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"
35 const nir_shader
*nir
;
39 /* maps pointer to index */
40 struct hash_table
*remap_table
;
42 /* the next index to assign to a NIR in-memory object */
45 /* Array of write_phi_fixup structs representing phi sources that need to
46 * be resolved in the second pass.
48 struct util_dynarray phi_fixups
;
54 struct blob_reader
*blob
;
56 /* the next index to assign to a NIR in-memory object */
59 /* The length of the index -> object table */
60 uintptr_t idx_table_len
;
62 /* map from index to deserialized pointer */
65 /* List of phi sources. */
66 struct list_head phi_srcs
;
71 write_add_object(write_ctx
*ctx
, const void *obj
)
73 uintptr_t index
= ctx
->next_idx
++;
74 _mesa_hash_table_insert(ctx
->remap_table
, obj
, (void *) index
);
78 write_lookup_object(write_ctx
*ctx
, const void *obj
)
80 struct hash_entry
*entry
= _mesa_hash_table_search(ctx
->remap_table
, obj
);
82 return (uintptr_t) entry
->data
;
86 write_object(write_ctx
*ctx
, const void *obj
)
88 blob_write_intptr(ctx
->blob
, write_lookup_object(ctx
, obj
));
92 read_add_object(read_ctx
*ctx
, void *obj
)
94 assert(ctx
->next_idx
< ctx
->idx_table_len
);
95 ctx
->idx_table
[ctx
->next_idx
++] = obj
;
99 read_lookup_object(read_ctx
*ctx
, uintptr_t idx
)
101 assert(idx
< ctx
->idx_table_len
);
102 return ctx
->idx_table
[idx
];
106 read_object(read_ctx
*ctx
)
108 return read_lookup_object(ctx
, blob_read_intptr(ctx
->blob
));
112 write_constant(write_ctx
*ctx
, const nir_constant
*c
)
114 blob_write_bytes(ctx
->blob
, c
->values
, sizeof(c
->values
));
115 blob_write_uint32(ctx
->blob
, c
->num_elements
);
116 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
117 write_constant(ctx
, c
->elements
[i
]);
120 static nir_constant
*
121 read_constant(read_ctx
*ctx
, nir_variable
*nvar
)
123 nir_constant
*c
= ralloc(nvar
, nir_constant
);
125 blob_copy_bytes(ctx
->blob
, (uint8_t *)c
->values
, sizeof(c
->values
));
126 c
->num_elements
= blob_read_uint32(ctx
->blob
);
127 c
->elements
= ralloc_array(nvar
, nir_constant
*, c
->num_elements
);
128 for (unsigned i
= 0; i
< c
->num_elements
; i
++)
129 c
->elements
[i
] = read_constant(ctx
, nvar
);
135 write_variable(write_ctx
*ctx
, const nir_variable
*var
)
137 write_add_object(ctx
, var
);
138 encode_type_to_blob(ctx
->blob
, var
->type
);
139 blob_write_uint32(ctx
->blob
, !!(var
->name
));
141 blob_write_string(ctx
->blob
, var
->name
);
142 blob_write_bytes(ctx
->blob
, (uint8_t *) &var
->data
, sizeof(var
->data
));
143 blob_write_uint32(ctx
->blob
, var
->num_state_slots
);
144 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
145 blob_write_bytes(ctx
->blob
, &var
->state_slots
[i
],
146 sizeof(var
->state_slots
[i
]));
148 blob_write_uint32(ctx
->blob
, !!(var
->constant_initializer
));
149 if (var
->constant_initializer
)
150 write_constant(ctx
, var
->constant_initializer
);
151 blob_write_uint32(ctx
->blob
, !!(var
->interface_type
));
152 if (var
->interface_type
)
153 encode_type_to_blob(ctx
->blob
, var
->interface_type
);
154 blob_write_uint32(ctx
->blob
, var
->num_members
);
155 if (var
->num_members
> 0) {
156 blob_write_bytes(ctx
->blob
, (uint8_t *) var
->members
,
157 var
->num_members
* sizeof(*var
->members
));
161 static nir_variable
*
162 read_variable(read_ctx
*ctx
)
164 nir_variable
*var
= rzalloc(ctx
->nir
, nir_variable
);
165 read_add_object(ctx
, var
);
167 var
->type
= decode_type_from_blob(ctx
->blob
);
168 bool has_name
= blob_read_uint32(ctx
->blob
);
170 const char *name
= blob_read_string(ctx
->blob
);
171 var
->name
= ralloc_strdup(var
, name
);
175 blob_copy_bytes(ctx
->blob
, (uint8_t *) &var
->data
, sizeof(var
->data
));
176 var
->num_state_slots
= blob_read_uint32(ctx
->blob
);
177 if (var
->num_state_slots
!= 0) {
178 var
->state_slots
= ralloc_array(var
, nir_state_slot
,
179 var
->num_state_slots
);
180 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
181 blob_copy_bytes(ctx
->blob
, &var
->state_slots
[i
],
182 sizeof(var
