2 * Copyright (c) 2013 Rob Clark <robdclark@gmail.com>
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 #include "util/u_debug.h"
32 #include "instr-a3xx.h"
33 #include "disasm.h" /* TODO move 'enum shader_t' somewhere else.. */
35 /* low level intermediate representation of an adreno shader program */
38 struct ir3_instruction
;
43 uint16_t instrs_count
; /* expanded to account for rpt's */
44 /* NOTE: max_reg, etc, does not include registers not touched
45 * by the shader (ie. vertex fetched via VFD_DECODE but not
48 int8_t max_reg
; /* highest GPR # used by shader */
55 IR3_REG_CONST
= 0x001,
56 IR3_REG_IMMED
= 0x002,
58 IR3_REG_RELATIV
= 0x008,
60 IR3_REG_NEGATE
= 0x020,
63 IR3_REG_POS_INF
= 0x100,
64 /* (ei) flag, end-input? Set on last bary, presumably to signal
65 * that the shader needs no more input:
68 /* meta-flags, for intermediate stages of IR, ie.
69 * before register assignment is done:
71 IR3_REG_SSA
= 0x1000, /* 'instr' is ptr to assigning instr */
72 IR3_REG_IA
= 0x2000, /* meta-input dst is "assigned" */
73 IR3_REG_ADDR
= 0x4000, /* register is a0.x */
77 * the component is in the low two bits of the reg #, so
78 * rN.x becomes: (N << 2) | x
89 /* for IR3_REG_SSA, src registers contain ptr back to
90 * assigning instruction.
92 struct ir3_instruction
*instr
;
95 /* used for cat5 instructions, but also for internal/IR level
96 * tracking of what registers are read/written by an instruction.
97 * wrmask may be a bad name since it is used to represent both
98 * src and dst that touch multiple adjacent registers.
101 /* for relative addressing, 32bits for array size is too small,
102 * but otoh we don't need to deal with disjoint sets, so instead
103 * use a simple size field (number of scalar components).
109 struct ir3_instruction
{
110 struct ir3_block
*block
;
114 /* (sy) flag is set on first instruction, and after sample
115 * instructions (probably just on RAW hazard).
117 IR3_INSTR_SY
= 0x001,
118 /* (ss) flag is set on first instruction, and first instruction
119 * to depend on the result of "long" instructions (RAW hazard):
121 * rcp, rsq, log2, exp2, sin, cos, sqrt
123 * It seems to synchronize until all in-flight instructions are
124 * completed, for example:
127 * add.f hr2.z, (neg)hr2.z, hc0.y
128 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
131 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
133 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
134 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
135 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
137 * The last mul.f does not have (ss) set, presumably because the
138 * (ss) on the previous instruction does the job.
140 * The blob driver also seems to set it on WAR hazards, although
141 * not really clear if this is needed or just blob compiler being
142 * sloppy. So far I haven't found a case where removing the (ss)
143 * causes problems for WAR hazard, but I could just be getting
147 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
150 IR3_INSTR_SS
= 0x002,
151 /* (jp) flag is set on jump targets:
153 IR3_INSTR_JP
= 0x004,
154 IR3_INSTR_UL
= 0x008,
155 IR3_INSTR_3D
= 0x010,
160 IR3_INSTR_S2EN
= 0x200,
161 /* meta-flags, for intermediate stages of IR, ie.
162 * before register assignment is done:
164 IR3_INSTR_MARK
= 0x1000,
171 struct ir3_register
**regs
;
179 type_t src_type
, dst_type
;
200 /* for meta-instructions, just used to hold extra data
201 * before instruction scheduling, etc
204 int off
; /* component/offset */
210 struct ir3_block
*if_block
, *else_block
;
213 struct ir3_block
*block
;
216 /* XXX keep this as big as all other union members! */
220 /* transient values used during various algorithms: */
222 /* The instruction depth is the max dependency distance to output.
224 * You can also think of it as the "cost", if we did any sort of
225 * optimization for register footprint. Ie. a value that is just
226 * result of moving a const to a reg would have a low cost, so to
227 * it could make sense to duplicate the instruction at various
228 * points where the result is needed to reduce register footprint.
230 * DEPTH_UNUSED used to mark unused instructions after depth
233 #define DEPTH_UNUSED ~0
237 /* Used during CP and RA stages. For fanin and shader inputs/
238 * outputs where we need a sequence of consecutive registers,
239 * keep track of each src instructions left (ie 'n-1') and right
240 * (ie 'n+1') neighbor. The front-end must insert enough mov's
241 * to ensure that each instruction has at most one left and at
242 * most one right neighbor. During the copy-propagation pass,
243 * we only remove mov's when we can preserve this constraint.
