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
, RegClass rc
) {
161 if (rc
.is_subdword())
162 fill_subdword(start
, rc
.bytes(), 0xFFFFFFFF);
164 fill(start
, rc
.size(), 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
&= ((int)i
<= 4 - (int)rc
.bytes() || reg_file
[entry
.first
+ 1] == 0);
386 PhysReg res
{entry
.first
};
393 stride
= 1; /* stride in full registers */
394 rc
= info
.rc
= RegClass(RegType::vgpr
, size
);
399 for (unsigned stride
= 8; stride
> 1; stride
/= 2) {
402 info
.stride
= stride
;
403 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
408 /* best fit algorithm: find the smallest gap to fit in the variable */
409 unsigned best_pos
= 0xFFFF;
410 unsigned gap_size
= 0xFFFF;
411 unsigned last_pos
= 0xFFFF;
413 for (unsigned current_reg
= lb
; current_reg
< ub
; current_reg
++) {
415 if (reg_file
[current_reg
] == 0 && !ctx
.war_hint
[current_reg
]) {
416 if (last_pos
== 0xFFFF)
417 last_pos
= current_reg
;
419 /* stop searching after max_used_gpr */
420 if (current_reg
== ctx
.max_used_sgpr
+ 1 || current_reg
== 256 + ctx
.max_used_vgpr
+ 1)
426 if (last_pos
== 0xFFFF)
429 /* early return on exact matches */
430 if (last_pos
+ size
== current_reg
) {
431 adjust_max_used_regs(ctx
, rc
, last_pos
);
432 return {PhysReg
{last_pos
}, true};
435 /* check if it fits and the gap size is smaller */
436 if (last_pos
+ size
< current_reg
&& current_reg
- last_pos
< gap_size
) {
438 gap_size
= current_reg
- last_pos
;
444 if (last_pos
+ size
<= ub
&& ub
- last_pos
< gap_size
) {
446 gap_size
= ub
- last_pos
;
449 if (best_pos
== 0xFFFF)
452 /* find best position within gap by leaving a good stride for other variables*/
453 unsigned buffer
= gap_size
- size
;
455 if (((best_pos
+ size
) % 8 != 0 && (best_pos
+ buffer
) % 8 == 0) ||
456 ((best_pos
+ size
) % 4 != 0 && (best_pos
+ buffer
) % 4 == 0) ||
457 ((best_pos
+ size
) % 2 != 0 && (best_pos
+ buffer
) % 2 == 0))
458 best_pos
= best_pos
+ buffer
;
461 adjust_max_used_regs(ctx
, rc
, best_pos
);
462 return {PhysReg
{best_pos
}, true};
466 unsigned reg_lo
= lb
;
467 unsigned reg_hi
= lb
+ size
- 1;
468 while (!found
&& reg_lo
+ size
<= ub
) {
469 if (reg_file
[reg_lo
] != 0) {
473 reg_hi
= reg_lo
+ size
- 1;
475 for (unsigned reg
= reg_lo
+ 1; found
&& reg
<= reg_hi
; reg
++) {
476 if (reg_file
[reg
] != 0 || ctx
.war_hint
[reg
])
480 adjust_max_used_regs(ctx
, rc
, reg_lo
);
481 return {PhysReg
{reg_lo
}, true};
490 /* collect variables from a register area and clear reg_file */
491 std::set
<std::pair
<unsigned, unsigned>> collect_vars(ra_ctx
& ctx
, RegisterFile
& reg_file
,
492 PhysReg reg
, unsigned size
)
494 std::set
<std::pair
<unsigned, unsigned>> vars
;
495 for (unsigned j
= reg
; j
< reg
+ size
; j
++) {
496 if (reg_file
.is_blocked(PhysReg
{j
}))
498 if (reg_file
[j
] == 0xF0000000) {
499 for (unsigned k
= 0; k
< 4; k
++) {
500 unsigned id
= reg_file
.subdword_regs
[j
][k
];
502 assignment
& var
= ctx
.assignments
[id
];
503 vars
.emplace(var
.rc
.bytes(), id
);
504 reg_file
.clear(var
.reg
, var
.rc
);
509 } else if (reg_file
[j
] != 0) {
510 unsigned id
= reg_file
[j
];
511 assignment
& var
= ctx
.assignments
[id
];
512 vars
.emplace(var
.rc
.bytes(), id
);
513 reg_file
.clear(var
.reg
, var
.rc
);
519 bool get_regs_for_copies(ra_ctx
& ctx
,
520 RegisterFile
& reg_file
,
521 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
522 const std::set
<std::pair
<unsigned, unsigned>> &vars
,
523 uint32_t lb
, uint32_t ub
,
524 aco_ptr
<Instruction
>& instr
,
529 /* variables are sorted from small sized to large */
530 /* NOTE: variables are also sorted by ID. this only affects a very small number of shaders slightly though. */
531 for (std::set
<std::pair
<unsigned, unsigned>>::const_reverse_iterator it
= vars
.rbegin(); it
!= vars
.rend(); ++it
) {
532 unsigned id
= it
->second
;
533 assignment
& var
= ctx
.assignments
[id
];
534 DefInfo info
= DefInfo(ctx
, ctx
.pseudo_dummy
, var
.rc
);
535 uint32_t size
= info
.size
;
537 /* check if this is a dead operand, then we can re-use the space from the definition */
538 bool is_dead_operand
= false;
539 for (unsigned i
= 0; !is_phi(instr
) && !is_dead_operand
&& (i
< instr
->operands
.size()); i
++) {
540 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isKillBeforeDef() && instr
->operands
[i
].tempId() == id
)
541 is_dead_operand
= true;
544 std::pair
<PhysReg
, bool> res
;
545 if (is_dead_operand
) {
546 if (instr
->opcode
== aco_opcode::p_create_vector
) {
547 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
548 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].tempId() == id
) {
549 PhysReg
reg(def_reg_lo
);
551 assert(!reg_file
.test(reg
, var
.rc
.bytes()));
557 info
.lb
= def_reg_lo
;
558 info
.ub
= def_reg_hi
+ 1;
559 res
= get_reg_simple(ctx
, reg_file
, info
);
563 info
.ub
= def_reg_lo
;
564 res
= get_reg_simple(ctx
, reg_file
, info
);
566 info
.lb
= (def_reg_hi
+ info
.stride
) & ~(info
.stride
- 1);
568 res
= get_reg_simple(ctx
, reg_file
, info
);
573 /* mark the area as blocked */
574 reg_file
.block(res
.first
, var
.rc
);
576 /* create parallelcopy pair (without definition id) */
577 Temp tmp
= Temp(id
, var
.rc
);
578 Operand pc_op
= Operand(tmp
);
579 pc_op
.setFixed(var
.reg
);
580 Definition pc_def
= Definition(res
.first
, pc_op
.regClass());
581 parallelcopies
.emplace_back(pc_op
, pc_def
);
585 unsigned best_pos
= lb
;
586 unsigned num_moves
= 0xFF;
587 unsigned num_vars
= 0;
589 /* we use a sliding window to find potential positions */
590 unsigned reg_lo
= lb
;
591 unsigned reg_hi
= lb
+ size
- 1;
592 unsigned stride
= var
.rc
.is_subdword() ? 1 : info
.stride
;
593 for (reg_lo
= lb
, reg_hi
= lb
+ size
- 1; reg_hi
< ub
; reg_lo
+= stride
, reg_hi
+= stride
) {
594 if (!is_dead_operand
&& ((reg_lo
>= def_reg_lo
&& reg_lo
<= def_reg_hi
) ||
595 (reg_hi
>= def_reg_lo
&& reg_hi
<= def_reg_hi
)))
598 /* second, check that we have at most k=num_moves elements in the window
599 * and no element is larger than the currently processed one */
602 unsigned last_var
= 0;
604 for (unsigned j
= reg_lo
; found
&& j
<= reg_hi
; j
++) {
605 if (reg_file
[j
] == 0 || reg_file
[j
] == last_var
)
608 if (reg_file
.is_blocked(PhysReg
{j
}) || k
> num_moves
) {
612 if (reg_file
[j
] == 0xF0000000) {
617 /* we cannot split live ranges of linear vgprs */
618 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6)) {
622 bool is_kill
= false;
623 for (const Operand
& op
: instr
->operands
) {
624 if (op
.isTemp() && op
.isKillBeforeDef() && op
.tempId() == reg_file
[j
]) {
629 if (!is_kill
&& ctx
.assignments
[reg_file
[j
]].rc
.size() >= size
) {
634 k
+= ctx
.assignments
[reg_file
[j
]].rc
.size();
635 last_var
= reg_file
[j
];
637 if (k
> num_moves
|| (k
== num_moves
&& n
<= num_vars
)) {
650 /* FIXME: we messed up and couldn't find space for the variables to be copied */
651 if (num_moves
== 0xFF)
655 reg_hi
= best_pos
+ size
- 1;
657 /* collect variables and block reg file */
658 std::set
<std::pair
<unsigned, unsigned>> new_vars
= collect_vars(ctx
, reg_file
, PhysReg
{reg_lo
}, size
);
660 /* mark the area as blocked */
661 reg_file
.block(PhysReg
{reg_lo
}, var
.rc
);
663 if (!get_regs_for_copies(ctx
, reg_file
, parallelcopies
, new_vars
