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_INTERFACE
:
494 case GLSL_TYPE_STRUCT
: {
495 vtn_assert(deref_chain
->link
[idx
].mode
== vtn_access_mode_literal
);
496 unsigned member
= deref_chain
->link
[idx
].id
;
497 offset
= nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[member
]);
498 type
= type
->members
[member
];
499 access
|= type
->access
;
504 vtn_fail("Invalid type for deref");
508 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
509 ptr
->mode
= base
->mode
;
511 ptr
->block_index
= block_index
;
512 ptr
->offset
= offset
;
513 ptr
->access
= access
;
518 /* Dereference the given base pointer by the access chain */
519 static struct vtn_pointer
*
520 vtn_pointer_dereference(struct vtn_builder
*b
,
521 struct vtn_pointer
*base
,
522 struct vtn_access_chain
*deref_chain
)
524 if (vtn_pointer_uses_ssa_offset(b
, base
)) {
525 return vtn_ssa_offset_pointer_dereference(b
, base
, deref_chain
);
527 return vtn_nir_deref_pointer_dereference(b
, base
, deref_chain
);
532 vtn_pointer_for_variable(struct vtn_builder
*b
,
533 struct vtn_variable
*var
, struct vtn_type
*ptr_type
)
535 struct vtn_pointer
*pointer
= rzalloc(b
, struct vtn_pointer
);
537 pointer
->mode
= var
->mode
;
538 pointer
->type
= var
->type
;
539 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
540 vtn_assert(ptr_type
->deref
->type
== var
->type
->type
);
541 pointer
->ptr_type
= ptr_type
;
543 pointer
->access
= var
->access
| var
->type
->access
;
548 /* Returns an atomic_uint type based on the original uint type. The returned
549 * type will be equivalent to the original one but will have an atomic_uint
550 * type as leaf instead of an uint.
552 * Manages uint scalars, arrays, and arrays of arrays of any nested depth.
554 static const struct glsl_type
*
555 repair_atomic_type(const struct glsl_type
*type
)
557 assert(glsl_get_base_type(glsl_without_array(type
)) == GLSL_TYPE_UINT
);
558 assert(glsl_type_is_scalar(glsl_without_array(type
)));
560 if (glsl_type_is_array(type
)) {
561 const struct glsl_type
*atomic
=
562 repair_atomic_type(glsl_get_array_element(type
));
564 return glsl_array_type(atomic
, glsl_get_length(type
),
565 glsl_get_explicit_stride(type
));
567 return glsl_atomic_uint_type();
572 vtn_pointer_to_deref(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
574 if (b
->wa_glslang_179
) {
575 /* Do on-the-fly copy propagation for samplers. */
576 if (ptr
->var
&& ptr
->var
->copy_prop_sampler
)
577 return vtn_pointer_to_deref(b
, ptr
->var
->copy_prop_sampler
);
580 vtn_assert(!vtn_pointer_uses_ssa_offset(b
, ptr
));
582 struct vtn_access_chain chain
= {
585 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
592 _vtn_local_load_store(struct vtn_builder
*b
, bool load
, nir_deref_instr
*deref
,
593 struct vtn_ssa_value
*inout
)
595 if (glsl_type_is_vector_or_scalar(deref
->type
)) {
597 inout
->def
= nir_load_deref(&b
->nb
, deref
);
599 nir_store_deref(&b
->nb
, deref
, inout
->def
, ~0);
601 } else if (glsl_type_is_array(deref
->type
) ||
602 glsl_type_is_matrix(deref
->type
)) {
603 unsigned elems
= glsl_get_length(deref
->type
);
604 for (unsigned i
= 0; i
< elems
; i
++) {
605 nir_deref_instr
*child
=
606 nir_build_deref_array_imm(&b
->nb
, deref
, i
);
607 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
]);
610 vtn_assert(glsl_type_is_struct_or_ifc(deref
->type
));
611 unsigned elems
= glsl_get_length(deref
->type
);
612 for (unsigned i
= 0; i
< elems
; i
++) {
613 nir_deref_instr
*child
= nir_build_deref_struct(&b
->nb
, deref
, i
);
614 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
]);
620 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
622 struct vtn_pointer
*ptr
= vtn_value(b
, id
, vtn_value_type_pointer
)->pointer
;
623 return vtn_pointer_to_deref(b
, ptr
);
627 * Gets the NIR-level deref tail, which may have as a child an array deref
628 * selecting which component due to OpAccessChain supporting per-component
629 * indexing in SPIR-V.
631 static nir_deref_instr
*
632 get_deref_tail(nir_deref_instr
*deref
)
634 if (deref
->deref_type
!= nir_deref_type_array
)
637 nir_deref_instr
*parent
=
638 nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
);
640 if (glsl_type_is_vector(parent
->type
))
646 struct vtn_ssa_value
*
647 vtn_local_load(struct vtn_builder
*b
, nir_deref_instr
*src
)
649 nir_deref_instr
*src_tail
= get_deref_tail(src
);
650 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
651 _vtn_local_load_store(b
, true, src_tail
, val
);
653 if (src_tail
!= src
) {
654 val
->type
= src
->type
;
655 if (nir_src_is_const(src
->arr
.index
))
656 val
->def
= vtn_vector_extract(b
, val
->def
,
657 nir_src_as_uint(src
->arr
.index
));
659 val
->def
= vtn_vector_extract_dynamic(b
, val
->def
, src
->arr
.index
.ssa
);
666 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
667 nir_deref_instr
*dest
)
669 nir_deref_instr
*dest_tail
= get_deref_tail(dest
);
671 if (dest_tail
!= dest
) {
672 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
673 _vtn_local_load_store(b
, true, dest_tail
, val
);
675 if (nir_src_is_const(dest
->arr
.index
))
676 val
->def
= vtn_vector_insert(b
, val
->def
, src
->def
,
677 nir_src_as_uint(dest
->arr
.index
));
679 val
->def
= vtn_vector_insert_dynamic(b
, val
->def
, src
->def
,
680 dest
->arr
.index
.ssa
);
681 _vtn_local_load_store(b
, false, dest_tail
, val
);
683 _vtn_local_load_store(b
, false, dest_tail
, src
);
688 vtn_pointer_to_offset(struct vtn_builder
*b
, struct vtn_pointer
*ptr
,
689 nir_ssa_def
**index_out
)
691 assert(vtn_pointer_uses_ssa_offset(b
, ptr
));
693 struct vtn_access_chain chain
= {
696 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
698 *index_out
= ptr
->block_index
;
702 /* Tries to compute the size of an interface block based on the strides and
703 * offsets that are provided to us in the SPIR-V source.
706 vtn_type_block_size(struct vtn_builder
*b
, struct vtn_type
*type
)
708 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
712 case GLSL_TYPE_UINT16
:
713 case GLSL_TYPE_INT16
:
714 case GLSL_TYPE_UINT8
:
716 case GLSL_TYPE_UINT64
:
717 case GLSL_TYPE_INT64
:
718 case GLSL_TYPE_FLOAT
:
719 case GLSL_TYPE_FLOAT16
:
721 case GLSL_TYPE_DOUBLE
: {
722 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
723 glsl_get_matrix_columns(type
->type
);
725 vtn_assert(type
->stride
> 0);
726 return type
->stride
* cols
;
728 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
729 return glsl_get_vector_elements(type
->type
) * type_size
;
733 case GLSL_TYPE_STRUCT
:
734 case GLSL_TYPE_INTERFACE
: {
736 unsigned num_fields
= glsl_get_length(type
->type
);
737 for (unsigned f
= 0; f
< num_fields
; f
++) {
738 unsigned field_end
= type
->offsets
[f
] +
739 vtn_type_block_size(b
, type
->members
[f
]);
740 size
= MAX2(size
, field_end
);
745 case GLSL_TYPE_ARRAY
:
746 vtn_assert(type
->stride
> 0);
747 vtn_assert(glsl_get_length(type
->type
) > 0);
748 return type
->stride
* glsl_get_length(type
->type
);
751 vtn_fail("Invalid block type");
757 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
758 nir_ssa_def
*index
, nir_ssa_def
*offset
,
759 unsigned access_offset
, unsigned access_size
,
760 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
,
761 enum gl_access_qualifier access
)
763 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
764 instr
->num_components
= glsl_get_vector_elements(type
);
766 /* Booleans usually shouldn't show up in external memory in SPIR-V.
767 * However, they do for certain older GLSLang versions and can for shared
768 * memory when we lower access chains internally.