244 * And during the RA stage, we use the neighbor information to
245 * allocate a block of registers in one shot.
247 * TODO: maybe just add something like:
248 * struct ir3_instruction_ref {
249 * struct ir3_instruction *instr;
253 * Or can we get away without the refcnt stuff? It seems like
254 * it should be overkill.. the problem is if, potentially after
255 * already eliminating some mov's, if you have a single mov that
256 * needs to be grouped with it's neighbors in two different
257 * places (ex. shader output and a fanin).
260 struct ir3_instruction
*left
, *right
;
261 uint16_t left_cnt
, right_cnt
;
264 /* an instruction can reference at most one address register amongst
265 * it's src/dst registers. Beyond that, you need to insert mov's.
267 struct ir3_instruction
*address
;
269 /* in case of a instruction with relative dst instruction, we need to
270 * capture the dependency on the fanin for the previous values of
271 * the array elements. Since we don't know at compile time actually
272 * which array elements are written, this serves to preserve the
273 * unconditional write to array elements prior to the conditional
276 * TODO only cat1 can do indirect write.. we could maybe move this
277 * into instr->cat1.fanin (but would require the frontend to insert
280 struct ir3_instruction
*fanin
;
282 struct ir3_instruction
*next
;
288 static inline struct ir3_instruction
*
289 ir3_neighbor_first(struct ir3_instruction
*instr
)
291 while (instr
->cp
.left
)
292 instr
= instr
->cp
.left
;
296 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
300 debug_assert(!instr
->cp
.left
);
302 while (instr
->cp
.right
) {
304 instr
= instr
->cp
.right
;
310 struct ir3_heap_chunk
;
313 unsigned instrs_count
, instrs_sz
;
314 struct ir3_instruction
**instrs
;
316 /* Track bary.f (and ldlv) instructions.. this is needed in
317 * scheduling to ensure that all varying fetches happen before
318 * any potential kill instructions. The hw gets grumpy if all
319 * threads in a group are killed before the last bary.f gets
320 * a chance to signal end of input (ei).
322 unsigned baryfs_count
, baryfs_sz
;
323 struct ir3_instruction
**baryfs
;
325 /* Track all indirect instructions (read and write). To avoid
326 * deadlock scenario where an address register gets scheduled,
327 * but other dependent src instructions cannot be scheduled due
328 * to dependency on a *different* address register value, the
329 * scheduler needs to ensure that all dependencies other than
330 * the instruction other than the address register are scheduled
331 * before the one that writes the address register. Having a
332 * convenient list of instructions that reference some address
333 * register simplifies this.
335 unsigned indirects_count
, indirects_sz
;
336 struct ir3_instruction
**indirects
;
338 struct ir3_block
*block
;
340 struct ir3_heap_chunk
*chunk
;
345 unsigned ntemporaries
, ninputs
, noutputs
;
346 /* maps TGSI_FILE_TEMPORARY index back to the assigning instruction: */
347 struct ir3_instruction
**temporaries
;
348 struct ir3_instruction
**inputs
;
349 struct ir3_instruction
**outputs
;
350 /* only a single address register: */
351 struct ir3_instruction
*address
;
352 struct ir3_block
*parent
;
353 struct ir3_instruction
*head
;
356 struct ir3
* ir3_create(void);
357 void ir3_destroy(struct ir3
*shader
);
358 void * ir3_assemble(struct ir3
*shader
,
359 struct ir3_info
*info
, uint32_t gpu_id
);
360 void * ir3_alloc(struct ir3
*shader
, int sz
);
362 struct ir3_block
* ir3_block_create(struct ir3
*shader
,
363 unsigned ntmp
, unsigned nin
, unsigned nout
);
365 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
,
366 int category
, opc_t opc
);
367 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
368 int category
, opc_t opc
, int nreg
);
369 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
370 const char *ir3_instr_name(struct ir3_instruction
*instr
);
372 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
376 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
378 if (instr
->flags
& IR3_INSTR_MARK
)
379 return true; /* already visited */
380 instr
->flags
|= IR3_INSTR_MARK
;
384 static inline void ir3_clear_mark(struct ir3
*shader
)
386 /* TODO would be nice to drop the instruction array.. for
387 * new compiler, _clear_mark() is all we use it for, and
388 * we could probably manage a linked list instead..
390 * Also, we'll probably want to mark instructions within
391 * a block, so tracking the list of instrs globally is
392 * unlikely to be what we want.
