, rhs.ty);
let result_type = lhs.ty;
$(if let Some(const_lhs) = self.builder.lookup_const(lhs) {
if let Some(const_rhs) = self.builder.lookup_const(rhs) {
match self.lookup_type(result_type) {
$(SpirvType::Integer(bits, signed) => {
let size = Size::from_bits(bits);
let as_u128 = |const_val| {
let x = match const_val {
SpirvConst::U32(x) => x as u128,
SpirvConst::U64(x) => x as u128,
_ => return None,
};
Some(if signed {
size.sign_extend(x)
} else {
size.truncate(x)
})
};
if let Some($fold_int_lhs) = as_u128(const_lhs) {
if let Some($fold_int_rhs) = as_u128(const_rhs) {
return self.const_uint_big(result_type, $fold_int);
}
}
})?
_ => {}
}
}
})?
self.emit()
.$inst_name(result_type, None, lhs.def(self), rhs.def(self))
.unwrap()
.with_type(result_type)
}
};
}
// shl and shr allow different types as their operands
macro_rules! simple_op_unchecked_type {
($func_name:ident, $inst_name:ident) => {
fn $func_name(&mut self, lhs: Self::Value, rhs: Self::Value) -> Self::Value {
self.emit()
.$inst_name(lhs.ty, None, lhs.def(self), rhs.def(self))
.unwrap()
.with_type(lhs.ty)
}
};
}
macro_rules! simple_uni_op {
($func_name:ident, $inst_name:ident) => {
fn $func_name(&mut self, val: Self::Value) -> Self::Value {
self.emit()
.$inst_name(val.ty, None, val.def(self))
.unwrap()
.with_type(val.ty)
}
};
}
fn memset_fill_u16(b: u8) -> u16 {
b as u16 | ((b as u16) << 8)
}
fn memset_fill_u32(b: u8) -> u32 {
b as u32 | ((b as u32) << 8) | ((b as u32) << 16) | ((b as u32) << 24)
}
fn memset_fill_u64(b: u8) -> u64 {
b as u64
| ((b as u64) << 8)
| ((b as u64) << 16)
| ((b as u64) << 24)
| ((b as u64) << 32)
| ((b as u64) << 40)
| ((b as u64) << 48)
| ((b as u64) << 56)
}
fn memset_dynamic_scalar(
builder: &Builder<'_, '_>,
fill_var: Word,
byte_width: usize,
is_float: bool,
) -> Word {
let composite_type = SpirvType::Vector {
element: SpirvType::Integer(8, false).def(builder.span(), builder),
count: byte_width as u32,
}
.def(builder.span(), builder);
let composite = builder
.emit()
.composite_construct(
composite_type,
None,
iter::repeat(fill_var).take(byte_width),
)
.unwrap();
let result_type = if is_float {
SpirvType::Float(byte_width as u32 * 8)
} else {
SpirvType::Integer(byte_width as u32 * 8, false)
};
builder
.emit()
.bitcast(result_type.def(builder.span(), builder), None, composite)
.unwrap()
}
impl<'a, 'tcx> Builder<'a, 'tcx> {
fn ordering_to_semantics_def(&self, ordering: AtomicOrdering) -> SpirvValue {
let mut invalid_seq_cst = false;
let semantics = match ordering {
AtomicOrdering::Unordered | AtomicOrdering::Relaxed => MemorySemantics::NONE,
// Note: rustc currently has AtomicOrdering::Consume commented out, if it ever becomes
// uncommented, it should be MakeVisible | Acquire.
