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).
138 * Stupid/simple growable array implementation:
140 #define DECLARE_ARRAY(type, name) \
141 unsigned name ## _count, name ## _sz; \
144 #define array_insert(ctx, arr, val) do { \
145 if (arr ## _count == arr ## _sz) { \
146 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
147 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
149 arr[arr ##_count++] = val; \
152 struct ir3_instruction
{
153 struct ir3_block
*block
;
156 /* (sy) flag is set on first instruction, and after sample
157 * instructions (probably just on RAW hazard).
159 IR3_INSTR_SY
= 0x001,
160 /* (ss) flag is set on first instruction, and first instruction
161 * to depend on the result of "long" instructions (RAW hazard):
163 * rcp, rsq, log2, exp2, sin, cos, sqrt
165 * It seems to synchronize until all in-flight instructions are
166 * completed, for example:
169 * add.f hr2.z, (neg)hr2.z, hc0.y
170 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
173 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
175 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
176 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
177 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
179 * The last mul.f does not have (ss) set, presumably because the
180 * (ss) on the previous instruction does the job.
182 * The blob driver also seems to set it on WAR hazards, although
183 * not really clear if this is needed or just blob compiler being
184 * sloppy. So far I haven't found a case where removing the (ss)
185 * causes problems for WAR hazard, but I could just be getting
189 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
192 IR3_INSTR_SS
= 0x002,
193 /* (jp) flag is set on jump targets:
195 IR3_INSTR_JP
= 0x004,
196 IR3_INSTR_UL
= 0x008,
197 IR3_INSTR_3D
= 0x010,
202 IR3_INSTR_S2EN
= 0x200,
204 /* meta-flags, for intermediate stages of IR, ie.
205 * before register assignment is done:
207 IR3_INSTR_MARK
= 0x1000,
208 IR3_INSTR_UNUSED
= 0x2000,
215 struct ir3_register
**regs
;
221 struct ir3_block
*target
;
224 type_t src_type
, dst_type
;
244 int iim_val
: 3; /* for ldgb/stgb, # of components */
249 unsigned w
: 1; /* write */
250 unsigned r
: 1; /* read */
251 unsigned l
: 1; /* local */
252 unsigned g
: 1; /* global */
254 /* for meta-instructions, just used to hold extra data
255 * before instruction scheduling, etc
258 int off
; /* component/offset */
261 /* used to temporarily hold reference to nir_phi_instr
262 * until we resolve the phi srcs
267 struct ir3_block
*block
;
271 /* transient values used during various algorithms: */
273 /* The instruction depth is the max dependency distance to output.
275 * You can also think of it as the "cost", if we did any sort of
276 * optimization for register footprint. Ie. a value that is just
277 * result of moving a const to a reg would have a low cost, so to
278 * it could make sense to duplicate the instruction at various
279 * points where the result is needed to reduce register footprint.
282 /* When we get to the RA stage, we no longer need depth, but
283 * we do need instruction's position/name:
291 /* used for per-pass extra instruction data.
295 /* Used during CP and RA stages. For fanin and shader inputs/
296 * outputs where we need a sequence of consecutive registers,
297 * keep track of each src instructions left (ie 'n-1') and right
298 * (ie 'n+1') neighbor. The front-end must insert enough mov's
299 * to ensure that each instruction has at most one left and at
300 * most one right neighbor. During the copy-propagation pass,
301 * we only remove mov's when we can preserve this constraint.
302 * And during the RA stage, we use the neighbor information to
303 * allocate a block of registers in one shot.
305 * TODO: maybe just add something like:
306 * struct ir3_instruction_ref {
307 * struct ir3_instruction *instr;
311 * Or can we get away without the refcnt stuff? It seems like
312 * it should be overkill.. the problem is if, potentially after
313 * already eliminating some mov's, if you have a single mov that
314 * needs to be grouped with it's neighbors in two different
315 * places (ex. shader output and a fanin).
318 struct ir3_instruction
*left
, *right
;
319 uint16_t left_cnt
, right_cnt
;
322 /* an instruction can reference at most one address register amongst
323 * it's src/dst registers. Beyond that, you need to insert mov's.
