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"
31 #include "util/list.h"
33 #include "instr-a3xx.h"
34 #include "disasm.h" /* TODO move 'enum shader_t' somewhere else.. */
36 /* low level intermediate representation of an adreno shader program */
40 struct ir3_instruction
;
46 uint16_t instrs_count
; /* expanded to account for rpt's */
47 /* NOTE: max_reg, etc, does not include registers not touched
48 * by the shader (ie. vertex fetched via VFD_DECODE but not
51 int8_t max_reg
; /* highest GPR # used by shader */
58 IR3_REG_CONST
= 0x001,
59 IR3_REG_IMMED
= 0x002,
61 /* high registers are used for some things in compute shaders,
62 * for example. Seems to be for things that are global to all
63 * threads in a wave, so possibly these are global/shared by
64 * all the threads in the wave?
67 IR3_REG_RELATIV
= 0x010,
69 /* Most instructions, it seems, can do float abs/neg but not
70 * integer. The CP pass needs to know what is intended (int or
71 * float) in order to do the right thing. For this reason the
72 * abs/neg flags are split out into float and int variants. In
73 * addition, .b (bitwise) operations, the negate is actually a
74 * bitwise not, so split that out into a new flag to make it
83 IR3_REG_POS_INF
= 0x1000,
84 /* (ei) flag, end-input? Set on last bary, presumably to signal
85 * that the shader needs no more input:
88 /* meta-flags, for intermediate stages of IR, ie.
89 * before register assignment is done:
91 IR3_REG_SSA
= 0x4000, /* 'instr' is ptr to assigning instr */
92 IR3_REG_ARRAY
= 0x8000,
93 IR3_REG_PHI_SRC
= 0x10000, /* phi src, regs[0]->instr points to phi */
98 * the component is in the low two bits of the reg #, so
99 * rN.x becomes: (N << 2) | x
113 /* For IR3_REG_SSA, src registers contain ptr back to assigning
116 * For IR3_REG_ARRAY, the pointer is back to the last dependent
117 * array access (although the net effect is the same, it points
118 * back to a previous instruction that we depend on).
120 struct ir3_instruction
*instr
;
123 /* used for cat5 instructions, but also for internal/IR level
124 * tracking of what registers are read/written by an instruction.
125 * wrmask may be a bad name since it is used to represent both
126 * src and dst that touch multiple adjacent registers.
129 /* for relative addressing, 32bits for array size is too small,
130 * but otoh we don't need to deal with disjoint sets, so instead
131 * use a simple size field (number of scalar components).
137 struct ir3_instruction
{
138 struct ir3_block
*block
;
141 /* (sy) flag is set on first instruction, and after sample
142 * instructions (probably just on RAW hazard).
144 IR3_INSTR_SY
= 0x001,
145 /* (ss) flag is set on first instruction, and first instruction
146 * to depend on the result of "long" instructions (RAW hazard):
148 * rcp, rsq, log2, exp2, sin, cos, sqrt
150 * It seems to synchronize until all in-flight instructions are
151 * completed, for example:
154 * add.f hr2.z, (neg)hr2.z, hc0.y
155 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
158 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
160 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
161 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
162 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
164 * The last mul.f does not have (ss) set, presumably because the
165 * (ss) on the previous instruction does the job.
167 * The blob driver also seems to set it on WAR hazards, although
168 * not really clear if this is needed or just blob compiler being
169 * sloppy. So far I haven't found a case where removing the (ss)
170 * causes problems for WAR hazard, but I could just be getting
174 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
177 IR3_INSTR_SS
= 0x002,
178 /* (jp) flag is set on jump targets:
180 IR3_INSTR_JP
= 0x004,
181 IR3_INSTR_UL
= 0x008,
182 IR3_INSTR_3D
= 0x010,
187 IR3_INSTR_S2EN
= 0x200,
189 /* meta-flags, for intermediate stages of IR, ie.
