aboutsummaryrefslogtreecommitdiffstats
path: root/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp
blob: 2c41d0208f25acc040ee94c7331ea79c47112a93 (plain)
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
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
/*
 * Copyright © 2011 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.
 */

/**
 * @file brw_vec4_copy_propagation.cpp
 *
 * Implements tracking of values copied between registers, and
 * optimizations based on that: copy propagation and constant
 * propagation.
 */

#include "brw_vec4.h"
extern "C" {
#include "main/macros.h"
}

namespace brw {

static bool
is_direct_copy(vec4_instruction *inst)
{
   return (inst->opcode == BRW_OPCODE_MOV &&
	   !inst->predicate &&
	   inst->dst.file == GRF &&
	   !inst->saturate &&
	   !inst->dst.reladdr &&
	   !inst->src[0].reladdr &&
	   inst->dst.type == inst->src[0].type);
}

static bool
is_dominated_by_previous_instruction(vec4_instruction *inst)
{
   return (inst->opcode != BRW_OPCODE_DO &&
	   inst->opcode != BRW_OPCODE_WHILE &&
	   inst->opcode != BRW_OPCODE_ELSE &&
	   inst->opcode != BRW_OPCODE_ENDIF);
}

static bool
is_channel_updated(vec4_instruction *inst, src_reg *values[4], int ch)
{
   const src_reg *src = values[ch];

   /* consider GRF only */
   assert(inst->dst.file == GRF);
   if (!src || src->file != GRF)
      return false;

   return (src->reg == inst->dst.reg &&
	   src->reg_offset == inst->dst.reg_offset &&
	   inst->dst.writemask & (1 << BRW_GET_SWZ(src->swizzle, ch)));
}

static bool
try_constant_propagate(struct brw_context *brw, vec4_instruction *inst,
                       int arg, src_reg *values[4])
{
   /* For constant propagation, we only handle the same constant
    * across all 4 channels.  Some day, we should handle the 8-bit
    * float vector format, which would let us constant propagate
    * vectors better.
    */
   src_reg value = *values[0];
   for (int i = 1; i < 4; i++) {
      if (!value.equals(*values[i]))
	 return false;
   }

   if (value.file != IMM)
      return false;

   if (inst->src[arg].abs) {
      if (value.type == BRW_REGISTER_TYPE_F) {
	 value.fixed_hw_reg.dw1.f = fabs(value.fixed_hw_reg.dw1.f);
      } else if (value.type == BRW_REGISTER_TYPE_D) {
	 if (value.fixed_hw_reg.dw1.d < 0)
	    value.fixed_hw_reg.dw1.d = -value.fixed_hw_reg.dw1.d;
      }
   }

   if (inst->src[arg].negate) {
      if (value.type == BRW_REGISTER_TYPE_F)
	 value.fixed_hw_reg.dw1.f = -value.fixed_hw_reg.dw1.f;
      else
	 value.fixed_hw_reg.dw1.ud = -value.fixed_hw_reg.dw1.ud;
   }

   switch (inst->opcode) {
   case BRW_OPCODE_MOV:
      inst->src[arg] = value;
      return true;

   case SHADER_OPCODE_POW:
   case SHADER_OPCODE_INT_QUOTIENT:
   case SHADER_OPCODE_INT_REMAINDER:
      if (brw->gen < 8)
         break;
      /* fallthrough */
   case BRW_OPCODE_DP2:
   case BRW_OPCODE_DP3:
   case BRW_OPCODE_DP4:
   case BRW_OPCODE_DPH:
   case BRW_OPCODE_BFI1:
   case BRW_OPCODE_ASR:
   case BRW_OPCODE_SHL:
   case BRW_OPCODE_SHR:
   case BRW_OPCODE_SUBB:
      if (arg == 1) {
         inst->src[arg] = value;
         return true;
      }
      break;

