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 */
396 /* best fit algorithm: find the smallest gap to fit in the variable */
399 if (rc
.type() == RegType::vgpr
&& (size
== 4 || size
== 8)) {
401 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
406 unsigned best_pos
= 0xFFFF;
407 unsigned gap_size
= 0xFFFF;
408 unsigned next_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 (next_pos
== 0xFFFF)
415 /* check if the variable fits */
416 if (next_pos
+ size
> current_reg
) {
421 /* check if the tested gap is smaller */
422 if (current_reg
- next_pos
< gap_size
) {
424 gap_size
= current_reg
- next_pos
;
430 if (next_pos
== 0xFFFF)
431 next_pos
= current_reg
;
435 if (next_pos
!= 0xFFFF &&
436 next_pos
+ size
<= ub
&&
437 ub
- next_pos
< gap_size
) {
439 gap_size
= ub
- next_pos
;
441 if (best_pos
!= 0xFFFF) {
442 adjust_max_used_regs(ctx
, rc
, best_pos
);
443 return {PhysReg
{best_pos
}, true};
449 unsigned reg_lo
= lb
;
450 unsigned reg_hi
= lb
+ size
- 1;
451 while (!found
&& reg_lo
+ size
<= ub
) {
452 if (reg_file
[reg_lo
] != 0) {
456 reg_hi
= reg_lo
+ size
- 1;
458 for (unsigned reg
= reg_lo
+ 1; found
&& reg
<= reg_hi
; reg
++) {
459 if (reg_file
[reg
] != 0 || ctx
.war_hint
[reg
])
463 adjust_max_used_regs(ctx
, rc
, reg_lo
);
464 return {PhysReg
{reg_lo
}, true};
473 /* collect variables from a register area and clear reg_file */
474 std::set
<std::pair
<unsigned, unsigned>> collect_vars(ra_ctx
& ctx
, RegisterFile
& reg_file
,
475 PhysReg reg
, unsigned size
)
477 std::set
<std::pair
<unsigned, unsigned>> vars
;
478 for (unsigned j
= reg
; j
< reg
+ size
; j
++) {
479 if (reg_file
.is_blocked(PhysReg
{j
}))
481 if (reg_file
[j
] == 0xF0000000) {
482 for (unsigned k
= 0; k
< 4; k
++) {
483 unsigned id
= reg_file
.subdword_regs
[j
][k
];
485 assignment
& var
= ctx
.assignments
[id
];
486 vars
.emplace(var
.rc
.bytes(), id
);
487 reg_file
.clear(var
.reg
, var
.rc
);
492 } else if (reg_file
[j
] != 0) {
493 unsigned id
= reg_file
[j
];
494 assignment
& var
= ctx
.assignments
[id
];
495 vars
.emplace(var
.rc
.bytes(), id
);
496 reg_file
.clear(var
.reg
, var
.rc
);
502 bool get_regs_for_copies(ra_ctx
& ctx
,
503 RegisterFile
& reg_file
,
504 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
505 const std::set
<std::pair
<unsigned, unsigned>> &vars
,
506 uint32_t lb
, uint32_t ub
,
507 aco_ptr
<Instruction
>& instr
,
512 /* variables are sorted from small sized to large */
513 /* NOTE: variables are also sorted by ID. this only affects a very small number of shaders slightly though. */
514 for (std::set
<std::pair
<unsigned, unsigned>>::const_reverse_iterator it
= vars
.rbegin(); it
!= vars
.rend(); ++it
) {
515 unsigned id
= it
->second
;
516 assignment
& var
= ctx
.assignments
[id
];
517 DefInfo info
= DefInfo(ctx
, ctx
.pseudo_dummy
, var
.rc
);
518 uint32_t size
= info
.size
;
520 /* check if this is a dead operand, then we can re-use the space from the definition */
521 bool is_dead_operand
= false;
522 for (unsigned i
= 0; !is_phi(instr
) && !is_dead_operand
&& (i
< instr
->operands
.size()); i
++) {
523 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isKillBeforeDef() && instr
->operands
[i
].tempId() == id
)
524 is_dead_operand
= true;
527 std::pair
<PhysReg
, bool> res
;
528 if (is_dead_operand
) {
529 if (instr
->opcode
== aco_opcode::p_create_vector
) {
530 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
531 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].tempId() == id
) {
532 PhysReg
reg(def_reg_lo
);
534 assert(!reg_file
.test(reg
, var
.rc
.bytes()));
540 info
.lb
= def_reg_lo
;
541 info
.ub
= def_reg_hi
+ 1;
542 res
= get_reg_simple(ctx
, reg_file
, info
);
546 info
.ub
= def_reg_lo
;
547 res
= get_reg_simple(ctx
, reg_file
, info
);
549 info
.lb
= (def_reg_hi
+ info
.stride
) & ~(info
.stride
- 1);
551 res
= get_reg_simple(ctx
, reg_file
, info
);
556 /* mark the area as blocked */
557 reg_file
.block(res
.first
, var
.rc
.bytes());
559 /* create parallelcopy pair (without definition id) */
560 Temp tmp
= Temp(id
, var
.rc
);
561 Operand pc_op
= Operand(tmp
);
562 pc_op
.setFixed(var
.reg
);
563 Definition pc_def
= Definition(res
.first
, pc_op
.regClass());
564 parallelcopies
.emplace_back(pc_op
, pc_def
);
568 unsigned best_pos
= lb
;
569 unsigned num_moves
= 0xFF;
570 unsigned num_vars
= 0;
572 /* we use a sliding window to find potential positions */
573 unsigned reg_lo
= lb
;
574 unsigned reg_hi
= lb
+ size
- 1;
575 unsigned stride
= var
.rc
.is_subdword() ? 1 : info
.stride
;
576 for (reg_lo
= lb
, reg_hi
= lb
+ size
- 1; reg_hi
< ub
; reg_lo
+= stride
, reg_hi
+= stride
) {
577 if (!is_dead_operand
&& ((reg_lo
>= def_reg_lo
&& reg_lo
<= def_reg_hi
) ||
578 (reg_hi
>= def_reg_lo
&& reg_hi
<= def_reg_hi
)))
581 /* second, check that we have at most k=num_moves elements in the window
582 * and no element is larger than the currently processed one */
585 unsigned last_var
= 0;
587 for (unsigned j
= reg_lo
; found
&& j
<= reg_hi
; j
++) {
588 if (reg_file
[j
] == 0 || reg_file
[j
] == last_var
)
591 if (reg_file
.is_blocked(PhysReg
{j
}) || k
> num_moves
) {
595 if (reg_file
[j
] == 0xF0000000) {
600 /* we cannot split live ranges of linear vgprs */
601 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6)) {
605 bool is_kill
= false;
606 for (const Operand
& op
: instr
->operands
) {
607 if (op
.isTemp() && op
.isKillBeforeDef() && op
.tempId() == reg_file
[j
]) {
612 if (!is_kill
&& ctx
.assignments
[reg_file
[j
]].rc
.size() >= size
) {
617 k
+= ctx
.assignments
[reg_file
[j
]].rc
.size();
618 last_var
= reg_file
[j
];
620 if (k
> num_moves
|| (k
== num_moves
&& n
<= num_vars
)) {
633 /* FIXME: we messed up and couldn't find space for the variables to be copied */
634 if (num_moves
== 0xFF)
638 reg_hi
= best_pos
+ size
- 1;
640 /* collect variables and block reg file */
641 std::set
<std::pair
<unsigned, unsigned>> new_vars
= collect_vars(ctx
, reg_file
, PhysReg
{reg_lo
}, size
);
643 /* mark the area as blocked */
644 reg_file
.block(PhysReg
{reg_lo
}, size
* 4);
646 if (!get_regs_for_copies(ctx
, reg_file
, parallelcopies
, new_vars
, lb
, ub
, instr
, def_reg_lo
, def_reg_hi
))
649 adjust_max_used_regs(ctx
, var
.rc
, reg_lo
);
651 /* create parallelcopy pair (without definition id) */
652 Temp tmp
= Temp(id
, var
.rc
);
653 Operand pc_op
= Operand(tmp
);
654 pc_op
.setFixed(var
.reg
);
655 Definition pc_def
= Definition(PhysReg
{reg_lo
}, pc_op
.regClass());
656 parallelcopies
.emplace_back(pc_op
, pc_def
);
663 std::pair
<PhysReg
, bool> get_reg_impl(ra_ctx
& ctx
,
664 RegisterFile
& reg_file
,
665 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
667 aco_ptr
<Instruction
>& instr
)
669 uint32_t lb
= info
.lb
;
670 uint32_t ub
= info
.ub
;
671 uint32_t size
= info
.size
;
672 uint32_t stride
= info
.stride
;
673 RegClass rc
= info
.rc
;
675 /* check how many free regs we have */
676 unsigned regs_free
= reg_file
.count_zero(PhysReg
{lb
}, ub
-lb
);
678 /* mark and count killed operands */
679 unsigned killed_ops
= 0;
680 for (unsigned j
= 0; !is_phi(instr
) && j
< instr
->operands
.size(); j
++) {
681 if (instr
->operands