, lb
, ub
, instr
, def_reg_lo
, def_reg_hi
))
666 adjust_max_used_regs(ctx
, var
.rc
, reg_lo
);
668 /* create parallelcopy pair (without definition id) */
669 Temp tmp
= Temp(id
, var
.rc
);
670 Operand pc_op
= Operand(tmp
);
671 pc_op
.setFixed(var
.reg
);
672 Definition pc_def
= Definition(PhysReg
{reg_lo
}, pc_op
.regClass());
673 parallelcopies
.emplace_back(pc_op
, pc_def
);
680 std::pair
<PhysReg
, bool> get_reg_impl(ra_ctx
& ctx
,
681 RegisterFile
& reg_file
,
682 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
684 aco_ptr
<Instruction
>& instr
)
686 uint32_t lb
= info
.lb
;
687 uint32_t ub
= info
.ub
;
688 uint32_t size
= info
.size
;
689 uint32_t stride
= info
.stride
;
690 RegClass rc
= info
.rc
;
692 /* check how many free regs we have */
693 unsigned regs_free
= reg_file
.count_zero(PhysReg
{lb
}, ub
-lb
);
695 /* mark and count killed operands */
696 unsigned killed_ops
= 0;
697 for (unsigned j
= 0; !is_phi(instr
) && j
< instr
->operands
.size(); j
++) {
698 if (instr
->operands
[j
].isTemp() &&
699 instr
->operands
[j
].isFirstKillBeforeDef() &&
700 instr
->operands
[j
].physReg() >= lb
&&
701 instr
->operands
[j
].physReg() < ub
) {
702 assert(instr
->operands
[j
].isFixed());
703 assert(!reg_file
.test(instr
->operands
[j
].physReg(), instr
->operands
[j
].bytes()));
704 reg_file
.block(instr
->operands
[j
].physReg(), instr
->operands
[j
].regClass());
705 killed_ops
+= instr
->operands
[j
].getTemp().size();
709 assert(regs_free
>= size
);
710 /* we might have to move dead operands to dst in order to make space */
711 unsigned op_moves
= 0;
713 if (size
> (regs_free
- killed_ops
))
714 op_moves
= size
- (regs_free
- killed_ops
);
716 /* find the best position to place the definition */
717 unsigned best_pos
= lb
;
718 unsigned num_moves
= 0xFF;
719 unsigned num_vars
= 0;
721 /* we use a sliding window to check potential positions */
722 unsigned reg_lo
= lb
;
723 unsigned reg_hi
= lb
+ size
- 1;
724 for (reg_lo
= lb
, reg_hi
= lb
+ size
- 1; reg_hi
< ub
; reg_lo
+= stride
, reg_hi
+= stride
) {
725 /* first check the edges: this is what we have to fix to allow for num_moves > size */
726 if (reg_lo
> lb
&& reg_file
[reg_lo
] != 0 && reg_file
[reg_lo
] == reg_file
[reg_lo
- 1])
728 if (reg_hi
< ub
- 1 && reg_file
[reg_hi
] != 0 && reg_file
[reg_hi
] == reg_file
[reg_hi
+ 1])
731 /* second, check that we have at most k=num_moves elements in the window
732 * and no element is larger than the currently processed one */
733 unsigned k
= op_moves
;
735 unsigned remaining_op_moves
= op_moves
;
736 unsigned last_var
= 0;
738 bool aligned
= rc
== RegClass::v4
&& reg_lo
% 4 == 0;
739 for (unsigned j
= reg_lo
; found
&& j
<= reg_hi
; j
++) {
740 if (reg_file
[j
] == 0 || reg_file
[j
] == last_var
)
743 /* dead operands effectively reduce the number of estimated moves */
744 if (reg_file
.is_blocked(PhysReg
{j
})) {
745 if (remaining_op_moves
) {
747 remaining_op_moves
--;
752 if (reg_file
[j
] == 0xF0000000) {
758 if (ctx
.assignments
[reg_file
[j
]].rc
.size() >= size
) {
763 /* we cannot split live ranges of linear vgprs */
764 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6)) {
769 k
+= ctx
.assignments
[reg_file
[j
]].rc
.size();
771 last_var
= reg_file
[j
];
774 if (!found
|| k
> num_moves
)
776 if (k
== num_moves
&& n
< num_vars
)
778 if (!aligned
&& k
== num_moves
&& n
== num_vars
)
788 if (num_moves
== 0xFF) {
789 /* remove killed operands from reg_file once again */
790 for (unsigned i
= 0; !is_phi(instr
) && i
< instr
->operands
.size(); i
++) {
791 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isFirstKillBeforeDef())
792 reg_file
.clear(instr
->operands
[i
]);
794 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
795 Definition def
= instr
->definitions
[i
];
796 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
802 RegisterFile register_file
= reg_file
;
804 /* now, we figured the placement for our definition */
805 std::set
<std::pair
<unsigned, unsigned>> vars
= collect_vars(ctx
, reg_file
, PhysReg
{best_pos
}, size
);
807 if (instr
->opcode
== aco_opcode::p_create_vector
) {
808 /* move killed operands which aren't yet at the correct position */
809 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].size(), i
++) {
810 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isFirstKillBeforeDef() &&
811 instr
->operands
[i
].getTemp().type() == rc
.type()) {
813 if (instr
->operands
[i
].physReg() != best_pos
+ offset
) {
814 vars
.emplace(instr
->operands
[i
].bytes(), instr
->operands
[i
].tempId());
815 reg_file
.clear(instr
->operands
[i
]);
817 reg_file
.fill(instr
->operands
[i
]);
822 /* re-enable the killed operands */
823 for (unsigned j
= 0; !is_phi(instr
) && j
< instr
->operands
.size(); j
++) {
824 if (instr
->operands
[j
].isTemp() && instr
->operands
[j
].isFirstKill())
825 reg_file
.fill(instr
->operands
[j
]);
829 std::vector
<std::pair
<Operand
, Definition
>> pc
;
830 if (!get_regs_for_copies(ctx
, reg_file
, pc
, vars
, lb
, ub
, instr
, best_pos
, best_pos
+ size
- 1)) {
831 reg_file
= std::move(register_file
);
832 /* remove killed operands from reg_file once again */
833 if (!is_phi(instr
)) {
834 for (const Operand
& op
: instr
->operands
) {
835 if (op
.isTemp() && op
.isFirstKillBeforeDef())
839 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
840 Definition
& def
= instr
->definitions
[i
];
841 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
847 parallelcopies
.insert(parallelcopies
.end(), pc
.begin(), pc
.end());
849 /* we set the definition regs == 0. the actual caller is responsible for correct setting */
850 reg_file
.clear(PhysReg
{best_pos
}, rc
);
852 update_renames(ctx
, reg_file
, parallelcopies
, instr
);
854 /* remove killed operands from reg_file once again */
855 for (unsigned i
= 0; !is_phi(instr
) && i
< instr
->operands
.size(); i
++) {
856 if (!instr
->operands
[i
].isTemp() || !instr
->operands
[i
].isFixed())
858 assert(!instr
->operands
[i
].isUndefined());
859 if (instr
->operands
[i
].isFirstKillBeforeDef())
860 reg_file
.clear(instr
->operands
[i
]);
862 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
863 Definition def
= instr
->definitions
[i
];
864 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
868 adjust_max_used_regs(ctx
, rc
, best_pos
);
869 return {PhysReg
{best_pos
}, true};
872 bool get_reg_specified(ra_ctx
& ctx
,
873 RegisterFile
& reg_file
,
875 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
876 aco_ptr
<Instruction
>& instr
,
879 if (rc
.is_subdword() && reg
.byte() && !instr_can_access_subdword(instr
))
881 if (!rc
.is_subdword() && reg
.byte())
884 uint32_t size
= rc
.size();
888 if (rc
.type() == RegType::vgpr
) {
890 ub
= 256 + ctx
.program
->max_reg_demand
.vgpr
;
896 if (reg
% stride
!= 0)
899 ub
= ctx
.program
->max_reg_demand
.sgpr
;
902 uint32_t reg_lo
= reg
.reg();
903 uint32_t reg_hi
= reg
+ (size
- 1);
905 if (reg_lo
< lb
|| reg_hi
>= ub
|| reg_lo
> reg_hi
)
908 if (reg_file
.test(reg
, rc
.bytes()))
911 adjust_max_used_regs(ctx
, rc
, reg_lo
);
915 PhysReg
get_reg(ra_ctx
& ctx
,
916 RegisterFile
& reg_file
,
918 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
919 aco_ptr
<Instruction
>& instr
)
921 if (ctx
.affinities
.find(temp
