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"
31 #include <vulkan/vulkan_core.h>
33 static struct vtn_access_chain
*
34 vtn_access_chain_create(struct vtn_builder
*b
, unsigned length
)
36 struct vtn_access_chain
*chain
;
38 /* Subtract 1 from the length since there's already one built in */
39 size_t size
= sizeof(*chain
) +
40 (MAX2(length
, 1) - 1) * sizeof(chain
->link
[0]);
41 chain
= rzalloc_size(b
, size
);
42 chain
->length
= length
;
48 vtn_pointer_uses_ssa_offset(struct vtn_builder
*b
,
49 struct vtn_pointer
*ptr
)
51 return ((ptr
->mode
== vtn_variable_mode_ubo
||
52 ptr
->mode
== vtn_variable_mode_ssbo
) &&
53 b
->options
->lower_ubo_ssbo_access_to_offsets
) ||
54 ptr
->mode
== vtn_variable_mode_push_constant
||
55 (ptr
->mode
== vtn_variable_mode_workgroup
&&
56 b
->options
->lower_workgroup_access_to_offsets
);
60 vtn_pointer_is_external_block(struct vtn_builder
*b
,
61 struct vtn_pointer
*ptr
)
63 return ptr
->mode
== vtn_variable_mode_ssbo
||
64 ptr
->mode
== vtn_variable_mode_ubo
||
65 ptr
->mode
== vtn_variable_mode_phys_ssbo
||
66 ptr
->mode
== vtn_variable_mode_push_constant
||
67 (ptr
->mode
== vtn_variable_mode_workgroup
&&
68 b
->options
->lower_workgroup_access_to_offsets
);
72 vtn_access_link_as_ssa(struct vtn_builder
*b
, struct vtn_access_link link
,
73 unsigned stride
, unsigned bit_size
)
75 vtn_assert(stride
> 0);
76 if (link
.mode
== vtn_access_mode_literal
) {
77 return nir_imm_intN_t(&b
->nb
, link
.id
* stride
, bit_size
);
79 nir_ssa_def
*ssa
= vtn_ssa_value(b
, link
.id
)->def
;
80 if (ssa
->bit_size
!= bit_size
)
81 ssa
= nir_i2i(&b
->nb
, ssa
, bit_size
);
82 return nir_imul_imm(&b
->nb
, ssa
, stride
);
86 static VkDescriptorType
87 vk_desc_type_for_mode(struct vtn_builder
*b
, enum vtn_variable_mode mode
)
90 case vtn_variable_mode_ubo
:
91 return VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
;
92 case vtn_variable_mode_ssbo
:
93 return VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
;
95 vtn_fail("Invalid mode for vulkan_resource_index");
99 static const struct glsl_type
*
100 vtn_ptr_type_for_mode(struct vtn_builder
*b
, enum vtn_variable_mode mode
)
103 case vtn_variable_mode_ubo
:
104 return b
->options
->ubo_ptr_type
;
105 case vtn_variable_mode_ssbo
:
106 return b
->options
->ssbo_ptr_type
;
108 vtn_fail("Invalid mode for vulkan_resource_index");
113 vtn_variable_resource_index(struct vtn_builder
*b
, struct vtn_variable
*var
,
114 nir_ssa_def
*desc_array_index
)
116 if (!desc_array_index
) {
117 vtn_assert(glsl_type_is_struct_or_ifc(var
->type
->type
));
118 desc_array_index
= nir_imm_int(&b
->nb
, 0);
121 nir_intrinsic_instr
*instr
=
122 nir_intrinsic_instr_create(b
->nb
.shader
,
123 nir_intrinsic_vulkan_resource_index
);
124 instr
->src
[0] = nir_src_for_ssa(desc_array_index
);
125 nir_intrinsic_set_desc_set(instr
, var
->descriptor_set
);
126 nir_intrinsic_set_binding(instr
, var
->binding
);
127 nir_intrinsic_set_desc_type(instr
, vk_desc_type_for_mode(b
, var
->mode
));
129 const struct glsl_type
*index_type
=
130 b
->options
->lower_ubo_ssbo_access_to_offsets
?
131 glsl_uint_type() : vtn_ptr_type_for_mode(b
, var
->mode
);
133 instr
->num_components
= glsl_get_vector_elements(index_type
);
134 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, instr
->num_components
,
135 glsl_get_bit_size(index_type
), NULL
);
136 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
138 return &instr
->dest
.ssa
;
142 vtn_resource_reindex(struct vtn_builder
*b
, enum vtn_variable_mode mode
,
143 nir_ssa_def
*base_index
, nir_ssa_def
*offset_index
)
145 nir_intrinsic_instr
*instr
=
146 nir_intrinsic_instr_create(b
->nb
.shader
,
147 nir_intrinsic_vulkan_resource_reindex
);
148 instr
->src
[0] = nir_src_for_ssa(base_index
);
149 instr
->src
[1] = nir_src_for_ssa(offset_index
);
150 nir_intrinsic_set_desc_type(instr
, vk_desc_type_for_mode(b
, mode
));
152 const struct glsl_type
*index_type
=
153 b
->options
->lower_ubo_ssbo_access_to_offsets
?
154 glsl_uint_type() : vtn_ptr_type_for_mode(b
, mode
);
156 instr
->num_components
= glsl_get_vector_elements(index_type
);
157 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, instr
->num_components
,
158 glsl_get_bit_size(index_type
), NULL
);
159 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
161 return &instr
->dest
.ssa
;
165 vtn_descriptor_load(struct vtn_builder
*b
, enum vtn_variable_mode mode
,
166 nir_ssa_def
*desc_index
)
168 nir_intrinsic_instr
*desc_load
=
169 nir_intrinsic_instr_create(b
->nb
.shader
,
170 nir_intrinsic_load_vulkan_descriptor
);
171 desc_load
->src
[0] = nir_src_for_ssa(desc_index
);
172 nir_intrinsic_set_desc_type(desc_load
, vk_desc_type_for_mode(b
, mode
));
174 const struct glsl_type
*ptr_type
= vtn_ptr_type_for_mode(b
, mode
);
176 desc_load
->num_components
= glsl_get_vector_elements(ptr_type
);
177 nir_ssa_dest_init(&desc_load
->instr
, &desc_load
->dest
,
178 desc_load
->num_components
,
179 glsl_get_bit_size(ptr_type
), NULL
);
180 nir_builder_instr_insert(&b
->nb
, &desc_load
->instr
);
182 return &desc_load
->dest
.ssa
;
185 /* Dereference the given base pointer by the access chain */
186 static struct vtn_pointer
*
187 vtn_nir_deref_pointer_dereference(struct vtn_builder
*b
,
188 struct vtn_pointer
*base
,
189 struct vtn_access_chain
*deref_chain
)
191 struct vtn_type
*type
= base
->type
;
192 enum gl_access_qualifier access
= base
->access
;
195 nir_deref_instr
*tail
;
198 } else if (vtn_pointer_is_external_block(b
, base
)) {
199 nir_ssa_def
*block_index
= base
->block_index
;
201 /* We dereferencing an external block pointer. Correctness of this
202 * operation relies on one particular line in the SPIR-V spec, section
203 * entitled "Validation Rules for Shader Capabilities":
205 * "Block and BufferBlock decorations cannot decorate a structure
206 * type that is nested at any level inside another structure type
207 * decorated with Block or BufferBlock."
209 * This means that we can detect the point where we cross over from
210 * descriptor indexing to buffer indexing by looking for the block
211 * decorated struct type. Anything before the block decorated struct
212 * type is a descriptor indexing operation and anything after the block
213 * decorated struct is a buffer offset operation.
216 /* Figure out the descriptor array index if any
218 * Some of the Vulkan CTS tests with hand-rolled SPIR-V have been known
219 * to forget the Block or BufferBlock decoration from time to time.
220 * It's more robust if we check for both !block_index and for the type
221 * to contain a block. This way there's a decent chance that arrays of
222 * UBOs/SSBOs will work correctly even if variable pointers are
225 nir_ssa_def
*desc_arr_idx
= NULL
;
226 if (!block_index
|| vtn_type_contains_block(b
, type
)) {
227 /* If our type contains a block, then we're still outside the block
228 * and we need to process enough levels of dereferences to get inside
231 if (deref_chain
->ptr_as_array
) {
232 unsigned aoa_size
= glsl_get_aoa_size(type
->type
);
233 desc_arr_idx
= vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
234 MAX2(aoa_size
, 1), 32);
238 for (; idx
< deref_chain
->length
; idx
++) {
239 if (type
->base_type
!= vtn_base_type_array
) {
240 vtn_assert(type
->base_type
== vtn_base_type_struct
);
244 unsigned aoa_size
= glsl_get_aoa_size(type
->array_element
->type
);
245 nir_ssa_def
*arr_offset
=
246 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
247 MAX2(aoa_size
, 1), 32);
249 desc_arr_idx
= nir_iadd(&b
->nb
, desc_arr_idx
, arr_offset
);
251 desc_arr_idx
= arr_offset
;
253 type
= type
->array_element
;
254 access
|= type
->access
;
259 vtn_assert(base
->var
&& base
->type
);
260 block_index
= vtn_variable_resource_index(b
, base
->var
, desc_arr_idx
);
261 } else if (desc_arr_idx
) {
262 block_index
= vtn_resource_reindex(b
, base
->mode
,
263 block_index
, desc_arr_idx
);
266 if (idx
== deref_chain
->length
) {
267 /* The entire deref was consumed in finding the block index. Return
268 * a pointer which just has a block index and a later access chain
269 * will dereference deeper.
271 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
272 ptr
->mode
= base
->mode
;
274 ptr
->block_index
= block_index
;
275 ptr
->access
= access
;
279 /* If we got here, there's more access chain to handle and we have the
280 * final block index. Insert a descriptor load and cast to a deref to
281 * start the deref chain.
283 nir_ssa_def
*desc
= vtn_descriptor_load(b
, base
->mode
, block_index
);
285 assert(base
->mode
== vtn_variable_mode_ssbo
||
286 base
->mode
== vtn_variable_mode_ubo
);
287 nir_variable_mode nir_mode
=
288 base
->mode
== vtn_variable_mode_ssbo
? nir_var_mem_ssbo
: nir_var_mem_ubo
;
290 tail
= nir_build_deref_cast(&b
->nb
, desc
, nir_mode
, type
->type
,
291 base
->ptr_type
->stride
);
293 assert(base
->var
&& base
->var
->var
);
294 tail
= nir_build_deref_var(&b
->nb
, base
->var
->var
);
295 if (base
->ptr_type
&& base
->ptr_type
->type
) {
296 tail
->dest
.ssa
.num_components
=
297 glsl_get_vector_elements(base
->ptr_type
->type
);
298 tail
->dest
.ssa
.bit_size
= glsl_get_bit_size(base
->ptr_type
->type
);
302 if (idx
== 0 && deref_chain
->ptr_as_array
) {
303 /* We start with a deref cast to get the stride. Hopefully, we'll be
304 * able to delete that cast eventually.
