2 * Copyright © 2017 Connor Abbott
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 #include "nir_serialize.h"
25 #include "nir_control_flow.h"
26 #include "util/u_dynarray.h"
27 #include "util/u_math.h"
29 #define NIR_SERIALIZE_FUNC_HAS_IMPL ((void *)(intptr_t)1)
30 #define MAX_OBJECT_IDS (1 << 20)
39 const nir_shader
*nir
;
43 /* maps pointer to index */
44 struct hash_table
*remap_table
;
46 /* the next index to assign to a NIR in-memory object */
49 /* Array of write_phi_fixup structs representing phi sources that need to
50 * be resolved in the second pass.
52 struct util_dynarray phi_fixups
;
54 /* The last serialized type. */
55 const struct glsl_type
*last_type
;
56 const struct glsl_type
*last_interface_type
;
57 struct nir_variable_data last_var_data
;
59 /* For skipping equal ALU headers (typical after scalarization). */
60 nir_instr_type last_instr_type
;
61 uintptr_t last_alu_header_offset
;
63 /* Don't write optional data such as variable names. */
70 struct blob_reader
*blob
;
72 /* the next index to assign to a NIR in-memory object */
75 /* The length of the index -> object table */
76 uint32_t idx_table_len
;
78 /* map from index to deserialized pointer */
81 /* List of phi sources. */
82 struct list_head phi_srcs
;
84 /* The last deserialized type. */
85 const struct glsl_type
*last_type
;
86 const struct glsl_type
*last_interface_type
;
87 struct nir_variable_data last_var_data
;
91 write_add_object(write_ctx
*ctx
, const void *obj
)
93 uint32_t index
= ctx
->next_idx
++;
94 assert(index
!= MAX_OBJECT_IDS
);
95 _mesa_hash_table_insert(ctx
->remap_table
, obj
, (void *)(uintptr_t) index
);
99 write_lookup_object(write_ctx
*ctx
, const void *obj
)
101 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->remap_table
, obj
);
103 return (uint32_t)(uintptr_t) entry
->data
;
107 read_add_object(read_ctx
*ctx
, void *obj
)
109 assert(ctx
->next_idx
< ctx
->idx_table_len
);
110 ctx
->idx_table
[ctx
->next_idx
++] = obj
;
114 read_lookup_object(read_ctx
*ctx
, uint32_t idx
)
116 assert(idx
< ctx
->idx_table_len
);
117 return ctx
->idx_table
[idx
];
121 read_object(read_ctx
*ctx
)
123 return read_lookup_object(ctx
, blob_read_uint32(ctx
->blob
));
127 encode_bit_size_3bits(uint8_t bit_size
)
129 /* Encode values of 0, 1, 2, 4, 8, 16, 32, 64 in 3 bits. */
130 assert(bit_size
<= 64 && util_is_power_of_two_or_zero(bit_size
));
132 return util_logbase2(bit_size
) + 1;
137 decode_bit_size_3bits(uint8_t bit_size
)
140 return 1 << (bit_size
- 1);
144 #define NUM_COMPONENTS_IS_SEPARATE_7 7
147 encode_num_components_in_3bits(uint8_t num_components
)
149 if (num_components
<= 4)
150 return num_components
;
151 if (num_components
== 8)
153 if (num_components
== 16)
156 /* special value indicating that num_components is in the next uint32 */
157 return NUM_COMPONENTS_IS_SEPARATE_7
;
161 decode_num_components_in_3bits(uint8_t value
)
170 unreachable("invalid num_components encoding");
175 write_constant(write_ctx
*ctx
, const nir_constant
*c
)
177 blob_write_bytes(ctx
->blob
, c
->values
, sizeof(c
->values
));
178 blob_write_uint32(ctx
->blob
, c
->num_elements
);
179 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
180 write_constant(ctx
, c
->elements
[i
]);
183 static nir_constant
*
184 read_constant(read_ctx
*ctx
, nir_variable
*nvar
)
186 nir_constant
*c
= ralloc(nvar
, nir_constant
);
188 blob_copy_bytes(ctx
->blob
, (uint8_t *)c
->values
, sizeof(c
->values
));
189 c
->num_elements
= blob_read_uint32(ctx
->blob
);
190 c
->elements
= ralloc_array(nvar
, nir_constant
*, c
->num_elements
);
191 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
192 c
->elements
[i
] = read_constant(ctx
, nvar
);
197 enum var_data_encoding
{
199 var_encode_shader_temp
,
200 var_encode_function_temp
,
201 var_encode_location_diff
,
208 unsigned has_constant_initializer
:1;
209 unsigned has_interface_type
:1;
210 unsigned num_state_slots
:7;
211 unsigned data_encoding
:2;
212 unsigned type_same_as_last
:1;
213 unsigned interface_type_same_as_last
:1;
215 unsigned num_members
:16;
219 union packed_var_data_diff
{
224 int driver_location
:16;
229 write_variable(write_ctx
*ctx
, const nir_variable
*var
)
231 write_add_object(ctx
, var
);
233 assert(var
->num_state_slots
< (1 << 7));
235 STATIC_ASSERT(sizeof(union packed_var
) == 4);
236 union packed_var flags
;
239 flags
.u
.has_name
= !ctx
->strip
&& var
->name
;
240 flags
.u
.has_constant_initializer
= !!(var
->constant_initializer
);
241 flags
.u
.has_interface_type
= !!(var
->interface_type
);
242 flags
.u
.type_same_as_last
= var
->type
== ctx
->last_type
;
243 flags
.u
.interface_type_same_as_last
=
244 var
->interface_type
&& var
->interface_type
== ctx
->last_interface_type
;
245 flags
.u
.num_state_slots
= var
->num_state_slots
;
246 flags
.u
.num_members
= var
->num_members
;
248 struct nir_variable_data data
= var
->data
;
250 /* When stripping, we expect that the location is no longer needed,
251 * which is typically after shaders are linked.
254 data
.mode
!= nir_var_shader_in
&&
255 data
.mode
!= nir_var_shader_out
)
258 /* Temporary variables don't serialize var->data. */
259 if (data
.mode
== nir_var_shader_temp
)
260 flags
.u
.data_encoding
= var_encode_shader_temp
;
261 else if (data
.mode
== nir_var_function_temp
)
262 flags
.u
.data_encoding
= var_encode_function_temp
;
264 struct nir_variable_data tmp
= data
;
266 tmp
.location
= ctx
->last_var_data
.location
;
267 tmp
.location_frac
= ctx
->last_var_data
.location_frac
;
268 tmp
.driver_location
= ctx
->last_var_data
.driver_location
;
270 /* See if we can encode only the difference in locations from the last
273 if (memcmp(&ctx
->last_var_data
, &tmp
, sizeof(tmp
)) == 0 &&
274 abs((int)data
.location
-
275 (int)ctx
->last_var_data
.location
) < (1 << 12) &&
276 abs((int)data
.driver_location
-
277 (int)ctx
->last_var_data
.driver_location
) < (1 << 15))
278 flags
.u
.data_encoding
= var_encode_location_diff
;
280 flags
.u
.data_encoding
= var_encode_full
;
283 blob_write_uint32(ctx
->blob
, flags
.u32
);
285 if (!flags
.u
.type_same_as_last
) {
286 encode_type_to_blob(ctx
->blob
, var
->type
);
287 ctx
->last_type
= var
->type
;
290 if (var
->interface_type
&& !flags
.u
.interface_type_same_as_last
) {
291 encode_type_to_blob(ctx
->blob
, var
->interface_type
);
292 ctx
->last_interface_type
= var
->interface_type
;
295 if (flags
.u
.has_name
)
296 blob_write_string(ctx
->blob
, var
->name
);
298 if (flags
.u
.data_encoding
== var_encode_full
||
299 flags
.u
.data_encoding
== var_encode_location_diff
) {
300 if (flags
.u
.data_encoding
== var_encode_full
) {
301 blob_write_bytes(ctx
->blob
, &data
, sizeof(data
));
303 /* Serialize only the difference in locations from the last variable.
