2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
28 #include "vtn_private.h"
29 #include "spirv_info.h"
30 #include "nir_deref.h"
32 static struct vtn_access_chain
*
33 vtn_access_chain_create(struct vtn_builder
*b
, unsigned length
)
35 struct vtn_access_chain
*chain
;
37 /* Subtract 1 from the length since there's already one built in */
38 size_t size
= sizeof(*chain
) +
39 (MAX2(length
, 1) - 1) * sizeof(chain
->link
[0]);
40 chain
= rzalloc_size(b
, size
);
41 chain
->length
= length
;
46 static struct vtn_access_chain
*
47 vtn_access_chain_extend(struct vtn_builder
*b
, struct vtn_access_chain
*old
,
50 struct vtn_access_chain
*chain
;
52 unsigned old_len
= old
? old
->length
: 0;
53 chain
= vtn_access_chain_create(b
, old_len
+ new_ids
);
55 for (unsigned i
= 0; i
< old_len
; i
++)
56 chain
->link
[i
] = old
->link
[i
];
62 vtn_pointer_uses_ssa_offset(struct vtn_builder
*b
,
63 struct vtn_pointer
*ptr
)
65 return ptr
->mode
== vtn_variable_mode_ubo
||
66 ptr
->mode
== vtn_variable_mode_ssbo
||
67 ptr
->mode
== vtn_variable_mode_push_constant
||
68 (ptr
->mode
== vtn_variable_mode_workgroup
&&
69 b
->options
->lower_workgroup_access_to_offsets
);
73 vtn_pointer_is_external_block(struct vtn_builder
*b
,
74 struct vtn_pointer
*ptr
)
76 return ptr
->mode
== vtn_variable_mode_ssbo
||
77 ptr
->mode
== vtn_variable_mode_ubo
||
78 ptr
->mode
== vtn_variable_mode_push_constant
||
79 (ptr
->mode
== vtn_variable_mode_workgroup
&&
80 b
->options
->lower_workgroup_access_to_offsets
);
83 /* Dereference the given base pointer by the access chain */
84 static struct vtn_pointer
*
85 vtn_access_chain_pointer_dereference(struct vtn_builder
*b
,
86 struct vtn_pointer
*base
,
87 struct vtn_access_chain
*deref_chain
)
89 struct vtn_access_chain
*chain
=
90 vtn_access_chain_extend(b
, base
->chain
, deref_chain
->length
);
91 struct vtn_type
*type
= base
->type
;
93 /* OpPtrAccessChain is only allowed on things which support variable
94 * pointers. For everything else, the client is expected to just pass us
95 * the right access chain.
97 vtn_assert(!deref_chain
->ptr_as_array
);
99 unsigned start
= base
->chain
? base
->chain
->length
: 0;
100 for (unsigned i
= 0; i
< deref_chain
->length
; i
++) {
101 chain
->link
[start
+ i
] = deref_chain
->link
[i
];
103 if (glsl_type_is_struct(type
->type
)) {
104 vtn_assert(deref_chain
->link
[i
].mode
== vtn_access_mode_literal
);
105 type
= type
->members
[deref_chain
->link
[i
].id
];
107 type
= type
->array_element
;
111 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
112 ptr
->mode
= base
->mode
;
114 ptr
->var
= base
->var
;
115 ptr
->deref
= base
->deref
;
122 vtn_access_link_as_ssa(struct vtn_builder
*b
, struct vtn_access_link link
,
125 vtn_assert(stride
> 0);
126 if (link
.mode
== vtn_access_mode_literal
) {
127 return nir_imm_int(&b
->nb
, link
.id
* stride
);
128 } else if (stride
== 1) {
129 nir_ssa_def
*ssa
= vtn_ssa_value(b
, link
.id
)->def
;
130 if (ssa
->bit_size
!= 32)
131 ssa
= nir_u2u32(&b
->nb
, ssa
);
134 nir_ssa_def
*src0
= vtn_ssa_value(b
, link
.id
)->def
;
135 if (src0
->bit_size
!= 32)
136 src0
= nir_u2u32(&b
->nb
, src0
);
137 return nir_imul(&b
->nb
, src0
, nir_imm_int(&b
->nb
, stride
));
142 vtn_variable_resource_index(struct vtn_builder
*b
, struct vtn_variable
*var
,
143 nir_ssa_def
*desc_array_index
)
145 if (!desc_array_index
) {
146 vtn_assert(glsl_type_is_struct(var
->type
->type
));
147 desc_array_index
= nir_imm_int(&b
->nb
, 0);
150 nir_intrinsic_instr
*instr
=
151 nir_intrinsic_instr_create(b
->nb
.shader
,
152 nir_intrinsic_vulkan_resource_index
);
153 instr
->src
[0] = nir_src_for_ssa(desc_array_index
);
154 nir_intrinsic_set_desc_set(instr
, var
->descriptor_set
);
155 nir_intrinsic_set_binding(instr
, var
->binding
);
157 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
158 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
160 return &instr
->dest
.ssa
;
164 vtn_resource_reindex(struct vtn_builder
*b
, nir_ssa_def
*base_index
,
165 nir_ssa_def
*offset_index
)
167 nir_intrinsic_instr
*instr
=
168 nir_intrinsic_instr_create(b
->nb
.shader
,
169 nir_intrinsic_vulkan_resource_reindex
);
170 instr
->src
[0] = nir_src_for_ssa(base_index
);
171 instr
->src
[1] = nir_src_for_ssa(offset_index
);
173 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
174 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
176 return &instr
->dest
.ssa
;
179 static struct vtn_pointer
*
180 vtn_ssa_offset_pointer_dereference(struct vtn_builder
*b
,
181 struct vtn_pointer
*base
,
182 struct vtn_access_chain
*deref_chain
)
184 nir_ssa_def
*block_index
= base
->block_index
;
185 nir_ssa_def
*offset
= base
->offset
;
186 struct vtn_type
*type
= base
->type
;
189 if (base
->mode
== vtn_variable_mode_ubo
||
190 base
->mode
== vtn_variable_mode_ssbo
) {
192 vtn_assert(base
->var
&& base
->type
);
193 nir_ssa_def
*desc_arr_idx
;
194 if (glsl_type_is_array(type
->type
)) {
195 if (deref_chain
->length
>= 1) {
197 vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1);
199 /* This consumes a level of type */
200 type
= type
->array_element
;
202 /* This is annoying. We've been asked for a pointer to the
203 * array of UBOs/SSBOs and not a specifc buffer. Return a
204 * pointer with a descriptor index of 0 and we'll have to do
205 * a reindex later to adjust it to the right thing.
207 desc_arr_idx
= nir_imm_int(&b
->nb
, 0);
209 } else if (deref_chain
->ptr_as_array
) {
210 /* You can't have a zero-length OpPtrAccessChain */
211 vtn_assert(deref_chain
->length
>= 1);
212 desc_arr_idx
= vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1);
214 /* We have a regular non-array SSBO. */
217 block_index
= vtn_variable_resource_index(b
, base
->var
, desc_arr_idx
);
218 } else if (deref_chain
->ptr_as_array
&&
219 type
->base_type
== vtn_base_type_struct
&& type
->block
) {
220 /* We are doing an OpPtrAccessChain on a pointer to a struct that is
221 * decorated block. This is an interesting corner in the SPIR-V
222 * spec. One interpretation would be that they client is clearly
223 * trying to treat that block as if it's an implicit array of blocks
224 * repeated in the buffer. However, the SPIR-V spec for the
225 * OpPtrAccessChain says:
227 * "Base is treated as the address of the first element of an
228 * array, and the Element element’s address is computed to be the
229 * base for the Indexes, as per OpAccessChain."
231 * Taken literally, that would mean that your struct type is supposed
232 * to be treated as an array of such a struct and, since it's
233 * decorated block, that means an array of blocks which corresponds
234 * to an array descriptor. Therefore, we need to do a reindex
235 * operation to add the index from the first link in the access chain
236 * to the index we recieved.
238 * The downside to this interpretation (there always is one) is that
239 * this might be somewhat surprising behavior to apps if they expect
240 * the implicit array behavior described above.