AtomicOrdering::Acquire => MemorySemantics::MAKE_VISIBLE | MemorySemantics::ACQUIRE,
AtomicOrdering::Release => MemorySemantics::MAKE_AVAILABLE | MemorySemantics::RELEASE,
AtomicOrdering::AcquireRelease => {
MemorySemantics::MAKE_AVAILABLE
| MemorySemantics::MAKE_VISIBLE
| MemorySemantics::ACQUIRE_RELEASE
}
AtomicOrdering::SequentiallyConsistent => {
let emit = self.emit();
let memory_model = emit.module_ref().memory_model.as_ref().unwrap();
if memory_model.operands[1].unwrap_memory_model() == MemoryModel::Vulkan {
invalid_seq_cst = true;
}
MemorySemantics::MAKE_AVAILABLE
| MemorySemantics::MAKE_VISIBLE
| MemorySemantics::SEQUENTIALLY_CONSISTENT
}
};
let semantics = self.constant_u32(self.span(), semantics.bits());
if invalid_seq_cst {
self.zombie(
semantics.def(self),
"cannot use AtomicOrdering=SequentiallyConsistent on Vulkan memory model \
(check if AcquireRelease fits your needs)",
);
}
semantics
}
fn memset_const_pattern(&self, ty: &SpirvType<'tcx>, fill_byte: u8) -> Word {
match *ty {
SpirvType::Void => self.fatal("memset invalid on void pattern"),
SpirvType::Bool => self.fatal("memset invalid on bool pattern"),
SpirvType::Integer(width, _signedness) => match width {
8 => self.constant_u8(self.span(), fill_byte).def(self),
16 => self
.constant_u16(self.span(), memset_fill_u16(fill_byte))
.def(self),
32 => self
.constant_u32(self.span(), memset_fill_u32(fill_byte))
.def(self),
64 => self
.constant_u64(self.span(), memset_fill_u64(fill_byte))
.def(self),
_ => self.fatal(format!(
"memset on integer width {width} not implemented yet"
)),
},
SpirvType::Float(width) => match width {
32 => self
.constant_f32(self.span(), f32::from_bits(memset_fill_u32(fill_byte)))
.def(self),
64 => self
.constant_f64(self.span(), f64::from_bits(memset_fill_u64(fill_byte)))
.def(self),
_ => self.fatal(format!("memset on float width {width} not implemented yet")),
},
SpirvType::Adt { .. } => self.fatal("memset on structs not implemented yet"),
SpirvType::Vector { element, count } | SpirvType::Matrix { element, count } => {
let elem_pat = self.memset_const_pattern(&self.lookup_type(element), fill_byte);
self.constant_composite(
ty.def(self.span(), self),
iter::repeat(elem_pat).take(count as usize),
)
.def(self)
}
SpirvType::Array { element, count } => {
let elem_pat = self.memset_const_pattern(&self.lookup_type(element), fill_byte);
let count = self.builder.lookup_const_u64(count).unwrap() as usize;
self.constant_composite(
ty.def(self.span(), self),
iter::repeat(elem_pat).take(count),
)
.def(self)
}
SpirvType::RuntimeArray { .. } => {
self.fatal("memset on runtime arrays not implemented yet")
}
SpirvType::Pointer { .. } => self.fatal("memset on pointers not implemented yet"),
SpirvType::Function { .. } => self.fatal("memset on functions not implemented yet"),
SpirvType::Image { .. } => self.fatal("cannot memset image"),
SpirvType::Sampler => self.fatal("cannot memset sampler"),
SpirvType::SampledImage { .. } => self.fatal("cannot memset sampled image"),
SpirvType::InterfaceBlock { .. } => self.fatal("cannot memset interface block"),
SpirvType::AccelerationStructureKhr => {
self.fatal("cannot memset acceleration structure")
}
SpirvType::RayQueryKhr => self.fatal("cannot memset ray query"),
}
}
fn memset_dynamic_pattern(&self, ty: &SpirvType<'tcx>, fill_var: Word) -> Word {
match *ty {
SpirvType::Void => self.fatal("memset invalid on void pattern"),
SpirvType::Bool => self.fatal("memset invalid on bool pattern"),
SpirvType::Integer(width, _signedness) => match width {
8 => fill_var,
16 => memset_dynamic_scalar(self, fill_var, 2, false),
32 => memset_dynamic_scalar(self, fill_var, 4, false),
64 => memset_dynamic_scalar(self, fill_var, 8, false),
_ => self.fatal(format!(
"memset on integer width {width} not implemented yet"
)),
},
SpirvType::Float(width) => match width {
32 => memset_dynamic_scalar(self, fill_var, 4, true),
64 => memset_dynamic_scalar(self, fill_var, 8, true),
_ => self.fatal(format!("memset on float width {width} not implemented yet")),
},
SpirvType::Adt { .. } => self.fatal("memset on structs not implemented yet"),
SpirvType::Array { element, count } => {
let elem_pat = self.memset_dynamic_pattern(&self.lookup_type(element), fill_var);