->state_slots
[i
]));
185 bool has_const_initializer
= blob_read_uint32(ctx
->blob
);
186 if (has_const_initializer
)
187 var
->constant_initializer
= read_constant(ctx
, var
);
189 var
->constant_initializer
= NULL
;
190 bool has_interface_type
= blob_read_uint32(ctx
->blob
);
191 if (has_interface_type
)
192 var
->interface_type
= decode_type_from_blob(ctx
->blob
);
194 var
->interface_type
= NULL
;
195 var
->num_members
= blob_read_uint32(ctx
->blob
);
196 if (var
->num_members
> 0) {
197 var
->members
= ralloc_array(var
, struct nir_variable_data
,
199 blob_copy_bytes(ctx
->blob
, (uint8_t *) var
->members
,
200 var
->num_members
* sizeof(*var
->members
));
207 write_var_list(write_ctx
*ctx
, const struct exec_list
*src
)
209 blob_write_uint32(ctx
->blob
, exec_list_length(src
));
210 foreach_list_typed(nir_variable
, var
, node
, src
) {
211 write_variable(ctx
, var
);
216 read_var_list(read_ctx
*ctx
, struct exec_list
*dst
)
218 exec_list_make_empty(dst
);
219 unsigned num_vars
= blob_read_uint32(ctx
->blob
);
220 for (unsigned i
= 0; i
< num_vars
; i
++) {
221 nir_variable
*var
= read_variable(ctx
);
222 exec_list_push_tail(dst
, &var
->node
);
227 write_register(write_ctx
*ctx
, const nir_register
*reg
)
229 write_add_object(ctx
, reg
);
230 blob_write_uint32(ctx
->blob
, reg
->num_components
);
231 blob_write_uint32(ctx
->blob
, reg
->bit_size
);
232 blob_write_uint32(ctx
->blob
, reg
->num_array_elems
);
233 blob_write_uint32(ctx
->blob
, reg
->index
);
234 blob_write_uint32(ctx
->blob
, !!(reg
->name
));
236 blob_write_string(ctx
->blob
, reg
->name
);
239 static nir_register
*
240 read_register(read_ctx
*ctx
)
242 nir_register
*reg
= ralloc(ctx
->nir
, nir_register
);
243 read_add_object(ctx
, reg
);
244 reg
->num_components
= blob_read_uint32(ctx
->blob
);
245 reg
->bit_size
= blob_read_uint32(ctx
->blob
);
246 reg
->num_array_elems
= blob_read_uint32(ctx
->blob
);
247 reg
->index
= blob_read_uint32(ctx
->blob
);
248 bool has_name
= blob_read_uint32(ctx
->blob
);
250 const char *name
= blob_read_string(ctx
->blob
);
251 reg
->name
= ralloc_strdup(reg
, name
);
256 list_inithead(®
->uses
);
257 list_inithead(®
->defs
);
258 list_inithead(®
->if_uses
);
264 write_reg_list(write_ctx
*ctx
, const struct exec_list
*src
)
266 blob_write_uint32(ctx
->blob
, exec_list_length(src
));
267 foreach_list_typed(nir_register
, reg
, node
, src
)
268 write_register(ctx
, reg
);
272 read_reg_list(read_ctx
*ctx
, struct exec_list
*dst
)
274 exec_list_make_empty(dst
);
275 unsigned num_regs
= blob_read_uint32(ctx
->blob
);
276 for (unsigned i
= 0; i
< num_regs
; i
++) {
277 nir_register
*reg
= read_register(ctx
);
278 exec_list_push_tail(dst
, ®
->node
);
283 write_src(write_ctx
*ctx
, const nir_src
*src
)
285 /* Since sources are very frequent, we try to save some space when storing
286 * them. In particular, we store whether the source is a register and
287 * whether the register has an indirect index in the low two bits. We can
288 * assume that the high two bits of the index are zero, since otherwise our
289 * address space would've been exhausted allocating the remap table!
292 uintptr_t idx
= write_lookup_object(ctx
, src
->ssa
) << 2;
294 blob_write_intptr(ctx
->blob
, idx
);
296 uintptr_t idx
= write_lookup_object(ctx
, src
->reg
.reg
) << 2;
297 if (src
->reg
.indirect
)
299 blob_write_intptr(ctx
->blob
, idx
);
300 blob_write_uint32(ctx
->blob
, src
->reg
.base_offset
);
301 if (src
->reg
.indirect
) {
302 write_src(ctx
, src
->reg
.indirect
);
308 read_src(read_ctx
*ctx
, nir_src
*src
, void *mem_ctx
)
310 uintptr_t val
= blob_read_intptr(ctx
->blob
);
311 uintptr_t idx
= val
>> 2;
312 src
->is_ssa
= val
& 0x1;
314 src
->ssa
= read_lookup_object(ctx
, idx
);
316 bool is_indirect
= val
& 0x2;
317 src
->reg
.reg
= read_lookup_object(ctx
, idx
);
318 src
->reg
.base_offset
= blob_read_uint32(ctx
->blob
);
320 src
->reg
.indirect
= ralloc(mem_ctx
, nir_src
);
321 read_src(ctx
, src
->reg
.indirect
, mem_ctx
);
323 src
->reg
.indirect