242 vtn_assert(deref_chain
->length
>= 1);
243 nir_ssa_def
*offset_index
=
244 vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1);
247 block_index
= vtn_resource_reindex(b
, block_index
, offset_index
);
252 if (base
->mode
== vtn_variable_mode_workgroup
) {
253 /* SLM doesn't need nor have a block index */
254 vtn_assert(!block_index
);
256 /* We need the variable for the base offset */
257 vtn_assert(base
->var
);
259 /* We need ptr_type for size and alignment */
260 vtn_assert(base
->ptr_type
);
262 /* Assign location on first use so that we don't end up bloating SLM
263 * address space for variables which are never statically used.
265 if (base
->var
->shared_location
< 0) {
266 vtn_assert(base
->ptr_type
->length
> 0 && base
->ptr_type
->align
> 0);
267 b
->shader
->num_shared
= vtn_align_u32(b
->shader
->num_shared
,
268 base
->ptr_type
->align
);
269 base
->var
->shared_location
= b
->shader
->num_shared
;
270 b
->shader
->num_shared
+= base
->ptr_type
->length
;
273 offset
= nir_imm_int(&b
->nb
, base
->var
->shared_location
);
274 } else if (base
->mode
== vtn_variable_mode_push_constant
) {
275 /* Push constants neither need nor have a block index */
276 vtn_assert(!block_index
);
278 /* Start off with at the start of the push constant block. */
279 offset
= nir_imm_int(&b
->nb
, 0);
281 /* The code above should have ensured a block_index when needed. */
282 vtn_assert(block_index
);
284 /* Start off with at the start of the buffer. */
285 offset
= nir_imm_int(&b
->nb
, 0);
289 if (deref_chain
->ptr_as_array
&& idx
== 0) {
290 /* We need ptr_type for the stride */
291 vtn_assert(base
->ptr_type
);
293 /* We need at least one element in the chain */
294 vtn_assert(deref_chain
->length
>= 1);
296 nir_ssa_def
*elem_offset
=
297 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
298 base
->ptr_type
->stride
);
299 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
303 for (; idx
< deref_chain
->length
; idx
++) {
304 switch (glsl_get_base_type(type
->type
)) {
307 case GLSL_TYPE_UINT16
:
308 case GLSL_TYPE_INT16
:
309 case GLSL_TYPE_UINT8
:
311 case GLSL_TYPE_UINT64
:
312 case GLSL_TYPE_INT64
:
313 case GLSL_TYPE_FLOAT
:
314 case GLSL_TYPE_FLOAT16
:
315 case GLSL_TYPE_DOUBLE
:
317 case GLSL_TYPE_ARRAY
: {
318 nir_ssa_def
*elem_offset
=
319 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
], type
->stride
);
320 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
321 type
= type
->array_element
;
325 case GLSL_TYPE_STRUCT
: {
326 vtn_assert(deref_chain
->link
[idx
].mode
== vtn_access_mode_literal
);
327 unsigned member
= deref_chain
->link
[idx
].id
;
328 nir_ssa_def
*mem_offset
= nir_imm_int(&b
->nb
, type
->offsets
[member
]);
329 offset
= nir_iadd(&b
->nb
, offset
, mem_offset
);
330 type
= type
->members
[member
];
335 vtn_fail("Invalid type for deref");
339 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
340 ptr
->mode
= base
->mode
;
342 ptr
->block_index
= block_index
;
343 ptr
->offset
= offset
;
348 /* Dereference the given base pointer by the access chain */
349 static struct vtn_pointer
*
350 vtn_pointer_dereference(struct vtn_builder
*b
,
351 struct vtn_pointer
*base
,
352 struct vtn_access_chain
*deref_chain
)
354 if (vtn_pointer_uses_ssa_offset(b
, base
)) {
355 return vtn_ssa_offset_pointer_dereference(b
, base
, deref_chain
);
357 return vtn_access_chain_pointer_dereference(b
, base
, deref_chain
);
362 vtn_pointer_for_variable(struct vtn_builder
*b
,
363 struct vtn_variable
*var
, struct vtn_type
*ptr_type
)
365 struct vtn_pointer
*pointer
= rzalloc(b
, struct vtn_pointer
);
367 pointer
->mode
= var
->mode
;
368 pointer
->type
= var
->type
;
369 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
370 vtn_assert(ptr_type
->deref
->type
== var
->type
->type
);
371 pointer
->ptr_type
= ptr_type
;
377 /* Returns an atomic_uint type based on the original uint type. The returned
378 * type will be equivalent to the original one but will have an atomic_uint
379 * type as leaf instead of an uint.
381 * Manages uint scalars, arrays, and arrays of arrays of any nested depth.
383 static const struct glsl_type
*
384 repair_atomic_type(const struct glsl_type
*type
)
386 assert(glsl_get_base_type(glsl_without_array(type
)) == GLSL_TYPE_UINT
);
387 assert(glsl_type_is_scalar(glsl_without_array(type
)));
389 if (glsl_type_is_array(type
)) {
390 const struct glsl_type
*atomic
=
391 repair_atomic_type(glsl_get_array_element(type
));
393 return glsl_array_type(atomic
, glsl_get_length(type
));
395 return glsl_atomic_uint_type();
400 vtn_pointer_to_deref(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
402 /* Do on-the-fly copy propagation for samplers. */
403 if (ptr
->var
&& ptr
->var
->copy_prop_sampler
)
404 return vtn_pointer_to_deref(b
, ptr
->var
->copy_prop_sampler
);
406 nir_deref_instr
*tail
;
410 assert(ptr
->var
&& ptr
->var
->var
);
411 tail
= nir_build_deref_var(&b
->nb
, ptr
->var
->var
);
414 /* Raw variable access */
418 struct vtn_access_chain
*chain
= ptr
->chain
;
421 for (unsigned i
= 0; i
< chain
->length
; i
++) {
422 if (glsl_type_is_struct(tail
->type
)) {
423 vtn_assert(chain
->link
[i
].mode
== vtn_access_mode_literal
);
424 unsigned idx
= chain
->link
[i
].id
;
425 tail
= nir_build_deref_struct(&b
->nb
, tail
, idx
);
428 if (chain
->link
[i
].mode
== vtn_access_mode_literal
) {
429 index
= nir_imm_int(&b
->nb
, chain
->link
[i
].id
);
431 vtn_assert(chain
->link
[i
].mode
== vtn_access_mode_id
);
432 index
= vtn_ssa_value(b
, chain
->link
[i
].id
)->def
;
434 tail
= nir_build_deref_array(&b
->nb
, tail
, index
);
442 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_instr
*deref
,
443 struct vtn_ssa_value
*inout
)
445 if (glsl_type_is_vector_or_scalar(deref
->type
)) {
447 inout
->def
= nir_load_deref(&b
->nb
, deref
);
449 nir_store_deref(&b
->nb
, deref
, inout
->def
, ~0);
451 } else if (glsl_type_is_array(deref
->type
) ||
452 glsl_type_is_matrix(deref
->type
)) {
453 unsigned elems
= glsl_get_length(deref
->type
);
454 for (unsigned i
= 0; i
< elems
; i
++) {
455 nir_deref_instr
*child
=
456 nir_build_deref_array(&b
->nb
, deref
, nir_imm_int(&b
->nb
, i
));
457 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
]);
460 vtn_assert(glsl_type_is_struct(deref
->type
));
461 unsigned elems
= glsl_get_length(deref
->type
);
462 for (unsigned i
= 0; i
< elems
; i
++) {
463 nir_deref_instr
*child
= nir_build_deref_struct(&b
->nb
, deref
, i
);
464 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
]);
470 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
472 struct vtn_pointer
*ptr
= vtn_value(b
, id
, vtn_value_type_pointer
)->pointer
;
473 return vtn_pointer_to_deref(b
, ptr
);
477 * Gets the NIR-level deref tail, which may have as a child an array deref
478 * selecting which component due to OpAccessChain supporting per-component
479 * indexing in SPIR-V.