let count = self.builder.lookup_const_u64(count).unwrap() as usize;
self.emit()
.composite_construct(
ty.def(self.span(), self),
None,
iter::repeat(elem_pat).take(count),
)
.unwrap()
}
SpirvType::Vector { element, count } | SpirvType::Matrix { element, count } => {
let elem_pat = self.memset_dynamic_pattern(&self.lookup_type(element), fill_var);
self.emit()
.composite_construct(
ty.def(self.span(), self),
None,
iter::repeat(elem_pat).take(count as usize),
)
.unwrap()
}
SpirvType::RuntimeArray { .. } => {
self.fatal("memset on runtime arrays not implemented yet")
}
SpirvType::Pointer { .. } => self.fatal("memset on pointers not implemented yet"),
SpirvType::Function { .. } => self.fatal("memset on functions not implemented yet"),
SpirvType::Image { .. } => self.fatal("cannot memset image"),
SpirvType::Sampler => self.fatal("cannot memset sampler"),
SpirvType::SampledImage { .. } => self.fatal("cannot memset sampled image"),
SpirvType::InterfaceBlock { .. } => self.fatal("cannot memset interface block"),
SpirvType::AccelerationStructureKhr => {
self.fatal("cannot memset acceleration structure")
}
SpirvType::RayQueryKhr => self.fatal("cannot memset ray query"),
}
}
fn memset_constant_size(&mut self, ptr: SpirvValue, pat: SpirvValue, size_bytes: u64) {
let size_elem = self
.lookup_type(pat.ty)
.sizeof(self)
.expect("Memset on unsized values not supported");
let count = size_bytes / size_elem.bytes();
if count == 1 {
self.store(pat, ptr, Align::from_bytes(0).unwrap());
} else {
for index in 0..count {
let const_index = self.constant_u32(self.span(), index as u32);
let gep_ptr = self.gep(pat.ty, ptr, &[const_index]);
self.store(pat, gep_ptr, Align::from_bytes(0).unwrap());
}
}
}
// TODO: Test this is correct
fn memset_dynamic_size(&mut self, ptr: SpirvValue, pat: SpirvValue, size_bytes: SpirvValue) {
let size_elem = self
.lookup_type(pat.ty)
.sizeof(self)
.expect("Unable to memset a dynamic sized object");
let size_elem_const = self.constant_int(size_bytes.ty, size_elem.bytes());
let zero = self.constant_int(size_bytes.ty, 0);
let one = self.constant_int(size_bytes.ty, 1);
let zero_align = Align::from_bytes(0).unwrap();
let header_bb = self.append_sibling_block("memset_header");
let body_bb = self.append_sibling_block("memset_body");
let exit_bb = self.append_sibling_block("memset_exit");
let count = self.udiv(size_bytes, size_elem_const);
let index = self.alloca(Size::from_bits(32), zero_align);
self.store(zero, index, zero_align);
self.br(header_bb);
self.switch_to_block(header_bb);
let current_index = self.load(count.ty, index, zero_align);
let cond = self.icmp(IntPredicate::IntULT, current_index, count);
self.cond_br(cond, body_bb, exit_bb);
self.switch_to_block(body_bb);
let gep_ptr = self.gep(pat.ty, ptr, &[current_index]);
self.store(pat, gep_ptr, zero_align);
let current_index_plus_1 = self.add(current_index, one);
self.store(current_index_plus_1, index, zero_align);
self.br(header_bb);
self.switch_to_block(exit_bb);
}
fn zombie_convert_ptr_to_u(&self, def: Word) {
self.zombie(def, "cannot convert pointers to integers");
}
fn zombie_convert_u_to_ptr(&self, def: Word) {
self.zombie(def, "cannot convert integers to pointers");
}
fn zombie_ptr_equal(&self, def: Word, inst: &str) {
if !self.builder.has_capability(Capability::VariablePointers) {
self.zombie(
def,
&format!("{inst} without OpCapability VariablePointers"),
);
}
}
/// Convenience wrapper for `adjust_pointer_for_sized_access`, falling back
/// on choosing `ty` as the leaf's type (and casting `ptr` to a pointer to it).
//
// HACK(eddyb) temporary workaround for untyped pointers upstream.
// FIXME(eddyb) replace with untyped memory SPIR-V + `qptr` or similar.
fn adjust_pointer_for_typed_access(
&mut self,
ptr: SpirvValue,
ty: <Self as BackendTypes>::Type,
) -> (SpirvValue, <Self as BackendTypes>::Type) {
self.lookup_type(ty)
.sizeof(self)
.and_then(|size| self.adjust_pointer_for_sized_access(ptr, size))
.unwrap_or_else(|| (self.pointercast(ptr, self.type_ptr_to(ty)), ty))
}
/// If `ptr`'s pointee type contains any prefix field/element of size `size`,
/// i.e. some leaf which can be used for all accesses of size `size`, return
/// `ptr` adjusted to point to the innermost such leaf, and the leaf's type.