190 * before register assignment is done:
192 IR3_INSTR_MARK
= 0x1000,
193 IR3_INSTR_UNUSED
= 0x2000,
200 struct ir3_register
**regs
;
206 struct ir3_block
*target
;
209 type_t src_type
, dst_type
;
229 int iim_val
; /* for ldgb/stgb, # of components */
231 /* for meta-instructions, just used to hold extra data
232 * before instruction scheduling, etc
235 int off
; /* component/offset */
238 /* used to temporarily hold reference to nir_phi_instr
239 * until we resolve the phi srcs
244 struct ir3_block
*block
;
248 /* transient values used during various algorithms: */
250 /* The instruction depth is the max dependency distance to output.
252 * You can also think of it as the "cost", if we did any sort of
253 * optimization for register footprint. Ie. a value that is just
254 * result of moving a const to a reg would have a low cost, so to
255 * it could make sense to duplicate the instruction at various
256 * points where the result is needed to reduce register footprint.
259 /* When we get to the RA stage, we no longer need depth, but
260 * we do need instruction's position/name:
268 /* used for per-pass extra instruction data.
272 /* Used during CP and RA stages. For fanin and shader inputs/
273 * outputs where we need a sequence of consecutive registers,
274 * keep track of each src instructions left (ie 'n-1') and right
275 * (ie 'n+1') neighbor. The front-end must insert enough mov's
276 * to ensure that each instruction has at most one left and at
277 * most one right neighbor. During the copy-propagation pass,
278 * we only remove mov's when we can preserve this constraint.
279 * And during the RA stage, we use the neighbor information to
280 * allocate a block of registers in one shot.
282 * TODO: maybe just add something like:
283 * struct ir3_instruction_ref {
284 * struct ir3_instruction *instr;
288 * Or can we get away without the refcnt stuff? It seems like
289 * it should be overkill.. the problem is if, potentially after
290 * already eliminating some mov's, if you have a single mov that
291 * needs to be grouped with it's neighbors in two different
292 * places (ex. shader output and a fanin).
295 struct ir3_instruction
*left
, *right
;
296 uint16_t left_cnt
, right_cnt
;
299 /* an instruction can reference at most one address register amongst
300 * it's src/dst registers. Beyond that, you need to insert mov's.
302 * NOTE: do not write this directly, use ir3_instr_set_address()
304 struct ir3_instruction
*address
;
306 /* Entry in ir3_block's instruction list: */
307 struct list_head node
;
314 static inline struct ir3_instruction
*
315 ir3_neighbor_first(struct ir3_instruction
*instr
)
318 while (instr
->cp
.left
) {
319 instr
= instr
->cp
.left
;
320 if (++cnt
> 0xffff) {
328 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
332 debug_assert(!instr
->cp
.left
);
334 while (instr
->cp
.right
) {
336 instr
= instr
->cp
.right
;
347 * Stupid/simple growable array implementation:
349 #define DECLARE_ARRAY(type, name) \
350 unsigned name ## _count, name ## _sz; \
353 #define array_insert(ctx, arr, val) do { \
354 if (arr ## _count == arr ## _sz) { \
355 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
356 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
358 arr[arr ##_count++] = val; \
362 struct ir3_compiler
*compiler
;
364 unsigned ninputs
, noutputs
;
365 struct ir3_instruction
**inputs
;
366 struct ir3_instruction
**outputs
;
368 /* Track bary.f (and ldlv) instructions.. this is needed in
369 * scheduling to ensure that all varying fetches happen before
370 * any potential kill instructions. The hw gets grumpy if all
371 * threads in a group are killed before the last bary.f gets
372 * a chance to signal end of input (ei).