305 union packed_var_data_diff diff
;
307 diff
.u
.location
= data
.location
- ctx
->last_var_data
.location
;
308 diff
.u
.location_frac
= data
.location_frac
-
309 ctx
->last_var_data
.location_frac
;
310 diff
.u
.driver_location
= data
.driver_location
-
311 ctx
->last_var_data
.driver_location
;
313 blob_write_uint32(ctx
->blob
, diff
.u32
);
316 ctx
->last_var_data
= data
;
319 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
320 blob_write_bytes(ctx
->blob
, &var
->state_slots
[i
],
321 sizeof(var
->state_slots
[i
]));
323 if (var
->constant_initializer
)
324 write_constant(ctx
, var
->constant_initializer
);
325 if (var
->num_members
> 0) {
326 blob_write_bytes(ctx
->blob
, (uint8_t *) var
->members
,
327 var
->num_members
* sizeof(*var
->members
));
331 static nir_variable
*
332 read_variable(read_ctx
*ctx
)
334 nir_variable
*var
= rzalloc(ctx
->nir
, nir_variable
);
335 read_add_object(ctx
, var
);
337 union packed_var flags
;
338 flags
.u32
= blob_read_uint32(ctx
->blob
);
340 if (flags
.u
.type_same_as_last
) {
341 var
->type
= ctx
->last_type
;
343 var
->type
= decode_type_from_blob(ctx
->blob
);
344 ctx
->last_type
= var
->type
;
347 if (flags
.u
.has_interface_type
) {
348 if (flags
.u
.interface_type_same_as_last
) {
349 var
->interface_type
= ctx
->last_interface_type
;
351 var
->interface_type
= decode_type_from_blob(ctx
->blob
);
352 ctx
->last_interface_type
= var
->interface_type
;
356 if (flags
.u
.has_name
) {
357 const char *name
= blob_read_string(ctx
->blob
);
358 var
->name
= ralloc_strdup(var
, name
);
363 if (flags
.u
.data_encoding
== var_encode_shader_temp
)
364 var
->data
.mode
= nir_var_shader_temp
;
365 else if (flags
.u
.data_encoding
== var_encode_function_temp
)
366 var
->data
.mode
= nir_var_function_temp
;
367 else if (flags
.u
.data_encoding
== var_encode_full
) {
368 blob_copy_bytes(ctx
->blob
, (uint8_t *) &var
->data
, sizeof(var
->data
));
369 ctx
->last_var_data
= var
->data
;
370 } else { /* var_encode_location_diff */
371 union packed_var_data_diff diff
;
372 diff
.u32
= blob_read_uint32(ctx
->blob
);
374 var
->data
= ctx
->last_var_data
;
375 var
->data
.location
+= diff
.u
.location
;
376 var
->data
.location_frac
+= diff
.u
.location_frac
;
377 var
->data
.driver_location
+= diff
.u
.driver_location
;
379 ctx
->last_var_data
= var
->data
;
382 var
->num_state_slots
= flags
.u
.num_state_slots
;
383 if (var
->num_state_slots
!= 0) {
384 var
->state_slots
= ralloc_array(var
, nir_state_slot
,
385 var
->num_state_slots
);
386 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
387 blob_copy_bytes(ctx
->blob
, &var
->state_slots
[i
],
388 sizeof(var
->state_slots
[i
]));
391 if (flags
.u
.has_constant_initializer
)
392 var
->constant_initializer
= read_constant(ctx
, var
);
394 var
->constant_initializer
= NULL
;
395 var
->num_members
= flags
.u
.num_members
;
396 if (var
->num_members
> 0) {
397 var
->members
= ralloc_array(var
, struct nir_variable_data
,
399 blob_copy_bytes(ctx
->blob
, (uint8_t *) var
->members
,
400 var
->num_members
* sizeof(*var
->members
));
407 write_var_list(write_ctx
*ctx
, const struct exec_list
*src
)
409 blob_write_uint32(ctx
->blob
, exec_list_length(src
));
410 foreach_list_typed(nir_variable
, var
, node
, src
) {
411 write_variable(ctx
, var
);
416 read_var_list(read_ctx
*ctx
, struct exec_list
*dst
)
418 exec_list_make_empty(dst
);
419 unsigned num_vars
= blob_read_uint32(ctx
->blob
);
420 for (unsigned i
= 0; i
< num_vars
; i
++) {
421 nir_variable
*var
= read_variable(ctx
);
422 exec_list_push_tail(dst
, &var
->node
);
427 write_register(write_ctx
*ctx
, const nir_register
*reg
)
429 write_add_object(ctx
, reg
);
430 blob_write_uint32(ctx
->blob
, reg
->num_components
);
431 blob_write_uint32(ctx
->blob
, reg
->bit_size
);
432 blob_write_uint32(ctx
->blob
, reg
->num_array_elems
);
433 blob_write_uint32(ctx
->blob
, reg
->index
);
434 blob_write_uint32(ctx
->blob
, !ctx
->strip
&& reg
->name
);
435 if (!ctx
->strip
&& reg
->name
)
436 blob_write_string(ctx
->blob
, reg
->name
);
439 static nir_register
*
440 read_register(read_ctx
*ctx
)
442 nir_register
*reg
= ralloc(ctx
->nir
, nir_register
);
443 read_add_object(ctx
, reg
);
444 reg
->num_components
= blob_read_uint32(ctx
->blob
);
445 reg
->bit_size
= blob_read_uint32(ctx
->blob
);
446 reg
->num_array_elems
= blob_read_uint32(ctx
->blob
);
447 reg
->index
= blob_read_uint32(ctx
->blob
);
448 bool has_name
= blob_read_uint32(ctx
->blob
);
450 const char *name
= blob_read_string(ctx
->blob
);
451 reg
->name
= ralloc_strdup(reg
, name
);
456 list_inithead(®
->uses
);
457 list_inithead(®
->defs
);
458 list_inithead(®
->if_uses
);
464 write_reg_list(write_ctx
*ctx
, const struct exec_list
*src
)
466 blob_write_uint32(ctx
->blob
, exec_list_length(src
));
467 foreach_list_typed(nir_register
, reg
, node
, src
)
468 write_register(ctx
, reg
);
472 read_reg_list(read_ctx
*ctx
, struct exec_list
*dst
)
474 exec_list_make_empty(dst
);
475 unsigned num_regs
= blob_read_uint32(ctx
->blob
);
476 for (unsigned i
= 0; i
< num_regs
; i
++) {
477 nir_register
*reg
= read_register(ctx
);
478 exec_list_push_tail(dst
, ®
->node
);
485 unsigned is_ssa
:1; /* <-- Header */
486 unsigned is_indirect
:1;
487 unsigned object_idx
:20;
488 unsigned _footer
:10; /* <-- Footer */
491 unsigned _header
:22; /* <-- Header */
492 unsigned negate
:1; /* <-- Footer */
494 unsigned swizzle_x
:2;
495 unsigned swizzle_y
:2;
496 unsigned swizzle_z
:2;
497 unsigned swizzle_w
:2;
500 unsigned _header
:22; /* <-- Header */
501 unsigned src_type
:5; /* <-- Footer */
507 write_src_full(write_ctx
*ctx
, const nir_src
*src
, union packed_src header
)
509 /* Since sources are very frequent, we try to save some space when storing
510 * them. In particular, we store whether the source is a register and
511 * whether the register has an indirect index in the low two bits. We can
512 * assume that the high two bits of the index are zero, since otherwise our
513 * address space would've been exhausted allocating the remap table!
515 header
.any
.is_ssa
= src
->is_ssa
;
517 header
.any
.object_idx
= write_lookup_object(ctx
, src
->ssa
);
518 blob_write_uint32(ctx
->blob
, header
.u32
);
520 header
.any
.object_idx
= write_lookup_object(ctx
, src
->reg
.reg
);
521 header
.any
.is_indirect
= !!src
->reg
.indirect
;
522 blob_write_uint32(ctx
->blob
, header
.u32
);
523 blob_write_uint32(ctx
->blob
, src
->reg
.base_offset
);
524 if (src
->reg
.indirect
) {
525 union packed_src header
= {0};
526 write_src_full(ctx
, src
->reg
.indirect
, header
);
532 write_src(write_ctx
*ctx
, const nir_src
*src
)
534 union packed_src header
= {0};
535 write_src_full(ctx
, src
, header
);
538 static union packed_src
539 read_src(read_ctx
*ctx
, nir_src
*src
, void *mem_ctx
)
541 STATIC_ASSERT(sizeof(union packed_src
) == 4);
542 union packed_src header
;
543 header
.u32
= blob_read_uint32(ctx
->blob
);
545 src
->is_ssa
= header
.any
.is_ssa
;
547 src
->ssa
= read_lookup_object(ctx
, header
.any
.object_idx
);
549 src
->reg
.reg
= read_lookup_object(ctx
, header
.any
.object_idx
);
550 src
->reg
.base_offset
= blob_read_uint32(ctx
->blob
);
551 if (header
.any
.is_indirect
) {
552 src
->reg
.indirect
= ralloc(mem_ctx
, nir_src
);
553 read_src(ctx
, src
->reg
.indirect
, mem_ctx
);
555 src
->reg
.indirect
= NULL
;
566 uint8_t num_components
:3;
571 uint8_t is_indirect
:1;
576 enum intrinsic_const_indices_encoding
{
577 /* Use the 9 bits of packed_const_indices to store 1-9 indices.
578 * 1 9-bit index, or 2 4-bit indices, or 3 3-bit indices, or
579 * 4 2-bit indices, or 5-9 1-bit indices.
581 * The common case for load_ubo is 0, 0, 0, which is trivially represented.