//
// FIXME(eddyb) technically this duplicates `pointercast`, but the main use
// of `pointercast` is being replaced by this, and this can be more efficient.
//
// HACK(eddyb) temporary workaround for untyped pointers upstream.
// FIXME(eddyb) replace with untyped memory SPIR-V + `qptr` or similar.
fn adjust_pointer_for_sized_access(
&mut self,
ptr: SpirvValue,
size: Size,
) -> Option<(SpirvValue, <Self as BackendTypes>::Type)> {
let ptr = ptr.strip_ptrcasts();
let mut leaf_ty = match self.lookup_type(ptr.ty) {
SpirvType::Pointer { pointee } => pointee,
other => self.fatal(format!("non-pointer type: {other:?}")),
};
// FIXME(eddyb) this isn't efficient, `recover_access_chain_from_offset`
// could instead be doing all the extra digging itself.
let mut indices = SmallVec::<[_; 8]>::new();
while let Some((inner_indices, inner_ty)) = self.recover_access_chain_from_offset(
leaf_ty,
Size::ZERO,
Some(size)..=Some(size),
None,
) {
indices.extend(inner_indices);
leaf_ty = inner_ty;
}
let leaf_ptr_ty = (self.lookup_type(leaf_ty).sizeof(self) == Some(size))
.then(|| self.type_ptr_to(leaf_ty))?;
let leaf_ptr = if indices.is_empty() {
assert_ty_eq!(self, ptr.ty, leaf_ptr_ty);
ptr
} else {
let indices = indices
.into_iter()
.map(|idx| self.constant_u32(self.span(), idx).def(self))
.collect::<Vec<_>>();
self.emit()
.in_bounds_access_chain(leaf_ptr_ty, None, ptr.def(self), indices)
.unwrap()
.with_type(leaf_ptr_ty)
};
Some((leaf_ptr, leaf_ty))
}
/// If possible, return the appropriate `OpAccessChain` indices for going
/// from a pointer to `ty`, to a pointer to some leaf field/element having
/// a size that fits `leaf_size_range` (and, optionally, the type `leaf_ty`),
/// while adding `offset` bytes.
///
/// That is, try to turn `((_: *T) as *u8).add(offset) as *Leaf` into a series
/// of struct field and array/vector element accesses.
fn recover_access_chain_from_offset(
&self,
mut ty: <Self as BackendTypes>::Type,
mut offset: Size,
// FIXME(eddyb) using `None` for "unsized" is a pretty bad design.
leaf_size_or_unsized_range: RangeInclusive<Option<Size>>,
leaf_ty: Option<<Self as BackendTypes>::Type>,
) -> Option<(SmallVec<[u32; 8]>, <Self as BackendTypes>::Type)> {
assert_ne!(Some(ty), leaf_ty);
// HACK(eddyb) this has the correct ordering (`Sized(_) < Unsized`).
#[derive(Copy, Clone, PartialEq, Eq, PartialOrd, Ord)]
enum MaybeSized {
Sized(Size),
Unsized,
}
let leaf_size_range = {
let r = leaf_size_or_unsized_range;
let [start, end] =
[r.start(), r.end()].map(|x| x.map_or(MaybeSized::Unsized, MaybeSized::Sized));
start..=end
};
// NOTE(eddyb) `ty` and `ty_kind`/`ty_size` should be kept in sync.
let mut ty_kind = self.lookup_type(ty);
let mut indices = SmallVec::new();
loop {
let ty_size;
match ty_kind {
SpirvType::Adt {
field_types,
field_offsets,
..
} => {
let (i, field_ty, field_ty_kind, field_ty_size, offset_in_field) = field_offsets
.iter()
.enumerate()
.find_map(|(i, &field_offset)| {
if field_offset > offset {
return None;
}
// Grab the actual field type to be able to confirm that
// the leaf is somewhere inside the field.