.id()) != ctx
.affinities
.end() &&
922 ctx
.assignments
[ctx
.affinities
[temp
.id()]].assigned
) {
923 PhysReg reg
= ctx
.assignments
[ctx
.affinities
[temp
.id()]].reg
;
924 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
928 if (ctx
.vectors
.find(temp
.id()) != ctx
.vectors
.end()) {
929 Instruction
* vec
= ctx
.vectors
[temp
.id()];
930 unsigned byte_offset
= 0;
931 for (const Operand
& op
: vec
->operands
) {
932 if (op
.isTemp() && op
.tempId() == temp
.id())
935 byte_offset
+= op
.bytes();
938 for (const Operand
& op
: vec
->operands
) {
940 op
.tempId() != temp
.id() &&
941 op
.getTemp().type() == temp
.type() &&
942 ctx
.assignments
[op
.tempId()].assigned
) {
943 PhysReg reg
= ctx
.assignments
[op
.tempId()].reg
;
944 reg
.reg_b
+= (byte_offset
- k
);
945 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
951 DefInfo
info(ctx
, ctx
.pseudo_dummy
, vec
->definitions
[0].regClass());
952 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
953 PhysReg reg
= res
.first
;
955 reg
.reg_b
+= byte_offset
;
956 /* make sure to only use byte offset if the instruction supports it */
957 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
962 DefInfo
info(ctx
, instr
, temp
.regClass());
964 /* try to find space without live-range splits */
965 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
970 /* try to find space with live-range splits */
971 res
= get_reg_impl(ctx
, reg_file
, parallelcopies
, info
, instr
);
976 /* try using more registers */
978 /* We should only fail here because keeping under the limit would require
980 assert(reg_file
.count_zero(PhysReg
{info
.lb
}, info
.ub
-info
.lb
) >= info
.size
);
982 uint16_t max_addressible_sgpr
= ctx
.program
->sgpr_limit
;
983 uint16_t max_addressible_vgpr
= ctx
.program
->vgpr_limit
;
984 if (info
.rc
.type() == RegType::vgpr
&& ctx
.program
->max_reg_demand
.vgpr
< max_addressible_vgpr
) {
985 update_vgpr_sgpr_demand(ctx
.program
, RegisterDemand(ctx
.program
->max_reg_demand
.vgpr
+ 1, ctx
.program
->max_reg_demand
.sgpr
));
986 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
987 } else if (info
.rc
.type() == RegType::sgpr
&& ctx
.program
->max_reg_demand
.sgpr
< max_addressible_sgpr
) {
988 update_vgpr_sgpr_demand(ctx
.program
, RegisterDemand(ctx
.program
->max_reg_demand
.vgpr
, ctx
.program
->max_reg_demand
.sgpr
+ 1));
989 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
992 //FIXME: if nothing helps, shift-rotate the registers to make space
994 fprintf(stderr
, "ACO: failed to allocate registers during shader compilation\n");
998 PhysReg
get_reg_create_vector(ra_ctx
& ctx
,
999 RegisterFile
& reg_file
,
1001 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
1002 aco_ptr
<Instruction
>& instr
)
1004 RegClass rc
= temp
.regClass();
1005 /* create_vector instructions have different costs w.r.t. register coalescing */
1006 uint32_t size
= rc
.size();
1007 uint32_t bytes
= rc
.bytes();
1008 uint32_t stride
= 1;
1010 if (rc
.type() == RegType::vgpr
) {
1012 ub
= 256 + ctx
.program
->max_reg_demand
.vgpr
;
1015 ub
= ctx
.program
->max_reg_demand
.sgpr
;
1022 //TODO: improve p_create_vector for sub-dword vectors
1024 unsigned best_pos
= -1;
1025 unsigned num_moves
= 0xFF;
1026 bool best_war_hint
= true;
1028 /* test for each operand which definition placement causes the least shuffle instructions */
1029 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
1030 // TODO: think about, if we can alias live operands on the same register
1031 if (!instr
->operands
[i
].isTemp() || !instr
->operands
[i
].isKillBeforeDef() || instr
->operands
[i
].getTemp().type() != rc
.type())
1034 if (offset
> instr
->operands
[i
].physReg().reg_b
)
1037 unsigned reg_lo
= instr
->operands
[i
].physReg().reg_b
- offset
;
1041 unsigned reg_hi
= reg_lo
+ size
- 1;
1044 /* no need to check multiple times */
1045 if (reg_lo
== best_pos
)
1049 // TODO: this can be improved */
1050 if (reg_lo
< lb
|| reg_hi
>= ub
|| reg_lo
% stride
!= 0)
1052 if (reg_lo
> lb
&& reg_file
[reg_lo
] != 0 && reg_file
[reg_lo
] == reg_file
[reg_lo
- 1])
1054 if (reg_hi
< ub
- 1 && reg_file
[reg_hi
] != 0 && reg_file
[reg_hi
] == reg_file
[reg_hi
+ 1])
1057 /* count variables to be moved and check war_hint */
1058 bool war_hint
= false;
1059 bool linear_vgpr
= false;
1060 for (unsigned j
= reg_lo
; j
<= reg_hi
&& !linear_vgpr
; j
++) {
1061 if (reg_file
[j
] != 0) {
1062 if (reg_file
[j
] == 0xF0000000) {
1065 unsigned bytes_left
= bytes
- (j
- reg_lo
) * 4;
1066 for (unsigned k
= 0; k
< MIN2(bytes_left
, 4); k
++, reg
.reg_b
++)
1067 k
+= reg_file
.test(reg
, 1);
1070 /* we cannot split live ranges of linear vgprs */
1071 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6))
1075 war_hint
|= ctx
.war_hint
[j
];
1077 if (linear_vgpr
|| (war_hint
&& !best_war_hint
))
1080 /* count operands in wrong positions */
1081 for (unsigned j
= 0, offset
= 0; j
< instr
->operands
.size(); offset
+= instr
->operands
[j
].bytes(), j
++) {
1083 !instr
->operands
[j
].isTemp() ||
1084 instr
->operands
[j
].getTemp().type() != rc
.type())
1086 if (instr
->operands
[j
].physReg().reg_b
!= reg_lo
* 4 + offset
)
1087 k
+= instr
->operands
[j
].bytes();
1089 bool aligned
= rc
== RegClass::v4
&& reg_lo
% 4 == 0;
1090 if (k
> num_moves
|| (!aligned
&& k
== num_moves
))
1095 best_war_hint
= war_hint
;
1098 if (num_moves
>= bytes
)
1099 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
1101 /* collect variables to be moved */
1102 std::set
<std::pair
<unsigned, unsigned>> vars
= collect_vars(ctx
, reg_file
, PhysReg
{best_pos
}, size
);
1104 /* move killed operands which aren't yet at the correct position */
1105 uint64_t moved_operand_mask
= 0;
1106 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
1107 if (instr
->operands
[i
].isTemp() &&
1108 instr
->operands
[i
].isFirstKillBeforeDef() &&
1109 instr
->operands
[i
].getTemp().type() == rc
.type() &&
1110 instr
->operands
[i
].physReg().reg_b
!= best_pos
* 4 + offset
) {
1111 vars
.emplace(instr
->operands
[i
].bytes(), instr
->operands
[i
].tempId());
1112 moved_operand_mask
|= (uint64_t)1 << i
;
1116 ASSERTED
bool success
= false;
1117 success
= get_regs_for_copies(ctx
, reg_file
, parallelcopies
, vars
, lb
, ub
, instr
, best_pos
, best_pos
+ size
- 1);
1120 update_renames(ctx
, reg_file
, parallelcopies
, instr
);
1121 adjust_max_used_regs(ctx
, rc
, best_pos
);
1123 while (moved_operand_mask
) {
1124 unsigned i
= u_bit_scan64(&moved_operand_mask
);
1125 assert(instr
->operands
[i
].isFirstKillBeforeDef());
1126 reg_file
.clear(instr
->operands
[i
]);
1129 return PhysReg
{best_pos
};
1132 void handle_pseudo(ra_ctx
& ctx
,
1133 const RegisterFile
& reg_file
,
1136 if (instr
->format
!= Format::PSEUDO
)
1139 /* all instructions which use handle_operands() need this information */
1140 switch (instr
->opcode
) {
1141 case aco_opcode::p_extract_vector
:
1142 case aco_opcode::p_create_vector
:
1143 case aco_opcode::p_split_vector
:
1144 case aco_opcode::p_parallelcopy
:
1145 case aco_opcode::p_wqm
:
1151 /* if all definitions are vgpr, no need to care for SCC */
1152 bool writes_sgpr
= false;
1153 for (Definition
& def
: instr
->definitions
) {
1154 if (def
.getTemp().type() == RegType::sgpr
) {
