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 /* Recursively constructs deref chains to split a copy instruction into
71 * multiple (if needed) copy instructions with full-length deref chains.
72 * External callers of this function should pass the tail and head of the
73 * deref chains found as the source and destination of the copy instruction
76 * \param old_copy The copy instruction we are splitting
77 * \param dest_head The head of the destination deref chain we are building
78 * \param src_head The head of the source deref chain we are building
79 * \param dest_tail The tail of the destination deref chain we are building
80 * \param src_tail The tail of the source deref chain we are building
81 * \param state The current split_var_copies_state object
84 split_var_copy_instr(nir_intrinsic_instr
*old_copy
,
85 nir_deref
*dest_head
, nir_deref
*src_head
,
86 nir_deref
*dest_tail
, nir_deref
*src_tail
,
87 struct split_var_copies_state
*state
)
89 assert(src_tail
->type
== dest_tail
->type
);
91 /* Make sure these really are the tails of the deref chains */
92 assert(dest_tail
->child
== NULL
);
93 assert(src_tail
->child
== NULL
);
95 switch (glsl_get_base_type(src_tail
->type
)) {
96 case GLSL_TYPE_ARRAY
: {
97 /* Make a wildcard dereference */
98 nir_deref_array
*deref
= nir_deref_array_create(state
->dead_ctx
);
99 deref
->deref
.type
= glsl_get_array_element(src_tail
->type
);
100 deref
->deref_array_type
= nir_deref_array_type_wildcard
;
102 /* Set the tail of both as the newly created wildcard deref. It is
103 * safe to use the same wildcard in both places because a) we will be
104 * copying it before we put it in an actual instruction and b)
105 * everything that will potentially add another link in the deref
106 * chain will also add the same thing to both chains.
108 src_tail
->child
= &deref
->deref
;
109 dest_tail
->child
= &deref
->deref
;
111 split_var_copy_instr(old_copy
, dest_head
, src_head
,
112 dest_tail
->child
, src_tail
->child
, state
);
114 /* Set it back to the way we found it */
115 src_tail
->child
= NULL
;
116 dest_tail
->child
= NULL
;
120 case GLSL_TYPE_STRUCT
:
121 /* This is the only part that actually does any interesting
122 * splitting. For array types, we just use wildcards and resolve
123 * them later. For structure types, we need to emit one copy
124 * instruction for every structure element. Because we may have
125 * structs inside structs, we just recurse and let the next level
126 * take care of any additional structures.
128 for (unsigned i
= 0; i
< glsl_get_length(src_tail
->type
); i
++) {
129 nir_deref_struct
*deref
= nir_deref_struct_create(state
->dead_ctx
, i
);
130 deref
->deref
.type
= glsl_get_struct_field(src_tail
->type
, i
);
132 /* Set the tail of both as the newly created structure deref. It
133 * is safe to use the same wildcard in both places because a) we
134 * will be copying it before we put it in an actual instruction
135 * and b) everything that will potentially add another link in the
136 * deref chain will also add the same thing to both chains.
138 src_tail
->child
= &deref
->deref
;
139 dest_tail
->child
= &deref
->deref
;
141 split_var_copy_instr(old_copy
, dest_head
, src_head
,
142 dest_tail
->child
, src_tail
->child
, state
);
144 /* Set it back to the way we found it */
145 src_tail
->child
= NULL
;
146 dest_tail
->child
= NULL
;
151 case GLSL_TYPE_FLOAT
:
152 case GLSL_TYPE_DOUBLE
:
154 if (glsl_type_is_matrix(src_tail
->type
)) {
155 nir_deref_array
*deref
= nir_deref_array_create(state
->dead_ctx
);
156 deref
->deref
.type
= glsl_get_column_type(src_tail
->type
);
157 deref
->deref_array_type
= nir_deref_array_type_wildcard
;
159 /* Set the tail of both as the newly created wildcard deref. It
160 * is safe to use the same wildcard in both places because a) we
161 * will be copying it before we put it in an actual instruction
162 * and b) everything that will potentially add another link in the
163 * deref chain will also add the same thing to both chains.