= NULL
;
329 write_dest(write_ctx
*ctx
, const nir_dest
*dst
)
331 uint32_t val
= dst
->is_ssa
;
333 val
|= !!(dst
->ssa
.name
) << 1;
334 val
|= dst
->ssa
.num_components
<< 2;
335 val
|= dst
->ssa
.bit_size
<< 5;
337 val
|= !!(dst
->reg
.indirect
) << 1;
339 blob_write_uint32(ctx
->blob
, val
);
341 write_add_object(ctx
, &dst
->ssa
);
343 blob_write_string(ctx
->blob
, dst
->ssa
.name
);
345 blob_write_intptr(ctx
->blob
, write_lookup_object(ctx
, dst
->reg
.reg
));
346 blob_write_uint32(ctx
->blob
, dst
->reg
.base_offset
);
347 if (dst
->reg
.indirect
)
348 write_src(ctx
, dst
->reg
.indirect
);
353 read_dest(read_ctx
*ctx
, nir_dest
*dst
, nir_instr
*instr
)
355 uint32_t val
= blob_read_uint32(ctx
->blob
);
356 bool is_ssa
= val
& 0x1;
358 bool has_name
= val
& 0x2;
359 unsigned num_components
= (val
>> 2) & 0x7;
360 unsigned bit_size
= val
>> 5;
361 char *name
= has_name
? blob_read_string(ctx
->blob
) : NULL
;
362 nir_ssa_dest_init(instr
, dst
, num_components
, bit_size
, name
);
363 read_add_object(ctx
, &dst
->ssa
);
365 bool is_indirect
= val
& 0x2;
366 dst
->reg
.reg
= read_object(ctx
);
367 dst
->reg
.base_offset
= blob_read_uint32(ctx
->blob
);
369 dst
->reg
.indirect
= ralloc(instr
, nir_src
);
370 read_src(ctx
, dst
->reg
.indirect
, instr
);
376 write_alu(write_ctx
*ctx
, const nir_alu_instr
*alu
)
378 blob_write_uint32(ctx
->blob
, alu
->op
);
379 uint32_t flags
= alu
->exact
;
380 flags
|= alu
->no_signed_wrap
<< 1;
381 flags
|= alu
->no_unsigned_wrap
<< 2;
382 flags
|= alu
->dest
.saturate
<< 3;
383 flags
|= alu
->dest
.write_mask
<< 4;
384 blob_write_uint32(ctx
->blob
, flags
);
386 write_dest(ctx
, &alu
->dest
.dest
);
388 for (unsigned i
= 0; i
< nir_op_infos
[alu
->op
].num_inputs
; i
++) {
389 write_src(ctx
, &alu
->src
[i
].src
);
390 flags
= alu
->src
[i
].negate
;
391 flags
|= alu
->src
[i
].abs
<< 1;
392 for (unsigned j
= 0; j
< 4; j
++)
393 flags
|= alu
->src
[i
].swizzle
[j
] << (2 + 2 * j
);
394 blob_write_uint32(ctx
->blob
, flags
);
398 static nir_alu_instr
*
399 read_alu(read_ctx
*ctx
)
401 nir_op op
= blob_read_uint32(ctx
->blob
);
402 nir_alu_instr
*alu
= nir_alu_instr_create(ctx
->nir
, op
);
404 uint32_t flags
= blob_read_uint32(ctx
->blob
);
405 alu
->exact
= flags
& 1;
406 alu
->no_signed_wrap
= flags
& 2;
407 alu
->no_unsigned_wrap
= flags
& 4;
408 alu
->dest
.saturate
= flags
& 8;
409 alu
->dest
.write_mask
= flags
>> 4;
411 read_dest(ctx
, &alu
->dest
.dest
, &alu
->instr
);
413 for (unsigned i
= 0; i
< nir_op_infos
[op
].num_inputs
; i
++) {
414 read_src(ctx
, &alu
->src
[i
].src
, &alu
->instr
);
415 flags
= blob_read_uint32(ctx
->blob
);
416 alu
->src
[i
].negate
= flags
& 1;
417 alu
->src
[i
].abs
= flags
& 2;
418 for (unsigned j
= 0; j
< 4; j
++)
419 alu
->src
[i
].swizzle
[j
] = (flags
>> (2 * j
+ 2)) & 3;
426 write_deref(write_ctx
*ctx
, const nir_deref_instr
*deref
)
428 blob_write_uint32(ctx
->blob
, deref
->deref_type
);
430 blob_write_uint32(ctx
->blob
, deref
->mode
);
431 encode_type_to_blob(ctx
->blob
, deref
->type
);
433 write_dest(ctx
, &deref
->dest
);
435 if (deref
->deref_type
== nir_deref_type_var
) {
436 write_object(ctx
, deref
->var
);
440 write_src(ctx
, &deref
->parent
);
442 switch (deref
->deref_type
) {
443 case nir_deref_type_struct
:
444 blob_write_uint32(ctx
->blob
, deref
->strct
.index
);
447 case nir_deref_type_array
:
448 case nir_deref_type_ptr_as_array
:
449 write_src(ctx
, &deref
->arr
.index
);
452 case nir_deref_type_cast
:
453 blob_write_uint32(ctx
->blob
, deref
->cast
.ptr_stride
);
456 case nir_deref_type_array_wildcard
:
461 unreachable("Invalid deref type");
465 static nir_deref_instr
*
466 read_deref(read_ctx
*ctx
)
468 nir_deref_type deref_type
= blob_read_uint32(ctx
->blob
);
469 nir_deref_instr
*deref
= nir_deref_instr_create(ctx
->nir
, deref_type
);
471 deref
->mode
= blob_read_uint32(ctx
->blob
);
472 deref
->type
= decode_type_from_blob(ctx
->blob
);
474 read_dest(ctx
, &deref
->dest
, &deref
->instr
);
476 if (deref_type
== nir_deref_type_var