395 for (i
= 0; i
< shader
->instrs_count
; i
++) {
396 struct ir3_instruction
*instr
= shader
->instrs
[i
];
397 instr
->flags
&= ~IR3_INSTR_MARK
;
401 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
402 struct ir3_register
*reg
)
405 for (i
= 0; i
< instr
->regs_count
; i
++)
406 if (reg
== instr
->regs
[i
])
412 #define MAX_ARRAYS 16
420 static inline uint32_t regid(int num
, int comp
)
422 return (num
<< 2) | (comp
& 0x3);
425 static inline uint32_t reg_num(struct ir3_register
*reg
)
427 return reg
->num
>> 2;
430 static inline uint32_t reg_comp(struct ir3_register
*reg
)
432 return reg
->num
& 0x3;
435 static inline bool is_flow(struct ir3_instruction
*instr
)
437 return (instr
->category
== 0);
440 static inline bool is_kill(struct ir3_instruction
*instr
)
442 return is_flow(instr
) && (instr
->opc
== OPC_KILL
);
445 static inline bool is_nop(struct ir3_instruction
*instr
)
447 return is_flow(instr
) && (instr
->opc
== OPC_NOP
);
450 static inline bool is_alu(struct ir3_instruction
*instr
)
452 return (1 <= instr
->category
) && (instr
->category
<= 3);
455 static inline bool is_sfu(struct ir3_instruction
*instr
)
457 return (instr
->category
== 4);
460 static inline bool is_tex(struct ir3_instruction
*instr
)
462 return (instr
->category
== 5);
465 static inline bool is_mem(struct ir3_instruction
*instr
)
467 return (instr
->category
== 6);
470 static inline bool is_input(struct ir3_instruction
*instr
)
472 /* in some cases, ldlv is used to fetch varying without
473 * interpolation.. fortunately inloc is the first src
474 * register in either case
476 if (is_mem(instr
) && (instr
->opc
== OPC_LDLV
))
478 return (instr
->category
== 2) && (instr
->opc
== OPC_BARY_F
);
481 static inline bool is_meta(struct ir3_instruction
*instr
)
483 /* TODO how should we count PHI (and maybe fan-in/out) which
484 * might actually contribute some instructions to the final
487 return (instr
->category
== -1);
490 static inline bool writes_addr(struct ir3_instruction
*instr
)
492 if (instr
->regs_count
> 0) {
493 struct ir3_register
*dst
= instr
->regs
[0];
494 return !!(dst
->flags
& IR3_REG_ADDR
);
499 static inline bool writes_pred(struct ir3_instruction
*instr
)
501 if (instr
->regs_count
> 0) {
502 struct ir3_register
*dst
= instr
->regs
[0];
503 return reg_num(dst
) == REG_P0
;
508 /* returns defining instruction for reg */
509 /* TODO better name */
510 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
512 if (reg
->flags
& IR3_REG_SSA
)
517 static inline bool reg_gpr(struct ir3_register
*r
)
519 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
| IR3_REG_ADDR
))
521 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
526 #define array_insert(arr, val) do { \
527 if (arr ## _count == arr ## _sz) { \
528 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
529 arr = realloc(arr, arr ## _sz * sizeof(arr[0])); \
531 arr[arr ##_count++] = val; \
534 /* iterator for an instructions's sources (reg), also returns src #: */
535 #define foreach_src_n(__srcreg, __n, __instr) \
536 if ((__instr)->regs_count) \
537 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
538 if ((__srcreg = (__instr)->regs[__n + 1]))
540 /* iterator for an instructions's sources (reg): */
541 #define foreach_src(__srcreg, __instr) \
542 foreach_src_n(__srcreg, __i, __instr)
544 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
547 return instr
->regs_count
+ 2;
549 return instr
->regs_count
+ 1;
550 return instr
->regs_count
;
553 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
555 if (n
== (instr
->regs_count
+ 1))
557 if (n
== (instr
->regs_count
+ 0))
558 return instr
->address
;
559 return ssa(instr
->regs
[n
]);
562 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
564 /* iterator for an instruction's SSA sources (instr), also returns src #: */
565 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
566 if ((__instr)->regs_count) \
567 for (unsigned __cnt = __ssa_src_cnt(__instr) - 1, __n = 0; __n < __cnt; __n++) \
568 if ((__srcinst = __ssa_src_n(__instr, __n + 1)))
570 /* iterator for an instruction's SSA sources (instr): */
571 #define foreach_ssa_src(__srcinst, __instr) \
572 foreach_ssa_src_n(__srcinst, __i, __instr)
577 void ir3_dump(struct ir3
*shader
, const char *name
,
578 struct ir3_block
*block
/* XXX maybe 'block' ptr should move to ir3? */,
580 void ir3_dump_instr_single(struct ir3_instruction
*instr
);
581 void ir3_dump_instr_list(struct ir3_instruction
*instr
);
583 /* flatten if/else: */
584 int ir3_block_flatten(struct ir3_block
*block
);
586 /* depth calculation: */
587 int ir3_delayslots(struct ir3_instruction
*assigner
,
588 struct ir3_instruction
*consumer
, unsigned n
);
589 void ir3_block_depth(struct ir3_block
*block
);
591 /* copy-propagate: */
592 void ir3_block_cp(struct ir3_block
*block
);
594 /* group neightbors and insert mov's to resolve conflicts: */