770 const unsigned data_bit_size
= glsl_type_is_boolean(type
) ? 32 :
771 glsl_get_bit_size(type
);
775 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
776 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
779 if (op
== nir_intrinsic_load_push_constant
) {
780 nir_intrinsic_set_base(instr
, access_offset
);
781 nir_intrinsic_set_range(instr
, access_size
);
784 if (op
== nir_intrinsic_load_ssbo
||
785 op
== nir_intrinsic_store_ssbo
) {
786 nir_intrinsic_set_access(instr
, access
);
789 /* With extensions like relaxed_block_layout, we really can't guarantee
790 * much more than scalar alignment.
792 if (op
!= nir_intrinsic_load_push_constant
)
793 nir_intrinsic_set_align(instr
, data_bit_size
/ 8, 0);
796 instr
->src
[src
++] = nir_src_for_ssa(index
);
798 if (op
== nir_intrinsic_load_push_constant
) {
799 /* We need to subtract the offset from where the intrinsic will load the
802 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
803 nir_imm_int(&b
->nb
, access_offset
)));
805 instr
->src
[src
++] = nir_src_for_ssa(offset
);
809 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
810 instr
->num_components
, data_bit_size
, NULL
);
811 (*inout
)->def
= &instr
->dest
.ssa
;
814 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
816 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
817 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
821 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
822 nir_ssa_def
*index
, nir_ssa_def
*offset
,
823 unsigned access_offset
, unsigned access_size
,
824 struct vtn_type
*type
, enum gl_access_qualifier access
,
825 struct vtn_ssa_value
**inout
)
827 if (load
&& *inout
== NULL
)
828 *inout
= vtn_create_ssa_value(b
, type
->type
);
830 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
834 case GLSL_TYPE_UINT16
:
835 case GLSL_TYPE_INT16
:
836 case GLSL_TYPE_UINT8
:
838 case GLSL_TYPE_UINT64
:
839 case GLSL_TYPE_INT64
:
840 case GLSL_TYPE_FLOAT
:
841 case GLSL_TYPE_FLOAT16
:
842 case GLSL_TYPE_DOUBLE
:
844 /* This is where things get interesting. At this point, we've hit
845 * a vector, a scalar, or a matrix.
847 if (glsl_type_is_matrix(type
->type
)) {
848 /* Loading the whole matrix */
849 struct vtn_ssa_value
*transpose
;
850 unsigned num_ops
, vec_width
, col_stride
;
851 if (type
->row_major
) {
852 num_ops
= glsl_get_vector_elements(type
->type
);
853 vec_width
= glsl_get_matrix_columns(type
->type
);
854 col_stride
= type
->array_element
->stride
;
856 const struct glsl_type
*transpose_type
=
857 glsl_matrix_type(base_type
, vec_width
, num_ops
);
858 *inout
= vtn_create_ssa_value(b
, transpose_type
);
860 transpose
= vtn_ssa_transpose(b
, *inout
);
864 num_ops
= glsl_get_matrix_columns(type
->type
);
865 vec_width
= glsl_get_vector_elements(type
->type
);
866 col_stride
= type
->stride
;
869 for (unsigned i
= 0; i
< num_ops
; i
++) {
870 nir_ssa_def
*elem_offset
=
871 nir_iadd_imm(&b
->nb
, offset
, i
* col_stride
);
872 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
873 access_offset
, access_size
,
875 glsl_vector_type(base_type
, vec_width
),
876 type
->access
| access
);
879 if (load
&& type
->row_major
)
880 *inout
= vtn_ssa_transpose(b
, *inout
);
882 unsigned elems
= glsl_get_vector_elements(type
->type
);
883 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
884 if (elems
== 1 || type
->stride
== type_size
) {
885 /* This is a tightly-packed normal scalar or vector load */
886 vtn_assert(glsl_type_is_vector_or_scalar(type
->type
));
887 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
888 access_offset
, access_size
,
890 type
->access
| access
);
892 /* This is a strided load. We have to load N things separately.
893 * This is the single column of a row-major matrix case.
895 vtn_assert(type
->stride
> type_size
);
896 vtn_assert(type
->stride
% type_size
== 0);
898 nir_ssa_def
*per_comp
[4];
899 for (unsigned i
= 0; i
< elems
; i
++) {
900 nir_ssa_def
*elem_offset
=
901 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
902 struct vtn_ssa_value
*comp
, temp_val
;
904 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
905 temp_val
.type
= glsl_scalar_type(base_type
);
908 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
909 access_offset
, access_size
,
910 &comp
, glsl_scalar_type(base_type
),
911 type
->access
| access
);
912 per_comp
[i
] = comp
->def
;
917 *inout
= vtn_create_ssa_value(b
, type
->type
);
918 (*inout
)->def
= nir_vec(&b
->nb
, per_comp
, elems
);
924 case GLSL_TYPE_ARRAY
: {
925 unsigned elems
= glsl_get_length(type
->type
);
926 for (unsigned i
= 0; i
< elems
; i
++) {
927 nir_ssa_def
*elem_off
=
928 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
929 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
930 access_offset
, access_size
,
932 type
->array_element
->access
| access
,
933 &(*inout
)->elems
[i
]);
938 case GLSL_TYPE_INTERFACE
:
939 case GLSL_TYPE_STRUCT
: {
940 unsigned elems
= glsl_get_length(type
->type
);
941 for (unsigned i
= 0; i
< elems
; i
++) {
942 nir_ssa_def
*elem_off
=
943 nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[i
]);
944 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
945 access_offset
, access_size
,
947 type
->members
[i
]->access
| access
,
948 &(*inout
)->elems
[i
]);
954 vtn_fail("Invalid block member type");
958 static struct vtn_ssa_value
*
959 vtn_block_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
962 unsigned access_offset
= 0, access_size
= 0;
964 case vtn_variable_mode_ubo
:
965 op
= nir_intrinsic_load_ubo
;
967 case vtn_variable_mode_ssbo
:
968 op
= nir_intrinsic_load_ssbo
;
970 case vtn_variable_mode_push_constant
:
971 op
= nir_intrinsic_load_push_constant
;
972 access_size
= b
->shader
->num_uniforms
;
974 case vtn_variable_mode_workgroup
:
975 op
= nir_intrinsic_load_shared
;
978 vtn_fail("Invalid block variable mode");
981 nir_ssa_def
*offset
, *index
= NULL
;
982 offset
= vtn_pointer_to_offset(b
, src
, &index
);
984 struct vtn_ssa_value
*value
= NULL
;
985 _vtn_block_load_store(b
, op
, true, index
, offset
,
986 access_offset
, access_size
,
987 src
->type
, src
->access
, &value
);
992 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
993 struct vtn_pointer
*dst
)
997 case vtn_variable_mode_ssbo
:
998 op
= nir_intrinsic_store_ssbo
;
1000 case vtn_variable_mode_workgroup
:
1001 op
= nir_intrinsic_store_shared
;
1004 vtn_fail("Invalid block variable mode");
1007 nir_ssa_def
*offset
, *index
= NULL
;
1008 offset
= vtn_pointer_to_offset(b
, dst
, &index
);
1010 _vtn_block_load_store(b
, op
, false, index
, offset
,
1011 0, 0, dst
->type
, dst
->access
, &src
);
1015 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
1016 struct vtn_pointer
*ptr
,
1017 struct vtn_ssa_value
**inout
)
1019 enum glsl_base_type base_type
= glsl_get_base_type(ptr
->type
->type
);
1020 switch (base_type
) {
1021 case GLSL_TYPE_UINT
:
1023 case GLSL_TYPE_UINT16
:
1024 case GLSL_TYPE_INT16
:
1025 case GLSL_TYPE_UINT8
:
1026 case GLSL_TYPE_INT8
:
1027 case GLSL_TYPE_UINT64
:
1028 case GLSL_TYPE_INT64
:
1029 case GLSL_TYPE_FLOAT
:
1030 case GLSL_TYPE_FLOAT16
:
1031 case GLSL_TYPE_BOOL
:
1032 case GLSL_TYPE_DOUBLE
:
1033 if (glsl_type_is_vector_or_scalar(ptr
->type
->type
)) {
1034 /* We hit a vector or scalar; go ahead and emit the load[s] */
1035 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
1036 if (vtn_pointer_is_external_block(b
, ptr
)) {
1037 /* If it's external, we call nir_load/store_deref directly. The
1038 * vtn_local_load/store helpers are too clever and do magic to
1039 * avoid array derefs of vectors. That magic is both less