) {
477 deref
->var
= read_object(ctx
);
481 read_src(ctx
, &deref
->parent
, &deref
->instr
);
483 switch (deref
->deref_type
) {
484 case nir_deref_type_struct
:
485 deref
->strct
.index
= blob_read_uint32(ctx
->blob
);
488 case nir_deref_type_array
:
489 case nir_deref_type_ptr_as_array
:
490 read_src(ctx
, &deref
->arr
.index
, &deref
->instr
);
493 case nir_deref_type_cast
:
494 deref
->cast
.ptr_stride
= blob_read_uint32(ctx
->blob
);
497 case nir_deref_type_array_wildcard
:
502 unreachable("Invalid deref type");
509 write_intrinsic(write_ctx
*ctx
, const nir_intrinsic_instr
*intrin
)
511 blob_write_uint32(ctx
->blob
, intrin
->intrinsic
);
513 unsigned num_srcs
= nir_intrinsic_infos
[intrin
->intrinsic
].num_srcs
;
514 unsigned num_indices
= nir_intrinsic_infos
[intrin
->intrinsic
].num_indices
;
516 blob_write_uint32(ctx
->blob
, intrin
->num_components
);
518 if (nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
)
519 write_dest(ctx
, &intrin
->dest
);
521 for (unsigned i
= 0; i
< num_srcs
; i
++)
522 write_src(ctx
, &intrin
->src
[i
]);
524 for (unsigned i
= 0; i
< num_indices
; i
++)
525 blob_write_uint32(ctx
->blob
, intrin
->const_index
[i
]);
528 static nir_intrinsic_instr
*
529 read_intrinsic(read_ctx
*ctx
)
531 nir_intrinsic_op op
= blob_read_uint32(ctx
->blob
);
533 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(ctx
->nir
, op
);
535 unsigned num_srcs
= nir_intrinsic_infos
[op
].num_srcs
;
536 unsigned num_indices
= nir_intrinsic_infos
[op
].num_indices
;
538 intrin
->num_components
= blob_read_uint32(ctx
->blob
);
540 if (nir_intrinsic_infos
[op
].has_dest
)
541 read_dest(ctx
, &intrin
->dest
, &intrin
->instr
);
543 for (unsigned i
= 0; i
< num_srcs
; i
++)
544 read_src(ctx
, &intrin
->src
[i
], &intrin
->instr
);
546 for (unsigned i
= 0; i
< num_indices
; i
++)
547 intrin
->const_index
[i
] = blob_read_uint32(ctx
->blob
);
553 write_load_const(write_ctx
*ctx
, const nir_load_const_instr
*lc
)
555 uint32_t val
= lc
->def
.num_components
;
556 val
|= lc
->def
.bit_size
<< 3;
557 blob_write_uint32(ctx
->blob
, val
);
558 blob_write_bytes(ctx
->blob
, lc
->value
, sizeof(*lc
->value
) * lc
->def
.num_components
);
559 write_add_object(ctx
, &lc
->def
);
562 static nir_load_const_instr
*
563 read_load_const(read_ctx
*ctx
)
565 uint32_t val
= blob_read_uint32(ctx
->blob
);
567 nir_load_const_instr
*lc
=
568 nir_load_const_instr_create(ctx
->nir
, val
& 0x7, val
>> 3);
570 blob_copy_bytes(ctx
->blob
, lc
->value
, sizeof(*lc
->value
) * lc
->def
.num_components
);
571 read_add_object(ctx
, &lc
->def
);
576 write_ssa_undef(write_ctx
*ctx
, const nir_ssa_undef_instr
*undef
)
578 uint32_t val
= undef
->def
.num_components
;
579 val
|= undef
->def
.bit_size
<< 3;
580 blob_write_uint32(ctx
->blob
, val
);
581 write_add_object(ctx
, &undef
->def
);
584 static nir_ssa_undef_instr
*
585 read_ssa_undef(read_ctx
*ctx
)
587 uint32_t val
= blob_read_uint32(ctx
->blob
);
589 nir_ssa_undef_instr
*undef
=
590 nir_ssa_undef_instr_create(ctx
->nir
, val
& 0x7, val
>> 3);
592 read_add_object(ctx
, &undef
->def
);
596 union packed_tex_data
{
599 enum glsl_sampler_dim sampler_dim
:4;
600 nir_alu_type dest_type
:8;
601 unsigned coord_components
:3;
603 unsigned is_shadow
:1;
604 unsigned is_new_style_shadow
:1;
605 unsigned component
:2;
606 unsigned unused
:10; /* Mark unused for valgrind. */
611 write_tex(write_ctx
*ctx
, const nir_tex_instr
*tex
)
613 blob_write_uint32(ctx
->blob
, tex
->num_srcs
);
614 blob_write_uint32(ctx
->blob
, tex
->op
);
615 blob_write_uint32(ctx
->blob
, tex
->texture_index
);
616 blob_write_uint32(ctx
->blob
, tex
->texture_array_size
);
617 blob_write_uint32(ctx
->blob
, tex
->sampler_index
);
618 blob_write_bytes(ctx
->blob
, tex
->tg4_offsets
, sizeof(tex
->tg4_offsets
));
620 STATIC_ASSERT(sizeof(union packed_tex_data
) == sizeof(uint32_t));
621 union packed_tex_data packed
= {
622 .u
.sampler_dim
= tex
->sampler_dim
,
623 .u
.dest_type
= tex
->dest_type
,
624 .u
.coord_components
= tex
->coord_components