481 static nir_deref_instr
*
482 get_deref_tail(nir_deref_instr
*deref
)
484 if (deref
->deref_type
!= nir_deref_type_array
)
487 nir_deref_instr
*parent
=
488 nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
);
490 if (glsl_type_is_vector(parent
->type
))
496 struct vtn_ssa_value
*
497 vtn_local_load(struct vtn_builder
*b
, nir_deref_instr
*src
)
499 nir_deref_instr
*src_tail
= get_deref_tail(src
);
500 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
501 _vtn_local_load_store(b
, true, src_tail
, val
);
503 if (src_tail
!= src
) {
504 val
->type
= src
->type
;
505 nir_const_value
*const_index
= nir_src_as_const_value(src
->arr
.index
);
507 val
->def
= vtn_vector_extract(b
, val
->def
, const_index
->u32
[0]);
509 val
->def
= vtn_vector_extract_dynamic(b
, val
->def
, src
->arr
.index
.ssa
);
516 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
517 nir_deref_instr
*dest
)
519 nir_deref_instr
*dest_tail
= get_deref_tail(dest
);
521 if (dest_tail
!= dest
) {
522 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
523 _vtn_local_load_store(b
, true, dest_tail
, val
);
525 nir_const_value
*const_index
= nir_src_as_const_value(dest
->arr
.index
);
527 val
->def
= vtn_vector_insert(b
, val
->def
, src
->def
,
528 const_index
->u32
[0]);
530 val
->def
= vtn_vector_insert_dynamic(b
, val
->def
, src
->def
,
531 dest
->arr
.index
.ssa
);
532 _vtn_local_load_store(b
, false, dest_tail
, val
);
534 _vtn_local_load_store(b
, false, dest_tail
, src
);
539 vtn_pointer_to_offset(struct vtn_builder
*b
, struct vtn_pointer
*ptr
,
540 nir_ssa_def
**index_out
)
542 assert(vtn_pointer_uses_ssa_offset(b
, ptr
));
544 struct vtn_access_chain chain
= {
547 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
549 *index_out
= ptr
->block_index
;
553 /* Tries to compute the size of an interface block based on the strides and
554 * offsets that are provided to us in the SPIR-V source.
557 vtn_type_block_size(struct vtn_builder
*b
, struct vtn_type
*type
)
559 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
563 case GLSL_TYPE_UINT16
:
564 case GLSL_TYPE_INT16
:
565 case GLSL_TYPE_UINT8
:
567 case GLSL_TYPE_UINT64
:
568 case GLSL_TYPE_INT64
:
569 case GLSL_TYPE_FLOAT
:
570 case GLSL_TYPE_FLOAT16
:
572 case GLSL_TYPE_DOUBLE
: {
573 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
574 glsl_get_matrix_columns(type
->type
);
576 vtn_assert(type
->stride
> 0);
577 return type
->stride
* cols
;
579 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
580 return glsl_get_vector_elements(type
->type
) * type_size
;
584 case GLSL_TYPE_STRUCT
:
585 case GLSL_TYPE_INTERFACE
: {
587 unsigned num_fields
= glsl_get_length(type
->type
);
588 for (unsigned f
= 0; f
< num_fields
; f
++) {
589 unsigned field_end
= type
->offsets
[f
] +
590 vtn_type_block_size(b
, type
->members
[f
]);
591 size
= MAX2(size
, field_end
);
596 case GLSL_TYPE_ARRAY
:
597 vtn_assert(type
->stride
> 0);
598 vtn_assert(glsl_get_length(type
->type
) > 0);
599 return type
->stride
* glsl_get_length(type
->type
);
602 vtn_fail("Invalid block type");
608 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
609 nir_ssa_def
*index
, nir_ssa_def
*offset
,
610 unsigned access_offset
, unsigned access_size
,
611 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
)
613 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
614 instr
->num_components
= glsl_get_vector_elements(type
);
618 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
619 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
622 if (op
== nir_intrinsic_load_push_constant
) {
623 nir_intrinsic_set_base(instr
, access_offset
);
624 nir_intrinsic_set_range(instr
, access_size
);
628 instr
->src
[src
++] = nir_src_for_ssa(index
);
630 if (op
== nir_intrinsic_load_push_constant
) {
631 /* We need to subtract the offset from where the intrinsic will load the
634 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
635 nir_imm_int(&b
->nb
, access_offset
)));
637 instr
->src
[src
++] = nir_src_for_ssa(offset
);
641 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
642 instr
->num_components
,
643 glsl_get_bit_size(type
), NULL
);
644 (*inout
)->def
= &instr
->dest
.ssa
;
647 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
649 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
650 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
654 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
655 nir_ssa_def
*index
, nir_ssa_def
*offset
,
656 unsigned access_offset
, unsigned access_size
,
657 struct vtn_type
*type
, struct vtn_ssa_value
**inout
)
659 if (load
&& *inout
== NULL
)
660 *inout
= vtn_create_ssa_value(b
, type
->type
);
662 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
666 case GLSL_TYPE_UINT16
:
667 case GLSL_TYPE_INT16
:
668 case GLSL_TYPE_UINT8
:
670 case GLSL_TYPE_UINT64
:
671 case GLSL_TYPE_INT64
:
672 case GLSL_TYPE_FLOAT
:
673 case GLSL_TYPE_FLOAT16
:
674 case GLSL_TYPE_DOUBLE
:
676 /* This is where things get interesting. At this point, we've hit
677 * a vector, a scalar, or a matrix.
679 if (glsl_type_is_matrix(type
->type
)) {
680 /* Loading the whole matrix */
681 struct vtn_ssa_value
*transpose
;
682 unsigned num_ops
, vec_width
, col_stride
;
683 if (type
->row_major
) {
684 num_ops
= glsl_get_vector_elements(type
->type
);
685 vec_width
= glsl_get_matrix_columns(type
->type
);
686 col_stride
= type
->array_element
->stride
;
688 const struct glsl_type
*transpose_type
=
689 glsl_matrix_type(base_type
, vec_width
, num_ops
);
690 *inout
= vtn_create_ssa_value(b
, transpose_type
);
692 transpose
= vtn_ssa_transpose(b
, *inout
);
696 num_ops
= glsl_get_matrix_columns(type
->type
);
697 vec_width
= glsl_get_vector_elements(type
->type
);
698 col_stride
= type
->stride
;
701 for (unsigned i
= 0; i
< num_ops
; i
++) {
702 nir_ssa_def
*elem_offset
=
703 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, i
* col_stride
));
704 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
705 access_offset
, access_size
,
707 glsl_vector_type(base_type
, vec_width
));
710 if (load
&& type
->row_major
)
711 *inout
= vtn_ssa_transpose(b
, *inout
);
713 unsigned elems
= glsl_get_vector_elements(type
->type
);
714 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
715 if (elems
== 1 || type
->stride
== type_size
) {
716 /* This is a tightly-packed normal scalar or vector load */
717 vtn_assert(glsl_type_is_vector_or_scalar(type
->type
));
718 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
719 access_offset
, access_size
,
722 /* This is a strided load. We have to load N things separately.
723 * This is the single column of a row-major matrix case.