595 void ir3_block_group(struct ir3_block
*block
);
598 int ir3_block_sched(struct ir3_block
*block
);
600 /* register assignment: */
601 int ir3_block_ra(struct ir3_block
*block
, enum shader_t type
,
602 bool frag_coord
, bool frag_face
);
605 void ir3_block_legalize(struct ir3_block
*block
,
606 bool *has_samp
, int *max_bary
);
608 /* ************************************************************************* */
609 /* instruction helpers */
611 static inline struct ir3_instruction
*
612 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
614 struct ir3_instruction
*instr
=
615 ir3_instr_create(block
, 1, 0);
616 ir3_reg_create(instr
, 0, 0); /* dst */
617 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
618 instr
->cat1
.src_type
= type
;
619 instr
->cat1
.dst_type
= type
;
623 static inline struct ir3_instruction
*
624 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
625 type_t src_type
, type_t dst_type
)
627 struct ir3_instruction
*instr
=
628 ir3_instr_create(block
, 1, 0);
629 ir3_reg_create(instr
, 0, 0); /* dst */
630 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
631 instr
->cat1
.src_type
= src_type
;
632 instr
->cat1
.dst_type
= dst_type
;
636 #define INSTR1(CAT, name) \
637 static inline struct ir3_instruction * \
638 ir3_##name(struct ir3_block *block, \
639 struct ir3_instruction *a, unsigned aflags) \
641 struct ir3_instruction *instr = \
642 ir3_instr_create(block, CAT, OPC_##name); \
643 ir3_reg_create(instr, 0, 0); /* dst */ \
644 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
648 #define INSTR2(CAT, name) \
649 static inline struct ir3_instruction * \
650 ir3_##name(struct ir3_block *block, \
651 struct ir3_instruction *a, unsigned aflags, \
652 struct ir3_instruction *b, unsigned bflags) \
654 struct ir3_instruction *instr = \
655 ir3_instr_create(block, CAT, OPC_##name); \
656 ir3_reg_create(instr, 0, 0); /* dst */ \
657 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
658 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
662 #define INSTR3(CAT, name) \
663 static inline struct ir3_instruction * \
664 ir3_##name(struct ir3_block *block, \
665 struct ir3_instruction *a, unsigned aflags, \
666 struct ir3_instruction *b, unsigned bflags, \
667 struct ir3_instruction *c, unsigned cflags) \
669 struct ir3_instruction *instr = \
670 ir3_instr_create(block, CAT, OPC_##name); \
671 ir3_reg_create(instr, 0, 0); /* dst */ \
672 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
673 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
674 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
678 /* cat0 instructions: */
681 /* cat2 instructions, most 2 src but some 1 src: */
729 /* cat3 instructions: */
747 /* cat4 instructions: */
756 /* cat5 instructions: */
760 /* cat6 instructions: */
763 /* ************************************************************************* */
764 /* split this out or find some helper to use.. like main/bitset.h.. */
770 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
772 static inline unsigned regmask_idx(struct ir3_register
*reg
)
774 unsigned num
= reg
->num
;
775 debug_assert(num
< MAX_REG
);
776 if (reg
->flags
& IR3_REG_HALF
)
781 static inline void regmask_init(regmask_t
*regmask
)
783 memset(regmask
, 0, sizeof(*regmask
));
786 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
788 unsigned idx
= regmask_idx(reg
);
789 if (reg
->flags
& IR3_REG_RELATIV
) {
791 for (i
= 0; i
< reg
->size
; i
++, idx
++)
792 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
795 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
797 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
801 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
804 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
805 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
808 /* set bits in a if not set in b, conceptually:
811 static inline void regmask_set_if_not(regmask_t
*a
,
812 struct ir3_register
*reg
, regmask_t
*b
)
814 unsigned idx
= regmask_idx(reg
);
815 if (reg
->flags
& IR3_REG_RELATIV
) {
817 for (i
= 0; i
< reg
->size
; i
++, idx
++)
818 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
819 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
822 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
824 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
825 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
829 static inline bool regmask_get(regmask_t
*regmask
,
830 struct ir3_register
*reg
)
832 unsigned idx
= regmask_idx(reg
);
833 if (reg
->flags
& IR3_REG_RELATIV
) {
835 for (i
= 0; i
< reg
->size
; i
++, idx
++)
836 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
840 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
842 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
848 /* ************************************************************************* */