let field_ty = field_types[i];
let field_ty_kind = self.lookup_type(field_ty);
let field_ty_size = field_ty_kind
.sizeof(self).map_or(MaybeSized::Unsized, MaybeSized::Sized);
let offset_in_field = offset - field_offset;
if MaybeSized::Sized(offset_in_field) < field_ty_size
// If the field is a zero sized type, check the
// expected size and type to get the correct entry
|| offset_in_field == Size::ZERO
&& leaf_size_range.contains(&MaybeSized::Sized(Size::ZERO)) && leaf_ty == Some(field_ty)
{
Some((i, field_ty, field_ty_kind, field_ty_size, offset_in_field))
} else {
None
}
})?;
ty = field_ty;
ty_kind = field_ty_kind;
ty_size = field_ty_size;
indices.push(i as u32);
offset = offset_in_field;
}
SpirvType::Vector { element, .. }
| SpirvType::Array { element, .. }
| SpirvType::RuntimeArray { element }
| SpirvType::Matrix { element, .. } => {
ty = element;
ty_kind = self.lookup_type(ty);
let stride = ty_kind.sizeof(self)?;
ty_size = MaybeSized::Sized(stride);
indices.push((offset.bytes() / stride.bytes()).try_into().ok()?);
offset = Size::from_bytes(offset.bytes() % stride.bytes());
}
_ => return None,
}
// Avoid digging beyond the point the leaf could actually fit.
if ty_size < *leaf_size_range.start() {
return None;
}
if offset == Size::ZERO
&& leaf_size_range.contains(&ty_size)
&& leaf_ty.map_or(true, |leaf_ty| leaf_ty == ty)
{
return Some((indices, ty));
}
}
}
fn maybe_inbounds_gep(
&mut self,
ty: Word,
ptr: SpirvValue,
combined_indices: &[SpirvValue],
is_inbounds: bool,
) -> SpirvValue {
let (&ptr_base_index, indices) = combined_indices.split_first().unwrap();
// The first index is an offset to the pointer, the rest are actual members.
// https://llvm.org/docs/GetElementPtr.html
// "An OpAccessChain instruction is the equivalent of an LLVM getelementptr instruction where the first index element is zero."
// https://github.com/gpuweb/gpuweb/issues/33
let mut result_pointee_type = ty;
let indices: Vec<_> = indices
.iter()
.map(|index| {
result_pointee_type = match self.lookup_type(result_pointee_type) {
SpirvType::Array { element, .. } | SpirvType::RuntimeArray { element } => {
element
}
_ => self.fatal(format!(
"GEP not implemented for type {}",
self.debug_type(result_pointee_type)
)),
};
index.def(self)
})
.collect();
// Special-case field accesses through a `pointercast`, to access the
// right field in the original type, for the `Logical` addressing model.
let ptr = ptr.strip_ptrcasts();
let ptr_id = ptr.def(self);
let original_pointee_ty = match self.lookup_type(ptr.ty) {
SpirvType::Pointer { pointee } => pointee,
other => self.fatal(format!("gep called on non-pointer type: {other:?}")),
};
// HACK(eddyb) `struct_gep` itself is falling out of use, as it's being
// replaced upstream by `ptr_add` (aka `inbounds_gep` with byte offsets).
//
// FIXME(eddyb) get rid of everything other than:
// - constant byte offset (`ptr_add`?)
// - dynamic indexing of a single array
let const_ptr_offset = self
.builder
.lookup_const_u64(ptr_base_index)
.and_then(|idx| Some(idx * self.lookup_type(ty).sizeof(self)?));
if let Some(const_ptr_offset) = const_ptr_offset {
if let Some((base_indices, base_pointee_ty)) = self.recover_access_chain_from_offset(
original_pointee_ty,
const_ptr_offset,
Some(Size::ZERO)..=None,
None,
) {
// FIXME(eddyb) this condition is pretty limiting, but
// eventually it shouldn't matter if GEPs are going away.
if ty == base_pointee_ty || indices.is_empty() {
let result_pointee_type = if indices.is_empty() {
base_pointee_ty
} else {
result_pointee_type
};
let indices = base_indices
.into_iter()
.map(|idx| self.constant_u32(self.span(), idx).def(self))
.chain(indices)
.collect();
return self.emit_access_chain(
self.type_ptr_to(result_pointee_type),
ptr_id,
None,
indices,
is_inbounds,
);
}
}
}
let result_type = self.type_ptr_to(result_pointee_type);
// Check if `ptr_id` is defined by an `OpAccessChain`, and if it is,
// grab its base pointer and indices.
//
// FIXME(eddyb) this could get ridiculously expensive, at the very least
// it could use `.rev()`, hoping the base pointer was recently defined?
let maybe_original_access_chain = if ty == original_pointee_ty {
let emit = self.emit();
let module = emit.module_ref();
let func = &module.functions[emit.selected_function().unwrap()];
let base_ptr_and_combined_indices = func
.all_inst_iter()
.find(|inst| inst.result_id == Some(ptr_id))
.and_then(|ptr_def_inst| {
if matches!(
ptr_def_inst.class.opcode,
Op::AccessChain | Op::InBoundsAccessChain
) {
let base_ptr = ptr_def_inst.operands[0].unwrap_id_ref();
let indices = ptr_def_inst.operands[1..]
