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_FB_READ_LOGICAL
:
75 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
76 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL
:
77 case FS_OPCODE_LINTERP
:
78 case SHADER_OPCODE_FIND_LIVE_CHANNEL
:
79 case SHADER_OPCODE_BROADCAST
:
80 case SHADER_OPCODE_MOV_INDIRECT
:
81 case SHADER_OPCODE_TEX_LOGICAL
:
82 case SHADER_OPCODE_TXD_LOGICAL
:
83 case SHADER_OPCODE_TXF_LOGICAL
:
84 case SHADER_OPCODE_TXL_LOGICAL
:
85 case SHADER_OPCODE_TXS_LOGICAL
:
86 case FS_OPCODE_TXB_LOGICAL
:
87 case SHADER_OPCODE_TXF_CMS_LOGICAL
:
88 case SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
89 case SHADER_OPCODE_TXF_UMS_LOGICAL
:
90 case SHADER_OPCODE_TXF_MCS_LOGICAL
:
91 case SHADER_OPCODE_LOD_LOGICAL
:
92 case SHADER_OPCODE_TG4_LOGICAL
:
93 case SHADER_OPCODE_TG4_OFFSET_LOGICAL
:
96 case SHADER_OPCODE_RCP
:
97 case SHADER_OPCODE_RSQ
:
98 case SHADER_OPCODE_SQRT
:
99 case SHADER_OPCODE_EXP2
:
100 case SHADER_OPCODE_LOG2
:
101 case SHADER_OPCODE_POW
:
102 case SHADER_OPCODE_INT_QUOTIENT
:
103 case SHADER_OPCODE_INT_REMAINDER
:
104 case SHADER_OPCODE_SIN
:
105 case SHADER_OPCODE_COS
:
106 return inst
->mlen
< 2;
107 case SHADER_OPCODE_LOAD_PAYLOAD
:
108 return !inst
->is_copy_payload(v
->alloc
);
110 return inst
->is_send_from_grf() && !inst
->has_side_effects() &&
111 !inst
->is_volatile();
116 operands_match(const fs_inst
*a
, const fs_inst
*b
, bool *negate
)
121 if (a
->opcode
== BRW_OPCODE_MAD
) {
122 return xs
[0].equals(ys
[0]) &&
123 ((xs
[1].equals(ys
[1]) && xs
[2].equals(ys
[2])) ||
124 (xs
[2].equals(ys
[1]) && xs
[1].equals(ys
[2])));
125 } else if (a
->opcode
== BRW_OPCODE_MUL
&& a
->dst
.type
== BRW_REGISTER_TYPE_F
) {
126 bool xs0_negate
= xs
[0].negate
;
127 bool xs1_negate
= xs
[1].file
== IMM
? xs
[1].f
< 0.0f
129 bool ys0_negate
= ys
[0].negate
;
130 bool ys1_negate
= ys
[1].file
== IMM
? ys
[1].f
< 0.0f
132 float xs1_imm
= xs
[1].f
;
133 float ys1_imm
= ys
[1].f
;
135 xs
[0].negate
= false;
136 xs
[1].negate
= false;
137 ys
[0].negate
= false;
138 ys
[1].negate
= false;
139 xs
[1].f
= fabsf(xs
[1].f
);
140 ys
[1].f
= fabsf(ys
[1].f
);
142 bool ret
= (xs
[0].equals(ys
[0]) && xs
[1].equals(ys
[1])) ||
143 (xs
[1].equals(ys
[0]) && xs
[0].equals(ys
[1]));
145 xs
[0].negate
= xs0_negate
;
146 xs
[1].negate
= xs
[1].file
== IMM
? false : xs1_negate
;
147 ys
[0].negate
= ys0_negate
;
148 ys
[1].negate
= ys
[1].file
== IMM
? false : ys1_negate
;
152 *negate
= (xs0_negate
!= xs1_negate
) != (ys0_negate
!= ys1_negate
);
153 if (*negate
&& (a
->saturate
|| b
->saturate
))
156 } else if (!a
->is_commutative()) {
158 for (int i
= 0; i
< a
->sources
; i
++) {
159 if (!xs
[i
].equals(ys
[i
])) {
166 return (xs
[0].equals(ys
[0]) && xs
[1].equals(ys
[1])) ||
167 (xs
[1].equals(ys
[0]) && xs