1159 /* if all operands are constant, no need to care either */
1160 bool reads_sgpr
= false;
1161 for (Operand
& op
: instr
->operands
) {
1162 if (op
.isTemp() && op
.getTemp().type() == RegType::sgpr
) {
1167 if (!(writes_sgpr
&& reads_sgpr
))
1170 Pseudo_instruction
*pi
= (Pseudo_instruction
*)instr
;
1171 if (reg_file
[scc
.reg()]) {
1172 pi
->tmp_in_scc
= true;
1174 int reg
= ctx
.max_used_sgpr
;
1175 for (; reg
>= 0 && reg_file
[reg
]; reg
--)
1178 reg
= ctx
.max_used_sgpr
+ 1;
1179 for (; reg
< ctx
.program
->max_reg_demand
.sgpr
&& reg_file
[reg
]; reg
++)
1181 assert(reg
< ctx
.program
->max_reg_demand
.sgpr
);
1184 adjust_max_used_regs(ctx
, s1
, reg
);
1185 pi
->scratch_sgpr
= PhysReg
{(unsigned)reg
};
1187 pi
->tmp_in_scc
= false;
1191 bool operand_can_use_reg(aco_ptr
<Instruction
>& instr
, unsigned idx
, PhysReg reg
)
1193 if (instr
->operands
[idx
].isFixed())
1194 return instr
->operands
[idx
].physReg() == reg
;
1196 if (!instr_can_access_subdword(instr
) && reg
.byte())
1199 switch (instr
->format
) {
1201 return reg
!= scc
&&
1203 (reg
!= m0
|| idx
== 1 || idx
== 3) && /* offset can be m0 */
1204 (reg
!= vcc
|| (instr
->definitions
.empty() && idx
== 2)); /* sdata can be vcc */
1206 // TODO: there are more instructions with restrictions on registers
1211 void get_reg_for_operand(ra_ctx
& ctx
, RegisterFile
& register_file
,
1212 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopy
,
1213 aco_ptr
<Instruction
>& instr
, Operand
& operand
)
1215 /* check if the operand is fixed */
1217 if (operand
.isFixed()) {
1218 assert(operand
.physReg() != ctx
.assignments
[operand
.tempId()].reg
);
1220 /* check if target reg is blocked, and move away the blocking var */
1221 if (register_file
[operand
.physReg().reg()]) {
1222 assert(register_file
[operand
.physReg()] != 0xF0000000);
1223 uint32_t blocking_id
= register_file
[operand
.physReg().reg()];
1224 RegClass rc
= ctx
.assignments
[blocking_id
].rc
;
1225 Operand pc_op
= Operand(Temp
{blocking_id
, rc
});
1226 pc_op
.setFixed(operand
.physReg());
1229 PhysReg reg
= get_reg(ctx
, register_file
, pc_op
.getTemp(), parallelcopy
, ctx
.pseudo_dummy
);
1230 Definition pc_def
= Definition(PhysReg
{reg
}, pc_op
.regClass());
1231 register_file
.clear(pc_op
);
1232 parallelcopy
.emplace_back(pc_op
, pc_def
);
1234 dst
= operand
.physReg();
1237 dst
= get_reg(ctx
, register_file
, operand
.getTemp(), parallelcopy
, instr
);
1240 Operand pc_op
= operand
;
1241 pc_op
.setFixed(ctx
.assignments
[operand
.tempId()].reg
);
1242 Definition pc_def
= Definition(dst
, pc_op
.regClass());
1243 register_file
.clear(pc_op
);
1244 parallelcopy
.emplace_back(pc_op
, pc_def
);
1245 update_renames(ctx
, register_file
, parallelcopy
, instr
);
1248 Temp
read_variable(ra_ctx
& ctx
, Temp val
, unsigned block_idx
)
1250 std::unordered_map
<unsigned, Temp
>::iterator it
= ctx
.renames
[block_idx
].find(val
.id());
1251 if (it
== ctx
.renames
[block_idx
].end())
1257 Temp
handle_live_in(ra_ctx
& ctx
, Temp val
, Block
* block
)
1259 std::vector
<unsigned>& preds
= val
.is_linear() ? block
->linear_preds
: block
->logical_preds
;
1260 if (preds
.size() == 0 || val
.regClass() == val
.regClass().as_linear())
1263 assert(preds
.size() > 0);
1266 if (!ctx
.sealed
[block
->index
]) {
1267 /* consider rename from already processed predecessor */
1268 Temp tmp
= read_variable(ctx
, val
, preds
[0]);
1270 /* if the block is not sealed yet, we create an incomplete phi (which might later get removed again) */
1271 new_val
= Temp
{ctx
.program
->allocateId(), val
.regClass()};
1272 ctx
.assignments
.emplace_back();
1273 aco_opcode opcode
= val
.is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1274 aco_ptr
<Instruction
> phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1275 phi
->definitions
[0] = Definition(new_val
);
1276 for (unsigned i
= 0; i
< preds
.size(); i
++)
1277 phi
->operands
[i
] = Operand(val
);
1278 if (tmp
.regClass() == new_val
.regClass())
1279 ctx
.affinities
[new_val
.id()] = tmp
.id();
1281 ctx
.phi_map
.emplace(new_val
.id(), phi_info
{phi
.get(), block
->index
});
1282 ctx
.incomplete_phis
[block
->index
].emplace_back(phi
.get());
1283 block
->instructions
.insert(block
->instructions
.begin(), std::move(phi
));
1285 } else if (preds
.size() == 1) {
1286 /* if the block has only one predecessor, just look there for the name */
1287 new_val
= read_variable(ctx
, val
, preds
[0]);
1289 /* there are multiple predecessors and the block is sealed */
1290 Temp ops
[preds
.size()];
1292 /* get the rename from each predecessor and check if they are the same */
1293 bool needs_phi
= false;
1294 for (unsigned i
= 0; i
< preds
.size(); i
++) {
1295 ops
[i
] = read_variable(ctx
, val
, preds
[i
]);
1299 needs_phi
|= !(new_val
== ops
[i
]);
1303 /* the variable has been renamed differently in the predecessors: we need to insert a phi */
1304 aco_opcode opcode
= val
.is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1305 aco_ptr
<Instruction
> phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1306 new_val
= Temp
{ctx
.program
->allocateId(), val
.regClass()};
1307 phi
->definitions
[0] = Definition(new_val
);
1308 for (unsigned i
= 0; i
< preds
.size(); i
++) {
1309 phi
->operands
[i
] = Operand(ops
[i
]);
1310 phi
->operands
[i
].setFixed(ctx
.assignments
[ops
[i
].id()].reg
);
1311 if (ops
[i
].regClass() == new_val
.regClass())
1312 ctx
.affinities
[new_val
.id()] = ops
[i
].id();
1314 ctx
.assignments
.emplace_back();
1315 assert(ctx
.assignments
.size() == ctx
.program
->peekAllocationId());
1316 ctx
.phi_map
.emplace(new_val
.id(), phi_info
{phi
.get(), block
->index
});
1317 block
->instructions
.insert(block
->instructions
.begin(), std::move(phi
));
1321 if (new_val
!= val
) {
1322 ctx
.renames
[block
->index
][val
.id()] = new_val
;
1323 ctx
.orig_names
[new_val
.id()] = val
;
1328 void try_remove_trivial_phi(ra_ctx
& ctx
, Temp temp
)
1330 std::unordered_map
<unsigned, phi_info
>::iterator info
= ctx
.phi_map
.find(temp
.id());
1332 if (info
== ctx
.phi_map
.end() || !ctx
.sealed
[info
->second
.block_idx
])
1335 assert(info
->second
.block_idx
!= 0);
1336 Instruction
* phi
= info
->second
.phi
;
1338 Definition def
= phi
->definitions
[0];
1340 /* a phi node is trivial if all operands are the same as the definition of the phi */
1341 for (const Operand
& op
: phi
->operands
) {
1342 const Temp t
= op
.getTemp();
1343 if (t
== same
|| t
== def
.getTemp()) {
1344 assert(t
== same
|| op
.physReg() == def
.physReg());
1352 assert(same
!= Temp() || same
== def
.getTemp());
1354 /* reroute all uses to same and remove phi */
1355 std::vector
<Temp
> phi_users
;
1356 std::unordered_map
<unsigned, phi_info
>::iterator same_phi_info
= ctx
.phi_map
.find(same
.id());
1357 for (Instruction
* instr
: info
->second
.uses
) {
1358 assert(phi
!= instr
);
1359 /* recursively try to remove trivial phis */
1360 if (is_phi(instr
)) {
1361 /* ignore if the phi was already flagged trivial */
1362 if (instr
->definitions
.empty())
1365 if (instr
->definitions
[0].getTemp() != temp
)
1366 phi_users
.emplace_back(instr
->definitions
[0].getTemp());
1368 for (Operand
& op
: instr
->operands
) {
1369 if (op
.isTemp() && op
.tempId() == def
.tempId()) {
1371 if (same_phi_info
!= ctx