1040 * efficient than the direct load/store and, in the case of
1041 * stores, is broken because it creates a race condition if two
1042 * threads are writing to different components of the same vector
1043 * due to the load+insert+store it uses to emulate the array
1047 *inout
= vtn_create_ssa_value(b
, ptr
->type
->type
);
1048 (*inout
)->def
= nir_load_deref(&b
->nb
, deref
);
1050 nir_store_deref(&b
->nb
, deref
, (*inout
)->def
, ~0);
1054 *inout
= vtn_local_load(b
, deref
);
1056 vtn_local_store(b
, *inout
, deref
);
1063 case GLSL_TYPE_INTERFACE
:
1064 case GLSL_TYPE_ARRAY
:
1065 case GLSL_TYPE_STRUCT
: {
1066 unsigned elems
= glsl_get_length(ptr
->type
->type
);
1068 vtn_assert(*inout
== NULL
);
1069 *inout
= rzalloc(b
, struct vtn_ssa_value
);
1070 (*inout
)->type
= ptr
->type
->type
;
1071 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
1074 struct vtn_access_chain chain
= {
1077 { .mode
= vtn_access_mode_literal
, },
1080 for (unsigned i
= 0; i
< elems
; i
++) {
1081 chain
.link
[0].id
= i
;
1082 struct vtn_pointer
*elem
= vtn_pointer_dereference(b
, ptr
, &chain
);
1083 _vtn_variable_load_store(b
, load
, elem
, &(*inout
)->elems
[i
]);
1089 vtn_fail("Invalid access chain type");
1093 struct vtn_ssa_value
*
1094 vtn_variable_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
1096 if (vtn_pointer_uses_ssa_offset(b
, src
)) {
1097 return vtn_block_load(b
, src
);
1099 struct vtn_ssa_value
*val
= NULL
;
1100 _vtn_variable_load_store(b
, true, src
, &val
);
1106 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1107 struct vtn_pointer
*dest
)
1109 if (vtn_pointer_uses_ssa_offset(b
, dest
)) {
1110 vtn_assert(dest
->mode
== vtn_variable_mode_ssbo
||
1111 dest
->mode
== vtn_variable_mode_workgroup
);
1112 vtn_block_store(b
, src
, dest
);
1114 _vtn_variable_load_store(b
, false, dest
, &src
);
1119 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1120 struct vtn_pointer
*src
)
1122 vtn_assert(src
->type
->type
== dest
->type
->type
);
1123 enum glsl_base_type base_type
= glsl_get_base_type(src
->type
->type
);
1124 switch (base_type
) {
1125 case GLSL_TYPE_UINT
:
1127 case GLSL_TYPE_UINT16
:
1128 case GLSL_TYPE_INT16
:
1129 case GLSL_TYPE_UINT8
:
1130 case GLSL_TYPE_INT8
:
1131 case GLSL_TYPE_UINT64
:
1132 case GLSL_TYPE_INT64
:
1133 case GLSL_TYPE_FLOAT
:
1134 case GLSL_TYPE_FLOAT16
:
1135 case GLSL_TYPE_DOUBLE
:
1136 case GLSL_TYPE_BOOL
:
1137 /* At this point, we have a scalar, vector, or matrix so we know that
1138 * there cannot be any structure splitting still in the way. By
1139 * stopping at the matrix level rather than the vector level, we
1140 * ensure that matrices get loaded in the optimal way even if they
1141 * are storred row-major in a UBO.
1143 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
1146 case GLSL_TYPE_INTERFACE
:
1147 case GLSL_TYPE_ARRAY
:
1148 case GLSL_TYPE_STRUCT
: {
1149 struct vtn_access_chain chain
= {
1152 { .mode
= vtn_access_mode_literal
, },
1155 unsigned elems
= glsl_get_length(src
->type
->type
);
1156 for (unsigned i
= 0; i
< elems
; i
++) {
1157 chain
.link
[0].id
= i
;
1158 struct vtn_pointer
*src_elem
=
1159 vtn_pointer_dereference(b
, src
, &chain
);
1160 struct vtn_pointer
*dest_elem
=
1161 vtn_pointer_dereference(b
, dest
, &chain
);
1163 _vtn_variable_copy(b
, dest_elem
, src_elem
);
1169 vtn_fail("Invalid access chain type");
1174 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1175 struct vtn_pointer
*src
)
1177 /* TODO: At some point, we should add a special-case for when we can
1178 * just emit a copy_var intrinsic.
1180 _vtn_variable_copy(b
, dest
, src
);
1184 set_mode_system_value(struct vtn_builder
*b
, nir_variable_mode
*mode
)
1186 vtn_assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
1187 *mode
= nir_var_system_value
;
1191 vtn_get_builtin_location(struct vtn_builder
*b
,
1192 SpvBuiltIn builtin
, int *location
,
1193 nir_variable_mode
*mode
)
1196 case SpvBuiltInPosition
:
1197 *location
= VARYING_SLOT_POS
;
1199 case SpvBuiltInPointSize
:
1200 *location
= VARYING_SLOT_PSIZ
;
1202 case SpvBuiltInClipDistance
:
1203 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
1205 case SpvBuiltInCullDistance
:
1206 *location
= VARYING_SLOT_CULL_DIST0
;
1208 case SpvBuiltInVertexId
:
1209 case SpvBuiltInVertexIndex
:
1210 /* The Vulkan spec defines VertexIndex to be non-zero-based and doesn't
1211 * allow VertexId. The ARB_gl_spirv spec defines VertexId to be the
1212 * same as gl_VertexID, which is non-zero-based, and removes
1213 * VertexIndex. Since they're both defined to be non-zero-based, we use
1214 * SYSTEM_VALUE_VERTEX_ID for both.
1216 *location
= SYSTEM_VALUE_VERTEX_ID
;
1217 set_mode_system_value(b
, mode
);
1219 case SpvBuiltInInstanceIndex
:
1220 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
1221 set_mode_system_value(b
, mode
);
1223 case SpvBuiltInInstanceId
:
1224 *location
= SYSTEM_VALUE_INSTANCE_ID
;
1225 set_mode_system_value(b
, mode
);
1227 case SpvBuiltInPrimitiveId
:
1228 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1229 vtn_assert(*mode
== nir_var_shader_in
);
1230 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1231 } else if (*mode
== nir_var_shader_out
) {
1232 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1234 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
1235 set_mode_system_value(b
, mode
);
1238 case SpvBuiltInInvocationId
:
1239 *location
= SYSTEM_VALUE_INVOCATION_ID
;
1240 set_mode_system_value(b
, mode
);
1242 case SpvBuiltInLayer
:
1243 *location
= VARYING_SLOT_LAYER
;
1244 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1245 *mode
= nir_var_shader_in
;
1246 else if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1247 *mode
= nir_var_shader_out
;
1248 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1249 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1250 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1251 *mode
= nir_var_shader_out
;
1253 vtn_fail("invalid stage for SpvBuiltInLayer");
1255 case SpvBuiltInViewportIndex
:
1256 *location
= VARYING_SLOT_VIEWPORT
;
1257 if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1258 *mode
= nir_var_shader_out
;
1259 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1260 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1261 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1262 *mode
= nir_var_shader_out
;
1263 else if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1264 *mode
= nir_var_shader_in
;
1266 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1268 case SpvBuiltInTessLevelOuter
:
1269 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
1271 case SpvBuiltInTessLevelInner
:
1272 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
1274 case SpvBuiltInTessCoord
:
1275 *location
= SYSTEM_VALUE_TESS_COORD
;
1276 set_mode_system_value(b
, mode
);
1278 case SpvBuiltInPatchVertices
:
1279 *location
= SYSTEM_VALUE_VERTICES_IN
;
1280 set_mode_system_value(b
, mode
);
1282 case SpvBuiltInFragCoord
:
1283 *location
= VARYING_SLOT_POS
;
1284 vtn_assert(*mode
== nir_var_shader_in
);
1286 case SpvBuiltInPointCoord
:
1287 *location
= VARYING_SLOT_PNTC
;
1288 vtn_assert(*mode
== nir_var_shader_in
);
1290 case SpvBuiltInFrontFacing
:
1291 *location
= SYSTEM_VALUE_FRONT_FACE
;
1292 set_mode_system_value(b
, mode
);
1294 case SpvBuiltInSampleId
:
1295 *location
= SYSTEM_VALUE_SAMPLE_ID
;
1296 set_mode_system_value(b
, mode
);
1298 case SpvBuiltInSamplePosition
:
1299 *location
= SYSTEM_VALUE_SAMPLE_POS
;
1300 set_mode_system_value(b
, mode
);
1302 case SpvBuiltInSampleMask
:
1303 if (*mode
== nir_var_shader_out