374 DECLARE_ARRAY(struct ir3_instruction
*, baryfs
);
376 /* Track all indirect instructions (read and write). To avoid
377 * deadlock scenario where an address register gets scheduled,
378 * but other dependent src instructions cannot be scheduled due
379 * to dependency on a *different* address register value, the
380 * scheduler needs to ensure that all dependencies other than
381 * the instruction other than the address register are scheduled
382 * before the one that writes the address register. Having a
383 * convenient list of instructions that reference some address
384 * register simplifies this.
386 DECLARE_ARRAY(struct ir3_instruction
*, indirects
);
388 /* and same for instructions that consume predicate register: */
389 DECLARE_ARRAY(struct ir3_instruction
*, predicates
);
391 /* Track texture sample instructions which need texture state
392 * patched in (for astc-srgb workaround):
394 DECLARE_ARRAY(struct ir3_instruction
*, astc_srgb
);
396 /* List of blocks: */
397 struct list_head block_list
;
399 /* List of ir3_array's: */
400 struct list_head array_list
;
403 typedef struct nir_variable nir_variable
;
406 struct list_head node
;
412 /* We track the last write and last access (read or write) to
413 * setup dependencies on instructions that read or write the
414 * array. Reads can be re-ordered wrt. other reads, but should
415 * not be re-ordered wrt. to writes. Writes cannot be reordered
416 * wrt. any other access to the array.
418 * So array reads depend on last write, and array writes depend
419 * on the last access.
421 struct ir3_instruction
*last_write
, *last_access
;
423 /* extra stuff used in RA pass: */
424 unsigned base
; /* base vreg name */
425 unsigned reg
; /* base physical reg */
426 uint16_t start_ip
, end_ip
;
429 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
431 typedef struct nir_block nir_block
;
434 struct list_head node
;
439 struct list_head instr_list
; /* list of ir3_instruction */
441 /* each block has either one or two successors.. in case of
442 * two successors, 'condition' decides which one to follow.
443 * A block preceding an if/else has two successors.
445 struct ir3_instruction
*condition
;
446 struct ir3_block
*successors
[2];
448 uint16_t start_ip
, end_ip
;
450 /* Track instructions which do not write a register but other-
451 * wise must not be discarded (such as kill, stg, etc)
453 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
455 /* used for per-pass extra block data. Mainly used right
456 * now in RA step to track livein/liveout.
465 static inline uint32_t
466 block_id(struct ir3_block
*block
)
469 return block
->serialno
;
471 return (uint32_t)(unsigned long)block
;
475 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
476 unsigned nin
, unsigned nout
);
477 void ir3_destroy(struct ir3
*shader
);
478 void * ir3_assemble(struct ir3
*shader
,
479 struct ir3_info
*info
, uint32_t gpu_id
);
480 void * ir3_alloc(struct ir3
*shader
, int sz
);
482 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
484 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
485 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
486 opc_t opc
, int nreg
);
487 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
488 const char *ir3_instr_name(struct ir3_instruction
*instr
);
490 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
492 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
493 struct ir3_register
*reg
);
495 void ir3_instr_set_address(struct ir3_instruction
*instr
,
496 struct ir3_instruction
*addr
);
498 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
500 if (instr
->flags
& IR3_INSTR_MARK
)
501 return true; /* already visited */
502 instr
->flags
|= IR3_INSTR_MARK
;
506 void ir3_block_clear_mark(struct ir3_block
*block
);
507 void ir3_clear_mark(struct ir3
*shader
);
509 unsigned ir3_count_instructions(struct ir3
*ir
);
511 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
512 struct ir3_register
*reg
)
515 for (i
= 0; i
< instr
->regs_count
; i
++)
516 if (reg
== instr
->regs
[i
])
522 #define MAX_ARRAYS 16
530 static inline uint32_t regid(int num
, int comp
)
532 return (num
<< 2) | (comp
& 0x3);
535 static inline uint32_t reg_num(struct ir3_register
*reg
)
537 return reg
->num
>> 2;
540 static inline uint32_t reg_comp(struct ir3_register
*reg
)
542 return reg
->num
& 0x3;
545 static inline bool is_flow(struct ir3_instruction
*instr
)
547 return (opc_cat(instr
->opc
) == 0);
550 static inline bool is_kill(struct ir3_instruction
*instr
)
552 return instr
->opc
== OPC_KILL
;
555 static inline bool is_nop(struct ir3_instruction
*instr
)
557 return instr
->opc
== OPC_NOP
;
560 /* Is it a non-transformative (ie. not type changing) mov? This can
561 * also include absneg.s/absneg.f, which for the most part can be
562 * treated as a mov (single src argument).