.phi_map
.end())
1372 same_phi_info
->second
.uses
.emplace(instr
);
1377 auto it
= ctx
.orig_names
.find(same
.id());
1378 unsigned orig_var
= it
!= ctx
.orig_names
.end() ? it
->second
.id() : same
.id();
1379 for (unsigned i
= 0; i
< ctx
.program
->blocks
.size(); i
++) {
1380 auto it
= ctx
.renames
[i
].find(orig_var
);
1381 if (it
!= ctx
.renames
[i
].end() && it
->second
== def
.getTemp())
1382 ctx
.renames
[i
][orig_var
] = same
;
1385 phi
->definitions
.clear(); /* this indicates that the phi can be removed */
1386 ctx
.phi_map
.erase(info
);
1387 for (Temp t
: phi_users
)
1388 try_remove_trivial_phi(ctx
, t
);
1393 } /* end namespace */
1396 void register_allocation(Program
*program
, std::vector
<TempSet
>& live_out_per_block
)
1398 ra_ctx
ctx(program
);
1399 std::vector
<std::vector
<Temp
>> phi_ressources
;
1400 std::unordered_map
<unsigned, unsigned> temp_to_phi_ressources
;
1402 for (std::vector
<Block
>::reverse_iterator it
= program
->blocks
.rbegin(); it
!= program
->blocks
.rend(); it
++) {
1405 /* first, compute the death points of all live vars within the block */
1406 TempSet
& live
= live_out_per_block
[block
.index
];
1408 std::vector
<aco_ptr
<Instruction
>>::reverse_iterator rit
;
1409 for (rit
= block
.instructions
.rbegin(); rit
!= block
.instructions
.rend(); ++rit
) {
1410 aco_ptr
<Instruction
>& instr
= *rit
;
1411 if (is_phi(instr
)) {
1412 live
.erase(instr
->definitions
[0].getTemp());
1413 if (instr
->definitions
[0].isKill() || instr
->definitions
[0].isFixed())
1415 /* collect information about affinity-related temporaries */
1416 std::vector
<Temp
> affinity_related
;
1417 /* affinity_related[0] is the last seen affinity-related temp */
1418 affinity_related
.emplace_back(instr
->definitions
[0].getTemp());
1419 affinity_related
.emplace_back(instr
->definitions
[0].getTemp());
1420 for (const Operand
& op
: instr
->operands
) {
1421 if (op
.isTemp() && op
.regClass() == instr
->definitions
[0].regClass()) {
1422 affinity_related
.emplace_back(op
.getTemp());
1423 temp_to_phi_ressources
[op
.tempId()] = phi_ressources
.size();
1426 phi_ressources
.emplace_back(std::move(affinity_related
));
1430 /* add vector affinities */
1431 if (instr
->opcode
== aco_opcode::p_create_vector
) {
1432 for (const Operand
& op
: instr
->operands
) {
1433 if (op
.isTemp() && op
.isFirstKill() && op
.getTemp().type() == instr
->definitions
[0].getTemp().type())
1434 ctx
.vectors
[op
.tempId()] = instr
.get();
1438 /* add operands to live variables */
1439 for (const Operand
& op
: instr
->operands
) {
1441 live
.emplace(op
.getTemp());
1444 /* erase definitions from live */
1445 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
1446 const Definition
& def
= instr
->definitions
[i
];
1449 live
.erase(def
.getTemp());
1450 /* mark last-seen phi operand */
1451 std::unordered_map
<unsigned, unsigned>::iterator it
= temp_to_phi_ressources
.find(def
.tempId());
1452 if (it
!= temp_to_phi_ressources
.end() && def
.regClass() == phi_ressources
[it
->second
][0].regClass()) {
1453 phi_ressources
[it
->second
][0] = def
.getTemp();
1454 /* try to coalesce phi affinities with parallelcopies */
1455 Operand op
= Operand();
1456 if (!def
.isFixed() && instr
->opcode
== aco_opcode::p_parallelcopy
)
1457 op
= instr
->operands
[i
];
1458 else if (instr
->opcode
== aco_opcode::v_mad_f32
&& !instr
->usesModifiers())
1459 op
= instr
->operands
[2];
1461 if (op
.isTemp() && op
.isFirstKillBeforeDef() && def
.regClass() == op
.regClass()) {
1462 phi_ressources
[it
->second
].emplace_back(op
.getTemp());
1463 temp_to_phi_ressources
[op
.tempId()] = it
->second
;
1469 /* create affinities */
1470 for (std::vector
<Temp
>& vec
: phi_ressources
) {
1471 assert(vec
.size() > 1);
1472 for (unsigned i
= 1; i
< vec
.size(); i
++)
1473 if (vec
[i
].id() != vec
[0].id())
1474 ctx
.affinities
[vec
[i
].id()] = vec
[0].id();
1477 /* state of register file after phis */
1478 std::vector
<std::bitset
<128>> sgpr_live_in(program
->blocks
.size());
1480 for (Block
& block
: program
->blocks
) {
1481 TempSet
& live
= live_out_per_block
[block
.index
];
1482 /* initialize register file */
1483 assert(block
.index
!= 0 || live
.empty());
1484 RegisterFile register_file
;
1485 ctx
.war_hint
.reset();
1487 for (Temp t
: live
) {
1488 Temp renamed
= handle_live_in(ctx
, t
, &block
);
1489 assignment
& var
= ctx
.assignments
[renamed
.id()];
1490 /* due to live-range splits, the live-in might be a phi, now */
1492 register_file
.fill(Definition(renamed
.id(), var
.reg
, var
.rc
));
1495 std::vector
<aco_ptr
<Instruction
>> instructions
;
1496 std::vector
<aco_ptr
<Instruction
>>::iterator it
;
1498 /* this is a slight adjustment from the paper as we already have phi nodes:
1499 * We consider them incomplete phis and only handle the definition. */
1501 /* handle fixed phi definitions */
1502 for (it
= block
.instructions
.begin(); it
!= block
.instructions
.end(); ++it
) {
1503 aco_ptr
<Instruction
>& phi
= *it
;
1506 Definition
& definition
= phi
->definitions
[0];
1507 if (!definition
.isFixed())
1510 /* check if a dead exec mask phi is needed */
1511 if (definition
.isKill()) {
1512 for (Operand
& op
: phi
->operands
) {
1513 assert(op
.isTemp());
1514 if (!ctx
.assignments
[op
.tempId()].assigned
||
1515 ctx
.assignments
[op
.tempId()].reg
!= exec
) {
1516 definition
.setKill(false);
1522 if (definition
.isKill())
1525 assert(definition
.physReg() == exec
);
1526 assert(!register_file
.test(definition
.physReg(), definition
.bytes()));
1527 register_file
.fill(definition
);
1528 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1531 /* look up the affinities */
1532 for (it
= block
.instructions
.begin(); it
!= block
.instructions
.end(); ++it
) {
1533 aco_ptr
<Instruction
>& phi
= *it
;
1536 Definition
& definition
= phi
->definitions
[0];
1537 if (definition
.isKill() || definition
.isFixed())
1540 if (ctx
.affinities
.find(definition
.tempId()) != ctx
.affinities
.end() &&
1541 ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].assigned
) {
1542 assert(ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].rc
== definition
.regClass());
1543 PhysReg reg
= ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].reg
;
1544 bool try_use_special_reg
= reg
== scc
|| reg
== exec
;
1545 if (try_use_special_reg
) {
1546 for (const Operand
& op
: phi
->operands
) {
1547 if (!(op
.isTemp() && ctx
.assignments
[op
.tempId()].assigned
&&
1548 ctx
.assignments
[op
.tempId()].reg
== reg
)) {
1549 try_use_special_reg
= false;
1553 if (!try_use_special_reg
)
1556 /* only assign if register is still free */
1557 if (!register_file
.test(reg
, definition
.bytes())) {
1558 definition
.setFixed(reg
);
1559 register_file
.fill(definition
);
1560 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1565 /* find registers for phis without affinity or where the register was blocked */
1566 for (it
= block
.instructions
.begin();it
!= block
.instructions
.end(); ++it
) {
1567 aco_ptr
<Instruction
>& phi
= *it
;
1571 Definition
& definition
= phi
->definitions
[0];
1572 if (definition
.isKill())
1575 if (!definition
.isFixed()) {
1576 std::vector
<std::pair
<Operand
, Definition
>> parallelcopy
;
1577 /* try to find a register that is used by at least one operand */