582 * The common cases for load_interpolated_input also fit here, e.g.: 7, 3
584 const_indices_9bit_all_combined
,
586 const_indices_8bit
, /* 8 bits per element */
587 const_indices_16bit
, /* 16 bits per element */
588 const_indices_32bit
, /* 32 bits per element */
591 enum load_const_packing
{
592 /* Constants are not packed and are stored in following dwords. */
595 /* packed_value contains high 19 bits, low bits are 0,
596 * good for floating-point decimals
598 load_const_scalar_hi_19bits
,
600 /* packed_value contains low 19 bits, high bits are sign-extended */
601 load_const_scalar_lo_19bits_sext
,
607 unsigned instr_type
:4; /* always present */
609 unsigned dest
:8; /* always last */
612 unsigned instr_type
:4;
614 unsigned no_signed_wrap
:1;
615 unsigned no_unsigned_wrap
:1;
617 /* Reg: writemask; SSA: swizzles for 2 srcs */
618 unsigned writemask_or_two_swizzles
:4;
620 unsigned packed_src_ssa_16bit
:1;
621 /* Scalarized ALUs always have the same header. */
622 unsigned num_followup_alu_sharing_header
:2;
626 unsigned instr_type
:4;
627 unsigned deref_type
:3;
628 unsigned cast_type_same_as_last
:1;
629 unsigned mode
:10; /* deref_var redefines this */
630 unsigned packed_src_ssa_16bit
:1; /* deref_var redefines this */
631 unsigned _pad
:5; /* deref_var redefines this */
635 unsigned instr_type
:4;
636 unsigned deref_type
:3;
638 unsigned object_idx
:16; /* if 0, the object ID is a separate uint32 */
642 unsigned instr_type
:4;
643 unsigned intrinsic
:9;
644 unsigned const_indices_encoding
:2;
645 unsigned packed_const_indices
:9;
649 unsigned instr_type
:4;
650 unsigned last_component
:4;
652 unsigned packing
:2; /* enum load_const_packing */
653 unsigned packed_value
:19; /* meaning determined by packing */
656 unsigned instr_type
:4;
657 unsigned last_component
:4;
662 unsigned instr_type
:4;
665 unsigned texture_array_size
:12;
669 unsigned instr_type
:4;
670 unsigned num_srcs
:20;
674 unsigned instr_type
:4;
680 /* Write "lo24" as low 24 bits in the first uint32. */
682 write_dest(write_ctx
*ctx
, const nir_dest
*dst
, union packed_instr header
,
683 nir_instr_type instr_type
)
685 STATIC_ASSERT(sizeof(union packed_dest
) == 1);
686 union packed_dest dest
;
689 dest
.ssa
.is_ssa
= dst
->is_ssa
;
691 dest
.ssa
.has_name
= !ctx
->strip
&& dst
->ssa
.name
;
692 dest
.ssa
.num_components
=
693 encode_num_components_in_3bits(dst
->ssa
.num_components
);
694 dest
.ssa
.bit_size
= encode_bit_size_3bits(dst
->ssa
.bit_size
);
696 dest
.reg
.is_indirect
= !!(dst
->reg
.indirect
);
698 header
.any
.dest
= dest
.u8
;
700 /* Check if the current ALU instruction has the same header as the previous
701 * instruction that is also ALU. If it is, we don't have to write
702 * the current header. This is a typical occurence after scalarization.
704 if (instr_type
== nir_instr_type_alu
) {
705 bool equal_header
= false;
707 if (ctx
->last_instr_type
== nir_instr_type_alu
) {
708 assert(ctx
->last_alu_header_offset
);
709 union packed_instr
*last_header
=
710 (union packed_instr
*)(ctx
->blob
->data
+
711 ctx
->last_alu_header_offset
);
713 /* Clear the field that counts ALUs with equal headers. */
714 union packed_instr clean_header
;
715 clean_header
.u32
= last_header
->u32
;
716 clean_header
.alu
.num_followup_alu_sharing_header
= 0;
718 /* There can be at most 4 consecutive ALU instructions
719 * sharing the same header.
721 if (last_header
->alu
.num_followup_alu_sharing_header
< 3 &&
722 header
.u32
== clean_header
.u32
) {
723 last_header
->alu
.num_followup_alu_sharing_header
++;
729 ctx
->last_alu_header_offset
= ctx
->blob
->size
;
730 blob_write_uint32(ctx
->blob
, header
.u32
);
733 blob_write_uint32(ctx
->blob
, header
.u32
);
736 if (dest
.ssa
.is_ssa
&&
737 dest
.ssa
.num_components
== NUM_COMPONENTS_IS_SEPARATE_7
)
738 blob_write_uint32(ctx
->blob
, dst
->ssa
.num_components
);
741 write_add_object(ctx
, &dst
->ssa
);
742 if (dest
.ssa
.has_name
)
743 blob_write_string(ctx
->blob
, dst
->ssa
.name
);
745 blob_write_uint32(ctx
->blob
, write_lookup_object(ctx
, dst
->reg
.reg
));
746 blob_write_uint32(ctx
->blob
, dst
->reg
.base_offset
);
747 if (dst
->reg
.indirect
)
748 write_src(ctx
, dst
->reg
.indirect
);
753 read_dest(read_ctx
*ctx
, nir_dest
*dst
, nir_instr
*instr
,
754 union packed_instr header
)
756 union packed_dest dest
;
757 dest
.u8
= header
.any
.dest
;
759 if (dest
.ssa
.is_ssa
) {
760 unsigned bit_size
= decode_bit_size_3bits(dest
.ssa
.bit_size
);
761 unsigned num_components
;
762 if (dest
.ssa
.num_components
== NUM_COMPONENTS_IS_SEPARATE_7
)
763 num_components
= blob_read_uint32(ctx
->blob
);
765 num_components
= decode_num_components_in_3bits(dest
.ssa
.num_components
);
766 char *name
= dest
.ssa
.has_name
? blob_read_string(ctx
->blob
) : NULL
;
767 nir_ssa_dest_init(instr
, dst
, num_components
, bit_size
, name
);
768 read_add_object(ctx
, &dst
->ssa
);
770 dst
->reg
.reg
= read_object(ctx
);
771 dst
->reg
.base_offset
= blob_read_uint32(ctx
->blob
);
772 if (dest
.reg
.is_indirect
) {
773 dst
->reg
.indirect
= ralloc(instr
, nir_src
);
774 read_src(ctx
, dst
->reg
.indirect
, instr
);
780 are_object_ids_16bit(write_ctx
*ctx
)
782 /* Check the highest object ID, because they are monotonic. */
783 return ctx
->next_idx
< (1 << 16);
787 is_alu_src_ssa_16bit(write_ctx
*ctx
, const nir_alu_instr
*alu
)
789 unsigned num_srcs
= nir_op_infos
[alu
->op
].num_inputs
;
791 for (unsigned i
= 0; i
< num_srcs
; i
++) {
792 if (!alu
->src
[i
].src
.is_ssa
|| alu
->src
[i
].abs
|| alu
->src
[i
].negate
)
795 unsigned src_components
= nir_ssa_alu_instr_src_components(alu
, i
);
797 for (unsigned chan
= 0; chan
< src_components
; chan
++) {
798 /* The swizzles for src0.x and src1.x are stored
799 * in writemask_or_two_swizzles for SSA ALUs.
801 if (alu
->dest
.dest
.is_ssa
&& i
< 2 && chan
== 0 &&
802 alu
->src
[i
].swizzle
[chan
] < 4)
805 if (alu
->src
[i
].swizzle
[chan
] != chan
)
810 return are_object_ids_16bit(ctx
);
814 write_alu(write_ctx
*ctx
, const nir_alu_instr
*alu
)
816 unsigned num_srcs
= nir_op_infos
[alu
->op
].num_inputs
;
817 unsigned dst_components
= nir_dest_num_components(alu
->dest
.dest
);
819 /* 9 bits for nir_op */
820 STATIC_ASSERT(nir_num_opcodes
<= 512);
821 union packed_instr header
;
824 header
.alu
.instr_type
= alu
->instr
.type
;
825 header
.alu
.exact
= alu
->exact
;
826 header
.alu
.no_signed_wrap
= alu
->no_signed_wrap
;
827 header
.alu
.no_unsigned_wrap
= alu
->no_unsigned_wrap
;
828 header
.alu
.saturate
= alu
->dest
.saturate
;
829 header
.alu
.op
= alu
->op
;
830 header
.alu
.packed_src_ssa_16bit
= is_alu_src_ssa_16bit(ctx
, alu
);
832 if (header
.alu
.packed_src_ssa_16bit
&&
833 alu
->dest
.dest
.is_ssa
) {
834 /* For packed srcs of SSA ALUs, this field stores the swizzles. */
835 header
.alu
.writemask_or_two_swizzles
= alu
->src
[0].swizzle
[0];
837 header
.alu
.writemask_or_two_swizzles
|= alu
->src
[1].swizzle
[0] << 2;
838 } else if (!alu
->dest
.dest
.is_ssa
&& dst_components
<= 4) {
839 /* For vec4 registers, this field is a writemask. */
840 header
.alu
.writemask_or_two_swizzles
= alu
->dest
.write_mask
;
843 write_dest(ctx
, &alu
->dest
.dest
, header
, alu
->instr
.type
);
845 if (!alu
->dest
.dest
.is_ssa
&& dst_components
> 4)
846 blob_write_uint32(ctx
->blob
, alu
->dest
.write_mask
);
848 if (header
.alu
.packed_src_ssa_16bit
) {
849 for (unsigned i
= 0; i
< num_srcs
; i
++) {
850 assert(alu
->src
[i
].src
.is_ssa
);
851 unsigned idx
= write_lookup_object(ctx
, alu
->src
[i
].src
.ssa
);
852 assert(idx
< (1 << 16));
853 blob_write_uint16(ctx
->blob
, idx
);
856 for (unsigned i
