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
27 /** @file brw_fs_cse.cpp
29 * Support for local common subexpression elimination.
31 * See Muchnick's Advanced Compiler Design and Implementation, section
36 struct aeb_entry
: public exec_node
{
37 /** The instruction that generates the expression value. */
40 /** The temporary where the value is stored. */
46 is_expression(const fs_visitor
*v
, const fs_inst
*const inst
)
48 switch (inst
->opcode
) {
71 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
72 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
73 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
74 case FS_OPCODE_CINTERP
:
75 case FS_OPCODE_LINTERP
:
76 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
77 case SHADER_OPCODE_BROADCAST
:
79 case SHADER_OPCODE_RCP
:
80 case SHADER_OPCODE_RSQ
:
81 case SHADER_OPCODE_SQRT
:
82 case SHADER_OPCODE_EXP2
:
83 case SHADER_OPCODE_LOG2
:
84 case SHADER_OPCODE_POW
:
85 case SHADER_OPCODE_INT_QUOTIENT
:
86 case SHADER_OPCODE_INT_REMAINDER
:
87 case SHADER_OPCODE_SIN
:
88 case SHADER_OPCODE_COS
:
89 return inst
->mlen
< 2;
90 case SHADER_OPCODE_LOAD_PAYLOAD
:
91 return !inst
->is_copy_payload(v
->alloc
);
93 return inst
->is_send_from_grf() && !inst
->has_side_effects();
98 operands_match(const fs_inst
*a
, const fs_inst
*b
, bool *negate
)
103 if (a
->opcode
== BRW_OPCODE_MAD
) {
104 return xs
[0].equals(ys
[0]) &&
105 ((xs
[1].equals(ys
[1]) && xs
[2].equals(ys
[2])) ||
106 (xs
[2].equals(ys
[1]) && xs
[1].equals(ys
[2])));
107 } else if (a
->opcode
== BRW_OPCODE_MUL
&& a
->dst
.type
== BRW_REGISTER_TYPE_F
) {
108 bool xs0_negate
= xs
[0].negate
;
109 bool xs1_negate
= xs
[1].file
== IMM
? xs
[1].fixed_hw_reg
.dw1
.f
< 0.0f
111 bool ys0_negate
= ys
[0].negate
;
112 bool ys1_negate
= ys
[1].file
== IMM
? ys
[1].fixed_hw_reg
.dw1
.f
< 0.0f
114 float xs1_imm
= xs
[1].fixed_hw_reg
.dw1
.f
;
115 float ys1_imm
= ys
[1].fixed_hw_reg
.dw1
.f
;
117 xs
[0].negate
= false;
118 xs
[1].negate
= false;
119 ys
[0].negate
= false;
120 ys
[1].negate
= false;
121 xs
[1].fixed_hw_reg
.dw1
.f
= fabsf(xs
[1].fixed_hw_reg
.dw1
.f
);
122 ys
[1].fixed_hw_reg
.dw1
.f
= fabsf(ys
[1].fixed_hw_reg
.dw1
.f
);
124 bool ret
= (xs
[0].equals(ys
[0]) && xs
[1].equals(ys
[1])) ||
125 (xs
[1].equals(ys
[0]) && xs
[0].equals(ys
[1]));
127 xs
[0].negate
= xs0_negate
;
128 xs
[1].negate
= xs
[1].file
== IMM
? false : xs1_negate
;
129 ys
[0].negate
= ys0_negate
;
130 ys
[1].negate
= ys
[1].file
== IMM
? false : ys1_negate
;
131 xs
[1].fixed_hw_reg
.dw1
.f
= xs1_imm
;
132 ys
[1].fixed_hw_reg
.dw1
.f
= ys1_imm
;
134 *negate
= (xs0_negate
!= xs1_negate
) != (ys0_negate
!= ys1_negate
);
136 } else if (!a
->is_commutative()) {
138 for (int i
= 0; i
< a
->sources
; i
++) {
139 if (!xs
[i
].equals(ys
[i
])) {
146 return (xs
[0].equals(ys
[0]) && xs
[1].equals(ys
[1])) ||
147 (xs
[1].equals(ys
[0]) && xs
[0].equals(ys
[1]));
152 instructions_match(fs_inst
*a
, fs_inst
*b
, bool *negate
)
154 return a
->opcode
== b
->opcode
&&
155 a
->saturate
== b
->saturate
&&
156 a
->predicate
== b
->predicate
&&
157 a
->predicate_inverse
== b
->predicate_inverse
&&
158 a
->conditional_mod
== b
->conditional_mod
&&
159 a
->dst
.type
== b
->dst
.type
&&
160 a
->sources
== b
->sources
&&
161 (a
->is_tex() ? (a
->offset
== b
->offset
&&
162 a
->mlen
== b
->mlen
&&
163 a
->regs_written
== b
->regs_written
&&
164 a
->base_mrf
== b
->base_mrf
&&
166 a
->header_size
== b
->header_size
&&
167 a
->shadow_compare
== b
->shadow_compare
)
169 operands_match(a
, b
, negate
);
173 create_copy_instr(fs_visitor
*v
, fs_inst
*inst
, fs_reg src
, bool negate
)
175 int written
= inst
->regs_written
;
176 int dst_width
= inst
->dst
.width
/ 8;
179 if (written
> dst_width
) {
181 int sources
, header_size
;
182 if (inst
->opcode
== SHADER_OPCODE_LOAD_PAYLOAD
) {
183 sources
= inst