1578 for (const Operand
& op
: phi
->operands
) {
1579 if (!(op
.isTemp() && ctx
.assignments
[op
.tempId()].assigned
))
1581 PhysReg reg
= ctx
.assignments
[op
.tempId()].reg
;
1582 /* we tried this already on the previous loop */
1583 if (reg
== scc
|| reg
== exec
)
1585 if (get_reg_specified(ctx
, register_file
, definition
.regClass(), parallelcopy
, phi
, reg
)) {
1586 definition
.setFixed(reg
);
1590 if (!definition
.isFixed())
1591 definition
.setFixed(get_reg(ctx
, register_file
, definition
.getTemp(), parallelcopy
, phi
));
1593 /* process parallelcopy */
1594 for (std::pair
<Operand
, Definition
> pc
: parallelcopy
) {
1595 /* see if it's a copy from a different phi */
1596 //TODO: prefer moving some previous phis over live-ins
1597 //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)
1598 Instruction
*prev_phi
= NULL
;
1599 std::vector
<aco_ptr
<Instruction
>>::iterator phi_it
;
1600 for (phi_it
= instructions
.begin(); phi_it
!= instructions
.end(); ++phi_it
) {
1601 if ((*phi_it
)->definitions
[0].tempId() == pc
.first
.tempId())
1602 prev_phi
= phi_it
->get();
1605 while (!prev_phi
&& is_phi(*++phi_it
)) {
1606 if ((*phi_it
)->definitions
[0].tempId() == pc
.first
.tempId())
1607 prev_phi
= phi_it
->get();
1610 /* if so, just update that phi's register */
1611 register_file
.clear(prev_phi
->definitions
[0]);
1612 prev_phi
->definitions
[0].setFixed(pc
.second
.physReg());
1613 ctx
.assignments
[prev_phi
->definitions
[0].tempId()] = {pc
.second
.physReg(), pc
.second
.regClass()};
1614 register_file
.fill(prev_phi
->definitions
[0]);
1619 std::unordered_map
<unsigned, Temp
>::iterator orig_it
= ctx
.orig_names
.find(pc
.first
.tempId());
1620 Temp orig
= pc
.first
.getTemp();
1621 if (orig_it
!= ctx
.orig_names
.end())
1622 orig
= orig_it
->second
;
1624 ctx
.orig_names
[pc
.second
.tempId()] = orig
;
1625 ctx
.renames
[block
.index
][orig
.id()] = pc
.second
.getTemp();
1627 /* otherwise, this is a live-in and we need to create a new phi
1628 * to move it in this block's predecessors */
1629 aco_opcode opcode
= pc
.first
.getTemp().is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1630 std::vector
<unsigned>& preds
= pc
.first
.getTemp().is_linear() ? block
.linear_preds
: block
.logical_preds
;
1631 aco_ptr
<Instruction
> new_phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1632 new_phi
->definitions
[0] = pc
.second
;
1633 for (unsigned i
= 0; i
< preds
.size(); i
++)
1634 new_phi
->operands
[i
] = Operand(pc
.first
);
1635 instructions
.emplace_back(std::move(new_phi
));
1638 register_file
.fill(definition
);
1639 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1641 live
.emplace(definition
.getTemp());
1643 /* update phi affinities */
1644 for (const Operand
& op
: phi
->operands
) {
1645 if (op
.isTemp() && op
.regClass() == phi
->definitions
[0].regClass())
1646 ctx
.affinities
[op
.tempId()] = definition
.tempId();
1649 instructions
.emplace_back(std::move(*it
));
1652 /* fill in sgpr_live_in */
1653 for (unsigned i
= 0; i
<= ctx
.max_used_sgpr
; i
++)
1654 sgpr_live_in
[block
.index
][i
] = register_file
[i
];
1655 sgpr_live_in
[block
.index
][127] = register_file
[scc
.reg()];
1657 /* Handle all other instructions of the block */
1658 for (; it
!= block
.instructions
.end(); ++it
) {
1659 aco_ptr
<Instruction
>& instr
= *it
;
1661 /* parallelcopies from p_phi are inserted here which means
1662 * live ranges of killed operands end here as well */
1663 if (instr
->opcode
== aco_opcode::p_logical_end
) {
1664 /* no need to process this instruction any further */
1665 if (block
.logical_succs
.size() != 1) {
1666 instructions
.emplace_back(std::move(instr
));
1670 Block
& succ
= program
->blocks
[block
.logical_succs
[0]];
1672 for (; idx
< succ
.logical_preds
.size(); idx
++) {
1673 if (succ
.logical_preds
[idx
] == block
.index
)
1676 for (aco_ptr
<Instruction
>& phi
: succ
.instructions
) {
1677 if (phi
->opcode
== aco_opcode::p_phi
) {
1678 if (phi
->operands
[idx
].isTemp() &&
1679 phi
->operands
[idx
].getTemp().type() == RegType::sgpr
&&
1680 phi
->operands
[idx
].isFirstKillBeforeDef()) {
1681 Temp phi_op
= read_variable(ctx
, phi
->operands
[idx
].getTemp(), block
.index
);
1682 PhysReg reg
= ctx
.assignments
[phi_op
.id()].reg
;
1683 assert(register_file
[reg
] == phi_op
.id());
1684 register_file
[reg
] = 0;
1686 } else if (phi
->opcode
!= aco_opcode::p_linear_phi
) {
1690 instructions
.emplace_back(std::move(instr
));
1694 std::vector
<std::pair
<Operand
, Definition
>> parallelcopy
;
1696 assert(!is_phi(instr
));
1698 /* handle operands */
1699 for (unsigned i
= 0; i
< instr
->operands
.size(); ++i
) {
1700 auto& operand
= instr
->operands
[i
];
1701 if (!operand
.isTemp())
1704 /* rename operands */
1705 operand
.setTemp(read_variable(ctx
, operand
.getTemp(), block
.index
));
1706 assert(ctx
.assignments
[operand
.tempId()].assigned
);
1708 PhysReg reg
= ctx
.assignments
[operand
.tempId()].reg
;
1709 if (operand_can_use_reg(instr
, i
, reg
))
1710 operand
.setFixed(reg
);
1712 get_reg_for_operand(ctx
, register_file
, parallelcopy
, instr
, operand
);
1714 if (instr
->format
== Format::EXP
||
1715 (instr
->isVMEM() && i
== 3 && ctx
.program
->chip_class
== GFX6
) ||
1716 (instr
->format
== Format::DS
&& static_cast<DS_instruction
*>(instr
.get())->gds
)) {
1717 for (unsigned j
= 0; j
< operand
.size(); j
++)
1718 ctx
.war_hint
.set(operand
.physReg().reg() + j
);
1721 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.getTemp().id());
1722 if (phi
!= ctx
.phi_map
.end())
1723 phi
->second
.uses
.emplace(instr
.get());
1726 /* remove dead vars from register file */
1727 for (const Operand
& op
: instr
->operands
) {
1728 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1729 register_file
.clear(op
);
1732 /* try to optimize v_mad_f32 -> v_mac_f32 */
1733 if (instr
->opcode
== aco_opcode::v_mad_f32
&&
1734 instr
->operands
[2].isTemp() &&
1735 instr
->operands
[2].isKillBeforeDef() &&
1736 instr
->operands
[2].getTemp().type() == RegType::vgpr
&&
1737 instr
->operands
[1].isTemp() &&
1738 instr
->operands
[1].getTemp().type() == RegType::vgpr
&&
1739 !instr
->usesModifiers()) {
1740 instr
->format
= Format::VOP2
;
1741 instr
->opcode
= aco_opcode::v_mac_f32
;
1744 /* handle definitions which must have the same register as an operand */
1745 if (instr
->opcode
== aco_opcode::v_interp_p2_f32
||
1746 instr
->opcode
== aco_opcode::v_mac_f32
||
1747 instr
->opcode
== aco_opcode::v_writelane_b32
||
1748 instr
->opcode
== aco_opcode::v_writelane_b32_e64
) {
1749 instr
->definitions
[0].setFixed(instr
->operands
[2].physReg());
1750 } else if (instr
->opcode
== aco_opcode::s_addk_i32
||
1751 instr
->opcode
== aco_opcode::s_mulk_i32
) {
1752 instr
->definitions
[0].setFixed(instr
->operands
[0].physReg());
1753 } else if (instr
->format
== Format::MUBUF
&&
1754 instr
->definitions
.size() == 1 &&
1755 instr
->operands
.size() == 4) {
1756 instr
->definitions
[0].setFixed(instr
->operands
[3].physReg());
1757 } else if (instr
->format
== Format::MIMG
&&
1758 instr
->definitions
.size() == 1 &&
1759 instr
->operands
[1].regClass().type() == RegType::vgpr
) {
1760 instr
->definitions
[0].setFixed(instr
->operands
