2 * Copyright © 2018 Valve 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
24 * Daniel Schürmann (daniel.schuermann@campus.tu-berlin.de)
25 * Bas Nieuwenhuizen (bas@basnieuwenhuizen.nl)
32 #include <unordered_map>
36 #include "util/u_math.h"
45 assignment() = default;
46 assignment(PhysReg reg
, RegClass rc
) : reg(reg
), rc(rc
), assigned(-1) {}
52 std::set
<Instruction
*> uses
;
56 std::bitset
<512> war_hint
;
58 std::vector
<assignment
> assignments
;
59 std::vector
<std::unordered_map
<unsigned, Temp
>> renames
;
60 std::vector
<std::vector
<Instruction
*>> incomplete_phis
;
61 std::vector
<bool> filled
;
62 std::vector
<bool> sealed
;
63 std::unordered_map
<unsigned, Temp
> orig_names
;
64 std::unordered_map
<unsigned, phi_info
> phi_map
;
65 std::unordered_map
<unsigned, unsigned> affinities
;
66 std::unordered_map
<unsigned, Instruction
*> vectors
;
67 aco_ptr
<Instruction
> pseudo_dummy
;
68 unsigned max_used_sgpr
= 0;
69 unsigned max_used_vgpr
= 0;
70 std::bitset
<64> defs_done
; /* see MAX_ARGS in aco_instruction_selection_setup.cpp */
72 ra_ctx(Program
* program
) : program(program
),
73 assignments(program
->peekAllocationId()),
74 renames(program
->blocks
.size()),
75 incomplete_phis(program
->blocks
.size()),
76 filled(program
->blocks
.size()),
77 sealed(program
->blocks
.size())
79 pseudo_dummy
.reset(create_instruction
<Instruction
>(aco_opcode::p_parallelcopy
, Format::PSEUDO
, 0, 0));
83 bool instr_can_access_subdword(aco_ptr
<Instruction
>& instr
)
85 return instr
->isSDWA() || instr
->format
== Format::PSEUDO
;
95 DefInfo(ra_ctx
& ctx
, aco_ptr
<Instruction
>& instr
, RegClass rc
) : rc(rc
) {
99 if (rc
.type() == RegType::vgpr
) {
101 ub
= 256 + ctx
.program
->max_reg_demand
.vgpr
;
104 ub
= ctx
.program
->max_reg_demand
.sgpr
;
111 if (rc
.is_subdword()) {
112 /* stride in bytes */
113 if(!instr_can_access_subdword(instr
))
115 else if (rc
.bytes() % 4 == 0)
117 else if (rc
.bytes() % 2 == 0)
125 RegisterFile() {regs
.fill(0);}
127 std::array
<uint32_t, 512> regs
;
128 std::map
<uint32_t, std::array
<uint32_t, 4>> subdword_regs
;
130 const uint32_t& operator [] (unsigned index
) const {
134 uint32_t& operator [] (unsigned index
) {
138 unsigned count_zero(PhysReg start
, unsigned size
) {
140 for (unsigned i
= 0; i
< size
; i
++)
141 res
+= !regs
[start
+ i
];
145 bool test(PhysReg start
, unsigned num_bytes
) {
146 for (PhysReg i
= start
; i
.reg_b
< start
.reg_b
+ num_bytes
; i
= PhysReg(i
+ 1)) {
147 if (regs
[i
] & 0x0FFFFFFF)
149 if (regs
[i
] == 0xF0000000) {
150 assert(subdword_regs
.find(i
) != subdword_regs
.end());
151 for (unsigned j
= i
.byte(); i
* 4 + j
< start
.reg_b
+ num_bytes
&& j
< 4; j
++) {
152 if (subdword_regs
[i
][j
])
160 void block(PhysReg start
, unsigned num_bytes
) {
161 if (start
.byte() || num_bytes
% 4)
162 fill_subdword(start
, num_bytes
, 0xFFFFFFFF);
164 fill(start
, num_bytes
/ 4, 0xFFFFFFFF);
167 bool is_blocked(PhysReg start
) {
168 if (regs
[start
] == 0xFFFFFFFF)
170 if (regs
[start
] == 0xF0000000) {
171 for (unsigned i
= start
.byte(); i
< 4; i
++)
172 if (subdword_regs
[start
][i
] == 0xFFFFFFFF)
178 void clear(PhysReg start
, RegClass rc
) {
179 if (rc
.is_subdword())
180 fill_subdword(start
, rc
.bytes(), 0);
182 fill(start
, rc
.size(), 0);
185 void fill(Operand op
) {
186 if (op
.regClass().is_subdword())
187 fill_subdword(op
.physReg(), op
.bytes(), op
.tempId());
189 fill(op
.physReg(), op
.size(), op
.tempId());
192 void clear(Operand op
) {
193 clear(op
.physReg(), op
.regClass());
196 void fill(Definition def
) {
197 if (def
.regClass().is_subdword())
198 fill_subdword(def
.physReg(), def
.bytes(), def
.tempId());
200 fill(def
.physReg(), def
.size(), def
.tempId());
203 void clear(Definition def
) {
204 clear(def
.physReg(), def
.regClass());
208 void fill(PhysReg start
, unsigned size
, uint32_t val
) {
209 for (unsigned i
= 0; i
< size
; i
++)
210 regs
[start
+ i
] = val
;
213 void fill_subdword(PhysReg start
, unsigned num_bytes
, uint32_t val
) {
214 fill(start
, DIV_ROUND_UP(num_bytes
, 4), 0xF0000000);
215 for (PhysReg i
= start
; i
.reg_b
< start
.reg_b
+ num_bytes
; i
= PhysReg(i
+ 1)) {
217 std::array
<uint32_t, 4>& sub
= subdword_regs
.emplace(i
, std::array
<uint32_t, 4>{0, 0, 0, 0}).first
->second
;
218 for (unsigned j
= i
.byte(); i
* 4 + j
< start
.reg_b
+ num_bytes
&& j
< 4; j
++)
221 if (sub
== std::array
<uint32_t, 4>{0, 0, 0, 0}) {
222 subdword_regs
.erase(i
);
230 /* helper function for debugging */
232 void print_regs(ra_ctx
& ctx
, bool vgprs
, RegisterFile
& reg_file
)
234 unsigned max
= vgprs
? ctx
.program
->max_reg_demand
.vgpr
: ctx
.program
->max_reg_demand
.sgpr
;
235 unsigned lb
= vgprs
? 256 : 0;
236 unsigned ub
= lb
+ max
;
237 char reg_char
= vgprs
? 'v' : 's';
241 for (unsigned i
= lb
; i
< ub
; i
+= 3) {
242 printf("%.2u ", i
- lb
);
247 printf("%cgprs: ", reg_char
);
248 unsigned free_regs
= 0;
250 bool char_select
= false;
251 for (unsigned i
= lb
; i
< ub
; i
++) {
252 if (reg_file
[i
] == 0xFFFF) {
254 } else if (reg_file
[i
]) {
255 if (reg_file
[i
] != prev
) {
257 char_select
= !char_select
;
259 printf(char_select
? "#" : "@");
267 printf("%u/%u used, %u/%u free\n", max
- free_regs
, max
, free_regs
, max
);
269 /* print assignments */
272 for (unsigned i
= lb
; i
< ub
; i
++) {
273 if (reg_file
[i
] != prev
) {
274 if (prev
&& size
> 1)
275 printf("-%d]\n", i
- 1 - lb
);
279 if (prev
&& prev
!= 0xFFFF) {
280 if (ctx
.orig_names
.count(reg_file
[i
]) && ctx
.orig_names
[reg_file
[i
]].id() != reg_file
[i
])
281 printf("%%%u (was %%%d) = %c[%d", reg_file
[i
], ctx
.orig_names
[reg_file
[i
]].id(), reg_char
, i
- lb
);
283 printf("%%%u = %c[%d", reg_file
[i
], reg_char
, i
- lb
);
290 if (prev
&& size
> 1)
291 printf("-%d]\n", ub
- lb
- 1);
298 void adjust_max_used_regs(ra_ctx
& ctx
, RegClass rc
, unsigned reg
)
300 unsigned max_addressible_sgpr
= ctx
.program
->sgpr_limit
;
301 unsigned size
= rc
.size();
302 if (rc
.type() == RegType::vgpr
) {
304 unsigned hi
= reg
- 256 + size
- 1;
305 ctx
.max_used_vgpr
= std::max(ctx
.max_used_vgpr
, hi
);
306 } else if (reg
+ rc
.size() <= max_addressible_sgpr
) {
307 unsigned hi
= reg
+ size
- 1;
308 ctx
.max_used_sgpr
= std::max(ctx
.max_used_sgpr
, std::min(hi
, max_addressible_sgpr
));
313 void update_renames(ra_ctx
& ctx
, RegisterFile
& reg_file
,
314 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
315 aco_ptr
<Instruction
>& instr
)
317 /* allocate id's and rename operands: this is done transparently here */
318 for (std::pair
<Operand
, Definition
>& copy
: parallelcopies
) {
319 /* the definitions with id are not from this function and already handled */
320 if (copy
.second
.isTemp())
323 /* check if we we moved another parallelcopy definition */
324 for (std::pair
<Operand
, Definition
>& other
: parallelcopies
) {
325 if (!other
.second
.isTemp())
327 if (copy
.first
.getTemp() == other
.second
.getTemp()) {
328 copy
.first
.setTemp(other
.first
.getTemp());
329 copy
.first
.setFixed(other
.first
.physReg());
332 // FIXME: if a definition got moved, change the target location and remove the parallelcopy
333 copy
.second
.setTemp(Temp(ctx
.program
->allocateId(), copy
.second
.regClass()));
334 ctx
.assignments
.emplace_back(copy
.second
.physReg(), copy
.second
.regClass());
335 assert(ctx
.assignments
.size() == ctx
.program
->peekAllocationId());
336 reg_file
.fill(copy
.second
);
338 /* check if we moved an operand */
339 for (Operand
& op
: instr
->operands
) {
342 if (op
.tempId() == copy
.first
.tempId()) {
343 bool omit_renaming
= instr
->opcode
== aco_opcode::p_create_vector
&& !op
.isKillBeforeDef();
344 for (std::pair
<Operand
, Definition
>& pc
: parallelcopies
) {
345 PhysReg def_reg
= pc
.second
.physReg();
346 omit_renaming
&= def_reg
> copy
.first
.physReg() ?