306 tail
= nir_build_deref_cast(&b
->nb
, &tail
->dest
.ssa
, tail
->mode
,
307 tail
->type
, base
->ptr_type
->stride
);
309 nir_ssa_def
*index
= vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1,
310 tail
->dest
.ssa
.bit_size
);
311 tail
= nir_build_deref_ptr_as_array(&b
->nb
, tail
, index
);
315 for (; idx
< deref_chain
->length
; idx
++) {
316 if (glsl_type_is_struct_or_ifc(type
->type
)) {
317 vtn_assert(deref_chain
->link
[idx
].mode
== vtn_access_mode_literal
);
318 unsigned field
= deref_chain
->link
[idx
].id
;
319 tail
= nir_build_deref_struct(&b
->nb
, tail
, field
);
320 type
= type
->members
[field
];
322 nir_ssa_def
*arr_index
=
323 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
], 1,
324 tail
->dest
.ssa
.bit_size
);
325 tail
= nir_build_deref_array(&b
->nb
, tail
, arr_index
);
326 type
= type
->array_element
;
329 access
|= type
->access
;
332 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
333 ptr
->mode
= base
->mode
;
335 ptr
->var
= base
->var
;
337 ptr
->access
= access
;
342 static struct vtn_pointer
*
343 vtn_ssa_offset_pointer_dereference(struct vtn_builder
*b
,
344 struct vtn_pointer
*base
,
345 struct vtn_access_chain
*deref_chain
)
347 nir_ssa_def
*block_index
= base
->block_index
;
348 nir_ssa_def
*offset
= base
->offset
;
349 struct vtn_type
*type
= base
->type
;
350 enum gl_access_qualifier access
= base
->access
;
353 if (base
->mode
== vtn_variable_mode_ubo
||
354 base
->mode
== vtn_variable_mode_ssbo
) {
356 vtn_assert(base
->var
&& base
->type
);
357 nir_ssa_def
*desc_arr_idx
;
358 if (glsl_type_is_array(type
->type
)) {
359 if (deref_chain
->length
>= 1) {
361 vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
363 /* This consumes a level of type */
364 type
= type
->array_element
;
365 access
|= type
->access
;
367 /* This is annoying. We've been asked for a pointer to the
368 * array of UBOs/SSBOs and not a specifc buffer. Return a
369 * pointer with a descriptor index of 0 and we'll have to do
370 * a reindex later to adjust it to the right thing.
372 desc_arr_idx
= nir_imm_int(&b
->nb
, 0);
374 } else if (deref_chain
->ptr_as_array
) {
375 /* You can't have a zero-length OpPtrAccessChain */
376 vtn_assert(deref_chain
->length
>= 1);
377 desc_arr_idx
= vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
379 /* We have a regular non-array SSBO. */
382 block_index
= vtn_variable_resource_index(b
, base
->var
, desc_arr_idx
);
383 } else if (deref_chain
->ptr_as_array
&&
384 type
->base_type
== vtn_base_type_struct
&& type
->block
) {
385 /* We are doing an OpPtrAccessChain on a pointer to a struct that is
386 * decorated block. This is an interesting corner in the SPIR-V
387 * spec. One interpretation would be that they client is clearly
388 * trying to treat that block as if it's an implicit array of blocks
389 * repeated in the buffer. However, the SPIR-V spec for the
390 * OpPtrAccessChain says:
392 * "Base is treated as the address of the first element of an
393 * array, and the Element element’s address is computed to be the
394 * base for the Indexes, as per OpAccessChain."
396 * Taken literally, that would mean that your struct type is supposed
397 * to be treated as an array of such a struct and, since it's
398 * decorated block, that means an array of blocks which corresponds
399 * to an array descriptor. Therefore, we need to do a reindex
400 * operation to add the index from the first link in the access chain
401 * to the index we recieved.
403 * The downside to this interpretation (there always is one) is that
404 * this might be somewhat surprising behavior to apps if they expect
405 * the implicit array behavior described above.
407 vtn_assert(deref_chain
->length
>= 1);
408 nir_ssa_def
*offset_index
=
409 vtn_access_link_as_ssa(b
, deref_chain
->link
[0], 1, 32);
412 block_index
= vtn_resource_reindex(b
, base
->mode
,
413 block_index
, offset_index
);
418 if (base
->mode
== vtn_variable_mode_workgroup
) {
419 /* SLM doesn't need nor have a block index */
420 vtn_assert(!block_index
);
422 /* We need the variable for the base offset */
423 vtn_assert(base
->var
);
425 /* We need ptr_type for size and alignment */
426 vtn_assert(base
->ptr_type
);
428 /* Assign location on first use so that we don't end up bloating SLM
429 * address space for variables which are never statically used.
431 if (base
->var
->shared_location
< 0) {
432 vtn_assert(base
->ptr_type
->length
> 0 && base
->ptr_type
->align
> 0);
433 b
->shader
->num_shared
= vtn_align_u32(b
->shader
->num_shared
,
434 base
->ptr_type
->align
);
435 base
->var
->shared_location
= b
->shader
->num_shared
;
436 b
->shader
->num_shared
+= base
->ptr_type
->length
;
439 offset
= nir_imm_int(&b
->nb
, base
->var
->shared_location
);
440 } else if (base
->mode
== vtn_variable_mode_push_constant
) {
441 /* Push constants neither need nor have a block index */
442 vtn_assert(!block_index
);
444 /* Start off with at the start of the push constant block. */
445 offset
= nir_imm_int(&b
->nb
, 0);
447 /* The code above should have ensured a block_index when needed. */
448 vtn_assert(block_index
);
450 /* Start off with at the start of the buffer. */
451 offset
= nir_imm_int(&b
->nb
, 0);
455 if (deref_chain
->ptr_as_array
&& idx
== 0) {
456 /* We need ptr_type for the stride */
457 vtn_assert(base
->ptr_type
);
459 /* We need at least one element in the chain */
460 vtn_assert(deref_chain
->length
>= 1);
462 nir_ssa_def
*elem_offset
=
463 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
464 base
->ptr_type
->stride
, offset
->bit_size
);
465 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
469 for (; idx
< deref_chain
->length
; idx
++) {
470 switch (glsl_get_base_type(type
->type
)) {
473 case GLSL_TYPE_UINT16
:
474 case GLSL_TYPE_INT16
:
475 case GLSL_TYPE_UINT8
:
477 case GLSL_TYPE_UINT64
:
478 case GLSL_TYPE_INT64
:
479 case GLSL_TYPE_FLOAT
:
480 case GLSL_TYPE_FLOAT16
:
481 case GLSL_TYPE_DOUBLE
:
483 case GLSL_TYPE_ARRAY
: {
484 nir_ssa_def
*elem_offset
=
485 vtn_access_link_as_ssa(b
, deref_chain
->link
[idx
],
486 type
->stride
, offset
->bit_size
);
487 offset
= nir_iadd(&b
->nb
, offset
, elem_offset
);
488 type
= type
->array_element
;
489 access
|= type
->access
;
493 case GLSL_TYPE_STRUCT
: {
494 vtn_assert(deref_chain
->link
[idx
].mode
== vtn_access_mode_literal
);
495 unsigned member
= deref_chain
->link
[idx
].id
;
496 offset
= nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[member
]);
497 type
= type
->members
[member
];
498 access
|= type
->access
;
503 vtn_fail("Invalid type for deref");
507 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
508 ptr
->mode
= base
->mode
;
510 ptr
->block_index
= block_index
;
511 ptr
->offset
= offset
;
512 ptr
->access
= access
;
517 /* Dereference the given base pointer by the access chain */
518 static struct vtn_pointer
*
519 vtn_pointer_dereference(struct vtn_builder
*b
,
520 struct vtn_pointer
*base
,
521 struct vtn_access_chain
*deref_chain
)
523 if (vtn_pointer_uses_ssa_offset(b
, base
)) {
524 return vtn_ssa_offset_pointer_dereference(b
, base
, deref_chain
);
526 return vtn_nir_deref_pointer_dereference(b
, base
, deref_chain
);
531 vtn_pointer_for_variable(struct vtn_builder
*b
,
532 struct vtn_variable
*var
, struct vtn_type
*ptr_type
)
534 struct vtn_pointer
*pointer
= rzalloc(b
, struct vtn_pointer
);
536 pointer
->mode
= var
->mode
;
537 pointer
->type
= var
->type
;
538 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
539 vtn_assert(ptr_type
->deref
->type
== var
->type
->type
);
540 pointer
->ptr_type
= ptr_type
;
542 pointer
->access
= var
->access
| var
->type
->access
;
547 /* Returns an atomic_uint type based on the original uint type. The returned
548 * type will be equivalent to the original one but will have an atomic_uint
549 * type as leaf instead of an uint.
551 * Manages uint scalars, arrays, and arrays of arrays of any nested depth.
553 static const struct glsl_type
*
554 repair_atomic_type(const struct glsl_type
*type
)
556 assert(glsl_get_base_type(glsl_without_array(type
)) == GLSL_TYPE_UINT
);
557 assert(glsl_type_is_scalar(glsl_without_array(type
)));
559 if (glsl_type_is_array(type
)) {
560 const struct glsl_type
*atomic
=
561 repair_atomic_type(glsl_get_array_element(type
));
563 return glsl_array_type(atomic
, glsl_get_length(type
),
564 glsl_get_explicit_stride(type
));
566 return glsl_atomic_uint_type();
571 vtn_pointer_to_deref(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
573 if (b
->wa_glslang_179
) {
574 /* Do on-the-fly copy propagation for samplers. */
575 if (ptr
->var
&& ptr
->var
->copy_prop_sampler
)
576 return vtn_pointer_to_deref(b
, ptr
->var
->copy_prop_sampler
);
579 vtn_assert(!vtn_pointer_uses_ssa_offset(b
, ptr
));
581 struct vtn_access_chain chain
= {
584 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
591 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_instr
*deref
,
592 struct vtn_ssa_value
*inout
)
594 if (glsl_type_is_vector_or_scalar(deref
->type
)) {
596 inout
->def
= nir_load_deref(&b
->nb
, deref
);
598 nir_store_deref(&b
->nb
, deref
, inout
->def
, ~0);
600 } else if (glsl_type_is_array(deref
->type
) ||
601 glsl_type_is_matrix(deref
->type
)) {
602 unsigned elems
= glsl_get_length(deref
->type
);
603 for (unsigned i
= 0; i
< elems
; i
++) {
604 nir_deref_instr
*child
=
605 nir_build_deref_array_imm(&b
->nb
, deref
, i
);
606 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
]);
609 vtn_assert(glsl_type_is_struct_or_ifc(deref
->type
));
610 unsigned elems
= glsl_get_length(deref
->type
);
611 for (unsigned i
= 0; i
< elems
; i
++) {
612 nir_deref_instr
*child
= nir_build_deref_struct(&b
->nb
, deref
, i
);
613 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
]);
619 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
621 struct vtn_pointer
*ptr
= vtn_value(b
, id
, vtn_value_type_pointer
)->pointer
;
622 return vtn_pointer_to_deref(b
, ptr
);
626 * Gets the NIR-level deref tail, which may have as a child an array deref
627 * selecting which component due to OpAccessChain supporting per-component
628 * indexing in SPIR-V.