564 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
566 struct ir3_register
*dst
= instr
->regs
[0];
568 /* mov's that write to a0.x or p0.x are special: */
569 if (dst
->num
== regid(REG_P0
, 0))
571 if (dst
->num
== regid(REG_A0
, 0))
574 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
577 switch (instr
->opc
) {
579 return instr
->cat1
.src_type
== instr
->cat1
.dst_type
;
588 static inline bool is_alu(struct ir3_instruction
*instr
)
590 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
593 static inline bool is_sfu(struct ir3_instruction
*instr
)
595 return (opc_cat(instr
->opc
) == 4);
598 static inline bool is_tex(struct ir3_instruction
*instr
)
600 return (opc_cat(instr
->opc
) == 5);
603 static inline bool is_mem(struct ir3_instruction
*instr
)
605 return (opc_cat(instr
->opc
) == 6);
609 is_store(struct ir3_instruction
*instr
)
611 /* these instructions, the "destination" register is
612 * actually a source, the address to store to.
614 switch (instr
->opc
) {
628 static inline bool is_load(struct ir3_instruction
*instr
)
630 switch (instr
->opc
) {
639 /* probably some others too.. */
646 static inline bool is_input(struct ir3_instruction
*instr
)
648 /* in some cases, ldlv is used to fetch varying without
649 * interpolation.. fortunately inloc is the first src
650 * register in either case
652 switch (instr
->opc
) {
661 static inline bool is_bool(struct ir3_instruction
*instr
)
663 switch (instr
->opc
) {
673 static inline bool is_meta(struct ir3_instruction
*instr
)
675 /* TODO how should we count PHI (and maybe fan-in/out) which
676 * might actually contribute some instructions to the final
679 return (opc_cat(instr
->opc
) == -1);
682 static inline bool writes_addr(struct ir3_instruction
*instr
)
684 if (instr
->regs_count
> 0) {
685 struct ir3_register
*dst
= instr
->regs
[0];
686 return reg_num(dst
) == REG_A0
;
691 static inline bool writes_pred(struct ir3_instruction
*instr
)
693 if (instr
->regs_count
> 0) {
694 struct ir3_register
*dst
= instr
->regs
[0];
695 return reg_num(dst
) == REG_P0
;
700 /* returns defining instruction for reg */
701 /* TODO better name */
702 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
704 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
705 debug_assert(!(reg
->instr
&& (reg
->instr
->flags
& IR3_INSTR_UNUSED
)));
711 static inline bool conflicts(struct ir3_instruction
*a
,
712 struct ir3_instruction
*b
)
714 return (a
&& b
) && (a
!= b
);
717 static inline bool reg_gpr(struct ir3_register
*r
)
719 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
721 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
726 static inline type_t
half_type(type_t type
)
729 case TYPE_F32
: return TYPE_F16
;
730 case TYPE_U32
: return TYPE_U16
;
731 case TYPE_S32
: return TYPE_S16
;
742 /* some cat2 instructions (ie. those which are not float) can embed an
745 static inline bool ir3_cat2_int(opc_t opc
)
786 /* map cat2 instruction to valid abs/neg flags: */
787 static inline unsigned ir3_cat2_absneg(opc_t opc
)
804 return IR3_REG_FABS
| IR3_REG_FNEG
;
825 return IR3_REG_SABS
| IR3_REG_SNEG
;
846 /* map cat3 instructions to valid abs/neg flags: */
847 static inline unsigned ir3_cat3_absneg(opc_t opc
)
866 /* neg *may* work on 3rd src.. */
876 /* iterator for an instructions's sources (reg), also returns src #: */
877 #define foreach_src_n(__srcreg, __n, __instr) \
878 if ((__instr)->regs_count) \
879 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
880 if ((__srcreg = (__instr)->regs[__n + 1]))
882 /* iterator for an instructions's sources (reg): */
883 #define foreach_src(__srcreg, __instr) \
884 foreach_src_n(__srcreg, __i, __instr)
886 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
889 return instr
->regs_count
+ 1;
890 return instr
->regs_count
;
893 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
895 if (n
== (instr
->regs_count
+ 0))
896 return instr
->address
;
897 return ssa(instr
->regs
[n
]);
900 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