= 0; i
< num_srcs
; i
++) {
857 unsigned src_channels
= nir_ssa_alu_instr_src_components(alu
, i
);
858 unsigned src_components
= nir_src_num_components(alu
->src
[i
].src
);
859 union packed_src src
;
860 bool packed
= src_components
<= 4 && src_channels
<= 4;
863 src
.alu
.negate
= alu
->src
[i
].negate
;
864 src
.alu
.abs
= alu
->src
[i
].abs
;
867 src
.alu
.swizzle_x
= alu
->src
[i
].swizzle
[0];
868 src
.alu
.swizzle_y
= alu
->src
[i
].swizzle
[1];
869 src
.alu
.swizzle_z
= alu
->src
[i
].swizzle
[2];
870 src
.alu
.swizzle_w
= alu
->src
[i
].swizzle
[3];
873 write_src_full(ctx
, &alu
->src
[i
].src
, src
);
875 /* Store swizzles for vec8 and vec16. */
877 for (unsigned o
= 0; o
< src_channels
; o
+= 8) {
880 for (unsigned j
= 0; j
< 8 && o
+ j
< src_channels
; j
++) {
881 value
|= (uint32_t)alu
->src
[i
].swizzle
[o
+ j
] <<
882 (4 * j
); /* 4 bits per swizzle */
885 blob_write_uint32(ctx
->blob
, value
);
892 static nir_alu_instr
*
893 read_alu(read_ctx
*ctx
, union packed_instr header
)
895 unsigned num_srcs
= nir_op_infos
[header
.alu
.op
].num_inputs
;
896 nir_alu_instr
*alu
= nir_alu_instr_create(ctx
->nir
, header
.alu
.op
);
898 alu
->exact
= header
.alu
.exact
;
899 alu
->no_signed_wrap
= header
.alu
.no_signed_wrap
;
900 alu
->no_unsigned_wrap
= header
.alu
.no_unsigned_wrap
;
901 alu
->dest
.saturate
= header
.alu
.saturate
;
903 read_dest(ctx
, &alu
->dest
.dest
, &alu
->instr
, header
);
905 unsigned dst_components
= nir_dest_num_components(alu
->dest
.dest
);
907 if (alu
->dest
.dest
.is_ssa
) {
908 alu
->dest
.write_mask
= u_bit_consecutive(0, dst_components
);
909 } else if (dst_components
<= 4) {
910 alu
->dest
.write_mask
= header
.alu
.writemask_or_two_swizzles
;
912 alu
->dest
.write_mask
= blob_read_uint32(ctx
->blob
);
915 if (header
.alu
.packed_src_ssa_16bit
) {
916 for (unsigned i
= 0; i
< num_srcs
; i
++) {
917 nir_alu_src
*src
= &alu
->src
[i
];
918 src
->src
.is_ssa
= true;
919 src
->src
.ssa
= read_lookup_object(ctx
, blob_read_uint16(ctx
->blob
));
921 memset(&src
->swizzle
, 0, sizeof(src
->swizzle
));
923 unsigned src_components
= nir_ssa_alu_instr_src_components(alu
, i
);
925 for (unsigned chan
= 0; chan
< src_components
; chan
++)
926 src
->swizzle
[chan
] = chan
;
929 for (unsigned i
= 0; i
< num_srcs
; i
++) {
930 union packed_src src
= read_src(ctx
, &alu
->src
[i
].src
, &alu
->instr
);
931 unsigned src_channels
= nir_ssa_alu_instr_src_components(alu
, i
);
932 unsigned src_components
= nir_src_num_components(alu
->src
[i
].src
);
933 bool packed
= src_components
<= 4 && src_channels
<= 4;
935 alu
->src
[i
].negate
= src
.alu
.negate
;
936 alu
->src
[i
].abs
= src
.alu
.abs
;
938 memset(&alu
->src
[i
].swizzle
, 0, sizeof(alu
->src
[i
].swizzle
));
941 alu
->src
[i
].swizzle
[0] = src
.alu
.swizzle_x
;
942 alu
->src
[i
].swizzle
[1] = src
.alu
.swizzle_y
;
943 alu
->src
[i
].swizzle
[2] = src
.alu
.swizzle_z
;
944 alu
->src
[i
].swizzle
[3] = src
.alu
.swizzle_w
;
946 /* Load swizzles for vec8 and vec16. */
947 for (unsigned o
= 0; o
< src_channels
; o
+= 8) {
948 unsigned value
= blob_read_uint32(ctx
->blob
);
950 for (unsigned j
= 0; j
< 8 && o
+ j
< src_channels
; j
++) {
951 alu
->src
[i
].swizzle
[o
+ j
] =
952 (value
>> (4 * j
)) & 0xf; /* 4 bits per swizzle */
959 if (header
.alu
.packed_src_ssa_16bit
&&
960 alu
->dest
.dest
.is_ssa
) {
961 alu
->src
[0].swizzle
[0] = header
.alu
.writemask_or_two_swizzles
& 0x3;
963 alu
->src
[1].swizzle
[0] = header
.alu
.writemask_or_two_swizzles
>> 2;
970 write_deref(write_ctx
*ctx
, const nir_deref_instr
*deref
)
972 assert(deref
->deref_type
< 8);
973 assert(deref
->mode
< (1 << 10));
975 union packed_instr header
;
978 header
.deref
.instr_type
= deref
->instr
.type
;
979 header
.deref
.deref_type
= deref
->deref_type
;
981 if (deref
->deref_type
== nir_deref_type_cast
) {
982 header
.deref
.mode
= deref
->mode
;
983 header
.deref
.cast_type_same_as_last
= deref
->type
== ctx
->last_type
;
986 unsigned var_idx
= 0;
987 if (deref
->deref_type
== nir_deref_type_var
) {
988 var_idx
= write_lookup_object(ctx
, deref
->var
);
989 if (var_idx
&& var_idx
< (1 << 16))
990 header
.deref_var
.object_idx
= var_idx
;
993 if (deref
->deref_type
== nir_deref_type_array
||
994 deref
->deref_type
== nir_deref_type_ptr_as_array
) {
995 header
.deref
.packed_src_ssa_16bit
=
996 deref
->parent
.is_ssa
&& deref
->arr
.index
.is_ssa
&&
997 are_object_ids_16bit(ctx
);
1000 write_dest(ctx
, &deref
->dest
, header
, deref
->instr
.type
);
1002 switch (deref
->deref_type
) {
1003 case nir_deref_type_var
:
1004 if (!header
.deref_var
.object_idx
)
1005 blob_write_uint32(ctx
->blob
, var_idx
);
1008 case nir_deref_type_struct
:
1009 write_src(ctx
, &deref
->parent
);
1010 blob_write_uint32(ctx
->blob
, deref
->strct
.index
);
1013 case nir_deref_type_array
:
1014 case nir_deref_type_ptr_as_array
:
1015 if (header
.deref
.packed_src_ssa_16bit
) {
1016 blob_write_uint16(ctx
->blob
,
1017 write_lookup_object(ctx
, deref
->parent
.ssa
));
1018 blob_write_uint16(ctx
->blob
,
1019 write_lookup_object(ctx
, deref
->arr
.index
.ssa
));
1021 write_src(ctx
, &deref
->parent
);
1022 write_src(ctx
, &deref
->arr
.index
);
1026 case nir_deref_type_cast
:
1027 write_src(ctx
, &deref
->parent
);
1028 blob_write_uint32(ctx
->blob
, deref
->cast
.ptr_stride
);
1029 if (!header
.deref
.cast_type_same_as_last
) {
1030 encode_type_to_blob(ctx
->blob
, deref
->type
);
1031 ctx
->last_type
= deref
->type
;
1035 case nir_deref_type_array_wildcard
:
1036 write_src(ctx
, &deref
->parent
);
1040 unreachable("Invalid deref type");
1044 static nir_deref_instr
*
1045 read_deref(read_ctx
*ctx
, union packed_instr header
)
1047 nir_deref_type deref_type
= header
.deref
.deref_type
;
1048 nir_deref_instr
*deref
= nir_deref_instr_create(ctx
->nir
, deref_type
);
1050 read_dest(ctx
, &deref
->dest
, &deref
->instr
, header
);
1052 nir_deref_instr
*parent
;
1054 switch (deref
->deref_type
) {
1055 case nir_deref_type_var
:
1056 if (header
.deref_var
.object_idx
)
1057 deref
->var
= read_lookup_object(ctx
, header
.deref_var
.object_idx
);
1059 deref
->var
= read_object(ctx
);
1061 deref
->type
= deref
->var
->type
;
1064 case nir_deref_type_struct
:
1065 read_src(ctx
, &deref
->parent
, &deref
->instr
);
1066 parent
= nir_src_as_deref(deref
->parent
);
1067 deref
->strct
.index
= blob_read_uint32(ctx
->blob
);
1068 deref
->type
= glsl_get_struct_field(parent
->type
, deref
->strct
.index
);
1071 case nir_deref_type_array
:
1072 case nir_deref_type_ptr_as_array
:
1073 if (header
.deref
.packed_src_ssa_16bit
) {
1074 deref
->parent
.is_ssa
= true;
1075 deref
->parent
.ssa
= read_lookup_object(ctx
, blob_read_uint16(ctx
->blob
));
1076 deref
->arr
.index
.is_ssa
= true;
1077 deref
->arr
.index
.ssa
= read_lookup_object(ctx
, blob_read_uint16(ctx
->blob
));
1079 read_src(ctx
, &deref
->parent
, &deref
->instr
);
1080 read_src(ctx
, &deref
->arr
.index
, &deref
->instr
);
1083 parent
= nir_src_as_deref(deref
->parent
);
1084 if (deref
->deref_type
== nir_deref_type_array
)
1085 deref
->type
= glsl_get_array_element(parent
->type
);
1087 deref
->type
= parent
->type
;
1090 case nir_deref_type_cast
:
1091 read_src(ctx
, &deref
->parent
, &deref
->instr
);
1092 deref
->cast
.ptr_stride
= blob_read_uint32(ctx
->blob
);
1093 if (header
.deref
.cast_type_same_as_last
) {
1094 deref
->type
= ctx
->last_type
;
1096 deref
->type
= decode_type_from_blob(ctx
->blob
);
1097 ctx
->last_type
= deref
->type
;
1101 case nir_deref_type_array_wildcard
:
1102 read_src(ctx
, &deref
->parent
, &deref
->instr
);
1103 parent
= nir_src_as_deref(deref
