2 * Copyright © 2015 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>
30 * This file implements an analysis pass that determines when we have to do
31 * a boolean resolve on Gen <= 5. Instructions that need a boolean resolve
32 * will have the booleans portion of the instr->pass_flags field set to
33 * BRW_NIR_BOOLEAN_NEEDS_RESOLVE.
37 /** Returns the resolve status for the given source
39 * If the source has a parent instruction then the resolve status is the
40 * status of the parent instruction. If the source does not have a parent
41 * instruction then we don't know so we return NON_BOOLEAN.
44 get_resolve_status_for_src(nir_src
*src
)
47 nir_instr
*src_instr
= src
->ssa
->parent_instr
;
48 uint8_t resolve_status
= src_instr
->pass_flags
& BRW_NIR_BOOLEAN_MASK
;
50 /* If the source instruction needs resolve, then from the perspective
51 * of the user, it's a true boolean.
53 if (resolve_status
== BRW_NIR_BOOLEAN_NEEDS_RESOLVE
)
54 resolve_status
= BRW_NIR_BOOLEAN_NO_RESOLVE
;
55 return resolve_status
;
57 return BRW_NIR_NON_BOOLEAN
;
61 /** Marks the given source as needing a resolve
63 * If the given source corresponds to an unresolved boolean it marks it as
64 * needing a resolve. Otherwise, we leave it alone.
67 src_mark_needs_resolve(nir_src
*src
, void *void_state
)
70 nir_instr
*src_instr
= src
->ssa
->parent_instr
;
71 uint8_t resolve_status
= src_instr
->pass_flags
& BRW_NIR_BOOLEAN_MASK
;
73 /* If the source instruction is unresolved, then mark it as needing
76 if (resolve_status
== BRW_NIR_BOOLEAN_UNRESOLVED
) {
77 src_instr
->pass_flags
&= ~BRW_NIR_BOOLEAN_MASK
;
78 src_instr
->pass_flags
|= BRW_NIR_BOOLEAN_NEEDS_RESOLVE
;
87 analyze_boolean_resolves_block(nir_block
*block
)
89 nir_foreach_instr(instr
, block
) {
90 switch (instr
->type
) {
91 case nir_instr_type_alu
: {
92 /* For ALU instructions, the resolve status is handled in a
95 * 1) Look at the instruction type and sources and determine if it
96 * can be left unresolved.
98 * 2) Look at the destination and see if we have to resolve
99 * anyway. (This is the case if this instruction is not the
100 * only instruction writing to a given register.)
102 * 3) If the instruction has a resolve status other than
103 * BOOL_UNRESOLVED or BOOL_NEEDS_RESOLVE then we walk through
104 * the sources and ensure that they are also resolved. This
105 * ensures that we don't end up with any stray unresolved
106 * booleans going into ADDs or something like that.
109 uint8_t resolve_status
;
110 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
112 case nir_op_b32all_fequal2
:
113 case nir_op_b32all_iequal2
:
114 case nir_op_b32all_fequal3
:
115 case nir_op_b32all_iequal3
:
116 case nir_op_b32all_fequal4
:
117 case nir_op_b32all_iequal4
:
118 case nir_op_b32any_fnequal2
:
119 case nir_op_b32any_inequal2
:
120 case nir_op_b32any_fnequal3
:
121 case nir_op_b32any_inequal3
:
122 case nir_op_b32any_fnequal4
:
123 case nir_op_b32any_inequal4
:
124 /* These are only implemented by the vec4 backend and its
125 * implementation emits resolved booleans. At some point in the
126 * future, this may change and we'll have to remove some of the
129 resolve_status
= BRW_NIR_BOOLEAN_NO_RESOLVE
;
134 /* This is a single-source instruction. Just copy the resolve
135 * status from the source.
137 resolve_status
= get_resolve_status_for_src(&alu
->src
[0].src
);
144 const unsigned first
= alu
->op
== nir_op_b32csel
? 1 : 0;
145 uint8_t src0_status
= get_resolve_status_for_src(&alu
->src
[first
+ 0].src
);
146 uint8_t src1_status
= get_resolve_status_for_src(&alu
->src
[first
+ 1].src
);
148 /* src0 of a bcsel is evaluated as a Boolean with the expectation
149 * that it has already been resolved. Mark it as such.
