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 "compiler/shader_enums.h"
32 #include "util/u_debug.h"
33 #include "util/list.h"
35 #include "instr-a3xx.h"
37 /* low level intermediate representation of an adreno shader program */
41 struct ir3_instruction
;
47 uint16_t instrs_count
; /* expanded to account for rpt's */
48 /* NOTE: max_reg, etc, does not include registers not touched
49 * by the shader (ie. vertex fetched via VFD_DECODE but not
52 int8_t max_reg
; /* highest GPR # used by shader */
56 /* number of sync bits: */
62 IR3_REG_CONST
= 0x001,
63 IR3_REG_IMMED
= 0x002,
65 /* high registers are used for some things in compute shaders,
66 * for example. Seems to be for things that are global to all
67 * threads in a wave, so possibly these are global/shared by
68 * all the threads in the wave?
71 IR3_REG_RELATIV
= 0x010,
73 /* Most instructions, it seems, can do float abs/neg but not
74 * integer. The CP pass needs to know what is intended (int or
75 * float) in order to do the right thing. For this reason the
76 * abs/neg flags are split out into float and int variants. In
77 * addition, .b (bitwise) operations, the negate is actually a
78 * bitwise not, so split that out into a new flag to make it
87 IR3_REG_POS_INF
= 0x1000,
88 /* (ei) flag, end-input? Set on last bary, presumably to signal
89 * that the shader needs no more input:
92 /* meta-flags, for intermediate stages of IR, ie.
93 * before register assignment is done:
95 IR3_REG_SSA
= 0x4000, /* 'instr' is ptr to assigning instr */
96 IR3_REG_ARRAY
= 0x8000,
101 * the component is in the low two bits of the reg #, so
102 * rN.x becomes: (N << 2) | x
117 /* For IR3_REG_SSA, src registers contain ptr back to assigning
120 * For IR3_REG_ARRAY, the pointer is back to the last dependent
121 * array access (although the net effect is the same, it points
122 * back to a previous instruction that we depend on).
124 struct ir3_instruction
*instr
;
127 /* used for cat5 instructions, but also for internal/IR level
128 * tracking of what registers are read/written by an instruction.
129 * wrmask may be a bad name since it is used to represent both
130 * src and dst that touch multiple adjacent registers.
133 /* for relative addressing, 32bits for array size is too small,
134 * but otoh we don't need to deal with disjoint sets, so instead
135 * use a simple size field (number of scalar components).
142 * Stupid/simple growable array implementation:
144 #define DECLARE_ARRAY(type, name) \
145 unsigned name ## _count, name ## _sz; \
148 #define array_insert(ctx, arr, val) do { \
149 if (arr ## _count == arr ## _sz) { \
150 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
151 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
153 arr[arr ##_count++] = val; \
156 struct ir3_instruction
{
157 struct ir3_block
*block
;
160 /* (sy) flag is set on first instruction, and after sample
161 * instructions (probably just on RAW hazard).
163 IR3_INSTR_SY
= 0x001,
164 /* (ss) flag is set on first instruction, and first instruction
165 * to depend on the result of "long" instructions (RAW hazard):
167 * rcp, rsq, log2, exp2, sin, cos, sqrt
169 * It seems to synchronize until all in-flight instructions are
170 * completed, for example:
173 * add.f hr2.z, (neg)hr2.z, hc0.y
174 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
177 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
179 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
180 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
181 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
183 * The last mul.f does not have (ss) set, presumably because the
184 * (ss) on the previous instruction does the job.
186 * The blob driver also seems to set it on WAR hazards, although
187 * not really clear if this is needed or just blob compiler being
188 * sloppy. So far I haven't found a case where removing the (ss)
189 * causes problems for WAR hazard, but I could just be getting
193 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
196 IR3_INSTR_SS
= 0x002,
197 /* (jp) flag is set on jump targets:
199 IR3_INSTR_JP
= 0x004,
200 IR3_INSTR_UL
= 0x008,
201 IR3_INSTR_3D
= 0x010,
206 IR3_INSTR_S2EN
= 0x200,
208 IR3_INSTR_SAT
= 0x800,
209 /* meta-flags, for intermediate stages of IR, ie.