725 vtn_assert(type
->stride
> type_size
);
726 vtn_assert(type
->stride
% type_size
== 0);
728 nir_ssa_def
*per_comp
[4];
729 for (unsigned i
= 0; i
< elems
; i
++) {
730 nir_ssa_def
*elem_offset
=
731 nir_iadd(&b
->nb
, offset
,
732 nir_imm_int(&b
->nb
, i
* type
->stride
));
733 struct vtn_ssa_value
*comp
, temp_val
;
735 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
736 temp_val
.type
= glsl_scalar_type(base_type
);
739 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
740 access_offset
, access_size
,
741 &comp
, glsl_scalar_type(base_type
));
742 per_comp
[i
] = comp
->def
;
747 *inout
= vtn_create_ssa_value(b
, type
->type
);
748 (*inout
)->def
= nir_vec(&b
->nb
, per_comp
, elems
);
754 case GLSL_TYPE_ARRAY
: {
755 unsigned elems
= glsl_get_length(type
->type
);
756 for (unsigned i
= 0; i
< elems
; i
++) {
757 nir_ssa_def
*elem_off
=
758 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, i
* type
->stride
));
759 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
760 access_offset
, access_size
,
761 type
->array_element
, &(*inout
)->elems
[i
]);
766 case GLSL_TYPE_STRUCT
: {
767 unsigned elems
= glsl_get_length(type
->type
);
768 for (unsigned i
= 0; i
< elems
; i
++) {
769 nir_ssa_def
*elem_off
=
770 nir_iadd(&b
->nb
, offset
, nir_imm_int(&b
->nb
, type
->offsets
[i
]));
771 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
772 access_offset
, access_size
,
773 type
->members
[i
], &(*inout
)->elems
[i
]);
779 vtn_fail("Invalid block member type");
783 static struct vtn_ssa_value
*
784 vtn_block_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
787 unsigned access_offset
= 0, access_size
= 0;
789 case vtn_variable_mode_ubo
:
790 op
= nir_intrinsic_load_ubo
;
792 case vtn_variable_mode_ssbo
:
793 op
= nir_intrinsic_load_ssbo
;
795 case vtn_variable_mode_push_constant
:
796 op
= nir_intrinsic_load_push_constant
;
797 access_size
= b
->shader
->num_uniforms
;
799 case vtn_variable_mode_workgroup
:
800 op
= nir_intrinsic_load_shared
;
803 vtn_fail("Invalid block variable mode");
806 nir_ssa_def
*offset
, *index
= NULL
;
807 offset
= vtn_pointer_to_offset(b
, src
, &index
);
809 struct vtn_ssa_value
*value
= NULL
;
810 _vtn_block_load_store(b
, op
, true, index
, offset
,
811 access_offset
, access_size
,
817 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
818 struct vtn_pointer
*dst
)
822 case vtn_variable_mode_ssbo
:
823 op
= nir_intrinsic_store_ssbo
;
825 case vtn_variable_mode_workgroup
:
826 op
= nir_intrinsic_store_shared
;
829 vtn_fail("Invalid block variable mode");
832 nir_ssa_def
*offset
, *index
= NULL
;
833 offset
= vtn_pointer_to_offset(b
, dst
, &index
);
835 _vtn_block_load_store(b
, op
, false, index
, offset
,
836 0, 0, dst
->type
, &src
);
840 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
841 struct vtn_pointer
*ptr
,
842 struct vtn_ssa_value
**inout
)
844 enum glsl_base_type base_type
= glsl_get_base_type(ptr
->type
->type
);
848 case GLSL_TYPE_UINT16
:
849 case GLSL_TYPE_INT16
:
850 case GLSL_TYPE_UINT8
:
852 case GLSL_TYPE_UINT64
:
853 case GLSL_TYPE_INT64
:
854 case GLSL_TYPE_FLOAT
:
855 case GLSL_TYPE_FLOAT16
:
857 case GLSL_TYPE_DOUBLE
:
858 /* At this point, we have a scalar, vector, or matrix so we know that
859 * there cannot be any structure splitting still in the way. By
860 * stopping at the matrix level rather than the vector level, we
861 * ensure that matrices get loaded in the optimal way even if they
862 * are storred row-major in a UBO.
865 *inout
= vtn_local_load(b
, vtn_pointer_to_deref(b
, ptr
));
867 vtn_local_store(b
, *inout
, vtn_pointer_to_deref(b
, ptr
));
871 case GLSL_TYPE_ARRAY
:
872 case GLSL_TYPE_STRUCT
: {
873 unsigned elems
= glsl_get_length(ptr
->type
->type
);
875 vtn_assert(*inout
== NULL
);
876 *inout
= rzalloc(b
, struct vtn_ssa_value
);
877 (*inout
)->type
= ptr
->type
->type
;
878 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
881 struct vtn_access_chain chain
= {
884 { .mode
= vtn_access_mode_literal
, },
887 for (unsigned i
= 0; i
< elems
; i
++) {
888 chain
.link
[0].id
= i
;
889 struct vtn_pointer
*elem
= vtn_pointer_dereference(b
, ptr
, &chain
);
890 _vtn_variable_load_store(b
, load
, elem
, &(*inout
)->elems
[i
]);
896 vtn_fail("Invalid access chain type");
900 struct vtn_ssa_value
*
901 vtn_variable_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
903 if (vtn_pointer_is_external_block(b
, src
)) {
904 return vtn_block_load(b
, src
);
906 struct vtn_ssa_value
*val
= NULL
;
907 _vtn_variable_load_store(b
, true, src
, &val
);
913 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
914 struct vtn_pointer
*dest
)
916 if (vtn_pointer_is_external_block(b
, dest
)) {
917 vtn_assert(dest
->mode
== vtn_variable_mode_ssbo
||
918 dest
->mode
== vtn_variable_mode_workgroup
);
919 vtn_block_store(b
, src
, dest
);
921 _vtn_variable_load_store(b
, false, dest
, &src
);
926 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
927 struct vtn_pointer
*src
)
929 vtn_assert(src
->type
->type
== dest
->type
->type
);
930 enum glsl_base_type base_type
= glsl_get_base_type(src
->type
->type
);
934 case GLSL_TYPE_UINT16
:
935 case GLSL_TYPE_INT16
:
936 case GLSL_TYPE_UINT8
:
938 case GLSL_TYPE_UINT64
:
939 case GLSL_TYPE_INT64
:
940 case GLSL_TYPE_FLOAT
:
941 case GLSL_TYPE_FLOAT16
:
942 case GLSL_TYPE_DOUBLE
:
944 /* At this point, we have a scalar, vector, or matrix so we know that
945 * there cannot be any structure splitting still in the way. By
946 * stopping at the matrix level rather than the vector level, we
947 * ensure that matrices get loaded in the optimal way even if they
948 * are storred row-major in a UBO.
950 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
953 case GLSL_TYPE_ARRAY
:
954 case GLSL_TYPE_STRUCT
: {
955 struct vtn_access_chain chain
= {
958 { .mode
= vtn_access_mode_literal
, },
961 unsigned elems
= glsl_get_length(src
->type
->type
);
962 for (unsigned i
= 0; i
< elems
; i
++) {
963 chain
.link
[0].id
= i
;
964 struct vtn_pointer
*src_elem
=
965 vtn_pointer_dereference(b
, src
, &chain
);
966 struct vtn_pointer
*dest_elem
=
967 vtn_pointer_dereference(b
, dest
, &chain
);
969 _vtn_variable_copy(b
, dest_elem
, src_elem
);
975 vtn_fail("Invalid access chain type");
980 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
981 struct vtn_pointer
*src
)
983 /* TODO: At some point, we should add a special-case for when we can
984 * just emit a copy_var intrinsic.
986 _vtn_variable_copy(b
, dest
, src
);
990 set_mode_system_value(struct vtn_builder
*b
, nir_variable_mode
*mode
)
992 vtn_assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
993 *mode
= nir_var_system_value
;
997 vtn_get_builtin_location(struct vtn_builder
*b
,
998 SpvBuiltIn builtin
, int *location
,
999 nir_variable_mode
*mode
)
1002 case SpvBuiltInPosition
:
1003 *location
= VARYING_SLOT_POS
;
1005 case SpvBuiltInPointSize
:
1006 *location
= VARYING_SLOT_PSIZ
;
1008 case SpvBuiltInClipDistance
:
1009 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
1011 case SpvBuiltInCullDistance
:
1012 *location
= VARYING_SLOT_CULL_DIST0
;
1014 case SpvBuiltInVertexIndex
:
1015 *location
= SYSTEM_VALUE_VERTEX_ID
;
1016 set_mode_system_value(b
, mode
);
1018 case SpvBuiltInVertexId
:
1019 /* Vulkan defines VertexID to be zero-based and reserves the new
1020 * builtin keyword VertexIndex to indicate the non-zero-based value.