[j
].isTemp() &&
682 instr
->operands
[j
].isFirstKillBeforeDef() &&
683 instr
->operands
[j
].physReg() >= lb
&&
684 instr
->operands
[j
].physReg() < ub
) {
685 assert(instr
->operands
[j
].isFixed());
686 assert(!reg_file
.test(instr
->operands
[j
].physReg(), instr
->operands
[j
].bytes()));
687 reg_file
.block(instr
->operands
[j
].physReg(), instr
->operands
[j
].bytes());
688 killed_ops
+= instr
->operands
[j
].getTemp().size();
692 assert(regs_free
>= size
);
693 /* we might have to move dead operands to dst in order to make space */
694 unsigned op_moves
= 0;
696 if (size
> (regs_free
- killed_ops
))
697 op_moves
= size
- (regs_free
- killed_ops
);
699 /* find the best position to place the definition */
700 unsigned best_pos
= lb
;
701 unsigned num_moves
= 0xFF;
702 unsigned num_vars
= 0;
704 /* we use a sliding window to check potential positions */
705 unsigned reg_lo
= lb
;
706 unsigned reg_hi
= lb
+ size
- 1;
707 for (reg_lo
= lb
, reg_hi
= lb
+ size
- 1; reg_hi
< ub
; reg_lo
+= stride
, reg_hi
+= stride
) {
708 /* first check the edges: this is what we have to fix to allow for num_moves > size */
709 if (reg_lo
> lb
&& reg_file
[reg_lo
] != 0 && reg_file
[reg_lo
] == reg_file
[reg_lo
- 1])
711 if (reg_hi
< ub
- 1 && reg_file
[reg_hi
] != 0 && reg_file
[reg_hi
] == reg_file
[reg_hi
+ 1])
714 /* second, check that we have at most k=num_moves elements in the window
715 * and no element is larger than the currently processed one */
716 unsigned k
= op_moves
;
718 unsigned remaining_op_moves
= op_moves
;
719 unsigned last_var
= 0;
721 bool aligned
= rc
== RegClass::v4
&& reg_lo
% 4 == 0;
722 for (unsigned j
= reg_lo
; found
&& j
<= reg_hi
; j
++) {
723 if (reg_file
[j
] == 0 || reg_file
[j
] == last_var
)
726 /* dead operands effectively reduce the number of estimated moves */
727 if (reg_file
.is_blocked(PhysReg
{j
})) {
728 if (remaining_op_moves
) {
730 remaining_op_moves
--;
735 if (reg_file
[j
] == 0xF0000000) {
741 if (ctx
.assignments
[reg_file
[j
]].rc
.size() >= size
) {
746 /* we cannot split live ranges of linear vgprs */
747 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6)) {
752 k
+= ctx
.assignments
[reg_file
[j
]].rc
.size();
754 last_var
= reg_file
[j
];
757 if (!found
|| k
> num_moves
)
759 if (k
== num_moves
&& n
< num_vars
)
761 if (!aligned
&& k
== num_moves
&& n
== num_vars
)
771 if (num_moves
== 0xFF) {
772 /* remove killed operands from reg_file once again */
773 for (unsigned i
= 0; !is_phi(instr
) && i
< instr
->operands
.size(); i
++) {
774 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isFirstKillBeforeDef())
775 reg_file
.clear(instr
->operands
[i
]);
777 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
778 Definition def
= instr
->definitions
[i
];
779 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
785 RegisterFile register_file
= reg_file
;
787 /* now, we figured the placement for our definition */
788 std::set
<std::pair
<unsigned, unsigned>> vars
= collect_vars(ctx
, reg_file
, PhysReg
{best_pos
}, size
);
790 if (instr
->opcode
== aco_opcode::p_create_vector
) {
791 /* move killed operands which aren't yet at the correct position */
792 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].size(), i
++) {
793 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].isFirstKillBeforeDef() &&
794 instr
->operands
[i
].getTemp().type() == rc
.type()) {
796 if (instr
->operands
[i
].physReg() != best_pos
+ offset
) {
797 vars
.emplace(instr
->operands
[i
].bytes(), instr
->operands
[i
].tempId());
798 reg_file
.clear(instr
->operands
[i
]);
800 reg_file
.fill(instr
->operands
[i
]);
805 /* re-enable the killed operands */
806 for (unsigned j
= 0; !is_phi(instr
) && j
< instr
->operands
.size(); j
++) {
807 if (instr
->operands
[j
].isTemp() && instr
->operands
[j
].isFirstKill())
808 reg_file
.fill(instr
->operands
[j
]);
812 std::vector
<std::pair
<Operand
, Definition
>> pc
;
813 if (!get_regs_for_copies(ctx
, reg_file
, pc
, vars
, lb
, ub
, instr
, best_pos
, best_pos
+ size
- 1)) {
814 reg_file
= std::move(register_file
);
815 /* remove killed operands from reg_file once again */
816 if (!is_phi(instr
)) {
817 for (const Operand
& op
: instr
->operands
) {
818 if (op
.isTemp() && op
.isFirstKillBeforeDef())
822 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
823 Definition
& def
= instr
->definitions
[i
];
824 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
830 parallelcopies
.insert(parallelcopies
.end(), pc
.begin(), pc
.end());
832 /* we set the definition regs == 0. the actual caller is responsible for correct setting */
833 reg_file
.clear(PhysReg
{best_pos
}, rc
);
835 update_renames(ctx
, reg_file
, parallelcopies
, instr
);
837 /* remove killed operands from reg_file once again */
838 for (unsigned i
= 0; !is_phi(instr
) && i
< instr
->operands
.size(); i
++) {
839 if (!instr
->operands
[i
].isTemp() || !instr
->operands
[i
].isFixed())
841 assert(!instr
->operands
[i
].isUndefined());
842 if (instr
->operands
[i
].isFirstKillBeforeDef())
843 reg_file
.clear(instr
->operands
[i
]);
845 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
846 Definition def
= instr
->definitions
[i
];
847 if (def
.isTemp() && def
.isFixed() && ctx
.defs_done
.test(i
))
851 adjust_max_used_regs(ctx
, rc
, best_pos
);
852 return {PhysReg
{best_pos
}, true};
855 bool get_reg_specified(ra_ctx
& ctx
,
856 RegisterFile
& reg_file
,
858 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
859 aco_ptr
<Instruction
>& instr
,
862 if (rc
.is_subdword() && reg
.byte() && !instr_can_access_subdword(instr
))
865 uint32_t size
= rc
.size();
869 if (rc
.type() == RegType::vgpr
) {
871 ub
= 256 + ctx
.program
->max_reg_demand
.vgpr
;
877 if (reg
% stride
!= 0)
880 ub
= ctx
.program
->max_reg_demand
.sgpr
;
883 uint32_t reg_lo
= reg
.reg();
884 uint32_t reg_hi
= reg
+ (size
- 1);
886 if (reg_lo
< lb
|| reg_hi
>= ub
|| reg_lo
> reg_hi
)
889 if (reg_file
.test(reg
, rc
.bytes()))
892 adjust_max_used_regs(ctx
, rc
, reg_lo
);
896 PhysReg
get_reg(ra_ctx
& ctx
,
897 RegisterFile
& reg_file
,
899 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
900 aco_ptr
<Instruction
>& instr
)
902 if (ctx
.affinities
.find(temp
.id()) != ctx
.affinities
.end() &&
903 ctx
.assignments
[ctx
.affinities
[temp
.id()]].assigned
) {
904 PhysReg reg
= ctx
.assignments
[ctx
.affinities
[temp
.id()]].reg
;
905 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
909 if (ctx
.vectors
.find(temp
.id()) != ctx
.vectors
.end()) {
910 Instruction
* vec
= ctx
.vectors
[temp
.id()];
911 unsigned byte_offset
= 0;
912 for (const Operand
& op
: vec
->operands
) {
913 if (op
.isTemp() && op
.tempId() == temp
.id())
916 byte_offset
+= op
.bytes();
919 for (const Operand
& op
: vec
->operands
) {
921 op
.tempId() != temp
.id() &&
922 op
.getTemp().type() == temp
.type() &&
923 ctx
.assignments
[op
.tempId()].assigned
) {
924 PhysReg reg
= ctx
.assignments
[op
.tempId()].reg
;
925 reg
.reg_b
+= (byte_offset
- k
);
926 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
932 DefInfo
info(ctx
, ctx
.pseudo_dummy
, vec
->definitions
[0].regClass());