630 static nir_deref_instr
*
631 get_deref_tail(nir_deref_instr
*deref
)
633 if (deref
->deref_type
!= nir_deref_type_array
)
636 nir_deref_instr
*parent
=
637 nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
);
639 if (glsl_type_is_vector(parent
->type
))
645 struct vtn_ssa_value
*
646 vtn_local_load(struct vtn_builder
*b
, nir_deref_instr
*src
)
648 nir_deref_instr
*src_tail
= get_deref_tail(src
);
649 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
650 _vtn_local_load_store(b
, true, src_tail
, val
);
652 if (src_tail
!= src
) {
653 val
->type
= src
->type
;
654 if (nir_src_is_const(src
->arr
.index
))
655 val
->def
= vtn_vector_extract(b
, val
->def
,
656 nir_src_as_uint(src
->arr
.index
));
658 val
->def
= vtn_vector_extract_dynamic(b
, val
->def
, src
->arr
.index
.ssa
);
665 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
666 nir_deref_instr
*dest
)
668 nir_deref_instr
*dest_tail
= get_deref_tail(dest
);
670 if (dest_tail
!= dest
) {
671 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
672 _vtn_local_load_store(b
, true, dest_tail
, val
);
674 if (nir_src_is_const(dest
->arr
.index
))
675 val
->def
= vtn_vector_insert(b
, val
->def
, src
->def
,
676 nir_src_as_uint(dest
->arr
.index
));
678 val
->def
= vtn_vector_insert_dynamic(b
, val
->def
, src
->def
,
679 dest
->arr
.index
.ssa
);
680 _vtn_local_load_store(b
, false, dest_tail
, val
);
682 _vtn_local_load_store(b
, false, dest_tail
, src
);
687 vtn_pointer_to_offset(struct vtn_builder
*b
, struct vtn_pointer
*ptr
,
688 nir_ssa_def
**index_out
)
690 assert(vtn_pointer_uses_ssa_offset(b
, ptr
));
692 struct vtn_access_chain chain
= {
695 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
697 *index_out
= ptr
->block_index
;
701 /* Tries to compute the size of an interface block based on the strides and
702 * offsets that are provided to us in the SPIR-V source.
705 vtn_type_block_size(struct vtn_builder
*b
, struct vtn_type
*type
)
707 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
711 case GLSL_TYPE_UINT16
:
712 case GLSL_TYPE_INT16
:
713 case GLSL_TYPE_UINT8
:
715 case GLSL_TYPE_UINT64
:
716 case GLSL_TYPE_INT64
:
717 case GLSL_TYPE_FLOAT
:
718 case GLSL_TYPE_FLOAT16
:
720 case GLSL_TYPE_DOUBLE
: {
721 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
722 glsl_get_matrix_columns(type
->type
);
724 vtn_assert(type
->stride
> 0);
725 return type
->stride
* cols
;
727 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
728 return glsl_get_vector_elements(type
->type
) * type_size
;
732 case GLSL_TYPE_STRUCT
:
733 case GLSL_TYPE_INTERFACE
: {
735 unsigned num_fields
= glsl_get_length(type
->type
);
736 for (unsigned f
= 0; f
< num_fields
; f
++) {
737 unsigned field_end
= type
->offsets
[f
] +
738 vtn_type_block_size(b
, type
->members
[f
]);
739 size
= MAX2(size
, field_end
);
744 case GLSL_TYPE_ARRAY
:
745 vtn_assert(type
->stride
> 0);
746 vtn_assert(glsl_get_length(type
->type
) > 0);
747 return type
->stride
* glsl_get_length(type
->type
);
750 vtn_fail("Invalid block type");
756 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
757 nir_ssa_def
*index
, nir_ssa_def
*offset
,
758 unsigned access_offset
, unsigned access_size
,
759 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
,
760 enum gl_access_qualifier access
)
762 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
763 instr
->num_components
= glsl_get_vector_elements(type
);
765 /* Booleans usually shouldn't show up in external memory in SPIR-V.
766 * However, they do for certain older GLSLang versions and can for shared
767 * memory when we lower access chains internally.
769 const unsigned data_bit_size
= glsl_type_is_boolean(type
) ? 32 :
770 glsl_get_bit_size(type
);
774 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
775 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
778 if (op
== nir_intrinsic_load_push_constant
) {
779 nir_intrinsic_set_base(instr
, access_offset
);
780 nir_intrinsic_set_range(instr
, access_size
);
783 if (op
== nir_intrinsic_load_ssbo
||
784 op
== nir_intrinsic_store_ssbo
) {
785 nir_intrinsic_set_access(instr
, access
);
788 /* With extensions like relaxed_block_layout, we really can't guarantee
789 * much more than scalar alignment.
791 if (op
!= nir_intrinsic_load_push_constant
)
792 nir_intrinsic_set_align(instr
, data_bit_size
/ 8, 0);
795 instr
->src
[src
++] = nir_src_for_ssa(index
);
797 if (op
== nir_intrinsic_load_push_constant
) {
798 /* We need to subtract the offset from where the intrinsic will load the
801 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
802 nir_imm_int(&b
->nb
, access_offset
)));
804 instr
->src
[src
++] = nir_src_for_ssa(offset
);
808 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
809 instr
->num_components
, data_bit_size
, NULL
);
810 (*inout
)->def
= &instr
->dest
.ssa
;
813 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
815 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
816 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
820 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
821 nir_ssa_def
*index
, nir_ssa_def
*offset
,
822 unsigned access_offset
, unsigned access_size
,
823 struct vtn_type
*type
, enum gl_access_qualifier access
,
824 struct vtn_ssa_value
**inout
)
826 if (load
&& *inout
== NULL
)
827 *inout
= vtn_create_ssa_value(b
, type
->type
);
829 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
833 case GLSL_TYPE_UINT16
:
834 case GLSL_TYPE_INT16
:
835 case GLSL_TYPE_UINT8
:
837 case GLSL_TYPE_UINT64
:
838 case GLSL_TYPE_INT64
:
839 case GLSL_TYPE_FLOAT
:
840 case GLSL_TYPE_FLOAT16
:
841 case GLSL_TYPE_DOUBLE
:
843 /* This is where things get interesting. At this point, we've hit
844 * a vector, a scalar, or a matrix.
846 if (glsl_type_is_matrix(type
->type
)) {
847 /* Loading the whole matrix */
848 struct vtn_ssa_value
*transpose
;
849 unsigned num_ops
, vec_width
, col_stride
;
850 if (type
->row_major
) {
851 num_ops
= glsl_get_vector_elements(type
->type
);
852 vec_width
= glsl_get_matrix_columns(type
->type
);
853 col_stride
= type
->array_element
->stride
;
855 const struct glsl_type
*transpose_type
=
856 glsl_matrix_type(base_type
, vec_width
, num_ops
);
857 *inout
= vtn_create_ssa_value(b
, transpose_type
);
859 transpose
= vtn_ssa_transpose(b
, *inout
);
863 num_ops
= glsl_get_matrix_columns(type
->type
);
864 vec_width
= glsl_get_vector_elements(type
->type
);
865 col_stride
= type
->stride
;
868 for (unsigned i
= 0; i
< num_ops
; i
++) {
869 nir_ssa_def
*elem_offset
=
870 nir_iadd_imm(&b
->nb
, offset
, i
* col_stride
);
871 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
872 access_offset
, access_size
,
874 glsl_vector_type(base_type
, vec_width
),
875 type
->access
| access
);
878 if (load
&& type
->row_major
)
879 *inout
= vtn_ssa_transpose(b
, *inout
);
881 unsigned elems
= glsl_get_vector_elements(type
->type
);
882 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
883 if (elems
== 1 || type
->stride
== type_size
) {
884 /* This is a tightly-packed normal scalar or vector load */
885 vtn_assert(glsl_type_is_vector_or_scalar(type
->type
));
886 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
887 access_offset
, access_size
,
889 type
->access
| access
);
891 /* This is a strided load. We have to load N things separately.
892 * This is the single column of a row-major matrix case.