210 * before register assignment is done:
212 IR3_INSTR_MARK
= 0x1000,
213 IR3_INSTR_UNUSED
= 0x2000,
221 struct ir3_register
**regs
;
227 struct ir3_block
*target
;
230 type_t src_type
, dst_type
;
250 int iim_val
: 3; /* for ldgb/stgb, # of components */
255 unsigned w
: 1; /* write */
256 unsigned r
: 1; /* read */
257 unsigned l
: 1; /* local */
258 unsigned g
: 1; /* global */
260 /* for meta-instructions, just used to hold extra data
261 * before instruction scheduling, etc
264 int off
; /* component/offset */
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 /* Tracking for additional dependent instructions. Used to handle
330 * barriers, WAR hazards for arrays/SSBOs/etc.
332 DECLARE_ARRAY(struct ir3_instruction
*, deps
);
335 * From PoV of instruction scheduling, not execution (ie. ignores global/
336 * local distinction):
337 * shared image atomic SSBO everything
338 * barrier()/ - R/W R/W R/W R/W X
339 * groupMemoryBarrier()
340 * memoryBarrier() - R/W R/W
341 * (but only images declared coherent?)
342 * memoryBarrierAtomic() - R/W
343 * memoryBarrierBuffer() - R/W
344 * memoryBarrierImage() - R/W
345 * memoryBarrierShared() - R/W
347 * TODO I think for SSBO/image/shared, in cases where we can determine
348 * which variable is accessed, we don't need to care about accesses to
349 * different variables (unless declared coherent??)
352 IR3_BARRIER_EVERYTHING
= 1 << 0,
353 IR3_BARRIER_SHARED_R
= 1 << 1,
354 IR3_BARRIER_SHARED_W
= 1 << 2,
355 IR3_BARRIER_IMAGE_R
= 1 << 3,
356 IR3_BARRIER_IMAGE_W
= 1 << 4,
357 IR3_BARRIER_BUFFER_R
= 1 << 5,
358 IR3_BARRIER_BUFFER_W
= 1 << 6,
359 IR3_BARRIER_ARRAY_R
= 1 << 7,
360 IR3_BARRIER_ARRAY_W
= 1 << 8,
361 } barrier_class
, barrier_conflict
;
363 /* Entry in ir3_block's instruction list: */
364 struct list_head node
;
366 int use_count
; /* currently just updated/used by cp */
373 static inline struct ir3_instruction
*
374 ir3_neighbor_first(struct ir3_instruction
*instr
)
377 while (instr
->cp
.left
) {
378 instr
= instr
->cp
.left
;
379 if (++cnt
> 0xffff) {
387 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
391 debug_assert(!instr
->cp
.left
);
393 while (instr
->cp
.right
) {
395 instr
= instr
->cp
.right
;
406 struct ir3_compiler
*compiler
;
408 unsigned ninputs
, noutputs
;
409 struct ir3_instruction
**inputs
;
410 struct ir3_instruction
**outputs
;
412 /* Track bary.f (and ldlv) instructions.. this is needed in
413 * scheduling to ensure that all varying fetches happen before
414 * any potential kill instructions. The hw gets grumpy if all
415 * threads in a group are killed before the last bary.f gets
416 * a chance to signal end of input (ei).
418 DECLARE_ARRAY(struct ir3_instruction
*, baryfs
);
420 /* Track all indirect instructions (read and write). To avoid
421 * deadlock scenario where an address register gets scheduled,
422 * but other dependent src instructions cannot be scheduled due
423 * to dependency on a *different* address register value, the
424 * scheduler needs to ensure that all dependencies other than
425 * the instruction other than the address register are scheduled
426 * before the one that writes the address register. Having a
427 * convenient list of instructions that reference some address
428 * register simplifies this.
430 DECLARE_ARRAY(struct ir3_instruction
*, indirects
);
432 /* and same for instructions that consume predicate register: */
433 DECLARE_ARRAY(struct ir3_instruction
*, predicates
);
435 /* Track texture sample instructions which need texture state
436 * patched in (for astc-srgb workaround):
438 DECLARE_ARRAY(struct ir3_instruction
*, astc_srgb
);
440 /* List of blocks: */
441 struct list_head block_list
;
443 /* List of ir3_array's: */
444 struct list_head array_list
;
447 unsigned block_count
, instr_count
;
452 struct list_head node
;
456 struct nir_register
*r
;
458 /* To avoid array write's from getting DCE'd, keep track of the
459 * most recent write. Any array access depends on the most
460 * recent write. This way, nothing depends on writes after the
461 * last read. But all the writes that happen before that have
462 * something depending on them
464 struct ir3_instruction
*last_write
;
466 /* extra stuff used in RA pass: */
467 unsigned base
; /* base vreg name */
468 unsigned reg
; /* base physical reg */
469 uint16_t start_ip
, end_ip
;
472 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
475 struct list_head node
;
478 const struct nir_block
*nblock
;
480 struct list_head instr_list
; /* list of ir3_instruction */
482 /* each block has either one or two successors.. in case of
483 * two successors, 'condition' decides which one to follow.
