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
51 block_check_for_allowed_instrs(nir_block
*block
)
53 nir_foreach_instr(block
, instr
) {
54 switch (instr
->type
) {
55 case nir_instr_type_intrinsic
: {
56 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
58 switch (intrin
->intrinsic
) {
59 case nir_intrinsic_load_var
:
60 switch (intrin
->variables
[0]->var
->data
.mode
) {
61 case nir_var_shader_in
:
77 case nir_instr_type_load_const
:
80 case nir_instr_type_alu
: {
81 nir_alu_instr
*mov
= nir_instr_as_alu(instr
);
92 /* It must be a move-like operation. */
98 /* Can't handle saturate */
99 if (mov
->dest
.saturate
)
103 if (!mov
->dest
.dest
.is_ssa
)
106 /* It cannot have any if-uses */
107 if (!list_empty(&mov
->dest
.dest
.ssa
.if_uses
))
110 /* The only uses of this definition must be phi's in the successor */
111 nir_foreach_use(&mov
->dest
.dest
.ssa
, use
) {
112 if (use
->parent_instr
->type
!= nir_instr_type_phi
||
113 use
->parent_instr
->block
!= block
->successors
[0])
128 nir_opt_peephole_select_block(nir_block
*block
, void *mem_ctx
)
130 /* If the block is empty, then it certainly doesn't have any phi nodes,
131 * so we can skip it. This also ensures that we do an early skip on the
132 * end block of the function which isn't actually attached to the CFG.
134 if (exec_list_is_empty(&block
->instr_list
))
137 if (nir_cf_node_is_first(&block
->cf_node
))
140 nir_cf_node
*prev_node
= nir_cf_node_prev(&block
->cf_node
);
141 if (prev_node
->type
!= nir_cf_node_if
)
144 nir_if
*if_stmt
= nir_cf_node_as_if(prev_node
);
145 nir_cf_node
*then_node
= nir_if_first_then_node(if_stmt
);
146 nir_cf_node
*else_node
= nir_if_first_else_node(if_stmt
);
148 /* We can only have one block in each side ... */
149 if (nir_if_last_then_node(if_stmt
) != then_node
||
150 nir_if_last_else_node(if_stmt
) != else_node
)
153 nir_block
*then_block
= nir_cf_node_as_block(then_node
);
154 nir_block
*else_block
= nir_cf_node_as_block(else_node
);
156 /* ... and those blocks must only contain "allowed" instructions. */
157 if (!block_check_for_allowed_instrs(then_block
) ||
158 !block_check_for_allowed_instrs(else_block
))
161 /* At this point, we know that the previous CFG node is an if-then
162 * statement containing only moves to phi nodes in this block. We can
163 * just remove that entire CF node and replace all of the phi nodes with
167 nir_block
*prev_block
= nir_cf_node_as_block(nir_cf_node_prev(prev_node
));
168 assert(prev_block
->cf_node
.type
== nir_cf_node_block
);
170 /* First, we move the remaining instructions from the blocks to the
171 * block before. We have already guaranteed that this is safe by
172 * calling block_check_for_allowed_instrs()
174 nir_foreach_instr_safe(then_block
, instr
) {
175 exec_node_remove(&instr
->node
);
176 instr
->block
= prev_block
;
177 exec_list_push_tail(&prev_block
->instr_list
, &instr
->node
);
180 nir_foreach_instr_safe(else_block
, instr
) {
181 exec_node_remove(&instr
->node
);
182 instr
->block
= prev_block
;
183 exec_list_push_tail(&prev_block
->instr_list
, &instr
->node
);
186 nir_foreach_instr_safe(block
, instr
) {
187 if (instr
->type
!= nir_instr_type_phi
)
190 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
191 nir_alu_instr
*sel
= nir_alu_instr_create(mem_ctx
, nir_op_bcsel
);
192 nir_src_copy(&sel
->src
[0].src
, &if_stmt
->condition
, sel
);
193 /* Splat the condition to all channels */
194 memset(sel
->src
[0].swizzle
, 0, sizeof sel
->src
[0].swizzle
);
196 assert(exec_list_length(&phi
->srcs
) == 2);
197 nir_foreach_phi_src(phi
, src
) {
198 assert(src
->pred
== then_block
|| src
->pred
== else_block
);
199 assert(src
->src
.is_ssa
);
201 unsigned idx
= src
->pred
== then_block
? 1 : 2;
202 nir_src_copy(&sel
->src
[idx
].src
, &src
->src
, sel
);
205 nir_ssa_dest_init(&sel
->instr
, &sel
->dest
.dest
,
206 phi
->dest
.ssa
.num_components
,
207 phi
->dest
.ssa
.bit_size
, phi
->dest
.ssa
.name
);
208 sel
->dest
.write_mask
= (1 << phi
->dest
.ssa
.num_components
) - 1;
210 nir_ssa_def_rewrite_uses(&phi
->dest
.ssa
,
211 nir_src_for_ssa(&sel
->dest
.dest
.ssa
));
213 nir_instr_insert_before(&phi
->instr
, &sel
->instr
);
214 nir_instr_remove(&phi
->instr
);
217 nir_cf_node_remove(&if_stmt
->cf_node
);
222 nir_opt_peephole_select_impl(nir_function_impl
*impl
)
224 void *mem_ctx
= ralloc_parent(impl
);
225 bool progress
= false;
227 nir_foreach_block_safe(block
, impl
) {
228 progress
|= nir_opt_peephole_select_block(block
, mem_ctx
);
232 nir_metadata_preserve(impl
, nir_metadata_none
);
238 nir_opt_peephole_select(nir_shader
*shader
)
240 bool progress
= false;
242 nir_foreach_function(shader
, function
) {
244 progress
|= nir_opt_peephole_select_impl(function
->impl
);