933 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
934 PhysReg reg
= res
.first
;
936 reg
.reg_b
+= byte_offset
;
937 /* make sure to only use byte offset if the instruction supports it */
938 if (get_reg_specified(ctx
, reg_file
, temp
.regClass(), parallelcopies
, instr
, reg
))
943 DefInfo
info(ctx
, instr
, temp
.regClass());
945 /* try to find space without live-range splits */
946 std::pair
<PhysReg
, bool> res
= get_reg_simple(ctx
, reg_file
, info
);
951 /* try to find space with live-range splits */
952 res
= get_reg_impl(ctx
, reg_file
, parallelcopies
, info
, instr
);
957 /* try using more registers */
959 /* We should only fail here because keeping under the limit would require
961 assert(reg_file
.count_zero(PhysReg
{info
.lb
}, info
.ub
-info
.lb
) >= info
.size
);
963 uint16_t max_addressible_sgpr
= ctx
.program
->sgpr_limit
;
964 uint16_t max_addressible_vgpr
= ctx
.program
->vgpr_limit
;
965 if (info
.rc
.type() == RegType::vgpr
&& ctx
.program
->max_reg_demand
.vgpr
< max_addressible_vgpr
) {
966 update_vgpr_sgpr_demand(ctx
.program
, RegisterDemand(ctx
.program
->max_reg_demand
.vgpr
+ 1, ctx
.program
->max_reg_demand
.sgpr
));
967 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
968 } else if (info
.rc
.type() == RegType::sgpr
&& ctx
.program
->max_reg_demand
.sgpr
< max_addressible_sgpr
) {
969 update_vgpr_sgpr_demand(ctx
.program
, RegisterDemand(ctx
.program
->max_reg_demand
.vgpr
, ctx
.program
->max_reg_demand
.sgpr
+ 1));
970 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
973 //FIXME: if nothing helps, shift-rotate the registers to make space
975 unreachable("did not find a register");
978 PhysReg
get_reg_create_vector(ra_ctx
& ctx
,
979 RegisterFile
& reg_file
,
981 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopies
,
982 aco_ptr
<Instruction
>& instr
)
984 RegClass rc
= temp
.regClass();
985 /* create_vector instructions have different costs w.r.t. register coalescing */
986 uint32_t size
= rc
.size();
987 uint32_t bytes
= rc
.bytes();
990 if (rc
.type() == RegType::vgpr
) {
992 ub
= 256 + ctx
.program
->max_reg_demand
.vgpr
;
995 ub
= ctx
.program
->max_reg_demand
.sgpr
;
1002 //TODO: improve p_create_vector for sub-dword vectors
1004 unsigned best_pos
= -1;
1005 unsigned num_moves
= 0xFF;
1006 bool best_war_hint
= true;
1008 /* test for each operand which definition placement causes the least shuffle instructions */
1009 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
1010 // TODO: think about, if we can alias live operands on the same register
1011 if (!instr
->operands
[i
].isTemp() || !instr
->operands
[i
].isKillBeforeDef() || instr
->operands
[i
].getTemp().type() != rc
.type())
1014 if (offset
> instr
->operands
[i
].physReg().reg_b
)
1017 unsigned reg_lo
= instr
->operands
[i
].physReg().reg_b
- offset
;
1021 unsigned reg_hi
= reg_lo
+ size
- 1;
1024 /* no need to check multiple times */
1025 if (reg_lo
== best_pos
)
1029 // TODO: this can be improved */
1030 if (reg_lo
< lb
|| reg_hi
>= ub
|| reg_lo
% stride
!= 0)
1032 if (reg_lo
> lb
&& reg_file
[reg_lo
] != 0 && reg_file
[reg_lo
] == reg_file
[reg_lo
- 1])
1034 if (reg_hi
< ub
- 1 && reg_file
[reg_hi
] != 0 && reg_file
[reg_hi
] == reg_file
[reg_hi
+ 1])
1037 /* count variables to be moved and check war_hint */
1038 bool war_hint
= false;
1039 bool linear_vgpr
= false;
1040 for (unsigned j
= reg_lo
; j
<= reg_hi
&& !linear_vgpr
; j
++) {
1041 if (reg_file
[j
] != 0) {
1042 if (reg_file
[j
] == 0xF0000000) {
1045 unsigned bytes_left
= bytes
- (j
- reg_lo
) * 4;
1046 for (unsigned k
= 0; k
< MIN2(bytes_left
, 4); k
++, reg
.reg_b
++)
1047 k
+= reg_file
.test(reg
, 1);
1050 /* we cannot split live ranges of linear vgprs */
1051 if (ctx
.assignments
[reg_file
[j
]].rc
& (1 << 6))
1055 war_hint
|= ctx
.war_hint
[j
];
1057 if (linear_vgpr
|| (war_hint
&& !best_war_hint
))
1060 /* count operands in wrong positions */
1061 for (unsigned j
= 0, offset
= 0; j
< instr
->operands
.size(); offset
+= instr
->operands
[j
].bytes(), j
++) {
1063 !instr
->operands
[j
].isTemp() ||
1064 instr
->operands
[j
].getTemp().type() != rc
.type())
1066 if (instr
->operands
[j
].physReg().reg_b
!= reg_lo
* 4 + offset
)
1067 k
+= instr
->operands
[j
].bytes();
1069 bool aligned
= rc
== RegClass::v4
&& reg_lo
% 4 == 0;
1070 if (k
> num_moves
|| (!aligned
&& k
== num_moves
))
1075 best_war_hint
= war_hint
;
1078 if (num_moves
>= bytes
)
1079 return get_reg(ctx
, reg_file
, temp
, parallelcopies
, instr
);
1081 /* collect variables to be moved */
1082 std::set
<std::pair
<unsigned, unsigned>> vars
= collect_vars(ctx
, reg_file
, PhysReg
{best_pos
}, size
);
1084 /* move killed operands which aren't yet at the correct position */
1085 uint64_t moved_operand_mask
= 0;
1086 for (unsigned i
= 0, offset
= 0; i
< instr
->operands
.size(); offset
+= instr
->operands
[i
].bytes(), i
++) {
1087 if (instr
->operands
[i
].isTemp() &&
1088 instr
->operands
[i
].isFirstKillBeforeDef() &&
1089 instr
->operands
[i
].getTemp().type() == rc
.type() &&
1090 instr
->operands
[i
].physReg().reg_b
!= best_pos
* 4 + offset
) {
1091 vars
.emplace(instr
->operands
[i
].bytes(), instr
->operands
[i
].tempId());
1092 moved_operand_mask
|= (uint64_t)1 << i
;
1096 ASSERTED
bool success
= false;
1097 success
= get_regs_for_copies(ctx
, reg_file
, parallelcopies
, vars
, lb
, ub
, instr
, best_pos
, best_pos
+ size
- 1);
1100 update_renames(ctx
, reg_file
, parallelcopies
, instr
);
1101 adjust_max_used_regs(ctx
, rc
, best_pos
);
1103 while (moved_operand_mask
) {
1104 unsigned i
= u_bit_scan64(&moved_operand_mask
);
1105 assert(instr
->operands
[i
].isFirstKillBeforeDef());
1106 reg_file
.clear(instr
->operands
[i
]);
1109 return PhysReg
{best_pos
};
1112 void handle_pseudo(ra_ctx
& ctx
,
1113 const RegisterFile
& reg_file
,
1116 if (instr
->format
!= Format::PSEUDO
)
1119 /* all instructions which use handle_operands() need this information */
1120 switch (instr
->opcode
) {
1121 case aco_opcode::p_extract_vector
:
1122 case aco_opcode::p_create_vector
:
1123 case aco_opcode::p_split_vector
:
1124 case aco_opcode::p_parallelcopy
:
1125 case aco_opcode::p_wqm
:
1131 /* if all definitions are vgpr, no need to care for SCC */
1132 bool writes_sgpr
= false;
1133 for (Definition
& def
: instr
->definitions
) {
1134 if (def
.getTemp().type() == RegType::sgpr
) {
1139 /* if all operands are constant, no need to care either */
1140 bool reads_sgpr
= false;
1141 for (Operand
& op
: instr
->operands
) {
1142 if (op
.isTemp() && op
.getTemp().type() == RegType::sgpr
) {
1147 if (!(writes_sgpr
&& reads_sgpr
))
1150 Pseudo_instruction
*pi
= (Pseudo_instruction
*)instr
;
1151 if (reg_file
[scc
.reg()]) {
1152 pi
->tmp_in_scc
= true;
1154 int reg
= ctx
.max_used_sgpr
;
1155 for (; reg
>= 0 && reg_file
[reg
]; reg
--)
1158 reg
= ctx
.max_used_sgpr
+ 1;
1159 for (; reg
< ctx
.program
->max_reg_demand
.sgpr
&& reg_file
[reg
]; reg
++)
1161 assert(reg
< ctx
.program
->max_reg_demand
.sgpr
);
1164 adjust_max_used_regs(ctx
, s1
, reg
);
1165 pi
->scratch_sgpr
= PhysReg
{(unsigned)reg
};
1167 pi
->tmp_in_scc
= false;
1171 bool operand_can_use_reg(aco_ptr
<Instruction
>& instr