347 (copy
.first
.physReg() + copy
.first
.size() <= def_reg
.reg()) :
348 (def_reg
+ pc
.second
.size() <= copy
.first
.physReg().reg());
352 op
.setTemp(copy
.second
.getTemp());
353 op
.setFixed(copy
.second
.physReg());
359 std::pair
<PhysReg
, bool> get_reg_simple(ra_ctx
& ctx
,
360 RegisterFile
& reg_file
,
363 uint32_t lb
= info
.lb
;
364 uint32_t ub
= info
.ub
;
365 uint32_t size
= info
.size
;
366 uint32_t stride
= info
.stride
;
367 RegClass rc
= info
.rc
;
369 if (rc
.is_subdword()) {
370 for (std::pair
<uint32_t, std::array
<uint32_t, 4>> entry
: reg_file
.subdword_regs
) {
371 assert(reg_file
[entry
.first
] == 0xF0000000);
372 if (lb
> entry
.first
|| entry
.first
>= ub
)
375 for (unsigned i
= 0; i
< 4; i
+= stride
) {
376 if (entry
.second
[i
] != 0)
379 bool reg_found
= true;
380 for (unsigned j
= 1; reg_found
&& i
+ j
< 4 && j
< rc
.bytes(); j
++)
381 reg_found
&= entry
.second
[i
+ j
] == 0;
383 /* check neighboring reg if needed */
384 reg_found
&= (i
<= 4 - rc
.bytes() || reg_file
[entry
.first
+ 1] == 0);
386 PhysReg res
{entry
.first
};
393 stride
= 1; /* stride in full registers */
398 if (rc
.type() == RegType::vgpr
&& (size
== 4 || size
== 8)) {
400 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
405 /* best fit algorithm: find the smallest gap to fit in the variable */
406 unsigned best_pos
= 0xFFFF;
407 unsigned gap_size
= 0xFFFF;
408 unsigned last_pos
= 0xFFFF;
410 for (unsigned current_reg
= lb
; current_reg
< ub
; current_reg
++) {
411 if (reg_file
[current_reg
] == 0 && !ctx
.war_hint
[current_reg
]) {
412 if (last_pos
== 0xFFFF)
413 last_pos
= current_reg
;
417 if (last_pos
== 0xFFFF)
420 /* early return on exact matches */
421 if (last_pos
+ size
== current_reg
) {
422 adjust_max_used_regs(ctx
, rc
, last_pos
);
423 return {PhysReg
{last_pos
}, true};
426 /* check if it fits and the gap size is smaller */
427 if (last_pos
+ size
< current_reg
&& current_reg
- last_pos
< gap_size
) {
429 gap_size
= current_reg
- last_pos
;
435 if (last_pos
+ size
<= ub
&& ub
- last_pos
< gap_size
)
438 if (best_pos
!= 0xFFFF) {
439 adjust_max_used_regs(ctx
, rc
, best_pos
);
440 return {PhysReg
{best_pos
}, true};
446 unsigned reg_lo
= lb
;
447 unsigned reg_hi
= lb
+ size
- 1;
448 while (!found
&& reg_lo
+ size
<= ub
) {
449 if (reg_file
[reg_lo
] != 0) {
453 reg_hi
= reg_lo
+ size
- 1;
455 for (unsigned reg
= reg_lo
+ 1; found
&& reg
<= reg_hi
; reg
++) {
456 if (reg_file
[reg
] != 0 || ctx
.war_hint
[reg
])
460 adjust_max_used_regs(ctx
, rc
, reg_lo
);
461 return {PhysReg
{reg_lo
}, true};
470 /* collect variables from a register area and clear reg_file */
471 std::set
<std::pair
<unsigned, unsigned>> collect_vars(ra_ctx
& ctx
, RegisterFile
& reg_file
,
472 PhysReg reg
, unsigned size
)
474 std::set
<std::pair
<unsigned, unsigned>> vars
;
475 for (unsigned j
= reg
; j
< reg
+ size
; j
++) {
476 if (reg_file
.is_blocked(PhysReg
{j
}))
478 if (reg_file
[j
] == 0xF0000000) {
479 for (unsigned k
= 0; k
< 4; k
++) {
480 unsigned id
= reg_file
.subdword_regs
[j
][k
];
482 assignment
& var
= ctx
.assignments
[id
];
483 vars
.emplace(var
.rc
.bytes(), id
);
484 reg_file
.clear(var
.reg
, var
.rc
);
489 } else if (reg_file
[j
] != 0) {
490 unsigned id
= reg_file
[j
];
491 assignment
& var
= ctx
.assignments
[id
];
492 vars
.emplace(var
.rc
.bytes(), id
);
493 reg_file
.clear(var
.reg
, var
.rc
);
499 bool get_regs_for_copies(ra_ctx
& ctx
,
500 RegisterFile
& reg_file
,
501 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
502 const std::set
<std::pair
<unsigned, unsigned>> &vars
,
503 uint32_t lb
, uint32_t ub
,
504 aco_ptr
<Instruction
>& instr
,
509 /* variables are sorted from small sized to large */
510 /* NOTE: variables are also sorted by ID. this only affects a very small number of shaders slightly though. */
511 for (std::set
<std::pair
<unsigned, unsigned>>::const_reverse_iterator it
= vars
.rbegin(); it
!= vars
.rend(); ++it
) {
512 unsigned id
= it
->second
;
513 assignment
& var
= ctx
.assignments
[id
];
514 DefInfo info
= DefInfo(ctx
, ctx
.pseudo_dummy
, var
.rc
);
515 uint32_t size
= info
.size
;
517 /* check if this is a dead operand, then we can re-use the space from the definition */
518 bool is_dead_operand
= false;
519 for (unsigned i
= 0; !is_phi(instr
) && !is_dead_operand
&& (i
< instr
->operands
.size()); i
++) {
520 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isKillBeforeDef() && instr
->operands
[i
].tempId() == id
)
521 is_dead_operand
= true;
524 std::pair
<PhysReg
, bool> res
;
525 if (is_dead_operand
) {
526 if (instr
->opcode
== aco_opcode::p_create_vector
) {
527 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
528 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].tempId() == id
) {
529 PhysReg
reg(def_reg_lo
);
531 assert(!reg_file
.test(reg
, var
.rc
.bytes()));
537 info
.lb
= def_reg_lo
;
538 info
.ub
= def_reg_hi
+ 1;
539 res
= get_reg_simple(ctx
, reg_file
, info
);
543 info
.ub
= def_reg_lo
;
544 res
= get_reg_simple(ctx
, reg_file
, info
);
546 info
.lb
= (def_reg_hi
+ info
.stride
) & ~(info
.stride
- 1);
548 res
= get_reg_simple(ctx
, reg_file
, info
);
553 /* mark the area as blocked */
554 reg_file
.block(res
.first
, var
.rc
.bytes());
556 /* create parallelcopy pair (without definition id) */
557 Temp tmp
= Temp(id
, var
.rc
);
558 Operand pc_op
= Operand(tmp
);
559 pc_op
.setFixed(var
.reg
);
560 Definition pc_def
= Definition(res
.first
, pc_op
.regClass());
561 parallelcopies
.emplace_back(pc_op
, pc_def
);
565 unsigned best_pos
= lb
;
566 unsigned num_moves
= 0xFF;
567 unsigned num_vars
= 0;
569 /* we use a sliding window to find potential positions */
570 unsigned reg_lo
= lb
;
571 unsigned reg_hi
= lb
+ size
- 1;
572 unsigned stride
= var
.rc
.is_subdword() ? 1 : info
.stride
;
573 for (reg_lo
= lb
, reg_hi
= lb
+ size
- 1; reg_hi
< ub
; reg_lo
+= stride
, reg_hi
+= stride
) {
574 if (!is_dead_operand
&& ((reg_lo
>= def_reg_lo
&& reg_lo
<= def_reg_hi
) ||
575 (reg_hi
>= def_reg_lo
&& reg_hi
<= def_reg_hi
)))
578 /* second, check that we have at most k=num_moves elements in the window
579 * and no element is larger than the currently processed one */
582 unsigned last_var
= 0;
584 for (unsigned j
= reg_lo
; found
&& j
<= reg_hi
; j
++) {
585 if (reg_file
[j
] == 0 || reg_file
[j
] == last_var
)
588 if (reg_file
.is_blocked(PhysReg
{j
}) || k
> num_moves
) {
592 if (reg_file
[j
] == 0xF0000000) {
597 /* we cannot split live ranges of linear vgprs */
598 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6)) {
602 bool is_kill
= false;
603 for (const Operand
& op
: instr
->operands
) {
604 if (op
.isTemp() && op
.isKillBeforeDef() && op
.tempId() == reg_file
[j
]) {
609 if (!is_kill
&& ctx
.assignments
[reg_file
[j
]].rc
.size() >= size
) {
614 k
+= ctx
.assignments
[reg_file
[j
]].rc
.size();
615 last_var
= reg_file
[j
];
617 if (k
> num_moves
|| (k
== num_moves
&& n
<= num_vars
)) {
630 /* FIXME: we messed up and couldn't find space for the variables to be copied */
631 if (num_moves
== 0xFF)
635 reg_hi
= best_pos
+ size
- 1;
637 /* collect variables and block reg file */
638 std::set
<std::pair
<unsigned, unsigned>> new_vars
= collect_vars(ctx
, reg_file
, PhysReg
{reg_lo
}, size
);
640 /* mark the area as blocked */
641 reg_file
.block(PhysReg
{reg_lo
}, size
* 4);
643 if (!get_regs_for_copies(ctx
, reg_file
, parallelcopies
, new_vars
, lb
, ub
, instr
, def_reg_lo
, def_reg_hi
))
646 adjust_max_used_regs(ctx
, var
.rc
, reg_lo
);
648 /* create parallelcopy pair (without definition id) */
649 Temp tmp
= Temp(id
, var
.rc
);
650 Operand pc_op
= Operand(tmp
);
651 pc_op
.setFixed(var
.reg
);
652 Definition pc_def
= Definition(PhysReg
{reg_lo
}, pc_op
.regClass());
653 parallelcopies
.emplace_back(pc_op
, pc_def
);
660 std::pair
<PhysReg
, bool> get_reg_impl(ra_ctx
& ctx
,
661 RegisterFile
& reg_file
,
662 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
664 aco_ptr
<Instruction
>& instr
)
666 uint32_t lb
= info
.lb
;
667 uint32_t ub
= info
.ub
;
668 uint32_t size
= info
.size
;
669 uint32_t stride
= info
.stride
;
670 RegClass rc
= info
.rc
;
672 /* check how many free regs we have */
673 unsigned regs_free
= reg_file
.count_zero(PhysReg
{lb
}, ub
-lb
);
675 /* mark and count killed operands */
676 unsigned killed_ops
= 0;
677 for (unsigned j
