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
,
594 enum gl_access_qualifier access
)
596 if (glsl_type_is_vector_or_scalar(deref
->type
)) {
598 inout
->def
= nir_load_deref_with_access(&b
->nb
, deref
, access
);
600 nir_store_deref_with_access(&b
->nb
, deref
, inout
->def
, ~0, access
);
602 } else if (glsl_type_is_array(deref
->type
) ||
603 glsl_type_is_matrix(deref
->type
)) {
604 unsigned elems
= glsl_get_length(deref
->type
);
605 for (unsigned i
= 0; i
< elems
; i
++) {
606 nir_deref_instr
*child
=
607 nir_build_deref_array_imm(&b
->nb
, deref
, i
);
608 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
], access
);
611 vtn_assert(glsl_type_is_struct_or_ifc(deref
->type
));
612 unsigned elems
= glsl_get_length(deref
->type
);
613 for (unsigned i
= 0; i
< elems
; i
++) {
614 nir_deref_instr
*child
= nir_build_deref_struct(&b
->nb
, deref
, i
);
615 _vtn_local_load_store(b
, load
, child
, inout
->elems
[i
], access
);
621 vtn_nir_deref(struct vtn_builder
*b
, uint32_t id
)
623 struct vtn_pointer
*ptr
= vtn_value(b
, id
, vtn_value_type_pointer
)->pointer
;
624 return vtn_pointer_to_deref(b
, ptr
);
628 * Gets the NIR-level deref tail, which may have as a child an array deref
629 * selecting which component due to OpAccessChain supporting per-component
630 * indexing in SPIR-V.
632 static nir_deref_instr
*
633 get_deref_tail(nir_deref_instr
*deref
)
635 if (deref
->deref_type
!= nir_deref_type_array
)
638 nir_deref_instr
*parent
=
639 nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
);
641 if (glsl_type_is_vector(parent
->type
))
647 struct vtn_ssa_value
*
648 vtn_local_load(struct vtn_builder
*b
, nir_deref_instr
*src
,
649 enum gl_access_qualifier access
)
651 nir_deref_instr
*src_tail
= get_deref_tail(src
);
652 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, src_tail
->type
);
653 _vtn_local_load_store(b
, true, src_tail
, val
, access
);
655 if (src_tail
!= src
) {
656 val
->type
= src
->type
;
657 if (nir_src_is_const(src
->arr
.index
))
658 val
->def
= vtn_vector_extract(b
, val
->def
,
659 nir_src_as_uint(src
->arr
.index
));
661 val
->def
= vtn_vector_extract_dynamic(b
, val
->def
, src
->arr
.index
.ssa
);
668 vtn_local_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
669 nir_deref_instr
*dest
, enum gl_access_qualifier access
)
671 nir_deref_instr
*dest_tail
= get_deref_tail(dest
);
673 if (dest_tail
!= dest
) {
674 struct vtn_ssa_value
*val
= vtn_create_ssa_value(b
, dest_tail
->type
);
675 _vtn_local_load_store(b
, true, dest_tail
, val
, access
);
677 if (nir_src_is_const(dest
->arr
.index
))
678 val
->def
= vtn_vector_insert(b
, val
->def
, src
->def
,
679 nir_src_as_uint(dest
->arr
.index
));
681 val
->def
= vtn_vector_insert_dynamic(b
, val
->def
, src
->def
,
682 dest
->arr
.index
.ssa
);
683 _vtn_local_load_store(b
, false, dest_tail
, val
, access
);
685 _vtn_local_load_store(b
, false, dest_tail
, src
, access
);
690 vtn_pointer_to_offset(struct vtn_builder
*b
, struct vtn_pointer
*ptr
,
691 nir_ssa_def
**index_out
)
693 assert(vtn_pointer_uses_ssa_offset(b
, ptr
));
695 struct vtn_access_chain chain
= {
698 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
700 *index_out
= ptr
->block_index
;
704 /* Tries to compute the size of an interface block based on the strides and
705 * offsets that are provided to us in the SPIR-V source.
708 vtn_type_block_size(struct vtn_builder
*b
, struct vtn_type
*type
)
710 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
714 case GLSL_TYPE_UINT16
:
715 case GLSL_TYPE_INT16
:
716 case GLSL_TYPE_UINT8
:
718 case GLSL_TYPE_UINT64
:
719 case GLSL_TYPE_INT64
:
720 case GLSL_TYPE_FLOAT
:
721 case GLSL_TYPE_FLOAT16
:
723 case GLSL_TYPE_DOUBLE
: {
724 unsigned cols
= type
->row_major
? glsl_get_vector_elements(type
->type
) :
725 glsl_get_matrix_columns(type
->type
);
727 vtn_assert(type
->stride
> 0);
728 return type
->stride
* cols
;
730 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
731 return glsl_get_vector_elements(type
->type
) * type_size
;
735 case GLSL_TYPE_STRUCT
:
736 case GLSL_TYPE_INTERFACE
: {
738 unsigned num_fields
= glsl_get_length(type
->type
);
739 for (unsigned f
= 0; f
< num_fields
; f
++) {
740 unsigned field_end
= type
->offsets
[f
] +
741 vtn_type_block_size(b
, type
->members
[f
]);
742 size
= MAX2(size
, field_end
);
747 case GLSL_TYPE_ARRAY
:
748 vtn_assert(type
->stride
> 0);
749 vtn_assert(glsl_get_length(type
->type
) > 0);
750 return type
->stride
* glsl_get_length(type
->type
);
753 vtn_fail("Invalid block type");
759 _vtn_load_store_tail(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
760 nir_ssa_def
*index
, nir_ssa_def
*offset
,
761 unsigned access_offset
, unsigned access_size
,
762 struct vtn_ssa_value
**inout
, const struct glsl_type
*type
,
763 enum gl_access_qualifier access
)
765 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(b
->nb
.shader
, op
);
766 instr
->num_components
= glsl_get_vector_elements(type
);
768 /* Booleans usually shouldn't show up in external memory in SPIR-V.
769 * However, they do for certain older GLSLang versions and can for shared
770 * memory when we lower access chains internally.
772 const unsigned data_bit_size
= glsl_type_is_boolean(type
) ? 32 :
773 glsl_get_bit_size(type
);
777 nir_intrinsic_set_write_mask(instr
, (1 << instr
->num_components
) - 1);
778 instr
->src
[src
++] = nir_src_for_ssa((*inout
)->def
);
781 if (op
== nir_intrinsic_load_push_constant
) {
782 nir_intrinsic_set_base(instr
, access_offset
);
783 nir_intrinsic_set_range(instr
, access_size
);
786 if (op
== nir_intrinsic_load_ubo
||
787 op
== nir_intrinsic_load_ssbo
||
788 op
== nir_intrinsic_store_ssbo
) {
789 nir_intrinsic_set_access(instr
, access
);
792 /* With extensions like relaxed_block_layout, we really can't guarantee
793 * much more than scalar alignment.
795 if (op
!= nir_intrinsic_load_push_constant
)
796 nir_intrinsic_set_align(instr
, data_bit_size
/ 8, 0);
799 instr
->src
[src
++] = nir_src_for_ssa(index
);
801 if (op
== nir_intrinsic_load_push_constant
) {
802 /* We need to subtract the offset from where the intrinsic will load the
805 nir_src_for_ssa(nir_isub(&b
->nb
, offset
,
806 nir_imm_int(&b
->nb
, access_offset
)));
808 instr
->src
[src
++] = nir_src_for_ssa(offset
);
812 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
813 instr
->num_components
, data_bit_size
, NULL
);
814 (*inout
)->def
= &instr
->dest
.ssa
;
817 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
819 if (load
&& glsl_get_base_type(type
) == GLSL_TYPE_BOOL
)
820 (*inout
)->def
= nir_ine(&b
->nb
, (*inout
)->def
, nir_imm_int(&b
->nb
, 0));
824 _vtn_block_load_store(struct vtn_builder
*b
, nir_intrinsic_op op
, bool load
,
825 nir_ssa_def
*index
, nir_ssa_def
*offset
,
826 unsigned access_offset
, unsigned access_size
,
827 struct vtn_type
*type
, enum gl_access_qualifier access
,
828 struct vtn_ssa_value
**inout
)
830 if (load
&& *inout
== NULL
)
831 *inout
= vtn_create_ssa_value(b
, type
->type
);
833 enum glsl_base_type base_type
= glsl_get_base_type(type
->type
);
837 case GLSL_TYPE_UINT16
:
838 case GLSL_TYPE_INT16
:
839 case GLSL_TYPE_UINT8
:
841 case GLSL_TYPE_UINT64
:
842 case GLSL_TYPE_INT64
:
843 case GLSL_TYPE_FLOAT
:
844 case GLSL_TYPE_FLOAT16
:
845 case GLSL_TYPE_DOUBLE
:
847 /* This is where things get interesting. At this point, we've hit
848 * a vector, a scalar, or a matrix.