) {
1304 *location
= FRAG_RESULT_SAMPLE_MASK
;
1306 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
1307 set_mode_system_value(b
, mode
);
1310 case SpvBuiltInFragDepth
:
1311 *location
= FRAG_RESULT_DEPTH
;
1312 vtn_assert(*mode
== nir_var_shader_out
);
1314 case SpvBuiltInHelperInvocation
:
1315 *location
= SYSTEM_VALUE_HELPER_INVOCATION
;
1316 set_mode_system_value(b
, mode
);
1318 case SpvBuiltInNumWorkgroups
:
1319 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
1320 set_mode_system_value(b
, mode
);
1322 case SpvBuiltInWorkgroupSize
:
1323 *location
= SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
1324 set_mode_system_value(b
, mode
);
1326 case SpvBuiltInWorkgroupId
:
1327 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
1328 set_mode_system_value(b
, mode
);
1330 case SpvBuiltInLocalInvocationId
:
1331 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
1332 set_mode_system_value(b
, mode
);
1334 case SpvBuiltInLocalInvocationIndex
:
1335 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
1336 set_mode_system_value(b
, mode
);
1338 case SpvBuiltInGlobalInvocationId
:
1339 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
1340 set_mode_system_value(b
, mode
);
1342 case SpvBuiltInGlobalLinearId
:
1343 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX
;
1344 set_mode_system_value(b
, mode
);
1346 case SpvBuiltInBaseVertex
:
1347 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1348 * semantic as SPIR-V BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1350 *location
= SYSTEM_VALUE_FIRST_VERTEX
;
1351 set_mode_system_value(b
, mode
);
1353 case SpvBuiltInBaseInstance
:
1354 *location
= SYSTEM_VALUE_BASE_INSTANCE
;
1355 set_mode_system_value(b
, mode
);
1357 case SpvBuiltInDrawIndex
:
1358 *location
= SYSTEM_VALUE_DRAW_ID
;
1359 set_mode_system_value(b
, mode
);
1361 case SpvBuiltInSubgroupSize
:
1362 *location
= SYSTEM_VALUE_SUBGROUP_SIZE
;
1363 set_mode_system_value(b
, mode
);
1365 case SpvBuiltInSubgroupId
:
1366 *location
= SYSTEM_VALUE_SUBGROUP_ID
;
1367 set_mode_system_value(b
, mode
);
1369 case SpvBuiltInSubgroupLocalInvocationId
:
1370 *location
= SYSTEM_VALUE_SUBGROUP_INVOCATION
;
1371 set_mode_system_value(b
, mode
);
1373 case SpvBuiltInNumSubgroups
:
1374 *location
= SYSTEM_VALUE_NUM_SUBGROUPS
;
1375 set_mode_system_value(b
, mode
);
1377 case SpvBuiltInDeviceIndex
:
1378 *location
= SYSTEM_VALUE_DEVICE_INDEX
;
1379 set_mode_system_value(b
, mode
);
1381 case SpvBuiltInViewIndex
:
1382 *location
= SYSTEM_VALUE_VIEW_INDEX
;
1383 set_mode_system_value(b
, mode
);
1385 case SpvBuiltInSubgroupEqMask
:
1386 *location
= SYSTEM_VALUE_SUBGROUP_EQ_MASK
,
1387 set_mode_system_value(b
, mode
);
1389 case SpvBuiltInSubgroupGeMask
:
1390 *location
= SYSTEM_VALUE_SUBGROUP_GE_MASK
,
1391 set_mode_system_value(b
, mode
);
1393 case SpvBuiltInSubgroupGtMask
:
1394 *location
= SYSTEM_VALUE_SUBGROUP_GT_MASK
,
1395 set_mode_system_value(b
, mode
);
1397 case SpvBuiltInSubgroupLeMask
:
1398 *location
= SYSTEM_VALUE_SUBGROUP_LE_MASK
,
1399 set_mode_system_value(b
, mode
);
1401 case SpvBuiltInSubgroupLtMask
:
1402 *location
= SYSTEM_VALUE_SUBGROUP_LT_MASK
,
1403 set_mode_system_value(b
, mode
);
1405 case SpvBuiltInFragStencilRefEXT
:
1406 *location
= FRAG_RESULT_STENCIL
;
1407 vtn_assert(*mode
== nir_var_shader_out
);
1409 case SpvBuiltInWorkDim
:
1410 *location
= SYSTEM_VALUE_WORK_DIM
;
1411 set_mode_system_value(b
, mode
);
1413 case SpvBuiltInGlobalSize
:
1414 *location
= SYSTEM_VALUE_GLOBAL_GROUP_SIZE
;
1415 set_mode_system_value(b
, mode
);
1418 vtn_fail("unsupported builtin: %u", builtin
);
1423 apply_var_decoration(struct vtn_builder
*b
,
1424 struct nir_variable_data
*var_data
,
1425 const struct vtn_decoration
*dec
)
1427 switch (dec
->decoration
) {
1428 case SpvDecorationRelaxedPrecision
:
1429 break; /* FIXME: Do nothing with this for now. */
1430 case SpvDecorationNoPerspective
:
1431 var_data
->interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1433 case SpvDecorationFlat
:
1434 var_data
->interpolation
= INTERP_MODE_FLAT
;
1436 case SpvDecorationCentroid
:
1437 var_data
->centroid
= true;
1439 case SpvDecorationSample
:
1440 var_data
->sample
= true;
1442 case SpvDecorationInvariant
:
1443 var_data
->invariant
= true;
1445 case SpvDecorationConstant
:
1446 var_data
->read_only
= true;
1448 case SpvDecorationNonReadable
:
1449 var_data
->image
.access
|= ACCESS_NON_READABLE
;
1451 case SpvDecorationNonWritable
:
1452 var_data
->read_only
= true;
1453 var_data
->image
.access
|= ACCESS_NON_WRITEABLE
;
1455 case SpvDecorationRestrict
:
1456 var_data
->image
.access
|= ACCESS_RESTRICT
;
1458 case SpvDecorationVolatile
:
1459 var_data
->image
.access
|= ACCESS_VOLATILE
;
1461 case SpvDecorationCoherent
:
1462 var_data
->image
.access
|= ACCESS_COHERENT
;
1464 case SpvDecorationComponent
:
1465 var_data
->location_frac
= dec
->literals
[0];
1467 case SpvDecorationIndex
:
1468 var_data
->index
= dec
->literals
[0];
1470 case SpvDecorationBuiltIn
: {
1471 SpvBuiltIn builtin
= dec
->literals
[0];
1473 nir_variable_mode mode
= var_data
->mode
;
1474 vtn_get_builtin_location(b
, builtin
, &var_data
->location
, &mode
);
1475 var_data
->mode
= mode
;
1478 case SpvBuiltInTessLevelOuter
:
1479 case SpvBuiltInTessLevelInner
:
1480 case SpvBuiltInClipDistance
:
1481 case SpvBuiltInCullDistance
:
1482 var_data
->compact
= true;
1489 case SpvDecorationSpecId
:
1490 case SpvDecorationRowMajor
:
1491 case SpvDecorationColMajor
:
1492 case SpvDecorationMatrixStride
:
1493 case SpvDecorationAliased
:
1494 case SpvDecorationUniform
:
1495 case SpvDecorationLinkageAttributes
:
1496 break; /* Do nothing with these here */
1498 case SpvDecorationPatch
:
1499 var_data
->patch
= true;
1502 case SpvDecorationLocation
:
1503 vtn_fail("Handled above");
1505 case SpvDecorationBlock
:
1506 case SpvDecorationBufferBlock
:
1507 case SpvDecorationArrayStride
:
1508 case SpvDecorationGLSLShared
:
1509 case SpvDecorationGLSLPacked
:
1510 break; /* These can apply to a type but we don't care about them */
1512 case SpvDecorationBinding
:
1513 case SpvDecorationDescriptorSet
:
1514 case SpvDecorationNoContraction
:
1515 case SpvDecorationInputAttachmentIndex
:
1516 vtn_warn("Decoration not allowed for variable or structure member: %s",
1517 spirv_decoration_to_string(dec
->decoration
));
1520 case SpvDecorationXfbBuffer
:
1521 var_data
->explicit_xfb_buffer
= true;
1522 var_data
->xfb_buffer
= dec
->literals
[0];
1523 var_data
->always_active_io
= true;
1525 case SpvDecorationXfbStride
:
1526 var_data
->explicit_xfb_stride
= true;
1527 var_data
->xfb_stride
= dec
->literals
[0];
1529 case SpvDecorationOffset
:
1530 var_data
->explicit_offset
= true;
1531 var_data
->offset
= dec
->literals
[0];
1534 case SpvDecorationStream
:
1535 var_data
->stream
= dec
->literals
[0];
1538 case SpvDecorationCPacked
:
1539 case SpvDecorationSaturatedConversion
:
1540 case SpvDecorationFuncParamAttr
:
1541 case SpvDecorationFPRoundingMode
:
1542 case SpvDecorationFPFastMathMode
:
1543 case SpvDecorationAlignment
:
1544 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
1545 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1546 spirv_decoration_to_string(dec
->decoration
));
1550 case SpvDecorationHlslSemanticGOOGLE
:
1551 /* HLSL semantic decorations can safely be ignored by the driver. */
1554 case SpvDecorationRestrictPointerEXT
:
1555 case SpvDecorationAliasedPointerEXT
:
1556 /* TODO: We should actually plumb alias information through NIR. */
1560 vtn_fail("Unhandled decoration");