= 0; !is_phi(instr
) && j
< instr
->operands
.size(); j
++) {
678 if (instr
->operands
[j
].isTemp() &&
679 instr
->operands
[j
].isFirstKillBeforeDef() &&
680 instr
->operands
[j
].physReg() >= lb
&&
681 instr
->operands
[j
].physReg() < ub
) {
682 assert(instr
->operands
[j
].isFixed());
683 assert(!reg_file
.test(instr
->operands
[j
].physReg(), instr
->operands
[j
].bytes()));
684 reg_file
.block(instr
->operands
[j
].physReg(), instr
->operands
[j
].bytes());
685 killed_ops
+= instr
->operands
[j
].getTemp().size();
689 assert(regs_free
>= size
);
690 /* we might have to move dead operands to dst in order to make space */
691 unsigned op_moves
= 0;
693 if (size
> (regs_free
- killed_ops
))
694 op_moves
= size
- (regs_free
- killed_ops
);
696 /* find the best position to place the definition */
697 unsigned best_pos
= lb
;
698 unsigned num_moves
= 0xFF;
699 unsigned num_vars
= 0;
701 /* we use a sliding window to check potential positions */
702 unsigned reg_lo
= lb
;
703 unsigned reg_hi
= lb
+ size
- 1;
704 for (reg_lo
= lb
, reg_hi
= lb
+ size
- 1; reg_hi
< ub
; reg_lo
+= stride
, reg_hi
+= stride
) {
705 /* first check the edges: this is what we have to fix to allow for num_moves > size */
706 if (reg_lo
> lb
&& reg_file
[reg_lo
] != 0 && reg_file
[reg_lo
] == reg_file
[reg_lo
- 1])
708 if (reg_hi
< ub
- 1 && reg_file
[reg_hi
] != 0 && reg_file
[reg_hi
] == reg_file
[reg_hi
+ 1])
711 /* second, check that we have at most k=num_moves elements in the window
712 * and no element is larger than the currently processed one */
713 unsigned k
= op_moves
;
715 unsigned remaining_op_moves
= op_moves
;
716 unsigned last_var
= 0;
718 bool aligned
= rc
== RegClass::v4
&& reg_lo
% 4 == 0;
719 for (unsigned j
= reg_lo
; found
&& j
<= reg_hi
; j
++) {
720 if (reg_file
[j
] == 0 || reg_file
[j
] == last_var
)
723 /* dead operands effectively reduce the number of estimated moves */
724 if (reg_file
.is_blocked(PhysReg
{j
})) {
725 if (remaining_op_moves
) {
727 remaining_op_moves
--;
732 if (reg_file
[j
] == 0xF0000000) {
738 if (ctx
.assignments
[reg_file
[j
]].rc
.size() >= size
) {
743 /* we cannot split live ranges of linear vgprs */
744 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6)) {
749 k
+= ctx
.assignments
[reg_file
[j
]].rc
.size();
751 last_var
= reg_file
[j
];
754 if (!found
|| k
> num_moves
)
756 if (k
== num_moves
&& n
< num_vars
)
758 if (!aligned
&& k
== num_moves
&& n
== num_vars
)
768 if (num_moves
== 0xFF) {
769 /* remove killed operands from reg_file once again */
770 for (unsigned i
= 0; !is_phi(instr
) && i
< instr
->operands
.size(); i
++) {
771 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isFirstKillBeforeDef())
772 reg_file
.clear(instr
->operands
[i
]);
774 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
775 Definition def
= instr
->definitions
[i
];
776 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
782 RegisterFile register_file
= reg_file
;
784 /* now, we figured the placement for our definition */
785 std::set
<std::pair
<unsigned, unsigned>> vars
= collect_vars(ctx
, reg_file
, PhysReg
{best_pos
}, size
);
787 if (instr
->opcode
== aco_opcode::p_create_vector
) {
788 /* move killed operands which aren't yet at the correct position */
789 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].size(), i
++) {
790 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isFirstKillBeforeDef() &&
791 instr
->operands
[i
].getTemp().type() == rc
.type()) {
793 if (instr
->operands
[i
].physReg() != best_pos
+ offset
) {
794 vars
.emplace(instr
->operands
[i
].bytes(), instr
->operands
[i
].tempId());
795 reg_file
.clear(instr
->operands
[i
]);
797 reg_file
.fill(instr
->operands
[i
]);
802 /* re-enable the killed operands */
803 for (unsigned j
= 0; !is_phi(instr
) && j
< instr
->operands
.size(); j
++) {
804 if (instr
->operands
[j
].isTemp() && instr
->operands
[j
].isFirstKill())
805 reg_file
.fill(instr
->operands
[j
]);
809 std::vector
<std::pair
<Operand
, Definition
>> pc
;
810 if (!get_regs_for_copies(ctx
, reg_file
, pc
, vars
, lb
, ub
, instr
, best_pos
, best_pos
+ size
- 1)) {
811 reg_file
= std::move(register_file
);
812 /* remove killed operands from reg_file once again */
813 if (!is_phi(instr
)) {
814 for (const Operand
& op
: instr
->operands
) {
815 if (op
.isTemp() && op
.isFirstKillBeforeDef())
819 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
820 Definition
& def
= instr
->definitions
[i
];
821 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
827 parallelcopies
.insert(parallelcopies
.end(), pc
.begin(), pc
.end());
829 /* we set the definition regs == 0. the actual caller is responsible for correct setting */
830 reg_file
.clear(PhysReg
{best_pos
}, rc
);
832 update_renames(ctx
, reg_file
, parallelcopies
, instr
);
834 /* remove killed operands from reg_file once again */
835 for (unsigned i
= 0; !is_phi(instr
) && i
< instr
->operands
.size(); i
++) {
836 if (!instr
->operands
[i
].isTemp() || !instr
->operands
[i
].isFixed())
838 assert(!instr
->operands
[i
].isUndefined());
839 if (instr
->operands
[i
].isFirstKillBeforeDef())
840 reg_file
.clear(instr
->operands
[i
]);
842 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
843 Definition def
= instr
->definitions
[i
];
844 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
848 adjust_max_used_regs(ctx
, rc
, best_pos
);
849 return {PhysReg
{best_pos
}, true};
852 bool get_reg_specified(ra_ctx
& ctx
,
853 RegisterFile
& reg_file
,
855 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
856 aco_ptr
<Instruction
>& instr
,
859 if (rc
.is_subdword() && reg
.byte() && !instr_can_access_subdword(instr
))
862 uint32_t size
= rc
.size();
866 if (rc
.type() == RegType::vgpr
) {
868 ub
= 256 + ctx
.program
->max_reg_demand
.vgpr
;
874 if (reg
% stride
!= 0)
877 ub
= ctx
.program
->max_reg_demand
.sgpr
;
880 uint32_t reg_lo
= reg
.reg();
881 uint32_t reg_hi
= reg
+ (size
- 1);
883 if (reg_lo
< lb
|| reg_hi
>= ub
|| reg_lo
> reg_hi
)
886 if (reg_file
.test(reg
, rc
.bytes()))
889 adjust_max_used_regs(ctx
, rc
, reg_lo
);
893 PhysReg
get_reg(ra_ctx
& ctx
,
894 RegisterFile
& reg_file
,
896 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
897 aco_ptr
<Instruction
>& instr
)
899 if (ctx
.affinities
.find(temp
.id()) != ctx
.affinities
.end() &&
900 ctx
.assignments
[ctx
.affinities
[temp
.id()]].assigned
) {
901 PhysReg reg
= ctx
.assignments
[ctx
.affinities
[temp
.id()]].reg
;
902 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
906 if (ctx
.vectors
.find(temp
.id()) != ctx
.vectors
.end()) {
907 Instruction
* vec
= ctx
.vectors
[temp
.id()];
908 unsigned byte_offset
= 0;
909 for (const Operand
& op
: vec
->operands
) {
910 if (op
.isTemp() && op
.tempId() == temp
.id())
913 byte_offset
+= op
.bytes();
916 for (const Operand
& op
: vec
->operands
) {
918 op
.tempId() != temp
.id() &&
919 op
.getTemp().type() == temp
.type() &&
920 ctx
.assignments
[op
.tempId()].assigned
) {
921 PhysReg reg
= ctx
.assignments
[op
.tempId()].reg
;
922 reg
.reg_b
+= (byte_offset
- k
);
923 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
929 DefInfo
info(ctx
, ctx
.pseudo_dummy
, vec
->definitions
[0].regClass());
930 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
931 PhysReg reg
= res
.first
;
933 reg
.reg_b
+= byte_offset
;
934 /* make sure to only use byte offset if the instruction supports it */
935 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
940 DefInfo
info(ctx
, instr
, temp
.regClass());
942 /* try to find space without live-range splits */
943 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
948 /* try to find space with live-range splits */
949 res
= get_reg_impl(ctx
, reg_file
, parallelcopies
, info
, instr
);
954 /* try using more registers */
956 /* We should only fail here because keeping under the limit would require
958 assert(reg_file
.count_zero(PhysReg
{info
.lb
}, info
.ub
-info
.lb
) >= info
.size
);
960 uint16_t max_addressible_sgpr
= ctx
.program
->sgpr_limit
;
961 uint16_t max_addressible_vgpr
= ctx
.program
->vgpr_limit
;
962 if (info
.rc
.type() == RegType::vgpr
&& ctx
.program
->max_reg_demand
.vgpr
< max_addressible_vgpr
) {
963 update_vgpr_sgpr_demand(ctx
.program
, RegisterDemand(ctx
.program
->max_reg_demand
.vgpr
+ 1, ctx
.program
->max_reg_demand
.sgpr
));