850 if (glsl_type_is_matrix(type
->type
)) {
851 /* Loading the whole matrix */
852 struct vtn_ssa_value
*transpose
;
853 unsigned num_ops
, vec_width
, col_stride
;
854 if (type
->row_major
) {
855 num_ops
= glsl_get_vector_elements(type
->type
);
856 vec_width
= glsl_get_matrix_columns(type
->type
);
857 col_stride
= type
->array_element
->stride
;
859 const struct glsl_type
*transpose_type
=
860 glsl_matrix_type(base_type
, vec_width
, num_ops
);
861 *inout
= vtn_create_ssa_value(b
, transpose_type
);
863 transpose
= vtn_ssa_transpose(b
, *inout
);
867 num_ops
= glsl_get_matrix_columns(type
->type
);
868 vec_width
= glsl_get_vector_elements(type
->type
);
869 col_stride
= type
->stride
;
872 for (unsigned i
= 0; i
< num_ops
; i
++) {
873 nir_ssa_def
*elem_offset
=
874 nir_iadd_imm(&b
->nb
, offset
, i
* col_stride
);
875 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
876 access_offset
, access_size
,
878 glsl_vector_type(base_type
, vec_width
),
879 type
->access
| access
);
882 if (load
&& type
->row_major
)
883 *inout
= vtn_ssa_transpose(b
, *inout
);
885 unsigned elems
= glsl_get_vector_elements(type
->type
);
886 unsigned type_size
= glsl_get_bit_size(type
->type
) / 8;
887 if (elems
== 1 || type
->stride
== type_size
) {
888 /* This is a tightly-packed normal scalar or vector load */
889 vtn_assert(glsl_type_is_vector_or_scalar(type
->type
));
890 _vtn_load_store_tail(b
, op
, load
, index
, offset
,
891 access_offset
, access_size
,
893 type
->access
| access
);
895 /* This is a strided load. We have to load N things separately.
896 * This is the single column of a row-major matrix case.
898 vtn_assert(type
->stride
> type_size
);
899 vtn_assert(type
->stride
% type_size
== 0);
901 nir_ssa_def
*per_comp
[4];
902 for (unsigned i
= 0; i
< elems
; i
++) {
903 nir_ssa_def
*elem_offset
=
904 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
905 struct vtn_ssa_value
*comp
, temp_val
;
907 temp_val
.def
= nir_channel(&b
->nb
, (*inout
)->def
, i
);
908 temp_val
.type
= glsl_scalar_type(base_type
);
911 _vtn_load_store_tail(b
, op
, load
, index
, elem_offset
,
912 access_offset
, access_size
,
913 &comp
, glsl_scalar_type(base_type
),
914 type
->access
| access
);
915 per_comp
[i
] = comp
->def
;
920 *inout
= vtn_create_ssa_value(b
, type
->type
);
921 (*inout
)->def
= nir_vec(&b
->nb
, per_comp
, elems
);
927 case GLSL_TYPE_ARRAY
: {
928 unsigned elems
= glsl_get_length(type
->type
);
929 for (unsigned i
= 0; i
< elems
; i
++) {
930 nir_ssa_def
*elem_off
=
931 nir_iadd_imm(&b
->nb
, offset
, i
* type
->stride
);
932 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
933 access_offset
, access_size
,
935 type
->array_element
->access
| access
,
936 &(*inout
)->elems
[i
]);
941 case GLSL_TYPE_INTERFACE
:
942 case GLSL_TYPE_STRUCT
: {
943 unsigned elems
= glsl_get_length(type
->type
);
944 for (unsigned i
= 0; i
< elems
; i
++) {
945 nir_ssa_def
*elem_off
=
946 nir_iadd_imm(&b
->nb
, offset
, type
->offsets
[i
]);
947 _vtn_block_load_store(b
, op
, load
, index
, elem_off
,
948 access_offset
, access_size
,
950 type
->members
[i
]->access
| access
,
951 &(*inout
)->elems
[i
]);
957 vtn_fail("Invalid block member type");
961 static struct vtn_ssa_value
*
962 vtn_block_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
965 unsigned access_offset
= 0, access_size
= 0;
967 case vtn_variable_mode_ubo
:
968 op
= nir_intrinsic_load_ubo
;
970 case vtn_variable_mode_ssbo
:
971 op
= nir_intrinsic_load_ssbo
;
973 case vtn_variable_mode_push_constant
:
974 op
= nir_intrinsic_load_push_constant
;
975 access_size
= b
->shader
->num_uniforms
;
977 case vtn_variable_mode_workgroup
:
978 op
= nir_intrinsic_load_shared
;
981 vtn_fail("Invalid block variable mode");
984 nir_ssa_def
*offset
, *index
= NULL
;
985 offset
= vtn_pointer_to_offset(b
, src
, &index
);
987 struct vtn_ssa_value
*value
= NULL
;
988 _vtn_block_load_store(b
, op
, true, index
, offset
,
989 access_offset
, access_size
,
990 src
->type
, src
->access
, &value
);
995 vtn_block_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
996 struct vtn_pointer
*dst
)
1000 case vtn_variable_mode_ssbo
:
1001 op
= nir_intrinsic_store_ssbo
;
1003 case vtn_variable_mode_workgroup
:
1004 op
= nir_intrinsic_store_shared
;
1007 vtn_fail("Invalid block variable mode");
1010 nir_ssa_def
*offset
, *index
= NULL
;
1011 offset
= vtn_pointer_to_offset(b
, dst
, &index
);
1013 _vtn_block_load_store(b
, op
, false, index
, offset
,
1014 0, 0, dst
->type
, dst
->access
, &src
);
1018 _vtn_variable_load_store(struct vtn_builder
*b
, bool load
,
1019 struct vtn_pointer
*ptr
,
1020 enum gl_access_qualifier access
,
1021 struct vtn_ssa_value
**inout
)
1023 enum glsl_base_type base_type
= glsl_get_base_type(ptr
->type
->type
);
1024 switch (base_type
) {
1025 case GLSL_TYPE_UINT
:
1027 case GLSL_TYPE_UINT16
:
1028 case GLSL_TYPE_INT16
:
1029 case GLSL_TYPE_UINT8
:
1030 case GLSL_TYPE_INT8
:
1031 case GLSL_TYPE_UINT64
:
1032 case GLSL_TYPE_INT64
:
1033 case GLSL_TYPE_FLOAT
:
1034 case GLSL_TYPE_FLOAT16
:
1035 case GLSL_TYPE_BOOL
:
1036 case GLSL_TYPE_DOUBLE
:
1037 if (glsl_type_is_vector_or_scalar(ptr
->type
->type
)) {
1038 /* We hit a vector or scalar; go ahead and emit the load[s] */
1039 nir_deref_instr
*deref
= vtn_pointer_to_deref(b
, ptr
);
1040 if (vtn_pointer_is_external_block(b
, ptr
)) {
1041 /* If it's external, we call nir_load/store_deref directly. The
1042 * vtn_local_load/store helpers are too clever and do magic to
1043 * avoid array derefs of vectors. That magic is both less
1044 * efficient than the direct load/store and, in the case of
1045 * stores, is broken because it creates a race condition if two
1046 * threads are writing to different components of the same vector
1047 * due to the load+insert+store it uses to emulate the array
1051 *inout
= vtn_create_ssa_value(b
, ptr
->type
->type
);
1052 (*inout
)->def
= nir_load_deref_with_access(&b
->nb
, deref
,
1053 ptr
->type
->access
| access
);
1055 nir_store_deref_with_access(&b
->nb
, deref
, (*inout
)->def
, ~0,
1056 ptr
->type
->access
| access
);
1060 *inout
= vtn_local_load(b
, deref
, ptr
->type
->access
| access
);
1062 vtn_local_store(b
, *inout
, deref
, ptr
->type
->access
| access
);
1069 case GLSL_TYPE_INTERFACE
:
1070 case GLSL_TYPE_ARRAY
:
1071 case GLSL_TYPE_STRUCT
: {
1072 unsigned elems
= glsl_get_length(ptr
->type
->type
);
1074 vtn_assert(*inout
== NULL
);
1075 *inout
= rzalloc(b
, struct vtn_ssa_value
);
1076 (*inout
)->type
= ptr
->type
->type
;
1077 (*inout
)->elems
= rzalloc_array(b
, struct vtn_ssa_value
*, elems
);
1080 struct vtn_access_chain chain
= {
1083 { .mode
= vtn_access_mode_literal
, },
1086 for (unsigned i
= 0; i
< elems
; i
++) {
1087 chain
.link
[0].id
= i
;
1088 struct vtn_pointer
*elem
= vtn_pointer_dereference(b
, ptr
, &chain
);
1089 _vtn_variable_load_store(b
, load
, elem
, ptr
->type
->access
| access
,
1090 &(*inout
)->elems
[i
]);
1096 vtn_fail("Invalid access chain type");
1100 struct vtn_ssa_value
*
1101 vtn_variable_load(struct vtn_builder
*b
, struct vtn_pointer
*src
)
1103 if (vtn_pointer_uses_ssa_offset(b
, src
)) {
1104 return vtn_block_load(b
, src
);
1106 struct vtn_ssa_value
*val
= NULL
;
1107 _vtn_variable_load_store(b
, true, src
, src
->access
, &val
);
1113 vtn_variable_store(struct vtn_builder
*b
, struct vtn_ssa_value
*src
,
1114 struct vtn_pointer
*dest
)
1116 if (vtn_pointer_uses_ssa_offset(b
, dest
)) {
1117 vtn_assert(dest
->mode
== vtn_variable_mode_ssbo
||
1118 dest
->mode
== vtn_variable_mode_workgroup
);
1119 vtn_block_store(b
, src
, dest
);
1121 _vtn_variable_load_store(b
, false, dest
, dest
->access
, &src
);
1126 _vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1127 struct vtn_pointer
*src
)
1129 vtn_assert(src
->type
->type
== dest
->type
->type
);
1130 enum glsl_base_type base_type
= glsl_get_base_type(src
->type
->type
);
1131 switch (base_type
) {
1132 case GLSL_TYPE_UINT
:
1134 case GLSL_TYPE_UINT16
:
1135 case GLSL_TYPE_INT16
:
1136 case GLSL_TYPE_UINT8
:
1137 case GLSL_TYPE_INT8
:
1138 case GLSL_TYPE_UINT64
:
1139 case GLSL_TYPE_INT64
:
1140 case GLSL_TYPE_FLOAT
:
1141 case GLSL_TYPE_FLOAT16
:
1142 case GLSL_TYPE_DOUBLE
:
1143 case GLSL_TYPE_BOOL
:
1144 /* At this point, we have a scalar, vector, or matrix so we know that
1145 * there cannot be any structure splitting still in the way. By
1146 * stopping at the matrix level rather than the vector level, we
1147 * ensure that matrices get loaded in the optimal way even if they
1148 * are storred row-major in a UBO.
