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 IR3_REG_RELATIV
= 0x008,
63 /* Most instructions, it seems, can do float abs/neg but not
64 * integer. The CP pass needs to know what is intended (int or
65 * float) in order to do the right thing. For this reason the
66 * abs/neg flags are split out into float and int variants. In
67 * addition, .b (bitwise) operations, the negate is actually a
68 * bitwise not, so split that out into a new flag to make it
77 IR3_REG_POS_INF
= 0x800,
78 /* (ei) flag, end-input? Set on last bary, presumably to signal
79 * that the shader needs no more input:
82 /* meta-flags, for intermediate stages of IR, ie.
83 * before register assignment is done:
85 IR3_REG_SSA
= 0x2000, /* 'instr' is ptr to assigning instr */
86 IR3_REG_ARRAY
= 0x4000,
87 IR3_REG_PHI_SRC
= 0x8000, /* phi src, regs[0]->instr points to phi */
92 * the component is in the low two bits of the reg #, so
93 * rN.x becomes: (N << 2) | x
107 /* For IR3_REG_SSA, src registers contain ptr back to assigning
110 * For IR3_REG_ARRAY, the pointer is back to the last dependent
111 * array access (although the net effect is the same, it points
112 * back to a previous instruction that we depend on).
114 struct ir3_instruction
*instr
;
117 /* used for cat5 instructions, but also for internal/IR level
118 * tracking of what registers are read/written by an instruction.
119 * wrmask may be a bad name since it is used to represent both
120 * src and dst that touch multiple adjacent registers.
123 /* for relative addressing, 32bits for array size is too small,
124 * but otoh we don't need to deal with disjoint sets, so instead
125 * use a simple size field (number of scalar components).
131 struct ir3_instruction
{
132 struct ir3_block
*block
;
135 /* (sy) flag is set on first instruction, and after sample
136 * instructions (probably just on RAW hazard).
138 IR3_INSTR_SY
= 0x001,
139 /* (ss) flag is set on first instruction, and first instruction
140 * to depend on the result of "long" instructions (RAW hazard):
142 * rcp, rsq, log2, exp2, sin, cos, sqrt
144 * It seems to synchronize until all in-flight instructions are
145 * completed, for example:
148 * add.f hr2.z, (neg)hr2.z, hc0.y
149 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
152 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
154 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
155 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
156 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
158 * The last mul.f does not have (ss) set, presumably because the
159 * (ss) on the previous instruction does the job.
161 * The blob driver also seems to set it on WAR hazards, although
162 * not really clear if this is needed or just blob compiler being
163 * sloppy. So far I haven't found a case where removing the (ss)
164 * causes problems for WAR hazard, but I could just be getting
168 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
171 IR3_INSTR_SS
= 0x002,
172 /* (jp) flag is set on jump targets:
174 IR3_INSTR_JP
= 0x004,
175 IR3_INSTR_UL
= 0x008,
176 IR3_INSTR_3D
= 0x010,
181 IR3_INSTR_S2EN
= 0x200,
183 /* meta-flags, for intermediate stages of IR, ie.
184 * before register assignment is done:
186 IR3_INSTR_MARK
= 0x1000,
187 IR3_INSTR_UNUSED
= 0x2000,
194 struct ir3_register
**regs
;
200 struct ir3_block
*target
;
203 type_t src_type
, dst_type
;
225 /* for meta-instructions, just used to hold extra data
226 * before instruction scheduling, etc
229 int off
; /* component/offset */
232 /* used to temporarily hold reference to nir_phi_instr
233 * until we resolve the phi srcs
238 struct ir3_block
*block
;
242 /* transient values used during various algorithms: */
244 /* The instruction depth is the max dependency distance to output.
246 * You can also think of it as the "cost", if we did any sort of
247 * optimization for register footprint. Ie. a value that is just
248 * result of moving a const to a reg would have a low cost, so to
249 * it could make sense to duplicate the instruction at various
250 * points where the result is needed to reduce register footprint.
253 /* When we get to the RA stage, we no longer need depth, but
254 * we do need instruction's position/name:
262 /* used for per-pass extra instruction data.
