2 * Copyright © 2012 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 * Eric Anholt <eric@anholt.net>
29 #include "brw_vec4_live_variables.h"
33 /** @file brw_vec4_live_variables.cpp
35 * Support for computing at the basic block level which variables
36 * (virtual GRFs in our case) are live at entry and exit.
38 * See Muchnick's Advanced Compiler Design and Implementation, section
43 * Sets up the use[] and def[] arrays.
45 * The basic-block-level live variable analysis needs to know which
46 * variables get used before they're completely defined, and which
47 * variables are completely defined before they're used.
49 * We independently track each channel of a vec4. This is because we need to
50 * be able to recognize a sequence like:
55 * MUL result.xy tmp.xy e.xy
58 * as having tmp live only across that sequence (assuming it's used nowhere
59 * else), because it's a common pattern. A more conservative approach that
60 * doesn't get tmp marked a deffed in this block will tend to result in
64 vec4_live_variables::setup_def_use()
68 foreach_block (block
, cfg
) {
69 assert(ip
== block
->start_ip
);
71 assert(cfg
->blocks
[block
->num
- 1]->end_ip
== ip
- 1);
73 foreach_inst_in_block(vec4_instruction
, inst
, block
) {
74 struct block_data
*bd
= &block_data
[block
->num
];
76 /* Set use[] for this instruction */
77 for (unsigned int i
= 0; i
< 3; i
++) {
78 if (inst
->src
[i
].file
== GRF
) {
79 int reg
= inst
->src
[i
].reg
;
81 for (int j
= 0; j
< 4; j
++) {
82 int c
= BRW_GET_SWZ(inst
->src
[i
].swizzle
, j
);
83 if (!BITSET_TEST(bd
->def
, reg
* 4 + c
))
84 BITSET_SET(bd
->use
, reg
* 4 + c
);
88 if (inst
->reads_flag()) {
89 if (!BITSET_TEST(bd
->flag_def
, 0)) {
90 BITSET_SET(bd
->flag_use
, 0);
94 /* Check for unconditional writes to whole registers. These
95 * are the things that screen off preceding definitions of a
96 * variable, and thus qualify for being in def[].
98 if (inst
->dst
.file
== GRF
&&
99 alloc
.sizes
[inst
->dst
.reg
] == 1 &&
101 for (int c
= 0; c
< 4; c
++) {
102 if (inst
->dst
.writemask
& (1 << c
)) {
103 int reg
= inst
->dst
.reg
;
104 if (!BITSET_TEST(bd
->use
, reg
* 4 + c
))
105 BITSET_SET(bd
->def
, reg
* 4 + c
);
109 if (inst
->writes_flag()) {
110 if (!BITSET_TEST(bd
->flag_use
, 0)) {
111 BITSET_SET(bd
->flag_def
, 0);
121 * The algorithm incrementally sets bits in liveout and livein,
122 * propagating it through control flow. It will eventually terminate
123 * because it only ever adds bits, and stops when no bits are added in
127 vec4_live_variables::compute_live_variables()
134 foreach_block (block
, cfg
) {
135 struct block_data
*bd
= &block_data
[block
->num
];
138 for (int i
= 0; i
< bitset_words
; i
++) {
139 BITSET_WORD new_livein
= (bd
->use
[i
] |
142 if (new_livein
& ~bd
->livein
[i
]) {
143 bd
->livein
[i
] |= new_livein
;
147 BITSET_WORD new_livein
= (bd
->flag_use
[0] |
148 (bd
->flag_liveout
[0] &
150 if (new_livein
& ~bd
->flag_livein
[0]) {
151 bd
->flag_livein
[0] |= new_livein
;
156 foreach_list_typed(bblock_link
, child_link
, link
, &block
->children
) {
157 struct block_data
*child_bd
= &block_data
[child_link
->block
->num
];
159 for (int i
= 0; i
< bitset_words
; i
++) {
160 BITSET_WORD new_liveout
= (child_bd
->livein
[i
] &
163 bd
->liveout
[i
] |= new_liveout
;
167 BITSET_WORD new_liveout
= (child_bd
->flag_livein
[0] &
168 ~bd
->flag_liveout
[0]);
170 bd
->flag_liveout
[0] |= new_liveout
;
178 vec4_live_variables::vec4_live_variables(const simple_allocator
&alloc
,
180 : alloc(alloc
), cfg(cfg
)
182 mem_ctx
= ralloc_context(NULL
);
184 num_vars
= alloc
.count
* 4;
185 block_data
= rzalloc_array(mem_ctx
, struct block_data
, cfg
->num_blocks
);
187 bitset_words
= BITSET_WORDS(num_vars
);
188 for (int i
= 0; i
< cfg
->num_blocks
; i
++) {
189 block_data
[i
].def
= rzalloc_array(mem_ctx
, BITSET_WORD
, bitset_words
);
190 block_data
[i
].use
= rzalloc_array(mem_ctx
, BITSET_WORD
, bitset_words
);
191 block_data
[i
].livein
= rzalloc_array(mem_ctx
, BITSET_WORD
, bitset_words
);
192 block_data
[i
].liveout
= rzalloc_array(mem_ctx
, BITSET_WORD
, bitset_words
);
194 block_data
[i
].flag_def
[0] = 0;
195 block_data
[i
].flag_use
[0] = 0;
196 block_data
[i
].flag_livein
[0] = 0;
197 block_data
[i
].flag_liveout
[0] = 0;
201 compute_live_variables();
204 vec4_live_variables::~vec4_live_variables()
206 ralloc_free(mem_ctx
);
209 #define MAX_INSTRUCTION (1 << 30)
212 * Computes a conservative start/end of the live intervals for each virtual GRF.