1150 vtn_variable_store(b
, vtn_variable_load(b
, src
), dest
);
1153 case GLSL_TYPE_INTERFACE
:
1154 case GLSL_TYPE_ARRAY
:
1155 case GLSL_TYPE_STRUCT
: {
1156 struct vtn_access_chain chain
= {
1159 { .mode
= vtn_access_mode_literal
, },
1162 unsigned elems
= glsl_get_length(src
->type
->type
);
1163 for (unsigned i
= 0; i
< elems
; i
++) {
1164 chain
.link
[0].id
= i
;
1165 struct vtn_pointer
*src_elem
=
1166 vtn_pointer_dereference(b
, src
, &chain
);
1167 struct vtn_pointer
*dest_elem
=
1168 vtn_pointer_dereference(b
, dest
, &chain
);
1170 _vtn_variable_copy(b
, dest_elem
, src_elem
);
1176 vtn_fail("Invalid access chain type");
1181 vtn_variable_copy(struct vtn_builder
*b
, struct vtn_pointer
*dest
,
1182 struct vtn_pointer
*src
)
1184 /* TODO: At some point, we should add a special-case for when we can
1185 * just emit a copy_var intrinsic.
1187 _vtn_variable_copy(b
, dest
, src
);
1191 set_mode_system_value(struct vtn_builder
*b
, nir_variable_mode
*mode
)
1193 vtn_assert(*mode
== nir_var_system_value
|| *mode
== nir_var_shader_in
);
1194 *mode
= nir_var_system_value
;
1198 vtn_get_builtin_location(struct vtn_builder
*b
,
1199 SpvBuiltIn builtin
, int *location
,
1200 nir_variable_mode
*mode
)
1203 case SpvBuiltInPosition
:
1204 *location
= VARYING_SLOT_POS
;
1206 case SpvBuiltInPointSize
:
1207 *location
= VARYING_SLOT_PSIZ
;
1209 case SpvBuiltInClipDistance
:
1210 *location
= VARYING_SLOT_CLIP_DIST0
; /* XXX CLIP_DIST1? */
1212 case SpvBuiltInCullDistance
:
1213 *location
= VARYING_SLOT_CULL_DIST0
;
1215 case SpvBuiltInVertexId
:
1216 case SpvBuiltInVertexIndex
:
1217 /* The Vulkan spec defines VertexIndex to be non-zero-based and doesn't
1218 * allow VertexId. The ARB_gl_spirv spec defines VertexId to be the
1219 * same as gl_VertexID, which is non-zero-based, and removes
1220 * VertexIndex. Since they're both defined to be non-zero-based, we use
1221 * SYSTEM_VALUE_VERTEX_ID for both.
1223 *location
= SYSTEM_VALUE_VERTEX_ID
;
1224 set_mode_system_value(b
, mode
);
1226 case SpvBuiltInInstanceIndex
:
1227 *location
= SYSTEM_VALUE_INSTANCE_INDEX
;
1228 set_mode_system_value(b
, mode
);
1230 case SpvBuiltInInstanceId
:
1231 *location
= SYSTEM_VALUE_INSTANCE_ID
;
1232 set_mode_system_value(b
, mode
);
1234 case SpvBuiltInPrimitiveId
:
1235 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
1236 vtn_assert(*mode
== nir_var_shader_in
);
1237 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1238 } else if (*mode
== nir_var_shader_out
) {
1239 *location
= VARYING_SLOT_PRIMITIVE_ID
;
1241 *location
= SYSTEM_VALUE_PRIMITIVE_ID
;
1242 set_mode_system_value(b
, mode
);
1245 case SpvBuiltInInvocationId
:
1246 *location
= SYSTEM_VALUE_INVOCATION_ID
;
1247 set_mode_system_value(b
, mode
);
1249 case SpvBuiltInLayer
:
1250 *location
= VARYING_SLOT_LAYER
;
1251 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1252 *mode
= nir_var_shader_in
;
1253 else 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
;
1260 vtn_fail("invalid stage for SpvBuiltInLayer");
1262 case SpvBuiltInViewportIndex
:
1263 *location
= VARYING_SLOT_VIEWPORT
;
1264 if (b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)
1265 *mode
= nir_var_shader_out
;
1266 else if (b
->options
&& b
->options
->caps
.shader_viewport_index_layer
&&
1267 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
1268 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
))
1269 *mode
= nir_var_shader_out
;
1270 else if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
)
1271 *mode
= nir_var_shader_in
;
1273 vtn_fail("invalid stage for SpvBuiltInViewportIndex");
1275 case SpvBuiltInTessLevelOuter
:
1276 *location
= VARYING_SLOT_TESS_LEVEL_OUTER
;
1278 case SpvBuiltInTessLevelInner
:
1279 *location
= VARYING_SLOT_TESS_LEVEL_INNER
;
1281 case SpvBuiltInTessCoord
:
1282 *location
= SYSTEM_VALUE_TESS_COORD
;
1283 set_mode_system_value(b
, mode
);
1285 case SpvBuiltInPatchVertices
:
1286 *location
= SYSTEM_VALUE_VERTICES_IN
;
1287 set_mode_system_value(b
, mode
);
1289 case SpvBuiltInFragCoord
:
1290 *location
= VARYING_SLOT_POS
;
1291 vtn_assert(*mode
== nir_var_shader_in
);
1293 case SpvBuiltInPointCoord
:
1294 *location
= VARYING_SLOT_PNTC
;
1295 vtn_assert(*mode
== nir_var_shader_in
);
1297 case SpvBuiltInFrontFacing
:
1298 *location
= SYSTEM_VALUE_FRONT_FACE
;
1299 set_mode_system_value(b
, mode
);
1301 case SpvBuiltInSampleId
:
1302 *location
= SYSTEM_VALUE_SAMPLE_ID
;
1303 set_mode_system_value(b
, mode
);
1305 case SpvBuiltInSamplePosition
:
1306 *location
= SYSTEM_VALUE_SAMPLE_POS
;
1307 set_mode_system_value(b
, mode
);
1309 case SpvBuiltInSampleMask
:
1310 if (*mode
== nir_var_shader_out
) {
1311 *location
= FRAG_RESULT_SAMPLE_MASK
;
1313 *location
= SYSTEM_VALUE_SAMPLE_MASK_IN
;
1314 set_mode_system_value(b
, mode
);
1317 case SpvBuiltInFragDepth
:
1318 *location
= FRAG_RESULT_DEPTH
;
1319 vtn_assert(*mode
== nir_var_shader_out
);
1321 case SpvBuiltInHelperInvocation
:
1322 *location
= SYSTEM_VALUE_HELPER_INVOCATION
;
1323 set_mode_system_value(b
, mode
);
1325 case SpvBuiltInNumWorkgroups
:
1326 *location
= SYSTEM_VALUE_NUM_WORK_GROUPS
;
1327 set_mode_system_value(b
, mode
);
1329 case SpvBuiltInWorkgroupSize
:
1330 *location
= SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
1331 set_mode_system_value(b
, mode
);
1333 case SpvBuiltInWorkgroupId
:
1334 *location
= SYSTEM_VALUE_WORK_GROUP_ID
;
1335 set_mode_system_value(b
, mode
);
1337 case SpvBuiltInLocalInvocationId
:
1338 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
1339 set_mode_system_value(b
, mode
);
1341 case SpvBuiltInLocalInvocationIndex
:
1342 *location
= SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
1343 set_mode_system_value(b
, mode
);
1345 case SpvBuiltInGlobalInvocationId
:
1346 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
1347 set_mode_system_value(b
, mode
);
1349 case SpvBuiltInGlobalLinearId
:
1350 *location
= SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX
;
1351 set_mode_system_value(b
, mode
);
1353 case SpvBuiltInBaseVertex
:
1354 /* OpenGL gl_BaseVertex (SYSTEM_VALUE_BASE_VERTEX) is not the same
1355 * semantic as SPIR-V BaseVertex (SYSTEM_VALUE_FIRST_VERTEX).