484 * A block preceding an if/else has two successors.
486 struct ir3_instruction
*condition
;
487 struct ir3_block
*successors
[2];
489 unsigned predecessors_count
;
490 struct ir3_block
**predecessors
;
492 uint16_t start_ip
, end_ip
;
494 /* Track instructions which do not write a register but other-
495 * wise must not be discarded (such as kill, stg, etc)
497 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
499 /* used for per-pass extra block data. Mainly used right
500 * now in RA step to track livein/liveout.
509 static inline uint32_t
510 block_id(struct ir3_block
*block
)
513 return block
->serialno
;
515 return (uint32_t)(unsigned long)block
;
519 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
520 unsigned nin
, unsigned nout
);
521 void ir3_destroy(struct ir3
*shader
);
522 void * ir3_assemble(struct ir3
*shader
,
523 struct ir3_info
*info
, uint32_t gpu_id
);
524 void * ir3_alloc(struct ir3
*shader
, int sz
);
526 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
528 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
529 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
530 opc_t opc
, int nreg
);
531 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
532 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
);
533 const char *ir3_instr_name(struct ir3_instruction
*instr
);
535 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
537 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
538 struct ir3_register
*reg
);
540 void ir3_instr_set_address(struct ir3_instruction
*instr
,
541 struct ir3_instruction
*addr
);
543 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
545 if (instr
->flags
& IR3_INSTR_MARK
)
546 return true; /* already visited */
547 instr
->flags
|= IR3_INSTR_MARK
;
551 void ir3_block_clear_mark(struct ir3_block
*block
);
552 void ir3_clear_mark(struct ir3
*shader
);
554 unsigned ir3_count_instructions(struct ir3
*ir
);
556 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
557 struct ir3_register
*reg
)
560 for (i
= 0; i
< instr
->regs_count
; i
++)
561 if (reg
== instr
->regs
[i
])
567 #define MAX_ARRAYS 16
575 static inline uint32_t regid(int num
, int comp
)
577 return (num
<< 2) | (comp
& 0x3);
580 static inline uint32_t reg_num(struct ir3_register
*reg
)
582 return reg
->num
>> 2;
585 static inline uint32_t reg_comp(struct ir3_register
*reg
)
587 return reg
->num
& 0x3;
590 static inline bool is_flow(struct ir3_instruction
*instr
)
592 return (opc_cat(instr
->opc
) == 0);
595 static inline bool is_kill(struct ir3_instruction
*instr
)
597 return instr
->opc
== OPC_KILL
;
600 static inline bool is_nop(struct ir3_instruction
*instr
)
602 return instr
->opc
== OPC_NOP
;
605 /* Is it a non-transformative (ie. not type changing) mov? This can
606 * also include absneg.s/absneg.f, which for the most part can be
607 * treated as a mov (single src argument).
609 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
611 struct ir3_register
*dst
;
613 switch (instr
->opc
) {
615 if (instr
->cat1
.src_type
!= instr
->cat1
.dst_type
)
620 if (instr
->flags
& IR3_INSTR_SAT
)
627 dst
= instr
->regs
[0];
629 /* mov's that write to a0.x or p0.x are special: */
630 if (dst
->num
== regid(REG_P0
, 0))
632 if (dst
->num
== regid(REG_A0
, 0))
635 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
641 static inline bool is_alu(struct ir3_instruction
*instr
)
643 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
646 static inline bool is_sfu(struct ir3_instruction
*instr
)
648 return (opc_cat(instr
->opc
) == 4);
651 static inline bool is_tex(struct ir3_instruction
*instr
)
653 return (opc_cat(instr
->opc
) == 5);
656 static inline bool is_mem(struct ir3_instruction
*instr
)
658 return (opc_cat(instr
->opc
) == 6);
661 static inline bool is_barrier(struct ir3_instruction
*instr
)
663 return (opc_cat(instr
->opc
) == 7);
667 is_store(struct ir3_instruction
*instr
)
669 /* these instructions, the "destination" register is
670 * actually a source, the address to store to.