->parent
);
1104 deref
->type
= glsl_get_array_element(parent
->type
);
1108 unreachable("Invalid deref type");
1111 if (deref_type
== nir_deref_type_var
) {
1112 deref
->mode
= deref
->var
->data
.mode
;
1113 } else if (deref
->deref_type
== nir_deref_type_cast
) {
1114 deref
->mode
= header
.deref
.mode
;
1116 assert(deref
->parent
.is_ssa
);
1117 deref
->mode
= nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
)->mode
;
1124 write_intrinsic(write_ctx
*ctx
, const nir_intrinsic_instr
*intrin
)
1126 /* 9 bits for nir_intrinsic_op */
1127 STATIC_ASSERT(nir_num_intrinsics
<= 512);
1128 unsigned num_srcs
= nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
;
1129 unsigned num_indices
= nir_intrinsic_infos
[intrin
->intrinsic
].num_indices
;
1130 assert(intrin
->intrinsic
< 512);
1132 union packed_instr header
;
1135 header
.intrinsic
.instr_type
= intrin
->instr
.type
;
1136 header
.intrinsic
.intrinsic
= intrin
->intrinsic
;
1138 /* Analyze constant indices to decide how to encode them. */
1140 unsigned max_bits
= 0;
1141 for (unsigned i
= 0; i
< num_indices
; i
++) {
1142 unsigned max
= util_last_bit(intrin
->const_index
[i
]);
1143 max_bits
= MAX2(max_bits
, max
);
1146 if (max_bits
* num_indices
<= 9) {
1147 header
.intrinsic
.const_indices_encoding
= const_indices_9bit_all_combined
;
1149 /* Pack all const indices into 6 bits. */
1150 unsigned bit_size
= 9 / num_indices
;
1151 for (unsigned i
= 0; i
< num_indices
; i
++) {
1152 header
.intrinsic
.packed_const_indices
|=
1153 intrin
->const_index
[i
] << (i
* bit_size
);
1155 } else if (max_bits
<= 8)
1156 header
.intrinsic
.const_indices_encoding
= const_indices_8bit
;
1157 else if (max_bits
<= 16)
1158 header
.intrinsic
.const_indices_encoding
= const_indices_16bit
;
1160 header
.intrinsic
.const_indices_encoding
= const_indices_32bit
;
1163 if (nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
)
1164 write_dest(ctx
, &intrin
->dest
, header
, intrin
->instr
.type
);
1166 blob_write_uint32(ctx
->blob
, header
.u32
);
1168 for (unsigned i
= 0; i
< num_srcs
; i
++)
1169 write_src(ctx
, &intrin
->src
[i
]);
1172 switch (header
.intrinsic
.const_indices_encoding
) {
1173 case const_indices_8bit
:
1174 for (unsigned i
= 0; i
< num_indices
; i
++)
1175 blob_write_uint8(ctx
->blob
, intrin
->const_index
[i
]);
1177 case const_indices_16bit
:
1178 for (unsigned i
= 0; i
< num_indices
; i
++)
1179 blob_write_uint16(ctx
->blob
, intrin
->const_index
[i
]);
1181 case const_indices_32bit
:
1182 for (unsigned i
= 0; i
< num_indices
; i
++)
1183 blob_write_uint32(ctx
->blob
, intrin
->const_index
[i
]);
1189 static nir_intrinsic_instr
*
1190 read_intrinsic(read_ctx
*ctx
, union packed_instr header
)
1192 nir_intrinsic_op op
= header
.intrinsic
.intrinsic
;
1193 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(ctx
->nir
, op
);
1195 unsigned num_srcs
= nir_intrinsic_infos
[op
].num_srcs
;
1196 unsigned num_indices
= nir_intrinsic_infos
[op
].num_indices
;
1198 if (nir_intrinsic_infos
[op
].has_dest
)
1199 read_dest(ctx
, &intrin
->dest
, &intrin
->instr
, header
);
1201 for (unsigned i
= 0; i
< num_srcs
; i
++)
1202 read_src(ctx
, &intrin
->src
[i
], &intrin
->instr
);
1204 /* Vectorized instrinsics have num_components same as dst or src that has
1205 * 0 components in the info. Find it.
1207 if (nir_intrinsic_infos
[op
].has_dest
&&
1208 nir_intrinsic_infos
[op
].dest_components
== 0) {
1209 intrin
->num_components
= nir_dest_num_components(intrin
->dest
);
1211 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1212 if (nir_intrinsic_infos
[op
].src_components
[i
] == 0) {
1213 intrin
->num_components
= nir_src_num_components(intrin
->src
[i
]);
1220 switch (header
.intrinsic
.const_indices_encoding
) {
1221 case const_indices_9bit_all_combined
: {
1222 unsigned bit_size
= 9 / num_indices
;
1223 unsigned bit_mask
= u_bit_consecutive(0, bit_size
);
1224 for (unsigned i
= 0; i
< num_indices
; i
++) {
1225 intrin
->const_index
[i
] =
1226 (header
.intrinsic
.packed_const_indices
>> (i
* bit_size
)) &
1231 case const_indices_8bit
:
1232 for (unsigned i
= 0; i
< num_indices
; i
++)
1233 intrin
->const_index
[i
] = blob_read_uint8(ctx
->blob
);
1235 case const_indices_16bit
:
1236 for (unsigned i
= 0; i
< num_indices
; i
++)
1237 intrin
->const_index
[i
] = blob_read_uint16(ctx
->blob
);
1239 case const_indices_32bit
:
1240 for (unsigned i
= 0; i
< num_indices
; i
++)
1241 intrin
->const_index
[i
] = blob_read_uint32(ctx
->blob
);
1250 write_load_const(write_ctx
*ctx
, const nir_load_const_instr
*lc
)
1252 assert(lc
->def
.num_components
>= 1 && lc
->def
.num_components
<= 16);
1253 union packed_instr header
;
1256 header
.load_const
.instr_type
= lc
->instr
.type
;
1257 header
.load_const
.last_component
= lc
->def
.num_components
- 1;
1258 header
.load_const
.bit_size
= encode_bit_size_3bits(lc
->def
.bit_size
);
1259 header
.load_const
.packing
= load_const_full
;
1261 /* Try to pack 1-component constants into the 19 free bits in the header. */
1262 if (lc
->def
.num_components
== 1) {
1263 switch (lc
->def
.bit_size
) {
1265 if ((lc
->value
[0].u64
& 0x1fffffffffffull
) == 0) {
1266 /* packed_value contains high 19 bits, low bits are 0 */
1267 header
.load_const
.packing
= load_const_scalar_hi_19bits
;
1268 header
.load_const
.packed_value
= lc
->value
[0].u64
>> 45;
1269 } else if (((lc
->value
[0].i64
<< 45) >> 45) == lc
->value
[0].i64
) {
1270 /* packed_value contains low 19 bits, high bits are sign-extended */
1271 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1272 header
.load_const
.packed_value
= lc
->value
[0].u64
;
1277 if ((lc
->value
[0].u32
& 0x1fff) == 0) {
1278 header
.load_const
.packing
= load_const_scalar_hi_19bits
;
1279 header
.load_const
.packed_value
= lc
->value
[0].u32
>> 13;
1280 } else if (((lc
->value
[0].i32
<< 13) >> 13) == lc
->value
[0].i32
) {
1281 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1282 header
.load_const
.packed_value
= lc
->value
[0].u32
;
1287 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1288 header
.load_const
.packed_value
= lc
->value
[0].u16
;
1291 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1292 header
.load_const
.packed_value
= lc
->value
[0].u8
;
1295 header
.load_const
.packing
= load_const_scalar_lo_19bits_sext
;
1296 header
.load_const
.packed_value
= lc
->value
[0].b
;
1299 unreachable("invalid bit_size");
1303 blob_write_uint32(ctx
->blob
, header
.u32
);
1305 if (header
.load_const
.packing
== load_const_full
) {
1306 switch (lc
->def
.bit_size
) {
1308 blob_write_bytes(ctx
->blob
, lc
->value
,
1309 sizeof(*lc
->value
) * lc
->def
.num_components
);
1313 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1314 blob_write_uint32(ctx
->blob
, lc
->value
[i
].u32
);
1318 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1319 blob_write_uint16(ctx
->blob
, lc
->value
[i
].u16
);
1323 assert(lc
->def
.bit_size
<= 8);
1324 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1325 blob_write_uint8(ctx
->blob
, lc
->value
[i
].u8
);
1330 write_add_object(ctx
, &lc
->def
);
1333 static nir_load_const_instr
*
1334 read_load_const(read_ctx
*ctx
, union packed_instr header
)
1336 nir_load_const_instr
*lc
=
1337 nir_load_const_instr_create(ctx
->nir
, header
.load_const
.last_component
+ 1,
1338 decode_bit_size_3bits(header
.load_const
.bit_size
));
1340 switch (header
.load_const
.packing
) {
1341 case load_const_scalar_hi_19bits
:
1342 switch (lc
->def
.bit_size
) {
1344 lc
->value
[0].u64
= (uint64_t)header
.load_const
.packed_value
<< 45;
1347 lc
->value
[0].u32
= (uint64_t)header
.load_const
.packed_value