1357 *location
= SYSTEM_VALUE_FIRST_VERTEX
;
1358 set_mode_system_value(b
, mode
);
1360 case SpvBuiltInBaseInstance
:
1361 *location
= SYSTEM_VALUE_BASE_INSTANCE
;
1362 set_mode_system_value(b
, mode
);
1364 case SpvBuiltInDrawIndex
:
1365 *location
= SYSTEM_VALUE_DRAW_ID
;
1366 set_mode_system_value(b
, mode
);
1368 case SpvBuiltInSubgroupSize
:
1369 *location
= SYSTEM_VALUE_SUBGROUP_SIZE
;
1370 set_mode_system_value(b
, mode
);
1372 case SpvBuiltInSubgroupId
:
1373 *location
= SYSTEM_VALUE_SUBGROUP_ID
;
1374 set_mode_system_value(b
, mode
);
1376 case SpvBuiltInSubgroupLocalInvocationId
:
1377 *location
= SYSTEM_VALUE_SUBGROUP_INVOCATION
;
1378 set_mode_system_value(b
, mode
);
1380 case SpvBuiltInNumSubgroups
:
1381 *location
= SYSTEM_VALUE_NUM_SUBGROUPS
;
1382 set_mode_system_value(b
, mode
);
1384 case SpvBuiltInDeviceIndex
:
1385 *location
= SYSTEM_VALUE_DEVICE_INDEX
;
1386 set_mode_system_value(b
, mode
);
1388 case SpvBuiltInViewIndex
:
1389 *location
= SYSTEM_VALUE_VIEW_INDEX
;
1390 set_mode_system_value(b
, mode
);
1392 case SpvBuiltInSubgroupEqMask
:
1393 *location
= SYSTEM_VALUE_SUBGROUP_EQ_MASK
,
1394 set_mode_system_value(b
, mode
);
1396 case SpvBuiltInSubgroupGeMask
:
1397 *location
= SYSTEM_VALUE_SUBGROUP_GE_MASK
,
1398 set_mode_system_value(b
, mode
);
1400 case SpvBuiltInSubgroupGtMask
:
1401 *location
= SYSTEM_VALUE_SUBGROUP_GT_MASK
,
1402 set_mode_system_value(b
, mode
);
1404 case SpvBuiltInSubgroupLeMask
:
1405 *location
= SYSTEM_VALUE_SUBGROUP_LE_MASK
,
1406 set_mode_system_value(b
, mode
);
1408 case SpvBuiltInSubgroupLtMask
:
1409 *location
= SYSTEM_VALUE_SUBGROUP_LT_MASK
,
1410 set_mode_system_value(b
, mode
);
1412 case SpvBuiltInFragStencilRefEXT
:
1413 *location
= FRAG_RESULT_STENCIL
;
1414 vtn_assert(*mode
== nir_var_shader_out
);
1416 case SpvBuiltInWorkDim
:
1417 *location
= SYSTEM_VALUE_WORK_DIM
;
1418 set_mode_system_value(b
, mode
);
1420 case SpvBuiltInGlobalSize
:
1421 *location
= SYSTEM_VALUE_GLOBAL_GROUP_SIZE
;
1422 set_mode_system_value(b
, mode
);
1425 vtn_fail("Unsupported builtin: %s (%u)",
1426 spirv_builtin_to_string(builtin
), builtin
);
1431 apply_var_decoration(struct vtn_builder
*b
,
1432 struct nir_variable_data
*var_data
,
1433 const struct vtn_decoration
*dec
)
1435 switch (dec
->decoration
) {
1436 case SpvDecorationRelaxedPrecision
:
1437 break; /* FIXME: Do nothing with this for now. */
1438 case SpvDecorationNoPerspective
:
1439 var_data
->interpolation
= INTERP_MODE_NOPERSPECTIVE
;
1441 case SpvDecorationFlat
:
1442 var_data
->interpolation
= INTERP_MODE_FLAT
;
1444 case SpvDecorationCentroid
:
1445 var_data
->centroid
= true;
1447 case SpvDecorationSample
:
1448 var_data
->sample
= true;
1450 case SpvDecorationInvariant
:
1451 var_data
->invariant
= true;
1453 case SpvDecorationConstant
:
1454 var_data
->read_only
= true;
1456 case SpvDecorationNonReadable
:
1457 var_data
->image
.access
|= ACCESS_NON_READABLE
;
1459 case SpvDecorationNonWritable
:
1460 var_data
->read_only
= true;
1461 var_data
->image
.access
|= ACCESS_NON_WRITEABLE
;
1463 case SpvDecorationRestrict
:
1464 var_data
->image
.access
|= ACCESS_RESTRICT
;
1466 case SpvDecorationVolatile
:
1467 var_data
->image
.access
|= ACCESS_VOLATILE
;
1469 case SpvDecorationCoherent
:
1470 var_data
->image
.access
|= ACCESS_COHERENT
;
1472 case SpvDecorationComponent
:
1473 var_data
->location_frac
= dec
->literals
[0];
1475 case SpvDecorationIndex
:
1476 var_data
->index
= dec
->literals
[0];
1478 case SpvDecorationBuiltIn
: {
1479 SpvBuiltIn builtin
= dec
->literals
[0];
1481 nir_variable_mode mode
= var_data
->mode
;
1482 vtn_get_builtin_location(b
, builtin
, &var_data
->location
, &mode
);
1483 var_data
->mode
= mode
;
1486 case SpvBuiltInTessLevelOuter
:
1487 case SpvBuiltInTessLevelInner
:
1488 case SpvBuiltInClipDistance
:
1489 case SpvBuiltInCullDistance
:
1490 var_data
->compact
= true;
1497 case SpvDecorationSpecId
:
1498 case SpvDecorationRowMajor
:
1499 case SpvDecorationColMajor
:
1500 case SpvDecorationMatrixStride
:
1501 case SpvDecorationAliased
:
1502 case SpvDecorationUniform
:
1503 case SpvDecorationLinkageAttributes
:
1504 break; /* Do nothing with these here */
1506 case SpvDecorationPatch
:
1507 var_data
->patch
= true;
1510 case SpvDecorationLocation
:
1511 vtn_fail("Handled above");
1513 case SpvDecorationBlock
:
1514 case SpvDecorationBufferBlock
:
1515 case SpvDecorationArrayStride
:
1516 case SpvDecorationGLSLShared
:
1517 case SpvDecorationGLSLPacked
:
1518 break; /* These can apply to a type but we don't care about them */
1520 case SpvDecorationBinding
:
1521 case SpvDecorationDescriptorSet
:
1522 case SpvDecorationNoContraction
:
1523 case SpvDecorationInputAttachmentIndex
:
1524 vtn_warn("Decoration not allowed for variable or structure member: %s",
1525 spirv_decoration_to_string(dec
->decoration
));
1528 case SpvDecorationXfbBuffer
:
1529 var_data
->explicit_xfb_buffer
= true;
1530 var_data
->xfb_buffer
= dec
->literals
[0];
1531 var_data
->always_active_io
= true;
1533 case SpvDecorationXfbStride
:
1534 var_data
->explicit_xfb_stride
= true;
1535 var_data
->xfb_stride
= dec
->literals
[0];
1537 case SpvDecorationOffset
:
1538 var_data
->explicit_offset
= true;
1539 var_data
->offset
= dec
->literals
[0];
1542 case SpvDecorationStream
:
1543 var_data
->stream
= dec
->literals
[0];
1546 case SpvDecorationCPacked
:
1547 case SpvDecorationSaturatedConversion
:
1548 case SpvDecorationFuncParamAttr
:
1549 case SpvDecorationFPRoundingMode
:
1550 case SpvDecorationFPFastMathMode
:
1551 case SpvDecorationAlignment
:
1552 if (b
->shader
->info
.stage
!= MESA_SHADER_KERNEL
) {
1553 vtn_warn("Decoration only allowed for CL-style kernels: %s",
1554 spirv_decoration_to_string(dec
->decoration
));
1558 case SpvDecorationHlslSemanticGOOGLE
:
1559 /* HLSL semantic decorations can safely be ignored by the driver. */
1562 case SpvDecorationRestrictPointerEXT
:
1563 case SpvDecorationAliasedPointerEXT
:
1564 /* TODO: We should actually plumb alias information through NIR. */
1568 vtn_fail_with_decoration("Unhandled decoration", dec
->decoration
);
1573 var_is_patch_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1574 const struct vtn_decoration
*dec
, void *out_is_patch
)
1576 if (dec
->decoration
== SpvDecorationPatch
) {
1577 *((bool *) out_is_patch
) = true;
1582 var_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1583 const struct vtn_decoration
*dec
, void *void_var
)
1585 struct vtn_variable
*vtn_var
= void_var
;
1587 /* Handle decorations that apply to a vtn_variable as a whole */
1588 switch (dec
->decoration
) {
1589 case SpvDecorationBinding
:
1590 vtn_var
->binding
= dec
->literals
[0];
1591 vtn_var
->explicit_binding
= true;
1593 case SpvDecorationDescriptorSet
:
1594 vtn_var
->descriptor_set
= dec
->literals
[0];
1596 case SpvDecorationInputAttachmentIndex
:
1597 vtn_var
->input_attachment_index
= dec
->literals
[0];
1599 case SpvDecorationPatch
:
1600 vtn_var
->patch
= true;
1602 case SpvDecorationOffset
:
1603 vtn_var
->offset
= dec
->literals
[0];
1605 case SpvDecorationNonWritable
:
1606 vtn_var
->access
|= ACCESS_NON_WRITEABLE
;
1608 case SpvDecorationNonReadable
:
1609 vtn_var
->access
|= ACCESS_NON_READABLE
;
1611 case SpvDecorationVolatile
:
1612 vtn_var
->access
|= ACCESS_VOLATILE
;
1614 case SpvDecorationCoherent
:
1615 vtn_var
->access
|= ACCESS_COHERENT
;
1617 case SpvDecorationHlslCounterBufferGOOGLE
:
1618 /* HLSL semantic decorations can safely be ignored by the driver. */
1624 if (val
->value_type
== vtn_value_type_pointer
) {
1625 assert(val
->pointer
->var
== void_var
);
1626 assert(member
== -1);
1628 assert(val
->value_type
== vtn_value_type_type
);
1631 /* Location is odd. If applied to a split structure, we have to walk the
1632 * whole thing and accumulate the location. It's easier to handle as a
1635 if (dec
->decoration
== SpvDecorationLocation
) {
1636 unsigned location
= dec
->literals
[0];
1637 if (b
->shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
1638 vtn_var
->mode
== vtn_variable_mode_output
) {
1639 location
+= FRAG_RESULT_DATA0
;
1640 } else if (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
&&
1641 vtn_var
->mode
== vtn_variable_mode_input
) {
1642 location
+= VERT_ATTRIB_GENERIC0
;
1643 } else if (vtn_var
->mode
== vtn_variable_mode_input
||
1644 vtn_var
->mode
== vtn_variable_mode_output
) {
1645 location
+= vtn_var
->patch
? VARYING_SLOT_PATCH0
: VARYING_SLOT_VAR0
;
1646 } else if (vtn_var
->mode
!= vtn_variable_mode_uniform
) {
1647 vtn_warn("Location must be on input, output, uniform, sampler or "
1652 if (vtn_var
->var
->num_members
== 0) {
1653 /* This handles the member and lone variable cases */
1654 vtn_var
->var
->data
.location
= location
;
1656 /* This handles the structure member case */
1657 assert(vtn_var
->var
->members
);
1660 vtn_var
->base_location
= location
;
1662 vtn_var
->var
->members
[member
].location
= location
;
1668 if (vtn_var
->var
->num_members
== 0) {
1669 /* We call this function on types as well as variables and not all
1670 * struct types get split so we can end up having stray member
1671 * decorations; just ignore them.
