1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
|
/*
* Copyright © 2014 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#include "brw_fs.h"
#include "brw_cfg.h"
#include "brw_eu.h"
/** @file brw_fs_cmod_propagation.cpp
*
* Implements a pass that propagates the conditional modifier from a CMP x 0.0
* instruction into the instruction that generated x. For instance, in this
* sequence
*
* add(8) g70<1>F g69<8,8,1>F 4096F
* cmp.ge.f0(8) null g70<8,8,1>F 0F
*
* we can do the comparison as part of the ADD instruction directly:
*
* add.ge.f0(8) g70<1>F g69<8,8,1>F 4096F
*
* If there had been a use of the flag register and another CMP using g70
*
* add.ge.f0(8) g70<1>F g69<8,8,1>F 4096F
* (+f0) sel(8) g71<F> g72<8,8,1>F g73<8,8,1>F
* cmp.ge.f0(8) null g70<8,8,1>F 0F
*
* we can recognize that the CMP is generating the flag value that already
* exists and therefore remove the instruction.
*/
static bool
cmod_propagate_cmp_to_add(const gen_device_info *devinfo, bblock_t *block,
fs_inst *inst, unsigned dispatch_width)
{
bool read_flag = false;
foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
if (scan_inst->opcode == BRW_OPCODE_ADD &&
!scan_inst->is_partial_var_write(dispatch_width) &&
scan_inst->exec_size == inst->exec_size) {
bool negate;
/* A CMP is basically a subtraction. The result of the
* subtraction must be the same as the result of the addition.
* This means that one of the operands must be negated. So (a +
* b) vs (a == -b) or (a + -b) vs (a == b).
*/
if ((inst->src[0].equals(scan_inst->src[0]) &&
inst->src[1].negative_equals(scan_inst->src[1])) ||
(inst->src[0].equals(scan_inst->src[1]) &&
inst->src[1].negative_equals(scan_inst->src[0]))) {
negate = false;
} else if ((inst->src[0].negative_equals(scan_inst->src[0]) &&
inst->src[1].equals(scan_inst->src[1])) ||
(inst->src[0].negative_equals(scan_inst->src[1]) &&
inst->src[1].equals(scan_inst->src[0]))) {
negate = true;
} else {
goto not_match;
}
/* From the Sky Lake PRM Vol. 7 "Assigning Conditional Mods":
*
* * Note that the [post condition signal] bits generated at
* the output of a compute are before the .sat.
*
* So we don't have to bail if scan_inst has saturate.
*/
/* Otherwise, try propagating the conditional. */
const enum brw_conditional_mod cond =
negate ? brw_swap_cmod(inst->conditional_mod)
: inst->conditional_mod;
if (scan_inst->can_do_cmod() &&
((!read_flag && scan_inst->conditional_mod == BRW_CONDITIONAL_NONE) ||
scan_inst->conditional_mod == cond)) {
scan_inst->conditional_mod = cond;
inst->remove(block);
return true;
}
break;
}
not_match:
if (scan_inst->flags_written())
break;
read_flag = read_flag || scan_inst->flags_read(devinfo);
}
return false;
}
/**
* Propagate conditional modifiers from NOT instructions
*
* Attempt to convert sequences like
*
* or(8) g78<8,8,1> g76<8,8,1>UD g77<8,8,1>UD
* ...