894 vtn_assert(type
->stride
> type_size
);
895 vtn_assert(type
->stride
% type_size
== 0);
897 nir_ssa_def
*per_comp
[4];
898 for (unsigned i
= 0; i
< elems
; i
++) {
899 nir_ssa_def
*elem_offset
=
900 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
901 struct vtn_ssa_value
*comp
, temp_val
;
903 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
904 temp_val
.type
= glsl_scalar_type(base_type
);
907 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
908 access_offset
, access_size
,
909 &comp
, glsl_scalar_type(base_type
),
910 type
->access
| access
);
911 per_comp
[i
] = comp
->def
;
916 *inout
= vtn_create_ssa_value(b
, type
->type
);
917 (*inout
)->def
= nir_vec(&b
->nb
, per_comp
, elems
);
923 case GLSL_TYPE_ARRAY
: {
924 unsigned elems
= glsl_get_length(type
->type
);
925 for (unsigned i
= 0; i
< elems
; i
++) {
926 nir_ssa_def
*elem_off
=
927 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
928 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
929 access_offset
, access_size
,
931 type
->array_element
->access
| access
,
932 &(*inout
)->elems
[i
]);
937 case GLSL_TYPE_STRUCT
: {
938 unsigned elems
= glsl_get_length(type
->type
);
939 for (unsigned i
= 0; i
< elems
; i
++) {
940 nir_ssa_def
*elem_off
=
941 nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[i
]);
942 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
943 access_offset
, access_size
,
945 type
->members
[i
]->access
| access
,
946 &(*inout
)->elems
[i
]);
952 vtn_fail("Invalid block member type");
956 static struct vtn_ssa_value
*
957 vtn_block_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
960 unsigned access_offset
= 0, access_size
= 0;
962 case vtn_variable_mode_ubo
:
963 op
= nir_intrinsic_load_ubo
;
965 case vtn_variable_mode_ssbo
:
966 op
= nir_intrinsic_load_ssbo
;
968 case vtn_variable_mode_push_constant
:
969 op
= nir_intrinsic_load_push_constant
;
970 access_size
= b
->shader
->num_uniforms
;
972 case vtn_variable_mode_workgroup
:
973 op
= nir_intrinsic_load_shared
;
976 vtn_fail("Invalid block variable mode");
979 nir_ssa_def
*offset
, *index
= NULL
;
980 offset
= vtn_pointer_to_offset(b
, src
, &index
);
982 struct vtn_ssa_value
*value
= NULL
;
983 _vtn_block_load_store(b
, op
, true, index
, offset
,
984 access_offset
, access_size
,
985 src
->type
, src
->access
, &value
);
990 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
991 struct vtn_pointer
*dst
)
995 case vtn_variable_mode_ssbo
:
996 op
= nir_intrinsic_store_ssbo
;
998 case vtn_variable_mode_workgroup
:
999 op
= nir_intrinsic_store_shared
;
1002 vtn_fail("Invalid block variable mode");
1005 nir_ssa_def
*offset
, *index
= NULL
;
1006 offset
= vtn_pointer_to_offset(b
, dst
, &index
);
1008 _vtn_block_load_store(b
, op
, false, index
, offset
,
1009 0, 0, dst
->type
, dst
->access
, &src
);
1013 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
1014 struct vtn_pointer
*ptr
,
1015 struct vtn_ssa_value
**inout
)
1017 enum glsl_base_type base_type
= glsl_get_base_type(ptr
->type
->type
);
1018 switch (base_type
) {
1019 case GLSL_TYPE_UINT
:
1021 case GLSL_TYPE_UINT16
:
1022 case GLSL_TYPE_INT16
:
1023 case GLSL_TYPE_UINT8
:
1024 case GLSL_TYPE_INT8
:
1025 case GLSL_TYPE_UINT64
:
1026 case GLSL_TYPE_INT64
:
1027 case GLSL_TYPE_FLOAT
:
1028 case GLSL_TYPE_FLOAT16
:
1029 case GLSL_TYPE_BOOL
:
1030 case GLSL_TYPE_DOUBLE
:
1031 if (glsl_type_is_vector_or_scalar(ptr
->type
->type
)) {
1032 /* We hit a vector or scalar; go ahead and emit the load[s] */
1033 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
1034 if (vtn_pointer_is_external_block(b
, ptr
)) {
1035 /* If it's external, we call nir_load/store_deref directly. The
1036 * vtn_local_load/store helpers are too clever and do magic to
1037 * avoid array derefs of vectors. That magic is both less
1038 * efficient than the direct load/store and, in the case of
1039 * stores, is broken because it creates a race condition if two
1040 * threads are writing to different components of the same vector
1041 * due to the load+insert+store it uses to emulate the array
1045 *inout
= vtn_create_ssa_value(b
, ptr
->type
->type
);
1046 (*inout
)->def
= nir_load_deref(&b
->nb
, deref
);
1048 nir_store_deref(&b
->nb
, deref
, (*inout
)->def
, ~0);
1052 *inout
= vtn_local_load(b
, deref
);
1054 vtn_local_store(b
, *inout
, deref
);
1061 case GLSL_TYPE_ARRAY
:
1062 case GLSL_TYPE_STRUCT
: {
1063 unsigned elems
= glsl_get_length(ptr
->type
->type
);
1065 vtn_assert(*inout
== NULL
);
1066 *inout
= rzalloc(b
, struct vtn_ssa_value
);
1067 (*inout
)->type
= ptr
->type
->type
;
1068 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
1071 struct vtn_access_chain chain
= {
1074 { .mode
= vtn_access_mode_literal
, },
1077 for (unsigned i
= 0; i
< elems
; i
++) {
1078 chain
.link
[0].id
= i
;
1079 struct vtn_pointer
*elem
= vtn_pointer_dereference(b
, ptr
, &chain
);
1080 _vtn_variable_load_store(b
, load
, elem
, &(*inout
)->elems
[i
]);
1086 vtn_fail("Invalid access chain type");
1090 struct vtn_ssa_value
*
1091 vtn_variable_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
1093 if (vtn_pointer_uses_ssa_offset(b
, src
)) {
1094 return vtn_block_load(b
, src
);
1096 struct vtn_ssa_value
*val
= NULL
;
1097 _vtn_variable_load_store(b
, true, src
, &val
);
1103 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1104 struct vtn_pointer
*dest
)
1106 if (vtn_pointer_uses_ssa_offset(b
, dest
)) {
1107 vtn_assert(dest
->mode
== vtn_variable_mode_ssbo
||
1108 dest
->mode
== vtn_variable_mode_workgroup
);
1109 vtn_block_store(b
, src
, dest
);
1111 _vtn_variable_load_store(b
, false, dest
, &src
);
1116 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1117 struct vtn_pointer
*src
)
1119 vtn_assert(src
->type
->type
== dest
->type
->type
);
1120 enum glsl_base_type base_type
= glsl_get_base_type(src
->type
->type
);
1121 switch (base_type
) {
1122 case GLSL_TYPE_UINT
:
1124 case GLSL_TYPE_UINT16
:
1125 case GLSL_TYPE_INT16
:
1126 case GLSL_TYPE_UINT8
:
1127 case GLSL_TYPE_INT8
:
1128 case GLSL_TYPE_UINT64
:
1129 case GLSL_TYPE_INT64
:
1130 case GLSL_TYPE_FLOAT
:
1131 case GLSL_TYPE_FLOAT16
:
1132 case GLSL_TYPE_DOUBLE
:
1133 case GLSL_TYPE_BOOL
:
1134 /* At this point, we have a scalar, vector, or matrix so we know that
1135 * there cannot be any structure splitting still in the way. By
1136 * stopping at the matrix level rather than the vector level, we
1137 * ensure that matrices get loaded in the optimal way even if they
1138 * are storred row-major in a UBO.
1140 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
1143 case GLSL_TYPE_ARRAY
:
1144 case GLSL_TYPE_STRUCT
: {
1145 struct vtn_access_chain chain
= {
1148 { .mode
= vtn_access_mode_literal
, },
1151 unsigned elems
= glsl_get_length(src
->type
->type
);
1152 for (unsigned i
= 0; i
< elems
; i
++) {
1153 chain
.link
[0].id
= i
;
1154 struct vtn_pointer
*src_elem
=
1155 vtn_pointer_dereference(b
, src
, &chain
);
1156 struct vtn_pointer
*dest_elem
=
1157 vtn_pointer_dereference(b
, dest
, &chain
);
1159 _vtn_variable_copy(b
, dest_elem
, src_elem
);
1165 vtn_fail("Invalid access chain type");
1170 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1171 struct vtn_pointer
*src
)
1173 /* TODO: At some point, we should add a special-case for when we can
1174 * just emit a copy_var intrinsic.
1176 _vtn_variable_copy(b
, dest
, src
);
1180 set_mode_system_value(struct vtn_builder
*b
, nir_variable_mode
*mode
)
1182 vtn_assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
1183 *mode
= nir_var_system_value
;
1187 vtn_get_builtin_location(struct vtn_builder
*b
,
1188 SpvBuiltIn builtin
, int *location
,
1189 nir_variable_mode
*mode
)
1192 case SpvBuiltInPosition
:
1193 *location
= VARYING_SLOT_POS
;
1195 case SpvBuiltInPointSize
:
1196 *location
= VARYING_SLOT_PSIZ
;
1198 case SpvBuiltInClipDistance
:
1199 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
1201 case SpvBuiltInCullDistance
:
1202 *location
= VARYING_SLOT_CULL_DIST0
;
1204 case SpvBuiltInVertexId
:
1205 case SpvBuiltInVertexIndex
:
1206 /* The Vulkan spec defines VertexIndex to be non-zero-based and doesn't
1207 * allow VertexId. The ARB_gl_spirv spec defines VertexId to be the
1208 * same as gl_VertexID, which is non-zero-based, and removes
1209 * VertexIndex. Since they're both defined to be non-zero-based, we use
1210 * SYSTEM_VALUE_VERTEX_ID for both.
1212 *location
= SYSTEM_VALUE_VERTEX_ID
;
1213 set_mode_system_value(b
, mode
);
1215 case SpvBuiltInInstanceIndex
:
1216 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
1217 set_mode_system_value(b
, mode
);
1219 case SpvBuiltInInstanceId
:
1220 *location
= SYSTEM_VALUE_INSTANCE_ID
;
1221 set_mode_system_value(b
, mode
);
1223 case SpvBuiltInPrimitiveId
:
1224 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1225 vtn_assert(*mode
== nir_var_shader_in
);
1226 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1227 } else if (*mode
== nir_var_shader_out
) {
1228 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1230 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
1231 set_mode_system_value(b
, mode
);
1234 case SpvBuiltInInvocationId
:
1235 *location
= SYSTEM_VALUE_INVOCATION_ID
;
1236 set_mode_system_value(b
, mode
);
1238 case SpvBuiltInLayer
:
1239 *location
= VARYING_SLOT_LAYER
;
1240 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1241 *mode
= nir_var_shader_in
;
1242 else if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1243 *mode
= nir_var_shader_out
;
1244 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1245 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1246 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1247 *mode
= nir_var_shader_out
;
1249 vtn_fail("invalid stage for SpvBuiltInLayer");
1251 case SpvBuiltInViewportIndex
:
1252 *location
= VARYING_SLOT_VIEWPORT
;
1253 if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1254 *mode
= nir_var_shader_out
;
1255 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1256 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1257 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1258 *mode
= nir_var_shader_out
;
1259 else if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1260 *mode
= nir_var_shader_in
;
1262 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1264 case SpvBuiltInTessLevelOuter
:
1265 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
1267 case SpvBuiltInTessLevelInner
:
1268 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
1270 case SpvBuiltInTessCoord
:
1271 *location
= SYSTEM_VALUE_TESS_COORD
;
1272 set_mode_system_value(b
, mode
);
1274 case SpvBuiltInPatchVertices
:
1275 *location
= SYSTEM_VALUE_VERTICES_IN
;
1276 set_mode_system_value(b
, mode
);
1278 case SpvBuiltInFragCoord
:
1279 *location
= VARYING_SLOT_POS
;
1280 vtn_assert(*mode
== nir_var_shader_in
);
1282 case SpvBuiltInPointCoord
:
1283 *location
= VARYING_SLOT_PNTC
;
1284 vtn_assert(*mode
== nir_var_shader_in
);
1286 case SpvBuiltInFrontFacing
:
1287 *location
= SYSTEM_VALUE_FRONT_FACE
;
1288 set_mode_system_value(b
, mode
);
1290 case SpvBuiltInSampleId
:
1291 *location
= SYSTEM_VALUE_SAMPLE_ID
;
1292 set_mode_system_value(b
, mode
);
1294 case SpvBuiltInSamplePosition
:
1295 *location
= SYSTEM_VALUE_SAMPLE_POS
;
1296 set_mode_system_value(b
, mode
);
1298 case SpvBuiltInSampleMask
:
1299 if (*mode
== nir_var_shader_out
) {
1300 *location
= FRAG_RESULT_SAMPLE_MASK
;
1302 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
1303 set_mode_system_value(b
, mode
);
1306 case SpvBuiltInFragDepth
:
1307 *location
= FRAG_RESULT_DEPTH
;
1308 vtn_assert(*mode
== nir_var_shader_out
);
1310 case SpvBuiltInHelperInvocation
:
1311 *location
= SYSTEM_VALUE_HELPER_INVOCATION
;
1312 set_mode_system_value(b
, mode
);
1314 case SpvBuiltInNumWorkgroups
:
1315 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
1316 set_mode_system_value(b
, mode
);
1318 case SpvBuiltInWorkgroupSize
:
1319 *location
= SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
1320 set_mode_system_value(b
, mode
);
1322 case SpvBuiltInWorkgroupId
:
1323 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
1324 set_mode_system_value(b
, mode
);
1326 case SpvBuiltInLocalInvocationId
:
1327 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
1328 set_mode_system_value(b
, mode
);
1330 case SpvBuiltInLocalInvocationIndex
:
1331 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
1332 set_mode_system_value(b
, mode
);
1334 case SpvBuiltInGlobalInvocationId
:
1335 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
1336 set_mode_system_value(b
, mode
);
1338 case SpvBuiltInGlobalLinearId
:
1339 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX
;
1340 set_mode_system_value(b
, mode
);
1342 case SpvBuiltInBaseVertex
:
1343 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1344 * semantic as SPIR-V BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1346 *location
= SYSTEM_VALUE_FIRST_VERTEX
;
1347 set_mode_system_value(b
, mode
);
1349 case SpvBuiltInBaseInstance
:
1350 *location
= SYSTEM_VALUE_BASE_INSTANCE
;
1351 set_mode_system_value(b
, mode
);
1353 case SpvBuiltInDrawIndex
:
1354 *location
= SYSTEM_VALUE_DRAW_ID
;
1355 set_mode_system_value(b
, mode
);
1357 case SpvBuiltInSubgroupSize
:
1358 *location
= SYSTEM_VALUE_SUBGROUP_SIZE
;
1359 set_mode_system_value(b
, mode
);
1361 case SpvBuiltInSubgroupId
:
1362 *location
= SYSTEM_VALUE_SUBGROUP_ID
;
1363 set_mode_system_value(b
, mode
);
1365 case SpvBuiltInSubgroupLocalInvocationId
:
1366 *location
= SYSTEM_VALUE_SUBGROUP_INVOCATION
;
1367 set_mode_system_value(b
, mode
);
1369 case SpvBuiltInNumSubgroups
:
1370 *location
= SYSTEM_VALUE_NUM_SUBGROUPS
;
1371 set_mode_system_value(b
, mode
);
1373 case SpvBuiltInDeviceIndex
:
1374 *location
= SYSTEM_VALUE_DEVICE_INDEX
;
1375 set_mode_system_value(b
, mode
);
1377 case SpvBuiltInViewIndex
:
1378 *location
= SYSTEM_VALUE_VIEW_INDEX
;
1379 set_mode_system_value(b
, mode
);
1381 case SpvBuiltInSubgroupEqMask
:
1382 *location
= SYSTEM_VALUE_SUBGROUP_EQ_MASK
,
1383 set_mode_system_value(b
, mode
);
1385 case SpvBuiltInSubgroupGeMask
:
1386 *location
= SYSTEM_VALUE_SUBGROUP_GE_MASK
,
1387 set_mode_system_value(b
, mode
);
1389 case SpvBuiltInSubgroupGtMask
:
1390 *location
= SYSTEM_VALUE_SUBGROUP_GT_MASK
,
1391 set_mode_system_value(b
, mode
);
1393 case SpvBuiltInSubgroupLeMask
:
1394 *location
= SYSTEM_VALUE_SUBGROUP_LE_MASK
,
1395 set_mode_system_value(b
, mode
);
1397 case SpvBuiltInSubgroupLtMask
:
1398 *location
= SYSTEM_VALUE_SUBGROUP_LT_MASK
,
1399 set_mode_system_value(b
, mode
);
1401 case SpvBuiltInFragStencilRefEXT
:
1402 *location
= FRAG_RESULT_STENCIL
;
1403 vtn_assert(*mode
== nir_var_shader_out
);
1405 case SpvBuiltInWorkDim
:
1406 *location
= SYSTEM_VALUE_WORK_DIM
;
1407 set_mode_system_value(b
, mode
);
1409 case SpvBuiltInGlobalSize
:
1410 *location
= SYSTEM_VALUE_GLOBAL_GROUP_SIZE
;
1411 set_mode_system_value(b
, mode
);
1414 vtn_fail("unsupported builtin: %u", builtin
);
1419 apply_var_decoration(struct vtn_builder
*b
,
1420 struct nir_variable_data
*var_data
,
1421 const struct vtn_decoration
*dec
)
1423 switch (dec
->decoration
) {
1424 case SpvDecorationRelaxedPrecision
:
1425 break; /* FIXME: Do nothing with this for now. */
1426 case SpvDecorationNoPerspective
:
1427 var_data
->interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1429 case SpvDecorationFlat
:
1430 var_data
->interpolation
= INTERP_MODE_FLAT
;
1432 case SpvDecorationCentroid
:
1433 var_data
->centroid
= true;
1435 case SpvDecorationSample
:
1436 var_data
->sample
= true;
1438 case SpvDecorationInvariant
:
1439 var_data
->invariant
= true;
1441 case SpvDecorationConstant
:
1442 var_data
->read_only
= true;
1444 case SpvDecorationNonReadable
:
1445 var_data
->image
.access
|= ACCESS_NON_READABLE
;
1447 case SpvDecorationNonWritable
:
1448 var_data
->read_only
= true;
1449 var_data
->image
.access
|= ACCESS_NON_WRITEABLE
;
1451 case SpvDecorationRestrict
:
1452 var_data
->image
.access
|= ACCESS_RESTRICT
;
1454 case SpvDecorationVolatile
:
1455 var_data
->image
.access
|= ACCESS_VOLATILE
;
1457 case SpvDecorationCoherent
:
1458 var_data
->image
.access
|= ACCESS_COHERENT
;
1460 case SpvDecorationComponent
:
1461 var_data
->location_frac
= dec
->literals
[0];
1463 case SpvDecorationIndex
:
1464 var_data
->index
= dec
->literals
[0];
1466 case SpvDecorationBuiltIn
: {
1467 SpvBuiltIn builtin
= dec
->literals
[0];
1469 nir_variable_mode mode
= var_data
->mode
;
1470 vtn_get_builtin_location(b
, builtin
, &var_data
->location
, &mode
);
1471 var_data
->mode
= mode
;
1474 case SpvBuiltInTessLevelOuter
:
1475 case SpvBuiltInTessLevelInner
:
1476 case SpvBuiltInClipDistance
:
1477 case SpvBuiltInCullDistance
:
1478 var_data
->compact
= true;
1485 case SpvDecorationSpecId
:
1486 case SpvDecorationRowMajor
:
1487 case SpvDecorationColMajor
:
1488 case SpvDecorationMatrixStride
:
1489 case SpvDecorationAliased
:
1490 case SpvDecorationUniform
:
1491 case SpvDecorationLinkageAttributes
:
1492 break; /* Do nothing with these here */
1494 case SpvDecorationPatch
:
1495 var_data
->patch
= true;
1498 case SpvDecorationLocation
:
1499 vtn_fail("Handled above");
1501 case SpvDecorationBlock
:
1502 case SpvDecorationBufferBlock
:
1503 case SpvDecorationArrayStride
:
1504 case SpvDecorationGLSLShared
:
1505 case SpvDecorationGLSLPacked
:
1506 break; /* These can apply to a type but we don't care about them */
1508 case SpvDecorationBinding
:
1509 case SpvDecorationDescriptorSet
:
1510 case SpvDecorationNoContraction
:
1511 case SpvDecorationInputAttachmentIndex
:
1512 vtn_warn("Decoration not allowed for variable or structure member: %s",
1513 spirv_decoration_to_string(dec
->decoration
));
1516 case SpvDecorationXfbBuffer
:
1517 var_data
->explicit_xfb_buffer
= true;
1518 var_data
->xfb_buffer
= dec
->literals
[0];
1519 var_data
->always_active_io
= true;
1521 case SpvDecorationXfbStride
:
1522 var_data
->explicit_xfb_stride
= true;
1523 var_data
->xfb_stride
= dec
->literals
[0];
1525 case SpvDecorationOffset
:
1526 var_data
->explicit_offset
= true;
1527 var_data
->offset
= dec
->literals
[0];
1530 case SpvDecorationStream
:
1531 var_data
->stream
= dec
->literals
[0];
1534 case SpvDecorationCPacked
:
1535 case SpvDecorationSaturatedConversion
:
1536 case SpvDecorationFuncParamAttr
:
1537 case SpvDecorationFPRoundingMode
:
1538 case SpvDecorationFPFastMathMode
:
1539 case SpvDecorationAlignment
:
1540 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
1541 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1542 spirv_decoration_to_string(dec
->decoration
));
1546 case SpvDecorationHlslSemanticGOOGLE
:
1547 /* HLSL semantic decorations can safely be ignored by the driver. */
1550 case SpvDecorationRestrictPointerEXT
:
1551 case SpvDecorationAliasedPointerEXT
:
1552 /* TODO: We should actually plumb alias information through NIR. */
1556 vtn_fail("Unhandled decoration");
1561 var_is_patch_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1562 const struct vtn_decoration
*dec
, void *out_is_patch
)
1564 if (dec
->decoration
== SpvDecorationPatch
) {
1565 *((bool *) out_is_patch
) = true;
1570 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1571 const struct vtn_decoration
*dec
, void *void_var
)
1573 struct vtn_variable
*vtn_var
= void_var
;
1575 /* Handle decorations that apply to a vtn_variable as a whole */
1576 switch (dec
->decoration
) {
1577 case SpvDecorationBinding
:
1578 vtn_var
->binding
= dec
->literals
[0];
1579 vtn_var
->explicit_binding
= true;
1581 case SpvDecorationDescriptorSet
:
1582 vtn_var
->descriptor_set
= dec
->literals
[0];
1584 case SpvDecorationInputAttachmentIndex
:
1585 vtn_var
->input_attachment_index
= dec
->literals
[0];
1587 case SpvDecorationPatch
:
1588 vtn_var
->patch
= true;
1590 case SpvDecorationOffset
:
1591 vtn_var
->offset
= dec
->literals
[0];
1593 case SpvDecorationNonWritable
:
1594 vtn_var
->access
|= ACCESS_NON_WRITEABLE
;
1596 case SpvDecorationNonReadable
:
1597 vtn_var
->access
|= ACCESS_NON_READABLE
;
1599 case SpvDecorationVolatile
:
1600 vtn_var
->access
|= ACCESS_VOLATILE
;
1602 case SpvDecorationCoherent
:
1603 vtn_var
->access
|= ACCESS_COHERENT
;
1605 case SpvDecorationHlslCounterBufferGOOGLE
:
1606 /* HLSL semantic decorations can safely be ignored by the driver. */
1612 if (val
->value_type
== vtn_value_type_pointer
) {
1613 assert(val
->pointer
->var
== void_var
);
1614 assert(member
== -1);
1616 assert(val
->value_type
== vtn_value_type_type
);
1619 /* Location is odd. If applied to a split structure, we have to walk the
1620 * whole thing and accumulate the location. It's easier to handle as a
1623 if (dec
->decoration
== SpvDecorationLocation
) {
1624 unsigned location
= dec
->literals
[0];
1625 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
1626 vtn_var
->mode
== vtn_variable_mode_output
) {
1627 location
+= FRAG_RESULT_DATA0
;
1628 } else if (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
&&
1629 vtn_var
->mode
== vtn_variable_mode_input
) {
1630 location
+= VERT_ATTRIB_GENERIC0
;
1631 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1632 vtn_var
->mode
== vtn_variable_mode_output
) {
1633 location
+= vtn_var
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
;
1634 } else if (vtn_var
->mode
!= vtn_variable_mode_uniform
) {
1635 vtn_warn("Location must be on input, output, uniform, sampler or "
1640 if (vtn_var
->var
->num_members
== 0) {
1641 /* This handles the member and lone variable cases */
1642 vtn_var
->var
->data
.location
= location
;
1644 /* This handles the structure member case */
1645 assert(vtn_var
->var
->members
);
1648 vtn_var
->base_location
= location
;
1650 vtn_var
->var
->members
[member
].location
= location
;
1656 if (vtn_var
->var
->num_members
== 0) {
1657 /* We call this function on types as well as variables and not all
1658 * struct types get split so we can end up having stray member
1659 * decorations; just ignore them.