1674 apply_var_decoration(b
, &vtn_var
->var
->data
, dec
);
1675 } else if (member
>= 0) {
1676 /* Member decorations must come from a type */
1677 assert(val
->value_type
== vtn_value_type_type
);
1678 apply_var_decoration(b
, &vtn_var
->var
->members
[member
], dec
);
1681 glsl_get_length(glsl_without_array(vtn_var
->type
->type
));
1682 for (unsigned i
= 0; i
< length
; i
++)
1683 apply_var_decoration(b
, &vtn_var
->var
->members
[i
], dec
);
1686 /* A few variables, those with external storage, have no actual
1687 * nir_variables associated with them. Fortunately, all decorations
1688 * we care about for those variables are on the type only.
1690 vtn_assert(vtn_var
->mode
== vtn_variable_mode_ubo
||
1691 vtn_var
->mode
== vtn_variable_mode_ssbo
||
1692 vtn_var
->mode
== vtn_variable_mode_push_constant
||
1693 (vtn_var
->mode
== vtn_variable_mode_workgroup
&&
1694 b
->options
->lower_workgroup_access_to_offsets
));
1700 ptr_decoration_cb(struct vtn_builder
*b
, struct vtn_value
*val
, int member
,
1701 const struct vtn_decoration
*dec
, void *void_ptr
)
1703 struct vtn_pointer
*ptr
= void_ptr
;
1705 switch (dec
->decoration
) {
1706 case SpvDecorationNonUniformEXT
:
1707 ptr
->access
|= ACCESS_NON_UNIFORM
;
1715 static enum vtn_variable_mode
1716 vtn_storage_class_to_mode(struct vtn_builder
*b
,
1717 SpvStorageClass
class,
1718 struct vtn_type
*interface_type
,
1719 nir_variable_mode
*nir_mode_out
)
1721 enum vtn_variable_mode mode
;
1722 nir_variable_mode nir_mode
;
1724 case SpvStorageClassUniform
:
1725 if (interface_type
->block
) {
1726 mode
= vtn_variable_mode_ubo
;
1727 nir_mode
= nir_var_mem_ubo
;
1728 } else if (interface_type
->buffer_block
) {
1729 mode
= vtn_variable_mode_ssbo
;
1730 nir_mode
= nir_var_mem_ssbo
;
1732 /* Default-block uniforms, coming from gl_spirv */
1733 mode
= vtn_variable_mode_uniform
;
1734 nir_mode
= nir_var_uniform
;
1737 case SpvStorageClassStorageBuffer
:
1738 mode
= vtn_variable_mode_ssbo
;
1739 nir_mode
= nir_var_mem_ssbo
;
1741 case SpvStorageClassPhysicalStorageBufferEXT
:
1742 mode
= vtn_variable_mode_phys_ssbo
;
1743 nir_mode
= nir_var_mem_global
;
1745 case SpvStorageClassUniformConstant
:
1746 mode
= vtn_variable_mode_uniform
;
1747 nir_mode
= nir_var_uniform
;
1749 case SpvStorageClassPushConstant
:
1750 mode
= vtn_variable_mode_push_constant
;
1751 nir_mode
= nir_var_uniform
;
1753 case SpvStorageClassInput
:
1754 mode
= vtn_variable_mode_input
;
1755 nir_mode
= nir_var_shader_in
;
1757 case SpvStorageClassOutput
:
1758 mode
= vtn_variable_mode_output
;
1759 nir_mode
= nir_var_shader_out
;
1761 case SpvStorageClassPrivate
:
1762 mode
= vtn_variable_mode_private
;
1763 nir_mode
= nir_var_shader_temp
;
1765 case SpvStorageClassFunction
:
1766 mode
= vtn_variable_mode_function
;
1767 nir_mode
= nir_var_function_temp
;
1769 case SpvStorageClassWorkgroup
:
1770 mode
= vtn_variable_mode_workgroup
;
1771 nir_mode
= nir_var_mem_shared
;
1773 case SpvStorageClassAtomicCounter
:
1774 mode
= vtn_variable_mode_uniform
;
1775 nir_mode
= nir_var_uniform
;
1777 case SpvStorageClassCrossWorkgroup
:
1778 mode
= vtn_variable_mode_cross_workgroup
;
1779 nir_mode
= nir_var_mem_global
;
1781 case SpvStorageClassGeneric
:
1783 vtn_fail("Unhandled variable storage class: %s (%u)",
1784 spirv_storageclass_to_string(class), class);
1788 *nir_mode_out
= nir_mode
;
1794 vtn_pointer_to_ssa(struct vtn_builder
*b
, struct vtn_pointer
*ptr
)
1796 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1797 /* This pointer needs to have a pointer type with actual storage */
1798 vtn_assert(ptr
->ptr_type
);
1799 vtn_assert(ptr
->ptr_type
->type
);
1802 /* If we don't have an offset then we must be a pointer to the variable
1805 vtn_assert(!ptr
->offset
&& !ptr
->block_index
);
1807 struct vtn_access_chain chain
= {
1810 ptr
= vtn_ssa_offset_pointer_dereference(b
, ptr
, &chain
);
1813 vtn_assert(ptr
->offset
);
1814 if (ptr
->block_index
) {
1815 vtn_assert(ptr
->mode
== vtn_variable_mode_ubo
||
1816 ptr
->mode
== vtn_variable_mode_ssbo
);
1817 return nir_vec2(&b
->nb
, ptr
->block_index
, ptr
->offset
);
1819 vtn_assert(ptr
->mode
== vtn_variable_mode_workgroup
);
1823 if (vtn_pointer_is_external_block(b
, ptr
) &&
1824 vtn_type_contains_block(b
, ptr
->type
) &&
1825 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
1826 /* In this case, we're looking for a block index and not an actual
1829 * For PhysicalStorageBufferEXT pointers, we don't have a block index
1830 * at all because we get the pointer directly from the client. This
1831 * assumes that there will never be a SSBO binding variable using the
1832 * PhysicalStorageBufferEXT storage class. This assumption appears
1833 * to be correct according to the Vulkan spec because the table,
1834 * "Shader Resource and Storage Class Correspondence," the only the
1835 * Uniform storage class with BufferBlock or the StorageBuffer
1836 * storage class with Block can be used.
1838 if (!ptr
->block_index
) {
1839 /* If we don't have a block_index then we must be a pointer to the
1842 vtn_assert(!ptr
->deref
);
1844 struct vtn_access_chain chain
= {
1847 ptr
= vtn_nir_deref_pointer_dereference(b
, ptr
, &chain
);
1850 return ptr
->block_index
;
1852 return &vtn_pointer_to_deref(b
, ptr
)->dest
.ssa
;
1857 struct vtn_pointer
*
1858 vtn_pointer_from_ssa(struct vtn_builder
*b
, nir_ssa_def
*ssa
,
1859 struct vtn_type
*ptr_type
)
1861 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
1863 struct vtn_type
*interface_type
= ptr_type
->deref
;
1864 while (interface_type
->base_type
== vtn_base_type_array
)
1865 interface_type
= interface_type
->array_element
;
1867 struct vtn_pointer
*ptr
= rzalloc(b
, struct vtn_pointer
);
1868 nir_variable_mode nir_mode
;
1869 ptr
->mode
= vtn_storage_class_to_mode(b
, ptr_type
->storage_class
,
1870 interface_type
, &nir_mode
);
1871 ptr
->type
= ptr_type
->deref
;
1872 ptr
->ptr_type
= ptr_type
;
1874 if (b
->wa_glslang_179
) {
1875 /* To work around https://github.com/KhronosGroup/glslang/issues/179 we
1876 * need to whack the mode because it creates a function parameter with
1877 * the Function storage class even though it's a pointer to a sampler.