   case BRW_OPCODE_MACH:
   case BRW_OPCODE_MUL:
   case BRW_OPCODE_ADD:
   case BRW_OPCODE_OR:
   case BRW_OPCODE_AND:
   case BRW_OPCODE_XOR:
   case BRW_OPCODE_ADDC:
      if (arg == 1) {
	 inst->src[arg] = value;
	 return true;
      } else if (arg == 0 && inst->src[1].file != IMM) {
	 /* Fit this constant in by commuting the operands.  Exception: we
	  * can't do this for 32-bit integer MUL/MACH because it's asymmetric.
	  */
	 if ((inst->opcode == BRW_OPCODE_MUL ||
              inst->opcode == BRW_OPCODE_MACH) &&
	     (inst->src[1].type == BRW_REGISTER_TYPE_D ||
	      inst->src[1].type == BRW_REGISTER_TYPE_UD))
	    break;
	 inst->src[0] = inst->src[1];
	 inst->src[1] = value;
	 return true;
      }
      break;

   case BRW_OPCODE_CMP:
      if (arg == 1) {
	 inst->src[arg] = value;
	 return true;
      } else if (arg == 0 && inst->src[1].file != IMM) {
	 enum brw_conditional_mod new_cmod;

	 new_cmod = brw_swap_cmod(inst->conditional_mod);
	 if (new_cmod != ~0u) {
	    /* Fit this constant in by swapping the operands and
	     * flipping the test.
	     */
	    inst->src[0] = inst->src[1];
	    inst->src[1] = value;
	    inst->conditional_mod = new_cmod;
	    return true;
	 }
      }
      break;

   case BRW_OPCODE_SEL:
      if (arg == 1) {
	 inst->src[arg] = value;
	 return true;
      } else if (arg == 0 && inst->src[1].file != IMM) {
	 inst->src[0] = inst->src[1];
	 inst->src[1] = value;

	 /* If this was predicated, flipping operands means
	  * we also need to flip the predicate.
	  */
	 if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
	    inst->predicate_inverse = !inst->predicate_inverse;
	 }
	 return true;
      }
      break;

   default:
      break;
   }

   return false;
}

static bool
is_logic_op(enum opcode opcode)
{
   return (opcode == BRW_OPCODE_AND ||
           opcode == BRW_OPCODE_OR  ||
           opcode == BRW_OPCODE_XOR ||
           opcode == BRW_OPCODE_NOT);
}

static bool
try_copy_propagate(struct brw_context *brw, vec4_instruction *inst,
                   int arg, src_reg *values[4])
{
   /* For constant propagation, we only handle the same constant
    * across all 4 channels.  Some day, we should handle the 8-bit
    * float vector format, which would let us constant propagate
    * vectors better.
    */
   src_reg value = *values[0];
   for (int i = 1; i < 4; i++) {
      /* This is equals() except we don't care about the swizzle. */
      if (value.file != values[i]->file ||
	  value.reg != values[i]->reg ||
	  value.reg_offset != values[i]->reg_offset ||
	  value.type != values[i]->type ||
	  value.negate != values[i]->negate ||
	  value.abs != values[i]->abs) {
	 return false;
      }
   }

   /* Compute the swizzle of the original register by swizzling the
    * component loaded from each value according to the swizzle of
    * operand we're going to change.
    */
   int s[4];
   for (int i = 0; i < 4; i++) {
      s[i] = BRW_GET_SWZ(values[i]->swizzle,
			 BRW_GET_SWZ(inst->src[arg].swizzle, i));
   }
   value.swizzle = BRW_SWIZZLE4(s[0], s[1], s[2], s[3]);

   if (value.file != UNIFORM &&
       value.file != GRF &&
       value.file != ATTR)
      return false;

   if (brw->gen >= 8) {
      if (value.negate) {
         if (is_logic_op(inst->opcode)) {
            return false;
         }
      }
   }

   if (inst->src[arg].abs) {
      value.negate = false;
      value.abs = true;
   }
   if (inst->src[arg].negate)
      value.negate = !value.negate;

   bool has_source_modifiers = value.negate || value.abs;

   /* gen6 math and gen7+ SENDs from GRFs ignore source modifiers on
    * instructions.
    */
   if ((has_source_modifiers || value.file == UNIFORM ||
        value.swizzle != BRW_SWIZZLE_XYZW) && !inst->can_do_source_mods(brw))
      return false;

   if (has_source_modifiers && value.type != inst->src[arg].type)
      return false;

   bool is_3src_inst = (inst->opcode == BRW_OPCODE_LRP ||
                        inst->opcode == BRW_OPCODE_MAD ||
                        inst->opcode == BRW_OPCODE_BFE ||
                        inst->opcode == BRW_OPCODE_BFI2);
   if (is_3src_inst && value.file == UNIFORM)
      return false;

   if (inst->is_send_from_grf())
      return false;