[1].physReg());
1763 ctx
.defs_done
.reset();
1765 /* handle fixed definitions first */
1766 for (unsigned i
= 0; i
< instr
->definitions
.size(); ++i
) {
1767 auto& definition
= instr
->definitions
[i
];
1768 if (!definition
.isFixed())
1771 adjust_max_used_regs(ctx
, definition
.regClass(), definition
.physReg());
1772 /* check if the target register is blocked */
1773 if (register_file
[definition
.physReg().reg()] != 0) {
1774 /* create parallelcopy pair to move blocking var */
1775 Temp tmp
= {register_file
[definition
.physReg()], ctx
.assignments
[register_file
[definition
.physReg()]].rc
};
1776 Operand pc_op
= Operand(tmp
);
1777 pc_op
.setFixed(ctx
.assignments
[register_file
[definition
.physReg().reg()]].reg
);
1778 RegClass rc
= pc_op
.regClass();
1779 tmp
= Temp
{program
->allocateId(), rc
};
1780 Definition pc_def
= Definition(tmp
);
1782 /* re-enable the killed operands, so that we don't move the blocking var there */
1783 for (const Operand
& op
: instr
->operands
) {
1784 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1785 register_file
.fill(op
);
1788 /* find a new register for the blocking variable */
1789 PhysReg reg
= get_reg(ctx
, register_file
, pc_op
.getTemp(), parallelcopy
, instr
);
1790 /* once again, disable killed operands */
1791 for (const Operand
& op
: instr
->operands
) {
1792 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1793 register_file
.clear(op
);
1795 for (unsigned k
= 0; k
< i
; k
++) {
1796 if (instr
->definitions
[k
].isTemp() && ctx
.defs_done
.test(k
) && !instr
->definitions
[k
].isKill())
1797 register_file
.fill(instr
->definitions
[k
]);
1799 pc_def
.setFixed(reg
);
1801 /* finish assignment of parallelcopy */
1802 ctx
.assignments
.emplace_back(reg
, pc_def
.regClass());
1803 assert(ctx
.assignments
.size() == ctx
.program
->peekAllocationId());
1804 parallelcopy
.emplace_back(pc_op
, pc_def
);
1806 /* add changes to reg_file */
1807 register_file
.clear(pc_op
);
1808 register_file
.fill(pc_def
);
1810 ctx
.defs_done
.set(i
);
1812 if (!definition
.isTemp())
1815 /* set live if it has a kill point */
1816 if (!definition
.isKill())
1817 live
.emplace(definition
.getTemp());
1819 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1820 register_file
.fill(definition
);
1823 /* handle all other definitions */
1824 for (unsigned i
= 0; i
< instr
->definitions
.size(); ++i
) {
1825 auto& definition
= instr
->definitions
[i
];
1827 if (definition
.isFixed() || !definition
.isTemp())
1831 if (definition
.hasHint() && register_file
[definition
.physReg().reg()] == 0)
1832 definition
.setFixed(definition
.physReg());
1833 else if (instr
->opcode
== aco_opcode::p_split_vector
) {
1834 PhysReg reg
= instr
->operands
[0].physReg();
1835 reg
.reg_b
+= i
* definition
.bytes();
1836 if (get_reg_specified(ctx
, register_file
, definition
.regClass(), parallelcopy
, instr
, reg
))
1837 definition
.setFixed(reg
);
1838 } else if (instr
->opcode
== aco_opcode::p_wqm
) {
1840 if (instr
->operands
[0].isKillBeforeDef() && instr
->operands
[0].getTemp().type() == definition
.getTemp().type()) {
1841 reg
= instr
->operands
[0].physReg();
1842 definition
.setFixed(reg
);
1843 assert(register_file
[reg
.reg()] == 0);
1845 } else if (instr
->opcode
== aco_opcode::p_extract_vector
) {
1847 if (instr
->operands
[0].isKillBeforeDef() &&
1848 instr
->operands
[0].getTemp().type() == definition
.getTemp().type()) {
1849 reg
= instr
->operands
[0].physReg();
1850 reg
.reg_b
+= definition
.bytes() * instr
->operands
[1].constantValue();
1851 assert(!register_file
.test(reg
, definition
.bytes()));
1852 definition
.setFixed(reg
);
1854 } else if (instr
->opcode
== aco_opcode::p_create_vector
) {
1855 PhysReg reg
= get_reg_create_vector(ctx
, register_file
, definition
.getTemp(),
1856 parallelcopy
, instr
);
1857 definition
.setFixed(reg
);
1860 if (!definition
.isFixed()) {
1861 Temp tmp
= definition
.getTemp();
1862 /* subdword instructions before RDNA write full registers */
1863 if (tmp
.regClass().is_subdword() &&
1864 !instr_can_access_subdword(instr
) &&
1865 ctx
.program
->chip_class
<= GFX9
) {
1866 assert(tmp
.bytes() <= 4);
1867 tmp
= Temp(definition
.tempId(), v1
);
1869 definition
.setFixed(get_reg(ctx
, register_file
, tmp
, parallelcopy
, instr
));
1872 assert(definition
.isFixed() && ((definition
.getTemp().type() == RegType::vgpr
&& definition
.physReg() >= 256) ||
1873 (definition
.getTemp().type() != RegType::vgpr
&& definition
.physReg() < 256)));
1874 ctx
.defs_done
.set(i
);
1876 /* set live if it has a kill point */
1877 if (!definition
.isKill())
1878 live
.emplace(definition
.getTemp());
1880 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1881 register_file
.fill(definition
);
1884 handle_pseudo(ctx
, register_file
, instr
.get());
1886 /* kill definitions and late-kill operands */
1887 for (const Definition
& def
: instr
->definitions
) {
1888 if (def
.isTemp() && def
.isKill())
1889 register_file
.clear(def
);
1891 for (const Operand
& op
: instr
->operands
) {
1892 if (op
.isTemp() && op
.isFirstKill() && op
.isLateKill())
1893 register_file
.clear(op
);
1896 /* emit parallelcopy */
1897 if (!parallelcopy
.empty()) {
1898 aco_ptr
<Pseudo_instruction
> pc
;
1899 pc
.reset(create_instruction
<Pseudo_instruction
>(aco_opcode::p_parallelcopy
, Format::PSEUDO
, parallelcopy
.size(), parallelcopy
.size()));
1900 bool temp_in_scc
= register_file
[scc
.reg()];
1901 bool sgpr_operands_alias_defs
= false;
1902 uint64_t sgpr_operands
[4] = {0, 0, 0, 0};
1903 for (unsigned i
= 0; i
< parallelcopy
.size(); i
++) {
1904 if (temp_in_scc
&& parallelcopy
[i
].first
.isTemp() && parallelcopy
[i
].first
.getTemp().type() == RegType::sgpr
) {
1905 if (!sgpr_operands_alias_defs
) {
1906 unsigned reg
= parallelcopy
[i
].first
.physReg().reg();
1907 unsigned size
= parallelcopy
[i
].first
.getTemp().size();
1908 sgpr_operands
[reg
/ 64u] |= ((1u << size
) - 1) << (reg
% 64u);
1910 reg
= parallelcopy
[i
].second
.physReg().reg();
1911 size
= parallelcopy
[i
].second
.getTemp().size();
1912 if (sgpr_operands
[reg
/ 64u] & ((1u << size
) - 1) << (reg
% 64u))
1913 sgpr_operands_alias_defs
= true;
1917 pc
->operands
[i
] = parallelcopy
[i
].first
;
1918 pc
->definitions
[i
] = parallelcopy
[i
].second
;
1919 assert(pc
->operands
[i
].size() == pc
->definitions
[i
].size());
1921 /* it might happen that the operand is already renamed. we have to restore the original name. */
1922 std::unordered_map
<unsigned, Temp
>::iterator it
= ctx
.orig_names
.find(pc
->operands
[i
].tempId());
1923 Temp orig
= it
!= ctx
.orig_names
.end() ? it
->second
: pc
->operands
[i
].getTemp();
1924 ctx
.orig_names
[pc
->definitions
[i
].tempId()] = orig
;
1925 ctx
.renames
[block
.index
][orig
.id()] = pc
->definitions
[i
].getTemp();
1927 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(pc
->operands
[i
].tempId());
1928 if (phi
!= ctx
.phi_map
.end())
1929 phi
->second
.uses
.emplace(pc
.get());
1932 if (temp_in_scc
&& sgpr_operands_alias_defs
) {
1933 /* disable definitions and re-enable operands */
1934 for (const Definition
& def
: instr
->definitions
) {
1935 if (def
.isTemp() && !def
.isKill())
1936 register_file
.clear(def
);
1938 for (const Operand
& op
: instr
->operands