, unsigned idx
, PhysReg reg
)
1173 if (instr
->operands
[idx
].isFixed())
1174 return instr
->operands
[idx
].physReg() == reg
;
1176 if (!instr_can_access_subdword(instr
) && reg
.byte())
1179 switch (instr
->format
) {
1181 return reg
!= scc
&&
1183 (reg
!= m0
|| idx
== 1 || idx
== 3) && /* offset can be m0 */
1184 (reg
!= vcc
|| (instr
->definitions
.empty() && idx
== 2)); /* sdata can be vcc */
1186 // TODO: there are more instructions with restrictions on registers
1191 void get_reg_for_operand(ra_ctx
& ctx
, RegisterFile
& register_file
,
1192 std::vector
<std::pair
<Operand
, Definition
>>& parallelcopy
,
1193 aco_ptr
<Instruction
>& instr
, Operand
& operand
)
1195 /* check if the operand is fixed */
1197 if (operand
.isFixed()) {
1198 assert(operand
.physReg() != ctx
.assignments
[operand
.tempId()].reg
);
1200 /* check if target reg is blocked, and move away the blocking var */
1201 if (register_file
[operand
.physReg().reg()]) {
1202 assert(register_file
[operand
.physReg()] != 0xF0000000);
1203 uint32_t blocking_id
= register_file
[operand
.physReg().reg()];
1204 RegClass rc
= ctx
.assignments
[blocking_id
].rc
;
1205 Operand pc_op
= Operand(Temp
{blocking_id
, rc
});
1206 pc_op
.setFixed(operand
.physReg());
1209 PhysReg reg
= get_reg(ctx
, register_file
, pc_op
.getTemp(), parallelcopy
, ctx
.pseudo_dummy
);
1210 Definition pc_def
= Definition(PhysReg
{reg
}, pc_op
.regClass());
1211 register_file
.clear(pc_op
);
1212 parallelcopy
.emplace_back(pc_op
, pc_def
);
1214 dst
= operand
.physReg();
1217 dst
= get_reg(ctx
, register_file
, operand
.getTemp(), parallelcopy
, instr
);
1220 Operand pc_op
= operand
;
1221 pc_op
.setFixed(ctx
.assignments
[operand
.tempId()].reg
);
1222 Definition pc_def
= Definition(dst
, pc_op
.regClass());
1223 register_file
.clear(pc_op
);
1224 parallelcopy
.emplace_back(pc_op
, pc_def
);
1225 update_renames(ctx
, register_file
, parallelcopy
, instr
);
1228 Temp
read_variable(ra_ctx
& ctx
, Temp val
, unsigned block_idx
)
1230 std::unordered_map
<unsigned, Temp
>::iterator it
= ctx
.renames
[block_idx
].find(val
.id());
1231 if (it
== ctx
.renames
[block_idx
].end())
1237 Temp
handle_live_in(ra_ctx
& ctx
, Temp val
, Block
* block
)
1239 std::vector
<unsigned>& preds
= val
.is_linear() ? block
->linear_preds
: block
->logical_preds
;
1240 if (preds
.size() == 0 || val
.regClass() == val
.regClass().as_linear())
1243 assert(preds
.size() > 0);
1246 if (!ctx
.sealed
[block
->index
]) {
1247 /* consider rename from already processed predecessor */
1248 Temp tmp
= read_variable(ctx
, val
, preds
[0]);
1250 /* if the block is not sealed yet, we create an incomplete phi (which might later get removed again) */
1251 new_val
= Temp
{ctx
.program
->allocateId(), val
.regClass()};
1252 ctx
.assignments
.emplace_back();
1253 aco_opcode opcode
= val
.is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1254 aco_ptr
<Instruction
> phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1255 phi
->definitions
[0] = Definition(new_val
);
1256 for (unsigned i
= 0; i
< preds
.size(); i
++)
1257 phi
->operands
[i
] = Operand(val
);
1258 if (tmp
.regClass() == new_val
.regClass())
1259 ctx
.affinities
[new_val
.id()] = tmp
.id();
1261 ctx
.phi_map
.emplace(new_val
.id(), phi_info
{phi
.get(), block
->index
});
1262 ctx
.incomplete_phis
[block
->index
].emplace_back(phi
.get());
1263 block
->instructions
.insert(block
->instructions
.begin(), std::move(phi
));
1265 } else if (preds
.size() == 1) {
1266 /* if the block has only one predecessor, just look there for the name */
1267 new_val
= read_variable(ctx
, val
, preds
[0]);
1269 /* there are multiple predecessors and the block is sealed */
1270 Temp ops
[preds
.size()];
1272 /* get the rename from each predecessor and check if they are the same */
1273 bool needs_phi
= false;
1274 for (unsigned i
= 0; i
< preds
.size(); i
++) {
1275 ops
[i
] = read_variable(ctx
, val
, preds
[i
]);
1279 needs_phi
|= !(new_val
== ops
[i
]);
1283 /* the variable has been renamed differently in the predecessors: we need to insert a phi */
1284 aco_opcode opcode
= val
.is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1285 aco_ptr
<Instruction
> phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1286 new_val
= Temp
{ctx
.program
->allocateId(), val
.regClass()};
1287 phi
->definitions
[0] = Definition(new_val
);
1288 for (unsigned i
= 0; i
< preds
.size(); i
++) {
1289 phi
->operands
[i
] = Operand(ops
[i
]);
1290 phi
->operands
[i
].setFixed(ctx
.assignments
[ops
[i
].id()].reg
);
1291 if (ops
[i
].regClass() == new_val
.regClass())
1292 ctx
.affinities
[new_val
.id()] = ops
[i
].id();
1294 ctx
.assignments
.emplace_back();
1295 assert(ctx
.assignments
.size() == ctx
.program
->peekAllocationId());
1296 ctx
.phi_map
.emplace(new_val
.id(), phi_info
{phi
.get(), block
->index
});
1297 block
->instructions
.insert(block
->instructions
.begin(), std::move(phi
));
1301 if (new_val
!= val
) {
1302 ctx
.renames
[block
->index
][val
.id()] = new_val
;
1303 ctx
.orig_names
[new_val
.id()] = val
;
1308 void try_remove_trivial_phi(ra_ctx
& ctx
, Temp temp
)
1310 std::unordered_map
<unsigned, phi_info
>::iterator info
= ctx
.phi_map
.find(temp
.id());
1312 if (info
== ctx
.phi_map
.end() || !ctx
.sealed
[info
->second
.block_idx
])
1315 assert(info
->second
.block_idx
!= 0);
1316 Instruction
* phi
= info
->second
.phi
;
1318 Definition def
= phi
->definitions
[0];
1320 /* a phi node is trivial if all operands are the same as the definition of the phi */
1321 for (const Operand
& op
: phi
->operands
) {
1322 const Temp t
= op
.getTemp();
1323 if (t
== same
|| t
== def
.getTemp()) {
1324 assert(t
== same
|| op
.physReg() == def
.physReg());
1332 assert(same
!= Temp() || same
== def
.getTemp());
1334 /* reroute all uses to same and remove phi */
1335 std::vector
<Temp
> phi_users
;
1336 std::unordered_map
<unsigned, phi_info
>::iterator same_phi_info
= ctx
.phi_map
.find(same
.id());
1337 for (Instruction
* instr
: info
->second
.uses
) {
1338 assert(phi
!= instr
);
1339 /* recursively try to remove trivial phis */
1340 if (is_phi(instr
)) {
1341 /* ignore if the phi was already flagged trivial */
1342 if (instr
->definitions
.empty())
1345 if (instr
->definitions
[0].getTemp() != temp
)
1346 phi_users
.emplace_back(instr
->definitions
[0].getTemp());
1348 for (Operand
& op
: instr
->operands
) {
1349 if (op
.isTemp() && op
.tempId() == def
.tempId()) {
1351 if (same_phi_info
!= ctx
.phi_map
.end())
1352 same_phi_info
->second
.uses
.emplace(instr
);
1357 auto it
= ctx
.orig_names
.find(same
.id());
1358 unsigned orig_var
= it
!= ctx
.orig_names
.end() ? it
->second
.id() : same
.id();
1359 for (unsigned i
= 0; i
< ctx
.program
->blocks
.size(); i
++) {
1360 auto it
= ctx
.renames
[i
].find(orig_var
);
1361 if (it
!= ctx
.renames
[i
].end() && it
->second
== def
.getTemp())
1362 ctx
.renames
[i
][orig_var
] = same
;
1365 phi
->definitions
.clear(); /* this indicates that the phi can be removed */
1366 ctx
.phi_map
.erase(info
);
1367 for (Temp t
: phi_users
)
1368 try_remove_trivial_phi(ctx
, t
);
1373 } /* end namespace */
1376 void register_allocation(Program