   /* We can't copy-propagate a UD negation into a condmod
    * instruction, because the condmod ends up looking at the 33-bit
    * signed accumulator value instead of the 32-bit value we wanted
    */
   if (inst->conditional_mod &&
       value.negate &&
       value.type == BRW_REGISTER_TYPE_UD)
      return false;

   /* Don't report progress if this is a noop. */
   if (value.equals(inst->src[arg]))
      return false;

   value.type = inst->src[arg].type;
   inst->src[arg] = value;
   return true;
}

bool
vec4_visitor::opt_copy_propagation()
{
   bool progress = false;
   src_reg *cur_value[virtual_grf_reg_count][4];

   memset(&cur_value, 0, sizeof(cur_value));

   foreach_in_list(vec4_instruction, inst, &instructions) {
      /* This pass only works on basic blocks.  If there's flow
       * control, throw out all our information and start from
       * scratch.
       *
       * This should really be fixed by using a structure like in
       * src/glsl/opt_copy_propagation.cpp to track available copies.
       */
      if (!is_dominated_by_previous_instruction(inst)) {
	 memset(cur_value, 0, sizeof(cur_value));
	 continue;
      }

      /* For each source arg, see if each component comes from a copy
       * from the same type file (IMM, GRF, UNIFORM), and try
       * optimizing out access to the copy result
       */
      for (int i = 2; i >= 0; i--) {
	 /* Copied values end up in GRFs, and we don't track reladdr
	  * accesses.
	  */
	 if (inst->src[i].file != GRF ||
	     inst->src[i].reladdr)
	    continue;

	 int reg = (virtual_grf_reg_map[inst->src[i].reg] +
		    inst->src[i].reg_offset);

	 /* Find the regs that each swizzle component came from.
	  */
	 src_reg *values[4];
	 int c;
	 for (c = 0; c < 4; c++) {
	    values[c] = cur_value[reg][BRW_GET_SWZ(inst->src[i].swizzle, c)];

	    /* If there's no available copy for this channel, bail.
	     * We could be more aggressive here -- some channels might
	     * not get used based on the destination writemask.
	     */
	    if (!values[c])
	       break;

	    /* We'll only be able to copy propagate if the sources are
	     * all from the same file -- there's no ability to swizzle
	     * 0 or 1 constants in with source registers like in i915.
	     */
	    if (c > 0 && values[c - 1]->file != values[c]->file)
	       break;
	 }

	 if (c != 4)
	    continue;

	 if (try_constant_propagate(brw, inst, i, values))
            progress = true;

	 if (try_copy_propagate(brw, inst, i, values))
	    progress = true;
      }

      /* Track available source registers. */
      if (inst->dst.file == GRF) {
	 const int reg =
	    virtual_grf_reg_map[inst->dst.reg] + inst->dst.reg_offset;

	 /* Update our destination's current channel values.  For a direct copy,
	  * the value is the newly propagated source.  Otherwise, we don't know
	  * the new value, so clear it.
	  */
	 bool direct_copy = is_direct_copy(inst);
	 for (int i = 0; i < 4; i++) {
	    if (inst->dst.writemask & (1 << i)) {
	       cur_value[reg][i] = direct_copy ? &inst->src[0] : NULL;
	    }
	 }

	 /* Clear the records for any registers whose current value came from
	  * our destination's updated channels, as the two are no longer equal.
	  */
	 if (inst->dst.reladdr)
	    memset(cur_value, 0, sizeof(cur_value));
	 else {
	    for (int i = 0; i < virtual_grf_reg_count; i++) {
	       for (int j = 0; j < 4; j++) {
		  if (is_channel_updated(inst, cur_value[i], j)){
		     cur_value[i][j] = NULL;
		  }
	       }
	    }
	 }
      }
   }

   if (progress)
      invalidate_live_intervals();

   return progress;
}

} /* namespace brw */