266 /* Used during CP and RA stages. For fanin and shader inputs/
267 * outputs where we need a sequence of consecutive registers,
268 * keep track of each src instructions left (ie 'n-1') and right
269 * (ie 'n+1') neighbor. The front-end must insert enough mov's
270 * to ensure that each instruction has at most one left and at
271 * most one right neighbor. During the copy-propagation pass,
272 * we only remove mov's when we can preserve this constraint.
273 * And during the RA stage, we use the neighbor information to
274 * allocate a block of registers in one shot.
276 * TODO: maybe just add something like:
277 * struct ir3_instruction_ref {
278 * struct ir3_instruction *instr;
282 * Or can we get away without the refcnt stuff? It seems like
283 * it should be overkill.. the problem is if, potentially after
284 * already eliminating some mov's, if you have a single mov that
285 * needs to be grouped with it's neighbors in two different
286 * places (ex. shader output and a fanin).
289 struct ir3_instruction
*left
, *right
;
290 uint16_t left_cnt
, right_cnt
;
293 /* an instruction can reference at most one address register amongst
294 * it's src/dst registers. Beyond that, you need to insert mov's.
296 * NOTE: do not write this directly, use ir3_instr_set_address()
298 struct ir3_instruction
*address
;
300 /* Entry in ir3_block's instruction list: */
301 struct list_head node
;
308 static inline struct ir3_instruction
*
309 ir3_neighbor_first(struct ir3_instruction
*instr
)
311 while (instr
->cp
.left
)
312 instr
= instr
->cp
.left
;
316 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
320 debug_assert(!instr
->cp
.left
);
322 while (instr
->cp
.right
) {
324 instr
= instr
->cp
.right
;
330 struct ir3_heap_chunk
;
333 struct ir3_compiler
*compiler
;
335 unsigned ninputs
, noutputs
;
336 struct ir3_instruction
**inputs
;
337 struct ir3_instruction
**outputs
;
339 /* Track bary.f (and ldlv) instructions.. this is needed in
340 * scheduling to ensure that all varying fetches happen before
341 * any potential kill instructions. The hw gets grumpy if all
342 * threads in a group are killed before the last bary.f gets
343 * a chance to signal end of input (ei).
345 unsigned baryfs_count
, baryfs_sz
;
346 struct ir3_instruction
**baryfs
;
348 /* Track all indirect instructions (read and write). To avoid
349 * deadlock scenario where an address register gets scheduled,
350 * but other dependent src instructions cannot be scheduled due
351 * to dependency on a *different* address register value, the
352 * scheduler needs to ensure that all dependencies other than
353 * the instruction other than the address register are scheduled
354 * before the one that writes the address register. Having a
355 * convenient list of instructions that reference some address
356 * register simplifies this.
358 unsigned indirects_count
, indirects_sz
;
359 struct ir3_instruction
**indirects
;
360 /* and same for instructions that consume predicate register: */
361 unsigned predicates_count
, predicates_sz
;
362 struct ir3_instruction
**predicates
;
364 /* Track instructions which do not write a register but other-
365 * wise must not be discarded (such as kill, stg, etc)
367 unsigned keeps_count
, keeps_sz
;
368 struct ir3_instruction
**keeps
;
370 /* List of blocks: */
371 struct list_head block_list
;
373 /* List of ir3_array's: */
374 struct list_head array_list
;
377 struct ir3_heap_chunk
*chunk
;
380 typedef struct nir_variable nir_variable
;
383 struct list_head node
;
389 /* We track the last write and last access (read or write) to
390 * setup dependencies on instructions that read or write the
391 * array. Reads can be re-ordered wrt. other reads, but should
392 * not be re-ordered wrt. to writes. Writes cannot be reordered
393 * wrt. any other access to the array.
395 * So array reads depend on last write, and array writes depend
396 * on the last access.
398 struct ir3_instruction
*last_write
, *last_access
;
400 /* extra stuff used in RA pass: */
401 unsigned base
; /* base vreg name */
402 unsigned reg
; /* base physical reg */
403 uint16_t start_ip
, end_ip
;
406 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
408 typedef struct nir_block nir_block
;
411 struct list_head node
;
416 struct list_head instr_list
; /* list of ir3_instruction */
418 /* each block has either one or two successors.. in case of
419 * two successors, 'condition' decides which one to follow.
420 * A block preceding an if/else has two successors.