964 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
965 } else if (info
.rc
.type() == RegType::sgpr
&& ctx
.program
->max_reg_demand
.sgpr
< max_addressible_sgpr
) {
966 update_vgpr_sgpr_demand(ctx
.program
, RegisterDemand(ctx
.program
->max_reg_demand
.vgpr
, ctx
.program
->max_reg_demand
.sgpr
+ 1));
967 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
970 //FIXME: if nothing helps, shift-rotate the registers to make space
972 unreachable("did not find a register");
975 PhysReg
get_reg_create_vector(ra_ctx
& ctx
,
976 RegisterFile
& reg_file
,
978 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
979 aco_ptr
<Instruction
>& instr
)
981 RegClass rc
= temp
.regClass();
982 /* create_vector instructions have different costs w.r.t. register coalescing */
983 uint32_t size
= rc
.size();
984 uint32_t bytes
= rc
.bytes();
987 if (rc
.type() == RegType::vgpr
) {
989 ub
= 256 + ctx
.program
->max_reg_demand
.vgpr
;
992 ub
= ctx
.program
->max_reg_demand
.sgpr
;
999 //TODO: improve p_create_vector for sub-dword vectors
1001 unsigned best_pos
= -1;
1002 unsigned num_moves
= 0xFF;
1003 bool best_war_hint
= true;
1005 /* test for each operand which definition placement causes the least shuffle instructions */
1006 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
1007 // TODO: think about, if we can alias live operands on the same register
1008 if (!instr
->operands
[i
].isTemp() || !instr
->operands
[i
].isKillBeforeDef() || instr
->operands
[i
].getTemp().type() != rc
.type())
1011 if (offset
> instr
->operands
[i
].physReg().reg_b
)
1014 unsigned reg_lo
= instr
->operands
[i
].physReg().reg_b
- offset
;
1018 unsigned reg_hi
= reg_lo
+ size
- 1;
1021 /* no need to check multiple times */
1022 if (reg_lo
== best_pos
)
1026 // TODO: this can be improved */
1027 if (reg_lo
< lb
|| reg_hi
>= ub
|| reg_lo
% stride
!= 0)
1029 if (reg_lo
> lb
&& reg_file
[reg_lo
] != 0 && reg_file
[reg_lo
] == reg_file
[reg_lo
- 1])
1031 if (reg_hi
< ub
- 1 && reg_file
[reg_hi
] != 0 && reg_file
[reg_hi
] == reg_file
[reg_hi
+ 1])
1034 /* count variables to be moved and check war_hint */
1035 bool war_hint
= false;
1036 bool linear_vgpr
= false;
1037 for (unsigned j
= reg_lo
; j
<= reg_hi
&& !linear_vgpr
; j
++) {
1038 if (reg_file
[j
] != 0) {
1039 if (reg_file
[j
] == 0xF0000000) {
1042 unsigned bytes_left
= bytes
- (j
- reg_lo
) * 4;
1043 for (unsigned k
= 0; k
< MIN2(bytes_left
, 4); k
++, reg
.reg_b
++)
1044 k
+= reg_file
.test(reg
, 1);
1047 /* we cannot split live ranges of linear vgprs */
1048 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6))
1052 war_hint
|= ctx
.war_hint
[j
];
1054 if (linear_vgpr
|| (war_hint
&& !best_war_hint
))
1057 /* count operands in wrong positions */
1058 for (unsigned j
= 0, offset
= 0; j
< instr
->operands
.size(); offset
+= instr
->operands
[j
].bytes(), j
++) {
1060 !instr
->operands
[j
].isTemp() ||
1061 instr
->operands
[j
].getTemp().type() != rc
.type())
1063 if (instr
->operands
[j
].physReg().reg_b
!= reg_lo
* 4 + offset
)
1064 k
+= instr
->operands
[j
].bytes();
1066 bool aligned
= rc
== RegClass::v4
&& reg_lo
% 4 == 0;
1067 if (k
> num_moves
|| (!aligned
&& k
== num_moves
))
1072 best_war_hint
= war_hint
;
1075 if (num_moves
>= bytes
)
1076 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
1078 /* collect variables to be moved */
1079 std::set
<std::pair
<unsigned, unsigned>> vars
= collect_vars(ctx
, reg_file
, PhysReg
{best_pos
}, size
);
1081 /* move killed operands which aren't yet at the correct position */
1082 uint64_t moved_operand_mask
= 0;
1083 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
1084 if (instr
->operands
[i
].isTemp() &&
1085 instr
->operands
[i
].isFirstKillBeforeDef() &&
1086 instr
->operands
[i
].getTemp().type() == rc
.type() &&
1087 instr
->operands
[i
].physReg().reg_b
!= best_pos
* 4 + offset
) {
1088 vars
.emplace(instr
->operands
[i
].bytes(), instr
->operands
[i
].tempId());
1089 moved_operand_mask
|= (uint64_t)1 << i
;
1093 ASSERTED
bool success
= false;
1094 success
= get_regs_for_copies(ctx
, reg_file
, parallelcopies
, vars
, lb
, ub
, instr
, best_pos
, best_pos
+ size
- 1);
1097 update_renames(ctx
, reg_file
, parallelcopies
, instr
);
1098 adjust_max_used_regs(ctx
, rc
, best_pos
);
1100 while (moved_operand_mask
) {
1101 unsigned i
= u_bit_scan64(&moved_operand_mask
);
1102 assert(instr
->operands
[i
].isFirstKillBeforeDef());
1103 reg_file
.clear(instr
->operands
[i
]);
1106 return PhysReg
{best_pos
};
1109 void handle_pseudo(ra_ctx
& ctx
,
1110 const RegisterFile
& reg_file
,
1113 if (instr
->format
!= Format::PSEUDO
)
1116 /* all instructions which use handle_operands() need this information */
1117 switch (instr
->opcode
) {
1118 case aco_opcode::p_extract_vector
:
1119 case aco_opcode::p_create_vector
:
1120 case aco_opcode::p_split_vector
:
1121 case aco_opcode::p_parallelcopy
:
1122 case aco_opcode::p_wqm
:
1128 /* if all definitions are vgpr, no need to care for SCC */
1129 bool writes_sgpr
= false;
1130 for (Definition
& def
: instr
->definitions
) {
1131 if (def
.getTemp().type() == RegType::sgpr
) {
1136 /* if all operands are constant, no need to care either */
1137 bool reads_sgpr
= false;
1138 for (Operand
& op
: instr
->operands
) {
1139 if (op
.isTemp() && op
.getTemp().type() == RegType::sgpr
) {
1144 if (!(writes_sgpr
&& reads_sgpr
))
1147 Pseudo_instruction
*pi
= (Pseudo_instruction
*)instr
;
1148 if (reg_file
[scc
.reg()]) {
1149 pi
->tmp_in_scc
= true;
1151 int reg
= ctx
.max_used_sgpr
;
1152 for (; reg
>= 0 && reg_file
[reg
]; reg
--)
1155 reg
= ctx
.max_used_sgpr
+ 1;
1156 for (; reg
< ctx
.program
->max_reg_demand
.sgpr
&& reg_file
[reg
]; reg
++)
1158 assert(reg
< ctx
.program
->max_reg_demand
.sgpr
);
1161 adjust_max_used_regs(ctx
, s1
, reg
);
1162 pi
->scratch_sgpr
= PhysReg
{(unsigned)reg
};
1164 pi
->tmp_in_scc
= false;
1168 bool operand_can_use_reg(aco_ptr
<Instruction
>& instr
, unsigned idx
, PhysReg reg
)
1170 if (instr
->operands
[idx
].isFixed())
1171 return instr
->operands
[idx
].physReg() == reg
;
1173 if (!instr_can_access_subdword(instr
) && reg
.byte())
1176 switch (instr
->format
) {
1178 return reg
!= scc
&&
1180 (reg
!= m0
|| idx
== 1 || idx
== 3) && /* offset can be m0 */
1181 (reg
!= vcc
|| (instr
->definitions
.empty() && idx
== 2)); /* sdata can be vcc */
1183 // TODO: there are more instructions with restrictions on registers
1188 void get_reg_for_operand(ra_ctx
& ctx
, RegisterFile
& register_file
,
1189 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopy
,
1190 aco_ptr
<Instruction
>& instr
, Operand
& operand
)
1192 /* check if the operand is fixed */
1194 if (operand
.isFixed()) {
1195 assert(operand
.physReg() != ctx
.assignments
[operand
.tempId()].reg
);
1197 /* check if target reg is blocked, and move away the blocking var */
1198 if (register_file
[operand
.physReg().reg()]) {
1199 assert(register_file
[operand
.physReg()] != 0xF0000000);
1200 uint32_t blocking_id
= register_file
[operand
.physReg().reg()];
1201 RegClass rc
= ctx
.assignments
[blocking_id
].rc
;
1202 Operand pc_op
= Operand(Temp
{blocking_id
, rc
});
1203 pc_op
.setFixed(operand
.physReg());
1206 PhysReg reg
= get_reg(ctx
, register_file
, pc_op
.getTemp(), parallelcopy
, ctx
.pseudo_dummy
);
1207 Definition pc_def
= Definition(PhysReg
{reg
}, pc_op
.regClass());
1208 register_file
.clear(pc_op
);
1209 parallelcopy
.emplace_back(pc_op
, pc_def
);
1211 dst
= operand
.physReg();
1214 dst
= get_reg(ctx
, register_file
, operand
.getTemp(), parallelcopy
, instr
);
1217 Operand pc_op
= operand
;
1218 pc_op
.setFixed(ctx
.assignments
[operand
.tempId()].reg
);
1219 Definition pc_def
= Definition(dst
, pc_op
.regClass());
1220 register_file
.clear(pc_op
);
1221 parallelcopy
.emplace_back(pc_op
, pc_def
);
1222 update_renames(ctx
, register_file
, parallelcopy
, instr
);
1225 Temp
read_variable(ra_ctx
& ctx
, Temp val
, unsigned block_idx
)
1227 std::unordered_map
<unsigned, Temp
>::iterator it
= ctx
.renames
[block_idx
].find(val
.id());
1228 if (it
== ctx
.renames
[block_idx
].end())
1234 Temp
handle_live_in(ra_ctx
& ctx
, Temp val
, Block
* block
)
1236 std::vector
<unsigned>& preds
= val
.is_linear() ? block
->linear_preds
: block
->logical_preds
;
1237 if (preds
.size() == 0 || val
.regClass() == val
.regClass().as_linear())
1240 assert(preds