1565 var_is_patch_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1566 const struct vtn_decoration
*dec
, void *out_is_patch
)
1568 if (dec
->decoration
== SpvDecorationPatch
) {
1569 *((bool *) out_is_patch
) = true;
1574 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1575 const struct vtn_decoration
*dec
, void *void_var
)
1577 struct vtn_variable
*vtn_var
= void_var
;
1579 /* Handle decorations that apply to a vtn_variable as a whole */
1580 switch (dec
->decoration
) {
1581 case SpvDecorationBinding
:
1582 vtn_var
->binding
= dec
->literals
[0];
1583 vtn_var
->explicit_binding
= true;
1585 case SpvDecorationDescriptorSet
:
1586 vtn_var
->descriptor_set
= dec
->literals
[0];
1588 case SpvDecorationInputAttachmentIndex
:
1589 vtn_var
->input_attachment_index
= dec
->literals
[0];
1591 case SpvDecorationPatch
:
1592 vtn_var
->patch
= true;
1594 case SpvDecorationOffset
:
1595 vtn_var
->offset
= dec
->literals
[0];
1597 case SpvDecorationNonWritable
:
1598 vtn_var
->access
|= ACCESS_NON_WRITEABLE
;
1600 case SpvDecorationNonReadable
:
1601 vtn_var
->access
|= ACCESS_NON_READABLE
;
1603 case SpvDecorationVolatile
:
1604 vtn_var
->access
|= ACCESS_VOLATILE
;
1606 case SpvDecorationCoherent
:
1607 vtn_var
->access
|= ACCESS_COHERENT
;
1609 case SpvDecorationHlslCounterBufferGOOGLE
:
1610 /* HLSL semantic decorations can safely be ignored by the driver. */
1616 if (val
->value_type
== vtn_value_type_pointer
) {
1617 assert(val
->pointer
->var
== void_var
);
1618 assert(member
== -1);
1620 assert(val
->value_type
== vtn_value_type_type
);
1623 /* Location is odd. If applied to a split structure, we have to walk the
1624 * whole thing and accumulate the location. It's easier to handle as a
1627 if (dec
->decoration
== SpvDecorationLocation
) {
1628 unsigned location
= dec
->literals
[0];
1629 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
1630 vtn_var
->mode
== vtn_variable_mode_output
) {
1631 location
+= FRAG_RESULT_DATA0
;
1632 } else if (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
&&
1633 vtn_var
->mode
== vtn_variable_mode_input
) {
1634 location
+= VERT_ATTRIB_GENERIC0
;
1635 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1636 vtn_var
->mode
== vtn_variable_mode_output
) {
1637 location
+= vtn_var
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
;
1638 } else if (vtn_var
->mode
!= vtn_variable_mode_uniform
) {
1639 vtn_warn("Location must be on input, output, uniform, sampler or "
1644 if (vtn_var
->var
->num_members
== 0) {
1645 /* This handles the member and lone variable cases */
1646 vtn_var
->var
->data
.location
= location
;
1648 /* This handles the structure member case */
1649 assert(vtn_var
->var
->members
);
1652 vtn_var
->base_location
= location
;
1654 vtn_var
->var
->members
[member
].location
= location
;
1660 if (vtn_var
->var
->num_members
== 0) {
1661 /* We call this function on types as well as variables and not all
1662 * struct types get split so we can end up having stray member
1663 * decorations; just ignore them.
1666 apply_var_decoration(b
, &vtn_var
->var
->data
, dec
);
1667 } else if (member
>= 0) {
1668 /* Member decorations must come from a type */
1669 assert(val
->value_type
== vtn_value_type_type
);
1670 apply_var_decoration(b
, &vtn_var
->var
->members
[member
], dec
);
1673 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1674 for (unsigned i
= 0; i
< length
; i
++)
1675 apply_var_decoration(b
, &vtn_var
->var
->members
[i
], dec
);
1678 /* A few variables, those with external storage, have no actual
1679 * nir_variables associated with them. Fortunately, all decorations
1680 * we care about for those variables are on the type only.
1682 vtn_assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1683 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1684 vtn_var
->mode
== vtn_variable_mode_push_constant
||
1685 (vtn_var
->mode
== vtn_variable_mode_workgroup
&&
1686 b
->options
->lower_workgroup_access_to_offsets
));
1691 static enum vtn_variable_mode
1692 vtn_storage_class_to_mode(struct vtn_builder
*b
,
1693 SpvStorageClass
class,
1694 struct vtn_type
*interface_type
,
1695 nir_variable_mode
*nir_mode_out
)
1697 enum vtn_variable_mode mode
;
1698 nir_variable_mode nir_mode
;
1700 case SpvStorageClassUniform
:
1701 if (interface_type
->block
) {
1702 mode
= vtn_variable_mode_ubo
;
1703 nir_mode
= nir_var_mem_ubo
;
1704 } else if (interface_type
->buffer_block
) {
1705 mode
= vtn_variable_mode_ssbo
;
1706 nir_mode
= nir_var_mem_ssbo
;
1708 /* Default-block uniforms, coming from gl_spirv */
1709 mode
= vtn_variable_mode_uniform
;
1710 nir_mode
= nir_var_uniform
;
1713 case SpvStorageClassStorageBuffer
:
1714 mode
= vtn_variable_mode_ssbo
;
1715 nir_mode
= nir_var_mem_ssbo
;
1717 case SpvStorageClassPhysicalStorageBufferEXT
:
1718 mode
= vtn_variable_mode_phys_ssbo
;
1719 nir_mode
= nir_var_mem_global
;
1721 case SpvStorageClassUniformConstant
:
1722 mode
= vtn_variable_mode_uniform
;
1723 nir_mode
= nir_var_uniform
;
1725 case SpvStorageClassPushConstant
:
1726 mode
= vtn_variable_mode_push_constant
;
1727 nir_mode
= nir_var_uniform
;
1729 case SpvStorageClassInput
:
1730 mode
= vtn_variable_mode_input
;
1731 nir_mode
= nir_var_shader_in
;
1733 case SpvStorageClassOutput
:
1734 mode
= vtn_variable_mode_output
;
1735 nir_mode
= nir_var_shader_out
;
1737 case SpvStorageClassPrivate
:
1738 mode
= vtn_variable_mode_private
;
1739 nir_mode
= nir_var_shader_temp
;
1741 case SpvStorageClassFunction
:
1742 mode
= vtn_variable_mode_function
;
1743 nir_mode
= nir_var_function_temp
;
1745 case SpvStorageClassWorkgroup
:
1746 mode
= vtn_variable_mode_workgroup
;
1747 nir_mode
= nir_var_mem_shared
;
1749 case SpvStorageClassAtomicCounter
:
1750 mode
= vtn_variable_mode_uniform
;
1751 nir_mode
= nir_var_uniform
;
1753 case SpvStorageClassCrossWorkgroup
:
1754 mode
= vtn_variable_mode_cross_workgroup
;
1755 nir_mode
= nir_var_mem_global
;
1757 case SpvStorageClassGeneric
:
1759 vtn_fail("Unhandled variable storage class");
1763 *nir_mode_out
= nir_mode
;
1769 vtn_pointer_to_ssa(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
1771 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1772 /* This pointer needs to have a pointer type with actual storage */
1773 vtn_assert(ptr
->ptr_type
);
1774 vtn_assert(ptr
->ptr_type
->type
);
1777 /* If we don't have an offset then we must be a pointer to the variable
1780 vtn_assert(!ptr
->offset
&& !ptr
->block_index
);
1782 struct vtn_access_chain chain
= {
1785 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
1788 vtn_assert(ptr
->offset
);
1789 if (ptr
->block_index
) {
1790 vtn_assert(ptr
->mode
== vtn_variable_mode_ubo
||
1791 ptr
->mode
== vtn_variable_mode_ssbo
);
1792 return nir_vec2(&b
->nb
, ptr
->block_index
, ptr
->offset
);
1794 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
);
1798 if (vtn_pointer_is_external_block(b
, ptr
) &&
1799 vtn_type_contains_block(b
, ptr
->type
) &&
1800 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
1801 /* In this case, we're looking for a block index and not an actual
1804 * For PhysicalStorageBufferEXT pointers, we don't have a block index
1805 * at all because we get the pointer directly from the client. This
1806 * assumes that there will never be a SSBO binding variable using the
1807 * PhysicalStorageBufferEXT storage class. This assumption appears
1808 * to be correct according to the Vulkan spec because the table,
1809 * "Shader Resource and Storage Class Correspondence," the only the
1810 * Uniform storage class with BufferBlock or the StorageBuffer
1811 * storage class with Block can be used.