,
625 .u
.is_array
= tex
->is_array
,
626 .u
.is_shadow
= tex
->is_shadow
,
627 .u
.is_new_style_shadow
= tex
->is_new_style_shadow
,
628 .u
.component
= tex
->component
,
630 blob_write_uint32(ctx
->blob
, packed
.u32
);
632 write_dest(ctx
, &tex
->dest
);
633 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
634 blob_write_uint32(ctx
->blob
, tex
->src
[i
].src_type
);
635 write_src(ctx
, &tex
->src
[i
].src
);
639 static nir_tex_instr
*
640 read_tex(read_ctx
*ctx
)
642 unsigned num_srcs
= blob_read_uint32(ctx
->blob
);
643 nir_tex_instr
*tex
= nir_tex_instr_create(ctx
->nir
, num_srcs
);
645 tex
->op
= blob_read_uint32(ctx
->blob
);
646 tex
->texture_index
= blob_read_uint32(ctx
->blob
);
647 tex
->texture_array_size
= blob_read_uint32(ctx
->blob
);
648 tex
->sampler_index
= blob_read_uint32(ctx
->blob
);
649 blob_copy_bytes(ctx
->blob
, tex
->tg4_offsets
, sizeof(tex
->tg4_offsets
));
651 union packed_tex_data packed
;
652 packed
.u32
= blob_read_uint32(ctx
->blob
);
653 tex
->sampler_dim
= packed
.u
.sampler_dim
;
654 tex
->dest_type
= packed
.u
.dest_type
;
655 tex
->coord_components
= packed
.u
.coord_components
;
656 tex
->is_array
= packed
.u
.is_array
;
657 tex
->is_shadow
= packed
.u
.is_shadow
;
658 tex
->is_new_style_shadow
= packed
.u
.is_new_style_shadow
;
659 tex
->component
= packed
.u
.component
;
661 read_dest(ctx
, &tex
->dest
, &tex
->instr
);
662 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
663 tex
->src
[i
].src_type
= blob_read_uint32(ctx
->blob
);
664 read_src(ctx
, &tex
->src
[i
].src
, &tex
->instr
);
671 write_phi(write_ctx
*ctx
, const nir_phi_instr
*phi
)
673 /* Phi nodes are special, since they may reference SSA definitions and
674 * basic blocks that don't exist yet. We leave two empty uintptr_t's here,
675 * and then store enough information so that a later fixup pass can fill
678 write_dest(ctx
, &phi
->dest
);
680 blob_write_uint32(ctx
->blob
, exec_list_length(&phi
->srcs
));
682 nir_foreach_phi_src(src
, phi
) {
683 assert(src
->src
.is_ssa
);
684 size_t blob_offset
= blob_reserve_intptr(ctx
->blob
);
685 ASSERTED
size_t blob_offset2
= blob_reserve_intptr(ctx
->blob
);
686 assert(blob_offset
+ sizeof(uintptr_t) == blob_offset2
);
687 write_phi_fixup fixup
= {
688 .blob_offset
= blob_offset
,
692 util_dynarray_append(&ctx
->phi_fixups
, write_phi_fixup
, fixup
);
697 write_fixup_phis(write_ctx
*ctx
)
699 util_dynarray_foreach(&ctx
->phi_fixups
, write_phi_fixup
, fixup
) {
700 uintptr_t *blob_ptr
= (uintptr_t *)(ctx
->blob
->data
+ fixup
->blob_offset
);
701 blob_ptr
[0] = write_lookup_object(ctx
, fixup
->src
);
702 blob_ptr
[1] = write_lookup_object(ctx
, fixup
->block
);
705 util_dynarray_clear(&ctx
->phi_fixups
);
708 static nir_phi_instr
*
709 read_phi(read_ctx
*ctx
, nir_block
*blk
)
711 nir_phi_instr
*phi
= nir_phi_instr_create(ctx
->nir
);
713 read_dest(ctx
, &phi
->dest
, &phi
->instr
);
715 unsigned num_srcs
= blob_read_uint32(ctx
->blob
);
717 /* For similar reasons as before, we just store the index directly into the
718 * pointer, and let a later pass resolve the phi sources.
720 * In order to ensure that the copied sources (which are just the indices
721 * from the blob for now) don't get inserted into the old shader's use-def
722 * lists, we have to add the phi instruction *before* we set up its
725 nir_instr_insert_after_block(blk
, &phi
->instr
);
727 for (unsigned i
= 0; i
< num_srcs
; i
++) {
728 nir_phi_src
*src
= ralloc(phi
, nir_phi_src
);
730 src
->src
.is_ssa
= true;
731 src
->src
.ssa
= (nir_ssa_def
*) blob_read_intptr(ctx
->blob
);
732 src
->pred
= (nir_block
*) blob_read_intptr(ctx
->blob
);
734 /* Since we're not letting nir_insert_instr handle use/def stuff for us,
735 * we have to set the parent_instr manually. It doesn't really matter
736 * when we do it, so we might as well do it here.
738 src
->src
.parent_instr
= &phi
->instr
;
740 /* Stash it in the list of phi sources. We'll walk this list and fix up
741 * sources at the very end of read_function_impl.