.size() > 0);
1243 if (!ctx
.sealed
[block
->index
]) {
1244 /* consider rename from already processed predecessor */
1245 Temp tmp
= read_variable(ctx
, val
, preds
[0]);
1247 /* if the block is not sealed yet, we create an incomplete phi (which might later get removed again) */
1248 new_val
= Temp
{ctx
.program
->allocateId(), val
.regClass()};
1249 ctx
.assignments
.emplace_back();
1250 aco_opcode opcode
= val
.is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1251 aco_ptr
<Instruction
> phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1252 phi
->definitions
[0] = Definition(new_val
);
1253 for (unsigned i
= 0; i
< preds
.size(); i
++)
1254 phi
->operands
[i
] = Operand(val
);
1255 if (tmp
.regClass() == new_val
.regClass())
1256 ctx
.affinities
[new_val
.id()] = tmp
.id();
1258 ctx
.phi_map
.emplace(new_val
.id(), phi_info
{phi
.get(), block
->index
});
1259 ctx
.incomplete_phis
[block
->index
].emplace_back(phi
.get());
1260 block
->instructions
.insert(block
->instructions
.begin(), std::move(phi
));
1262 } else if (preds
.size() == 1) {
1263 /* if the block has only one predecessor, just look there for the name */
1264 new_val
= read_variable(ctx
, val
, preds
[0]);
1266 /* there are multiple predecessors and the block is sealed */
1267 Temp ops
[preds
.size()];
1269 /* get the rename from each predecessor and check if they are the same */
1270 bool needs_phi
= false;
1271 for (unsigned i
= 0; i
< preds
.size(); i
++) {
1272 ops
[i
] = read_variable(ctx
, val
, preds
[i
]);
1276 needs_phi
|= !(new_val
== ops
[i
]);
1280 /* the variable has been renamed differently in the predecessors: we need to insert a phi */
1281 aco_opcode opcode
= val
.is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1282 aco_ptr
<Instruction
> phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1283 new_val
= Temp
{ctx
.program
->allocateId(), val
.regClass()};
1284 phi
->definitions
[0] = Definition(new_val
);
1285 for (unsigned i
= 0; i
< preds
.size(); i
++) {
1286 phi
->operands
[i
] = Operand(ops
[i
]);
1287 phi
->operands
[i
].setFixed(ctx
.assignments
[ops
[i
].id()].reg
);
1288 if (ops
[i
].regClass() == new_val
.regClass())
1289 ctx
.affinities
[new_val
.id()] = ops
[i
].id();
1291 ctx
.assignments
.emplace_back();
1292 assert(ctx
.assignments
.size() == ctx
.program
->peekAllocationId());
1293 ctx
.phi_map
.emplace(new_val
.id(), phi_info
{phi
.get(), block
->index
});
1294 block
->instructions
.insert(block
->instructions
.begin(), std::move(phi
));
1298 if (new_val
!= val
) {
1299 ctx
.renames
[block
->index
][val
.id()] = new_val
;
1300 ctx
.orig_names
[new_val
.id()] = val
;
1305 void try_remove_trivial_phi(ra_ctx
& ctx
, Temp temp
)
1307 std::unordered_map
<unsigned, phi_info
>::iterator info
= ctx
.phi_map
.find(temp
.id());
1309 if (info
== ctx
.phi_map
.end() || !ctx
.sealed
[info
->second
.block_idx
])
1312 assert(info
->second
.block_idx
!= 0);
1313 Instruction
* phi
= info
->second
.phi
;
1315 Definition def
= phi
->definitions
[0];
1317 /* a phi node is trivial if all operands are the same as the definition of the phi */
1318 for (const Operand
& op
: phi
->operands
) {
1319 const Temp t
= op
.getTemp();
1320 if (t
== same
|| t
== def
.getTemp()) {
1321 assert(t
== same
|| op
.physReg() == def
.physReg());
1329 assert(same
!= Temp() || same
== def
.getTemp());
1331 /* reroute all uses to same and remove phi */
1332 std::vector
<Temp
> phi_users
;
1333 std::unordered_map
<unsigned, phi_info
>::iterator same_phi_info
= ctx
.phi_map
.find(same
.id());
1334 for (Instruction
* instr
: info
->second
.uses
) {
1335 assert(phi
!= instr
);
1336 /* recursively try to remove trivial phis */
1337 if (is_phi(instr
)) {
1338 /* ignore if the phi was already flagged trivial */
1339 if (instr
->definitions
.empty())
1342 if (instr
->definitions
[0].getTemp() != temp
)
1343 phi_users
.emplace_back(instr
->definitions
[0].getTemp());
1345 for (Operand
& op
: instr
->operands
) {
1346 if (op
.isTemp() && op
.tempId() == def
.tempId()) {
1348 if (same_phi_info
!= ctx
.phi_map
.end())
1349 same_phi_info
->second
.uses
.emplace(instr
);
1354 auto it
= ctx
.orig_names
.find(same
.id());
1355 unsigned orig_var
= it
!= ctx
.orig_names
.end() ? it
->second
.id() : same
.id();
1356 for (unsigned i
= 0; i
< ctx
.program
->blocks
.size(); i
++) {
1357 auto it
= ctx
.renames
[i
].find(orig_var
);
1358 if (it
!= ctx
.renames
[i
].end() && it
->second
== def
.getTemp())
1359 ctx
.renames
[i
][orig_var
] = same
;
1362 phi
->definitions
.clear(); /* this indicates that the phi can be removed */
1363 ctx
.phi_map
.erase(info
);
1364 for (Temp t
: phi_users
)
1365 try_remove_trivial_phi(ctx
, t
);
1370 } /* end namespace */
1373 void register_allocation(Program
*program
, std::vector
<TempSet
>& live_out_per_block
)
1375 ra_ctx
ctx(program
);
1376 std::vector
<std::vector
<Temp
>> phi_ressources
;
1377 std::unordered_map
<unsigned, unsigned> temp_to_phi_ressources
;
1379 for (std::vector
<Block
>::reverse_iterator it
= program
->blocks
.rbegin(); it
!= program
->blocks
.rend(); it
++) {
1382 /* first, compute the death points of all live vars within the block */
1383 TempSet
& live
= live_out_per_block
[block
.index
];
1385 std::vector
<aco_ptr
<Instruction
>>::reverse_iterator rit
;
1386 for (rit
= block
.instructions
.rbegin(); rit
!= block
.instructions
.rend(); ++rit
) {
1387 aco_ptr
<Instruction
>& instr
= *rit
;
1388 if (is_phi(instr
)) {
1389 live
.erase(instr
->definitions
[0].getTemp());
1390 if (instr
->definitions
[0].isKill() || instr
->definitions
[0].isFixed())
1392 /* collect information about affinity-related temporaries */
1393 std::vector
<Temp
> affinity_related
;
1394 /* affinity_related[0] is the last seen affinity-related temp */
1395 affinity_related
.emplace_back(instr
->definitions
[0].getTemp());
1396 affinity_related
.emplace_back(instr
->definitions
[0].getTemp());
1397 for (const Operand
& op
: instr
->operands
) {
1398 if (op
.isTemp() && op
.regClass() == instr
->definitions
[0].regClass()) {
1399 affinity_related
.emplace_back(op
.getTemp());
1400 temp_to_phi_ressources
[op
.tempId()] = phi_ressources
.size();
1403 phi_ressources
.emplace_back(std::move(affinity_related
));
1407 /* add vector affinities */
1408 if (instr
->opcode
== aco_opcode::p_create_vector
) {
1409 for (const Operand
& op
: instr
->operands
) {
1410 if (op
.isTemp() && op
.isFirstKill() && op
.getTemp().type() == instr
->definitions
[0].getTemp().type())
1411 ctx
.vectors
[op
.tempId()] = instr
.get();
1415 /* add operands to live variables */
1416 for (const Operand
& op
: instr
->operands
) {
1418 live
.emplace(op
.getTemp());
1421 /* erase definitions from live */
1422 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
1423 const Definition
& def
= instr
->definitions
[i
];
1426 live
.erase(def
.getTemp());
1427 /* mark last-seen phi operand */
1428 std::unordered_map
<unsigned, unsigned>::iterator it
= temp_to_phi_ressources
.find(def
.tempId());
1429 if (it
!= temp_to_phi_ressources
.end() && def
.regClass() == phi_ressources
[it
->second
][0].regClass()) {
1430 phi_ressources
[it
->second
][0] = def
.getTemp();
1431 /* try to coalesce phi affinities with parallelcopies */
1432 if (!def
.isFixed() && instr
->opcode
== aco_opcode::p_parallelcopy
) {
1433 Operand op
= instr
->operands
[i
];
1434 if (op
.isTemp() && op
.isFirstKillBeforeDef() && def
.regClass() == op
.regClass()) {
1435 phi_ressources
[it
->second
].emplace_back(op
.getTemp());
1436 temp_to_phi_ressources
[op
.tempId()] = it
->second
;
1443 /* create affinities */
1444 for (std::vector
<Temp
>& vec
: phi_ressources
) {
1445 assert(vec
.size() > 1);
1446 for (unsigned i
= 1; i
< vec
.size(); i
++)
1447 if (vec
[i
].id() != vec
[0].id())
1448 ctx
.affinities
[vec
[i
].id()] = vec
[0].id();
1451 /* state of register file after phis */
1452 std::vector
<std::bitset
<128>> sgpr_live_in(program
->blocks
.size());
1454 for (Block
& block
: program
->blocks
) {
1455 TempSet
& live
= live_out_per_block
[block
.index
];
1456 /* initialize register file */
1457 assert(block
.index
!= 0 || live
.empty());
1458 RegisterFile register_file
;
1459 ctx
.war_hint
.reset();
1461 for (Temp t
: live
) {
1462 Temp renamed
= handle_live_in(ctx
, t
, &block
);
1463 assignment
& var
= ctx
.assignments
[renamed