1878 * If we don't do this, then NIR won't get rid of the deref_cast for us.
1880 if (ptr
->mode
== vtn_variable_mode_function
&&
1881 (ptr
->type
->base_type
== vtn_base_type_sampler
||
1882 ptr
->type
->base_type
== vtn_base_type_sampled_image
)) {
1883 ptr
->mode
= vtn_variable_mode_uniform
;
1884 nir_mode
= nir_var_uniform
;
1888 if (vtn_pointer_uses_ssa_offset(b
, ptr
)) {
1889 /* This pointer type needs to have actual storage */
1890 vtn_assert(ptr_type
->type
);
1891 if (ptr
->mode
== vtn_variable_mode_ubo
||
1892 ptr
->mode
== vtn_variable_mode_ssbo
) {
1893 vtn_assert(ssa
->num_components
== 2);
1894 ptr
->block_index
= nir_channel(&b
->nb
, ssa
, 0);
1895 ptr
->offset
= nir_channel(&b
->nb
, ssa
, 1);
1897 vtn_assert(ssa
->num_components
== 1);
1898 ptr
->block_index
= NULL
;
1902 const struct glsl_type
*deref_type
= ptr_type
->deref
->type
;
1903 if (!vtn_pointer_is_external_block(b
, ptr
)) {
1904 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
1906 } else if (vtn_type_contains_block(b
, ptr
->type
) &&
1907 ptr
->mode
!= vtn_variable_mode_phys_ssbo
) {
1908 /* This is a pointer to somewhere in an array of blocks, not a
1909 * pointer to somewhere inside the block. Set the block index
1910 * instead of making a cast.
1912 ptr
->block_index
= ssa
;
1914 /* This is a pointer to something internal or a pointer inside a
1915 * block. It's just a regular cast.
1917 * For PhysicalStorageBufferEXT pointers, we don't have a block index
1918 * at all because we get the pointer directly from the client. This
1919 * assumes that there will never be a SSBO binding variable using the
1920 * PhysicalStorageBufferEXT storage class. This assumption appears
1921 * to be correct according to the Vulkan spec because the table,
1922 * "Shader Resource and Storage Class Correspondence," the only the
1923 * Uniform storage class with BufferBlock or the StorageBuffer
1924 * storage class with Block can be used.
1926 ptr
->deref
= nir_build_deref_cast(&b
->nb
, ssa
, nir_mode
,
1927 ptr_type
->deref
->type
,
1929 ptr
->deref
->dest
.ssa
.num_components
=
1930 glsl_get_vector_elements(ptr_type
->type
);
1931 ptr
->deref
->dest
.ssa
.bit_size
= glsl_get_bit_size(ptr_type
->type
);
1939 is_per_vertex_inout(const struct vtn_variable
*var
, gl_shader_stage stage
)
1941 if (var
->patch
|| !glsl_type_is_array(var
->type
->type
))
1944 if (var
->mode
== vtn_variable_mode_input
) {
1945 return stage
== MESA_SHADER_TESS_CTRL
||
1946 stage
== MESA_SHADER_TESS_EVAL
||
1947 stage
== MESA_SHADER_GEOMETRY
;
1950 if (var
->mode
== vtn_variable_mode_output
)
1951 return stage
== MESA_SHADER_TESS_CTRL
;
1957 assign_missing_member_locations(struct vtn_variable
*var
)
1960 glsl_get_length(glsl_without_array(var
->type
->type
));
1961 int location
= var
->base_location
;
1963 for (unsigned i
= 0; i
< length
; i
++) {
1964 /* From the Vulkan spec:
1966 * “If the structure type is a Block but without a Location, then each
1967 * of its members must have a Location decoration.”
1970 if (var
->type
->block
) {
1971 assert(var
->base_location
!= -1 ||
1972 var
->var
->members
[i
].location
!= -1);
1975 /* From the Vulkan spec:
1977 * “Any member with its own Location decoration is assigned that
1978 * location. Each remaining member is assigned the location after the
1979 * immediately preceding member in declaration order.”
1981 if (var
->var
->members
[i
].location
!= -1)
1982 location
= var
->var
->members
[i
].location
;
1984 var
->var
->members
[i
].location
= location
;
1986 /* Below we use type instead of interface_type, because interface_type
1987 * is only available when it is a Block. This code also supports
1988 * input/outputs that are just structs
1990 const struct glsl_type
*member_type
=
1991 glsl_get_struct_field(glsl_without_array(var
->type
->type
), i
);
1994 glsl_count_attribute_slots(member_type
,
1995 false /* is_gl_vertex_input */);
2001 vtn_create_variable(struct vtn_builder
*b
, struct vtn_value
*val
,
2002 struct vtn_type
*ptr_type
, SpvStorageClass storage_class
,
2003 nir_constant
*initializer
)
2005 vtn_assert(ptr_type
->base_type
== vtn_base_type_pointer
);
2006 struct vtn_type
*type
= ptr_type
->deref
;
2008 struct vtn_type
*without_array
= type
;
2009 while(glsl_type_is_array(without_array
->type
))
2010 without_array
= without_array
->array_element
;
2012 enum vtn_variable_mode mode
;
2013 nir_variable_mode nir_mode
;
2014 mode
= vtn_storage_class_to_mode(b
, storage_class
, without_array
, &nir_mode
);
2017 case vtn_variable_mode_ubo
:
2018 /* There's no other way to get vtn_variable_mode_ubo */
2019 vtn_assert(without_array
->block
);
2020 b
->shader
->info
.num_ubos
++;
2022 case vtn_variable_mode_ssbo
:
2023 if (storage_class
== SpvStorageClassStorageBuffer
&&
2024 !without_array
->block
) {
2025 if (b
->variable_pointers
) {
2026 vtn_fail("Variables in the StorageBuffer storage class must "
2027 "have a struct type with the Block decoration");
2029 /* If variable pointers are not present, it's still malformed
2030 * SPIR-V but we can parse it and do the right thing anyway.
2031 * Since some of the 8-bit storage tests have bugs in this are,
2032 * just make it a warning for now.
2034 vtn_warn("Variables in the StorageBuffer storage class must "
2035 "have a struct type with the Block decoration");
2038 b
->shader
->info
.num_ssbos
++;
2040 case vtn_variable_mode_uniform
:
2041 if (glsl_type_is_image(without_array
->type
))
2042 b
->shader
->info
.num_images
++;
2043 else if (glsl_type_is_sampler(without_array
->type
))
2044 b
->shader
->info
.num_textures
++;
2046 case vtn_variable_mode_push_constant
:
2047 b
->shader
->num_uniforms
= vtn_type_block_size(b
, type
);
2050 case vtn_variable_mode_phys_ssbo
:
2051 vtn_fail("Cannot create a variable with the "
2052 "PhysicalStorageBufferEXT storage class");
2056 /* No tallying is needed */
2060 struct vtn_variable
*var
= rzalloc(b
, struct vtn_variable
);
2063 var
->base_location
= -1;
2065 vtn_assert(val
->value_type
== vtn_value_type_pointer
);
2066 val
->pointer
= vtn_pointer_for_variable(b
, var
, ptr_type
);
2068 switch (var
->mode
) {
2069 case vtn_variable_mode_function
:
2070 case vtn_variable_mode_private
:
2071 case vtn_variable_mode_uniform
:
2072 /* For these, we create the variable normally */
2073 var
->var
= rzalloc(b
->shader
, nir_variable
);
2074 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2076 if (storage_class
== SpvStorageClassAtomicCounter
) {
2077 /* Need to tweak the nir type here as at vtn_handle_type we don't
2078 * have the access to storage_class, that is the one that points us
2079 * that is an atomic uint.
2081 var
->var
->type
= repair_atomic_type(var
->type
->type
);
2083 /* Private variables don't have any explicit layout but some layouts
2084 * may have leaked through due to type deduplication in the SPIR-V.
2086 var
->var
->type
= var
->type
->type
;
2088 var
->var
->data
.mode
= nir_mode
;
2089 var
->var
->data
.location
= -1;
2090 var
->var
->interface_type
= NULL
;
2093 case vtn_variable_mode_workgroup
:
2094 if (b
->options
->lower_workgroup_access_to_offsets
) {
2095 var
->shared_location
= -1;
2097 /* Create the variable normally */
2098 var
->var
= rzalloc(b
->shader
, nir_variable
);
2099 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2100 /* Workgroup variables don't have any explicit layout but some
2101 * layouts may have leaked through due to type deduplication in the
2104 var
->var
->type
= var
->type
->type
;
2105 var
->var
->data
.mode
= nir_var_mem_shared
;
2109 case vtn_variable_mode_input
:
2110 case vtn_variable_mode_output
: {
2111 /* In order to know whether or not we're a per-vertex inout, we need
2112 * the patch qualifier. This means walking the variable decorations
2113 * early before we actually create any variables. Not a big deal.
2115 * GLSLang really likes to place decorations in the most interior
2116 * thing it possibly can. In particular, if you have a struct, it
2117 * will place the patch decorations on the struct members. This
2118 * should be handled by the variable splitting below just fine.
2120 * If you have an array-of-struct, things get even more weird as it
2121 * will place the patch decorations on the struct even though it's
2122 * inside an array and some of the members being patch and others not
2123 * makes no sense whatsoever. Since the only sensible thing is for
2124 * it to be all or nothing, we'll call it patch if any of the members
2125 * are declared patch.
