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
38 struct aeb_entry
: public exec_node
{
39 /** The instruction that generates the expression value. */
42 /** The temporary where the value is stored. */
48 is_expression(const fs_visitor
*v
, const fs_inst
*const inst
)
50 switch (inst
->opcode
) {
64 case SHADER_OPCODE_MULH
:
74 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
75 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
:
76 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
77 case FS_OPCODE_CINTERP
:
78 case FS_OPCODE_LINTERP
:
79 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
80 case SHADER_OPCODE_BROADCAST
:
81 case SHADER_OPCODE_MOV_INDIRECT
:
84 case SHADER_OPCODE_RCP
:
85 case SHADER_OPCODE_RSQ
:
86 case SHADER_OPCODE_SQRT
:
87 case SHADER_OPCODE_EXP2
:
88 case SHADER_OPCODE_LOG2
:
89 case SHADER_OPCODE_POW
:
90 case SHADER_OPCODE_INT_QUOTIENT
:
91 case SHADER_OPCODE_INT_REMAINDER
:
92 case SHADER_OPCODE_SIN
:
93 case SHADER_OPCODE_COS
:
94 return inst
->mlen
< 2;
95 case SHADER_OPCODE_LOAD_PAYLOAD
:
96 return !inst
->is_copy_payload(v
->alloc
);
98 return inst
->is_send_from_grf() && !inst
->has_side_effects() &&
104 operands_match(const fs_inst
*a
, const fs_inst
*b
, bool *negate
)
109 if (a
->opcode
== BRW_OPCODE_MAD
) {
110 return xs
[0].equals(ys
[0]) &&
111 ((xs
[1].equals(ys
[1]) && xs
[2].equals(ys
[2])) ||
112 (xs
[2].equals(ys
[1]) && xs
[1].equals(ys
[2])));
113 } else if (a
->opcode
== BRW_OPCODE_MUL
&& a
->dst
.type
== BRW_REGISTER_TYPE_F
) {
114 bool xs0_negate
= xs
[0].negate
;
115 bool xs1_negate
= xs
[1].file
== IMM
? xs
[1].f
< 0.0f
117 bool ys0_negate
= ys
[0].negate
;
118 bool ys1_negate
= ys
[1].file
== IMM
? ys
[1].f
< 0.0f
120 float xs1_imm
= xs
[1].f
;
121 float ys1_imm
= ys
[1].f
;
123 xs
[0].negate
= false;
124 xs
[1].negate
= false;
125 ys
[0].negate
= false;
126 ys
[1].negate
= false;
127 xs
[1].f
= fabsf(xs
[1].f
);
128 ys
[1].f
= fabsf(ys
[1].f
);
130 bool ret
= (xs
[0].equals(ys
[0]) && xs
[1].equals(ys
[1])) ||
131 (xs
[1].equals(ys
[0]) && xs
[0].equals(ys
[1]));
133 xs
[0].negate
= xs0_negate
;
134 xs
[1].negate
= xs
[1].file
== IMM
? false : xs1_negate
;
135 ys
[0].negate
= ys0_negate
;
136 ys
[1].negate
= ys
[1].file
== IMM
? false : ys1_negate
;
140 *negate
= (xs0_negate
!= xs1_negate
) != (ys0_negate
!= ys1_negate
);
141 if (*negate
&& (a
->saturate
|| b
->saturate
))
144 } else if (!a
->is_commutative()) {
146 for (int i
= 0; i
< a
->sources
; i
++) {
147 if (!xs
[i
].equals(ys
[i
])) {
154 return (xs
[0].equals(ys
[0]) && xs
[1].equals(ys
[1])) ||
155 (xs
[1].equals(ys
[0]) && xs
[0].equals(ys
[1]));
160 instructions_match(fs_inst
*a
, fs_inst
*b
, bool *negate
)
162 return a
->opcode
== b
->opcode
&&
163 a
->force_writemask_all
== b
->force_writemask_all
&&
164 a
->exec_size
== b
->exec_size
&&
165 a
->group
== b
->group
&&
166 a
->saturate
== b
->saturate
&&
167 a
->predicate
== b
->predicate
&&
168 a
->predicate_inverse
== b
->predicate_inverse
&&
169 a
->conditional_mod
== b
->conditional_mod
&&
170 a
->flag_subreg
== b
->flag_subreg
&&
171 a
->dst
.type
== b
->dst
.type
&&
172 a
->offset
== b
->offset
&&
173 a
->mlen
== b
->mlen
&&
174 a
->regs_written
== b
->regs_written
&&
175 a
->base_mrf
== b
->base_mrf
&&
177 a
->header_size
== b
->header_size
&&
178 a
->shadow_compare
== b
->shadow_compare
&&
179 a
->pi_noperspective
== b
->pi_noperspective
&&
180 a
->sources
== b
->sources
&&
181 operands_match(a
, b
, negate
);
185 create_copy_instr(const fs_builder
&bld
, fs_inst
*inst
, fs_reg src
, bool negate
)
187 int written
= inst
->regs_written
;
189 DIV_ROUND_UP(inst
->dst
.component_size(inst
->exec_size
), REG_SIZE
);
192 if (inst
->opcode
== SHADER_OPCODE_LOAD_PAYLOAD
||
193 written
!= dst_width
) {
195 int sources
, header_size
;
196 if (inst
->opcode
== SHADER_OPCODE_LOAD_PAYLOAD
) {