325 * NOTE: do not write this directly, use ir3_instr_set_address()
327 struct ir3_instruction
*address
;
329 /* Entry in ir3_block's instruction list: */
330 struct list_head node
;
337 static inline struct ir3_instruction
*
338 ir3_neighbor_first(struct ir3_instruction
*instr
)
341 while (instr
->cp
.left
) {
342 instr
= instr
->cp
.left
;
343 if (++cnt
> 0xffff) {
351 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
355 debug_assert(!instr
->cp
.left
);
357 while (instr
->cp
.right
) {
359 instr
= instr
->cp
.right
;
370 struct ir3_compiler
*compiler
;
372 unsigned ninputs
, noutputs
;
373 struct ir3_instruction
**inputs
;
374 struct ir3_instruction
**outputs
;
376 /* Track bary.f (and ldlv) instructions.. this is needed in
377 * scheduling to ensure that all varying fetches happen before
378 * any potential kill instructions. The hw gets grumpy if all
379 * threads in a group are killed before the last bary.f gets
380 * a chance to signal end of input (ei).
382 DECLARE_ARRAY(struct ir3_instruction
*, baryfs
);
384 /* Track all indirect instructions (read and write). To avoid
385 * deadlock scenario where an address register gets scheduled,
386 * but other dependent src instructions cannot be scheduled due
387 * to dependency on a *different* address register value, the
388 * scheduler needs to ensure that all dependencies other than
389 * the instruction other than the address register are scheduled
390 * before the one that writes the address register. Having a
391 * convenient list of instructions that reference some address
392 * register simplifies this.
394 DECLARE_ARRAY(struct ir3_instruction
*, indirects
);
396 /* and same for instructions that consume predicate register: */
397 DECLARE_ARRAY(struct ir3_instruction
*, predicates
);
399 /* Track texture sample instructions which need texture state
400 * patched in (for astc-srgb workaround):
402 DECLARE_ARRAY(struct ir3_instruction
*, astc_srgb
);
404 /* List of blocks: */
405 struct list_head block_list
;
407 /* List of ir3_array's: */
408 struct list_head array_list
;
411 typedef struct nir_register nir_register
;
414 struct list_head node
;
420 /* We track the last write and last access (read or write) to
421 * setup dependencies on instructions that read or write the
422 * array. Reads can be re-ordered wrt. other reads, but should
423 * not be re-ordered wrt. to writes. Writes cannot be reordered
424 * wrt. any other access to the array.
426 * So array reads depend on last write, and array writes depend
427 * on the last access.
429 struct ir3_instruction
*last_write
, *last_access
;
431 /* extra stuff used in RA pass: */
432 unsigned base
; /* base vreg name */
433 unsigned reg
; /* base physical reg */
434 uint16_t start_ip
, end_ip
;
437 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
439 typedef struct nir_block nir_block
;
442 struct list_head node
;
447 struct list_head instr_list
; /* list of ir3_instruction */
449 /* each block has either one or two successors.. in case of
450 * two successors, 'condition' decides which one to follow.
451 * A block preceding an if/else has two successors.
453 struct ir3_instruction
*condition
;
454 struct ir3_block
*successors
[2];
456 uint16_t start_ip
, end_ip
;
458 /* Track instructions which do not write a register but other-
459 * wise must not be discarded (such as kill, stg, etc)
461 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
463 /* used for per-pass extra block data. Mainly used right
464 * now in RA step to track livein/liveout.
473 static inline uint32_t
474 block_id(struct ir3_block
*block
)
477 return block
->serialno
;
479 return (uint32_t)(unsigned long)block
;
483 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
484 unsigned nin
, unsigned nout
);
485 void ir3_destroy(struct ir3
*shader
);
486 void * ir3_assemble(struct ir3
*shader
,
487 struct ir3_info
*info
, uint32_t gpu_id
);
488 void * ir3_alloc(struct ir3
*shader
, int sz
);
490 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
492 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
493 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
494 opc_t opc
, int nreg
);
495 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
496 const char *ir3_instr_name(struct ir3_instruction
*instr
);
498 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
500 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
501 struct ir3_register
*reg
);
503 void ir3_instr_set_address(struct ir3_instruction
*instr
,
504 struct ir3_instruction
*addr
);
506 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
508 if (instr
->flags
& IR3_INSTR_MARK
)
509 return true; /* already visited */
510 instr
->flags
|= IR3_INSTR_MARK
;
514 void ir3_block_clear_mark(struct ir3_block
*block
);
515 void ir3_clear_mark(struct ir3
*shader
);
517 unsigned ir3_count_instructions(struct ir3
*ir
);
519 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
520 struct ir3_register
*reg
)
523 for (i
= 0; i
< instr
->regs_count
; i
++)
524 if (reg
== instr
->regs
[i
])
530 #define MAX_ARRAYS 16
538 static inline uint32_t regid(int num
, int comp
)
540 return (num
<< 2) | (comp
& 0x3);
543 static inline uint32_t reg_num(struct ir3_register
*reg
)
545 return reg
->num
>> 2;
548 static inline uint32_t reg_comp(struct ir3_register
*reg
)
550 return reg
->num
& 0x3;
553 static inline bool is_flow(struct ir3_instruction
*instr
)
555 return (opc_cat(instr
->opc
) == 0);
558 static inline bool is_kill(struct ir3_instruction
*instr
)
560 return instr
->opc
== OPC_KILL
;
563 static inline bool is_nop(struct ir3_instruction
*instr
)
565 return instr
->opc
== OPC_NOP
;
568 /* Is it a non-transformative (ie. not type changing) mov? This can
569 * also include absneg.s/absneg.f, which for the most part can be
570 * treated as a mov (single src argument).