422 struct ir3_instruction
*condition
;
423 struct ir3_block
*successors
[2];
425 uint16_t start_ip
, end_ip
;
427 /* used for per-pass extra block data. Mainly used right
428 * now in RA step to track livein/liveout.
437 static inline uint32_t
438 block_id(struct ir3_block
*block
)
441 return block
->serialno
;
443 return (uint32_t)(unsigned long)block
;
447 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
448 unsigned nin
, unsigned nout
);
449 void ir3_destroy(struct ir3
*shader
);
450 void * ir3_assemble(struct ir3
*shader
,
451 struct ir3_info
*info
, uint32_t gpu_id
);
452 void * ir3_alloc(struct ir3
*shader
, int sz
);
454 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
456 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
457 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
458 opc_t opc
, int nreg
);
459 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
460 const char *ir3_instr_name(struct ir3_instruction
*instr
);
462 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
464 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
465 struct ir3_register
*reg
);
467 void ir3_instr_set_address(struct ir3_instruction
*instr
,
468 struct ir3_instruction
*addr
);
470 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
472 if (instr
->flags
& IR3_INSTR_MARK
)
473 return true; /* already visited */
474 instr
->flags
|= IR3_INSTR_MARK
;
478 void ir3_block_clear_mark(struct ir3_block
*block
);
479 void ir3_clear_mark(struct ir3
*shader
);
481 unsigned ir3_count_instructions(struct ir3
*ir
);
483 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
484 struct ir3_register
*reg
)
487 for (i
= 0; i
< instr
->regs_count
; i
++)
488 if (reg
== instr
->regs
[i
])
494 #define MAX_ARRAYS 16
502 static inline uint32_t regid(int num
, int comp
)
504 return (num
<< 2) | (comp
& 0x3);
507 static inline uint32_t reg_num(struct ir3_register
*reg
)
509 return reg
->num
>> 2;
512 static inline uint32_t reg_comp(struct ir3_register
*reg
)
514 return reg
->num
& 0x3;
517 static inline bool is_flow(struct ir3_instruction
*instr
)
519 return (opc_cat(instr
->opc
) == 0);
522 static inline bool is_kill(struct ir3_instruction
*instr
)
524 return instr
->opc
== OPC_KILL
;
527 static inline bool is_nop(struct ir3_instruction
*instr
)
529 return instr
->opc
== OPC_NOP
;
532 /* Is it a non-transformative (ie. not type changing) mov? This can
533 * also include absneg.s/absneg.f, which for the most part can be
534 * treated as a mov (single src argument).
536 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
538 struct ir3_register
*dst
= instr
->regs
[0];
540 /* mov's that write to a0.x or p0.x are special: */
541 if (dst
->num
== regid(REG_P0
, 0))
543 if (dst
->num
== regid(REG_A0
, 0))
546 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
549 switch (instr
->opc
) {
551 return instr
->cat1
.src_type
== instr
->cat1
.dst_type
;
560 static inline bool is_alu(struct ir3_instruction
*instr
)
562 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
565 static inline bool is_sfu(struct ir3_instruction
*instr
)
567 return (opc_cat(instr
->opc
) == 4);
570 static inline bool is_tex(struct ir3_instruction
*instr
)
572 return (opc_cat(instr
->opc
) == 5);
575 static inline bool is_mem(struct ir3_instruction
*instr
)
577 return (opc_cat(instr
->opc
) == 6);
581 is_store(struct ir3_instruction
*instr
)
583 /* these instructions, the "destination" register is
584 * actually a source, the address to store to.