197 sources
= inst
->sources
;
198 header_size
= inst
->header_size
;
200 assert(written
% dst_width
== 0);
201 sources
= written
/ dst_width
;
205 assert(src
.file
== VGRF
);
206 payload
= ralloc_array(bld
.shader
->mem_ctx
, fs_reg
, sources
);
207 for (int i
= 0; i
< header_size
; i
++) {
211 for (int i
= header_size
; i
< sources
; i
++) {
213 src
= offset(src
, bld
, 1);
215 copy
= bld
.LOAD_PAYLOAD(inst
->dst
, payload
, sources
, header_size
);
217 copy
= bld
.MOV(inst
->dst
, src
);
218 copy
->group
= inst
->group
;
219 copy
->force_writemask_all
= inst
->force_writemask_all
;
220 copy
->src
[0].negate
= negate
;
222 assert(copy
->regs_written
== written
);
226 fs_visitor::opt_cse_local(bblock_t
*block
)
228 bool progress
= false;
231 void *cse_ctx
= ralloc_context(NULL
);
233 int ip
= block
->start_ip
;
234 foreach_inst_in_block(fs_inst
, inst
, block
) {
235 /* Skip some cases. */
236 if (is_expression(this, inst
) && !inst
->is_partial_write() &&
237 ((inst
->dst
.file
!= ARF
&& inst
->dst
.file
!= FIXED_GRF
) ||
238 inst
->dst
.is_null()))
243 foreach_in_list_use_after(aeb_entry
, entry
, &aeb
) {
244 /* Match current instruction's expression against those in AEB. */
245 if (!(entry
->generator
->dst
.is_null() && !inst
->dst
.is_null()) &&
246 instructions_match(inst
, entry
->generator
, &negate
)) {
254 if (inst
->opcode
!= BRW_OPCODE_MOV
||
255 (inst
->opcode
== BRW_OPCODE_MOV
&&
256 inst
->src
[0].file
== IMM
&&
257 inst
->src
[0].type
== BRW_REGISTER_TYPE_VF
)) {
258 /* Our first sighting of this expression. Create an entry. */
259 aeb_entry
*entry
= ralloc(cse_ctx
, aeb_entry
);
260 entry
->tmp
= reg_undef
;
261 entry
->generator
= inst
;
262 aeb
.push_tail(entry
);
265 /* This is at least our second sighting of this expression.
266 * If we don't have a temporary already, make one.
268 bool no_existing_temp
= entry
->tmp
.file
== BAD_FILE
;
269 if (no_existing_temp
&& !entry
->generator
->dst
.is_null()) {
270 const fs_builder ibld
= fs_builder(this, block
, entry
->generator
)
271 .at(block
, entry
->generator
->next
);
272 int written
= entry
->generator
->regs_written
;
274 entry
->tmp
= fs_reg(VGRF
, alloc
.allocate(written
),
275 entry
->generator
->dst
.type
);
277 create_copy_instr(ibld
, entry
->generator
, entry
->tmp
, false);
279 entry
->generator
->dst
= entry
->tmp
;
283 if (!inst
->dst
.is_null()) {
284 assert(inst
->regs_written
== entry
->generator
->regs_written
);
285 assert(inst
->dst
.type
== entry
->tmp
.type
);
286 const fs_builder
ibld(this, block
, inst
);
288 create_copy_instr(ibld
, inst
, entry
->tmp
, negate
);
291 /* Set our iterator so that next time through the loop inst->next
292 * will get the instruction in the basic block after the one we've
295 fs_inst
*prev
= (fs_inst
*)inst
->prev
;
302 foreach_in_list_safe(aeb_entry
, entry
, &aeb
) {
303 /* Kill all AEB entries that write a different value to or read from
304 * the flag register if we just wrote it.
306 if (inst
->flags_written()) {
307 bool negate
; /* dummy */
308 if (entry
->generator
->flags_read(devinfo
) ||
309 (entry
->generator
->flags_written() &&
310 !instructions_match(inst
, entry
->generator
, &negate
))) {
317 for (int i
= 0; i
< entry
->generator
->sources
; i
++) {
318 fs_reg
*src_reg
= &entry
->generator
->src
[i
];
320 /* Kill all AEB entries that use the destination we just
323 if (inst
->overwrites_reg(entry
->generator
->src
[i
])) {
329 /* Kill any AEB entries using registers that don't get reused any
330 * more -- a sure sign they'll fail operands_match().
332 if (src_reg
->file
== VGRF
&& virtual_grf_end
[src_reg
->nr
] < ip
) {
343 ralloc_free(cse_ctx
);
349 fs_visitor::opt_cse()
351 bool progress
= false;
353 calculate_live_intervals();
355 foreach_block (block
, cfg
) {
356 progress
= opt_cse_local(block
) || progress
;
360 invalidate_live_intervals();