2 * Copyright © 2014 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
31 * Implements "copy splitting" which is similar to structure splitting only
32 * it works on copy operations rather than the datatypes themselves. The
33 * GLSL language allows you to copy one variable to another an entire
34 * structure (which may contain arrays or other structures) at a time.
35 * Normally, in a language such as C this would be handled by a "structure
36 * splitting" pass that breaks up the structures. Unfortunately for us,
37 * structures used in inputs or outputs can't be split. Therefore,
38 * regardlesss of what we do, we have to be able to copy to/from
41 * The primary purpose of structure splitting is to allow you to better
42 * optimize variable access and lower things to registers where you can.
43 * The primary issue here is that, if you lower the copy to a bunch of
44 * loads and stores, you loose a lot of information about the copy
45 * operation that you would like to keep around. To solve this problem, we
46 * have a "copy splitting" pass that, instead of splitting the structures
47 * or lowering the copy into loads and storres, splits the copy operation
48 * into a bunch of copy operations one for each leaf of the structure tree.
49 * If an intermediate array is encountered, it is referenced with a
50 * wildcard reference to indicate that the entire array is to be copied.
52 * As things become direct, array copies may be able to be losslessly
53 * lowered to having fewer and fewer wildcards. However, until that
54 * happens we want to keep the information about the arrays intact.
56 * Prior to the copy splitting pass, there are no wildcard references but
57 * there may be incomplete references where the tail of the deref chain is
58 * an array or a structure and not a specific element. After the copy
59 * splitting pass has completed, every variable deref will be a full-length
60 * dereference pointing to a single leaf in the structure type tree with
61 * possibly a few wildcard array dereferences.
64 struct split_var_copies_state
{
70 get_deref_tail(nir_deref
*deref
)
72 while (deref
->child
!= NULL
)
77 /* Recursively constructs deref chains to split a copy instruction into
78 * multiple (if needed) copy instructions with full-length deref chains.
79 * External callers of this function should pass the tail and head of the
80 * deref chains found as the source and destination of the copy instruction
83 * \param old_copy The copy instruction we are splitting
84 * \param dest_head The head of the destination deref chain we are building
85 * \param src_head The head of the source deref chain we are building
86 * \param dest_tail The tail of the destination deref chain we are building
87 * \param src_tail The tail of the source deref chain we are building
88 * \param state The current split_var_copies_state object
91 split_var_copy_instr(nir_intrinsic_instr
*old_copy
,
92 nir_deref
*dest_head
, nir_deref
*src_head
,
93 nir_deref
*dest_tail
, nir_deref
*src_tail
,
94 struct split_var_copies_state
*state
)
96 assert(src_tail
->type
== dest_tail
->type
);
98 /* Make sure these really are the tails of the deref chains */
99 assert(dest_tail
->child
== NULL
);
100 assert(src_tail
->child
== NULL
);
102 switch (glsl_get_base_type(src_tail
->type
)) {
103 case GLSL_TYPE_ARRAY
: {
104 /* Make a wildcard dereference */
105 nir_deref_array
*deref
= nir_deref_array_create(state
->dead_ctx
);
106 deref
->deref
.type
= glsl_get_array_element(src_tail
->type
);
107 deref
->deref_array_type
= nir_deref_array_type_wildcard
;
109 /* Set the tail of both as the newly created wildcard deref. It is
110 * safe to use the same wildcard in both places because a) we will be
111 * copying it before we put it in an actual instruction and b)
112 * everything that will potentially add another link in the deref
113 * chain will also add the same thing to both chains.
115 src_tail
->child
= &deref
->deref
;
116 dest_tail
->child
= &deref
->deref
;
118 split_var_copy_instr(old_copy
, dest_head
, src_head
,
119 dest_tail
->child
, src_tail
->child
, state
);
121 /* Set it back to the way we found it */
122 src_tail
->child
= NULL
;
123 dest_tail
->child
= NULL
;
127 case GLSL_TYPE_STRUCT
:
128 /* This is the only part that actually does any interesting
129 * splitting. For array types, we just use wildcards and resolve
130 * them later. For structure types, we need to emit one copy
131 * instruction for every structure element. Because we may have
132 * structs inside structs, we just recurse and let the next level
133 * take care of any additional structures.
135 for (unsigned i
= 0; i
< glsl_get_length(src_tail
->type
); i
++) {
136 nir_deref_struct
*deref
= nir_deref_struct_create(state
->dead_ctx
, i
);
137 deref
->deref
.type
= glsl_get_struct_field(src_tail
->type
, i
);
139 /* Set the tail of both as the newly created structure deref. It
140 * is safe to use the same wildcard in both places because a) we
141 * will be copying it before we put it in an actual instruction
142 * and b) everything that will potentially add another link in the
143 * deref chain will also add the same thing to both chains.