) {
1939 if (op
.isTemp() && op
.isFirstKill())
1940 register_file
.block(op
.physReg(), op
.regClass());
1943 handle_pseudo(ctx
, register_file
, pc
.get());
1945 /* re-enable live vars */
1946 for (const Operand
& op
: instr
->operands
) {
1947 if (op
.isTemp() && op
.isFirstKill())
1948 register_file
.clear(op
);
1950 for (const Definition
& def
: instr
->definitions
) {
1951 if (def
.isTemp() && !def
.isKill())
1952 register_file
.fill(def
);
1955 pc
->tmp_in_scc
= false;
1958 instructions
.emplace_back(std::move(pc
));
1961 /* some instructions need VOP3 encoding if operand/definition is not assigned to VCC */
1962 bool instr_needs_vop3
= !instr
->isVOP3() &&
1963 ((instr
->format
== Format::VOPC
&& !(instr
->definitions
[0].physReg() == vcc
)) ||
1964 (instr
->opcode
== aco_opcode::v_cndmask_b32
&& !(instr
->operands
[2].physReg() == vcc
)) ||
1965 ((instr
->opcode
== aco_opcode::v_add_co_u32
||
1966 instr
->opcode
== aco_opcode::v_addc_co_u32
||
1967 instr
->opcode
== aco_opcode::v_sub_co_u32
||
1968 instr
->opcode
== aco_opcode::v_subb_co_u32
||
1969 instr
->opcode
== aco_opcode::v_subrev_co_u32
||
1970 instr
->opcode
== aco_opcode::v_subbrev_co_u32
) &&
1971 !(instr
->definitions
[1].physReg() == vcc
)) ||
1972 ((instr
->opcode
== aco_opcode::v_addc_co_u32
||
1973 instr
->opcode
== aco_opcode::v_subb_co_u32
||
1974 instr
->opcode
== aco_opcode::v_subbrev_co_u32
) &&
1975 !(instr
->operands
[2].physReg() == vcc
)));
1976 if (instr_needs_vop3
) {
1978 /* if the first operand is a literal, we have to move it to a reg */
1979 if (instr
->operands
.size() && instr
->operands
[0].isLiteral() && program
->chip_class
< GFX10
) {
1980 bool can_sgpr
= true;
1981 /* check, if we have to move to vgpr */
1982 for (const Operand
& op
: instr
->operands
) {
1983 if (op
.isTemp() && op
.getTemp().type() == RegType::sgpr
) {
1988 /* disable definitions and re-enable operands */
1989 for (const Definition
& def
: instr
->definitions
)
1990 register_file
.clear(def
);
1991 for (const Operand
& op
: instr
->operands
) {
1992 if (op
.isTemp() && op
.isFirstKill())
1993 register_file
.block(op
.physReg(), op
.regClass());
1995 Temp tmp
= {program
->allocateId(), can_sgpr
? s1
: v1
};
1996 ctx
.assignments
.emplace_back();
1997 PhysReg reg
= get_reg(ctx
, register_file
, tmp
, parallelcopy
, instr
);
1999 aco_ptr
<Instruction
> mov
;
2001 mov
.reset(create_instruction
<SOP1_instruction
>(aco_opcode::s_mov_b32
, Format::SOP1
, 1, 1));
2003 mov
.reset(create_instruction
<VOP1_instruction
>(aco_opcode::v_mov_b32
, Format::VOP1
, 1, 1));
2004 mov
->operands
[0] = instr
->operands
[0];
2005 mov
->definitions
[0] = Definition(tmp
);
2006 mov
->definitions
[0].setFixed(reg
);
2008 instr
->operands
[0] = Operand(tmp
);
2009 instr
->operands
[0].setFixed(reg
);
2010 instructions
.emplace_back(std::move(mov
));
2011 /* re-enable live vars */
2012 for (const Operand
& op
: instr
->operands
) {
2013 if (op
.isTemp() && op
.isFirstKill())
2014 register_file
.clear(op
);
2016 for (const Definition
& def
: instr
->definitions
) {
2017 if (def
.isTemp() && !def
.isKill())
2018 register_file
.fill(def
);
2022 /* change the instruction to VOP3 to enable an arbitrary register pair as dst */
2023 aco_ptr
<Instruction
> tmp
= std::move(instr
);
2024 Format format
= asVOP3(tmp
->format
);
2025 instr
.reset(create_instruction
<VOP3A_instruction
>(tmp
->opcode
, format
, tmp
->operands
.size(), tmp
->definitions
.size()));
2026 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
2027 Operand
& operand
= tmp
->operands
[i
];
2028 instr
->operands
[i
] = operand
;
2029 /* keep phi_map up to date */
2030 if (operand
.isTemp()) {
2031 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.tempId());
2032 if (phi
!= ctx
.phi_map
.end()) {
2033 phi
->second
.uses
.erase(tmp
.get());
2034 phi
->second
.uses
.emplace(instr
.get());
2038 std::copy(tmp
->definitions
.begin(), tmp
->definitions
.end(), instr
->definitions
.begin());
2040 instructions
.emplace_back(std::move(*it
));
2042 } /* end for Instr */
2044 block
.instructions
= std::move(instructions
);
2046 ctx
.filled
[block
.index
] = true;
2047 for (unsigned succ_idx
: block
.linear_succs
) {
2048 Block
& succ
= program
->blocks
[succ_idx
];
2049 /* seal block if all predecessors are filled */
2050 bool all_filled
= true;
2051 for (unsigned pred_idx
: succ
.linear_preds
) {
2052 if (!ctx
.filled
[pred_idx
]) {
2058 ctx
.sealed
[succ_idx
] = true;
2060 /* finish incomplete phis and check if they became trivial */
2061 for (Instruction
* phi
: ctx
.incomplete_phis
[succ_idx
]) {
2062 std::vector
<unsigned> preds
= phi
->definitions
[0].getTemp().is_linear() ? succ
.linear_preds
: succ
.logical_preds
;
2063 for (unsigned i
= 0; i
< phi
->operands
.size(); i
++) {
2064 phi
->operands
[i
].setTemp(read_variable(ctx
, phi
->operands
[i
].getTemp(), preds
[i
]));
2065 phi
->operands
[i
].setFixed(ctx
.assignments
[phi
->operands
[i
].tempId()].reg
);
2067 try_remove_trivial_phi(ctx
, phi
->definitions
[0].getTemp());
2069 /* complete the original phi nodes, but no need to check triviality */
2070 for (aco_ptr
<Instruction
>& instr
: succ
.instructions
) {
2073 std::vector
<unsigned> preds
= instr
->opcode
== aco_opcode::p_phi
? succ
.logical_preds
: succ
.linear_preds
;
2075 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
2076 auto& operand
= instr
->operands
[i
];
2077 if (!operand
.isTemp())
2079 operand
.setTemp(read_variable(ctx
, operand
.getTemp(), preds
[i
]));
2080 operand
.setFixed(ctx
.assignments
[operand
.tempId()].reg
);
2081 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.getTemp().id());
2082 if (phi
!= ctx
.phi_map
.end())
2083 phi
->second
.uses
.emplace(instr
.get());
2090 /* remove trivial phis */
2091 for (Block
& block
: program
->blocks
) {
2092 auto end
= std::find_if(block
.instructions
.begin(), block
.instructions
.end(),
2093 [](aco_ptr
<Instruction
>& instr
) { return !is_phi(instr
);});
2094 auto middle
= std::remove_if(block
.instructions
.begin(), end
,
2095 [](const aco_ptr
<Instruction
>& instr
) { return instr
->definitions
.empty();});
2096 block
.instructions
.erase(middle
, end
);
2099 /* find scc spill registers which may be needed for parallelcopies created by phis */
2100 for (Block
& block
: program
->blocks
) {
2101 if (block
.linear_preds
.size() <= 1)
2104 std::bitset
<128> regs
= sgpr_live_in
[block
.index
];
2108 /* choose a register */
2110 for (; reg
< ctx
.program
->max_reg_demand
.sgpr
&& regs
[reg
]; reg
++)
2112 assert(reg
< ctx
.program
->max_reg_demand
.sgpr
);
2113 adjust_max_used_regs(ctx
, s1
, reg
);
2115 /* update predecessors */
2116 for (unsigned& pred_index
: block
.linear_preds
) {
2117 Block
& pred
= program
->blocks
[pred_index
];
2118 pred
.scc_live_out
= true;
2119 pred
.scratch_sgpr
= PhysReg
{(uint16_t)reg
};
2123 /* num_gpr = rnd_up(max_used_gpr + 1) */
2124 program
->config
->num_vgprs
= align(ctx
.max_used_vgpr
+ 1, 4);
2125 if (program
->family
== CHIP_TONGA
|| program
->family
== CHIP_ICELAND
) /* workaround hardware bug */
2126 program
->config
->num_sgprs
= get_sgpr_alloc(program
, program
->sgpr_limit
);
2128 program
->config
->num_sgprs
= align(ctx
.max_used_sgpr
+ 1 + get_extra_sgprs(program
), 8);