1662 apply_var_decoration(b
, &vtn_var
->var
->data
, dec
);
1663 } else if (member
>= 0) {
1664 /* Member decorations must come from a type */
1665 assert(val
->value_type
== vtn_value_type_type
);
1666 apply_var_decoration(b
, &vtn_var
->var
->members
[member
], dec
);
1669 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1670 for (unsigned i
= 0; i
< length
; i
++)
1671 apply_var_decoration(b
, &vtn_var
->var
->members
[i
], dec
);
1674 /* A few variables, those with external storage, have no actual
1675 * nir_variables associated with them. Fortunately, all decorations
1676 * we care about for those variables are on the type only.
1678 vtn_assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1679 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1680 vtn_var
->mode
== vtn_variable_mode_push_constant
||
1681 (vtn_var
->mode
== vtn_variable_mode_workgroup
&&
1682 b
->options
->lower_workgroup_access_to_offsets
));
1687 static enum vtn_variable_mode
1688 vtn_storage_class_to_mode(struct vtn_builder
*b
,
1689 SpvStorageClass
class,
1690 struct vtn_type
*interface_type
,
1691 nir_variable_mode
*nir_mode_out
)
1693 enum vtn_variable_mode mode
;
1694 nir_variable_mode nir_mode
;
1696 case SpvStorageClassUniform
:
1697 if (interface_type
->block
) {
1698 mode
= vtn_variable_mode_ubo
;
1699 nir_mode
= nir_var_mem_ubo
;
1700 } else if (interface_type
->buffer_block
) {
1701 mode
= vtn_variable_mode_ssbo
;
1702 nir_mode
= nir_var_mem_ssbo
;
1704 /* Default-block uniforms, coming from gl_spirv */
1705 mode
= vtn_variable_mode_uniform
;
1706 nir_mode
= nir_var_uniform
;
1709 case SpvStorageClassStorageBuffer
:
1710 mode
= vtn_variable_mode_ssbo
;
1711 nir_mode
= nir_var_mem_ssbo
;
1713 case SpvStorageClassPhysicalStorageBufferEXT
:
1714 mode
= vtn_variable_mode_phys_ssbo
;
1715 nir_mode
= nir_var_mem_global
;
1717 case SpvStorageClassUniformConstant
:
1718 mode
= vtn_variable_mode_uniform
;
1719 nir_mode
= nir_var_uniform
;
1721 case SpvStorageClassPushConstant
:
1722 mode
= vtn_variable_mode_push_constant
;
1723 nir_mode
= nir_var_uniform
;
1725 case SpvStorageClassInput
:
1726 mode
= vtn_variable_mode_input
;
1727 nir_mode
= nir_var_shader_in
;
1729 case SpvStorageClassOutput
:
1730 mode
= vtn_variable_mode_output
;
1731 nir_mode
= nir_var_shader_out
;
1733 case SpvStorageClassPrivate
:
1734 mode
= vtn_variable_mode_private
;
1735 nir_mode
= nir_var_shader_temp
;
1737 case SpvStorageClassFunction
:
1738 mode
= vtn_variable_mode_function
;
1739 nir_mode
= nir_var_function_temp
;
1741 case SpvStorageClassWorkgroup
:
1742 mode
= vtn_variable_mode_workgroup
;
1743 nir_mode
= nir_var_mem_shared
;
1745 case SpvStorageClassAtomicCounter
:
1746 mode
= vtn_variable_mode_uniform
;
1747 nir_mode
= nir_var_uniform
;
1749 case SpvStorageClassCrossWorkgroup
:
1750 mode
= vtn_variable_mode_cross_workgroup
;
1751 nir_mode
= nir_var_mem_global
;
1753 case SpvStorageClassGeneric
:
1755 vtn_fail("Unhandled variable storage class");
1759 *nir_mode_out
= nir_mode
;
1765 vtn_pointer_to_ssa(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
1767 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1768 /* This pointer needs to have a pointer type with actual storage */
1769 vtn_assert(ptr
->ptr_type
);
1770 vtn_assert(ptr
->ptr_type
->type
);
1773 /* If we don't have an offset then we must be a pointer to the variable
1776 vtn_assert(!ptr
->offset
&& !ptr
->block_index
);
1778 struct vtn_access_chain chain
= {
1781 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
1784 vtn_assert(ptr
->offset
);
1785 if (ptr
->block_index
) {
1786 vtn_assert(ptr
->mode
== vtn_variable_mode_ubo
||
1787 ptr
->mode
== vtn_variable_mode_ssbo
);
1788 return nir_vec2(&b
->nb
, ptr
->block_index
, ptr
->offset
);
1790 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
);
1794 if (vtn_pointer_is_external_block(b
, ptr
) &&
1795 vtn_type_contains_block(b
, ptr
->type
) &&
1796 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
1797 /* In this case, we're looking for a block index and not an actual
1800 * For PhysicalStorageBufferEXT pointers, we don't have a block index
1801 * at all because we get the pointer directly from the client. This
1802 * assumes that there will never be a SSBO binding variable using the
1803 * PhysicalStorageBufferEXT storage class. This assumption appears
1804 * to be correct according to the Vulkan spec because the table,
1805 * "Shader Resource and Storage Class Correspondence," the only the
1806 * Uniform storage class with BufferBlock or the StorageBuffer
1807 * storage class with Block can be used.
1809 if (!ptr
->block_index
) {
1810 /* If we don't have a block_index then we must be a pointer to the
1813 vtn_assert(!ptr
->deref
);
1815 struct vtn_access_chain chain
= {
1818 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
1821 return ptr
->block_index
;
1823 return &vtn_pointer_to_deref(b
, ptr
)->dest
.ssa
;
1828 struct vtn_pointer
*
1829 vtn_pointer_from_ssa(struct vtn_builder
*b
, nir_ssa_def
*ssa
,
1830 struct vtn_type
*ptr_type
)
1832 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1834 struct vtn_type
*interface_type
= ptr_type
->deref
;
1835 while (interface_type
->base_type
== vtn_base_type_array
)
1836 interface_type
= interface_type
->array_element
;
1838 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
1839 nir_variable_mode nir_mode
;
1840 ptr
->mode
= vtn_storage_class_to_mode(b
, ptr_type
->storage_class
,
1841 interface_type
, &nir_mode
);
1842 ptr
->type
= ptr_type
->deref
;
1843 ptr
->ptr_type
= ptr_type
;
1845 if (b
->wa_glslang_179
) {
1846 /* To work around https://github.com/KhronosGroup/glslang/issues/179 we
1847 * need to whack the mode because it creates a function parameter with
1848 * the Function storage class even though it's a pointer to a sampler.
1849 * If we don't do this, then NIR won't get rid of the deref_cast for us.
1851 if (ptr
->mode
== vtn_variable_mode_function
&&
1852 (ptr
->type
->base_type
== vtn_base_type_sampler
||
1853 ptr
->type
->base_type
== vtn_base_type_sampled_image
)) {
1854 ptr
->mode
= vtn_variable_mode_uniform
;
1855 nir_mode
= nir_var_uniform
;
1859 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1860 /* This pointer type needs to have actual storage */
1861 vtn_assert(ptr_type
->type
);
1862 if (ptr
->mode
== vtn_variable_mode_ubo
||
1863 ptr
->mode
== vtn_variable_mode_ssbo
) {
1864 vtn_assert(ssa
->num_components
== 2);
1865 ptr
->block_index
= nir_channel(&b
->nb
, ssa
, 0);
1866 ptr
->offset
= nir_channel(&b
->nb
, ssa
, 1);
1868 vtn_assert(ssa
->num_components
== 1);
1869 ptr
->block_index
= NULL
;
1873 const struct glsl_type
*deref_type
= ptr_type
->deref
->type
;
1874 if (!vtn_pointer_is_external_block(b
, ptr
)) {
1875 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
1877 } else if (vtn_type_contains_block(b
, ptr
->type
) &&
1878 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
1879 /* This is a pointer to somewhere in an array of blocks, not a
1880 * pointer to somewhere inside the block. Set the block index
1881 * instead of making a cast.
1883 ptr
->block_index
= ssa
;
1885 /* This is a pointer to something internal or a pointer inside a
1886 * block. It's just a regular cast.
1888 * For PhysicalStorageBufferEXT pointers, we don't have a block index
1889 * at all because we get the pointer directly from the client. This
1890 * assumes that there will never be a SSBO binding variable using the
1891 * PhysicalStorageBufferEXT storage class. This assumption appears
1892 * to be correct according to the Vulkan spec because the table,
1893 * "Shader Resource and Storage Class Correspondence," the only the
1894 * Uniform storage class with BufferBlock or the StorageBuffer
1895 * storage class with Block can be used.
1897 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
1898 ptr_type
->deref
->type
,
1900 ptr
->deref
->dest
.ssa
.num_components
=
1901 glsl_get_vector_elements(ptr_type
->type
);
1902 ptr
->deref
->dest
.ssa
.bit_size
= glsl_get_bit_size(ptr_type
->type
);
1910 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
1912 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
1915 if (var
->mode
== vtn_variable_mode_input
) {
1916 return stage
== MESA_SHADER_TESS_CTRL
||
1917 stage
== MESA_SHADER_TESS_EVAL
||
1918 stage
== MESA_SHADER_GEOMETRY
;
1921 if (var
->mode
== vtn_variable_mode_output
)
1922 return stage
== MESA_SHADER_TESS_CTRL
;
1928 assign_missing_member_locations(struct vtn_variable
*var
)
1931 glsl_get_length(glsl_without_array(var
->type
->type
));
1932 int location
= var
->base_location
;
1934 for (unsigned i
= 0; i
< length
; i
++) {
1935 /* From the Vulkan spec:
1937 * “If the structure type is a Block but without a Location, then each
1938 * of its members must have a Location decoration.”