*program
, std::vector
<TempSet
>& live_out_per_block
)
1378 ra_ctx
ctx(program
);
1379 std::vector
<std::vector
<Temp
>> phi_ressources
;
1380 std::unordered_map
<unsigned, unsigned> temp_to_phi_ressources
;
1382 for (std::vector
<Block
>::reverse_iterator it
= program
->blocks
.rbegin(); it
!= program
->blocks
.rend(); it
++) {
1385 /* first, compute the death points of all live vars within the block */
1386 TempSet
& live
= live_out_per_block
[block
.index
];
1388 std::vector
<aco_ptr
<Instruction
>>::reverse_iterator rit
;
1389 for (rit
= block
.instructions
.rbegin(); rit
!= block
.instructions
.rend(); ++rit
) {
1390 aco_ptr
<Instruction
>& instr
= *rit
;
1391 if (is_phi(instr
)) {
1392 live
.erase(instr
->definitions
[0].getTemp());
1393 if (instr
->definitions
[0].isKill() || instr
->definitions
[0].isFixed())
1395 /* collect information about affinity-related temporaries */
1396 std::vector
<Temp
> affinity_related
;
1397 /* affinity_related[0] is the last seen affinity-related temp */
1398 affinity_related
.emplace_back(instr
->definitions
[0].getTemp());
1399 affinity_related
.emplace_back(instr
->definitions
[0].getTemp());
1400 for (const Operand
& op
: instr
->operands
) {
1401 if (op
.isTemp() && op
.regClass() == instr
->definitions
[0].regClass()) {
1402 affinity_related
.emplace_back(op
.getTemp());
1403 temp_to_phi_ressources
[op
.tempId()] = phi_ressources
.size();
1406 phi_ressources
.emplace_back(std::move(affinity_related
));
1410 /* add vector affinities */
1411 if (instr
->opcode
== aco_opcode::p_create_vector
) {
1412 for (const Operand
& op
: instr
->operands
) {
1413 if (op
.isTemp() && op
.isFirstKill() && op
.getTemp().type() == instr
->definitions
[0].getTemp().type())
1414 ctx
.vectors
[op
.tempId()] = instr
.get();
1418 /* add operands to live variables */
1419 for (const Operand
& op
: instr
->operands
) {
1421 live
.emplace(op
.getTemp());
1424 /* erase definitions from live */
1425 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
1426 const Definition
& def
= instr
->definitions
[i
];
1429 live
.erase(def
.getTemp());
1430 /* mark last-seen phi operand */
1431 std::unordered_map
<unsigned, unsigned>::iterator it
= temp_to_phi_ressources
.find(def
.tempId());
1432 if (it
!= temp_to_phi_ressources
.end() && def
.regClass() == phi_ressources
[it
->second
][0].regClass()) {
1433 phi_ressources
[it
->second
][0] = def
.getTemp();
1434 /* try to coalesce phi affinities with parallelcopies */
1435 if (!def
.isFixed() && instr
->opcode
== aco_opcode::p_parallelcopy
) {
1436 Operand op
= instr
->operands
[i
];
1437 if (op
.isTemp() && op
.isFirstKillBeforeDef() && def
.regClass() == op
.regClass()) {
1438 phi_ressources
[it
->second
].emplace_back(op
.getTemp());
1439 temp_to_phi_ressources
[op
.tempId()] = it
->second
;
1446 /* create affinities */
1447 for (std::vector
<Temp
>& vec
: phi_ressources
) {
1448 assert(vec
.size() > 1);
1449 for (unsigned i
= 1; i
< vec
.size(); i
++)
1450 if (vec
[i
].id() != vec
[0].id())
1451 ctx
.affinities
[vec
[i
].id()] = vec
[0].id();
1454 /* state of register file after phis */
1455 std::vector
<std::bitset
<128>> sgpr_live_in(program
->blocks
.size());
1457 for (Block
& block
: program
->blocks
) {
1458 TempSet
& live
= live_out_per_block
[block
.index
];
1459 /* initialize register file */
1460 assert(block
.index
!= 0 || live
.empty());
1461 RegisterFile register_file
;
1462 ctx
.war_hint
.reset();
1464 for (Temp t
: live
) {
1465 Temp renamed
= handle_live_in(ctx
, t
, &block
);
1466 assignment
& var
= ctx
.assignments
[renamed
.id()];
1467 /* due to live-range splits, the live-in might be a phi, now */
1469 register_file
.fill(Definition(renamed
.id(), var
.reg
, var
.rc
));
1472 std::vector
<aco_ptr
<Instruction
>> instructions
;
1473 std::vector
<aco_ptr
<Instruction
>>::iterator it
;
1475 /* this is a slight adjustment from the paper as we already have phi nodes:
1476 * We consider them incomplete phis and only handle the definition. */
1478 /* handle fixed phi definitions */
1479 for (it
= block
.instructions
.begin(); it
!= block
.instructions
.end(); ++it
) {
1480 aco_ptr
<Instruction
>& phi
= *it
;
1483 Definition
& definition
= phi
->definitions
[0];
1484 if (!definition
.isFixed())
1487 /* check if a dead exec mask phi is needed */
1488 if (definition
.isKill()) {
1489 for (Operand
& op
: phi
->operands
) {
1490 assert(op
.isTemp());
1491 if (!ctx
.assignments
[op
.tempId()].assigned
||
1492 ctx
.assignments
[op
.tempId()].reg
!= exec
) {
1493 definition
.setKill(false);
1499 if (definition
.isKill())
1502 assert(definition
.physReg() == exec
);
1503 assert(!register_file
.test(definition
.physReg(), definition
.bytes()));
1504 register_file
.fill(definition
);
1505 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1508 /* look up the affinities */
1509 for (it
= block
.instructions
.begin(); it
!= block
.instructions
.end(); ++it
) {
1510 aco_ptr
<Instruction
>& phi
= *it
;
1513 Definition
& definition
= phi
->definitions
[0];
1514 if (definition
.isKill() || definition
.isFixed())
1517 if (ctx
.affinities
.find(definition
.tempId()) != ctx
.affinities
.end() &&
1518 ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].assigned
) {
1519 assert(ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].rc
== definition
.regClass());
1520 PhysReg reg
= ctx
.assignments
[ctx
.affinities
[definition
.tempId()]].reg
;
1521 bool try_use_special_reg
= reg
== scc
|| reg
== exec
;
1522 if (try_use_special_reg
) {
1523 for (const Operand
& op
: phi
->operands
) {
1524 if (!(op
.isTemp() && ctx
.assignments
[op
.tempId()].assigned
&&
1525 ctx
.assignments
[op
.tempId()].reg
== reg
)) {
1526 try_use_special_reg
= false;
1530 if (!try_use_special_reg
)
1533 /* only assign if register is still free */
1534 if (!register_file
.test(reg
, definition
.bytes())) {
1535 definition
.setFixed(reg
);
1536 register_file
.fill(definition
);
1537 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1542 /* find registers for phis without affinity or where the register was blocked */
1543 for (it
= block
.instructions
.begin();it
!= block
.instructions
.end(); ++it
) {
1544 aco_ptr
<Instruction
>& phi
= *it
;
1548 Definition
& definition
= phi
->definitions
[0];
1549 if (definition
.isKill())
1552 if (!definition
.isFixed()) {
1553 std::vector
<std::pair
<Operand
, Definition
>> parallelcopy
;
1554 /* try to find a register that is used by at least one operand */
1555 for (const Operand
& op
: phi
->operands
) {
1556 if (!(op
.isTemp() && ctx
.assignments
[op
.tempId()].assigned
))
1558 PhysReg reg
= ctx
.assignments
[op
.tempId()].reg
;
1559 /* we tried this already on the previous loop */
1560 if (reg
== scc
|| reg
== exec
)
1562 if (get_reg_specified(ctx
, register_file
, definition
.regClass(), parallelcopy
, phi
, reg
)) {
1563 definition
.setFixed(reg
);
1567 if (!definition
.isFixed())
1568 definition
.setFixed(get_reg(ctx
, register_file
, definition
.getTemp(), parallelcopy
, phi
));
1570 /* process parallelcopy */
1571 for (std::pair
<Operand
, Definition
> pc
: parallelcopy
) {
1572 /* see if it's a copy from a different phi */