->sources
;
184 header_size
= inst
->header_size
;
186 assert(written
% dst_width
== 0);
187 sources
= written
/ dst_width
;
191 assert(src
.file
== GRF
);
192 payload
= ralloc_array(v
->mem_ctx
, fs_reg
, sources
);
193 for (int i
= 0; i
< header_size
; i
++) {
195 payload
[i
].width
= 8;
198 for (int i
= header_size
; i
< sources
; i
++) {
200 src
= offset(src
, 1);
202 copy
= v
->LOAD_PAYLOAD(inst
->dst
, payload
, sources
, header_size
);
204 copy
= v
->MOV(inst
->dst
, src
);
205 copy
->force_writemask_all
= inst
->force_writemask_all
;
206 copy
->src
[0].negate
= negate
;
208 assert(copy
->regs_written
== written
);
214 fs_visitor::opt_cse_local(bblock_t
*block
)
216 bool progress
= false;
219 void *cse_ctx
= ralloc_context(NULL
);
221 int ip
= block
->start_ip
;
222 foreach_inst_in_block(fs_inst
, inst
, block
) {
223 /* Skip some cases. */
224 if (is_expression(this, inst
) && !inst
->is_partial_write() &&
225 (inst
->dst
.file
!= HW_REG
|| inst
->dst
.is_null()))
230 foreach_in_list_use_after(aeb_entry
, entry
, &aeb
) {
231 /* Match current instruction's expression against those in AEB. */
232 if (!(entry
->generator
->dst
.is_null() && !inst
->dst
.is_null()) &&
233 instructions_match(inst
, entry
->generator
, &negate
)) {
241 if (inst
->opcode
!= BRW_OPCODE_MOV
||
242 (inst
->opcode
== BRW_OPCODE_MOV
&&
243 inst
->src
[0].file
== IMM
&&
244 inst
->src
[0].type
== BRW_REGISTER_TYPE_VF
)) {
245 /* Our first sighting of this expression. Create an entry. */
246 aeb_entry
*entry
= ralloc(cse_ctx
, aeb_entry
);
247 entry
->tmp
= reg_undef
;
248 entry
->generator
= inst
;
249 aeb
.push_tail(entry
);
252 /* This is at least our second sighting of this expression.
253 * If we don't have a temporary already, make one.
255 bool no_existing_temp
= entry
->tmp
.file
== BAD_FILE
;
256 if (no_existing_temp
&& !entry
->generator
->dst
.is_null()) {
257 int written
= entry
->generator
->regs_written
;
258 assert((written
* 8) % entry
->generator
->dst
.width
== 0);
260 entry
->tmp
= fs_reg(GRF
, alloc
.allocate(written
),
261 entry
->generator
->dst
.type
,
262 entry
->generator
->dst
.width
);
264 fs_inst
*copy
= create_copy_instr(this, entry
->generator
,
266 entry
->generator
->insert_after(block
, copy
);
268 entry
->generator
->dst
= entry
->tmp
;
272 if (!inst
->dst
.is_null()) {
273 assert(inst
->regs_written
== entry
->generator
->regs_written
);
274 assert(inst
->dst
.width
== entry
->generator
->dst
.width
);
275 assert(inst
->dst
.type
== entry
->tmp
.type
);
277 fs_inst
*copy
= create_copy_instr(this, inst
,
279 inst
->insert_before(block
, copy
);
282 /* Set our iterator so that next time through the loop inst->next
283 * will get the instruction in the basic block after the one we've
286 fs_inst
*prev
= (fs_inst
*)inst
->prev
;
293 foreach_in_list_safe(aeb_entry
, entry
, &aeb
) {
294 /* Kill all AEB entries that write a different value to or read from
295 * the flag register if we just wrote it.
297 if (inst
->writes_flag()) {
298 bool negate
; /* dummy */
299 if (entry
->generator
->reads_flag() ||
300 (entry
->generator
->writes_flag() &&
301 !instructions_match(inst
, entry
->generator
, &negate
))) {
308 for (int i
= 0; i
< entry
->generator
->sources
; i
++) {
309 fs_reg
*src_reg
= &entry
->generator
->src
[i
];
311 /* Kill all AEB entries that use the destination we just
314 if (inst
->overwrites_reg(entry
->generator
->src
[i
])) {
320 /* Kill any AEB entries using registers that don't get reused any
321 * more -- a sure sign they'll fail operands_match().
323 if (src_reg
->file
== GRF
&& virtual_grf_end
[src_reg
->reg
] < ip
) {
334 ralloc_free(cse_ctx
);
340 fs_visitor::opt_cse()
342 bool progress
= false;
344 calculate_live_intervals();
346 foreach_block (block
, cfg
) {
347 progress
= opt_cse_local(block
) || progress
;
351 invalidate_live_intervals();