.id()];
1464 /* due to live-range splits, the live-in might be a phi, now */
1466 register_file
.fill(Definition(renamed
.id(), var
.reg
, var
.rc
));
1469 std::vector
<aco_ptr
<Instruction
>> instructions
;
1470 std::vector
<aco_ptr
<Instruction
>>::iterator it
;
1472 /* this is a slight adjustment from the paper as we already have phi nodes:
1473 * We consider them incomplete phis and only handle the definition. */
1475 /* handle fixed phi definitions */
1476 for (it
= block
.instructions
.begin(); it
!= block
.instructions
.end(); ++it
) {
1477 aco_ptr
<Instruction
>& phi
= *it
;
1480 Definition
& definition
= phi
->definitions
[0];
1481 if (!definition
.isFixed())
1484 /* check if a dead exec mask phi is needed */
1485 if (definition
.isKill()) {
1486 for (Operand
& op
: phi
->operands
) {
1487 assert(op
.isTemp());
1488 if (!ctx
.assignments
[op
.tempId()].assigned
||
1489 ctx
.assignments
[op
.tempId()].reg
!= exec
) {
1490 definition
.setKill(false);
1496 if (definition
.isKill())
1499 assert(definition
.physReg() == exec
);
1500 assert(!register_file
.test(definition
.physReg(), definition
.bytes()));
1501 register_file
.fill(definition
);
1502 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1505 /* look up the affinities */
1506 for (it
= block
.instructions
.begin(); it
!= block
.instructions
.end(); ++it
) {
1507 aco_ptr
<Instruction
>& phi
= *it
;
1510 Definition
& definition
= phi
->definitions
[0];
1511 if (definition
.isKill() || definition
.isFixed())
1514 if (ctx
.affinities
.find(definition
.tempId()) != ctx
.affinities
.end() &&
1515 ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].assigned
) {
1516 assert(ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].rc
== definition
.regClass());
1517 PhysReg reg
= ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].reg
;
1518 bool try_use_special_reg
= reg
== scc
|| reg
== exec
;
1519 if (try_use_special_reg
) {
1520 for (const Operand
& op
: phi
->operands
) {
1521 if (!(op
.isTemp() && ctx
.assignments
[op
.tempId()].assigned
&&
1522 ctx
.assignments
[op
.tempId()].reg
== reg
)) {
1523 try_use_special_reg
= false;
1527 if (!try_use_special_reg
)
1530 /* only assign if register is still free */
1531 if (!register_file
.test(reg
, definition
.bytes())) {
1532 definition
.setFixed(reg
);
1533 register_file
.fill(definition
);
1534 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1539 /* find registers for phis without affinity or where the register was blocked */
1540 for (it
= block
.instructions
.begin();it
!= block
.instructions
.end(); ++it
) {
1541 aco_ptr
<Instruction
>& phi
= *it
;
1545 Definition
& definition
= phi
->definitions
[0];
1546 if (definition
.isKill())
1549 if (!definition
.isFixed()) {
1550 std::vector
<std::pair
<Operand
, Definition
>> parallelcopy
;
1551 /* try to find a register that is used by at least one operand */
1552 for (const Operand
& op
: phi
->operands
) {
1553 if (!(op
.isTemp() && ctx
.assignments
[op
.tempId()].assigned
))
1555 PhysReg reg
= ctx
.assignments
[op
.tempId()].reg
;
1556 /* we tried this already on the previous loop */
1557 if (reg
== scc
|| reg
== exec
)
1559 if (get_reg_specified(ctx
, register_file
, definition
.regClass(), parallelcopy
, phi
, reg
)) {
1560 definition
.setFixed(reg
);
1564 if (!definition
.isFixed())
1565 definition
.setFixed(get_reg(ctx
, register_file
, definition
.getTemp(), parallelcopy
, phi
));
1567 /* process parallelcopy */
1568 for (std::pair
<Operand
, Definition
> pc
: parallelcopy
) {
1569 /* see if it's a copy from a different phi */
1570 //TODO: prefer moving some previous phis over live-ins
1571 //TODO: somehow prevent phis fixed before the RA from being updated (shouldn't be a problem in practice since they can only be fixed to exec)
1572 Instruction
*prev_phi
= NULL
;
1573 std::vector
<aco_ptr
<Instruction
>>::iterator phi_it
;
1574 for (phi_it
= instructions
.begin(); phi_it
!= instructions
.end(); ++phi_it
) {
1575 if ((*phi_it
)->definitions
[0].tempId() == pc
.first
.tempId())
1576 prev_phi
= phi_it
->get();
1579 while (!prev_phi
&& is_phi(*++phi_it
)) {
1580 if ((*phi_it
)->definitions
[0].tempId() == pc
.first
.tempId())
1581 prev_phi
= phi_it
->get();
1584 /* if so, just update that phi's register */
1585 register_file
.clear(prev_phi
->definitions
[0]);
1586 prev_phi
->definitions
[0].setFixed(pc
.second
.physReg());
1587 ctx
.assignments
[prev_phi
->definitions
[0].tempId()] = {pc
.second
.physReg(), pc
.second
.regClass()};
1588 register_file
.fill(prev_phi
->definitions
[0]);
1593 std::unordered_map
<unsigned, Temp
>::iterator orig_it
= ctx
.orig_names
.find(pc
.first
.tempId());
1594 Temp orig
= pc
.first
.getTemp();
1595 if (orig_it
!= ctx
.orig_names
.end())
1596 orig
= orig_it
->second
;
1598 ctx
.orig_names
[pc
.second
.tempId()] = orig
;
1599 ctx
.renames
[block
.index
][orig
.id()] = pc
.second
.getTemp();
1601 /* otherwise, this is a live-in and we need to create a new phi
1602 * to move it in this block's predecessors */
1603 aco_opcode opcode
= pc
.first
.getTemp().is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1604 std::vector
<unsigned>& preds
= pc
.first
.getTemp().is_linear() ? block
.linear_preds
: block
.logical_preds
;
1605 aco_ptr
<Instruction
> new_phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1606 new_phi
->definitions
[0] = pc
.second
;
1607 for (unsigned i
= 0; i
< preds
.size(); i
++)
1608 new_phi
->operands
[i
] = Operand(pc
.first
);
1609 instructions
.emplace_back(std::move(new_phi
));
1612 register_file
.fill(definition
);
1613 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1615 live
.emplace(definition
.getTemp());
1617 /* update phi affinities */
1618 for (const Operand
& op
: phi
->operands
) {
1619 if (op
.isTemp() && op
.regClass() == phi
->definitions
[0].regClass())
1620 ctx
.affinities
[op
.tempId()] = definition
.tempId();
1623 instructions
.emplace_back(std::move(*it
));
1626 /* fill in sgpr_live_in */
1627 for (unsigned i
= 0; i
<= ctx
.max_used_sgpr
; i
++)
1628 sgpr_live_in
[block
.index
][i
] = register_file
[i
];
1629 sgpr_live_in
[block
.index
][127] = register_file
[scc
.reg()];
1631 /* Handle all other instructions of the block */
1632 for (; it
!= block
.instructions
.end(); ++it
) {
1633 aco_ptr
<Instruction
>& instr
= *it
;
1635 /* parallelcopies from p_phi are inserted here which means
1636 * live ranges of killed operands end here as well */
1637 if (instr
->opcode
== aco_opcode::p_logical_end
) {
1638 /* no need to process this instruction any further */
1639 if (block
.logical_succs
.size() != 1) {
1640 instructions
.emplace_back(std::move(instr
));
1644 Block
& succ
= program
->blocks
[block
.logical_succs
[0]];
1646 for (; idx
< succ
.logical_preds
.size(); idx
++) {
1647 if (succ
.logical_preds
[idx
] == block
.index
)
1650 for (aco_ptr
<Instruction
>& phi
: succ
.instructions
) {
1651 if (phi
->opcode
== aco_opcode::p_phi
) {
1652 if (phi
->operands
[idx
].isTemp() &&
1653 phi
->operands
[idx
].getTemp().type() == RegType::sgpr
&&
1654 phi
->operands
[idx
].isFirstKillBeforeDef()) {
1655 Temp phi_op
= read_variable(ctx
, phi
->operands
[idx
].getTemp(), block
.index
);
1656 PhysReg reg
= ctx
.assignments
[phi_op
.id()].reg
;
1657 assert(register_file
[reg
] == phi_op
.id());
1658 register_file
[reg
] = 0;
1660 } else if (phi
->opcode
!= aco_opcode::p_linear_phi
) {
1664 instructions
.emplace_back(std::move(instr
));
1668 std::vector
<std::pair
<Operand
, Definition
>> parallelcopy
;
1670 assert(!is_phi(instr
));
1672 /* handle operands */
1673 for (unsigned i
= 0; i
< instr
->operands
.size(); ++i
) {
1674 auto& operand
= instr
->operands
[i
];
1675 if (!operand
.isTemp())
1678 /* rename operands */
1679 operand
.setTemp(read_variable(ctx
, operand
.getTemp(), block
.index
));
1680 assert(ctx
.assignments
[operand
.tempId()].assigned
);
1682 PhysReg reg
= ctx
.assignments
[operand
.tempId()].reg
;
1683 if (operand_can_use_reg(instr
, i
, reg
))
1684 operand
.setFixed(reg
);
1686 get_reg_for_operand(ctx
, register_file
, parallelcopy
, instr
, operand
);
1688 if (instr
->format
== Format::EXP
||
1689 (instr
->isVMEM() && i
== 3 && ctx
.program
->chip_class
== GFX6
) ||
1690 (instr
->format
== Format::DS
&& static_cast<DS_instruction
*>(instr
.get())->gds
)) {
1691 for (unsigned j
= 0; j