672 switch (instr
->opc
) {
687 static inline bool is_load(struct ir3_instruction
*instr
)
689 switch (instr
->opc
) {
699 /* probably some others too.. */
706 static inline bool is_input(struct ir3_instruction
*instr
)
708 /* in some cases, ldlv is used to fetch varying without
709 * interpolation.. fortunately inloc is the first src
710 * register in either case
712 switch (instr
->opc
) {
721 static inline bool is_bool(struct ir3_instruction
*instr
)
723 switch (instr
->opc
) {
733 static inline bool is_meta(struct ir3_instruction
*instr
)
735 /* TODO how should we count PHI (and maybe fan-in/out) which
736 * might actually contribute some instructions to the final
739 return (opc_cat(instr
->opc
) == -1);
742 static inline bool writes_addr(struct ir3_instruction
*instr
)
744 if (instr
->regs_count
> 0) {
745 struct ir3_register
*dst
= instr
->regs
[0];
746 return reg_num(dst
) == REG_A0
;
751 static inline bool writes_pred(struct ir3_instruction
*instr
)
753 if (instr
->regs_count
> 0) {
754 struct ir3_register
*dst
= instr
->regs
[0];
755 return reg_num(dst
) == REG_P0
;
760 /* returns defining instruction for reg */
761 /* TODO better name */
762 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
764 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
770 static inline bool conflicts(struct ir3_instruction
*a
,
771 struct ir3_instruction
*b
)
773 return (a
&& b
) && (a
!= b
);
776 static inline bool reg_gpr(struct ir3_register
*r
)
778 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
780 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
785 static inline type_t
half_type(type_t type
)
788 case TYPE_F32
: return TYPE_F16
;
789 case TYPE_U32
: return TYPE_U16
;
790 case TYPE_S32
: return TYPE_S16
;
801 /* some cat2 instructions (ie. those which are not float) can embed an
804 static inline bool ir3_cat2_int(opc_t opc
)
845 /* map cat2 instruction to valid abs/neg flags: */
846 static inline unsigned ir3_cat2_absneg(opc_t opc
)
863 return IR3_REG_FABS
| IR3_REG_FNEG
;
884 return IR3_REG_SABS
| IR3_REG_SNEG
;
905 /* map cat3 instructions to valid abs/neg flags: */
906 static inline unsigned ir3_cat3_absneg(opc_t opc
)
925 /* neg *may* work on 3rd src.. */
935 #define MASK(n) ((1 << (n)) - 1)
937 /* iterator for an instructions's sources (reg), also returns src #: */
938 #define foreach_src_n(__srcreg, __n, __instr) \
939 if ((__instr)->regs_count) \
940 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
941 if ((__srcreg = (__instr)->regs[__n + 1]))
943 /* iterator for an instructions's sources (reg): */
944 #define foreach_src(__srcreg, __instr) \
945 foreach_src_n(__srcreg, __i, __instr)
947 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
949 unsigned cnt
= instr
->regs_count
+ instr
->deps_count
;
955 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
957 if (n
== (instr
->regs_count
+ instr
->deps_count
))
958 return instr
->address
;
959 if (n
>= instr
->regs_count
)
960 return instr
->deps
[n
- instr
->regs_count
];
961 return ssa(instr
->regs
[n
]);
964 static inline bool __is_false_dep(struct ir3_instruction
*instr
, unsigned n
)
966 if (n
== (instr
->regs_count
+ instr
->deps_count
))
968 if (n
>= instr
->regs_count
)
973 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
975 /* iterator for an instruction's SSA sources (instr), also returns src #: */
976 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
977 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
978 if ((__srcinst = __ssa_src_n(__instr, __n)))
980 /* iterator for an instruction's SSA sources (instr): */
981 #define foreach_ssa_src(__srcinst, __instr) \
982 foreach_ssa_src_n(__srcinst, __i, __instr)
986 void ir3_print(struct ir3
*ir
);
987 void ir3_print_instr(struct ir3_instruction
*instr
);
989 /* depth calculation: */