* not.nz.f0(8) null g78<8,8,1>UD
*
* into
*
* or.z.f0(8) g78<8,8,1> g76<8,8,1>UD g77<8,8,1>UD
*/
static bool
cmod_propagate_not(const gen_device_info *devinfo, bblock_t *block,
fs_inst *inst, unsigned dispatch_width)
{
const enum brw_conditional_mod cond = brw_negate_cmod(inst->conditional_mod);
bool read_flag = false;
if (cond != BRW_CONDITIONAL_Z && cond != BRW_CONDITIONAL_NZ)
return false;
foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
if (regions_overlap(scan_inst->dst, scan_inst->size_written,
inst->src[0], inst->size_read(0))) {
if (scan_inst->opcode != BRW_OPCODE_OR &&
scan_inst->opcode != BRW_OPCODE_AND)
break;
if (scan_inst->is_partial_var_write(dispatch_width) ||
scan_inst->dst.offset != inst->src[0].offset ||
scan_inst->exec_size != inst->exec_size)
break;
if (scan_inst->can_do_cmod() &&
((!read_flag && scan_inst->conditional_mod == BRW_CONDITIONAL_NONE) ||
scan_inst->conditional_mod == cond)) {
scan_inst->conditional_mod = cond;
inst->remove(block);
return true;
}
break;
}
if (scan_inst->flags_written())
break;
read_flag = read_flag || scan_inst->flags_read(devinfo);
}
return false;
}
static bool
opt_cmod_propagation_local(const gen_device_info *devinfo,
bblock_t *block,
unsigned dispatch_width)
{
bool progress = false;
int ip = block->end_ip + 1;
foreach_inst_in_block_reverse_safe(fs_inst, inst, block) {
ip--;
if ((inst->opcode != BRW_OPCODE_AND &&
inst->opcode != BRW_OPCODE_CMP &&
inst->opcode != BRW_OPCODE_MOV &&
inst->opcode != BRW_OPCODE_NOT) ||
inst->predicate != BRW_PREDICATE_NONE ||
!inst->dst.is_null() ||
(inst->src[0].file != VGRF && inst->src[0].file != ATTR &&
inst->src[0].file != UNIFORM))
continue;
/* An ABS source modifier can only be handled when processing a compare
* with a value other than zero.
*/
if (inst->src[0].abs &&
(inst->opcode != BRW_OPCODE_CMP || inst->src[1].is_zero()))
continue;
/* Only an AND.NZ can be propagated. Many AND.Z instructions are
* generated (for ir_unop_not in fs_visitor::emit_bool_to_cond_code).
* Propagating those would require inverting the condition on the CMP.
* This changes both the flag value and the register destination of the
* CMP. That result may be used elsewhere, so we can't change its value
* on a whim.
*/
if (inst->opcode == BRW_OPCODE_AND &&
!(inst->src[1].is_one() &&
inst->conditional_mod == BRW_CONDITIONAL_NZ &&
!inst->src[0].negate))
continue;
if (inst->opcode == BRW_OPCODE_MOV &&
inst->conditional_mod != BRW_CONDITIONAL_NZ)
continue;
/* A CMP with a second source of zero can match with anything. A CMP
* with a second source that is not zero can only match with an ADD
* instruction.
*
* Only apply this optimization to float-point sources. It can fail for
* integers. For inputs a = 0x80000000, b = 4, int(0x80000000) < 4, but
* int(0x80000000) - 4 overflows and results in 0x7ffffffc. that's not
* less than zero, so the flags get set differently than for (a < b).
*/
if (inst->opcode == BRW_OPCODE_CMP && !inst->src[1].is_zero()) {
if (brw_reg_type_is_floating_point(inst->src[0].type) &&
cmod_propagate_cmp_to_add(devinfo, block, inst, dispatch_width))
progress = true;
continue;
}
if (inst->opcode == BRW_OPCODE_NOT) {
progress = cmod_propagate_not(devinfo, block, inst, dispatch_width) || progress;
continue;
}
bool read_flag = false;
foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
if (regions_overlap(scan_inst->dst, scan_inst->size_written,
inst->src[0], inst->size_read(0))) {
if (scan_inst->is_partial_var_write(dispatch_width) ||
scan_inst->dst.offset != inst->src[0].offset ||
scan_inst->exec_size != inst->exec_size)
break;
/* CMP's result is the same regardless of dest type. */
if (inst->conditional_mod == BRW_CONDITIONAL_NZ &&
scan_inst->opcode == BRW_OPCODE_CMP &&
brw_reg_type_is_integer(inst->dst.type)) {
inst->remove(block);
progress = true;
break;
}
/* If the AND wasn't handled by the previous case, it isn't safe
* to remove it.