1813 if (!ptr
->block_index
) {
1814 /* If we don't have a block_index then we must be a pointer to the
1817 vtn_assert(!ptr
->deref
);
1819 struct vtn_access_chain chain
= {
1822 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
1825 return ptr
->block_index
;
1827 return &vtn_pointer_to_deref(b
, ptr
)->dest
.ssa
;
1832 struct vtn_pointer
*
1833 vtn_pointer_from_ssa(struct vtn_builder
*b
, nir_ssa_def
*ssa
,
1834 struct vtn_type
*ptr_type
)
1836 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1838 struct vtn_type
*interface_type
= ptr_type
->deref
;
1839 while (interface_type
->base_type
== vtn_base_type_array
)
1840 interface_type
= interface_type
->array_element
;
1842 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
1843 nir_variable_mode nir_mode
;
1844 ptr
->mode
= vtn_storage_class_to_mode(b
, ptr_type
->storage_class
,
1845 interface_type
, &nir_mode
);
1846 ptr
->type
= ptr_type
->deref
;
1847 ptr
->ptr_type
= ptr_type
;
1849 if (b
->wa_glslang_179
) {
1850 /* To work around https://github.com/KhronosGroup/glslang/issues/179 we
1851 * need to whack the mode because it creates a function parameter with
1852 * the Function storage class even though it's a pointer to a sampler.
1853 * If we don't do this, then NIR won't get rid of the deref_cast for us.
1855 if (ptr
->mode
== vtn_variable_mode_function
&&
1856 (ptr
->type
->base_type
== vtn_base_type_sampler
||
1857 ptr
->type
->base_type
== vtn_base_type_sampled_image
)) {
1858 ptr
->mode
= vtn_variable_mode_uniform
;
1859 nir_mode
= nir_var_uniform
;
1863 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1864 /* This pointer type needs to have actual storage */
1865 vtn_assert(ptr_type
->type
);
1866 if (ptr
->mode
== vtn_variable_mode_ubo
||
1867 ptr
->mode
== vtn_variable_mode_ssbo
) {
1868 vtn_assert(ssa
->num_components
== 2);
1869 ptr
->block_index
= nir_channel(&b
->nb
, ssa
, 0);
1870 ptr
->offset
= nir_channel(&b
->nb
, ssa
, 1);
1872 vtn_assert(ssa
->num_components
== 1);
1873 ptr
->block_index
= NULL
;
1877 const struct glsl_type
*deref_type
= ptr_type
->deref
->type
;
1878 if (!vtn_pointer_is_external_block(b
, ptr
)) {
1879 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
1881 } else if (vtn_type_contains_block(b
, ptr
->type
) &&
1882 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
1883 /* This is a pointer to somewhere in an array of blocks, not a
1884 * pointer to somewhere inside the block. Set the block index
1885 * instead of making a cast.
1887 ptr
->block_index
= ssa
;
1889 /* This is a pointer to something internal or a pointer inside a
1890 * block. It's just a regular cast.
1892 * For PhysicalStorageBufferEXT pointers, we don't have a block index
1893 * at all because we get the pointer directly from the client. This
1894 * assumes that there will never be a SSBO binding variable using the
1895 * PhysicalStorageBufferEXT storage class. This assumption appears
1896 * to be correct according to the Vulkan spec because the table,
1897 * "Shader Resource and Storage Class Correspondence," the only the
1898 * Uniform storage class with BufferBlock or the StorageBuffer
1899 * storage class with Block can be used.
1901 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
1902 ptr_type
->deref
->type
,
1904 ptr
->deref
->dest
.ssa
.num_components
=
1905 glsl_get_vector_elements(ptr_type
->type
);
1906 ptr
->deref
->dest
.ssa
.bit_size
= glsl_get_bit_size(ptr_type
->type
);
1914 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
1916 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
1919 if (var
->mode
== vtn_variable_mode_input
) {
1920 return stage
== MESA_SHADER_TESS_CTRL
||
1921 stage
== MESA_SHADER_TESS_EVAL
||
1922 stage
== MESA_SHADER_GEOMETRY
;
1925 if (var
->mode
== vtn_variable_mode_output
)
1926 return stage
== MESA_SHADER_TESS_CTRL
;
1932 assign_missing_member_locations(struct vtn_variable
*var
)
1935 glsl_get_length(glsl_without_array(var
->type
->type
));
1936 int location
= var
->base_location
;
1938 for (unsigned i
= 0; i
< length
; i
++) {
1939 /* From the Vulkan spec:
1941 * “If the structure type is a Block but without a Location, then each
1942 * of its members must have a Location decoration.”
1945 if (var
->type
->block
) {
1946 assert(var
->base_location
!= -1 ||
1947 var
->var
->members
[i
].location
!= -1);
1950 /* From the Vulkan spec:
1952 * “Any member with its own Location decoration is assigned that
1953 * location. Each remaining member is assigned the location after the
1954 * immediately preceding member in declaration order.”
1956 if (var
->var
->members
[i
].location
!= -1)
1957 location
= var
->var
->members
[i
].location
;
1959 var
->var
->members
[i
].location
= location
;
1961 /* Below we use type instead of interface_type, because interface_type
1962 * is only available when it is a Block. This code also supports
1963 * input/outputs that are just structs
1965 const struct glsl_type
*member_type
=
1966 glsl_get_struct_field(glsl_without_array(var
->type
->type
), i
);
1969 glsl_count_attribute_slots(member_type
,
1970 false /* is_gl_vertex_input */);
1976 vtn_create_variable(struct vtn_builder
*b
, struct vtn_value
*val
,
1977 struct vtn_type
*ptr_type
, SpvStorageClass storage_class
,
1978 nir_constant
*initializer
)
1980 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1981 struct vtn_type
*type
= ptr_type
->deref
;
1983 struct vtn_type
*without_array
= type
;
1984 while(glsl_type_is_array(without_array
->type
))
1985 without_array
= without_array
->array_element
;
1987 enum vtn_variable_mode mode
;
1988 nir_variable_mode nir_mode
;
1989 mode
= vtn_storage_class_to_mode(b
, storage_class
, without_array
, &nir_mode
);
1992 case vtn_variable_mode_ubo
:
1993 /* There's no other way to get vtn_variable_mode_ubo */
1994 vtn_assert(without_array
->block
);
1995 b
->shader
->info
.num_ubos
++;
1997 case vtn_variable_mode_ssbo
:
1998 if (storage_class
== SpvStorageClassStorageBuffer
&&
1999 !without_array
->block
) {
2000 if (b
->variable_pointers
) {
2001 vtn_fail("Variables in the StorageBuffer storage class must "
2002 "have a struct type with the Block decoration");
2004 /* If variable pointers are not present, it's still malformed
2005 * SPIR-V but we can parse it and do the right thing anyway.
2006 * Since some of the 8-bit storage tests have bugs in this are,
2007 * just make it a warning for now.
2009 vtn_warn("Variables in the StorageBuffer storage class must "
2010 "have a struct type with the Block decoration");
2013 b
->shader
->info
.num_ssbos
++;
2015 case vtn_variable_mode_uniform
:
2016 if (glsl_type_is_image(without_array
->type
))
2017 b
->shader
->info
.num_images
++;
2018 else if (glsl_type_is_sampler(without_array
->type
))
2019 b
->shader
->info
.num_textures
++;
2021 case vtn_variable_mode_push_constant
:
2022 b
->shader
->num_uniforms
= vtn_type_block_size(b
, type
);
2025 case vtn_variable_mode_phys_ssbo
:
2026 vtn_fail("Cannot create a variable with the "
2027 "PhysicalStorageBufferEXT storage class");
2031 /* No tallying is needed */
2035 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
2038 var
->base_location
= -1;
2040 vtn_assert(val
->value_type
== vtn_value_type_pointer
);
2041 val
->pointer
= vtn_pointer_for_variable(b
, var
, ptr_type
);
2043 switch (var
->mode
) {
2044 case vtn_variable_mode_function
:
2045 case vtn_variable_mode_private
:
2046 case vtn_variable_mode_uniform
:
2047 /* For these, we create the variable normally */
2048 var
->var
= rzalloc(b
->shader
, nir_variable
);
2049 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2051 if (storage_class
== SpvStorageClassAtomicCounter
) {
2052 /* Need to tweak the nir type here as at vtn_handle_type we don't
2053 * have the access to storage_class, that is the one that points us
2054 * that is an atomic uint.
2056 var
->var
->type
= repair_atomic_type(var
->type
->type
);
2058 /* Private variables don't have any explicit layout but some layouts
2059 * may have leaked through due to type deduplication in the SPIR-V.