743 list_add(&src
->src
.use_link
, &ctx
->phi_srcs
);
745 exec_list_push_tail(&phi
->srcs
, &src
->node
);
752 read_fixup_phis(read_ctx
*ctx
)
754 list_for_each_entry_safe(nir_phi_src
, src
, &ctx
->phi_srcs
, src
.use_link
) {
755 src
->pred
= read_lookup_object(ctx
, (uintptr_t)src
->pred
);
756 src
->src
.ssa
= read_lookup_object(ctx
, (uintptr_t)src
->src
.ssa
);
758 /* Remove from this list */
759 list_del(&src
->src
.use_link
);
761 list_addtail(&src
->src
.use_link
, &src
->src
.ssa
->uses
);
763 assert(list_is_empty(&ctx
->phi_srcs
));
767 write_jump(write_ctx
*ctx
, const nir_jump_instr
*jmp
)
769 blob_write_uint32(ctx
->blob
, jmp
->type
);
772 static nir_jump_instr
*
773 read_jump(read_ctx
*ctx
)
775 nir_jump_type type
= blob_read_uint32(ctx
->blob
);
776 nir_jump_instr
*jmp
= nir_jump_instr_create(ctx
->nir
, type
);
781 write_call(write_ctx
*ctx
, const nir_call_instr
*call
)
783 blob_write_intptr(ctx
->blob
, write_lookup_object(ctx
, call
->callee
));
785 for (unsigned i
= 0; i
< call
->num_params
; i
++)
786 write_src(ctx
, &call
->params
[i
]);
789 static nir_call_instr
*
790 read_call(read_ctx
*ctx
)
792 nir_function
*callee
= read_object(ctx
);
793 nir_call_instr
*call
= nir_call_instr_create(ctx
->nir
, callee
);
795 for (unsigned i
= 0; i
< call
->num_params
; i
++)
796 read_src(ctx
, &call
->params
[i
], call
);
802 write_instr(write_ctx
*ctx
, const nir_instr
*instr
)
804 blob_write_uint32(ctx
->blob
, instr
->type
);
805 switch (instr
->type
) {
806 case nir_instr_type_alu
:
807 write_alu(ctx
, nir_instr_as_alu(instr
));
809 case nir_instr_type_deref
:
810 write_deref(ctx
, nir_instr_as_deref(instr
));
812 case nir_instr_type_intrinsic
:
813 write_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
815 case nir_instr_type_load_const
:
816 write_load_const(ctx
, nir_instr_as_load_const(instr
));
818 case nir_instr_type_ssa_undef
:
819 write_ssa_undef(ctx
, nir_instr_as_ssa_undef(instr
));
821 case nir_instr_type_tex
:
822 write_tex(ctx
, nir_instr_as_tex(instr
));
824 case nir_instr_type_phi
:
825 write_phi(ctx
, nir_instr_as_phi(instr
));
827 case nir_instr_type_jump
:
828 write_jump(ctx
, nir_instr_as_jump(instr
));
830 case nir_instr_type_call
:
831 write_call(ctx
, nir_instr_as_call(instr
));
833 case nir_instr_type_parallel_copy
:
834 unreachable("Cannot write parallel copies");
836 unreachable("bad instr type");
841 read_instr(read_ctx
*ctx
, nir_block
*block
)
843 nir_instr_type type
= blob_read_uint32(ctx
->blob
);
846 case nir_instr_type_alu
:
847 instr
= &read_alu(ctx
)->instr
;
849 case nir_instr_type_deref
:
850 instr
= &read_deref(ctx
)->instr
;
852 case nir_instr_type_intrinsic
:
853 instr
= &read_intrinsic(ctx
)->instr
;
855 case nir_instr_type_load_const
:
856 instr
= &read_load_const(ctx
)->instr
;
858 case nir_instr_type_ssa_undef
:
859 instr
= &read_ssa_undef(ctx
)->instr
;
861 case nir_instr_type_tex
:
862 instr
= &read_tex(ctx
)->instr
;
864 case nir_instr_type_phi
:
865 /* Phi instructions are a bit of a special case when reading because we
866 * don't want inserting the instruction to automatically handle use/defs
867 * for us. Instead, we need to wait until all the blocks/instructions
868 * are read so that we can set their sources up.
870 read_phi(ctx
, block
);
872 case nir_instr_type_jump
:
873 instr
= &read_jump(ctx
)->instr
;
875 case nir_instr_type_call
:
876 instr
= &read_call(ctx
)->instr
;
878 case nir_instr_type_parallel_copy
:
879 unreachable("Cannot read parallel copies");
881 unreachable("bad instr type");
884 nir_instr_insert_after_block(block
, instr
);
888 write_block(write_ctx
*ctx
, const nir_block
*block
)
890 write_add_object(ctx
, block
);
891 blob_write_uint32(ctx
->blob
, exec_list_length(&block
->instr_list
));
892 nir_foreach_instr(instr
, block
)
893 write_instr(ctx
, instr
);
897 read_block(read_ctx
*ctx
, struct exec_list
*cf_list
)
899 /* Don't actually create a new block. Just use the one from the tail of
900 * the list. NIR guarantees that the tail of the list is a block and that
901 * no two blocks are side-by-side in the IR; It should be empty.