1022 *location
= SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
1023 set_mode_system_value(b
, mode
);
1025 case SpvBuiltInInstanceIndex
:
1026 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
1027 set_mode_system_value(b
, mode
);
1029 case SpvBuiltInInstanceId
:
1030 *location
= SYSTEM_VALUE_INSTANCE_ID
;
1031 set_mode_system_value(b
, mode
);
1033 case SpvBuiltInPrimitiveId
:
1034 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1035 vtn_assert(*mode
== nir_var_shader_in
);
1036 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1037 } else if (*mode
== nir_var_shader_out
) {
1038 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1040 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
1041 set_mode_system_value(b
, mode
);
1044 case SpvBuiltInInvocationId
:
1045 *location
= SYSTEM_VALUE_INVOCATION_ID
;
1046 set_mode_system_value(b
, mode
);
1048 case SpvBuiltInLayer
:
1049 *location
= VARYING_SLOT_LAYER
;
1050 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1051 *mode
= nir_var_shader_in
;
1052 else if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1053 *mode
= nir_var_shader_out
;
1054 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1055 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1056 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1057 *mode
= nir_var_shader_out
;
1059 vtn_fail("invalid stage for SpvBuiltInLayer");
1061 case SpvBuiltInViewportIndex
:
1062 *location
= VARYING_SLOT_VIEWPORT
;
1063 if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1064 *mode
= nir_var_shader_out
;
1065 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1066 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1067 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1068 *mode
= nir_var_shader_out
;
1069 else if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1070 *mode
= nir_var_shader_in
;
1072 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1074 case SpvBuiltInTessLevelOuter
:
1075 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
1077 case SpvBuiltInTessLevelInner
:
1078 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
1080 case SpvBuiltInTessCoord
:
1081 *location
= SYSTEM_VALUE_TESS_COORD
;
1082 set_mode_system_value(b
, mode
);
1084 case SpvBuiltInPatchVertices
:
1085 *location
= SYSTEM_VALUE_VERTICES_IN
;
1086 set_mode_system_value(b
, mode
);
1088 case SpvBuiltInFragCoord
:
1089 *location
= VARYING_SLOT_POS
;
1090 vtn_assert(*mode
== nir_var_shader_in
);
1092 case SpvBuiltInPointCoord
:
1093 *location
= VARYING_SLOT_PNTC
;
1094 vtn_assert(*mode
== nir_var_shader_in
);
1096 case SpvBuiltInFrontFacing
:
1097 *location
= SYSTEM_VALUE_FRONT_FACE
;
1098 set_mode_system_value(b
, mode
);
1100 case SpvBuiltInSampleId
:
1101 *location
= SYSTEM_VALUE_SAMPLE_ID
;
1102 set_mode_system_value(b
, mode
);
1104 case SpvBuiltInSamplePosition
:
1105 *location
= SYSTEM_VALUE_SAMPLE_POS
;
1106 set_mode_system_value(b
, mode
);
1108 case SpvBuiltInSampleMask
:
1109 if (*mode
== nir_var_shader_out
) {
1110 *location
= FRAG_RESULT_SAMPLE_MASK
;
1112 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
1113 set_mode_system_value(b
, mode
);
1116 case SpvBuiltInFragDepth
:
1117 *location
= FRAG_RESULT_DEPTH
;
1118 vtn_assert(*mode
== nir_var_shader_out
);
1120 case SpvBuiltInHelperInvocation
:
1121 *location
= SYSTEM_VALUE_HELPER_INVOCATION
;
1122 set_mode_system_value(b
, mode
);
1124 case SpvBuiltInNumWorkgroups
:
1125 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
1126 set_mode_system_value(b
, mode
);
1128 case SpvBuiltInWorkgroupSize
:
1129 *location
= SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
1130 set_mode_system_value(b
, mode
);
1132 case SpvBuiltInWorkgroupId
:
1133 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
1134 set_mode_system_value(b
, mode
);
1136 case SpvBuiltInLocalInvocationId
:
1137 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
1138 set_mode_system_value(b
, mode
);
1140 case SpvBuiltInLocalInvocationIndex
:
1141 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
1142 set_mode_system_value(b
, mode
);
1144 case SpvBuiltInGlobalInvocationId
:
1145 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
1146 set_mode_system_value(b
, mode
);
1148 case SpvBuiltInBaseVertex
:
1149 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1150 * semantic as SPIR-V BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1152 *location
= SYSTEM_VALUE_FIRST_VERTEX
;
1153 set_mode_system_value(b
, mode
);
1155 case SpvBuiltInBaseInstance
:
1156 *location
= SYSTEM_VALUE_BASE_INSTANCE
;
1157 set_mode_system_value(b
, mode
);
1159 case SpvBuiltInDrawIndex
:
1160 *location
= SYSTEM_VALUE_DRAW_ID
;
1161 set_mode_system_value(b
, mode
);
1163 case SpvBuiltInSubgroupSize
:
1164 *location
= SYSTEM_VALUE_SUBGROUP_SIZE
;
1165 set_mode_system_value(b
, mode
);
1167 case SpvBuiltInSubgroupId
:
1168 *location
= SYSTEM_VALUE_SUBGROUP_ID
;
1169 set_mode_system_value(b
, mode
);
1171 case SpvBuiltInSubgroupLocalInvocationId
:
1172 *location
= SYSTEM_VALUE_SUBGROUP_INVOCATION
;
1173 set_mode_system_value(b
, mode
);
1175 case SpvBuiltInNumSubgroups
:
1176 *location
= SYSTEM_VALUE_NUM_SUBGROUPS
;
1177 set_mode_system_value(b
, mode
);
1179 case SpvBuiltInDeviceIndex
:
1180 *location
= SYSTEM_VALUE_DEVICE_INDEX
;
1181 set_mode_system_value(b
, mode
);
1183 case SpvBuiltInViewIndex
:
1184 *location
= SYSTEM_VALUE_VIEW_INDEX
;
1185 set_mode_system_value(b
, mode
);
1187 case SpvBuiltInSubgroupEqMask
:
1188 *location
= SYSTEM_VALUE_SUBGROUP_EQ_MASK
,
1189 set_mode_system_value(b
, mode
);
1191 case SpvBuiltInSubgroupGeMask
:
1192 *location
= SYSTEM_VALUE_SUBGROUP_GE_MASK
,
1193 set_mode_system_value(b
, mode
);
1195 case SpvBuiltInSubgroupGtMask
:
1196 *location
= SYSTEM_VALUE_SUBGROUP_GT_MASK
,
1197 set_mode_system_value(b
, mode
);
1199 case SpvBuiltInSubgroupLeMask
:
1200 *location
= SYSTEM_VALUE_SUBGROUP_LE_MASK
,
1201 set_mode_system_value(b
, mode
);
1203 case SpvBuiltInSubgroupLtMask
:
1204 *location
= SYSTEM_VALUE_SUBGROUP_LT_MASK
,
1205 set_mode_system_value(b
, mode
);
1207 case SpvBuiltInFragStencilRefEXT
:
1208 *location
= FRAG_RESULT_STENCIL
;
1209 vtn_assert(*mode
== nir_var_shader_out
);
1211 case SpvBuiltInGlobalSize
:
1212 *location
= SYSTEM_VALUE_GLOBAL_GROUP_SIZE
;
1213 set_mode_system_value(b
, mode
);
1216 vtn_fail("unsupported builtin");
1221 apply_var_decoration(struct vtn_builder
*b
,
1222 struct nir_variable_data
*var_data
,
1223 const struct vtn_decoration
*dec
)
1225 switch (dec
->decoration
) {
1226 case SpvDecorationRelaxedPrecision
:
1227 break; /* FIXME: Do nothing with this for now. */
1228 case SpvDecorationNoPerspective
:
1229 var_data
->interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1231 case SpvDecorationFlat
:
1232 var_data
->interpolation
= INTERP_MODE_FLAT
;
1234 case SpvDecorationCentroid
:
1235 var_data
->centroid
= true;
1237 case SpvDecorationSample
:
1238 var_data
->sample
= true;
1240 case SpvDecorationInvariant
:
1241 var_data
->invariant
= true;
1243 case SpvDecorationConstant
:
1244 var_data
->read_only
= true;
1246 case SpvDecorationNonReadable
:
1247 var_data
->image
.write_only
= true;
1249 case SpvDecorationNonWritable
:
1250 var_data
->read_only
= true;
1251 var_data
->image
.read_only
= true;
1253 case SpvDecorationRestrict
:
1254 var_data
->image
.restrict_flag
= true;
1256 case SpvDecorationVolatile
:
1257 var_data
->image
._volatile
= true;
1259 case SpvDecorationCoherent
:
1260 var_data
->image
.coherent
= true;
1262 case SpvDecorationComponent
:
1263 var_data
->location_frac
= dec
->literals
[0];
1265 case SpvDecorationIndex
:
1266 var_data
->index
= dec
->literals
[0];
1268 case SpvDecorationBuiltIn
: {
1269 SpvBuiltIn builtin
= dec
->literals
[0];
1271 nir_variable_mode mode