.iter()
.map(|op| op.unwrap_id_ref())
.collect::<Vec<_>>();
Some((base_ptr, indices))
} else {
None
}
});
base_ptr_and_combined_indices
} else {
None
};
if let Some((original_ptr, mut original_indices)) = maybe_original_access_chain {
// Transform the following:
// OpAccessChain original_ptr [a, b, c]
// OpPtrAccessChain ptr base [d, e, f]
// into
// OpAccessChain original_ptr [a, b, c + base, d, e, f]
// to remove the need for OpPtrAccessChain
let last = original_indices.last_mut().unwrap();
*last = self
.add(last.with_type(ptr_base_index.ty), ptr_base_index)
.def(self);
original_indices.extend(indices);
return self.emit_access_chain(
result_type,
original_ptr,
None,
original_indices,
is_inbounds,
);
}
// HACK(eddyb) temporary workaround for untyped pointers upstream.
// FIXME(eddyb) replace with untyped memory SPIR-V + `qptr` or similar.
let ptr = self.pointercast(ptr, self.type_ptr_to(ty));
let ptr_id = ptr.def(self);
self.emit_access_chain(
result_type,
ptr_id,
Some(ptr_base_index),
indices,
is_inbounds,
)
}
fn emit_access_chain(
&self,
result_type: <Self as BackendTypes>::Type,
pointer: Word,
ptr_base_index: Option<SpirvValue>,
indices: Vec<Word>,
is_inbounds: bool,
) -> SpirvValue {
let mut emit = self.emit();
let non_zero_ptr_base_index =
ptr_base_index.filter(|&idx| self.builder.lookup_const_u64(idx) != Some(0));
if let Some(ptr_base_index) = non_zero_ptr_base_index {
let result = if is_inbounds {
emit.in_bounds_ptr_access_chain(
result_type,
None,
pointer,
ptr_base_index.def(self),
indices,
)
} else {
emit.ptr_access_chain(
result_type,
None,
pointer,
ptr_base_index.def(self),
indices,
)
}
.unwrap();
self.zombie(result, "cannot offset a pointer to an arbitrary element");
result
} else {
if is_inbounds {
emit.in_bounds_access_chain(result_type, None, pointer, indices)
} else {
emit.access_chain(result_type, None, pointer, indices)
}
.unwrap()
}
.with_type(result_type)
}
fn fptoint_sat(
&mut self,
signed: bool,
val: SpirvValue,
dest_ty: <Self as BackendTypes>::Type,
) -> SpirvValue {
// This uses the old llvm emulation to implement saturation
let src_ty = self.cx.val_ty(val);
let (float_ty, int_ty) = if self.cx.type_kind(src_ty) == TypeKind::Vector {
assert_eq!(
self.cx.vector_length(src_ty),
self.cx.vector_length(dest_ty)
);
(self.cx.element_type(src_ty), self.cx.element_type(dest_ty))
} else {
(src_ty, dest_ty)
};
let int_width = self.cx().int_width(int_ty);
let float_width = self.cx().float_width(float_ty);
// LLVM's fpto[su]i returns undef when the input x is infinite, NaN, or does not fit into the
// destination integer type after rounding towards zero. This `undef` value can cause UB in
// safe code (see issue #10184), so we implement a saturating conversion on top of it:
// Semantically, the mathematical value of the input is rounded towards zero to the next
// mathematical integer, and then the result is clamped into the range of the destination
// integer type. Positive and negative infinity are mapped to the maximum and minimum value of
// the destination integer type. NaN is mapped to 0.
//
// Define f_min and f_max as the largest and smallest (finite) floats that are exactly equal to
// a value representable in int_ty.
// They are exactly equal to int_ty::{MIN,MAX} if float_ty has enough significand bits.
// Otherwise, int_ty::MAX must be rounded towards zero, as it is one less than a power of two.
// int_ty::MIN, however, is either zero or a negative power of two and is thus exactly
// representable. Note that this only works if float_ty's exponent range is sufficiently large.
// f16 or 256 bit integers would break this property. Right now the smallest float type is f32
// with exponents ranging up to 127, which is barely enough for i128::MIN = -2^127.
// On the other hand, f_max works even if int_ty::MAX is greater than float_ty::MAX. Because
// we're rounding towards zero, we just get float_ty::MAX (which is always an integer).