[0].equals(ys
[1]));
172 instructions_match(fs_inst
*a
, fs_inst
*b
, bool *negate
)
174 return a
->opcode
== b
->opcode
&&
175 a
->force_writemask_all
== b
->force_writemask_all
&&
176 a
->exec_size
== b
->exec_size
&&
177 a
->group
== b
->group
&&
178 a
->saturate
== b
->saturate
&&
179 a
->predicate
== b
->predicate
&&
180 a
->predicate_inverse
== b
->predicate_inverse
&&
181 a
->conditional_mod
== b
->conditional_mod
&&
182 a
->flag_subreg
== b
->flag_subreg
&&
183 a
->dst
.type
== b
->dst
.type
&&
184 a
->offset
== b
->offset
&&
185 a
->mlen
== b
->mlen
&&
186 a
->ex_mlen
== b
->ex_mlen
&&
187 a
->sfid
== b
->sfid
&&
188 a
->desc
== b
->desc
&&
189 a
->size_written
== b
->size_written
&&
190 a
->base_mrf
== b
->base_mrf
&&
191 a
->check_tdr
== b
->check_tdr
&&
192 a
->send_has_side_effects
== b
->send_has_side_effects
&&
194 a
->header_size
== b
->header_size
&&
195 a
->shadow_compare
== b
->shadow_compare
&&
196 a
->pi_noperspective
== b
->pi_noperspective
&&
197 a
->target
== b
->target
&&
198 a
->sources
== b
->sources
&&
199 operands_match(a
, b
, negate
);
203 create_copy_instr(const fs_builder
&bld
, fs_inst
*inst
, fs_reg src
, bool negate
)
205 unsigned written
= regs_written(inst
);
207 DIV_ROUND_UP(inst
->dst
.component_size(inst
->exec_size
), REG_SIZE
);
210 if (inst
->opcode
== SHADER_OPCODE_LOAD_PAYLOAD
) {
211 assert(src
.file
== VGRF
);
212 fs_reg
*payload
= ralloc_array(bld
.shader
->mem_ctx
, fs_reg
,
214 for (int i
= 0; i
< inst
->header_size
; i
++) {
216 src
.offset
+= REG_SIZE
;
218 for (int i
= inst
->header_size
; i
< inst
->sources
; i
++) {
219 src
.type
= inst
->src
[i
].type
;
221 src
= offset(src
, bld
, 1);
223 copy
= bld
.LOAD_PAYLOAD(inst
->dst
, payload
, inst
->sources
,
225 } else if (written
!= dst_width
) {
226 assert(src
.file
== VGRF
);
227 assert(written
% dst_width
== 0);
228 const int sources
= written
/ dst_width
;
229 fs_reg
*payload
= ralloc_array(bld
.shader
->mem_ctx
, fs_reg
, sources
);
230 for (int i
= 0; i
< sources
; i
++) {
232 src
= offset(src
, bld
, 1);
234 copy
= bld
.LOAD_PAYLOAD(inst
->dst
, payload
, sources
, 0);
236 copy
= bld
.MOV(inst
->dst
, src
);
237 copy
->group
= inst
->group
;
238 copy
->force_writemask_all
= inst
->force_writemask_all
;
239 copy
->src
[0].negate
= negate
;
241 assert(regs_written(copy
) == written
);
245 fs_visitor::opt_cse_local(bblock_t
*block
)
247 bool progress
= false;
250 void *cse_ctx
= ralloc_context(NULL
);
252 int ip
= block
->start_ip
;
253 foreach_inst_in_block(fs_inst
, inst
, block
) {
254 /* Skip some cases. */
255 if (is_expression(this, inst
) && !inst
->is_partial_write() &&
256 ((inst
->dst
.file
!= ARF
&& inst
->dst
.file
!= FIXED_GRF
) ||
257 inst
->dst
.is_null()))
262 foreach_in_list_use_after(aeb_entry
, entry
, &aeb
) {