902 /* iterator for an instruction's SSA sources (instr), also returns src #: */
903 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
904 if ((__instr)->regs_count) \
905 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
906 if ((__srcinst = __ssa_src_n(__instr, __n)))
908 /* iterator for an instruction's SSA sources (instr): */
909 #define foreach_ssa_src(__srcinst, __instr) \
910 foreach_ssa_src_n(__srcinst, __i, __instr)
914 void ir3_print(struct ir3
*ir
);
915 void ir3_print_instr(struct ir3_instruction
*instr
);
917 /* depth calculation: */
918 int ir3_delayslots(struct ir3_instruction
*assigner
,
919 struct ir3_instruction
*consumer
, unsigned n
);
920 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
921 void ir3_depth(struct ir3
*ir
);
923 /* copy-propagate: */
924 struct ir3_shader_variant
;
925 void ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
927 /* group neighbors and insert mov's to resolve conflicts: */
928 void ir3_group(struct ir3
*ir
);
931 int ir3_sched(struct ir3
*ir
);
933 /* register assignment: */
934 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(void *memctx
);
935 int ir3_ra(struct ir3
*ir3
, enum shader_t type
,
936 bool frag_coord
, bool frag_face
);
939 void ir3_legalize(struct ir3
*ir
, bool *has_samp
, bool *has_ssbo
, int *max_bary
);
941 /* ************************************************************************* */
942 /* instruction helpers */
944 static inline struct ir3_instruction
*
945 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
947 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
948 ir3_reg_create(instr
, 0, 0); /* dst */
949 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
950 struct ir3_register
*src_reg
=
951 ir3_reg_create(instr
, 0, IR3_REG_ARRAY
);
952 src_reg
->array
= src
->regs
[0]->array
;
953 src_reg
->instr
= src
;
955 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
957 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
958 instr
->cat1
.src_type
= type
;
959 instr
->cat1
.dst_type
= type
;
963 static inline struct ir3_instruction
*
964 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
965 type_t src_type
, type_t dst_type
)
967 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
968 ir3_reg_create(instr
, 0, 0); /* dst */
969 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
970 instr
->cat1
.src_type
= src_type
;
971 instr
->cat1
.dst_type
= dst_type
;
972 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
976 static inline struct ir3_instruction
*
977 ir3_NOP(struct ir3_block
*block
)
979 return ir3_instr_create(block
, OPC_NOP
);
982 #define INSTR0(name) \
983 static inline struct ir3_instruction * \
984 ir3_##name(struct ir3_block *block) \
986 struct ir3_instruction *instr = \
987 ir3_instr_create(block, OPC_##name); \
991 #define INSTR1(name) \
992 static inline struct ir3_instruction * \
993 ir3_##name(struct ir3_block *block, \
994 struct ir3_instruction *a, unsigned aflags) \
996 struct ir3_instruction *instr = \
997 ir3_instr_create(block, OPC_##name); \
998 ir3_reg_create(instr, 0, 0); /* dst */ \
999 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1003 #define INSTR2(name) \
1004 static inline struct ir3_instruction * \
1005 ir3_##name(struct ir3_block *block, \
1006 struct ir3_instruction *a, unsigned aflags, \
1007 struct ir3_instruction *b, unsigned bflags) \
1009 struct ir3_instruction *instr = \
1010 ir3_instr_create(block, OPC_##name); \
1011 ir3_reg_create(instr, 0, 0); /* dst */ \
1012 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1013 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1017 #define INSTR3(name) \
1018 static inline struct ir3_instruction * \
1019 ir3_##name(struct ir3_block *block, \
1020 struct ir3_instruction *a, unsigned aflags, \
1021 struct ir3_instruction *b, unsigned bflags, \
1022 struct ir3_instruction *c, unsigned cflags) \
1024 struct ir3_instruction *instr = \
1025 ir3_instr_create(block, OPC_##name); \
1026 ir3_reg_create(instr, 0, 0); /* dst */ \