586 switch (instr
->opc
) {
599 static inline bool is_load(struct ir3_instruction
*instr
)
601 switch (instr
->opc
) {
609 /* probably some others too.. */
616 static inline bool is_input(struct ir3_instruction
*instr
)
618 /* in some cases, ldlv is used to fetch varying without
619 * interpolation.. fortunately inloc is the first src
620 * register in either case
622 switch (instr
->opc
) {
631 static inline bool is_meta(struct ir3_instruction
*instr
)
633 /* TODO how should we count PHI (and maybe fan-in/out) which
634 * might actually contribute some instructions to the final
637 return (opc_cat(instr
->opc
) == -1);
640 static inline bool writes_addr(struct ir3_instruction
*instr
)
642 if (instr
->regs_count
> 0) {
643 struct ir3_register
*dst
= instr
->regs
[0];
644 return reg_num(dst
) == REG_A0
;
649 static inline bool writes_pred(struct ir3_instruction
*instr
)
651 if (instr
->regs_count
> 0) {
652 struct ir3_register
*dst
= instr
->regs
[0];
653 return reg_num(dst
) == REG_P0
;
658 /* returns defining instruction for reg */
659 /* TODO better name */
660 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
662 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
663 debug_assert(!(reg
->instr
&& (reg
->instr
->flags
& IR3_INSTR_UNUSED
)));
669 static inline bool conflicts(struct ir3_instruction
*a
,
670 struct ir3_instruction
*b
)
672 return (a
&& b
) && (a
!= b
);
675 static inline bool reg_gpr(struct ir3_register
*r
)
677 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
679 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
684 static inline type_t
half_type(type_t type
)
687 case TYPE_F32
: return TYPE_F16
;
688 case TYPE_U32
: return TYPE_U16
;
689 case TYPE_S32
: return TYPE_S16
;
700 /* some cat2 instructions (ie. those which are not float) can embed an
703 static inline bool ir3_cat2_int(opc_t opc
)
744 /* map cat2 instruction to valid abs/neg flags: */
745 static inline unsigned ir3_cat2_absneg(opc_t opc
)
762 return IR3_REG_FABS
| IR3_REG_FNEG
;
783 return IR3_REG_SABS
| IR3_REG_SNEG
;
804 /* map cat3 instructions to valid abs/neg flags: */
805 static inline unsigned ir3_cat3_absneg(opc_t opc
)
824 /* neg *may* work on 3rd src.. */
834 #define array_insert(arr, val) do { \
835 if (arr ## _count == arr ## _sz) { \
836 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
837 arr = realloc(arr, arr ## _sz * sizeof(arr[0])); \
839 arr[arr ##_count++] = val; \
842 /* iterator for an instructions's sources (reg), also returns src #: */
843 #define foreach_src_n(__srcreg, __n, __instr) \
844 if ((__instr)->regs_count) \
845 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
846 if ((__srcreg = (__instr)->regs[__n + 1]))
848 /* iterator for an instructions's sources (reg): */
849 #define foreach_src(__srcreg, __instr) \
850 foreach_src_n(__srcreg, __i, __instr)
852 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
855 return instr
->regs_count
+ 1;
856 return instr
->regs_count
;
859 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
861 if (n
== (instr
->regs_count
+ 0))
862 return instr
->address
;
863 return ssa(instr
->regs
[n
]);
866 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
868 /* iterator for an instruction's SSA sources (instr), also returns src #: */
869 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
870 if ((__instr)->regs_count) \
871 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
872 if ((__srcinst = __ssa_src_n(__instr, __n)))
874 /* iterator for an instruction's SSA sources (instr): */
875 #define foreach_ssa_src(__srcinst, __instr) \
876 foreach_ssa_src_n(__srcinst, __i, __instr)
880 void ir3_print(struct ir3
*ir
);
881 void ir3_print_instr(struct ir3_instruction
*instr
);
883 /* depth calculation: */
884 int ir3_delayslots(struct ir3_instruction
*assigner
,
885 struct ir3_instruction
*consumer
, unsigned n
);
886 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
887 void ir3_depth(struct ir3
*ir
);
889 /* copy-propagate: */
890 void ir3_cp(struct ir3
*ir
);
892 /* group neighbors and insert mov's to resolve conflicts: */
893 void ir3_group(struct ir3
*ir
);
896 int ir3_sched(struct ir3
*ir
);
898 /* register assignment: */
899 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(void *memctx
);
900 int ir3_ra(struct ir3
*ir3
, enum shader_t type
,
901 bool frag_coord
, bool frag_face
);
904 void ir3_legalize(struct ir3
*ir
, bool *has_samp
, int *max_bary
);
906 /* ************************************************************************* */
907 /* instruction helpers */
909 static inline struct ir3_instruction
*
910 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
912 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
913 ir3_reg_create(instr
, 0, 0); /* dst */
914 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
915 struct ir3_register
*src_reg
=
916 ir3_reg_create(instr
, 0, IR3_REG_ARRAY
);
917 src_reg
->array
= src
->regs
[0]->array
;
918 src_reg
->instr
= src
;
920 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
922 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
923 instr
->cat1
.src_type
= type
;
924 instr
->cat1
.dst_type
= type
;
928 static inline struct ir3_instruction