< operand
.size(); j
++)
1692 ctx
.war_hint
.set(operand
.physReg().reg() + j
);
1695 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.getTemp().id());
1696 if (phi
!= ctx
.phi_map
.end())
1697 phi
->second
.uses
.emplace(instr
.get());
1700 /* remove dead vars from register file */
1701 for (const Operand
& op
: instr
->operands
) {
1702 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1703 register_file
.clear(op
);
1706 /* try to optimize v_mad_f32 -> v_mac_f32 */
1707 if (instr
->opcode
== aco_opcode::v_mad_f32
&&
1708 instr
->operands
[2].isTemp() &&
1709 instr
->operands
[2].isKillBeforeDef() &&
1710 instr
->operands
[2].getTemp().type() == RegType::vgpr
&&
1711 instr
->operands
[1].isTemp() &&
1712 instr
->operands
[1].getTemp().type() == RegType::vgpr
) { /* TODO: swap src0 and src1 in this case */
1713 VOP3A_instruction
* vop3
= static_cast<VOP3A_instruction
*>(instr
.get());
1714 bool can_use_mac
= !(vop3
->abs
[0] || vop3
->abs
[1] || vop3
->abs
[2] ||
1715 vop3
->neg
[0] || vop3
->neg
[1] || vop3
->neg
[2] ||
1716 vop3
->clamp
|| vop3
->omod
|| vop3
->opsel
);
1718 instr
->format
= Format::VOP2
;
1719 instr
->opcode
= aco_opcode::v_mac_f32
;
1723 /* handle definitions which must have the same register as an operand */
1724 if (instr
->opcode
== aco_opcode::v_interp_p2_f32
||
1725 instr
->opcode
== aco_opcode::v_mac_f32
||
1726 instr
->opcode
== aco_opcode::v_writelane_b32
||
1727 instr
->opcode
== aco_opcode::v_writelane_b32_e64
) {
1728 instr
->definitions
[0].setFixed(instr
->operands
[2].physReg());
1729 } else if (instr
->opcode
== aco_opcode::s_addk_i32
||
1730 instr
->opcode
== aco_opcode::s_mulk_i32
) {
1731 instr
->definitions
[0].setFixed(instr
->operands
[0].physReg());
1732 } else if (instr
->format
== Format::MUBUF
&&
1733 instr
->definitions
.size() == 1 &&
1734 instr
->operands
.size() == 4) {
1735 instr
->definitions
[0].setFixed(instr
->operands
[3].physReg());
1736 } else if (instr
->format
== Format::MIMG
&&
1737 instr
->definitions
.size() == 1 &&
1738 instr
->operands
[1].regClass().type() == RegType::vgpr
) {
1739 instr
->definitions
[0].setFixed(instr
->operands
[1].physReg());
1742 ctx
.defs_done
.reset();
1744 /* handle fixed definitions first */
1745 for (unsigned i
= 0; i
< instr
->definitions
.size(); ++i
) {
1746 auto& definition
= instr
->definitions
[i
];
1747 if (!definition
.isFixed())
1750 adjust_max_used_regs(ctx
, definition
.regClass(), definition
.physReg());
1751 /* check if the target register is blocked */
1752 if (register_file
[definition
.physReg().reg()] != 0) {
1753 /* create parallelcopy pair to move blocking var */
1754 Temp tmp
= {register_file
[definition
.physReg()], ctx
.assignments
[register_file
[definition
.physReg()]].rc
};
1755 Operand pc_op
= Operand(tmp
);
1756 pc_op
.setFixed(ctx
.assignments
[register_file
[definition
.physReg().reg()]].reg
);
1757 RegClass rc
= pc_op
.regClass();
1758 tmp
= Temp
{program
->allocateId(), rc
};
1759 Definition pc_def
= Definition(tmp
);
1761 /* re-enable the killed operands, so that we don't move the blocking var there */
1762 for (const Operand
& op
: instr
->operands
) {
1763 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1764 register_file
.fill(op
);
1767 /* find a new register for the blocking variable */
1768 PhysReg reg
= get_reg(ctx
, register_file
, pc_op
.getTemp(), parallelcopy
, instr
);
1769 /* once again, disable killed operands */
1770 for (const Operand
& op
: instr
->operands
) {
1771 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1772 register_file
.clear(op
);
1774 for (unsigned k
= 0; k
< i
; k
++) {
1775 if (instr
->definitions
[k
].isTemp() && ctx
.defs_done
.test(k
) && !instr
->definitions
[k
].isKill())
1776 register_file
.fill(instr
->definitions
[k
]);
1778 pc_def
.setFixed(reg
);
1780 /* finish assignment of parallelcopy */
1781 ctx
.assignments
.emplace_back(reg
, pc_def
.regClass());
1782 assert(ctx
.assignments
.size() == ctx
.program
->peekAllocationId());
1783 parallelcopy
.emplace_back(pc_op
, pc_def
);
1785 /* add changes to reg_file */
1786 register_file
.clear(pc_op
);
1787 register_file
.fill(pc_def
);
1789 ctx
.defs_done
.set(i
);
1791 if (!definition
.isTemp())
1794 /* set live if it has a kill point */
1795 if (!definition
.isKill())
1796 live
.emplace(definition
.getTemp());
1798 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1799 register_file
.fill(definition
);
1802 /* handle all other definitions */
1803 for (unsigned i
= 0; i
< instr
->definitions
.size(); ++i
) {
1804 auto& definition
= instr
->definitions
[i
];
1806 if (definition
.isFixed() || !definition
.isTemp())
1810 if (definition
.hasHint() && register_file
[definition
.physReg().reg()] == 0)
1811 definition
.setFixed(definition
.physReg());
1812 else if (instr
->opcode
== aco_opcode::p_split_vector
) {
1813 PhysReg reg
= instr
->operands
[0].physReg();
1814 reg
.reg_b
+= i
* definition
.bytes();
1815 if (get_reg_specified(ctx
, register_file
, definition
.regClass(), parallelcopy
, instr
, reg
))
1816 definition
.setFixed(reg
);
1817 } else if (instr
->opcode
== aco_opcode::p_wqm
) {
1819 if (instr
->operands
[0].isKillBeforeDef() && instr
->operands
[0].getTemp().type() == definition
.getTemp().type()) {
1820 reg
= instr
->operands
[0].physReg();
1821 definition
.setFixed(reg
);
1822 assert(register_file
[reg
.reg()] == 0);
1824 } else if (instr
->opcode
== aco_opcode::p_extract_vector
) {
1826 if (instr
->operands
[0].isKillBeforeDef() &&
1827 instr
->operands
[0].getTemp().type() == definition
.getTemp().type()) {
1828 reg
= instr
->operands
[0].physReg();
1829 reg
.reg_b
+= definition
.bytes() * instr
->operands
[1].constantValue();
1830 assert(!register_file
.test(reg
, definition
.bytes()));
1831 definition
.setFixed(reg
);
1833 } else if (instr
->opcode
== aco_opcode::p_create_vector
) {
1834 PhysReg reg
= get_reg_create_vector(ctx
, register_file
, definition
.getTemp(),
1835 parallelcopy
, instr
);
1836 definition
.setFixed(reg
);
1839 if (!definition
.isFixed()) {
1840 Temp tmp
= definition
.getTemp();
1841 /* subdword instructions before RDNA write full registers */
1842 if (tmp
.regClass().is_subdword() &&
1843 !instr_can_access_subdword(instr
) &&
1844 ctx
.program
->chip_class
<= GFX9
) {
1845 assert(tmp
.bytes() <= 4);
1846 tmp
= Temp(definition
.tempId(), v1
);
1848 definition
.setFixed(get_reg(ctx
, register_file
, tmp
, parallelcopy
, instr
));
1851 assert(definition
.isFixed() && ((definition
.getTemp().type() == RegType::vgpr
&& definition
.physReg() >= 256) ||
1852 (definition
.getTemp().type() != RegType::vgpr
&& definition
.physReg() < 256)));
1853 ctx
.defs_done
.set(i
);
1855 /* set live if it has a kill point */
1856 if (!definition
.isKill())
1857 live
.emplace(definition
.getTemp());
1859 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1860 register_file
.fill(definition
);
1863 handle_pseudo(ctx
, register_file
, instr
.get());
1865 /* kill definitions and late-kill operands */
1866 for (const Definition
& def
: instr
->definitions
) {
1867 if (def
.isTemp() && def
.isKill())
1868 register_file
.clear(def
);
1870 for (const Operand
& op
: instr
->operands
) {
1871 if (op
.isTemp() && op
.isFirstKill() && op
.isLateKill())
1872 register_file
.clear(op
);
1875 /* emit parallelcopy */
1876 if (!parallelcopy
.empty()) {
1877 aco_ptr
<Pseudo_instruction
> pc
;
1878 pc
.reset(create_instruction
<Pseudo_instruction
>(aco_opcode::p_parallelcopy
, Format::PSEUDO
, parallelcopy
.size(), parallelcopy
.size()));
1879 bool temp_in_scc
= register_file
[scc
.reg()];
1880 bool sgpr_operands_alias_defs
= false;
1881 uint64_t sgpr_operands
[4] = {0, 0, 0, 0};
1882 for (unsigned i
= 0; i
< parallelcopy
.size(); i
++) {
1883 if (temp_in_scc
&& parallelcopy
[i
].first
.isTemp() && parallelcopy
[i
].first
.getTemp().type() == RegType::sgpr
) {
1884 if (!sgpr_operands_alias_defs
) {
1885 unsigned reg
= parallelcopy
[i
].first
.physReg().reg();
1886 unsigned size
= parallelcopy
[i
].first
.getTemp().size();
1887 sgpr_operands
[reg
/ 64u] |= ((1u << size
) - 1) << (reg
% 64u);
1889 reg
= parallelcopy
[i
].second
.physReg().reg();
1890 size
= parallelcopy
[i
].second
.getTemp().size();
1891 if (sgpr_operands
[reg
/ 64u] & ((1u << size
) - 1) << (reg
% 64u))
1892 sgpr_operands_alias_defs
= true;
1896 pc