1941 if (var
->type
->block
) {
1942 assert(var
->base_location
!= -1 ||
1943 var
->var
->members
[i
].location
!= -1);
1946 /* From the Vulkan spec:
1948 * “Any member with its own Location decoration is assigned that
1949 * location. Each remaining member is assigned the location after the
1950 * immediately preceding member in declaration order.”
1952 if (var
->var
->members
[i
].location
!= -1)
1953 location
= var
->var
->members
[i
].location
;
1955 var
->var
->members
[i
].location
= location
;
1957 /* Below we use type instead of interface_type, because interface_type
1958 * is only available when it is a Block. This code also supports
1959 * input/outputs that are just structs
1961 const struct glsl_type
*member_type
=
1962 glsl_get_struct_field(glsl_without_array(var
->type
->type
), i
);
1965 glsl_count_attribute_slots(member_type
,
1966 false /* is_gl_vertex_input */);
1972 vtn_create_variable(struct vtn_builder
*b
, struct vtn_value
*val
,
1973 struct vtn_type
*ptr_type
, SpvStorageClass storage_class
,
1974 nir_constant
*initializer
)
1976 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1977 struct vtn_type
*type
= ptr_type
->deref
;
1979 struct vtn_type
*without_array
= type
;
1980 while(glsl_type_is_array(without_array
->type
))
1981 without_array
= without_array
->array_element
;
1983 enum vtn_variable_mode mode
;
1984 nir_variable_mode nir_mode
;
1985 mode
= vtn_storage_class_to_mode(b
, storage_class
, without_array
, &nir_mode
);
1988 case vtn_variable_mode_ubo
:
1989 /* There's no other way to get vtn_variable_mode_ubo */
1990 vtn_assert(without_array
->block
);
1991 b
->shader
->info
.num_ubos
++;
1993 case vtn_variable_mode_ssbo
:
1994 if (storage_class
== SpvStorageClassStorageBuffer
&&
1995 !without_array
->block
) {
1996 if (b
->variable_pointers
) {
1997 vtn_fail("Variables in the StorageBuffer storage class must "
1998 "have a struct type with the Block decoration");
2000 /* If variable pointers are not present, it's still malformed
2001 * SPIR-V but we can parse it and do the right thing anyway.
2002 * Since some of the 8-bit storage tests have bugs in this are,
2003 * just make it a warning for now.
2005 vtn_warn("Variables in the StorageBuffer storage class must "
2006 "have a struct type with the Block decoration");
2009 b
->shader
->info
.num_ssbos
++;
2011 case vtn_variable_mode_uniform
:
2012 if (glsl_type_is_image(without_array
->type
))
2013 b
->shader
->info
.num_images
++;
2014 else if (glsl_type_is_sampler(without_array
->type
))
2015 b
->shader
->info
.num_textures
++;
2017 case vtn_variable_mode_push_constant
:
2018 b
->shader
->num_uniforms
= vtn_type_block_size(b
, type
);
2021 case vtn_variable_mode_phys_ssbo
:
2022 vtn_fail("Cannot create a variable with the "
2023 "PhysicalStorageBufferEXT storage class");
2027 /* No tallying is needed */
2031 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
2034 var
->base_location
= -1;
2036 vtn_assert(val
->value_type
== vtn_value_type_pointer
);
2037 val
->pointer
= vtn_pointer_for_variable(b
, var
, ptr_type
);
2039 switch (var
->mode
) {
2040 case vtn_variable_mode_function
:
2041 case vtn_variable_mode_private
:
2042 case vtn_variable_mode_uniform
:
2043 /* For these, we create the variable normally */
2044 var
->var
= rzalloc(b
->shader
, nir_variable
);
2045 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2047 if (storage_class
== SpvStorageClassAtomicCounter
) {
2048 /* Need to tweak the nir type here as at vtn_handle_type we don't
2049 * have the access to storage_class, that is the one that points us
2050 * that is an atomic uint.
2052 var
->var
->type
= repair_atomic_type(var
->type
->type
);
2054 /* Private variables don't have any explicit layout but some layouts
2055 * may have leaked through due to type deduplication in the SPIR-V.
2057 var
->var
->type
= var
->type
->type
;
2059 var
->var
->data
.mode
= nir_mode
;
2060 var
->var
->data
.location
= -1;
2061 var
->var
->interface_type
= NULL
;
2064 case vtn_variable_mode_workgroup
:
2065 if (b
->options
->lower_workgroup_access_to_offsets
) {
2066 var
->shared_location
= -1;
2068 /* Create the variable normally */
2069 var
->var
= rzalloc(b
->shader
, nir_variable
);
2070 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2071 /* Workgroup variables don't have any explicit layout but some
2072 * layouts may have leaked through due to type deduplication in the
2075 var
->var
->type
= var
->type
->type
;
2076 var
->var
->data
.mode
= nir_var_mem_shared
;
2080 case vtn_variable_mode_input
:
2081 case vtn_variable_mode_output
: {
2082 /* In order to know whether or not we're a per-vertex inout, we need
2083 * the patch qualifier. This means walking the variable decorations
2084 * early before we actually create any variables. Not a big deal.
2086 * GLSLang really likes to place decorations in the most interior
2087 * thing it possibly can. In particular, if you have a struct, it
2088 * will place the patch decorations on the struct members. This
2089 * should be handled by the variable splitting below just fine.
2091 * If you have an array-of-struct, things get even more weird as it
2092 * will place the patch decorations on the struct even though it's
2093 * inside an array and some of the members being patch and others not
2094 * makes no sense whatsoever. Since the only sensible thing is for
2095 * it to be all or nothing, we'll call it patch if any of the members
2096 * are declared patch.
2099 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
2100 if (glsl_type_is_array(var
->type
->type
) &&
2101 glsl_type_is_struct_or_ifc(without_array
->type
)) {
2102 vtn_foreach_decoration(b
, vtn_value(b
, without_array
->id
,
2103 vtn_value_type_type
),
2104 var_is_patch_cb
, &var
->patch
);
2107 /* For inputs and outputs, we immediately split structures. This
2108 * is for a couple of reasons. For one, builtins may all come in
2109 * a struct and we really want those split out into separate
2110 * variables. For another, interpolation qualifiers can be
2111 * applied to members of the top-level struct ane we need to be
2112 * able to preserve that information.
2115 struct vtn_type
*per_vertex_type
= var
->type
;
2116 if (is_per_vertex_inout(var
, b
->shader
->info
.stage
)) {
2117 /* In Geometry shaders (and some tessellation), inputs come
2118 * in per-vertex arrays. However, some builtins come in
2119 * non-per-vertex, hence the need for the is_array check. In
2120 * any case, there are no non-builtin arrays allowed so this
2121 * check should be sufficient.
2123 per_vertex_type
= var
->type
->array_element
;
2126 var
->var
= rzalloc(b
->shader
, nir_variable
);
2127 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2128 /* In Vulkan, shader I/O variables don't have any explicit layout but
2129 * some layouts may have leaked through due to type deduplication in
2130 * the SPIR-V. We do, however, keep the layouts in the variable's
2131 * interface_type because we need offsets for XFB arrays of blocks.
2133 var
->var
->type
= var
->type
->type
;
2134 var
->var
->data
.mode
= nir_mode
;
2135 var
->var
->data
.patch
= var
->patch
;
2137 /* Figure out the interface block type. */
2138 struct vtn_type
*iface_type
= per_vertex_type
;
2139 if (var
->mode
== vtn_variable_mode_output
&&
2140 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
2141 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
||
2142 b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)) {
2143 /* For vertex data outputs, we can end up with arrays of blocks for
2144 * transform feedback where each array element corresponds to a
2145 * different XFB output buffer.
2147 while (iface_type
->base_type
== vtn_base_type_array
)
2148 iface_type
= iface_type
->array_element
;
2150 if (iface_type
->base_type
== vtn_base_type_struct
&& iface_type
->block
)
2151 var
->var
->interface_type
= iface_type
->type
;
2153 if (per_vertex_type
->base_type
== vtn_base_type_struct
&&
2154 per_vertex_type
->block
) {
2155 /* It's a struct. Set it up as per-member. */
2156 var
->var
->num_members
= glsl_get_length(per_vertex_type
->type
);
2157 var
->var
->members
= rzalloc_array(var
->var
, struct nir_variable_data
,
2158 var
->var
->num_members
);
2160 for (unsigned i
= 0; i
< var
->var
->num_members
; i
++) {
2161 var
->var
->members
[i
].mode
= nir_mode
;
2162 var
->var
->members
[i
].patch
= var
->patch
;
2163 var
->var
->members
[i
].location
= -1;
2167 /* For inputs and outputs, we need to grab locations and builtin
2168 * information from the per-vertex type.
2170 vtn_foreach_decoration(b
, vtn_value(b
, per_vertex_type
->id
,
2171 vtn_value_type_type
),
2172 var_decoration_cb
, var
);
2176 case vtn_variable_mode_ubo
:
2177 case vtn_variable_mode_ssbo
:
2178 case vtn_variable_mode_push_constant
:
2179 case vtn_variable_mode_cross_workgroup
:
2180 /* These don't need actual variables. */
2183 case vtn_variable_mode_phys_ssbo
:
2184 unreachable("Should have been caught before");
2188 var
->var
->constant_initializer
=
2189 nir_constant_clone(initializer
, var
->var
);
2192 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
2194 if ((var
->mode
== vtn_variable_mode_input
||
2195 var
->mode
== vtn_variable_mode_output
) &&
2196 var
->var
->members
) {
2197 assign_missing_member_locations(var
);
2200 if (var
->mode
== vtn_variable_mode_uniform
) {
2201 /* XXX: We still need the binding information in the nir_variable
2202 * for these. We should fix that.
2204 var
->var
->data
.binding
= var
->binding
;
2205 var
->var
->data
.explicit_binding
= var
->explicit_binding
;
2206 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
2207 var
->var
->data
.index
= var
->input_attachment_index
;
2208 var
->var
->data
.offset
= var
->offset
;
2210 if (glsl_type_is_image(without_array
->type
))
2211 var
->var
->data
.image
.format
= without_array
->image_format
;
2214 if (var
->mode
== vtn_variable_mode_function
) {
2215 vtn_assert(var
->var
!= NULL
&& var
->var
->members
== NULL
);
2216 nir_function_impl_add_variable(b
->nb
.impl
, var
->var
);
2217 } else if (var
->var
) {
2218 nir_shader_add_variable(b
->shader
, var
->var
);
2220 vtn_assert(vtn_pointer_is_external_block(b
, val
->pointer
));
2225 vtn_assert_types_equal(struct vtn_builder
*b
, SpvOp opcode
,
2226 struct vtn_type
*dst_type
,
2227 struct vtn_type
*src_type
)
2229 if (dst_type
->id
== src_type
->id
)
2232 if (vtn_types_compatible(b
, dst_type
, src_type
)) {
2233 /* Early versions of GLSLang would re-emit types unnecessarily and you
2234 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
2235 * mismatched source and destination types.