1573 //TODO: prefer moving some previous phis over live-ins
1574 //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)
1575 Instruction
*prev_phi
= NULL
;
1576 std::vector
<aco_ptr
<Instruction
>>::iterator phi_it
;
1577 for (phi_it
= instructions
.begin(); phi_it
!= instructions
.end(); ++phi_it
) {
1578 if ((*phi_it
)->definitions
[0].tempId() == pc
.first
.tempId())
1579 prev_phi
= phi_it
->get();
1582 while (!prev_phi
&& is_phi(*++phi_it
)) {
1583 if ((*phi_it
)->definitions
[0].tempId() == pc
.first
.tempId())
1584 prev_phi
= phi_it
->get();
1587 /* if so, just update that phi's register */
1588 register_file
.clear(prev_phi
->definitions
[0]);
1589 prev_phi
->definitions
[0].setFixed(pc
.second
.physReg());
1590 ctx
.assignments
[prev_phi
->definitions
[0].tempId()] = {pc
.second
.physReg(), pc
.second
.regClass()};
1591 register_file
.fill(prev_phi
->definitions
[0]);
1596 std::unordered_map
<unsigned, Temp
>::iterator orig_it
= ctx
.orig_names
.find(pc
.first
.tempId());
1597 Temp orig
= pc
.first
.getTemp();
1598 if (orig_it
!= ctx
.orig_names
.end())
1599 orig
= orig_it
->second
;
1601 ctx
.orig_names
[pc
.second
.tempId()] = orig
;
1602 ctx
.renames
[block
.index
][orig
.id()] = pc
.second
.getTemp();
1604 /* otherwise, this is a live-in and we need to create a new phi
1605 * to move it in this block's predecessors */
1606 aco_opcode opcode
= pc
.first
.getTemp().is_linear() ? aco_opcode::p_linear_phi
: aco_opcode::p_phi
;
1607 std::vector
<unsigned>& preds
= pc
.first
.getTemp().is_linear() ? block
.linear_preds
: block
.logical_preds
;
1608 aco_ptr
<Instruction
> new_phi
{create_instruction
<Pseudo_instruction
>(opcode
, Format::PSEUDO
, preds
.size(), 1)};
1609 new_phi
->definitions
[0] = pc
.second
;
1610 for (unsigned i
= 0; i
< preds
.size(); i
++)
1611 new_phi
->operands
[i
] = Operand(pc
.first
);
1612 instructions
.emplace_back(std::move(new_phi
));
1615 register_file
.fill(definition
);
1616 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1618 live
.emplace(definition
.getTemp());
1620 /* update phi affinities */
1621 for (const Operand
& op
: phi
->operands
) {
1622 if (op
.isTemp() && op
.regClass() == phi
->definitions
[0].regClass())
1623 ctx
.affinities
[op
.tempId()] = definition
.tempId();
1626 instructions
.emplace_back(std::move(*it
));
1629 /* fill in sgpr_live_in */
1630 for (unsigned i
= 0; i
<= ctx
.max_used_sgpr
; i
++)
1631 sgpr_live_in
[block
.index
][i
] = register_file
[i
];
1632 sgpr_live_in
[block
.index
][127] = register_file
[scc
.reg()];
1634 /* Handle all other instructions of the block */
1635 for (; it
!= block
.instructions
.end(); ++it
) {
1636 aco_ptr
<Instruction
>& instr
= *it
;
1638 /* parallelcopies from p_phi are inserted here which means
1639 * live ranges of killed operands end here as well */
1640 if (instr
->opcode
== aco_opcode::p_logical_end
) {
1641 /* no need to process this instruction any further */
1642 if (block
.logical_succs
.size() != 1) {
1643 instructions
.emplace_back(std::move(instr
));
1647 Block
& succ
= program
->blocks
[block
.logical_succs
[0]];
1649 for (; idx
< succ
.logical_preds
.size(); idx
++) {
1650 if (succ
.logical_preds
[idx
] == block
.index
)
1653 for (aco_ptr
<Instruction
>& phi
: succ
.instructions
) {
1654 if (phi
->opcode
== aco_opcode::p_phi
) {
1655 if (phi
->operands
[idx
].isTemp() &&
1656 phi
->operands
[idx
].getTemp().type() == RegType::sgpr
&&
1657 phi
->operands
[idx
].isFirstKillBeforeDef()) {
1658 Temp phi_op
= read_variable(ctx
, phi
->operands
[idx
].getTemp(), block
.index
);
1659 PhysReg reg
= ctx
.assignments
[phi_op
.id()].reg
;
1660 assert(register_file
[reg
] == phi_op
.id());
1661 register_file
[reg
] = 0;
1663 } else if (phi
->opcode
!= aco_opcode::p_linear_phi
) {
1667 instructions
.emplace_back(std::move(instr
));
1671 std::vector
<std::pair
<Operand
, Definition
>> parallelcopy
;
1673 assert(!is_phi(instr
));
1675 /* handle operands */
1676 for (unsigned i
= 0; i
< instr
->operands
.size(); ++i
) {
1677 auto& operand
= instr
->operands
[i
];
1678 if (!operand
.isTemp())
1681 /* rename operands */
1682 operand
.setTemp(read_variable(ctx
, operand
.getTemp(), block
.index
));
1683 assert(ctx
.assignments
[operand
.tempId()].assigned
);
1685 PhysReg reg
= ctx
.assignments
[operand
.tempId()].reg
;
1686 if (operand_can_use_reg(instr
, i
, reg
))
1687 operand
.setFixed(reg
);
1689 get_reg_for_operand(ctx
, register_file
, parallelcopy
, instr
, operand
);
1691 if (instr
->format
== Format::EXP
||
1692 (instr
->isVMEM() && i
== 3 && ctx
.program
->chip_class
== GFX6
) ||
1693 (instr
->format
== Format::DS
&& static_cast<DS_instruction
*>(instr
.get())->gds
)) {
1694 for (unsigned j
= 0; j
< operand
.size(); j
++)
1695 ctx
.war_hint
.set(operand
.physReg().reg() + j
);
1698 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.getTemp().id());
1699 if (phi
!= ctx
.phi_map
.end())
1700 phi
->second
.uses
.emplace(instr
.get());
1703 /* remove dead vars from register file */
1704 for (const Operand
& op
: instr
->operands
) {
1705 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1706 register_file
.clear(op
);
1709 /* try to optimize v_mad_f32 -> v_mac_f32 */
1710 if (instr
->opcode
== aco_opcode::v_mad_f32
&&
1711 instr
->operands
[2].isTemp() &&
1712 instr
->operands
[2].isKillBeforeDef() &&
1713 instr
->operands
[2].getTemp().type() == RegType::vgpr
&&
1714 instr
->operands
[1].isTemp() &&
1715 instr
->operands
[1].getTemp().type() == RegType::vgpr
) { /* TODO: swap src0 and src1 in this case */
1716 VOP3A_instruction
* vop3
= static_cast<VOP3A_instruction
*>(instr
.get());
1717 bool can_use_mac
= !(vop3
->abs
[0] || vop3
->abs
[1] || vop3
->abs
[2] ||
1718 vop3
->neg
[0] || vop3
->neg
[1] || vop3
->neg
[2] ||
1719 vop3
->clamp
|| vop3
->omod
|| vop3
->opsel
);
1721 instr
->format
= Format::VOP2
;
1722 instr
->opcode
= aco_opcode::v_mac_f32
;
1726 /* handle definitions which must have the same register as an operand */
1727 if (instr
->opcode
== aco_opcode::v_interp_p2_f32
||
1728 instr
->opcode
== aco_opcode::v_mac_f32
||
1729 instr
->opcode
== aco_opcode::v_writelane_b32
||
1730 instr
->opcode
== aco_opcode::v_writelane_b32_e64
) {
1731 instr
->definitions
[0].setFixed(instr
->operands
[2].physReg());
1732 } else if (instr
->opcode
== aco_opcode::s_addk_i32
||
1733 instr
->opcode
== aco_opcode::s_mulk_i32
) {
1734 instr
->definitions
[0].setFixed(instr
->operands
[0].physReg());
1735 } else if (instr
->format
== Format::MUBUF
&&
1736 instr
->definitions
.size() == 1 &&
1737 instr
->operands
.size() == 4) {
1738 instr
->definitions
[0].setFixed(instr
->operands
[3].physReg());
1739 } else if (instr
->format
== Format::MIMG
&&
1740 instr
->definitions
.size() == 1 &&
1741 instr
->operands
[1].regClass().type() == RegType::vgpr
) {
1742 instr
->definitions
[0].setFixed(instr
->operands
[1].physReg());
1745 ctx
.defs_done
.reset();
1747 /* handle fixed definitions first */
1748 for (unsigned i
= 0; i
< instr
->definitions
.size(); ++i
) {
1749 auto& definition
= instr
->definitions
[i
];
1750 if (!definition
.isFixed())
1753 adjust_max_used_regs(ctx
