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)
29 #include "nir_control_flow.h"
32 * Implements a small peephole optimization that looks for
43 * and replaces it with a series of selects. It can also handle the case
44 * where, instead of being empty, the if may contain some move operations
45 * whose only use is one of the following phi nodes. This happens all the
46 * time when the SSA form comes from a conditional assignment with a
50 struct peephole_select_state
{
56 block_check_for_allowed_instrs(nir_block
*block
)
58 nir_foreach_instr(block
, instr
) {
59 switch (instr
->type
) {
60 case nir_instr_type_intrinsic
: {
61 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
63 switch (intrin
->intrinsic
) {
64 case nir_intrinsic_load_var
:
65 switch (intrin
->variables
[0]->var
->data
.mode
) {
66 case nir_var_shader_in
:
82 case nir_instr_type_load_const
:
85 case nir_instr_type_alu
: {
86 nir_alu_instr
*mov
= nir_instr_as_alu(instr
);
97 /* It must be a move-like operation. */
103 /* Can't handle saturate */
104 if (mov
->dest
.saturate
)
108 if (!mov
->dest
.dest
.is_ssa
)
111 /* It cannot have any if-uses */
112 if (!list_empty(&mov
->dest
.dest
.ssa
.if_uses
))
115 /* The only uses of this definition must be phi's in the successor */
116 nir_foreach_use(&mov
->dest
.dest
.ssa
, use
) {
117 if (use
->parent_instr
->type
!= nir_instr_type_phi
||
118 use
->parent_instr
->block
!= block
->successors
[0])
133 nir_opt_peephole_select_block(nir_block
*block
, void *void_state
)
135 struct peephole_select_state
*state
= void_state
;
137 /* If the block is empty, then it certainly doesn't have any phi nodes,
138 * so we can skip it. This also ensures that we do an early skip on the
139 * end block of the function which isn't actually attached to the CFG.
141 if (exec_list_is_empty(&block
->instr_list
))
144 if (nir_cf_node_is_first(&block
->cf_node
))
147 nir_cf_node
*prev_node
= nir_cf_node_prev(&block
->cf_node
);
148 if (prev_node
->type
!= nir_cf_node_if
)
151 nir_if
*if_stmt
= nir_cf_node_as_if(prev_node
);
152 nir_cf_node
*then_node
= nir_if_first_then_node(if_stmt
);
153 nir_cf_node
*else_node
= nir_if_first_else_node(if_stmt
);
155 /* We can only have one block in each side ... */
156 if (nir_if_last_then_node(if_stmt
) != then_node
||
157 nir_if_last_else_node(if_stmt
) != else_node
)
160 nir_block
*then_block
= nir_cf_node_as_block(then_node
);
161 nir_block
*else_block
= nir_cf_node_as_block(else_node
);
163 /* ... and those blocks must only contain "allowed" instructions. */
164 if (!block_check_for_allowed_instrs(then_block
) ||
165 !block_check_for_allowed_instrs(else_block
))
168 /* At this point, we know that the previous CFG node is an if-then
169 * statement containing only moves to phi nodes in this block. We can
170 * just remove that entire CF node and replace all of the phi nodes with
174 nir_block
*prev_block
= nir_cf_node_as_block(nir_cf_node_prev(prev_node
));
175 assert(prev_block
->cf_node
.type
== nir_cf_node_block
);
177 /* First, we move the remaining instructions from the blocks to the
178 * block before. We have already guaranteed that this is safe by
179 * calling block_check_for_allowed_instrs()
181 nir_foreach_instr_safe(then_block
, instr
) {
182 exec_node_remove(&instr
->node
);
183 instr
->block
= prev_block
;
184 exec_list_push_tail(&prev_block
->instr_list
, &instr
->node
);
187 nir_foreach_instr_safe(else_block
, instr
) {
188 exec_node_remove(&instr
->node
);
189 instr
->block
= prev_block
;
190 exec_list_push_tail(&prev_block
->instr_list
, &instr
->node
);
193 nir_foreach_instr_safe(block
, instr
) {
194 if (instr
->type
!= nir_instr_type_phi
)
197 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
198 nir_alu_instr
*sel
= nir_alu_instr_create(state
->mem_ctx
, nir_op_bcsel
);
199 nir_src_copy(&sel
->src
[0].src
, &if_stmt
->condition
, sel
);
200 /* Splat the condition to all channels */
201 memset(sel
->src
[0].swizzle
, 0, sizeof sel
->src
[0].swizzle
);
203 assert(exec_list_length(&phi
->srcs
) == 2);
204 nir_foreach_phi_src(phi
, src
) {
205 assert(src
->pred
== then_block
|| src
->pred
== else_block
);
206 assert(src
->src
.is_ssa
);
208 unsigned idx
= src
->pred
== then_block
? 1 : 2;
209 nir_src_copy(&sel
->src
[idx
].src
, &src
->src
, sel
);
212 nir_ssa_dest_init(&sel
->instr
, &sel
->dest
.dest
,
213 phi
->dest
.ssa
.num_components
, phi
->dest
.ssa
.name
);
214 sel
->dest
.write_mask
= (1 << phi
->dest
.ssa
.num_components
) - 1;
216 nir_ssa_def_rewrite_uses(&phi
->dest
.ssa
,
217 nir_src_for_ssa(&sel
->dest
.dest
.ssa
));
219 nir_instr_insert_before(&phi
->instr
, &sel
->instr
);
220 nir_instr_remove(&phi
->instr
);
223 nir_cf_node_remove(&if_stmt
->cf_node
);
224 state
->progress
= true;
230 nir_opt_peephole_select_impl(nir_function_impl
*impl
)
232 struct peephole_select_state state
;
234 state
.mem_ctx
= ralloc_parent(impl
);
235 state
.progress
= false;
237 nir_foreach_block(impl
, nir_opt_peephole_select_block
, &state
);
240 nir_metadata_preserve(impl
, nir_metadata_none
);
242 return state
.progress
;
246 nir_opt_peephole_select(nir_shader
*shader
)
248 bool progress
= false;
250 nir_foreach_function(shader
, function
) {
252 progress
|= nir_opt_peephole_select_impl(function
->impl
);