214 * We could expose per-channel live intervals to the consumer based on the
215 * information we computed in vec4_live_variables, except that our only
216 * current user is virtual_grf_interferes(). So we instead union the
217 * per-channel ranges into a per-vgrf range for virtual_grf_start[] and
220 * We could potentially have virtual_grf_interferes() do the test per-channel,
221 * which would let some interesting register allocation occur (particularly on
222 * code-generated GLSL sequences from the Cg compiler which does register
223 * allocation at the GLSL level and thus reuses components of the variable
224 * with distinct lifetimes). But right now the complexity of doing so doesn't
225 * seem worth it, since having virtual_grf_interferes() be cheap is important
226 * for register allocation performance.
229 vec4_visitor::calculate_live_intervals()
231 if (this->live_intervals
)
234 int *start
= ralloc_array(mem_ctx
, int, this->alloc
.count
* 4);
235 int *end
= ralloc_array(mem_ctx
, int, this->alloc
.count
* 4);
236 ralloc_free(this->virtual_grf_start
);
237 ralloc_free(this->virtual_grf_end
);
238 this->virtual_grf_start
= start
;
239 this->virtual_grf_end
= end
;
241 for (unsigned i
= 0; i
< this->alloc
.count
* 4; i
++) {
242 start
[i
] = MAX_INSTRUCTION
;
246 /* Start by setting up the intervals with no knowledge of control
250 foreach_block_and_inst(block
, vec4_instruction
, inst
, cfg
) {
251 for (unsigned int i
= 0; i
< 3; i
++) {
252 if (inst
->src
[i
].file
== GRF
) {
253 int reg
= inst
->src
[i
].reg
;
255 for (int j
= 0; j
< 4; j
++) {
256 int c
= BRW_GET_SWZ(inst
->src
[i
].swizzle
, j
);
258 start
[reg
* 4 + c
] = MIN2(start
[reg
* 4 + c
], ip
);
259 end
[reg
* 4 + c
] = ip
;
264 if (inst
->dst
.file
== GRF
) {
265 int reg
= inst
->dst
.reg
;
267 for (int c
= 0; c
< 4; c
++) {
268 if (inst
->dst
.writemask
& (1 << c
)) {
269 start
[reg
* 4 + c
] = MIN2(start
[reg
* 4 + c
], ip
);
270 end
[reg
* 4 + c
] = ip
;
278 /* Now, extend those intervals using our analysis of control flow.
280 * The control flow-aware analysis was done at a channel level, while at
281 * this point we're distilling it down to vgrfs.
283 this->live_intervals
= new(mem_ctx
) vec4_live_variables(alloc
, cfg
);
285 foreach_block (block
, cfg
) {
286 struct block_data
*bd
= &live_intervals
->block_data
[block
->num
];
288 for (int i
= 0; i
< live_intervals
->num_vars
; i
++) {
289 if (BITSET_TEST(bd
->livein
, i
)) {
290 start
[i
] = MIN2(start
[i
], block
->start_ip
);
291 end
[i
] = MAX2(end
[i
], block
->start_ip
);
294 if (BITSET_TEST(bd
->liveout
, i
)) {
295 start
[i
] = MIN2(start
[i
], block
->end_ip
);
296 end
[i
] = MAX2(end
[i
], block
->end_ip
);
303 vec4_visitor::invalidate_live_intervals()
305 ralloc_free(live_intervals
);
306 live_intervals
= NULL
;
310 vec4_visitor::virtual_grf_interferes(int a
, int b
)
312 int start_a
= MIN2(MIN2(virtual_grf_start
[a
* 4 + 0],
313 virtual_grf_start
[a
* 4 + 1]),
314 MIN2(virtual_grf_start
[a
* 4 + 2],
315 virtual_grf_start
[a
* 4 + 3]));
316 int start_b
= MIN2(MIN2(virtual_grf_start
[b
* 4 + 0],
317 virtual_grf_start
[b
* 4 + 1]),
318 MIN2(virtual_grf_start
[b
* 4 + 2],
319 virtual_grf_start
[b
* 4 + 3]));
320 int end_a
= MAX2(MAX2(virtual_grf_end
[a
* 4 + 0],
321 virtual_grf_end
[a
* 4 + 1]),
322 MAX2(virtual_grf_end
[a
* 4 + 2],
323 virtual_grf_end
[a
* 4 + 3]));
324 int end_b
= MAX2(MAX2(virtual_grf_end
[b
* 4 + 0],
325 virtual_grf_end
[b
* 4 + 1]),
326 MAX2(virtual_grf_end
[b
* 4 + 2],
327 virtual_grf_end
[b
* 4 + 3]));
328 return !(end_a
<= start_b
||