, definition
.regClass(), definition
.physReg());
1754 /* check if the target register is blocked */
1755 if (register_file
[definition
.physReg().reg()] != 0) {
1756 /* create parallelcopy pair to move blocking var */
1757 Temp tmp
= {register_file
[definition
.physReg()], ctx
.assignments
[register_file
[definition
.physReg()]].rc
};
1758 Operand pc_op
= Operand(tmp
);
1759 pc_op
.setFixed(ctx
.assignments
[register_file
[definition
.physReg().reg()]].reg
);
1760 RegClass rc
= pc_op
.regClass();
1761 tmp
= Temp
{program
->allocateId(), rc
};
1762 Definition pc_def
= Definition(tmp
);
1764 /* re-enable the killed operands, so that we don't move the blocking var there */
1765 for (const Operand
& op
: instr
->operands
) {
1766 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1767 register_file
.fill(op
);
1770 /* find a new register for the blocking variable */
1771 PhysReg reg
= get_reg(ctx
, register_file
, pc_op
.getTemp(), parallelcopy
, instr
);
1772 /* once again, disable killed operands */
1773 for (const Operand
& op
: instr
->operands
) {
1774 if (op
.isTemp() && op
.isFirstKillBeforeDef())
1775 register_file
.clear(op
);
1777 for (unsigned k
= 0; k
< i
; k
++) {
1778 if (instr
->definitions
[k
].isTemp() && ctx
.defs_done
.test(k
) && !instr
->definitions
[k
].isKill())
1779 register_file
.fill(instr
->definitions
[k
]);
1781 pc_def
.setFixed(reg
);
1783 /* finish assignment of parallelcopy */
1784 ctx
.assignments
.emplace_back(reg
, pc_def
.regClass());
1785 assert(ctx
.assignments
.size() == ctx
.program
->peekAllocationId());
1786 parallelcopy
.emplace_back(pc_op
, pc_def
);
1788 /* add changes to reg_file */
1789 register_file
.clear(pc_op
);
1790 register_file
.fill(pc_def
);
1792 ctx
.defs_done
.set(i
);
1794 if (!definition
.isTemp())
1797 /* set live if it has a kill point */
1798 if (!definition
.isKill())
1799 live
.emplace(definition
.getTemp());
1801 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1802 register_file
.fill(definition
);
1805 /* handle all other definitions */
1806 for (unsigned i
= 0; i
< instr
->definitions
.size(); ++i
) {
1807 auto& definition
= instr
->definitions
[i
];
1809 if (definition
.isFixed() || !definition
.isTemp())
1813 if (definition
.hasHint() && register_file
[definition
.physReg().reg()] == 0)
1814 definition
.setFixed(definition
.physReg());
1815 else if (instr
->opcode
== aco_opcode::p_split_vector
) {
1816 PhysReg reg
= instr
->operands
[0].physReg();
1817 reg
.reg_b
+= i
* definition
.bytes();
1818 if (get_reg_specified(ctx
, register_file
, definition
.regClass(), parallelcopy
, instr
, reg
))
1819 definition
.setFixed(reg
);
1820 } else if (instr
->opcode
== aco_opcode::p_wqm
) {
1822 if (instr
->operands
[0].isKillBeforeDef() && instr
->operands
[0].getTemp().type() == definition
.getTemp().type()) {
1823 reg
= instr
->operands
[0].physReg();
1824 definition
.setFixed(reg
);
1825 assert(register_file
[reg
.reg()] == 0);
1827 } else if (instr
->opcode
== aco_opcode::p_extract_vector
) {
1829 if (instr
->operands
[0].isKillBeforeDef() &&
1830 instr
->operands
[0].getTemp().type() == definition
.getTemp().type()) {
1831 reg
= instr
->operands
[0].physReg();
1832 reg
.reg_b
+= definition
.bytes() * instr
->operands
[1].constantValue();
1833 assert(!register_file
.test(reg
, definition
.bytes()));
1834 definition
.setFixed(reg
);
1836 } else if (instr
->opcode
== aco_opcode::p_create_vector
) {
1837 PhysReg reg
= get_reg_create_vector(ctx
, register_file
, definition
.getTemp(),
1838 parallelcopy
, instr
);
1839 definition
.setFixed(reg
);
1842 if (!definition
.isFixed()) {
1843 Temp tmp
= definition
.getTemp();
1844 /* subdword instructions before RDNA write full registers */
1845 if (tmp
.regClass().is_subdword() &&
1846 !instr_can_access_subdword(instr
) &&
1847 ctx
.program
->chip_class
<= GFX9
) {
1848 assert(tmp
.bytes() <= 4);
1849 tmp
= Temp(definition
.tempId(), v1
);
1851 definition
.setFixed(get_reg(ctx
, register_file
, tmp
, parallelcopy
, instr
));
1854 assert(definition
.isFixed() && ((definition
.getTemp().type() == RegType::vgpr
&& definition
.physReg() >= 256) ||
1855 (definition
.getTemp().type() != RegType::vgpr
&& definition
.physReg() < 256)));
1856 ctx
.defs_done
.set(i
);
1858 /* set live if it has a kill point */
1859 if (!definition
.isKill())
1860 live
.emplace(definition
.getTemp());
1862 ctx
.assignments
[definition
.tempId()] = {definition
.physReg(), definition
.regClass()};
1863 register_file
.fill(definition
);
1866 handle_pseudo(ctx
, register_file
, instr
.get());
1868 /* kill definitions and late-kill operands */
1869 for (const Definition
& def
: instr
->definitions
) {
1870 if (def
.isTemp() && def
.isKill())
1871 register_file
.clear(def
);
1873 for (const Operand
& op
: instr
->operands
) {
1874 if (op
.isTemp() && op
.isFirstKill() && op
.isLateKill())
1875 register_file
.clear(op
);
1878 /* emit parallelcopy */
1879 if (!parallelcopy
.empty()) {
1880 aco_ptr
<Pseudo_instruction
> pc
;
1881 pc
.reset(create_instruction
<Pseudo_instruction
>(aco_opcode::p_parallelcopy
, Format::PSEUDO
, parallelcopy
.size(), parallelcopy
.size()));
1882 bool temp_in_scc
= register_file
[scc
.reg()];
1883 bool sgpr_operands_alias_defs
= false;
1884 uint64_t sgpr_operands
[4] = {0, 0, 0, 0};
1885 for (unsigned i
= 0; i
< parallelcopy
.size(); i
++) {
1886 if (temp_in_scc
&& parallelcopy
[i
].first
.isTemp() && parallelcopy
[i
].first
.getTemp().type() == RegType::sgpr
) {
1887 if (!sgpr_operands_alias_defs
) {
1888 unsigned reg
= parallelcopy
[i
].first
.physReg().reg();
1889 unsigned size
= parallelcopy
[i
].first
.getTemp().size();
1890 sgpr_operands
[reg
/ 64u] |= ((1u << size
) - 1) << (reg
% 64u);
1892 reg
= parallelcopy
[i
].second
.physReg().reg();
1893 size
= parallelcopy
[i
].second
.getTemp().size();
1894 if (sgpr_operands
[reg
/ 64u] & ((1u << size
) - 1) << (reg
% 64u))
1895 sgpr_operands_alias_defs
= true;
1899 pc
->operands
[i
] = parallelcopy
[i
].first
;
1900 pc
->definitions
[i
] = parallelcopy
[i
].second
;
1901 assert(pc
->operands
[i
].size() == pc
->definitions
[i
].size());
1903 /* it might happen that the operand is already renamed. we have to restore the original name. */
1904 std::unordered_map
<unsigned, Temp
>::iterator it
= ctx
.orig_names
.find(pc
->operands
[i
].tempId());
1905 Temp orig
= it
!= ctx
.orig_names
.end() ? it
->second
: pc
->operands
[i
].getTemp();
1906 ctx
.orig_names
[pc
->definitions
[i
].tempId()] = orig
;
1907 ctx
.renames
[block
.index
][orig
.id()] = pc
->definitions
[i
].getTemp();
1909 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(pc
->operands
[i
].tempId());
1910 if (phi
!= ctx
.phi_map
.end())
1911 phi
->second
.uses
.emplace(pc
.get());
1914 if (temp_in_scc
&& sgpr_operands_alias_defs
) {
1915 /* disable definitions and re-enable operands */
1916 for (const Definition
& def
: instr
->definitions
) {
1917 if (def
.isTemp() && !def
.isKill())
1918 register_file
.clear(def
);
1920 for (const Operand
& op
: instr
->operands
) {
1921 if (op
.isTemp() && op
.isFirstKill())
1922 register_file
.block(op
.physReg(), op
.bytes());
1925 handle_pseudo(ctx