2128 vtn_foreach_decoration(b
, val
, var_is_patch_cb
, &var
->patch
);
2129 if (glsl_type_is_array(var
->type
->type
) &&
2130 glsl_type_is_struct_or_ifc(without_array
->type
)) {
2131 vtn_foreach_decoration(b
, vtn_value(b
, without_array
->id
,
2132 vtn_value_type_type
),
2133 var_is_patch_cb
, &var
->patch
);
2136 /* For inputs and outputs, we immediately split structures. This
2137 * is for a couple of reasons. For one, builtins may all come in
2138 * a struct and we really want those split out into separate
2139 * variables. For another, interpolation qualifiers can be
2140 * applied to members of the top-level struct ane we need to be
2141 * able to preserve that information.
2144 struct vtn_type
*per_vertex_type
= var
->type
;
2145 if (is_per_vertex_inout(var
, b
->shader
->info
.stage
)) {
2146 /* In Geometry shaders (and some tessellation), inputs come
2147 * in per-vertex arrays. However, some builtins come in
2148 * non-per-vertex, hence the need for the is_array check. In
2149 * any case, there are no non-builtin arrays allowed so this
2150 * check should be sufficient.
2152 per_vertex_type
= var
->type
->array_element
;
2155 var
->var
= rzalloc(b
->shader
, nir_variable
);
2156 var
->var
->name
= ralloc_strdup(var
->var
, val
->name
);
2157 /* In Vulkan, shader I/O variables don't have any explicit layout but
2158 * some layouts may have leaked through due to type deduplication in
2159 * the SPIR-V. We do, however, keep the layouts in the variable's
2160 * interface_type because we need offsets for XFB arrays of blocks.
2162 var
->var
->type
= var
->type
->type
;
2163 var
->var
->data
.mode
= nir_mode
;
2164 var
->var
->data
.patch
= var
->patch
;
2166 /* Figure out the interface block type. */
2167 struct vtn_type
*iface_type
= per_vertex_type
;
2168 if (var
->mode
== vtn_variable_mode_output
&&
2169 (b
->shader
->info
.stage
== MESA_SHADER_VERTEX
||
2170 b
->shader
->info
.stage
== MESA_SHADER_TESS_EVAL
||
2171 b
->shader
->info
.stage
== MESA_SHADER_GEOMETRY
)) {
2172 /* For vertex data outputs, we can end up with arrays of blocks for
2173 * transform feedback where each array element corresponds to a
2174 * different XFB output buffer.
2176 while (iface_type
->base_type
== vtn_base_type_array
)
2177 iface_type
= iface_type
->array_element
;
2179 if (iface_type
->base_type
== vtn_base_type_struct
&& iface_type
->block
)
2180 var
->var
->interface_type
= iface_type
->type
;
2182 if (per_vertex_type
->base_type
== vtn_base_type_struct
&&
2183 per_vertex_type
->block
) {
2184 /* It's a struct. Set it up as per-member. */
2185 var
->var
->num_members
= glsl_get_length(per_vertex_type
->type
);
2186 var
->var
->members
= rzalloc_array(var
->var
, struct nir_variable_data
,
2187 var
->var
->num_members
);
2189 for (unsigned i
= 0; i
< var
->var
->num_members
; i
++) {
2190 var
->var
->members
[i
].mode
= nir_mode
;
2191 var
->var
->members
[i
].patch
= var
->patch
;
2192 var
->var
->members
[i
].location
= -1;
2196 /* For inputs and outputs, we need to grab locations and builtin
2197 * information from the per-vertex type.
2199 vtn_foreach_decoration(b
, vtn_value(b
, per_vertex_type
->id
,
2200 vtn_value_type_type
),
2201 var_decoration_cb
, var
);
2205 case vtn_variable_mode_ubo
:
2206 case vtn_variable_mode_ssbo
:
2207 case vtn_variable_mode_push_constant
:
2208 case vtn_variable_mode_cross_workgroup
:
2209 /* These don't need actual variables. */
2212 case vtn_variable_mode_phys_ssbo
:
2213 unreachable("Should have been caught before");
2217 var
->var
->constant_initializer
=
2218 nir_constant_clone(initializer
, var
->var
);
2221 vtn_foreach_decoration(b
, val
, var_decoration_cb
, var
);
2222 vtn_foreach_decoration(b
, val
, ptr_decoration_cb
, val
->pointer
);
2224 if ((var
->mode
== vtn_variable_mode_input
||
2225 var
->mode
== vtn_variable_mode_output
) &&
2226 var
->var
->members
) {
2227 assign_missing_member_locations(var
);
2230 if (var
->mode
== vtn_variable_mode_uniform
) {
2231 /* XXX: We still need the binding information in the nir_variable
2232 * for these. We should fix that.
2234 var
->var
->data
.binding
= var
->binding
;
2235 var
->var
->data
.explicit_binding
= var
->explicit_binding
;
2236 var
->var
->data
.descriptor_set
= var
->descriptor_set
;
2237 var
->var
->data
.index
= var
->input_attachment_index
;
2238 var
->var
->data
.offset
= var
->offset
;
2240 if (glsl_type_is_image(without_array
->type
))
2241 var
->var
->data
.image
.format
= without_array
->image_format
;
2244 if (var
->mode
== vtn_variable_mode_function
) {
2245 vtn_assert(var
->var
!= NULL
&& var
->var
->members
== NULL
);
2246 nir_function_impl_add_variable(b
->nb
.impl
, var
->var
);
2247 } else if (var
->var
) {
2248 nir_shader_add_variable(b
->shader
, var
->var
);
2250 vtn_assert(vtn_pointer_is_external_block(b
, val
->pointer
));
2255 vtn_assert_types_equal(struct vtn_builder
*b
, SpvOp opcode
,
2256 struct vtn_type
*dst_type
,
2257 struct vtn_type
*src_type
)
2259 if (dst_type
->id
== src_type
->id
)
2262 if (vtn_types_compatible(b
, dst_type
, src_type
)) {
2263 /* Early versions of GLSLang would re-emit types unnecessarily and you
2264 * would end up with OpLoad, OpStore, or OpCopyMemory opcodes which have
2265 * mismatched source and destination types.
2267 * https://github.com/KhronosGroup/glslang/issues/304
2268 * https://github.com/KhronosGroup/glslang/issues/307
2269 * https://bugs.freedesktop.org/show_bug.cgi?id=104338
2270 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2272 vtn_warn("Source and destination types of %s do not have the same "
2273 "ID (but are compatible): %u vs %u",
2274 spirv_op_to_string(opcode
), dst_type
->id
, src_type
->id
);
2278 vtn_fail("Source and destination types of %s do not match: %s vs. %s",
2279 spirv_op_to_string(opcode
),
2280 glsl_get_type_name(dst_type
->type
),
2281 glsl_get_type_name(src_type
->type
));
2284 static nir_ssa_def
*
2285 nir_shrink_zero_pad_vec(nir_builder
*b
, nir_ssa_def
*val
,
2286 unsigned num_components
)
2288 if (val
->num_components
== num_components
)
2291 nir_ssa_def
*comps
[NIR_MAX_VEC_COMPONENTS
];
2292 for (unsigned i
= 0; i
< num_components
; i
++) {
2293 if (i
< val
->num_components
)
2294 comps
[i
] = nir_channel(b
, val
, i
);
2296 comps
[i
] = nir_imm_intN_t(b
, 0, val
->bit_size
);
2298 return nir_vec(b
, comps
, num_components
);
2301 static nir_ssa_def
*
2302 nir_sloppy_bitcast(nir_builder
*b
, nir_ssa_def
*val
,
2303 const struct glsl_type
*type
)
2305 const unsigned num_components
= glsl_get_vector_elements(type
);
2306 const unsigned bit_size
= glsl_get_bit_size(type
);
2308 /* First, zero-pad to ensure that the value is big enough that when we
2309 * bit-cast it, we don't loose anything.
2311 if (val
->bit_size
< bit_size
) {
2312 const unsigned src_num_components_needed
=
2313 vtn_align_u32(val
->num_components
, bit_size
/ val
->bit_size
);
2314 val
= nir_shrink_zero_pad_vec(b
, val
, src_num_components_needed
);
2317 val
= nir_bitcast_vector(b
, val
, bit_size
);
2319 return nir_shrink_zero_pad_vec(b
, val
, num_components
);
2323 vtn_handle_variables(struct vtn_builder
*b
, SpvOp opcode
,
2324 const uint32_t *w
, unsigned count
)
2328 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_undef
);
2329 val
->type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2333 case SpvOpVariable
: {
2334 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2336 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2338 SpvStorageClass storage_class
= w
[3];
2339 nir_constant
*initializer
= NULL
;
2341 initializer
= vtn_value(b
, w
[4], vtn_value_type_constant
)->constant
;
2343 vtn_create_variable(b
, val
, ptr_type
, storage_class
, initializer
);
2347 case SpvOpAccessChain
:
2348 case SpvOpPtrAccessChain
:
2349 case SpvOpInBoundsAccessChain
:
2350 case SpvOpInBoundsPtrAccessChain
: {
2351 struct vtn_access_chain
*chain
= vtn_access_chain_create(b
, count
- 4);
2352 chain
->ptr_as_array
= (opcode
== SpvOpPtrAccessChain
|| opcode
== SpvOpInBoundsPtrAccessChain
);
2355 for (int i
= 4; i
< count
; i
++) {
2356 struct vtn_value
*link_val
= vtn_untyped_value(b
, w
[i
]);
2357 if (link_val
->value_type
== vtn_value_type_constant
) {
2358 chain
->link
[idx
].mode
= vtn_access_mode_literal
;
2359 const unsigned bit_size
= glsl_get_bit_size(link_val
->type
->type
);
2362 chain
->link
[idx
].id
= link_val
->constant
->values
[0][0].i8
;
2365 chain
->link
[idx
].id
= link_val
->constant
->values
[0][0].i16
;
2368 chain
->link
[idx
].id
= link_val
->constant
->values
[0][0].i32
;
2371 chain
->link
[idx
].id
= link_val
->constant
->values
[0][0].i64
;
2374 vtn_fail("Invalid bit size: %u", bit_size
);
2377 chain
->link
[idx
].mode
= vtn_access_mode_id
;
2378 chain
->link
[idx
].id
= w
[i
];
2384 struct vtn_type
*ptr_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2385 struct vtn_value
*base_val
= vtn_untyped_value(b
, w
[3]);
2386 if (base_val
->value_type
== vtn_value_type_sampled_image
) {
2387 /* This is rather insane. SPIR-V allows you to use OpSampledImage
2388 * to combine an array of images with a single sampler to get an
2389 * array of sampled images that all share the same sampler.