// This already happens today with u128::MAX = 2^128 - 1 > f32::MAX.
let int_max = |signed: bool, int_width: u64| -> u128 {
let shift_amount = 128 - int_width;
if signed {
i128::MAX as u128 >> shift_amount
} else {
u128::MAX >> shift_amount
}
};
let int_min = |signed: bool, int_width: u64| -> i128 {
if signed {
i128::MIN >> (128 - int_width)
} else {
0
}
};
let compute_clamp_bounds_single = |signed: bool, int_width: u64| -> (u128, u128) {
let rounded_min =
ieee::Single::from_i128_r(int_min(signed, int_width), Round::TowardZero);
assert_eq!(rounded_min.status, Status::OK);
let rounded_max =
ieee::Single::from_u128_r(int_max(signed, int_width), Round::TowardZero);
assert!(rounded_max.value.is_finite());
(rounded_min.value.to_bits(), rounded_max.value.to_bits())
};
let compute_clamp_bounds_double = |signed: bool, int_width: u64| -> (u128, u128) {
let rounded_min =
ieee::Double::from_i128_r(int_min(signed, int_width), Round::TowardZero);
assert_eq!(rounded_min.status, Status::OK);
let rounded_max =
ieee::Double::from_u128_r(int_max(signed, int_width), Round::TowardZero);
assert!(rounded_max.value.is_finite());
(rounded_min.value.to_bits(), rounded_max.value.to_bits())
};
// To implement saturation, we perform the following steps:
//
// 1. Cast x to an integer with fpto[su]i. This may result in undef.
// 2. Compare x to f_min and f_max, and use the comparison results to select:
// a) int_ty::MIN if x < f_min or x is NaN
// b) int_ty::MAX if x > f_max
// c) the result of fpto[su]i otherwise
// 3. If x is NaN, return 0.0, otherwise return the result of step 2.
//
// This avoids resulting undef because values in range [f_min, f_max] by definition fit into the
// destination type. It creates an undef temporary, but *producing* undef is not UB. Our use of
// undef does not introduce any non-determinism either.
// More importantly, the above procedure correctly implements saturating conversion.
// Proof (sketch):
// If x is NaN, 0 is returned by definition.
// Otherwise, x is finite or infinite and thus can be compared with f_min and f_max.
// This yields three cases to consider:
// (1) if x in [f_min, f_max], the result of fpto[su]i is returned, which agrees with
// saturating conversion for inputs in that range.
// (2) if x > f_max, then x is larger than int_ty::MAX. This holds even if f_max is rounded
// (i.e., if f_max < int_ty::MAX) because in those cases, nextUp(f_max) is already larger
// than int_ty::MAX. Because x is larger than int_ty::MAX, the return value of int_ty::MAX
// is correct.
// (3) if x < f_min, then x is smaller than int_ty::MIN. As shown earlier, f_min exactly equals
// int_ty::MIN and therefore the return value of int_ty::MIN is correct.
// QED.
let float_bits_to_llval = |bx: &mut Self, bits| {
let bits_llval = match float_width {
32 => bx.cx().const_u32(bits as u32),
64 => bx.cx().const_u64(bits as u64),
n => bug!("unsupported float width {}", n),
};
bx.bitcast(bits_llval, float_ty)
};
let (f_min, f_max) = match float_width {
32 => compute_clamp_bounds_single(signed, int_width),
64 => compute_clamp_bounds_double(signed, int_width),
n => bug!("unsupported float width {}", n),
};
let f_min = float_bits_to_llval(self, f_min);
let f_max = float_bits_to_llval(self, f_max);
let int_max = self.cx().const_uint_big(int_ty, int_max(signed, int_width));
let int_min = self
.cx()
.const_uint_big(int_ty, int_min(signed, int_width) as u128);
let zero = self.cx().const_uint(int_ty, 0);
// If we're working with vectors, constants must be "splatted": the constant is duplicated
// into each lane of the vector. The algorithm stays the same, we are just using the
// same constant across all lanes.
let maybe_splat = |bx: &mut Self, val| {
if bx.cx().type_kind(dest_ty) == TypeKind::Vector {
bx.vector_splat(bx.vector_length(dest_ty), val)
} else {
val
}
};
let f_min = maybe_splat(self, f_min);
let f_max = maybe_splat(self, f_max);
let int_max = maybe_splat(self, int_max);
let int_min = maybe_splat(self, int_min);
let zero = maybe_splat(self, zero);
// Step 1 ...