<< 13;
1350 unreachable("invalid bit_size");
1354 case load_const_scalar_lo_19bits_sext
:
1355 switch (lc
->def
.bit_size
) {
1357 lc
->value
[0].i64
= ((int64_t)header
.load_const
.packed_value
<< 45) >> 45;
1360 lc
->value
[0].i32
= ((int32_t)header
.load_const
.packed_value
<< 13) >> 13;
1363 lc
->value
[0].u16
= header
.load_const
.packed_value
;
1366 lc
->value
[0].u8
= header
.load_const
.packed_value
;
1369 lc
->value
[0].b
= header
.load_const
.packed_value
;
1372 unreachable("invalid bit_size");
1376 case load_const_full
:
1377 switch (lc
->def
.bit_size
) {
1379 blob_copy_bytes(ctx
->blob
, lc
->value
, sizeof(*lc
->value
) * lc
->def
.num_components
);
1383 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1384 lc
->value
[i
].u32
= blob_read_uint32(ctx
->blob
);
1388 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1389 lc
->value
[i
].u16
= blob_read_uint16(ctx
->blob
);
1393 assert(lc
->def
.bit_size
<= 8);
1394 for (unsigned i
= 0; i
< lc
->def
.num_components
; i
++)
1395 lc
->value
[i
].u8
= blob_read_uint8(ctx
->blob
);
1401 read_add_object(ctx
, &lc
->def
);
1406 write_ssa_undef(write_ctx
*ctx
, const nir_ssa_undef_instr
*undef
)
1408 assert(undef
->def
.num_components
>= 1 && undef
->def
.num_components
<= 16);
1410 union packed_instr header
;
1413 header
.undef
.instr_type
= undef
->instr
.type
;
1414 header
.undef
.last_component
= undef
->def
.num_components
- 1;
1415 header
.undef
.bit_size
= encode_bit_size_3bits(undef
->def
.bit_size
);
1417 blob_write_uint32(ctx
->blob
, header
.u32
);
1418 write_add_object(ctx
, &undef
->def
);
1421 static nir_ssa_undef_instr
*
1422 read_ssa_undef(read_ctx
*ctx
, union packed_instr header
)
1424 nir_ssa_undef_instr
*undef
=
1425 nir_ssa_undef_instr_create(ctx
->nir
, header
.undef
.last_component
+ 1,
1426 decode_bit_size_3bits(header
.undef
.bit_size
));
1428 read_add_object(ctx
, &undef
->def
);
1432 union packed_tex_data
{
1435 enum glsl_sampler_dim sampler_dim
:4;
1436 nir_alu_type dest_type
:8;
1437 unsigned coord_components
:3;
1438 unsigned is_array
:1;
1439 unsigned is_shadow
:1;
1440 unsigned is_new_style_shadow
:1;
1441 unsigned component
:2;
1442 unsigned texture_non_uniform
:1;
1443 unsigned sampler_non_uniform
:1;
1444 unsigned unused
:8; /* Mark unused for valgrind. */
1449 write_tex(write_ctx
*ctx
, const nir_tex_instr
*tex
)
1451 assert(tex
->num_srcs
< 16);
1452 assert(tex
->op
< 16);
1453 assert(tex
->texture_array_size
< 1024);
1455 union packed_instr header
;
1458 header
.tex
.instr_type
= tex
->instr
.type
;
1459 header
.tex
.num_srcs
= tex
->num_srcs
;
1460 header
.tex
.op
= tex
->op
;
1461 header
.tex
.texture_array_size
= tex
->texture_array_size
;
1463 write_dest(ctx
, &tex
->dest
, header
, tex
->instr
.type
);
1465 blob_write_uint32(ctx
->blob
, tex
->texture_index
);
1466 blob_write_uint32(ctx
->blob
, tex
->sampler_index
);
1467 if (tex
->op
== nir_texop_tg4
)
1468 blob_write_bytes(ctx
->blob
, tex
->tg4_offsets
, sizeof(tex
->tg4_offsets
));
1470 STATIC_ASSERT(sizeof(union packed_tex_data
) == sizeof(uint32_t));
1471 union packed_tex_data packed
= {
1472 .u
.sampler_dim
= tex
->sampler_dim
,
1473 .u
.dest_type
= tex
->dest_type
,
1474 .u
.coord_components
= tex
->coord_components
,
1475 .u
.is_array
= tex
->is_array
,
1476 .u
.is_shadow
= tex
->is_shadow
,
1477 .u
.is_new_style_shadow
= tex
->is_new_style_shadow
,
1478 .u
.component
= tex
->component
,
1479 .u
.texture_non_uniform
= tex
->texture_non_uniform
,
1480 .u
.sampler_non_uniform
= tex
->sampler_non_uniform
,
1482 blob_write_uint32(ctx
->blob
, packed
.u32
);
1484 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
1485 union packed_src src
;
1487 src
.tex
.src_type
= tex
->src
[i
].src_type
;
1488 write_src_full(ctx
, &tex
->src
[i
].src
, src
);
1492 static nir_tex_instr
*
1493 read_tex(read_ctx
*ctx
, union packed_instr header
)
1495 nir_tex_instr
*tex
= nir_tex_instr_create(ctx
->nir
, header
.tex
.num_srcs
);
1497 read_dest(ctx
, &tex
->dest
, &tex
->instr
, header
);
1499 tex
->op
= header
.tex
.op
;
1500 tex
->texture_index
= blob_read_uint32(ctx
->blob
);
1501 tex
->texture_array_size
= header
.tex
.texture_array_size
;
1502 tex
->sampler_index
= blob_read_uint32(ctx
->blob
);
1503 if (tex
->op
== nir_texop_tg4
)
1504 blob_copy_bytes(ctx
->blob
, tex
->tg4_offsets
, sizeof(tex
->tg4_offsets
));
1506 union packed_tex_data packed
;
1507 packed
.u32
= blob_read_uint32(ctx
->blob
);
1508 tex
->sampler_dim
= packed
.u
.sampler_dim
;
1509 tex
->dest_type
= packed
.u
.dest_type
;
1510 tex
->coord_components
= packed
.u
.coord_components
;
1511 tex
->is_array
= packed
.u
.is_array
;
1512 tex
->is_shadow
= packed
.u
.is_shadow
;
1513 tex
->is_new_style_shadow
= packed
.u
.is_new_style_shadow
;
1514 tex
->component
= packed
.u
.component
;
1515 tex
->texture_non_uniform
= packed
.u
.texture_non_uniform
;
1516 tex
->sampler_non_uniform
= packed
.u
.sampler_non_uniform
;
1518 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
1519 union packed_src src
= read_src(ctx
, &tex
->src
[i
].src
, &tex
->instr
);
1520 tex
->src
[i
].src_type
= src
.tex
.src_type
;
1527 write_phi(write_ctx
*ctx
, const nir_phi_instr
*phi
)
1529 union packed_instr header
;
1532 header
.phi
.instr_type
= phi
->instr
.type
;
1533 header
.phi
.num_srcs
= exec_list_length(&phi
->srcs
);
1535 /* Phi nodes are special, since they may reference SSA definitions and
1536 * basic blocks that don't exist yet. We leave two empty uint32_t's here,
1537 * and then store enough information so that a later fixup pass can fill
1538 * them in correctly.
1540 write_dest(ctx
, &phi
->dest
, header
, phi
->instr
.type
);
1542 nir_foreach_phi_src(src
, phi
) {
1543 assert(src
->src
.is_ssa
);
1544 size_t blob_offset
= blob_reserve_uint32(ctx
->blob
);
1545 ASSERTED
size_t blob_offset2
= blob_reserve_uint32(ctx
->blob
);
1546 assert(blob_offset
+ sizeof(uint32_t) == blob_offset2
);
1547 write_phi_fixup fixup
= {
1548 .blob_offset
= blob_offset
,
1549 .src
= src
->src
.ssa
,
1552 util_dynarray_append(&ctx
->phi_fixups
, write_phi_fixup
, fixup
);
1557 write_fixup_phis(write_ctx
*ctx
)
1559 util_dynarray_foreach(&ctx
->phi_fixups
, write_phi_fixup
, fixup
) {
1560 uint32_t *blob_ptr
= (uint32_t *)(ctx
->blob
->data
+ fixup
->blob_offset
);
1561 blob_ptr
[0] = write_lookup_object(ctx
, fixup
->src
);
1562 blob_ptr
[1] = write_lookup_object(ctx
, fixup
->block
);
1565 util_dynarray_clear(&ctx
->phi_fixups
);
1568 static nir_phi_instr
*
1569 read_phi(read_ctx
*ctx
, nir_block
*blk
, union packed_instr header
)
1571 nir_phi_instr
*phi
= nir_phi_instr_create(ctx
->nir
);
1573 read_dest(ctx
, &phi
->dest
, &phi
->instr
, header
);
1575 /* For similar reasons as before, we just store the index directly into the
1576 * pointer, and let a later pass resolve the phi sources.
1578 * In order to ensure that the copied sources (which are just the indices
1579 * from the blob for now) don't get inserted into the old shader's use-def
1580 * lists, we have to add the phi instruction *before* we set up its
1583 nir_instr_insert_after_block(blk
, &phi
->instr
);
1585 for (unsigned i
= 0; i
< header
.phi
.num_srcs
; i
++) {
1586 nir_phi_src
*src
= ralloc(phi
, nir_phi_src
);
1588 src
->src
.is_ssa
= true;
1589 src
->src
.ssa
= (nir_ssa_def
*)(uintptr_t) blob_read_uint32(ctx
->blob
);
1590 src
->pred
= (nir_block
*)(uintptr_t) blob_read_uint32(ctx
->blob
);
1592 /* Since we're not letting nir_insert_instr handle use/def stuff for us,
1593 * we have to set the parent_instr manually. It doesn't really matter
1594 * when we do it, so we might as well do it here.