2237 * https://github.com/KhronosGroup/glslang/issues/304
2238 * https://github.com/KhronosGroup/glslang/issues/307
2239 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
2240 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2242 vtn_warn("Source and destination types of %s do not have the same "
2243 "ID (but are compatible): %u vs %u",
2244 spirv_op_to_string(opcode
), dst_type
->id
, src_type
->id
);
2248 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
2249 spirv_op_to_string(opcode
),
2250 glsl_get_type_name(dst_type
->type
),
2251 glsl_get_type_name(src_type
->type
));
2254 static nir_ssa_def
*
2255 nir_shrink_zero_pad_vec(nir_builder
*b
, nir_ssa_def
*val
,
2256 unsigned num_components
)
2258 if (val
->num_components
== num_components
)
2261 nir_ssa_def
*comps
[NIR_MAX_VEC_COMPONENTS
];
2262 for (unsigned i
= 0; i
< num_components
; i
++) {
2263 if (i
< val
->num_components
)
2264 comps
[i
] = nir_channel(b
, val
, i
);
2266 comps
[i
] = nir_imm_intN_t(b
, 0, val
->bit_size
);
2268 return nir_vec(b
, comps
, num_components
);
2271 static nir_ssa_def
*
2272 nir_sloppy_bitcast(nir_builder
*b
, nir_ssa_def
*val
,
2273 const struct glsl_type
*type
)
2275 const unsigned num_components
= glsl_get_vector_elements(type
);
2276 const unsigned bit_size
= glsl_get_bit_size(type
);
2278 /* First, zero-pad to ensure that the value is big enough that when we
2279 * bit-cast it, we don't loose anything.
2281 if (val
->bit_size
< bit_size
) {
2282 const unsigned src_num_components_needed
=
2283 vtn_align_u32(val
->num_components
, bit_size
/ val
->bit_size
);
2284 val
= nir_shrink_zero_pad_vec(b
, val
, src_num_components_needed
);
2287 val
= nir_bitcast_vector(b
, val
, bit_size
);
2289 return nir_shrink_zero_pad_vec(b
, val
, num_components
);
2293 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
2294 const uint32_t *w
, unsigned count
)
2298 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2299 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2303 case SpvOpVariable
: {
2304 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2306 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2308 SpvStorageClass storage_class
= w
[3];
2309 nir_constant
*initializer
= NULL
;
2311 initializer
= vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
2313 vtn_create_variable(b
, val
, ptr_type
, storage_class
, initializer
);
2317 case SpvOpAccessChain
:
2318 case SpvOpPtrAccessChain
:
2319 case SpvOpInBoundsAccessChain
:
2320 case SpvOpInBoundsPtrAccessChain
: {
2321 struct vtn_access_chain
*chain
= vtn_access_chain_create(b
, count
- 4);
2322 chain
->ptr_as_array
= (opcode
== SpvOpPtrAccessChain
|| opcode
== SpvOpInBoundsPtrAccessChain
);
2325 for (int i
= 4; i
< count
; i
++) {
2326 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
2327 if (link_val
->value_type
== vtn_value_type_constant
) {
2328 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
2329 switch (glsl_get_bit_size(link_val
->type
->type
)) {
2331 chain
->link
[idx
].id
= link_val
->constant
->values
[0].i8
[0];
2334 chain
->link
[idx
].id
= link_val
->constant
->values
[0].i16
[0];
2337 chain
->link
[idx
].id
= link_val
->constant
->values
[0].i32
[0];
2340 chain
->link
[idx
].id
= link_val
->constant
->values
[0].i64
[0];
2343 vtn_fail("Invalid bit size");
2346 chain
->link
[idx
].mode
= vtn_access_mode_id
;
2347 chain
->link
[idx
].id
= w
[i
];
2353 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2354 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
2355 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
2356 /* This is rather insane. SPIR-V allows you to use OpSampledImage
2357 * to combine an array of images with a single sampler to get an
2358 * array of sampled images that all share the same sampler.
2359 * Fortunately, this means that we can more-or-less ignore the
2360 * sampler when crawling the access chain, but it does leave us
2361 * with this rather awkward little special-case.
2363 struct vtn_value
*val
=
2364 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2365 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2366 val
->sampled_image
->type
= base_val
->sampled_image
->type
;
2367 val
->sampled_image
->image
=
2368 vtn_pointer_dereference(b
, base_val
->sampled_image
->image
, chain
);
2369 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
2371 vtn_assert(base_val
->value_type
== vtn_value_type_pointer
);
2372 struct vtn_value
*val
=
2373 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2374 val
->pointer
= vtn_pointer_dereference(b
, base_val
->pointer
, chain
);
2375 val
->pointer
->ptr_type
= ptr_type
;
2380 case SpvOpCopyMemory
: {
2381 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2382 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_pointer
);
2384 vtn_assert_types_equal(b
, opcode
, dest
->type
->deref
, src
->type
->deref
);
2386 vtn_variable_copy(b
, dest
->pointer
, src
->pointer
);
2391 struct vtn_type
*res_type
=
2392 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2393 struct vtn_value
*src_val
= vtn_value(b
, w
[3], vtn_value_type_pointer
);
2394 struct vtn_pointer
*src
= src_val
->pointer
;
2396 vtn_assert_types_equal(b
, opcode
, res_type
, src_val
->type
->deref
);
2398 if (glsl_type_is_image(res_type
->type
) ||
2399 glsl_type_is_sampler(res_type
->type
)) {
2400 vtn_push_value(b
, w
[2], vtn_value_type_pointer
)->pointer
= src
;
2404 vtn_push_ssa(b
, w
[2], res_type
, vtn_variable_load(b
, src
));
2409 struct vtn_value
*dest_val
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2410 struct vtn_pointer
*dest
= dest_val
->pointer
;
2411 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[2]);
2413 /* OpStore requires us to actually have a storage type */
2414 vtn_fail_if(dest
->type
->type
== NULL
,
2415 "Invalid destination type for OpStore");
2417 if (glsl_get_base_type(dest
->type
->type
) == GLSL_TYPE_BOOL
&&
2418 glsl_get_base_type(src_val
->type
->type
) == GLSL_TYPE_UINT
) {
2419 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2420 * would then store them to a local variable as bool. Work around
2421 * the issue by doing an implicit conversion.
2423 * https://github.com/KhronosGroup/glslang/issues/170
2424 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2426 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2427 "OpTypeBool. Doing an implicit conversion to work around "
2429 struct vtn_ssa_value
*bool_ssa
=
2430 vtn_create_ssa_value(b
, dest
->type
->type
);
2431 bool_ssa
->def
= nir_i2b(&b
->nb
, vtn_ssa_value(b
, w
[2])->def
);
2432 vtn_variable_store(b
, bool_ssa
, dest
);
2436 vtn_assert_types_equal(b
, opcode
, dest_val
->type
->deref
, src_val
->type
);
2438 if (glsl_type_is_sampler(dest
->type
->type
)) {
2439 if (b
->wa_glslang_179
) {
2440 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2441 "propagation to workaround the problem.");
2442 vtn_assert(dest
->var
->copy_prop_sampler
== NULL
);
2443 dest
->var
->copy_prop_sampler
=
2444 vtn_value(b
, w
[2], vtn_value_type_pointer
)->pointer
;
2446 vtn_fail("Vulkan does not allow OpStore of a sampler or image.");
2451 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
2452 vtn_variable_store(b
, src
, dest
);
2456 case SpvOpArrayLength
: {
2457 struct vtn_pointer
*ptr
=
2458 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2459 const uint32_t field
= w
[4];
2461 vtn_fail_if(ptr
->type
->base_type
!= vtn_base_type_struct
,
2462 "OpArrayLength must take a pointer to a structure type");
2463 vtn_fail_if(field
!= ptr
->type
->length
- 1 ||
2464 ptr
->type
->members
[field
]->base_type
!= vtn_base_type_array
,
2465 "OpArrayLength must reference the last memeber of the "
2466 "structure and that must be an array");
2468 const uint32_t offset
= ptr
->type
->offsets
[field
];
2469 const uint32_t stride
= ptr
->type
->members
[field
]->stride
;
2471 if (!ptr
->block_index
) {
2472 struct vtn_access_chain chain
= {
2475 ptr
= vtn_pointer_dereference(b
, ptr
, &chain
);
2476 vtn_assert(ptr
->block_index
);
2479 nir_intrinsic_instr
*instr
=
2480 nir_intrinsic_instr_create(b
->nb
.shader
,
2481 nir_intrinsic_get_buffer_size
);
2482 instr
->src
[0] = nir_src_for_ssa(ptr
->block_index
);
2483 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
2484 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2485 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
2487 /* array_length = max(buffer_size - offset, 0) / stride */
2488 nir_ssa_def
*array_length
=
2493 nir_imm_int(&b
->nb
, offset
)),
2494 nir_imm_int(&b
->nb
, 0u)),
2495 nir_imm_int(&b
->nb
, stride
));
2497 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2498 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
2499 val
->ssa
->def
= array_length
;
2503 case SpvOpConvertPtrToU
: {
2504 struct vtn_value
*u_val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2506 vtn_fail_if(u_val
->type
->base_type
!= vtn_base_type_vector
&&
2507 u_val
->type
->base_type
!= vtn_base_type_scalar
,
2508 "OpConvertPtrToU can only be used to cast to a vector or "
2511 /* The pointer will be converted to an SSA value automatically */
2512 nir_ssa_def
*ptr_ssa
= vtn_ssa_value(b
, w
[3])->def
;
2514 u_val
->ssa
= vtn_create_ssa_value(b
, u_val
->type
->type
);
2515 u_val
->ssa
->def
= nir_sloppy_bitcast(&b
->nb
, ptr_ssa
, u_val
->type
->type
);
2519 case SpvOpConvertUToPtr
: {
2520 struct vtn_value
*ptr_val
=
2521 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2522 struct vtn_value
*u_val
= vtn_value(b
, w
[3], vtn_value_type_ssa
);
2524 vtn_fail_if(ptr_val
->type
->type
== NULL
,
2525 "OpConvertUToPtr can only be used on physical pointers");
2527 vtn_fail_if(u_val
->type
->base_type
!= vtn_base_type_vector
&&
2528 u_val
->type
->base_type
!= vtn_base_type_scalar
,
2529 "OpConvertUToPtr can only be used to cast from a vector or "
2532 nir_ssa_def
*ptr_ssa
= nir_sloppy_bitcast(&b
->nb
, u_val
->ssa
->def
,
2533 ptr_val
->type
->type
);
2534 ptr_val
->pointer
= vtn_pointer_from_ssa(b
, ptr_ssa
, ptr_val
->type
);
2538 case SpvOpCopyMemorySized
:
2540 vtn_fail("Unhandled opcode");