*
929 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
930 type_t src_type
, type_t dst_type
)
932 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
933 ir3_reg_create(instr
, 0, 0); /* dst */
934 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
935 instr
->cat1
.src_type
= src_type
;
936 instr
->cat1
.dst_type
= dst_type
;
937 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
941 static inline struct ir3_instruction
*
942 ir3_NOP(struct ir3_block
*block
)
944 return ir3_instr_create(block
, OPC_NOP
);
947 #define INSTR0(name) \
948 static inline struct ir3_instruction * \
949 ir3_##name(struct ir3_block *block) \
951 struct ir3_instruction *instr = \
952 ir3_instr_create(block, OPC_##name); \
956 #define INSTR1(name) \
957 static inline struct ir3_instruction * \
958 ir3_##name(struct ir3_block *block, \
959 struct ir3_instruction *a, unsigned aflags) \
961 struct ir3_instruction *instr = \
962 ir3_instr_create(block, OPC_##name); \
963 ir3_reg_create(instr, 0, 0); /* dst */ \
964 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
968 #define INSTR2(name) \
969 static inline struct ir3_instruction * \
970 ir3_##name(struct ir3_block *block, \
971 struct ir3_instruction *a, unsigned aflags, \
972 struct ir3_instruction *b, unsigned bflags) \
974 struct ir3_instruction *instr = \
975 ir3_instr_create(block, OPC_##name); \
976 ir3_reg_create(instr, 0, 0); /* dst */ \
977 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
978 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
982 #define INSTR3(name) \
983 static inline struct ir3_instruction * \
984 ir3_##name(struct ir3_block *block, \
985 struct ir3_instruction *a, unsigned aflags, \
986 struct ir3_instruction *b, unsigned bflags, \
987 struct ir3_instruction *c, unsigned cflags) \
989 struct ir3_instruction *instr = \
990 ir3_instr_create(block, OPC_##name); \
991 ir3_reg_create(instr, 0, 0); /* dst */ \
992 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
993 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
994 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
998 /* cat0 instructions: */
1004 /* cat2 instructions, most 2 src but some 1 src: */
1052 /* cat3 instructions: */
1070 /* cat4 instructions: */
1079 /* cat5 instructions: */
1083 static inline struct ir3_instruction
*
1084 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1085 unsigned wrmask
, unsigned flags
, unsigned samp
, unsigned tex
,
1086 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1088 struct ir3_instruction
*sam
;
1089 struct ir3_register
*reg
;
1091 sam
= ir3_instr_create(block
, opc
);
1092 sam
->flags
|= flags
;
1093 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1095 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1096 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1100 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1102 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1104 sam
->cat5
.samp
= samp
;
1105 sam
->cat5
.tex
= tex
;
1106 sam
->cat5
.type
= type
;
1111 /* cat6 instructions: */
1116 /* ************************************************************************* */
1117 /* split this out or find some helper to use.. like main/bitset.h.. */
1123 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1125 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1127 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1128 debug_assert(num
< MAX_REG
);
1129 if (reg
->flags
& IR3_REG_HALF
)
1134 static inline void regmask_init(regmask_t
*regmask
)
1136 memset(regmask
, 0, sizeof(*regmask
));
1139 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1141 unsigned idx
= regmask_idx(reg
);
1142 if (reg
->flags
& IR3_REG_RELATIV
) {
1144 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1145 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1148 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1150 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1154 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1157 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1158 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1161 /* set bits in a if not set in b, conceptually:
1164 static inline void regmask_set_if_not(regmask_t
*a
,
1165 struct ir3_register
*reg
, regmask_t
*b
)
1167 unsigned idx
= regmask_idx(reg
);
1168 if (reg
->flags
& IR3_REG_RELATIV
) {
1170 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1171 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1172 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1175 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1177 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1178 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1182 static inline bool regmask_get(regmask_t
*regmask
,
1183 struct ir3_register
*reg
)
1185 unsigned idx
= regmask_idx(reg
);
1186 if (reg
->flags
& IR3_REG_RELATIV
) {
1188 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1189 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1193 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1195 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1201 /* ************************************************************************* */