1027 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1028 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1029 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1033 #define INSTR4(name) \
1034 static inline struct ir3_instruction * \
1035 ir3_##name(struct ir3_block *block, \
1036 struct ir3_instruction *a, unsigned aflags, \
1037 struct ir3_instruction *b, unsigned bflags, \
1038 struct ir3_instruction *c, unsigned cflags, \
1039 struct ir3_instruction *d, unsigned dflags) \
1041 struct ir3_instruction *instr = \
1042 ir3_instr_create2(block, OPC_##name, 5); \
1043 ir3_reg_create(instr, 0, 0); /* dst */ \
1044 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1045 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1046 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1047 ir3_reg_create(instr, 0, IR3_REG_SSA | dflags)->instr = d; \
1051 /* cat0 instructions: */
1057 /* cat2 instructions, most 2 src but some 1 src: */
1105 /* cat3 instructions: */
1123 /* cat4 instructions: */
1132 /* cat5 instructions: */
1136 static inline struct ir3_instruction
*
1137 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1138 unsigned wrmask
, unsigned flags
, unsigned samp
, unsigned tex
,
1139 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1141 struct ir3_instruction
*sam
;
1142 struct ir3_register
*reg
;
1144 sam
= ir3_instr_create(block
, opc
);
1145 sam
->flags
|= flags
;
1146 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1148 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1149 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1153 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1155 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1157 sam
->cat5
.samp
= samp
;
1158 sam
->cat5
.tex
= tex
;
1159 sam
->cat5
.type
= type
;
1164 /* cat6 instructions: */
1172 INSTR4(ATOMIC_XCHG
);
1175 INSTR4(ATOMIC_CMPXCHG
);
1182 /* ************************************************************************* */
1183 /* split this out or find some helper to use.. like main/bitset.h.. */
1189 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1191 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1193 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1194 debug_assert(num
< MAX_REG
);
1195 if (reg
->flags
& IR3_REG_HALF
)
1200 static inline void regmask_init(regmask_t
*regmask
)
1202 memset(regmask
, 0, sizeof(*regmask
));
1205 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1207 unsigned idx
= regmask_idx(reg
);
1208 if (reg
->flags
& IR3_REG_RELATIV
) {
1210 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1211 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1214 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1216 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1220 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1223 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1224 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1227 /* set bits in a if not set in b, conceptually:
1230 static inline void regmask_set_if_not(regmask_t
*a
,
1231 struct ir3_register
*reg
, regmask_t
*b
)
1233 unsigned idx
= regmask_idx(reg
);
1234 if (reg
->flags
& IR3_REG_RELATIV
) {
1236 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1237 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1238 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1241 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1243 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1244 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1248 static inline bool regmask_get(regmask_t
*regmask
,
1249 struct ir3_register
*reg
)
1251 unsigned idx
= regmask_idx(reg
);
1252 if (reg
->flags
& IR3_REG_RELATIV
) {
1254 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1255 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1259 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1261 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1267 /* ************************************************************************* */