*/
if (inst->opcode == BRW_OPCODE_AND)
break;
/* Not safe to use inequality operators if the types are different
*/
if (scan_inst->dst.type != inst->src[0].type &&
inst->conditional_mod != BRW_CONDITIONAL_Z &&
inst->conditional_mod != BRW_CONDITIONAL_NZ)
break;
/* Comparisons operate differently for ints and floats */
if (scan_inst->dst.type != inst->dst.type) {
/* Comparison result may be altered if the bit-size changes
* since that affects range, denorms, etc
*/
if (type_sz(scan_inst->dst.type) != type_sz(inst->dst.type))
break;
/* We should propagate from a MOV to another instruction in a
* sequence like:
*
* and(16) g31<1>UD g20<8,8,1>UD g22<8,8,1>UD
* mov.nz.f0(16) null<1>F g31<8,8,1>D
*/
if (inst->opcode == BRW_OPCODE_MOV) {
if ((inst->src[0].type != BRW_REGISTER_TYPE_D &&
inst->src[0].type != BRW_REGISTER_TYPE_UD) ||
(scan_inst->dst.type != BRW_REGISTER_TYPE_D &&
scan_inst->dst.type != BRW_REGISTER_TYPE_UD)) {
break;
}
} else if (brw_reg_type_is_floating_point(scan_inst->dst.type) !=
brw_reg_type_is_floating_point(inst->dst.type)) {
break;
}
}
/* If the instruction generating inst's source also wrote the
* flag, and inst is doing a simple .nz comparison, then inst
* is redundant - the appropriate value is already in the flag
* register. Delete inst.
*/
if (inst->conditional_mod == BRW_CONDITIONAL_NZ &&
!inst->src[0].negate &&
scan_inst->flags_written()) {
inst->remove(block);
progress = true;
break;
}
/* The conditional mod of the CMP/CMPN instructions behaves
* specially because the flag output is not calculated from the
* result of the instruction, but the other way around, which
* means that even if the condmod to propagate and the condmod
* from the CMP instruction are the same they will in general give
* different results because they are evaluated based on different
* inputs.
*/
if (scan_inst->opcode == BRW_OPCODE_CMP ||
scan_inst->opcode == BRW_OPCODE_CMPN)
break;
/* From the Sky Lake PRM Vol. 7 "Assigning Conditional Mods":
*
* * Note that the [post condition signal] bits generated at
* the output of a compute are before the .sat.
*/
if (scan_inst->saturate)
break;
/* From the Sky Lake PRM, Vol 2a, "Multiply":
*
* "When multiplying integer data types, if one of the sources
* is a DW, the resulting full precision data is stored in
* the accumulator. However, if the destination data type is
* either W or DW, the low bits of the result are written to
* the destination register and the remaining high bits are
* discarded. This results in undefined Overflow and Sign
* flags. Therefore, conditional modifiers and saturation
* (.sat) cannot be used in this case."
*
* We just disallow cmod propagation on all integer multiplies.
*/
if (!brw_reg_type_is_floating_point(scan_inst->dst.type) &&
scan_inst->opcode == BRW_OPCODE_MUL)
break;
/* Otherwise, try propagating the conditional. */
enum brw_conditional_mod cond =
inst->src[0].negate ? brw_swap_cmod(inst->conditional_mod)
: inst->conditional_mod;
if (scan_inst->can_do_cmod() &&
((!read_flag && scan_inst->conditional_mod == BRW_CONDITIONAL_NONE) ||
scan_inst->conditional_mod == cond)) {
scan_inst->conditional_mod = cond;
inst->remove(block);
progress = true;
}
break;
}
if (scan_inst->flags_written())
break;
read_flag = read_flag || scan_inst->flags_read(devinfo);
}
}
return progress;
}
bool
fs_visitor::opt_cmod_propagation()
{
bool progress = false;
foreach_block_reverse(block, cfg) {
progress = opt_cmod_propagation_local(devinfo, block, dispatch_width) || progress;
}
if (progress)
invalidate_live_intervals();
return progress;
}
|