263 /* Match current instruction's expression against those in AEB. */
264 if (!(entry
->generator
->dst
.is_null() && !inst
->dst
.is_null()) &&
265 instructions_match(inst
, entry
->generator
, &negate
)) {
273 if (inst
->opcode
!= BRW_OPCODE_MOV
||
274 (inst
->opcode
== BRW_OPCODE_MOV
&&
275 inst
->src
[0].file
== IMM
&&
276 inst
->src
[0].type
== BRW_REGISTER_TYPE_VF
)) {
277 /* Our first sighting of this expression. Create an entry. */
278 aeb_entry
*entry
= ralloc(cse_ctx
, aeb_entry
);
279 entry
->tmp
= reg_undef
;
280 entry
->generator
= inst
;
281 aeb
.push_tail(entry
);
284 /* This is at least our second sighting of this expression.
285 * If we don't have a temporary already, make one.
287 bool no_existing_temp
= entry
->tmp
.file
== BAD_FILE
;
288 if (no_existing_temp
&& !entry
->generator
->dst
.is_null()) {
289 const fs_builder ibld
= fs_builder(this, block
, entry
->generator
)
290 .at(block
, entry
->generator
->next
);
291 int written
= regs_written(entry
->generator
);
293 entry
->tmp
= fs_reg(VGRF
, alloc
.allocate(written
),
294 entry
->generator
->dst
.type
);
296 create_copy_instr(ibld
, entry
->generator
, entry
->tmp
, false);
298 entry
->generator
->dst
= entry
->tmp
;
302 if (!inst
->dst
.is_null()) {
303 assert(inst
->size_written
== entry
->generator
->size_written
);
304 assert(inst
->dst
.type
== entry
->tmp
.type
);
305 const fs_builder
ibld(this, block
, inst
);
307 create_copy_instr(ibld
, inst
, entry
->tmp
, negate
);
310 /* Set our iterator so that next time through the loop inst->next
311 * will get the instruction in the basic block after the one we've
314 fs_inst
*prev
= (fs_inst
*)inst
->prev
;
321 foreach_in_list_safe(aeb_entry
, entry
, &aeb
) {
322 /* Kill all AEB entries that write a different value to or read from
323 * the flag register if we just wrote it.
325 if (inst
->flags_written()) {
326 bool negate
; /* dummy */
327 if (entry
->generator
->flags_read(devinfo
) ||
328 (entry
->generator
->flags_written() &&
329 !instructions_match(inst
, entry
->generator
, &negate
))) {
336 for (int i
= 0; i
< entry
->generator
->sources
; i
++) {
337 fs_reg
*src_reg
= &entry
->generator
->src
[i
];
339 /* Kill all AEB entries that use the destination we just
342 if (regions_overlap(inst
->dst
, inst
->size_written
,
343 entry
->generator
->src
[i
],
344 entry
->generator
->size_read(i
))) {
350 /* Kill any AEB entries using registers that don't get reused any
351 * more -- a sure sign they'll fail operands_match().
353 if (src_reg
->file
== VGRF
&& virtual_grf_end
[src_reg
->nr
] < ip
) {
364 ralloc_free(cse_ctx
);
370 fs_visitor::opt_cse()
372 bool progress
= false;
374 calculate_live_intervals();
376 foreach_block (block
, cfg
) {
377 progress
= opt_cse_local(block
) || progress
;
381 invalidate_live_intervals();