= var_data
->mode
;
1272 vtn_get_builtin_location(b
, builtin
, &var_data
->location
, &mode
);
1273 var_data
->mode
= mode
;
1276 case SpvBuiltInTessLevelOuter
:
1277 case SpvBuiltInTessLevelInner
:
1278 var_data
->compact
= true;
1280 case SpvBuiltInFragCoord
:
1281 var_data
->pixel_center_integer
= b
->pixel_center_integer
;
1283 case SpvBuiltInSamplePosition
:
1284 var_data
->origin_upper_left
= b
->origin_upper_left
;
1291 case SpvDecorationSpecId
:
1292 case SpvDecorationRowMajor
:
1293 case SpvDecorationColMajor
:
1294 case SpvDecorationMatrixStride
:
1295 case SpvDecorationAliased
:
1296 case SpvDecorationUniform
:
1297 case SpvDecorationStream
:
1298 case SpvDecorationOffset
:
1299 case SpvDecorationLinkageAttributes
:
1300 break; /* Do nothing with these here */
1302 case SpvDecorationPatch
:
1303 var_data
->patch
= true;
1306 case SpvDecorationLocation
:
1307 vtn_fail("Handled above");
1309 case SpvDecorationBlock
:
1310 case SpvDecorationBufferBlock
:
1311 case SpvDecorationArrayStride
:
1312 case SpvDecorationGLSLShared
:
1313 case SpvDecorationGLSLPacked
:
1314 break; /* These can apply to a type but we don't care about them */
1316 case SpvDecorationBinding
:
1317 case SpvDecorationDescriptorSet
:
1318 case SpvDecorationNoContraction
:
1319 case SpvDecorationInputAttachmentIndex
:
1320 vtn_warn("Decoration not allowed for variable or structure member: %s",
1321 spirv_decoration_to_string(dec
->decoration
));
1324 case SpvDecorationXfbBuffer
:
1325 case SpvDecorationXfbStride
:
1326 vtn_warn("Vulkan does not have transform feedback: %s",
1327 spirv_decoration_to_string(dec
->decoration
));
1330 case SpvDecorationCPacked
:
1331 case SpvDecorationSaturatedConversion
:
1332 case SpvDecorationFuncParamAttr
:
1333 case SpvDecorationFPRoundingMode
:
1334 case SpvDecorationFPFastMathMode
:
1335 case SpvDecorationAlignment
:
1336 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1337 spirv_decoration_to_string(dec
->decoration
));
1341 vtn_fail("Unhandled decoration");
1346 var_is_patch_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1347 const struct vtn_decoration
*dec
, void *out_is_patch
)
1349 if (dec
->decoration
== SpvDecorationPatch
) {
1350 *((bool *) out_is_patch
) = true;
1355 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1356 const struct vtn_decoration
*dec
, void *void_var
)
1358 struct vtn_variable
*vtn_var
= void_var
;
1360 /* Handle decorations that apply to a vtn_variable as a whole */
1361 switch (dec
->decoration
) {
1362 case SpvDecorationBinding
:
1363 vtn_var
->binding
= dec
->literals
[0];
1364 vtn_var
->explicit_binding
= true;
1366 case SpvDecorationDescriptorSet
:
1367 vtn_var
->descriptor_set
= dec
->literals
[0];
1369 case SpvDecorationInputAttachmentIndex
:
1370 vtn_var
->input_attachment_index
= dec
->literals
[0];
1372 case SpvDecorationPatch
:
1373 vtn_var
->patch
= true;
1375 case SpvDecorationOffset
:
1376 vtn_var
->offset
= dec
->literals
[0];
1382 if (val
->value_type
== vtn_value_type_pointer
) {
1383 assert(val
->pointer
->var
== void_var
);
1384 assert(val
->pointer
->chain
== NULL
);
1385 assert(member
== -1);
1387 assert(val
->value_type
== vtn_value_type_type
);
1390 /* Location is odd. If applied to a split structure, we have to walk the
1391 * whole thing and accumulate the location. It's easier to handle as a
1394 if (dec
->decoration
== SpvDecorationLocation
) {
1395 unsigned location
= dec
->literals
[0];
1396 bool is_vertex_input
;
1397 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
1398 vtn_var
->mode
== vtn_variable_mode_output
) {
1399 is_vertex_input
= false;
1400 location
+= FRAG_RESULT_DATA0
;
1401 } else if (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
&&
1402 vtn_var
->mode
== vtn_variable_mode_input
) {
1403 is_vertex_input
= true;
1404 location
+= VERT_ATTRIB_GENERIC0
;
1405 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1406 vtn_var
->mode
== vtn_variable_mode_output
) {
1407 is_vertex_input
= false;
1408 location
+= vtn_var
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
;
1409 } else if (vtn_var
->mode
!= vtn_variable_mode_uniform
) {
1410 vtn_warn("Location must be on input, output, uniform, sampler or "
1415 if (vtn_var
->var
->num_members
== 0) {
1416 /* This handles the member and lone variable cases */
1417 vtn_var
->var
->data
.location
= location
;
1419 /* This handles the structure member case */
1420 assert(vtn_var
->var
->members
);
1421 for (unsigned i
= 0; i
< vtn_var
->var
->num_members
; i
++) {
1422 vtn_var
->var
->members
[i
].location
= location
;
1423 const struct glsl_type
*member_type
=
1424 glsl_get_struct_field(vtn_var
->var
->interface_type
, i
);
1425 location
+= glsl_count_attribute_slots(member_type
,
1432 if (vtn_var
->var
->num_members
== 0) {
1433 assert(member
== -1);
1434 apply_var_decoration(b
, &vtn_var
->var
->data
, dec
);
1435 } else if (member
>= 0) {
1436 /* Member decorations must come from a type */
1437 assert(val
->value_type
== vtn_value_type_type
);
1438 apply_var_decoration(b
, &vtn_var
->var
->members
[member
], dec
);
1441 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1442 for (unsigned i
= 0; i
< length
; i
++)
1443 apply_var_decoration(b
, &vtn_var
->var
->members
[i
], dec
);
1446 /* A few variables, those with external storage, have no actual
1447 * nir_variables associated with them. Fortunately, all decorations
1448 * we care about for those variables are on the type only.
1450 vtn_assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1451 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1452 vtn_var
->mode
== vtn_variable_mode_push_constant
||
1453 (vtn_var
->mode
== vtn_variable_mode_workgroup
&&
1454 b
->options
->lower_workgroup_access_to_offsets
));
1459 static enum vtn_variable_mode
1460 vtn_storage_class_to_mode(struct vtn_builder
*b
,
1461 SpvStorageClass
class,
1462 struct vtn_type
*interface_type
,
1463 nir_variable_mode
*nir_mode_out
)
1465 enum vtn_variable_mode mode
;
1466 nir_variable_mode nir_mode
;
1468 case SpvStorageClassUniform
:
1469 if (interface_type
->block
) {
1470 mode
= vtn_variable_mode_ubo
;
1472 } else if (interface_type
->buffer_block
) {
1473 mode
= vtn_variable_mode_ssbo
;
1476 /* Default-block uniforms, coming from gl_spirv */
1477 mode
= vtn_variable_mode_uniform
;
1478 nir_mode
= nir_var_uniform
;
1481 case SpvStorageClassStorageBuffer
:
1482 mode
= vtn_variable_mode_ssbo
;
1485 case SpvStorageClassUniformConstant
:
1486 mode
= vtn_variable_mode_uniform
;
1487 nir_mode
= nir_var_uniform
;
1489 case SpvStorageClassPushConstant
:
1490 mode
= vtn_variable_mode_push_constant
;
1491 nir_mode
= nir_var_uniform
;
1493 case SpvStorageClassInput
:
1494 mode
= vtn_variable_mode_input
;
1495 nir_mode
= nir_var_shader_in
;
1497 case SpvStorageClassOutput
:
1498 mode
= vtn_variable_mode_output
;
1499 nir_mode
= nir_var_shader_out
;
1501 case SpvStorageClassPrivate
:
1502 mode
= vtn_variable_mode_global
;
1503 nir_mode
= nir_var_global
;
1505 case SpvStorageClassFunction
:
1506 mode
= vtn_variable_mode_local
;
1507 nir_mode
= nir_var_local
;
1509 case SpvStorageClassWorkgroup
:
1510 mode
= vtn_variable_mode_workgroup
;
1511 nir_mode
= nir_var_shared
;
1513 case SpvStorageClassAtomicCounter
:
1514 mode
= vtn_variable_mode_uniform
;
1515 nir_mode
= nir_var_uniform
;
1517 case SpvStorageClassCrossWorkgroup
:
1518 case SpvStorageClassGeneric
:
1520 vtn_fail("Unhandled variable storage class");
1524 *nir_mode_out
= nir_mode
;
1530 vtn_pointer_to_ssa(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
1532 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1533 /* This pointer needs to have a pointer type with actual storage */
1534 vtn_assert(ptr
->ptr_type
);
1535 vtn_assert(ptr
->ptr_type
->type
);