904 exec_node_data(nir_block
, exec_list_get_tail(cf_list
), cf_node
.node
);
906 read_add_object(ctx
, block
);
907 unsigned num_instrs
= blob_read_uint32(ctx
->blob
);
908 for (unsigned i
= 0; i
< num_instrs
; i
++) {
909 read_instr(ctx
, block
);
914 write_cf_list(write_ctx
*ctx
, const struct exec_list
*cf_list
);
917 read_cf_list(read_ctx
*ctx
, struct exec_list
*cf_list
);
920 write_if(write_ctx
*ctx
, nir_if
*nif
)
922 write_src(ctx
, &nif
->condition
);
924 write_cf_list(ctx
, &nif
->then_list
);
925 write_cf_list(ctx
, &nif
->else_list
);
929 read_if(read_ctx
*ctx
, struct exec_list
*cf_list
)
931 nir_if
*nif
= nir_if_create(ctx
->nir
);
933 read_src(ctx
, &nif
->condition
, nif
);
935 nir_cf_node_insert_end(cf_list
, &nif
->cf_node
);
937 read_cf_list(ctx
, &nif
->then_list
);
938 read_cf_list(ctx
, &nif
->else_list
);
942 write_loop(write_ctx
*ctx
, nir_loop
*loop
)
944 write_cf_list(ctx
, &loop
->body
);
948 read_loop(read_ctx
*ctx
, struct exec_list
*cf_list
)
950 nir_loop
*loop
= nir_loop_create(ctx
->nir
);
952 nir_cf_node_insert_end(cf_list
, &loop
->cf_node
);
954 read_cf_list(ctx
, &loop
->body
);
958 write_cf_node(write_ctx
*ctx
, nir_cf_node
*cf
)
960 blob_write_uint32(ctx
->blob
, cf
->type
);
963 case nir_cf_node_block
:
964 write_block(ctx
, nir_cf_node_as_block(cf
));
967 write_if(ctx
, nir_cf_node_as_if(cf
));
969 case nir_cf_node_loop
:
970 write_loop(ctx
, nir_cf_node_as_loop(cf
));
973 unreachable("bad cf type");
978 read_cf_node(read_ctx
*ctx
, struct exec_list
*list
)
980 nir_cf_node_type type
= blob_read_uint32(ctx
->blob
);
983 case nir_cf_node_block
:
984 read_block(ctx
, list
);
989 case nir_cf_node_loop
:
990 read_loop(ctx
, list
);
993 unreachable("bad cf type");
998 write_cf_list(write_ctx
*ctx
, const struct exec_list
*cf_list
)
1000 blob_write_uint32(ctx
->blob
, exec_list_length(cf_list
));
1001 foreach_list_typed(nir_cf_node
, cf
, node
, cf_list
) {
1002 write_cf_node(ctx
, cf
);
1007 read_cf_list(read_ctx
*ctx
, struct exec_list
*cf_list
)
1009 uint32_t num_cf_nodes
= blob_read_uint32(ctx
->blob
);
1010 for (unsigned i
= 0; i
< num_cf_nodes
; i
++)
1011 read_cf_node(ctx
, cf_list
);
1015 write_function_impl(write_ctx
*ctx
, const nir_function_impl
*fi
)
1017 write_var_list(ctx
, &fi
->locals
);
1018 write_reg_list(ctx
, &fi
->registers
);
1019 blob_write_uint32(ctx
->blob
, fi
->reg_alloc
);
1021 write_cf_list(ctx
, &fi
->body
);
1022 write_fixup_phis(ctx
);
1025 static nir_function_impl
*
1026 read_function_impl(read_ctx
*ctx
, nir_function
*fxn
)
1028 nir_function_impl
*fi
= nir_function_impl_create_bare(ctx
->nir
);
1031 read_var_list(ctx
, &fi
->locals
);
1032 read_reg_list(ctx
, &fi
->registers
);
1033 fi
->reg_alloc
= blob_read_uint32(ctx
->blob
);
1035 read_cf_list(ctx
, &fi
->body
);
1036 read_fixup_phis(ctx
);
1038 fi
->valid_metadata
= 0;
1044 write_function(write_ctx
*ctx
, const nir_function
*fxn
)
1046 blob_write_uint32(ctx
->blob
, !!(fxn
->name
));
1048 blob_write_string(ctx
->blob
, fxn
->name
);
1050 write_add_object(ctx
, fxn
);
1052 blob_write_uint32(ctx
->blob
, fxn
->num_params
);
1053 for (unsigned i
= 0; i
< fxn
->num_params
; i
++) {
1055 ((uint32_t)fxn
->params
[i
].num_components
) |
1056 ((uint32_t)fxn
->params
[i
].bit_size
) << 8;
1057 blob_write_uint32(ctx
->blob
, val
);
1060 blob_write_uint32(ctx
->blob
, fxn
->is_entrypoint
);
1062 /* At first glance, it looks like we should write the function_impl here.
1063 * However, call instructions need to be able to reference at least the
1064 * function and those will get processed as we write the function_impls.
1065 * We stop here and write function_impls as a second pass.
1070 read_function(read_ctx
*ctx
)
1072 bool has_name
= blob_read_uint32(ctx
->blob
);
1073 char *name
= has_name
? blob_read_string(ctx
->blob
) : NULL
;
1075 nir_function
*fxn
= nir_function_create(ctx
->nir
, name
);
1077 read_add_object(ctx
, fxn
);
1079 fxn
->num_params
= blob_read_uint32(ctx
->blob
);
1080 fxn
->params
= ralloc_array(fxn
, nir_parameter
, fxn
->num_params
);
1081 for (unsigned i
= 0; i
< fxn
->num_params
; i
++) {
1082 uint32_t val
= blob_read_uint32(ctx
->blob
);
1083 fxn
->params
[i
].num_components
= val
& 0xff;
1084 fxn
->params
[i
].bit_size
= (val
>> 8) & 0xff;
1087 fxn
->is_entrypoint
= blob_read_uint32(ctx
->blob
);
1091 nir_serialize(struct blob
*blob
, const nir_shader
*nir
, bool strip
)
1093 nir_shader
*stripped
= NULL
;
1096 /* Drop unnecessary information (like variable names), so the serialized
1097 * NIR is smaller, and also to let us detect more isomorphic shaders
1098 * when hashing, increasing cache hits.