let fptosui_result = if signed {
self.fptosi(val, dest_ty)
} else {
self.fptoui(val, dest_ty)
};
let less_or_nan = self.fcmp(RealPredicate::RealULT, val, f_min);
let greater = self.fcmp(RealPredicate::RealOGT, val, f_max);
// Step 2: We use two comparisons and two selects, with %s1 being the
// result:
// %less_or_nan = fcmp ult %x, %f_min
// %greater = fcmp olt %x, %f_max
// %s0 = select %less_or_nan, int_ty::MIN, %fptosi_result
// %s1 = select %greater, int_ty::MAX, %s0
// Note that %less_or_nan uses an *unordered* comparison. This
// comparison is true if the operands are not comparable (i.e., if x is
// NaN). The unordered comparison ensures that s1 becomes int_ty::MIN if
// x is NaN.
//
// Performance note: Unordered comparison can be lowered to a "flipped"
// comparison and a negation, and the negation can be merged into the
// select. Therefore, it not necessarily any more expensive than an
// ordered ("normal") comparison. Whether these optimizations will be
// performed is ultimately up to the backend, but at least x86 does
// perform them.
let s0 = self.select(less_or_nan, int_min, fptosui_result);
let s1 = self.select(greater, int_max, s0);
// Step 3: NaN replacement.
// For unsigned types, the above step already yielded int_ty::MIN == 0 if x is NaN.
// Therefore we only need to execute this step for signed integer types.
if signed {
// LLVM has no isNaN predicate, so we use (x == x) instead
let cmp = self.fcmp(RealPredicate::RealOEQ, val, val);
self.select(cmp, s1, zero)
} else {
s1
}
}
}
impl<'a, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'tcx> {
fn build(cx: &'a Self::CodegenCx, llbb: Self::BasicBlock) -> Self {
let cursor = cx.builder.select_block_by_id(llbb);
// FIXME(eddyb) change `Self::Function` to be more like a function index.
let current_fn = {
let emit = cx.emit_with_cursor(cursor);
let selected_function = emit.selected_function().unwrap();
let selected_function = &emit.module_ref().functions[selected_function];
let def_inst = selected_function.def.as_ref().unwrap();
let def = def_inst.result_id.unwrap();
let ty = def_inst.operands[1].unwrap_id_ref();
def.with_type(ty)
};
Self {
cx,
cursor,
current_fn,
basic_block: llbb,
current_span: Default::default(),
}
}
fn cx(&self) -> &Self::CodegenCx {
self.cx
}
fn llbb(&self) -> Self::BasicBlock {
self.basic_block
}
fn set_span(&mut self, span: Span) {
// HACK(eddyb) this is what `#[track_caller]` does, and we need it to be
// able to point at e.g. a use of `panic!`, instead of its implementation,
// but it should be more fine-grained and/or include macro backtraces in
// debuginfo (so the decision to use them can be deferred).
let span = span.ctxt().outer_expn().expansion_cause().unwrap_or(span);
let old_span = self.current_span.replace(span);
// FIXME(eddyb) enable this once cross-block interactions are figured out
// (in particular, every block starts off with no debuginfo active).
if false {
// Avoid redundant debuginfo.
if old_span == Some(span) {
return;
}
}
// HACK(eddyb) this is only to aid testing (and to not remove the old code).
let use_custom_insts = true;
if use_custom_insts {
// FIXME(eddyb) this should be cached more efficiently.
let void_ty = SpirvType::Void.def(rustc_span::DUMMY_SP, self);
// We may not always have valid spans.
// FIXME(eddyb) reduce the sources of this as much as possible.
if span.is_dummy() {
self.custom_inst(void_ty, CustomInst::ClearDebugSrcLoc);
} else {
let (file, line_col_range) = self.builder.file_line_col_range_for_debuginfo(span);
let ((line_start, col_start), (line_end, col_end)) =
(line_col_range.start, line_col_range.end);
self.custom_inst(
void_ty,
CustomInst::SetDebugSrcLoc {
file: Operand::IdRef(file.file_name_op_string_id),
line_start: Operand::IdRef(self.const_u32(line_start).def(self)),
line_end: Operand::IdRef(self.const_u32(line_end).def(self)),
col_start: Operand::IdRef(self.const_u32(col_start).def(self)),
col_end: Operand::IdRef(self.const_u32(col_end).def(self)),
},
);
}
// HACK(eddyb) remove the previous instruction if made irrelevant.
let mut builder = self.emit();
if let (Some(func_idx), Some(block_idx)) =