2390 * Fortunately, this means that we can more-or-less ignore the
2391 * sampler when crawling the access chain, but it does leave us
2392 * with this rather awkward little special-case.
2394 struct vtn_value
*val
=
2395 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
2396 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
2397 val
->sampled_image
->type
= base_val
->sampled_image
->type
;
2398 val
->sampled_image
->image
=
2399 vtn_pointer_dereference(b
, base_val
->sampled_image
->image
, chain
);
2400 val
->sampled_image
->sampler
= base_val
->sampled_image
->sampler
;
2401 vtn_foreach_decoration(b
, val
, ptr_decoration_cb
,
2402 val
->sampled_image
->image
);
2403 vtn_foreach_decoration(b
, val
, ptr_decoration_cb
,
2404 val
->sampled_image
->sampler
);
2406 vtn_assert(base_val
->value_type
== vtn_value_type_pointer
);
2407 struct vtn_value
*val
=
2408 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2409 val
->pointer
= vtn_pointer_dereference(b
, base_val
->pointer
, chain
);
2410 val
->pointer
->ptr_type
= ptr_type
;
2411 vtn_foreach_decoration(b
, val
, ptr_decoration_cb
, val
->pointer
);
2416 case SpvOpCopyMemory
: {
2417 struct vtn_value
*dest
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2418 struct vtn_value
*src
= vtn_value(b
, w
[2], vtn_value_type_pointer
);
2420 vtn_assert_types_equal(b
, opcode
, dest
->type
->deref
, src
->type
->deref
);
2422 vtn_variable_copy(b
, dest
->pointer
, src
->pointer
);
2427 struct vtn_type
*res_type
=
2428 vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
2429 struct vtn_value
*src_val
= vtn_value(b
, w
[3], vtn_value_type_pointer
);
2430 struct vtn_pointer
*src
= src_val
->pointer
;
2432 vtn_assert_types_equal(b
, opcode
, res_type
, src_val
->type
->deref
);
2434 if (glsl_type_is_image(res_type
->type
) ||
2435 glsl_type_is_sampler(res_type
->type
)) {
2436 vtn_push_value(b
, w
[2], vtn_value_type_pointer
)->pointer
= src
;
2440 vtn_push_ssa(b
, w
[2], res_type
, vtn_variable_load(b
, src
));
2445 struct vtn_value
*dest_val
= vtn_value(b
, w
[1], vtn_value_type_pointer
);
2446 struct vtn_pointer
*dest
= dest_val
->pointer
;
2447 struct vtn_value
*src_val
= vtn_untyped_value(b
, w
[2]);
2449 /* OpStore requires us to actually have a storage type */
2450 vtn_fail_if(dest
->type
->type
== NULL
,
2451 "Invalid destination type for OpStore");
2453 if (glsl_get_base_type(dest
->type
->type
) == GLSL_TYPE_BOOL
&&
2454 glsl_get_base_type(src_val
->type
->type
) == GLSL_TYPE_UINT
) {
2455 /* Early versions of GLSLang would use uint types for UBOs/SSBOs but
2456 * would then store them to a local variable as bool. Work around
2457 * the issue by doing an implicit conversion.
2459 * https://github.com/KhronosGroup/glslang/issues/170
2460 * https://bugs.freedesktop.org/show_bug.cgi?id=104424
2462 vtn_warn("OpStore of value of type OpTypeInt to a pointer to type "
2463 "OpTypeBool. Doing an implicit conversion to work around "
2465 struct vtn_ssa_value
*bool_ssa
=
2466 vtn_create_ssa_value(b
, dest
->type
->type
);
2467 bool_ssa
->def
= nir_i2b(&b
->nb
, vtn_ssa_value(b
, w
[2])->def
);
2468 vtn_variable_store(b
, bool_ssa
, dest
);
2472 vtn_assert_types_equal(b
, opcode
, dest_val
->type
->deref
, src_val
->type
);
2474 if (glsl_type_is_sampler(dest
->type
->type
)) {
2475 if (b
->wa_glslang_179
) {
2476 vtn_warn("OpStore of a sampler detected. Doing on-the-fly copy "
2477 "propagation to workaround the problem.");
2478 vtn_assert(dest
->var
->copy_prop_sampler
== NULL
);
2479 dest
->var
->copy_prop_sampler
=
2480 vtn_value(b
, w
[2], vtn_value_type_pointer
)->pointer
;
2482 vtn_fail("Vulkan does not allow OpStore of a sampler or image.");
2487 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[2]);
2488 vtn_variable_store(b
, src
, dest
);
2492 case SpvOpArrayLength
: {
2493 struct vtn_pointer
*ptr
=
2494 vtn_value(b
, w
[3], vtn_value_type_pointer
)->pointer
;
2495 const uint32_t field
= w
[4];
2497 vtn_fail_if(ptr
->type
->base_type
!= vtn_base_type_struct
,
2498 "OpArrayLength must take a pointer to a structure type");
2499 vtn_fail_if(field
!= ptr
->type
->length
- 1 ||
2500 ptr
->type
->members
[field
]->base_type
!= vtn_base_type_array
,
2501 "OpArrayLength must reference the last memeber of the "
2502 "structure and that must be an array");
2504 const uint32_t offset
= ptr
->type
->offsets
[field
];
2505 const uint32_t stride
= ptr
->type
->members
[field
]->stride
;
2507 if (!ptr
->block_index
) {
2508 struct vtn_access_chain chain
= {
2511 ptr
= vtn_pointer_dereference(b
, ptr
, &chain
);
2512 vtn_assert(ptr
->block_index
);
2515 nir_intrinsic_instr
*instr
=
2516 nir_intrinsic_instr_create(b
->nb
.shader
,
2517 nir_intrinsic_get_buffer_size
);
2518 instr
->src
[0] = nir_src_for_ssa(ptr
->block_index
);
2519 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
2520 nir_builder_instr_insert(&b
->nb
, &instr
->instr
);
2521 nir_ssa_def
*buf_size
= &instr
->dest
.ssa
;
2523 /* array_length = max(buffer_size - offset, 0) / stride */
2524 nir_ssa_def
*array_length
=
2529 nir_imm_int(&b
->nb
, offset
)),
2530 nir_imm_int(&b
->nb
, 0u)),
2531 nir_imm_int(&b
->nb
, stride
));
2533 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2534 val
->ssa
= vtn_create_ssa_value(b
, glsl_uint_type());
2535 val
->ssa
->def
= array_length
;
2539 case SpvOpConvertPtrToU
: {
2540 struct vtn_value
*u_val
= vtn_push_value(b
, w
[2], vtn_value_type_ssa
);
2542 vtn_fail_if(u_val
->type
->base_type
!= vtn_base_type_vector
&&
2543 u_val
->type
->base_type
!= vtn_base_type_scalar
,
2544 "OpConvertPtrToU can only be used to cast to a vector or "
2547 /* The pointer will be converted to an SSA value automatically */
2548 nir_ssa_def
*ptr_ssa
= vtn_ssa_value(b
, w
[3])->def
;
2550 u_val
->ssa
= vtn_create_ssa_value(b
, u_val
->type
->type
);
2551 u_val
->ssa
->def
= nir_sloppy_bitcast(&b
->nb
, ptr_ssa
, u_val
->type
->type
);
2555 case SpvOpConvertUToPtr
: {
2556 struct vtn_value
*ptr_val
=
2557 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
2558 struct vtn_value
*u_val
= vtn_value(b
, w
[3], vtn_value_type_ssa
);
2560 vtn_fail_if(ptr_val
->type
->type
== NULL
,
2561 "OpConvertUToPtr can only be used on physical pointers");
2563 vtn_fail_if(u_val
->type
->base_type
!= vtn_base_type_vector
&&
2564 u_val
->type
->base_type
!= vtn_base_type_scalar
,
2565 "OpConvertUToPtr can only be used to cast from a vector or "
2568 nir_ssa_def
*ptr_ssa
= nir_sloppy_bitcast(&b
->nb
, u_val
->ssa
->def
,
2569 ptr_val
->type
->type
);
2570 ptr_val
->pointer
= vtn_pointer_from_ssa(b
, ptr_ssa
, ptr_val
->type
);
2574 case SpvOpCopyMemorySized
:
2576 vtn_fail_with_opcode("Unhandled opcode", opcode
);