1100 stripped
= nir_shader_clone(NULL
, nir
);
1101 nir_strip(stripped
);
1106 ctx
.remap_table
= _mesa_pointer_hash_table_create(NULL
);
1110 util_dynarray_init(&ctx
.phi_fixups
, NULL
);
1112 size_t idx_size_offset
= blob_reserve_intptr(blob
);
1114 struct shader_info info
= nir
->info
;
1115 uint32_t strings
= 0;
1120 blob_write_uint32(blob
, strings
);
1122 blob_write_string(blob
, info
.name
);
1124 blob_write_string(blob
, info
.label
);
1125 info
.name
= info
.label
= NULL
;
1126 blob_write_bytes(blob
, (uint8_t *) &info
, sizeof(info
));
1128 write_var_list(&ctx
, &nir
->uniforms
);
1129 write_var_list(&ctx
, &nir
->inputs
);
1130 write_var_list(&ctx
, &nir
->outputs
);
1131 write_var_list(&ctx
, &nir
->shared
);
1132 write_var_list(&ctx
, &nir
->globals
);
1133 write_var_list(&ctx
, &nir
->system_values
);
1135 blob_write_uint32(blob
, nir
->num_inputs
);
1136 blob_write_uint32(blob
, nir
->num_uniforms
);
1137 blob_write_uint32(blob
, nir
->num_outputs
);
1138 blob_write_uint32(blob
, nir
->num_shared
);
1139 blob_write_uint32(blob
, nir
->scratch_size
);
1141 blob_write_uint32(blob
, exec_list_length(&nir
->functions
));
1142 nir_foreach_function(fxn
, nir
) {
1143 write_function(&ctx
, fxn
);
1146 nir_foreach_function(fxn
, nir
) {
1147 write_function_impl(&ctx
, fxn
->impl
);
1150 blob_write_uint32(blob
, nir
->constant_data_size
);
1151 if (nir
->constant_data_size
> 0)
1152 blob_write_bytes(blob
, nir
->constant_data
, nir
->constant_data_size
);
1154 *(uintptr_t *)(blob
->data
+ idx_size_offset
) = ctx
.next_idx
;
1156 _mesa_hash_table_destroy(ctx
.remap_table
, NULL
);
1157 util_dynarray_fini(&ctx
.phi_fixups
);
1160 ralloc_free(stripped
);
1164 nir_deserialize(void *mem_ctx
,
1165 const struct nir_shader_compiler_options
*options
,
1166 struct blob_reader
*blob
)
1170 list_inithead(&ctx
.phi_srcs
);
1171 ctx
.idx_table_len
= blob_read_intptr(blob
);
1172 ctx
.idx_table
= calloc(ctx
.idx_table_len
, sizeof(uintptr_t));
1175 uint32_t strings
= blob_read_uint32(blob
);
1176 char *name
= (strings
& 0x1) ? blob_read_string(blob
) : NULL
;
1177 char *label
= (strings
& 0x2) ? blob_read_string(blob
) : NULL
;
1179 struct shader_info info
;
1180 blob_copy_bytes(blob
, (uint8_t *) &info
, sizeof(info
));
1182 ctx
.nir
= nir_shader_create(mem_ctx
, info
.stage
, options
, NULL
);
1184 info
.name
= name
? ralloc_strdup(ctx
.nir
, name
) : NULL
;
1185 info
.label
= label
? ralloc_strdup(ctx
.nir
, label
) : NULL
;
1187 ctx
.nir
->info
= info
;
1189 read_var_list(&ctx
, &ctx
.nir
->uniforms
);
1190 read_var_list(&ctx
, &ctx
.nir
->inputs
);
1191 read_var_list(&ctx
, &ctx
.nir
->outputs
);
1192 read_var_list(&ctx
, &ctx
.nir
->shared
);
1193 read_var_list(&ctx
, &ctx
.nir
->globals
);
1194 read_var_list(&ctx
, &ctx
.nir
->system_values
);
1196 ctx
.nir
->num_inputs
= blob_read_uint32(blob
);
1197 ctx
.nir
->num_uniforms
= blob_read_uint32(blob
);
1198 ctx
.nir
->num_outputs
= blob_read_uint32(blob
);
1199 ctx
.nir
->num_shared
= blob_read_uint32(blob
);
1200 ctx
.nir
->scratch_size
= blob_read_uint32(blob
);
1202 unsigned num_functions
= blob_read_uint32(blob
);
1203 for (unsigned i
= 0; i
< num_functions
; i
++)
1204 read_function(&ctx
);
1206 nir_foreach_function(fxn
, ctx
.nir
)
1207 fxn
->impl
= read_function_impl(&ctx
, fxn
);
1209 ctx
.nir
->constant_data_size
= blob_read_uint32(blob
);
1210 if (ctx
.nir
->constant_data_size
> 0) {
1211 ctx
.nir
->constant_data
=
1212 ralloc_size(ctx
.nir
, ctx
.nir
->constant_data_size
);
1213 blob_copy_bytes(blob
, ctx
.nir
->constant_data
,
1214 ctx
.nir
->constant_data_size
);
1217 free(ctx
.idx_table
);
1223 nir_shader_serialize_deserialize(nir_shader
*shader
)
1225 const struct nir_shader_compiler_options
*options
= shader
->options
;
1229 nir_serialize(&writer
, shader
, false);
1231 /* Delete all of dest's ralloc children but leave dest alone */
1232 void *dead_ctx
= ralloc_context(NULL
);
1233 ralloc_adopt(dead_ctx
, shader
);
1234 ralloc_free(dead_ctx
);
1236 dead_ctx
= ralloc_context(NULL
);
1238 struct blob_reader reader
;
1239 blob_reader_init(&reader
, writer
.data
, writer
.size
);
1240 nir_shader
*copy
= nir_deserialize(dead_ctx
, options
, &reader
);
1242 blob_finish(&writer
);
1244 nir_shader_replace(shader
, copy
);
1245 ralloc_free(dead_ctx
);