990 int ir3_delayslots(struct ir3_instruction
*assigner
,
991 struct ir3_instruction
*consumer
, unsigned n
);
992 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
993 void ir3_depth(struct ir3
*ir
);
995 /* copy-propagate: */
996 struct ir3_shader_variant
;
997 void ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
999 /* group neighbors and insert mov's to resolve conflicts: */
1000 void ir3_group(struct ir3
*ir
);
1003 void ir3_sched_add_deps(struct ir3
*ir
);
1004 int ir3_sched(struct ir3
*ir
);
1006 void ir3_a6xx_fixup_atomic_dests(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1008 /* register assignment: */
1009 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(struct ir3_compiler
*compiler
);
1010 int ir3_ra(struct ir3
*ir3
, gl_shader_stage type
,
1011 bool frag_coord
, bool frag_face
);
1014 void ir3_legalize(struct ir3
*ir
, int *num_samp
, bool *has_ssbo
, int *max_bary
);
1016 /* ************************************************************************* */
1017 /* instruction helpers */
1019 static inline struct ir3_instruction
*
1020 create_immed_typed(struct ir3_block
*block
, uint32_t val
, type_t type
)
1022 struct ir3_instruction
*mov
;
1023 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1025 mov
= ir3_instr_create(block
, OPC_MOV
);
1026 mov
->cat1
.src_type
= type
;
1027 mov
->cat1
.dst_type
= type
;
1028 ir3_reg_create(mov
, 0, flags
);
1029 ir3_reg_create(mov
, 0, IR3_REG_IMMED
)->uim_val
= val
;
1034 static inline struct ir3_instruction
*
1035 create_immed(struct ir3_block
*block
, uint32_t val
)
1037 return create_immed_typed(block
, val
, TYPE_U32
);
1040 static inline struct ir3_instruction
*
1041 create_uniform(struct ir3_block
*block
, unsigned n
)
1043 struct ir3_instruction
*mov
;
1045 mov
= ir3_instr_create(block
, OPC_MOV
);
1046 /* TODO get types right? */
1047 mov
->cat1
.src_type
= TYPE_F32
;
1048 mov
->cat1
.dst_type
= TYPE_F32
;
1049 ir3_reg_create(mov
, 0, 0);
1050 ir3_reg_create(mov
, n
, IR3_REG_CONST
);
1055 static inline struct ir3_instruction
*
1056 create_uniform_indirect(struct ir3_block
*block
, int n
,
1057 struct ir3_instruction
*address
)
1059 struct ir3_instruction
*mov
;
1061 mov
= ir3_instr_create(block
, OPC_MOV
);
1062 mov
->cat1
.src_type
= TYPE_U32
;
1063 mov
->cat1
.dst_type
= TYPE_U32
;
1064 ir3_reg_create(mov
, 0, 0);
1065 ir3_reg_create(mov
, 0, IR3_REG_CONST
| IR3_REG_RELATIV
)->array
.offset
= n
;
1067 ir3_instr_set_address(mov
, address
);
1072 /* creates SSA src of correct type (ie. half vs full precision) */
1073 static inline struct ir3_register
* __ssa_src(struct ir3_instruction
*instr
,
1074 struct ir3_instruction
*src
, unsigned flags
)
1076 struct ir3_register
*reg
;
1077 if (src
->regs
[0]->flags
& IR3_REG_HALF
)
1078 flags
|= IR3_REG_HALF
;
1079 reg
= ir3_reg_create(instr
, 0, IR3_REG_SSA
| flags
);
1081 reg
->wrmask
= src
->regs
[0]->wrmask
;
1085 static inline struct ir3_instruction
*
1086 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
1088 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1089 ir3_reg_create(instr
, 0, 0); /* dst */
1090 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
1091 struct ir3_register
*src_reg
= __ssa_src(instr
, src
, IR3_REG_ARRAY
);
1092 src_reg
->array
= src
->regs
[0]->array
;
1094 __ssa_src(instr
, src
, src
->regs
[0]->flags
& IR3_REG_HIGH
);
1096 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
1097 instr
->cat1
.src_type
= type
;
1098 instr
->cat1
.dst_type
= type
;
1102 static inline struct ir3_instruction
*
1103 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
1104 type_t src_type
, type_t dst_type
)
1106 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1107 unsigned dst_flags
= (type_size(dst_type
) < 32) ? IR3_REG_HALF
: 0;
1108 unsigned src_flags
