1 /*
2  * Copyright © 2013 Intel Corporation
3  *
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:
10  *
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
13  * Software.
14  *
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
21  * IN THE SOFTWARE.
22  */
23 
24 #include "brw_fs.h"
25 #include "brw_fs_live_variables.h"
26 #include "brw_cfg.h"
27 
28 /** @file brw_fs_saturate_propagation.cpp
29  *
30  * Implements a pass that propagates the SAT modifier from a MOV.SAT into the
31  * instruction that produced the source of the MOV.SAT, thereby allowing the
32  * MOV's src and dst to be coalesced and the MOV removed.
33  *
34  * For instance,
35  *
36  *    ADD     tmp, src0, src1
37  *    MOV.SAT dst, tmp
38  *
39  * would be transformed into
40  *
41  *    ADD.SAT tmp, src0, src1
42  *    MOV     dst, tmp
43  */
44 
45 static bool
46 opt_saturate_propagation_local(fs_visitor *v, bblock_t *block)
47 {
48    bool progress = false;
49    int ip = block->end_ip + 1;
50 
51    foreach_inst_in_block_reverse(fs_inst, inst, block) {
52       ip--;
53 
54       if (inst->opcode != BRW_OPCODE_MOV ||
55           !inst->saturate ||
56           inst->dst.file != VGRF ||
57           inst->dst.type != inst->src[0].type ||
58           inst->src[0].file != VGRF ||
59           inst->src[0].abs)
60          continue;
61 
62       int src_var = v->live_intervals->var_from_reg(inst->src[0]);
63       int src_end_ip = v->live_intervals->end[src_var];
64 
65       bool interfered = false;
66       foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
67          if (regions_overlap(scan_inst->dst, scan_inst->size_written,
68                              inst->src[0], inst->size_read(0))) {
69             if (scan_inst->is_partial_write() ||
70                 (scan_inst->dst.type != inst->dst.type &&
71                  !scan_inst->can_change_types()))
72                break;
73 
74             if (scan_inst->saturate) {
75                inst->saturate = false;
76                progress = true;
77             } else if (src_end_ip == ip || inst->dst.equals(inst->src[0])) {
78                if (scan_inst->can_do_saturate()) {
79                   if (scan_inst->dst.type != inst->dst.type) {
80                      scan_inst->dst.type = inst->dst.type;
81                      for (int i = 0; i < scan_inst->sources; i++) {
82                         scan_inst->src[i].type = inst->dst.type;
83                      }
84                   }
85 
86                   if (inst->src[0].negate) {
87                      if (scan_inst->opcode == BRW_OPCODE_MUL) {
88                         scan_inst->src[0].negate = !scan_inst->src[0].negate;
89                         inst->src[0].negate = false;
90                      } else if (scan_inst->opcode == BRW_OPCODE_MAD) {
91                         scan_inst->src[0].negate = !scan_inst->src[0].negate;
92                         scan_inst->src[1].negate = !scan_inst->src[1].negate;
93                         inst->src[0].negate = false;
94                      } else if (scan_inst->opcode == BRW_OPCODE_ADD) {
95                         if (scan_inst->src[1].file == IMM) {
96                            if (!brw_negate_immediate(scan_inst->src[1].type,
97                                                      &scan_inst->src[1].as_brw_reg())) {
98                               break;
99                            }
100                         } else {
101                            scan_inst->src[1].negate = !scan_inst->src[1].negate;
102                         }
103                         scan_inst->src[0].negate = !scan_inst->src[0].negate;
104                         inst->src[0].negate = false;
105                      } else {
106                         break;
107                      }
108                   }
109 
110                   scan_inst->saturate = true;
111                   inst->saturate = false;
112                   progress = true;
113                }
114             }
115             break;
116          }
117          for (int i = 0; i < scan_inst->sources; i++) {
118             if (scan_inst->src[i].file == VGRF &&
119                 scan_inst->src[i].nr == inst->src[0].nr &&
120                 scan_inst->src[i].offset / REG_SIZE ==
121                  inst->src[0].offset / REG_SIZE) {
122                if (scan_inst->opcode != BRW_OPCODE_MOV ||
123                    !scan_inst->saturate ||
124                    scan_inst->src[0].abs ||
125                    scan_inst->src[0].negate ||
126                    scan_inst->src[0].abs != inst->src[0].abs ||
127                    scan_inst->src[0].negate != inst->src[0].negate) {
128                   interfered = true;
129                   break;
130                }
131             }
132          }
133 
134          if (interfered)
135             break;
136       }
137    }
138 
139    return progress;
140 }
141 
142 bool
143 fs_visitor::opt_saturate_propagation()
144 {
145    bool progress = false;
146 
147    calculate_live_intervals();
148 
149    foreach_block (block, cfg) {
150       progress = opt_saturate_propagation_local(this, block) || progress;
151    }
152 
153    /* Live intervals are still valid. */
154 
155    return progress;
156 }
157