572 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
574 struct ir3_register
*dst
;
576 switch (instr
->opc
) {
578 if (instr
->cat1
.src_type
!= instr
->cat1
.dst_type
)
588 dst
= instr
->regs
[0];
590 /* mov's that write to a0.x or p0.x are special: */
591 if (dst
->num
== regid(REG_P0
, 0))
593 if (dst
->num
== regid(REG_A0
, 0))
596 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
602 static inline bool is_alu(struct ir3_instruction
*instr
)
604 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
607 static inline bool is_sfu(struct ir3_instruction
*instr
)
609 return (opc_cat(instr
->opc
) == 4);
612 static inline bool is_tex(struct ir3_instruction
*instr
)
614 return (opc_cat(instr
->opc
) == 5);
617 static inline bool is_mem(struct ir3_instruction
*instr
)
619 return (opc_cat(instr
->opc
) == 6);
622 static inline bool is_barrier(struct ir3_instruction
*instr
)
624 return (opc_cat(instr
->opc
) == 7);
628 is_store(struct ir3_instruction
*instr
)
630 /* these instructions, the "destination" register is
631 * actually a source, the address to store to.
633 switch (instr
->opc
) {
648 static inline bool is_load(struct ir3_instruction
*instr
)
650 switch (instr
->opc
) {
659 /* probably some others too.. */
666 static inline bool is_input(struct ir3_instruction
*instr
)
668 /* in some cases, ldlv is used to fetch varying without
669 * interpolation.. fortunately inloc is the first src
670 * register in either case
672 switch (instr
->opc
) {
681 static inline bool is_bool(struct ir3_instruction
*instr
)
683 switch (instr
->opc
) {
693 static inline bool is_meta(struct ir3_instruction
*instr
)
695 /* TODO how should we count PHI (and maybe fan-in/out) which
696 * might actually contribute some instructions to the final
699 return (opc_cat(instr
->opc
) == -1);
702 static inline bool writes_addr(struct ir3_instruction
*instr
)
704 if (instr
->regs_count
> 0) {
705 struct ir3_register
*dst
= instr
->regs
[0];
706 return reg_num(dst
) == REG_A0
;
711 static inline bool writes_pred(struct ir3_instruction
*instr
)
713 if (instr
->regs_count
> 0) {
714 struct ir3_register
*dst
= instr
->regs
[0];
715 return reg_num(dst
) == REG_P0
;
720 /* returns defining instruction for reg */
721 /* TODO better name */
722 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
724 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
725 debug_assert(!(reg
->instr
&& (reg
->instr
->flags
& IR3_INSTR_UNUSED
)));
731 static inline bool conflicts(struct ir3_instruction
*a
,
732 struct ir3_instruction
*b
)
734 return (a
&& b
) && (a
!= b
);
737 static inline bool reg_gpr(struct ir3_register
*r
)
739 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
741 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
746 static inline type_t
half_type(type_t type
)
749 case TYPE_F32
: return TYPE_F16
;
750 case TYPE_U32
: return TYPE_U16
;
751 case TYPE_S32
: return TYPE_S16
;
762 /* some cat2 instructions (ie. those which are not float) can embed an
765 static inline bool ir3_cat2_int(opc_t opc
)
806 /* map cat2 instruction to valid abs/neg flags: */
807 static inline unsigned ir3_cat2_absneg(opc_t opc
)
824 return IR3_REG_FABS
| IR3_REG_FNEG
;
845 return IR3_REG_SABS
| IR3_REG_SNEG
;
866 /* map cat3 instructions to valid abs/neg flags: */
867 static inline unsigned ir3_cat3_absneg(opc_t opc
)
886 /* neg *may* work on 3rd src.. */
896 #define MASK(n) ((1 << (n)) - 1)
898 /* iterator for an instructions's sources (reg), also returns src #: */
899 #define foreach_src_n(__srcreg, __n, __instr) \
900 if ((__instr)->regs_count) \
901 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
902 if ((__srcreg = (__instr)->regs[__n + 1]))
904 /* iterator for an instructions's sources (reg): */
905 #define foreach_src(__srcreg, __instr) \
906 foreach_src_n(__srcreg, __i, __instr)
908 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
911 return instr
->regs_count
+ 1;
912 return instr
->regs_count
;
915 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
917 if (n
== (instr
->regs_count
+ 0))
918 return instr
->address
;
919 return ssa(instr
->regs
[n
]);
922 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
924 /* iterator for an instruction's SSA sources (instr), also returns src #: */
925 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
926 if ((__instr)->regs_count) \
927 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
928 if ((__srcinst = __ssa_src_n(__instr, __n)))