145 src_tail
->child
= &deref
->deref
;
146 dest_tail
->child
= &deref
->deref
;
148 split_var_copy_instr(old_copy
, dest_head
, src_head
,
149 dest_tail
->child
, src_tail
->child
, state
);
151 /* Set it back to the way we found it */
152 src_tail
->child
= NULL
;
153 dest_tail
->child
= NULL
;
158 case GLSL_TYPE_FLOAT
:
160 if (glsl_type_is_matrix(src_tail
->type
)) {
161 nir_deref_array
*deref
= nir_deref_array_create(state
->dead_ctx
);
162 deref
->deref
.type
= glsl_get_column_type(src_tail
->type
);
163 deref
->deref_array_type
= nir_deref_array_type_wildcard
;
165 /* Set the tail of both as the newly created wildcard deref. It
166 * is safe to use the same wildcard in both places because a) we
167 * will be copying it before we put it in an actual instruction
168 * and b) everything that will potentially add another link in the
169 * deref chain will also add the same thing to both chains.
171 src_tail
->child
= &deref
->deref
;
172 dest_tail
->child
= &deref
->deref
;
174 split_var_copy_instr(old_copy
, dest_head
, src_head
,
175 dest_tail
->child
, src_tail
->child
, state
);
177 /* Set it back to the way we found it */
178 src_tail
->child
= NULL
;
179 dest_tail
->child
= NULL
;
181 /* At this point, we have fully built our deref chains and can
182 * actually add the new copy instruction.
184 nir_intrinsic_instr
*new_copy
=
185 nir_intrinsic_instr_create(state
->mem_ctx
, nir_intrinsic_copy_var
);
187 /* We need to make copies because a) this deref chain actually
188 * belongs to the copy instruction and b) the deref chains may
189 * have some of the same links due to the way we constructed them
191 nir_deref
*src
= nir_copy_deref(new_copy
, src_head
);
192 nir_deref
*dest
= nir_copy_deref(new_copy
, dest_head
);
194 new_copy
->variables
[0] = nir_deref_as_var(dest
);
195 new_copy
->variables
[1] = nir_deref_as_var(src
);
197 /* Emit the copy instruction after the old instruction. We'll
198 * remove the old one later.
200 nir_instr_insert_after(&old_copy
->instr
, &new_copy
->instr
);
204 case GLSL_TYPE_SAMPLER
:
205 case GLSL_TYPE_IMAGE
:
206 case GLSL_TYPE_ATOMIC_UINT
:
207 case GLSL_TYPE_INTERFACE
:
209 unreachable("Cannot copy these types");
214 split_var_copies_block(nir_block
*block
, void *void_state
)
216 struct split_var_copies_state
*state
= void_state
;
218 nir_foreach_instr_safe(block
, instr
) {
219 if (instr
->type
!= nir_instr_type_intrinsic
)
222 nir_intrinsic_instr
*intrinsic
= nir_instr_as_intrinsic(instr
);
223 if (intrinsic
->intrinsic
!= nir_intrinsic_copy_var
)
226 nir_deref
*dest_head
= &intrinsic
->variables
[0]->deref
;
227 nir_deref
*src_head
= &intrinsic
->variables
[1]->deref
;
228 nir_deref
*dest_tail
= get_deref_tail(dest_head
);
229 nir_deref
*src_tail
= get_deref_tail(src_head
);
231 switch (glsl_get_base_type(src_tail
->type
)) {
232 case GLSL_TYPE_ARRAY
:
233 case GLSL_TYPE_STRUCT
:
234 split_var_copy_instr(intrinsic
, dest_head
, src_head
,
235 dest_tail
, src_tail
, state
);
236 nir_instr_remove(&intrinsic
->instr
);
237 ralloc_steal(state
->dead_ctx
, instr
);
239 case GLSL_TYPE_FLOAT
:
243 if (glsl_type_is_matrix(src_tail
->type
)) {
244 split_var_copy_instr(intrinsic
, dest_head
, src_head
,
245 dest_tail
, src_tail
, state
);
246 nir_instr_remove(&intrinsic
->instr
);
247 ralloc_steal(state
->dead_ctx
, instr
);
251 unreachable("Invalid type");
260 split_var_copies_impl(nir_function_impl
*impl
)
262 struct split_var_copies_state state
;
264 state
.mem_ctx
= ralloc_parent(impl
);
265 state
.dead_ctx
= ralloc_context(NULL
);
267 nir_foreach_block(impl
, split_var_copies_block
, &state
);
269 ralloc_free(state
.dead_ctx
);
273 nir_split_var_copies(nir_shader
*shader
)
275 nir_foreach_overload(shader
, overload
) {
277 split_var_copies_impl(overload
->impl
);