165 src_tail
->child
= &deref
->deref
;
166 dest_tail
->child
= &deref
->deref
;
168 split_var_copy_instr(old_copy
, dest_head
, src_head
,
169 dest_tail
->child
, src_tail
->child
, state
);
171 /* Set it back to the way we found it */
172 src_tail
->child
= NULL
;
173 dest_tail
->child
= NULL
;
175 /* At this point, we have fully built our deref chains and can
176 * actually add the new copy instruction.
178 nir_intrinsic_instr
*new_copy
=
179 nir_intrinsic_instr_create(state
->mem_ctx
, nir_intrinsic_copy_var
);
181 /* We need to make copies because a) this deref chain actually
182 * belongs to the copy instruction and b) the deref chains may
183 * have some of the same links due to the way we constructed them
185 nir_deref
*src
= nir_copy_deref(new_copy
, src_head
);
186 nir_deref
*dest
= nir_copy_deref(new_copy
, dest_head
);
188 new_copy
->variables
[0] = nir_deref_as_var(dest
);
189 new_copy
->variables
[1] = nir_deref_as_var(src
);
191 /* Emit the copy instruction after the old instruction. We'll
192 * remove the old one later.
194 nir_instr_insert_after(&old_copy
->instr
, &new_copy
->instr
);
195 state
->progress
= true;
199 case GLSL_TYPE_SAMPLER
:
200 case GLSL_TYPE_IMAGE
:
201 case GLSL_TYPE_ATOMIC_UINT
:
202 case GLSL_TYPE_INTERFACE
:
204 unreachable("Cannot copy these types");
209 split_var_copies_block(nir_block
*block
, struct split_var_copies_state
*state
)
211 nir_foreach_instr_safe(instr
, block
) {
212 if (instr
->type
!= nir_instr_type_intrinsic
)
215 nir_intrinsic_instr
*intrinsic
= nir_instr_as_intrinsic(instr
);
216 if (intrinsic
->intrinsic
!= nir_intrinsic_copy_var
)
219 nir_deref
*dest_head
= &intrinsic
->variables
[0]->deref
;
220 nir_deref
*src_head
= &intrinsic
->variables
[1]->deref
;
221 nir_deref
*dest_tail
= nir_deref_tail(dest_head
);
222 nir_deref
*src_tail
= nir_deref_tail(src_head
);
224 switch (glsl_get_base_type(src_tail
->type
)) {
225 case GLSL_TYPE_ARRAY
:
226 case GLSL_TYPE_STRUCT
:
227 split_var_copy_instr(intrinsic
, dest_head
, src_head
,
228 dest_tail
, src_tail
, state
);
229 nir_instr_remove(&intrinsic
->instr
);
230 ralloc_steal(state
->dead_ctx
, instr
);
232 case GLSL_TYPE_FLOAT
:
233 case GLSL_TYPE_DOUBLE
:
237 if (glsl_type_is_matrix(src_tail
->type
)) {
238 split_var_copy_instr(intrinsic
, dest_head
, src_head
,
239 dest_tail
, src_tail
, state
);
240 nir_instr_remove(&intrinsic
->instr
);
241 ralloc_steal(state
->dead_ctx
, instr
);
245 unreachable("Invalid type");
254 split_var_copies_impl(nir_function_impl
*impl
)
256 struct split_var_copies_state state
;
258 state
.mem_ctx
= ralloc_parent(impl
);
259 state
.dead_ctx
= ralloc_context(NULL
);
260 state
.progress
= false;
262 nir_foreach_block(block
, impl
) {
263 split_var_copies_block(block
, &state
);
266 ralloc_free(state
.dead_ctx
);
268 if (state
.progress
) {
269 nir_metadata_preserve(impl
, nir_metadata_block_index
|
270 nir_metadata_dominance
);
273 return state
.progress
;
277 nir_split_var_copies(nir_shader
*shader
)
279 bool progress
= false;
281 nir_foreach_function(function
, shader
) {
283 progress
= split_var_copies_impl(function
->impl
) || progress
;