930 /* iterator for an instruction's SSA sources (instr): */
931 #define foreach_ssa_src(__srcinst, __instr) \
932 foreach_ssa_src_n(__srcinst, __i, __instr)
936 void ir3_print(struct ir3
*ir
);
937 void ir3_print_instr(struct ir3_instruction
*instr
);
939 /* depth calculation: */
940 int ir3_delayslots(struct ir3_instruction
*assigner
,
941 struct ir3_instruction
*consumer
, unsigned n
);
942 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
943 void ir3_depth(struct ir3
*ir
);
945 /* copy-propagate: */
946 struct ir3_shader_variant
;
947 void ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
949 /* group neighbors and insert mov's to resolve conflicts: */
950 void ir3_group(struct ir3
*ir
);
953 int ir3_sched(struct ir3
*ir
);
955 /* register assignment: */
956 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(void *memctx
);
957 int ir3_ra(struct ir3
*ir3
, enum shader_t type
,
958 bool frag_coord
, bool frag_face
);
961 void ir3_legalize(struct ir3
*ir
, bool *has_samp
, bool *has_ssbo
, int *max_bary
);
963 /* ************************************************************************* */
964 /* instruction helpers */
966 static inline struct ir3_instruction
*
967 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
969 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
970 ir3_reg_create(instr
, 0, 0); /* dst */
971 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
972 struct ir3_register
*src_reg
=
973 ir3_reg_create(instr
, 0, IR3_REG_ARRAY
);
974 src_reg
->array
= src
->regs
[0]->array
;
975 src_reg
->instr
= src
;
977 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
979 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
980 instr
->cat1
.src_type
= type
;
981 instr
->cat1
.dst_type
= type
;
985 static inline struct ir3_instruction
*
986 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
987 type_t src_type
, type_t dst_type
)
989 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
990 ir3_reg_create(instr
, 0, 0); /* dst */
991 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
992 instr
->cat1
.src_type
= src_type
;
993 instr
->cat1
.dst_type
= dst_type
;
994 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
998 static inline struct ir3_instruction
*
999 ir3_NOP(struct ir3_block
*block
)
1001 return ir3_instr_create(block
, OPC_NOP
);
1004 #define INSTR0(name) \
1005 static inline struct ir3_instruction * \
1006 ir3_##name(struct ir3_block *block) \
1008 struct ir3_instruction *instr = \
1009 ir3_instr_create(block, OPC_##name); \
1013 #define INSTR1(name) \
1014 static inline struct ir3_instruction * \
1015 ir3_##name(struct ir3_block *block, \
1016 struct ir3_instruction *a, unsigned aflags) \
1018 struct ir3_instruction *instr = \
1019 ir3_instr_create(block, OPC_##name); \
1020 ir3_reg_create(instr, 0, 0); /* dst */ \
1021 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1025 #define INSTR2(name) \
1026 static inline struct ir3_instruction * \
1027 ir3_##name(struct ir3_block *block, \
1028 struct ir3_instruction *a, unsigned aflags, \
1029 struct ir3_instruction *b, unsigned bflags) \
1031 struct ir3_instruction *instr = \
1032 ir3_instr_create(block, OPC_##name); \
1033 ir3_reg_create(instr, 0, 0); /* dst */ \
1034 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1035 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1039 #define INSTR3(name) \
1040 static inline struct ir3_instruction * \
1041 ir3_##name(struct ir3_block *block, \
1042 struct ir3_instruction *a, unsigned aflags, \
1043 struct ir3_instruction *b, unsigned bflags, \
1044 struct ir3_instruction *c, unsigned cflags) \
1046 struct ir3_instruction *instr = \
1047 ir3_instr_create(block, OPC_##name); \
1048 ir3_reg_create(instr, 0, 0); /* dst */ \
1049 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1050 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1051 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1055 #define INSTR4(name) \
1056 static inline struct ir3_instruction * \
1057 ir3_##name(struct ir3_block *block, \
1058 struct ir3_instruction *a, unsigned aflags, \
1059 struct ir3_instruction *b, unsigned bflags, \
1060 struct ir3_instruction *c, unsigned cflags, \
1061 struct ir3_instruction *d, unsigned dflags) \
1063 struct ir3_instruction *instr = \
1064 ir3_instr_create2(block, OPC_##name, 5); \
1065 ir3_reg_create(instr, 0, 0); /* dst */ \