2061 var
->var
->type
= var
->type
->type
;
2063 var
->var
->data
.mode
= nir_mode
;
2064 var
->var
->data
.location
= -1;
2065 var
->var
->interface_type
= NULL
;
2068 case vtn_variable_mode_workgroup
:
2069 if (b
->options
->lower_workgroup_access_to_offsets
) {
2070 var
->shared_location
= -1;
2072 /* Create the variable normally */
2073 var
->var
= rzalloc(b
->shader
, nir_variable
);
2074 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2075 /* Workgroup variables don't have any explicit layout but some
2076 * layouts may have leaked through due to type deduplication in the
2079 var
->var
->type
= var
->type
->type
;
2080 var
->var
->data
.mode
= nir_var_mem_shared
;
2084 case vtn_variable_mode_input
:
2085 case vtn_variable_mode_output
: {
2086 /* In order to know whether or not we're a per-vertex inout, we need
2087 * the patch qualifier. This means walking the variable decorations
2088 * early before we actually create any variables. Not a big deal.
2090 * GLSLang really likes to place decorations in the most interior
2091 * thing it possibly can. In particular, if you have a struct, it
2092 * will place the patch decorations on the struct members. This
2093 * should be handled by the variable splitting below just fine.
2095 * If you have an array-of-struct, things get even more weird as it
2096 * will place the patch decorations on the struct even though it's
2097 * inside an array and some of the members being patch and others not
2098 * makes no sense whatsoever. Since the only sensible thing is for
2099 * it to be all or nothing, we'll call it patch if any of the members
2100 * are declared patch.
2103 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
2104 if (glsl_type_is_array(var
->type
->type
) &&
2105 glsl_type_is_struct_or_ifc(without_array
->type
)) {
2106 vtn_foreach_decoration(b
, vtn_value(b
, without_array
->id
,
2107 vtn_value_type_type
),
2108 var_is_patch_cb
, &var
->patch
);
2111 /* For inputs and outputs, we immediately split structures. This
2112 * is for a couple of reasons. For one, builtins may all come in
2113 * a struct and we really want those split out into separate
2114 * variables. For another, interpolation qualifiers can be
2115 * applied to members of the top-level struct ane we need to be
2116 * able to preserve that information.
2119 struct vtn_type
*per_vertex_type
= var
->type
;
2120 if (is_per_vertex_inout(var
, b
->shader
->info
.stage
)) {
2121 /* In Geometry shaders (and some tessellation), inputs come
2122 * in per-vertex arrays. However, some builtins come in
2123 * non-per-vertex, hence the need for the is_array check. In
2124 * any case, there are no non-builtin arrays allowed so this
2125 * check should be sufficient.
2127 per_vertex_type
= var
->type
->array_element
;
2130 var
->var
= rzalloc(b
->shader
, nir_variable
);
2131 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2132 /* In Vulkan, shader I/O variables don't have any explicit layout but
2133 * some layouts may have leaked through due to type deduplication in
2134 * the SPIR-V. We do, however, keep the layouts in the variable's
2135 * interface_type because we need offsets for XFB arrays of blocks.
2137 var
->var
->type
= var
->type
->type
;
2138 var
->var
->data
.mode
= nir_mode
;
2139 var
->var
->data
.patch
= var
->patch
;
2141 /* Figure out the interface block type. */
2142 struct vtn_type
*iface_type
= per_vertex_type
;
2143 if (var
->mode
== vtn_variable_mode_output
&&
2144 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
2145 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
||
2146 b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)) {
2147 /* For vertex data outputs, we can end up with arrays of blocks for
2148 * transform feedback where each array element corresponds to a
2149 * different XFB output buffer.
2151 while (iface_type
->base_type
== vtn_base_type_array
)
2152 iface_type
= iface_type
->array_element
;
2154 if (iface_type
->base_type
== vtn_base_type_struct
&& iface_type
->block
)
2155 var
->var
->interface_type
= iface_type
->type
;
2157 if (per_vertex_type
->base_type
== vtn_base_type_struct
&&
2158 per_vertex_type
->block
) {
2159 /* It's a struct. Set it up as per-member. */
2160 var
->var
->num_members
= glsl_get_length(per_vertex_type
->type
);
2161 var
->var
->members
= rzalloc_array(var
->var
, struct nir_variable_data
,
2162 var
->var
->num_members
);
2164 for (unsigned i
= 0; i
< var
->var
->num_members
; i
++) {
2165 var
->var
->members
[i
].mode
= nir_mode
;
2166 var
->var
->members
[i
].patch
= var
->patch
;
2167 var
->var
->members
[i
].location
= -1;
2171 /* For inputs and outputs, we need to grab locations and builtin
2172 * information from the per-vertex type.
2174 vtn_foreach_decoration(b
, vtn_value(b
, per_vertex_type
->id
,
2175 vtn_value_type_type
),
2176 var_decoration_cb
, var
);
2180 case vtn_variable_mode_ubo
:
2181 case vtn_variable_mode_ssbo
:
2182 case vtn_variable_mode_push_constant
:
2183 case vtn_variable_mode_cross_workgroup
:
2184 /* These don't need actual variables. */
2187 case vtn_variable_mode_phys_ssbo
:
2188 unreachable("Should have been caught before");
2192 var
->var
->constant_initializer
=
2193 nir_constant_clone(initializer
, var
->var
);
2196 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
2198 if ((var
->mode
== vtn_variable_mode_input
||
2199 var
->mode
== vtn_variable_mode_output
) &&
2200 var
->var
->members
) {
2201 assign_missing_member_locations(var
);
2204 if (var
->mode
== vtn_variable_mode_uniform
) {
2205 /* XXX: We still need the binding information in the nir_variable
2206 * for these. We should fix that.
2208 var
->var
->data
.binding
= var
->binding
;
2209 var
->var
->data
.explicit_binding
= var
->explicit_binding
;
2210 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
2211 var
->var
->data
.index
= var
->input_attachment_index
;
2212 var
->var
->data
.offset
= var
->offset
;
2214 if (glsl_type_is_image(without_array
->type
))
2215 var
->var
->data
.image
.format
= without_array
->image_format
;
2218 if (var
->mode
== vtn_variable_mode_function
) {
2219 vtn_assert(var
->var
!= NULL
&& var
->var
->members
== NULL
);
2220 nir_function_impl_add_variable(b
->nb
.impl
, var
->var
);
2221 } else if (var
->var
) {
2222 nir_shader_add_variable(b
->shader
, var
->var
);
2224 vtn_assert(vtn_pointer_is_external_block(b
, val
->pointer
));
2229 vtn_assert_types_equal(struct vtn_builder
*b
, SpvOp opcode
,
2230 struct vtn_type
*dst_type
,
2231 struct vtn_type
*src_type
)
2233 if (dst_type
->id
== src_type
->id
)
2236 if (vtn_types_compatible(b
, dst_type
, src_type
)) {
2237 /* Early versions of GLSLang would re-emit types unnecessarily and you
2238 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
2239 * mismatched source and destination types.
2241 * https://github.com/KhronosGroup/glslang/issues/304
2242 * https://github.com/KhronosGroup/glslang/issues/307
2243 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
2244 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2246 vtn_warn("Source and destination types of %s do not have the same "
2247 "ID (but are compatible): %u vs %u",
2248 spirv_op_to_string(opcode
), dst_type
->id
, src_type
->id
);
2252 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
2253 spirv_op_to_string(opcode
),
2254 glsl_get_type_name(dst_type
->type
),
2255 glsl_get_type_name(src_type
->type
));
2258 static nir_ssa_def
*
2259 nir_shrink_zero_pad_vec(nir_builder
*b
, nir_ssa_def
*val
,
2260 unsigned num_components
)
2262 if (val
->num_components
== num_components
)
2265 nir_ssa_def
*comps
[NIR_MAX_VEC_COMPONENTS
];
2266 for (unsigned i
= 0; i
< num_components
; i
++) {
2267 if (i
< val
->num_components
)
2268 comps
[i
] = nir_channel(b
, val
, i
);
2270 comps
[i
] = nir_imm_intN_t(b
, 0, val
->bit_size
);
2272 return nir_vec(b
, comps
, num_components
);
2275 static nir_ssa_def
*
2276 nir_sloppy_bitcast(nir_builder
*b
, nir_ssa_def
*val
,
2277 const struct glsl_type
*type
)
2279 const unsigned num_components
= glsl_get_vector_elements(type
);
2280 const unsigned bit_size
= glsl_get_bit_size(type
);
2282 /* First, zero-pad to ensure that the value is big enough that when we
2283 * bit-cast it, we don't loose anything.