1538 /* If we don't have an offset then we must be a pointer to the variable
1541 vtn_assert(!ptr
->offset
&& !ptr
->block_index
);
1543 struct vtn_access_chain chain
= {
1546 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
1549 vtn_assert(ptr
->offset
);
1550 if (ptr
->block_index
) {
1551 vtn_assert(ptr
->mode
== vtn_variable_mode_ubo
||
1552 ptr
->mode
== vtn_variable_mode_ssbo
);
1553 return nir_vec2(&b
->nb
, ptr
->block_index
, ptr
->offset
);
1555 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
);
1559 return &vtn_pointer_to_deref(b
, ptr
)->dest
.ssa
;
1563 struct vtn_pointer
*
1564 vtn_pointer_from_ssa(struct vtn_builder
*b
, nir_ssa_def
*ssa
,
1565 struct vtn_type
*ptr_type
)
1567 vtn_assert(ssa
->num_components
<= 2 && ssa
->bit_size
== 32);
1568 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1570 struct vtn_type
*interface_type
= ptr_type
->deref
;
1571 while (interface_type
->base_type
== vtn_base_type_array
)
1572 interface_type
= interface_type
->array_element
;
1574 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
1575 nir_variable_mode nir_mode
;
1576 ptr
->mode
= vtn_storage_class_to_mode(b
, ptr_type
->storage_class
,
1577 interface_type
, &nir_mode
);
1578 ptr
->type
= ptr_type
->deref
;
1579 ptr
->ptr_type
= ptr_type
;
1581 if (ptr
->mode
== vtn_variable_mode_ubo
||
1582 ptr
->mode
== vtn_variable_mode_ssbo
) {
1583 /* This pointer type needs to have actual storage */
1584 vtn_assert(ptr_type
->type
);
1585 vtn_assert(ssa
->num_components
== 2);
1586 ptr
->block_index
= nir_channel(&b
->nb
, ssa
, 0);
1587 ptr
->offset
= nir_channel(&b
->nb
, ssa
, 1);
1588 } else if (ptr
->mode
== vtn_variable_mode_workgroup
||
1589 ptr
->mode
== vtn_variable_mode_push_constant
) {
1590 /* This pointer type needs to have actual storage */
1591 vtn_assert(ptr_type
->type
);
1592 vtn_assert(ssa
->num_components
== 1);
1593 ptr
->block_index
= NULL
;
1596 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
1597 ptr_type
->deref
->type
);
1604 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
1606 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
1609 if (var
->mode
== vtn_variable_mode_input
) {
1610 return stage
== MESA_SHADER_TESS_CTRL
||
1611 stage
== MESA_SHADER_TESS_EVAL
||
1612 stage
== MESA_SHADER_GEOMETRY
;
1615 if (var
->mode
== vtn_variable_mode_output
)
1616 return stage
== MESA_SHADER_TESS_CTRL
;
1622 vtn_create_variable(struct vtn_builder
*b
, struct vtn_value
*val
,
1623 struct vtn_type
*ptr_type
, SpvStorageClass storage_class
,
1624 nir_constant
*initializer
)
1626 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1627 struct vtn_type
*type
= ptr_type
->deref
;
1629 struct vtn_type
*without_array
= type
;
1630 while(glsl_type_is_array(without_array
->type
))
1631 without_array
= without_array
->array_element
;
1633 enum vtn_variable_mode mode
;
1634 nir_variable_mode nir_mode
;
1635 mode
= vtn_storage_class_to_mode(b
, storage_class
, without_array
, &nir_mode
);
1638 case vtn_variable_mode_ubo
:
1639 b
->shader
->info
.num_ubos
++;
1641 case vtn_variable_mode_ssbo
:
1642 b
->shader
->info
.num_ssbos
++;
1644 case vtn_variable_mode_uniform
:
1645 if (glsl_type_is_image(without_array
->type
))
1646 b
->shader
->info
.num_images
++;
1647 else if (glsl_type_is_sampler(without_array
->type
))
1648 b
->shader
->info
.num_textures
++;
1650 case vtn_variable_mode_push_constant
:
1651 b
->shader
->num_uniforms
= vtn_type_block_size(b
, type
);
1654 /* No tallying is needed */
1658 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
1662 vtn_assert(val
->value_type
== vtn_value_type_pointer
);
1663 val
->pointer
= vtn_pointer_for_variable(b
, var
, ptr_type
);
1665 switch (var
->mode
) {
1666 case vtn_variable_mode_local
:
1667 case vtn_variable_mode_global
:
1668 case vtn_variable_mode_uniform
:
1669 /* For these, we create the variable normally */
1670 var
->var
= rzalloc(b
->shader
, nir_variable
);
1671 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1673 /* Need to tweak the nir type here as at vtn_handle_type we don't have
1674 * the access to storage_class, that is the one that points us that is
1677 if (storage_class
== SpvStorageClassAtomicCounter
) {
1678 var
->var
->type
= repair_atomic_type(var
->type
->type
);
1680 var
->var
->type
= var
->type
->type
;
1682 var
->var
->data
.mode
= nir_mode
;
1683 var
->var
->data
.location
= -1;
1684 var
->var
->interface_type
= NULL
;
1687 case vtn_variable_mode_workgroup
:
1688 if (b
->options
->lower_workgroup_access_to_offsets
) {
1689 var
->shared_location
= -1;
1691 /* Create the variable normally */
1692 var
->var
= rzalloc(b
->shader
, nir_variable
);
1693 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1694 var
->var
->type
= var
->type
->type
;
1695 var
->var
->data
.mode
= nir_var_shared
;
1699 case vtn_variable_mode_input
:
1700 case vtn_variable_mode_output
: {
1701 /* In order to know whether or not we're a per-vertex inout, we need
1702 * the patch qualifier. This means walking the variable decorations
1703 * early before we actually create any variables. Not a big deal.
1705 * GLSLang really likes to place decorations in the most interior
1706 * thing it possibly can. In particular, if you have a struct, it
1707 * will place the patch decorations on the struct members. This
1708 * should be handled by the variable splitting below just fine.
1710 * If you have an array-of-struct, things get even more weird as it
1711 * will place the patch decorations on the struct even though it's
1712 * inside an array and some of the members being patch and others not
1713 * makes no sense whatsoever. Since the only sensible thing is for
1714 * it to be all or nothing, we'll call it patch if any of the members
1715 * are declared patch.
1718 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
1719 if (glsl_type_is_array(var
->type
->type
) &&
1720 glsl_type_is_struct(without_array
->type
)) {
1721 vtn_foreach_decoration(b
, vtn_value(b
, without_array
->id
,
1722 vtn_value_type_type
),
1723 var_is_patch_cb
, &var
->patch
);
1726 /* For inputs and outputs, we immediately split structures. This
1727 * is for a couple of reasons. For one, builtins may all come in
1728 * a struct and we really want those split out into separate
1729 * variables. For another, interpolation qualifiers can be
1730 * applied to members of the top-level struct ane we need to be
1731 * able to preserve that information.
1734 struct vtn_type
*interface_type
= var
->type
;
1735 if (is_per_vertex_inout(var
, b
->shader
->info
.stage
)) {
1736 /* In Geometry shaders (and some tessellation), inputs come
1737 * in per-vertex arrays. However, some builtins come in
1738 * non-per-vertex, hence the need for the is_array check. In
1739 * any case, there are no non-builtin arrays allowed so this
1740 * check should be sufficient.
1742 interface_type
= var
->type
->array_element
;
1745 var
->var
= rzalloc(b
->shader
, nir_variable
);
1746 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
1747 var
->var
->type
= var
->type
->type
;
1748 var
->var
->interface_type
= interface_type
->type
;
1749 var
->var
->data
.mode
= nir_mode
;
1750 var
->var
->data
.patch
= var
->patch
;
1752 if (glsl_type_is_struct(interface_type
->type
)) {
1753 /* It's a struct. Set it up as per-member. */
1754 var
->var
->num_members
= glsl_get_length(interface_type
->type
);
1755 var
->var
->members
= rzalloc_array(var
->var
, struct nir_variable_data
,
1756 var
->var
->num_members
);
1758 for (unsigned i
= 0; i
< var
->var
->num_members
; i
++) {
1759 var
->var
->members
[i
].mode
= nir_mode
;
1760 var
->var
->members
[i
].patch
= var
->patch
;
1764 /* For inputs and outputs, we need to grab locations and builtin
1765 * information from the interface type.