1596 src
->src
.parent_instr
= &phi
->instr
;
1598 /* Stash it in the list of phi sources. We'll walk this list and fix up
1599 * sources at the very end of read_function_impl.
1601 list_add(&src
->src
.use_link
, &ctx
->phi_srcs
);
1603 exec_list_push_tail(&phi
->srcs
, &src
->node
);
1610 read_fixup_phis(read_ctx
*ctx
)
1612 list_for_each_entry_safe(nir_phi_src
, src
, &ctx
->phi_srcs
, src
.use_link
) {
1613 src
->pred
= read_lookup_object(ctx
, (uintptr_t)src
->pred
);
1614 src
->src
.ssa
= read_lookup_object(ctx
, (uintptr_t)src
->src
.ssa
);
1616 /* Remove from this list */
1617 list_del(&src
->src
.use_link
);
1619 list_addtail(&src
->src
.use_link
, &src
->src
.ssa
->uses
);
1621 assert(list_is_empty(&ctx
->phi_srcs
));
1625 write_jump(write_ctx
*ctx
, const nir_jump_instr
*jmp
)
1627 assert(jmp
->type
< 4);
1629 union packed_instr header
;
1632 header
.jump
.instr_type
= jmp
->instr
.type
;
1633 header
.jump
.type
= jmp
->type
;
1635 blob_write_uint32(ctx
->blob
, header
.u32
);
1638 static nir_jump_instr
*
1639 read_jump(read_ctx
*ctx
, union packed_instr header
)
1641 nir_jump_instr
*jmp
= nir_jump_instr_create(ctx
->nir
, header
.jump
.type
);
1646 write_call(write_ctx
*ctx
, const nir_call_instr
*call
)
1648 blob_write_uint32(ctx
->blob
, write_lookup_object(ctx
, call
->callee
));
1650 for (unsigned i
= 0; i
< call
->num_params
; i
++)
1651 write_src(ctx
, &call
->params
[i
]);
1654 static nir_call_instr
*
1655 read_call(read_ctx
*ctx
)
1657 nir_function
*callee
= read_object(ctx
);
1658 nir_call_instr
*call
= nir_call_instr_create(ctx
->nir
, callee
);
1660 for (unsigned i
= 0; i
< call
->num_params
; i
++)
1661 read_src(ctx
, &call
->params
[i
], call
);
1667 write_instr(write_ctx
*ctx
, const nir_instr
*instr
)
1669 /* We have only 4 bits for the instruction type. */
1670 assert(instr
->type
< 16);
1672 switch (instr
->type
) {
1673 case nir_instr_type_alu
:
1674 write_alu(ctx
, nir_instr_as_alu(instr
));
1676 case nir_instr_type_deref
:
1677 write_deref(ctx
, nir_instr_as_deref(instr
));
1679 case nir_instr_type_intrinsic
:
1680 write_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1682 case nir_instr_type_load_const
:
1683 write_load_const(ctx
, nir_instr_as_load_const(instr
));
1685 case nir_instr_type_ssa_undef
:
1686 write_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
1688 case nir_instr_type_tex
:
1689 write_tex(ctx
, nir_instr_as_tex(instr
));
1691 case nir_instr_type_phi
:
1692 write_phi(ctx
, nir_instr_as_phi(instr
));
1694 case nir_instr_type_jump
:
1695 write_jump(ctx
, nir_instr_as_jump(instr
));
1697 case nir_instr_type_call
:
1698 blob_write_uint32(ctx
->blob
, instr
->type
);
1699 write_call(ctx
, nir_instr_as_call(instr
));
1701 case nir_instr_type_parallel_copy
:
1702 unreachable("Cannot write parallel copies");
1704 unreachable("bad instr type");
1708 /* Return the number of instructions read. */
1710 read_instr(read_ctx
*ctx
, nir_block
*block
)
1712 STATIC_ASSERT(sizeof(union packed_instr
) == 4);
1713 union packed_instr header
;
1714 header
.u32
= blob_read_uint32(ctx
->blob
);
1717 switch (header
.any
.instr_type
) {
1718 case nir_instr_type_alu
:
1719 for (unsigned i
= 0; i
<= header
.alu
.num_followup_alu_sharing_header
; i
++)
1720 nir_instr_insert_after_block(block
, &read_alu(ctx
, header
)->instr
);
1721 return header
.alu
.num_followup_alu_sharing_header
+ 1;
1722 case nir_instr_type_deref
:
1723 instr
= &read_deref(ctx
, header
)->instr
;
1725 case nir_instr_type_intrinsic
:
1726 instr
= &read_intrinsic(ctx
, header
)->instr
;
1728 case nir_instr_type_load_const
:
1729 instr
= &read_load_const(ctx
, header
)->instr
;
1731 case nir_instr_type_ssa_undef
:
1732 instr
= &read_ssa_undef(ctx
, header
)->instr
;
1734 case nir_instr_type_tex
:
1735 instr
= &read_tex(ctx
, header
)->instr
;
1737 case nir_instr_type_phi
:
1738 /* Phi instructions are a bit of a special case when reading because we
1739 * don't want inserting the instruction to automatically handle use/defs
1740 * for us. Instead, we need to wait until all the blocks/instructions
1741 * are read so that we can set their sources up.
1743 read_phi(ctx
, block
, header
);
1745 case nir_instr_type_jump
:
1746 instr
= &read_jump(ctx
, header
)->instr
;
1748 case nir_instr_type_call
:
1749 instr
= &read_call(ctx
)->instr
;
1751 case nir_instr_type_parallel_copy
:
1752 unreachable("Cannot read parallel copies");
1754 unreachable("bad instr type");
1757 nir_instr_insert_after_block(block
, instr
);
1762 write_block(write_ctx
*ctx
, const nir_block
*block
)
1764 write_add_object(ctx
, block
);
1765 blob_write_uint32(ctx
->blob
, exec_list_length(&block
->instr_list
));
1767 ctx
->last_instr_type
= ~0;
1768 ctx
->last_alu_header_offset
= 0;
1770 nir_foreach_instr(instr
, block
) {
1771 write_instr(ctx
, instr
);
1772 ctx
->last_instr_type
= instr
->type
;
1777 read_block(read_ctx
*ctx
, struct exec_list
*cf_list
)
1779 /* Don't actually create a new block. Just use the one from the tail of
1780 * the list. NIR guarantees that the tail of the list is a block and that
1781 * no two blocks are side-by-side in the IR; It should be empty.
1784 exec_node_data(nir_block
, exec_list_get_tail(cf_list
), cf_node
.node
);
1786 read_add_object(ctx
, block
);
1787 unsigned num_instrs
= blob_read_uint32(ctx
->blob
);
1788 for (unsigned i
= 0; i
< num_instrs
;) {
1789 i
+= read_instr(ctx
, block
);
1794 write_cf_list(write_ctx
*ctx
, const struct exec_list
*cf_list
);
1797 read_cf_list(read_ctx
*ctx
, struct exec_list
*cf_list
);
1800 write_if(write_ctx
*ctx
, nir_if
*nif
)
1802 write_src(ctx
, &nif
->condition
);
1804 write_cf_list(ctx
, &nif
->then_list
);
1805 write_cf_list(ctx
, &nif
->else_list
);
1809 read_if(read_ctx
*ctx
, struct exec_list
*cf_list
)
1811 nir_if
*nif
= nir_if_create(ctx
->nir
);
1813 read_src(ctx
, &nif
->condition
, nif
);
1815 nir_cf_node_insert_end(cf_list
, &nif
->cf_node
);
1817 read_cf_list(ctx
, &nif
->then_list
);
1818 read_cf_list(ctx
, &nif
->else_list
);
1822 write_loop(write_ctx
*ctx
, nir_loop
*loop
)
1824 write_cf_list(ctx
, &loop
->body
);
1828 read_loop(read_ctx
*ctx
, struct exec_list
*cf_list
)
1830 nir_loop
*loop
= nir_loop_create(ctx
->nir
);
1832 nir_cf_node_insert_end(cf_list
, &loop
->cf_node
);
1834 read_cf_list(ctx
, &loop
->body
);
1838 write_cf_node(write_ctx
*ctx
, nir_cf_node
*cf
)
1840 blob_write_uint32(ctx
->blob
, cf
->type
);
1843 case nir_cf_node_block
:
1844 write_block(ctx
, nir_cf_node_as_block(cf
));
1846 case nir_cf_node_if
:
1847 write_if(ctx
, nir_cf_node_as_if(cf
));
1849 case nir_cf_node_loop
:
1850 write_loop(ctx
, nir_cf_node_as_loop(cf
));
1853 unreachable("bad cf type");
1858 read_cf_node(read_ctx
*ctx
, struct exec_list
*list
)
1860 nir_cf_node_type type
= blob_read_uint32(ctx
->blob
);
1863 case nir_cf_node_block
:
1864 read_block(ctx
, list
);
1866 case nir_cf_node_if
:
1869 case nir_cf_node_loop
:
1870 read_loop(ctx
, list
);
1873 unreachable("bad cf type");
1878 write_cf_list(write_ctx
*ctx
, const struct exec_list
*cf_list
)
1880 blob_write_uint32(ctx
->blob
, exec_list_length(cf_list
));
1881 foreach_list_typed(nir_cf_node
, cf
, node
, cf_list
) {
1882 write_cf_node(ctx
, cf
);
1887 read_cf_list(read_ctx
*ctx
, struct exec_list
*cf_list
)
1889 uint32_t num_cf_nodes
= blob_read_uint32(ctx
->blob
);
1890 for (unsigned i
= 0; i
< num_cf_nodes
; i
++)
1891 read_cf_node(ctx
, cf_list
);
1895 write_function_impl(write_ctx
*ctx
, const nir_function_impl
*fi
)
1897 write_var_list(ctx
, &fi
->locals
);
1898 write_reg_list(ctx
, &fi
->registers
);
1899 blob_write_uint32(ctx
->blob
, fi
->reg_alloc
);
1901 write_cf_list(ctx
, &fi
->body
);
1902 write_fixup_phis(ctx
);
1905 static nir_function_impl
*
1906 read_function_impl(read_ctx
*ctx
, nir_function
*fxn
)
1908 nir_function_impl
*fi
= nir_function_impl_create_bare(ctx
->nir
);
1911 read_var_list(ctx
, &fi
->locals
);
1912 read_reg_list(ctx
, &fi
->registers
);
1913 fi
->reg_alloc
= blob_read_uint32(ctx
->blob
);
1915 read_cf_list(ctx
, &fi
->body
);
1916 read_fixup_phis(ctx
);
1918 fi
->valid_metadata
= 0;
1924 write_function(write_ctx
*ctx
, const nir_function
*fxn
)
1926 uint32_t flags
= fxn
->is_entrypoint
;
1931 blob_write_uint32(ctx
->blob
, flags
);
1933 blob_write_string(ctx
->blob
, fxn
->name
);
1935 write_add_object(ctx
, fxn
);
1937 blob_write_uint32(ctx
->blob
, fxn
->num_params
);
1938 for (unsigned i
= 0; i
< fxn
->num_params
; i
++) {
1940 ((uint32_t)fxn
->params
[i
].num_components
) |
1941 ((uint32_t)fxn
->params
[i
].bit_size
) << 8;
1942 blob_write_uint32(ctx
->blob
, val
);
1945 /* At first glance, it looks like we should write the function_impl here.