151 if (alu
->op
== nir_op_b32csel
)
152 src_mark_needs_resolve(&alu
->src
[0].src
, NULL
);
154 if (src0_status
== src1_status
) {
155 resolve_status
= src0_status
;
156 } else if (src0_status
== BRW_NIR_NON_BOOLEAN
||
157 src1_status
== BRW_NIR_NON_BOOLEAN
) {
158 /* If one of the sources is a non-boolean then the whole
159 * thing is a non-boolean.
161 resolve_status
= BRW_NIR_NON_BOOLEAN
;
163 /* At this point one of them is a true boolean and one is a
164 * boolean that needs a resolve. We could either resolve the
165 * unresolved source or we could resolve here. If we resolve
166 * the unresolved source then we get two resolves for the price
167 * of one. Just set this one to BOOLEAN_NO_RESOLVE and we'll
168 * let the code below force a resolve on the unresolved source.
170 resolve_status
= BRW_NIR_BOOLEAN_NO_RESOLVE
;
176 if (nir_alu_type_get_base_type(nir_op_infos
[alu
->op
].output_type
) == nir_type_bool
) {
177 /* This instructions will turn into a CMP when we actually emit
178 * them so the result will have to be resolved before it can be
181 resolve_status
= BRW_NIR_BOOLEAN_UNRESOLVED
;
183 /* Even though the destination is allowed to be left
184 * unresolved, the sources are treated as regular integers or
185 * floats so they need to be resolved.
187 nir_foreach_src(instr
, src_mark_needs_resolve
, NULL
);
189 resolve_status
= BRW_NIR_NON_BOOLEAN
;
193 /* If the destination is SSA, go ahead allow unresolved booleans.
194 * If the destination register doesn't have a well-defined parent_instr
195 * we need to resolve immediately.
197 if (!alu
->dest
.dest
.is_ssa
&&
198 resolve_status
== BRW_NIR_BOOLEAN_UNRESOLVED
) {
199 resolve_status
= BRW_NIR_BOOLEAN_NEEDS_RESOLVE
;
202 instr
->pass_flags
= (instr
->pass_flags
& ~BRW_NIR_BOOLEAN_MASK
) |
205 /* Finally, resolve sources if it's needed */
206 switch (resolve_status
) {
207 case BRW_NIR_BOOLEAN_NEEDS_RESOLVE
:
208 case BRW_NIR_BOOLEAN_UNRESOLVED
:
209 /* This instruction is either unresolved or we're doing the
210 * resolve here; leave the sources alone.
214 case BRW_NIR_BOOLEAN_NO_RESOLVE
:
215 case BRW_NIR_NON_BOOLEAN
:
216 nir_foreach_src(instr
, src_mark_needs_resolve
, NULL
);
220 unreachable("Invalid boolean flag");
226 case nir_instr_type_load_const
: {
227 nir_load_const_instr
*load
= nir_instr_as_load_const(instr
);
229 /* For load_const instructions, it's a boolean exactly when it holds
230 * one of the values NIR_TRUE or NIR_FALSE.
232 * Since load_const instructions don't have any sources, we don't
233 * have to worry about resolving them.
235 instr
->pass_flags
&= ~BRW_NIR_BOOLEAN_MASK
;
236 if (load
->value
[0].u32
== NIR_TRUE
|| load
->value
[0].u32
== NIR_FALSE
) {
237 instr
->pass_flags
|= BRW_NIR_BOOLEAN_NO_RESOLVE
;
239 instr
->pass_flags
|= BRW_NIR_NON_BOOLEAN
;
245 /* Everything else is an unknown non-boolean value and needs to
246 * have all sources resolved.
248 instr
->pass_flags
= (instr
->pass_flags
& ~BRW_NIR_BOOLEAN_MASK
) |
250 nir_foreach_src(instr
, src_mark_needs_resolve
, NULL
);
255 nir_if
*following_if
= nir_block_get_following_if(block
);
257 src_mark_needs_resolve(&following_if
->condition
, NULL
);
263 analyze_boolean_resolves_impl(nir_function_impl
*impl
)
265 nir_foreach_block(block
, impl
) {
266 analyze_boolean_resolves_block(block
);
271 brw_nir_analyze_boolean_resolves(nir_shader
*shader
)
273 nir_foreach_function(function
, shader
) {
275 analyze_boolean_resolves_impl(function
->impl
);