1767 vtn_foreach_decoration(b
, vtn_value(b
, interface_type
->id
,
1768 vtn_value_type_type
),
1769 var_decoration_cb
, var
);
1773 case vtn_variable_mode_ubo
:
1774 case vtn_variable_mode_ssbo
:
1775 case vtn_variable_mode_push_constant
:
1776 /* These don't need actual variables. */
1781 var
->var
->constant_initializer
=
1782 nir_constant_clone(initializer
, var
->var
);
1785 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
1787 if (var
->mode
== vtn_variable_mode_uniform
) {
1788 /* XXX: We still need the binding information in the nir_variable
1789 * for these. We should fix that.
1791 var
->var
->data
.binding
= var
->binding
;
1792 var
->var
->data
.explicit_binding
= var
->explicit_binding
;
1793 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
1794 var
->var
->data
.index
= var
->input_attachment_index
;
1795 var
->var
->data
.offset
= var
->offset
;
1797 if (glsl_type_is_image(without_array
->type
))
1798 var
->var
->data
.image
.format
= without_array
->image_format
;
1801 if (var
->mode
== vtn_variable_mode_local
) {
1802 vtn_assert(var
->var
!= NULL
&& var
->var
->members
== NULL
);
1803 nir_function_impl_add_variable(b
->nb
.impl
, var
->var
);
1804 } else if (var
->var
) {
1805 nir_shader_add_variable(b
->shader
, var
->var
);
1807 vtn_assert(vtn_pointer_is_external_block(b
, val
->pointer
));
1812 vtn_assert_types_equal(struct vtn_builder
*b
, SpvOp opcode
,
1813 struct vtn_type
*dst_type
,
1814 struct vtn_type
*src_type
)
1816 if (dst_type
->id
== src_type
->id
)
1819 if (vtn_types_compatible(b
, dst_type
, src_type
)) {
1820 /* Early versions of GLSLang would re-emit types unnecessarily and you
1821 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
1822 * mismatched source and destination types.
1824 * https://github.com/KhronosGroup/glslang/issues/304
1825 * https://github.com/KhronosGroup/glslang/issues/307
1826 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
1827 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
1829 vtn_warn("Source and destination types of %s do not have the same "
1830 "ID (but are compatible): %u vs %u",
1831 spirv_op_to_string(opcode
), dst_type
->id
, src_type
->id
);
1835 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
1836 spirv_op_to_string(opcode
),
1837 glsl_get_type_name(dst_type
->type
),
1838 glsl_get_type_name(src_type
->type
));
1842 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
1843 const uint32_t *w
, unsigned count
)
1847 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
1848 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1852 case SpvOpVariable
: {
1853 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1855 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1857 SpvStorageClass storage_class
= w
[3];
1858 nir_constant
*initializer
= NULL
;
1860 initializer
= vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
1862 vtn_create_variable(b
, val
, ptr_type
, storage_class
, initializer
);
1866 case SpvOpAccessChain
:
1867 case SpvOpPtrAccessChain
:
1868 case SpvOpInBoundsAccessChain
: {
1869 struct vtn_access_chain
*chain
= vtn_access_chain_create(b
, count
- 4);
1870 chain
->ptr_as_array
= (opcode
== SpvOpPtrAccessChain
);
1873 for (int i
= 4; i
< count
; i
++) {
1874 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
1875 if (link_val
->value_type
== vtn_value_type_constant
) {
1876 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
1877 chain
->link
[idx
].id
= link_val
->constant
->values
[0].u32
[0];
1879 chain
->link
[idx
].mode
= vtn_access_mode_id
;
1880 chain
->link
[idx
].id
= w
[i
];
1886 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1887 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
1888 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
1889 /* This is rather insane. SPIR-V allows you to use OpSampledImage
1890 * to combine an array of images with a single sampler to get an
1891 * array of sampled images that all share the same sampler.
1892 * Fortunately, this means that we can more-or-less ignore the
1893 * sampler when crawling the access chain, but it does leave us
1894 * with this rather awkward little special-case.
1896 struct vtn_value
*val
=
1897 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
1898 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
1899 val
->sampled_image
->type
= base_val
->sampled_image
->type
;
1900 val
->sampled_image
->image
=
1901 vtn_pointer_dereference(b
, base_val
->sampled_image
->image
, chain
);
1902 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
1904 vtn_assert(base_val
->value_type
== vtn_value_type_pointer
);
1905 struct vtn_value
*val
=
1906 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
1907 val
->pointer
= vtn_pointer_dereference(b
, base_val
->pointer
, chain
);
1908 val
->pointer
->ptr_type
= ptr_type
;
1913 case SpvOpCopyMemory
: {
1914 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
1915 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_pointer
);
1917 vtn_assert_types_equal(b
, opcode
, dest
->type
->deref
, src
->type
->deref
);
1919 vtn_variable_copy(b
, dest
->pointer
, src
->pointer
);
1924 struct vtn_type
*res_type
=
1925 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
1926 struct vtn_value
*src_val
= vtn_value(b
, w
[3], vtn_value_type_pointer
);
1927 struct vtn_pointer
*src
= src_val
->pointer
;
1929 vtn_assert_types_equal(b
, opcode
, res_type
, src_val
->type
->deref
);
1931 if (glsl_type_is_image(res_type
->type
) ||
1932 glsl_type_is_sampler(res_type
->type
)) {
1933 vtn_push_value(b
, w
[2], vtn_value_type_pointer
)->pointer
= src
;
1937 vtn_push_ssa(b
, w
[2], res_type
, vtn_variable_load(b
, src
));
1942 struct vtn_value
*dest_val
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
1943 struct vtn_pointer
*dest
= dest_val
->pointer
;
1944 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[2]);
1946 /* OpStore requires us to actually have a storage type */
1947 vtn_fail_if(dest
->type
->type
== NULL
,
1948 "Invalid destination type for OpStore");
1950 if (glsl_get_base_type(dest
->type
->type
) == GLSL_TYPE_BOOL
&&
1951 glsl_get_base_type(src_val
->type
->type
) == GLSL_TYPE_UINT
) {
1952 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
1953 * would then store them to a local variable as bool. Work around
1954 * the issue by doing an implicit conversion.
1956 * https://github.com/KhronosGroup/glslang/issues/170
1957 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
1959 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
1960 "OpTypeBool. Doing an implicit conversion to work around "
1962 struct vtn_ssa_value
*bool_ssa
=
1963 vtn_create_ssa_value(b
, dest
->type
->type
);
1964 bool_ssa
->def
= nir_i2b(&b
->nb
, vtn_ssa_value(b
, w
[2])->def
);
1965 vtn_variable_store(b
, bool_ssa
, dest
);
1969 vtn_assert_types_equal(b
, opcode
, dest_val
->type
->deref
, src_val
->type
);
1971 if (glsl_type_is_sampler(dest
->type
->type
)) {
1972 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
1973 "propagation to workaround the problem.");
1974 vtn_assert(dest
->var
->copy_prop_sampler
== NULL
);
1975 dest
->var
->copy_prop_sampler
=
1976 vtn_value(b
, w
[2], vtn_value_type_pointer
)->pointer
;
1980 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
1981 vtn_variable_store(b
, src
, dest
);
1985 case SpvOpArrayLength
: {
1986 struct vtn_pointer
*ptr
=
1987 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
1989 const uint32_t offset
= ptr
->var
->type
->offsets
[w
[4]];
1990 const uint32_t stride
= ptr
->var
->type
->members
[w
[4]]->stride
;
1992 if (!ptr
->block_index
) {
1993 struct vtn_access_chain chain
= {
1996 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
1997 vtn_assert(ptr
->block_index
);
2000 nir_intrinsic_instr
*instr
=
2001 nir_intrinsic_instr_create(b
->nb
.shader
,
2002 nir_intrinsic_get_buffer_size
);
2003 instr
->src
[0] = nir_src_for_ssa(ptr
->block_index
);
2004 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
2005 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2006 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
2008 /* array_length = max(buffer_size - offset, 0) / stride */
2009 nir_ssa_def
*array_length
=
2014 nir_imm_int(&b
->nb
, offset
)),
2015 nir_imm_int(&b
->nb
, 0u)),
2016 nir_imm_int(&b
->nb
, stride
));
2018 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2019 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
2020 val
->ssa
->def
= array_length
;
2024 case SpvOpCopyMemorySized
:
2026 vtn_fail("Unhandled opcode");