= (type_size(src_type
) < 32) ? IR3_REG_HALF
: 0;
1110 debug_assert((src
->regs
[0]->flags
& IR3_REG_HALF
) == src_flags
);
1112 ir3_reg_create(instr
, 0, dst_flags
); /* dst */
1113 __ssa_src(instr
, src
, 0);
1114 instr
->cat1
.src_type
= src_type
;
1115 instr
->cat1
.dst_type
= dst_type
;
1116 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
1120 static inline struct ir3_instruction
*
1121 ir3_NOP(struct ir3_block
*block
)
1123 return ir3_instr_create(block
, OPC_NOP
);
1126 #define INSTR0(name) \
1127 static inline struct ir3_instruction * \
1128 ir3_##name(struct ir3_block *block) \
1130 struct ir3_instruction *instr = \
1131 ir3_instr_create(block, OPC_##name); \
1135 #define INSTR1(name) \
1136 static inline struct ir3_instruction * \
1137 ir3_##name(struct ir3_block *block, \
1138 struct ir3_instruction *a, unsigned aflags) \
1140 struct ir3_instruction *instr = \
1141 ir3_instr_create(block, OPC_##name); \
1142 ir3_reg_create(instr, 0, 0); /* dst */ \
1143 __ssa_src(instr, a, aflags); \
1147 #define INSTR2(name) \
1148 static inline struct ir3_instruction * \
1149 ir3_##name(struct ir3_block *block, \
1150 struct ir3_instruction *a, unsigned aflags, \
1151 struct ir3_instruction *b, unsigned bflags) \
1153 struct ir3_instruction *instr = \
1154 ir3_instr_create(block, OPC_##name); \
1155 ir3_reg_create(instr, 0, 0); /* dst */ \
1156 __ssa_src(instr, a, aflags); \
1157 __ssa_src(instr, b, bflags); \
1161 #define INSTR3(name) \
1162 static inline struct ir3_instruction * \
1163 ir3_##name(struct ir3_block *block, \
1164 struct ir3_instruction *a, unsigned aflags, \
1165 struct ir3_instruction *b, unsigned bflags, \
1166 struct ir3_instruction *c, unsigned cflags) \
1168 struct ir3_instruction *instr = \
1169 ir3_instr_create(block, OPC_##name); \
1170 ir3_reg_create(instr, 0, 0); /* dst */ \
1171 __ssa_src(instr, a, aflags); \
1172 __ssa_src(instr, b, bflags); \
1173 __ssa_src(instr, c, cflags); \
1177 #define INSTR3F(f, name) \
1178 static inline struct ir3_instruction * \
1179 ir3_##name##_##f(struct ir3_block *block, \
1180 struct ir3_instruction *a, unsigned aflags, \
1181 struct ir3_instruction *b, unsigned bflags, \
1182 struct ir3_instruction *c, unsigned cflags) \
1184 struct ir3_instruction *instr = \
1185 ir3_instr_create2(block, OPC_##name, 5); \
1186 ir3_reg_create(instr, 0, 0); /* dst */ \
1187 __ssa_src(instr, a, aflags); \
1188 __ssa_src(instr, b, bflags); \
1189 __ssa_src(instr, c, cflags); \
1190 instr->flags |= IR3_INSTR_##f; \
1194 #define INSTR4(name) \
1195 static inline struct ir3_instruction * \
1196 ir3_##name(struct ir3_block *block, \
1197 struct ir3_instruction *a, unsigned aflags, \
1198 struct ir3_instruction *b, unsigned bflags, \
1199 struct ir3_instruction *c, unsigned cflags, \
1200 struct ir3_instruction *d, unsigned dflags) \
1202 struct ir3_instruction *instr = \
1203 ir3_instr_create2(block, OPC_##name, 5); \
1204 ir3_reg_create(instr, 0, 0); /* dst */ \
1205 __ssa_src(instr, a, aflags); \
1206 __ssa_src(instr, b, bflags); \
1207 __ssa_src(instr, c, cflags); \
1208 __ssa_src(instr, d, dflags); \
1212 #define INSTR4F(f, name) \
1213 static inline struct ir3_instruction * \
1214 ir3_##name##_##f(struct ir3_block *block, \
1215 struct ir3_instruction *a, unsigned aflags, \
1216 struct ir3_instruction *b, unsigned bflags, \
1217 struct ir3_instruction *c, unsigned cflags, \
1218 struct ir3_instruction *d, unsigned dflags) \
1220 struct ir3_instruction *instr = \
1221 ir3_instr_create2(block, OPC_##name, 5); \
1222 ir3_reg_create(instr, 0, 0); /* dst */ \
1223 __ssa_src(instr, a, aflags); \
1224 __ssa_src(instr, b, bflags); \
1225 __ssa_src(instr, c, cflags); \
1226 __ssa_src(instr, d, dflags); \
1227 instr->flags |= IR3_INSTR_##f; \
1231 /* cat0 instructions: */
