1 /*
2 * Copyright © 2014 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_cfg.h"
26 #include "brw_eu.h"
27
28 /** @file brw_fs_cmod_propagation.cpp
29 *
30 * Implements a pass that propagates the conditional modifier from a CMP x 0.0
31 * instruction into the instruction that generated x. For instance, in this
32 * sequence
33 *
34 * add(8) g70<1>F g69<8,8,1>F 4096F
35 * cmp.ge.f0(8) null g70<8,8,1>F 0F
36 *
37 * we can do the comparison as part of the ADD instruction directly:
38 *
39 * add.ge.f0(8) g70<1>F g69<8,8,1>F 4096F
40 *
41 * If there had been a use of the flag register and another CMP using g70
42 *
43 * add.ge.f0(8) g70<1>F g69<8,8,1>F 4096F
44 * (+f0) sel(8) g71<F> g72<8,8,1>F g73<8,8,1>F
45 * cmp.ge.f0(8) null g70<8,8,1>F 0F
46 *
47 * we can recognize that the CMP is generating the flag value that already
48 * exists and therefore remove the instruction.
49 */
50
51 static bool
opt_cmod_propagation_local(const gen_device_info * devinfo,bblock_t * block)52 opt_cmod_propagation_local(const gen_device_info *devinfo, bblock_t *block)
53 {
54 bool progress = false;
55 int ip = block->end_ip + 1;
56
57 foreach_inst_in_block_reverse_safe(fs_inst, inst, block) {
58 ip--;
59
60 if ((inst->opcode != BRW_OPCODE_AND &&
61 inst->opcode != BRW_OPCODE_CMP &&
62 inst->opcode != BRW_OPCODE_MOV) ||
63 inst->predicate != BRW_PREDICATE_NONE ||
64 !inst->dst.is_null() ||
65 inst->src[0].file != VGRF ||
66 inst->src[0].abs)
67 continue;
68
69 /* Only an AND.NZ can be propagated. Many AND.Z instructions are
70 * generated (for ir_unop_not in fs_visitor::emit_bool_to_cond_code).
71 * Propagating those would require inverting the condition on the CMP.
72 * This changes both the flag value and the register destination of the
73 * CMP. That result may be used elsewhere, so we can't change its value
74 * on a whim.
75 */
76 if (inst->opcode == BRW_OPCODE_AND &&
77 !(inst->src[1].is_one() &&
78 inst->conditional_mod == BRW_CONDITIONAL_NZ &&
79 !inst->src[0].negate))
80 continue;
81
82 if (inst->opcode == BRW_OPCODE_CMP && !inst->src[1].is_zero())
83 continue;
84
85 if (inst->opcode == BRW_OPCODE_MOV &&
86 inst->conditional_mod != BRW_CONDITIONAL_NZ)
87 continue;
88
89 bool read_flag = false;
90 foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
91 if (regions_overlap(scan_inst->dst, scan_inst->size_written,
92 inst->src[0], inst->size_read(0))) {
93 if (scan_inst->is_partial_write() ||
94 scan_inst->dst.offset != inst->src[0].offset ||
95 scan_inst->exec_size != inst->exec_size)
96 break;
97
98 /* CMP's result is the same regardless of dest type. */
99 if (inst->conditional_mod == BRW_CONDITIONAL_NZ &&
100 scan_inst->opcode == BRW_OPCODE_CMP &&
101 (inst->dst.type == BRW_REGISTER_TYPE_D ||
102 inst->dst.type == BRW_REGISTER_TYPE_UD)) {
103 inst->remove(block);
104 progress = true;
105 break;
106 }
107
108 /* If the AND wasn't handled by the previous case, it isn't safe
109 * to remove it.
110 */
111 if (inst->opcode == BRW_OPCODE_AND)
112 break;
113
114 /* Comparisons operate differently for ints and floats */
115 if (scan_inst->dst.type != inst->dst.type &&
116 (scan_inst->dst.type == BRW_REGISTER_TYPE_F ||
117 inst->dst.type == BRW_REGISTER_TYPE_F))
118 break;
119
120 /* If the instruction generating inst's source also wrote the
121 * flag, and inst is doing a simple .nz comparison, then inst
122 * is redundant - the appropriate value is already in the flag
123 * register. Delete inst.
124 */
125 if (inst->conditional_mod == BRW_CONDITIONAL_NZ &&
126 !inst->src[0].negate &&
127 scan_inst->flags_written()) {
128 inst->remove(block);
129 progress = true;
130 break;
131 }
132
133 /* The conditional mod of the CMP/CMPN instructions behaves
134 * specially because the flag output is not calculated from the
135 * result of the instruction, but the other way around, which
136 * means that even if the condmod to propagate and the condmod
137 * from the CMP instruction are the same they will in general give
138 * different results because they are evaluated based on different
139 * inputs.
140 */
141 if (scan_inst->opcode == BRW_OPCODE_CMP ||
142 scan_inst->opcode == BRW_OPCODE_CMPN)
143 break;
144
145 /* From the Sky Lake PRM Vol. 7 "Assigning Conditional Mods":
146 *
147 * * Note that the [post condition signal] bits generated at
148 * the output of a compute are before the .sat.
149 */
150 if (scan_inst->saturate)
151 break;
152
153 /* From the Sky Lake PRM, Vol 2a, "Multiply":
154 *
155 * "When multiplying integer data types, if one of the sources
156 * is a DW, the resulting full precision data is stored in
157 * the accumulator. However, if the destination data type is
158 * either W or DW, the low bits of the result are written to
159 * the destination register and the remaining high bits are
160 * discarded. This results in undefined Overflow and Sign
161 * flags. Therefore, conditional modifiers and saturation
162 * (.sat) cannot be used in this case."
163 *
164 * We just disallow cmod propagation on all integer multiplies.
165 */
166 if (!brw_reg_type_is_floating_point(scan_inst->dst.type) &&
167 scan_inst->opcode == BRW_OPCODE_MUL)
168 break;
169
170 /* Otherwise, try propagating the conditional. */
171 enum brw_conditional_mod cond =
172 inst->src[0].negate ? brw_swap_cmod(inst->conditional_mod)
173 : inst->conditional_mod;
174
175 if (scan_inst->can_do_cmod() &&
176 ((!read_flag && scan_inst->conditional_mod == BRW_CONDITIONAL_NONE) ||
177 scan_inst->conditional_mod == cond)) {
178 scan_inst->conditional_mod = cond;
179 inst->remove(block);
180 progress = true;
181 }
182 break;
183 }
184
185 if (scan_inst->flags_written())
186 break;
187
188 read_flag = read_flag || scan_inst->flags_read(devinfo);
189 }
190 }
191
192 return progress;
193 }
194
195 bool
opt_cmod_propagation()196 fs_visitor::opt_cmod_propagation()
197 {
198 bool progress = false;
199
200 foreach_block_reverse(block, cfg) {
201 progress = opt_cmod_propagation_local(devinfo, block) || progress;
202 }
203
204 if (progress)
205 invalidate_live_intervals();
206
207 return progress;
208 }
209