1946 * However, call instructions need to be able to reference at least the
1947 * function and those will get processed as we write the function_impls.
1948 * We stop here and write function_impls as a second pass.
1953 read_function(read_ctx
*ctx
)
1955 uint32_t flags
= blob_read_uint32(ctx
->blob
);
1956 bool has_name
= flags
& 0x2;
1957 char *name
= has_name
? blob_read_string(ctx
->blob
) : NULL
;
1959 nir_function
*fxn
= nir_function_create(ctx
->nir
, name
);
1961 read_add_object(ctx
, fxn
);
1963 fxn
->num_params
= blob_read_uint32(ctx
->blob
);
1964 fxn
->params
= ralloc_array(fxn
, nir_parameter
, fxn
->num_params
);
1965 for (unsigned i
= 0; i
< fxn
->num_params
; i
++) {
1966 uint32_t val
= blob_read_uint32(ctx
->blob
);
1967 fxn
->params
[i
].num_components
= val
& 0xff;
1968 fxn
->params
[i
].bit_size
= (val
>> 8) & 0xff;
1971 fxn
->is_entrypoint
= flags
& 0x1;
1973 fxn
->impl
= NIR_SERIALIZE_FUNC_HAS_IMPL
;
1977 * Serialize NIR into a binary blob.
1979 * \param strip Don't serialize information only useful for debugging,
1980 * such as variable names, making cache hits from similar
1981 * shaders more likely.
1984 nir_serialize(struct blob
*blob
, const nir_shader
*nir
, bool strip
)
1986 write_ctx ctx
= {0};
1987 ctx
.remap_table
= _mesa_pointer_hash_table_create(NULL
);
1991 util_dynarray_init(&ctx
.phi_fixups
, NULL
);
1993 size_t idx_size_offset
= blob_reserve_uint32(blob
);
1995 struct shader_info info
= nir
->info
;
1996 uint32_t strings
= 0;
1997 if (!strip
&& info
.name
)
1999 if (!strip
&& info
.label
)
2001 blob_write_uint32(blob
, strings
);
2002 if (!strip
&& info
.name
)
2003 blob_write_string(blob
, info
.name
);
2004 if (!strip
&& info
.label
)
2005 blob_write_string(blob
, info
.label
);
2006 info
.name
= info
.label
= NULL
;
2007 blob_write_bytes(blob
, (uint8_t *) &info
, sizeof(info
));
2009 write_var_list(&ctx
, &nir
->uniforms
);
2010 write_var_list(&ctx
, &nir
->inputs
);
2011 write_var_list(&ctx
, &nir
->outputs
);
2012 write_var_list(&ctx
, &nir
->shared
);
2013 write_var_list(&ctx
, &nir
->globals
);
2014 write_var_list(&ctx
, &nir
->system_values
);
2016 blob_write_uint32(blob
, nir
->num_inputs
);
2017 blob_write_uint32(blob
, nir
->num_uniforms
);
2018 blob_write_uint32(blob
, nir
->num_outputs
);
2019 blob_write_uint32(blob
, nir
->num_shared
);
2020 blob_write_uint32(blob
, nir
->scratch_size
);
2022 blob_write_uint32(blob
, exec_list_length(&nir
->functions
));
2023 nir_foreach_function(fxn
, nir
) {
2024 write_function(&ctx
, fxn
);
2027 nir_foreach_function(fxn
, nir
) {
2029 write_function_impl(&ctx
, fxn
->impl
);
2032 blob_write_uint32(blob
, nir
->constant_data_size
);
2033 if (nir
->constant_data_size
> 0)
2034 blob_write_bytes(blob
, nir
->constant_data
, nir
->constant_data_size
);
2036 *(uint32_t *)(blob
->data
+ idx_size_offset
) = ctx
.next_idx
;
2038 _mesa_hash_table_destroy(ctx
.remap_table
, NULL
);
2039 util_dynarray_fini(&ctx
.phi_fixups
);
2043 nir_deserialize(void *mem_ctx
,
2044 const struct nir_shader_compiler_options
*options
,
2045 struct blob_reader
*blob
)
2049 list_inithead(&ctx
.phi_srcs
);
2050 ctx
.idx_table_len
= blob_read_uint32(blob
);
2051 ctx
.idx_table
= calloc(ctx
.idx_table_len
, sizeof(uintptr_t));
2053 uint32_t strings
= blob_read_uint32(blob
);
2054 char *name
= (strings
& 0x1) ? blob_read_string(blob
) : NULL
;
2055 char *label
= (strings
& 0x2) ? blob_read_string(blob
) : NULL
;
2057 struct shader_info info
;
2058 blob_copy_bytes(blob
, (uint8_t *) &info
, sizeof(info
));
2060 ctx
.nir
= nir_shader_create(mem_ctx
, info
.stage
, options
, NULL
);
2062 info
.name
= name
? ralloc_strdup(ctx
.nir
, name
) : NULL
;
2063 info
.label
= label
? ralloc_strdup(ctx
.nir
, label
) : NULL
;
2065 ctx
.nir
->info
= info
;
2067 read_var_list(&ctx
, &ctx
.nir
->uniforms
);
2068 read_var_list(&ctx
, &ctx
.nir
->inputs
);
2069 read_var_list(&ctx
, &ctx
.nir
->outputs
);
2070 read_var_list(&ctx
, &ctx
.nir
->shared
);
2071 read_var_list(&ctx
, &ctx
.nir
->globals
);
2072 read_var_list(&ctx
, &ctx
.nir
->system_values
);
2074 ctx
.nir
->num_inputs
= blob_read_uint32(blob
);
2075 ctx
.nir
->num_uniforms
= blob_read_uint32(blob
);
2076 ctx
.nir
->num_outputs
= blob_read_uint32(blob
);
2077 ctx
.nir
->num_shared
= blob_read_uint32(blob
);
2078 ctx
.nir
->scratch_size
= blob_read_uint32(blob
);
2080 unsigned num_functions
= blob_read_uint32(blob
);
2081 for (unsigned i
= 0; i
< num_functions
; i
++)
2082 read_function(&ctx
);
2084 nir_foreach_function(fxn
, ctx
.nir
) {
2085 if (fxn
->impl
== NIR_SERIALIZE_FUNC_HAS_IMPL
)
2086 fxn
->impl
= read_function_impl(&ctx
, fxn
);
2089 ctx
.nir
->constant_data_size
= blob_read_uint32(blob
);
2090 if (ctx
.nir
->constant_data_size
> 0) {
2091 ctx
.nir
->constant_data
=
2092 ralloc_size(ctx
.nir
, ctx
.nir
->constant_data_size
);
2093 blob_copy_bytes(blob
, ctx
.nir
->constant_data
,
2094 ctx
.nir
->constant_data_size
);
2097 free(ctx
.idx_table
);
2103 nir_shader_serialize_deserialize(nir_shader
*shader
)
2105 const struct nir_shader_compiler_options
*options
= shader
->options
;
2109 nir_serialize(&writer
, shader
, false);
2111 /* Delete all of dest's ralloc children but leave dest alone */
2112 void *dead_ctx
= ralloc_context(NULL
);
2113 ralloc_adopt(dead_ctx
, shader
);
2114 ralloc_free(dead_ctx
);
2116 dead_ctx
= ralloc_context(NULL
);
2118 struct blob_reader reader
;
2119 blob_reader_init(&reader
, writer
.data
, writer
.size
);
2120 nir_shader
*copy
= nir_deserialize(dead_ctx
, options
, &reader
);
2122 blob_finish(&writer
);
2124 nir_shader_replace(shader
, copy
);
2125 ralloc_free(dead_ctx
);