1237 /* cat2 instructions, most 2 src but some 1 src: */
1285 /* cat3 instructions: */
1303 /* cat4 instructions: */
1312 /* cat5 instructions: */
1316 static inline struct ir3_instruction
*
1317 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1318 unsigned wrmask
, unsigned flags
, unsigned samp
, unsigned tex
,
1319 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1321 struct ir3_instruction
*sam
;
1322 struct ir3_register
*reg
;
1324 sam
= ir3_instr_create(block
, opc
);
1325 sam
->flags
|= flags
;
1326 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1328 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1329 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1333 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1335 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1337 sam
->cat5
.samp
= samp
;
1338 sam
->cat5
.tex
= tex
;
1339 sam
->cat5
.type
= type
;
1344 /* cat6 instructions: */
1357 INSTR2(ATOMIC_CMPXCHG
)
1366 INSTR3F(G
, ATOMIC_ADD
)
1367 INSTR3F(G
, ATOMIC_SUB
)
1368 INSTR3F(G
, ATOMIC_XCHG
)
1369 INSTR3F(G
, ATOMIC_INC
)
1370 INSTR3F(G
, ATOMIC_DEC
)
1371 INSTR3F(G
, ATOMIC_CMPXCHG
)
1372 INSTR3F(G
, ATOMIC_MIN
)
1373 INSTR3F(G
, ATOMIC_MAX
)
1374 INSTR3F(G
, ATOMIC_AND
)
1375 INSTR3F(G
, ATOMIC_OR
)
1376 INSTR3F(G
, ATOMIC_XOR
)
1381 INSTR4F(G
, ATOMIC_ADD
)
1382 INSTR4F(G
, ATOMIC_SUB
)
1383 INSTR4F(G
, ATOMIC_XCHG
)
1384 INSTR4F(G
, ATOMIC_INC
)
1385 INSTR4F(G
, ATOMIC_DEC
)
1386 INSTR4F(G
, ATOMIC_CMPXCHG
)
1387 INSTR4F(G
, ATOMIC_MIN
)
1388 INSTR4F(G
, ATOMIC_MAX
)
1389 INSTR4F(G
, ATOMIC_AND
)
1390 INSTR4F(G
, ATOMIC_OR
)
1391 INSTR4F(G
, ATOMIC_XOR
)
1394 /* cat7 instructions: */
1398 /* ************************************************************************* */
1399 /* split this out or find some helper to use.. like main/bitset.h.. */
1405 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1407 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1409 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1410 debug_assert(num
< MAX_REG
);
1411 if (reg
->flags
& IR3_REG_HALF
)
1416 static inline void regmask_init(regmask_t
*regmask
)
1418 memset(regmask
, 0, sizeof(*regmask
));
1421 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1423 unsigned idx
= regmask_idx(reg
);
1424 if (reg
->flags
& IR3_REG_RELATIV
) {
1426 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1427 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1430 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1432 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1436 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1439 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1440 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1443 /* set bits in a if not set in b, conceptually:
1446 static inline void regmask_set_if_not(regmask_t
*a
,
1447 struct ir3_register
*reg
, regmask_t
*b
)
1449 unsigned idx
= regmask_idx(reg
);
1450 if (reg
->flags
& IR3_REG_RELATIV
) {
1452 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1453 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1454 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1457 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1459 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1460 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1464 static inline bool regmask_get(regmask_t
*regmask
,
1465 struct ir3_register
*reg
)
1467 unsigned idx
= regmask_idx(reg
);
1468 if (reg
->flags
& IR3_REG_RELATIV
) {
1470 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1471 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1475 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1477 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1483 /* ************************************************************************* */