1066 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1067 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1068 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1069 ir3_reg_create(instr, 0, IR3_REG_SSA | dflags)->instr = d; \
1073 #define INSTR4F(f, name) \
1074 static inline struct ir3_instruction * \
1075 ir3_##name##_##f(struct ir3_block *block, \
1076 struct ir3_instruction *a, unsigned aflags, \
1077 struct ir3_instruction *b, unsigned bflags, \
1078 struct ir3_instruction *c, unsigned cflags, \
1079 struct ir3_instruction *d, unsigned dflags) \
1081 struct ir3_instruction *instr = \
1082 ir3_instr_create2(block, OPC_##name, 5); \
1083 ir3_reg_create(instr, 0, 0); /* dst */ \
1084 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1085 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1086 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1087 ir3_reg_create(instr, 0, IR3_REG_SSA | dflags)->instr = d; \
1088 instr->flags |= IR3_INSTR_##f; \
1092 /* cat0 instructions: */
1098 /* cat2 instructions, most 2 src but some 1 src: */
1146 /* cat3 instructions: */
1164 /* cat4 instructions: */
1173 /* cat5 instructions: */
1177 static inline struct ir3_instruction
*
1178 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1179 unsigned wrmask
, unsigned flags
, unsigned samp
, unsigned tex
,
1180 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1182 struct ir3_instruction
*sam
;
1183 struct ir3_register
*reg
;
1185 sam
= ir3_instr_create(block
, opc
);
1186 sam
->flags
|= flags
;
1187 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1189 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1190 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1194 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1196 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1198 sam
->cat5
.samp
= samp
;
1199 sam
->cat5
.tex
= tex
;
1200 sam
->cat5
.type
= type
;
1205 /* cat6 instructions: */
1221 INSTR2(ATOMIC_CMPXCHG
)
1227 INSTR4F(G
, ATOMIC_ADD
)
1228 INSTR4F(G
, ATOMIC_SUB
)
1229 INSTR4F(G
, ATOMIC_XCHG
)
1230 INSTR4F(G
, ATOMIC_INC
)
1231 INSTR4F(G
, ATOMIC_DEC
)
1232 INSTR4F(G
, ATOMIC_CMPXCHG
)
1233 INSTR4F(G
, ATOMIC_MIN
)
1234 INSTR4F(G
, ATOMIC_MAX
)
1235 INSTR4F(G
, ATOMIC_AND
)
1236 INSTR4F(G
, ATOMIC_OR
)
1237 INSTR4F(G
, ATOMIC_XOR
)
1239 /* cat7 instructions: */
1243 /* ************************************************************************* */
1244 /* split this out or find some helper to use.. like main/bitset.h.. */
1250 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1252 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1254 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1255 debug_assert(num
< MAX_REG
);
1256 if (reg
->flags
& IR3_REG_HALF
)
1261 static inline void regmask_init(regmask_t
*regmask
)
1263 memset(regmask
, 0, sizeof(*regmask
));
1266 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1268 unsigned idx
= regmask_idx(reg
);
1269 if (reg
->flags
& IR3_REG_RELATIV
) {
1271 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1272 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1275 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1277 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1281 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1284 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1285 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1288 /* set bits in a if not set in b, conceptually:
1291 static inline void regmask_set_if_not(regmask_t
*a
,
1292 struct ir3_register
*reg
, regmask_t
*b
)
1294 unsigned idx
= regmask_idx(reg
);
1295 if (reg
->flags
& IR3_REG_RELATIV
) {
1297 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1298 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1299 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1302 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1304 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1305 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1309 static inline bool regmask_get(regmask_t
*regmask
,
1310 struct ir3_register
*reg
)
1312 unsigned idx
= regmask_idx(reg
);
1313 if (reg
->flags
& IR3_REG_RELATIV
) {
1315 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1316 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1320 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1322 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1328 /* ************************************************************************* */