->operands
[i
] = parallelcopy
[i
].first
;
1897 pc
->definitions
[i
] = parallelcopy
[i
].second
;
1898 assert(pc
->operands
[i
].size() == pc
->definitions
[i
].size());
1900 /* it might happen that the operand is already renamed. we have to restore the original name. */
1901 std::unordered_map
<unsigned, Temp
>::iterator it
= ctx
.orig_names
.find(pc
->operands
[i
].tempId());
1902 Temp orig
= it
!= ctx
.orig_names
.end() ? it
->second
: pc
->operands
[i
].getTemp();
1903 ctx
.orig_names
[pc
->definitions
[i
].tempId()] = orig
;
1904 ctx
.renames
[block
.index
][orig
.id()] = pc
->definitions
[i
].getTemp();
1906 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(pc
->operands
[i
].tempId());
1907 if (phi
!= ctx
.phi_map
.end())
1908 phi
->second
.uses
.emplace(pc
.get());
1911 if (temp_in_scc
&& sgpr_operands_alias_defs
) {
1912 /* disable definitions and re-enable operands */
1913 for (const Definition
& def
: instr
->definitions
) {
1914 if (def
.isTemp() && !def
.isKill())
1915 register_file
.clear(def
);
1917 for (const Operand
& op
: instr
->operands
) {
1918 if (op
.isTemp() && op
.isFirstKill())
1919 register_file
.block(op
.physReg(), op
.bytes());
1922 handle_pseudo(ctx
, register_file
, pc
.get());
1924 /* re-enable live vars */
1925 for (const Operand
& op
: instr
->operands
) {
1926 if (op
.isTemp() && op
.isFirstKill())
1927 register_file
.clear(op
);
1929 for (const Definition
& def
: instr
->definitions
) {
1930 if (def
.isTemp() && !def
.isKill())
1931 register_file
.fill(def
);
1934 pc
->tmp_in_scc
= false;
1937 instructions
.emplace_back(std::move(pc
));
1940 /* some instructions need VOP3 encoding if operand/definition is not assigned to VCC */
1941 bool instr_needs_vop3
= !instr
->isVOP3() &&
1942 ((instr
->format
== Format::VOPC
&& !(instr
->definitions
[0].physReg() == vcc
)) ||
1943 (instr
->opcode
== aco_opcode::v_cndmask_b32
&& !(instr
->operands
[2].physReg() == vcc
)) ||
1944 ((instr
->opcode
== aco_opcode::v_add_co_u32
||
1945 instr
->opcode
== aco_opcode::v_addc_co_u32
||
1946 instr
->opcode
== aco_opcode::v_sub_co_u32
||
1947 instr
->opcode
== aco_opcode::v_subb_co_u32
||
1948 instr
->opcode
== aco_opcode::v_subrev_co_u32
||
1949 instr
->opcode
== aco_opcode::v_subbrev_co_u32
) &&
1950 !(instr
->definitions
[1].physReg() == vcc
)) ||
1951 ((instr
->opcode
== aco_opcode::v_addc_co_u32
||
1952 instr
->opcode
== aco_opcode::v_subb_co_u32
||
1953 instr
->opcode
== aco_opcode::v_subbrev_co_u32
) &&
1954 !(instr
->operands
[2].physReg() == vcc
)));
1955 if (instr_needs_vop3
) {
1957 /* if the first operand is a literal, we have to move it to a reg */
1958 if (instr
->operands
.size() && instr
->operands
[0].isLiteral() && program
->chip_class
< GFX10
) {
1959 bool can_sgpr
= true;
1960 /* check, if we have to move to vgpr */
1961 for (const Operand
& op
: instr
->operands
) {
1962 if (op
.isTemp() && op
.getTemp().type() == RegType::sgpr
) {
1967 /* disable definitions and re-enable operands */
1968 for (const Definition
& def
: instr
->definitions
)
1969 register_file
.clear(def
);
1970 for (const Operand
& op
: instr
->operands
) {
1971 if (op
.isTemp() && op
.isFirstKill())
1972 register_file
.block(op
.physReg(), op
.bytes());
1974 Temp tmp
= {program
->allocateId(), can_sgpr
? s1
: v1
};
1975 ctx
.assignments
.emplace_back();
1976 PhysReg reg
= get_reg(ctx
, register_file
, tmp
, parallelcopy
, instr
);
1978 aco_ptr
<Instruction
> mov
;
1980 mov
.reset(create_instruction
<SOP1_instruction
>(aco_opcode::s_mov_b32
, Format::SOP1
, 1, 1));
1982 mov
.reset(create_instruction
<VOP1_instruction
>(aco_opcode::v_mov_b32
, Format::VOP1
, 1, 1));
1983 mov
->operands
[0] = instr
->operands
[0];
1984 mov
->definitions
[0] = Definition(tmp
);
1985 mov
->definitions
[0].setFixed(reg
);
1987 instr
->operands
[0] = Operand(tmp
);
1988 instr
->operands
[0].setFixed(reg
);
1989 instructions
.emplace_back(std::move(mov
));
1990 /* re-enable live vars */
1991 for (const Operand
& op
: instr
->operands
) {
1992 if (op
.isTemp() && op
.isFirstKill())
1993 register_file
.clear(op
);
1995 for (const Definition
& def
: instr
->definitions
) {
1996 if (def
.isTemp() && !def
.isKill())
1997 register_file
.fill(def
);
2001 /* change the instruction to VOP3 to enable an arbitrary register pair as dst */
2002 aco_ptr
<Instruction
> tmp
= std::move(instr
);
2003 Format format
= asVOP3(tmp
->format
);
2004 instr
.reset(create_instruction
<VOP3A_instruction
>(tmp
->opcode
, format
, tmp
->operands
.size(), tmp
->definitions
.size()));
2005 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
2006 Operand
& operand
= tmp
->operands
[i
];
2007 instr
->operands
[i
] = operand
;
2008 /* keep phi_map up to date */
2009 if (operand
.isTemp()) {
2010 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.tempId());
2011 if (phi
!= ctx
.phi_map
.end()) {
2012 phi
->second
.uses
.erase(tmp
.get());
2013 phi
->second
.uses
.emplace(instr
.get());
2017 std::copy(tmp
->definitions
.begin(), tmp
->definitions
.end(), instr
->definitions
.begin());
2019 instructions
.emplace_back(std::move(*it
));
2021 } /* end for Instr */
2023 block
.instructions
= std::move(instructions
);
2025 ctx
.filled
[block
.index
] = true;
2026 for (unsigned succ_idx
: block
.linear_succs
) {
2027 Block
& succ
= program
->blocks
[succ_idx
];
2028 /* seal block if all predecessors are filled */
2029 bool all_filled
= true;
2030 for (unsigned pred_idx
: succ
.linear_preds
) {
2031 if (!ctx
.filled
[pred_idx
]) {
2037 ctx
.sealed
[succ_idx
] = true;
2039 /* finish incomplete phis and check if they became trivial */
2040 for (Instruction
* phi
: ctx
.incomplete_phis
[succ_idx
]) {
2041 std::vector
<unsigned> preds
= phi
->definitions
[0].getTemp().is_linear() ? succ
.linear_preds
: succ
.logical_preds
;
2042 for (unsigned i
= 0; i
< phi
->operands
.size(); i
++) {
2043 phi
->operands
[i
].setTemp(read_variable(ctx
, phi
->operands
[i
].getTemp(), preds
[i
]));
2044 phi
->operands
[i
].setFixed(ctx
.assignments
[phi
->operands
[i
].tempId()].reg
);
2046 try_remove_trivial_phi(ctx
, phi
->definitions
[0].getTemp());
2048 /* complete the original phi nodes, but no need to check triviality */
2049 for (aco_ptr
<Instruction
>& instr
: succ
.instructions
) {
2052 std::vector
<unsigned> preds
= instr
->opcode
== aco_opcode::p_phi
? succ
.logical_preds
: succ
.linear_preds
;
2054 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
2055 auto& operand
= instr
->operands
[i
];
2056 if (!operand
.isTemp())
2058 operand
.setTemp(read_variable(ctx
, operand
.getTemp(), preds
[i
]));
2059 operand
.setFixed(ctx
.assignments
[operand
.tempId()].reg
);
2060 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.getTemp().id());
2061 if (phi
!= ctx
.phi_map
.end())
2062 phi
->second
.uses
.emplace(instr
.get());
2069 /* remove trivial phis */
2070 for (Block
& block
: program
->blocks
) {
2071 auto end
= std::find_if(block
.instructions
.begin(), block
.instructions
.end(),
2072 [](aco_ptr
<Instruction
>& instr
) { return !is_phi(instr
);});
2073 auto middle
= std::remove_if(block
.instructions
.begin(), end
,
2074 [](const aco_ptr
<Instruction
>& instr
) { return instr
->definitions
.empty();});
2075 block
.instructions
.erase(middle
, end
);
2078 /* find scc spill registers which may be needed for parallelcopies created by phis */
2079 for (Block
& block
: program
->blocks
) {
2080 if (block
.linear_preds
.size() <= 1)
2083 std::bitset
<128> regs
= sgpr_live_in
[block
.index
];
2087 /* choose a register */
2089 for (; reg
< ctx
.program
->max_reg_demand
.sgpr
&& regs
[reg
]; reg
++)
2091 assert(reg
< ctx
.program
->max_reg_demand
.sgpr
);
2092 adjust_max_used_regs(ctx
, s1
, reg
);
2094 /* update predecessors */
2095 for (unsigned& pred_index
: block
.linear_preds
) {
2096 Block
& pred
= program
->blocks
[pred_index
];
2097 pred
.scc_live_out
= true;
2098 pred
.scratch_sgpr
= PhysReg
{(uint16_t)reg
};
2102 /* num_gpr = rnd_up(max_used_gpr + 1) */
2103 program
->config
->num_vgprs
= align(ctx
.max_used_vgpr
+ 1, 4);
2104 if (program
->family
== CHIP_TONGA
|| program
->family
== CHIP_ICELAND
) /* workaround hardware bug */
2105 program
->config
->num_sgprs
= get_sgpr_alloc(program
, program
->sgpr_limit
);
2107 program
->config
->num_sgprs
= align(ctx
.max_used_sgpr
+ 1 + get_extra_sgprs(program
), 8);