2285 if (val
->bit_size
< bit_size
) {
2286 const unsigned src_num_components_needed
=
2287 vtn_align_u32(val
->num_components
, bit_size
/ val
->bit_size
);
2288 val
= nir_shrink_zero_pad_vec(b
, val
, src_num_components_needed
);
2291 val
= nir_bitcast_vector(b
, val
, bit_size
);
2293 return nir_shrink_zero_pad_vec(b
, val
, num_components
);
2297 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
2298 const uint32_t *w
, unsigned count
)
2302 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2303 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2307 case SpvOpVariable
: {
2308 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2310 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2312 SpvStorageClass storage_class
= w
[3];
2313 nir_constant
*initializer
= NULL
;
2315 initializer
= vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
2317 vtn_create_variable(b
, val
, ptr_type
, storage_class
, initializer
);
2321 case SpvOpAccessChain
:
2322 case SpvOpPtrAccessChain
:
2323 case SpvOpInBoundsAccessChain
:
2324 case SpvOpInBoundsPtrAccessChain
: {
2325 struct vtn_access_chain
*chain
= vtn_access_chain_create(b
, count
- 4);
2326 chain
->ptr_as_array
= (opcode
== SpvOpPtrAccessChain
|| opcode
== SpvOpInBoundsPtrAccessChain
);
2329 for (int i
= 4; i
< count
; i
++) {
2330 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
2331 if (link_val
->value_type
== vtn_value_type_constant
) {
2332 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
2333 switch (glsl_get_bit_size(link_val
->type
->type
)) {
2335 chain
->link
[idx
].id
= link_val
->constant
->values
[0].i8
[0];
2338 chain
->link
[idx
].id
= link_val
->constant
->values
[0].i16
[0];
2341 chain
->link
[idx
].id
= link_val
->constant
->values
[0].i32
[0];
2344 chain
->link
[idx
].id
= link_val
->constant
->values
[0].i64
[0];
2347 vtn_fail("Invalid bit size");
2350 chain
->link
[idx
].mode
= vtn_access_mode_id
;
2351 chain
->link
[idx
].id
= w
[i
];
2357 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2358 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
2359 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
2360 /* This is rather insane. SPIR-V allows you to use OpSampledImage
2361 * to combine an array of images with a single sampler to get an
2362 * array of sampled images that all share the same sampler.
2363 * Fortunately, this means that we can more-or-less ignore the
2364 * sampler when crawling the access chain, but it does leave us
2365 * with this rather awkward little special-case.
2367 struct vtn_value
*val
=
2368 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2369 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2370 val
->sampled_image
->type
= base_val
->sampled_image
->type
;
2371 val
->sampled_image
->image
=
2372 vtn_pointer_dereference(b
, base_val
->sampled_image
->image
, chain
);
2373 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
2375 vtn_assert(base_val
->value_type
== vtn_value_type_pointer
);
2376 struct vtn_value
*val
=
2377 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2378 val
->pointer
= vtn_pointer_dereference(b
, base_val
->pointer
, chain
);
2379 val
->pointer
->ptr_type
= ptr_type
;
2384 case SpvOpCopyMemory
: {
2385 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2386 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_pointer
);
2388 vtn_assert_types_equal(b
, opcode
, dest
->type
->deref
, src
->type
->deref
);
2390 vtn_variable_copy(b
, dest
->pointer
, src
->pointer
);
2395 struct vtn_type
*res_type
=
2396 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2397 struct vtn_value
*src_val
= vtn_value(b
, w
[3], vtn_value_type_pointer
);
2398 struct vtn_pointer
*src
= src_val
->pointer
;
2400 vtn_assert_types_equal(b
, opcode
, res_type
, src_val
->type
->deref
);
2402 if (glsl_type_is_image(res_type
->type
) ||
2403 glsl_type_is_sampler(res_type
->type
)) {
2404 vtn_push_value(b
, w
[2], vtn_value_type_pointer
)->pointer
= src
;
2408 vtn_push_ssa(b
, w
[2], res_type
, vtn_variable_load(b
, src
));
2413 struct vtn_value
*dest_val
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2414 struct vtn_pointer
*dest
= dest_val
->pointer
;
2415 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[2]);
2417 /* OpStore requires us to actually have a storage type */
2418 vtn_fail_if(dest
->type
->type
== NULL
,
2419 "Invalid destination type for OpStore");
2421 if (glsl_get_base_type(dest
->type
->type
) == GLSL_TYPE_BOOL
&&
2422 glsl_get_base_type(src_val
->type
->type
) == GLSL_TYPE_UINT
) {
2423 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2424 * would then store them to a local variable as bool. Work around
2425 * the issue by doing an implicit conversion.
2427 * https://github.com/KhronosGroup/glslang/issues/170
2428 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2430 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2431 "OpTypeBool. Doing an implicit conversion to work around "
2433 struct vtn_ssa_value
*bool_ssa
=
2434 vtn_create_ssa_value(b
, dest
->type
->type
);
2435 bool_ssa
->def
= nir_i2b(&b
->nb
, vtn_ssa_value(b
, w
[2])->def
);
2436 vtn_variable_store(b
, bool_ssa
, dest
);
2440 vtn_assert_types_equal(b
, opcode
, dest_val
->type
->deref
, src_val
->type
);
2442 if (glsl_type_is_sampler(dest
->type
->type
)) {
2443 if (b
->wa_glslang_179
) {
2444 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2445 "propagation to workaround the problem.");
2446 vtn_assert(dest
->var
->copy_prop_sampler
== NULL
);
2447 dest
->var
->copy_prop_sampler
=
2448 vtn_value(b
, w
[2], vtn_value_type_pointer
)->pointer
;
2450 vtn_fail("Vulkan does not allow OpStore of a sampler or image.");
2455 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
2456 vtn_variable_store(b
, src
, dest
);
2460 case SpvOpArrayLength
: {
2461 struct vtn_pointer
*ptr
=
2462 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2463 const uint32_t field
= w
[4];
2465 vtn_fail_if(ptr
->type
->base_type
!= vtn_base_type_struct
,
2466 "OpArrayLength must take a pointer to a structure type");
2467 vtn_fail_if(field
!= ptr
->type
->length
- 1 ||
2468 ptr
->type
->members
[field
]->base_type
!= vtn_base_type_array
,
2469 "OpArrayLength must reference the last memeber of the "
2470 "structure and that must be an array");
2472 const uint32_t offset
= ptr
->type
->offsets
[field
];
2473 const uint32_t stride
= ptr
->type
->members
[field
]->stride
;
2475 if (!ptr
->block_index
) {
2476 struct vtn_access_chain chain
= {
2479 ptr
= vtn_pointer_dereference(b
, ptr
, &chain
);
2480 vtn_assert(ptr
->block_index
);
2483 nir_intrinsic_instr
*instr
=
2484 nir_intrinsic_instr_create(b
->nb
.shader
,
2485 nir_intrinsic_get_buffer_size
);
2486 instr
->src
[0] = nir_src_for_ssa(ptr
->block_index
);
2487 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
2488 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2489 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
2491 /* array_length = max(buffer_size - offset, 0) / stride */
2492 nir_ssa_def
*array_length
=
2497 nir_imm_int(&b
->nb
, offset
)),
2498 nir_imm_int(&b
->nb
, 0u)),
2499 nir_imm_int(&b
->nb
, stride
));
2501 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2502 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
2503 val
->ssa
->def
= array_length
;
2507 case SpvOpConvertPtrToU
: {
2508 struct vtn_value
*u_val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2510 vtn_fail_if(u_val
->type
->base_type
!= vtn_base_type_vector
&&
2511 u_val
->type
->base_type
!= vtn_base_type_scalar
,
2512 "OpConvertPtrToU can only be used to cast to a vector or "
2515 /* The pointer will be converted to an SSA value automatically */
2516 nir_ssa_def
*ptr_ssa
= vtn_ssa_value(b
, w
[3])->def
;
2518 u_val
->ssa
= vtn_create_ssa_value(b
, u_val
->type
->type
);
2519 u_val
->ssa
->def
= nir_sloppy_bitcast(&b
->nb
, ptr_ssa
, u_val
->type
->type
);
2523 case SpvOpConvertUToPtr
: {
2524 struct vtn_value
*ptr_val
=
2525 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2526 struct vtn_value
*u_val
= vtn_value(b
, w
[3], vtn_value_type_ssa
);
2528 vtn_fail_if(ptr_val
->type
->type
== NULL
,
2529 "OpConvertUToPtr can only be used on physical pointers");
2531 vtn_fail_if(u_val
->type
->base_type
!= vtn_base_type_vector
&&
2532 u_val
->type
->base_type
!= vtn_base_type_scalar
,
2533 "OpConvertUToPtr can only be used to cast from a vector or "
2536 nir_ssa_def
*ptr_ssa
= nir_sloppy_bitcast(&b
->nb
, u_val
->ssa
->def
,
2537 ptr_val
->type
->type
);
2538 ptr_val
->pointer
= vtn_pointer_from_ssa(b
, ptr_ssa
, ptr_val
->type
);
2542 case SpvOpCopyMemorySized
:
2544 vtn_fail("Unhandled opcode");