, register_file
, pc
.get());
1927 /* re-enable live vars */
1928 for (const Operand
& op
: instr
->operands
) {
1929 if (op
.isTemp() && op
.isFirstKill())
1930 register_file
.clear(op
);
1932 for (const Definition
& def
: instr
->definitions
) {
1933 if (def
.isTemp() && !def
.isKill())
1934 register_file
.fill(def
);
1937 pc
->tmp_in_scc
= false;
1940 instructions
.emplace_back(std::move(pc
));
1943 /* some instructions need VOP3 encoding if operand/definition is not assigned to VCC */
1944 bool instr_needs_vop3
= !instr
->isVOP3() &&
1945 ((instr
->format
== Format::VOPC
&& !(instr
->definitions
[0].physReg() == vcc
)) ||
1946 (instr
->opcode
== aco_opcode::v_cndmask_b32
&& !(instr
->operands
[2].physReg() == vcc
)) ||
1947 ((instr
->opcode
== aco_opcode::v_add_co_u32
||
1948 instr
->opcode
== aco_opcode::v_addc_co_u32
||
1949 instr
->opcode
== aco_opcode::v_sub_co_u32
||
1950 instr
->opcode
== aco_opcode::v_subb_co_u32
||
1951 instr
->opcode
== aco_opcode::v_subrev_co_u32
||
1952 instr
->opcode
== aco_opcode::v_subbrev_co_u32
) &&
1953 !(instr
->definitions
[1].physReg() == vcc
)) ||
1954 ((instr
->opcode
== aco_opcode::v_addc_co_u32
||
1955 instr
->opcode
== aco_opcode::v_subb_co_u32
||
1956 instr
->opcode
== aco_opcode::v_subbrev_co_u32
) &&
1957 !(instr
->operands
[2].physReg() == vcc
)));
1958 if (instr_needs_vop3
) {
1960 /* if the first operand is a literal, we have to move it to a reg */
1961 if (instr
->operands
.size() && instr
->operands
[0].isLiteral() && program
->chip_class
< GFX10
) {
1962 bool can_sgpr
= true;
1963 /* check, if we have to move to vgpr */
1964 for (const Operand
& op
: instr
->operands
) {
1965 if (op
.isTemp() && op
.getTemp().type() == RegType::sgpr
) {
1970 /* disable definitions and re-enable operands */
1971 for (const Definition
& def
: instr
->definitions
)
1972 register_file
.clear(def
);
1973 for (const Operand
& op
: instr
->operands
) {
1974 if (op
.isTemp() && op
.isFirstKill())
1975 register_file
.block(op
.physReg(), op
.bytes());
1977 Temp tmp
= {program
->allocateId(), can_sgpr
? s1
: v1
};
1978 ctx
.assignments
.emplace_back();
1979 PhysReg reg
= get_reg(ctx
, register_file
, tmp
, parallelcopy
, instr
);
1981 aco_ptr
<Instruction
> mov
;
1983 mov
.reset(create_instruction
<SOP1_instruction
>(aco_opcode::s_mov_b32
, Format::SOP1
, 1, 1));
1985 mov
.reset(create_instruction
<VOP1_instruction
>(aco_opcode::v_mov_b32
, Format::VOP1
, 1, 1));
1986 mov
->operands
[0] = instr
->operands
[0];
1987 mov
->definitions
[0] = Definition(tmp
);
1988 mov
->definitions
[0].setFixed(reg
);
1990 instr
->operands
[0] = Operand(tmp
);
1991 instr
->operands
[0].setFixed(reg
);
1992 instructions
.emplace_back(std::move(mov
));
1993 /* re-enable live vars */
1994 for (const Operand
& op
: instr
->operands
) {
1995 if (op
.isTemp() && op
.isFirstKill())
1996 register_file
.clear(op
);
1998 for (const Definition
& def
: instr
->definitions
) {
1999 if (def
.isTemp() && !def
.isKill())
2000 register_file
.fill(def
);
2004 /* change the instruction to VOP3 to enable an arbitrary register pair as dst */
2005 aco_ptr
<Instruction
> tmp
= std::move(instr
);
2006 Format format
= asVOP3(tmp
->format
);
2007 instr
.reset(create_instruction
<VOP3A_instruction
>(tmp
->opcode
, format
, tmp
->operands
.size(), tmp
->definitions
.size()));
2008 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
2009 Operand
& operand
= tmp
->operands
[i
];
2010 instr
->operands
[i
] = operand
;
2011 /* keep phi_map up to date */
2012 if (operand
.isTemp()) {
2013 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.tempId());
2014 if (phi
!= ctx
.phi_map
.end()) {
2015 phi
->second
.uses
.erase(tmp
.get());
2016 phi
->second
.uses
.emplace(instr
.get());
2020 std::copy(tmp
->definitions
.begin(), tmp
->definitions
.end(), instr
->definitions
.begin());
2022 instructions
.emplace_back(std::move(*it
));
2024 } /* end for Instr */
2026 block
.instructions
= std::move(instructions
);
2028 ctx
.filled
[block
.index
] = true;
2029 for (unsigned succ_idx
: block
.linear_succs
) {
2030 Block
& succ
= program
->blocks
[succ_idx
];
2031 /* seal block if all predecessors are filled */
2032 bool all_filled
= true;
2033 for (unsigned pred_idx
: succ
.linear_preds
) {
2034 if (!ctx
.filled
[pred_idx
]) {
2040 ctx
.sealed
[succ_idx
] = true;
2042 /* finish incomplete phis and check if they became trivial */
2043 for (Instruction
* phi
: ctx
.incomplete_phis
[succ_idx
]) {
2044 std::vector
<unsigned> preds
= phi
->definitions
[0].getTemp().is_linear() ? succ
.linear_preds
: succ
.logical_preds
;
2045 for (unsigned i
= 0; i
< phi
->operands
.size(); i
++) {
2046 phi
->operands
[i
].setTemp(read_variable(ctx
, phi
->operands
[i
].getTemp(), preds
[i
]));
2047 phi
->operands
[i
].setFixed(ctx
.assignments
[phi
->operands
[i
].tempId()].reg
);
2049 try_remove_trivial_phi(ctx
, phi
->definitions
[0].getTemp());
2051 /* complete the original phi nodes, but no need to check triviality */
2052 for (aco_ptr
<Instruction
>& instr
: succ
.instructions
) {
2055 std::vector
<unsigned> preds
= instr
->opcode
== aco_opcode::p_phi
? succ
.logical_preds
: succ
.linear_preds
;
2057 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
2058 auto& operand
= instr
->operands
[i
];
2059 if (!operand
.isTemp())
2061 operand
.setTemp(read_variable(ctx
, operand
.getTemp(), preds
[i
]));
2062 operand
.setFixed(ctx
.assignments
[operand
.tempId()].reg
);
2063 std::unordered_map
<unsigned, phi_info
>::iterator phi
= ctx
.phi_map
.find(operand
.getTemp().id());
2064 if (phi
!= ctx
.phi_map
.end())
2065 phi
->second
.uses
.emplace(instr
.get());
2072 /* remove trivial phis */
2073 for (Block
& block
: program
->blocks
) {
2074 auto end
= std::find_if(block
.instructions
.begin(), block
.instructions
.end(),
2075 [](aco_ptr
<Instruction
>& instr
) { return !is_phi(instr
);});
2076 auto middle
= std::remove_if(block
.instructions
.begin(), end
,
2077 [](const aco_ptr
<Instruction
>& instr
) { return instr
->definitions
.empty();});
2078 block
.instructions
.erase(middle
, end
);
2081 /* find scc spill registers which may be needed for parallelcopies created by phis */
2082 for (Block
& block
: program
->blocks
) {
2083 if (block
.linear_preds
.size() <= 1)
2086 std::bitset
<128> regs
= sgpr_live_in
[block
.index
];
2090 /* choose a register */
2092 for (; reg
< ctx
.program
->max_reg_demand
.sgpr
&& regs
[reg
]; reg
++)
2094 assert(reg
< ctx
.program
->max_reg_demand
.sgpr
);
2095 adjust_max_used_regs(ctx
, s1
, reg
);
2097 /* update predecessors */
2098 for (unsigned& pred_index
: block
.linear_preds
) {
2099 Block
& pred
= program
->blocks
[pred_index
];
2100 pred
.scc_live_out
= true;
2101 pred
.scratch_sgpr
= PhysReg
{(uint16_t)reg
};
2105 /* num_gpr = rnd_up(max_used_gpr + 1) */
2106 program
->config
->num_vgprs
= align(ctx
.max_used_vgpr
+ 1, 4);
2107 if (program
->family
== CHIP_TONGA
|| program
->family
== CHIP_ICELAND
) /* workaround hardware bug */
2108 program
->config
->num_sgprs
= get_sgpr_alloc(program
, program
->sgpr_limit
);
2110 program
->config
->num_sgprs
= align(ctx
.max_used_sgpr
+ 1 + get_extra_sgprs(program
), 8);