4bd7601b8dd4f79b9d1e24e8970b72ebb3eb12ac
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/bitscan.h"
33 #include "util/list.h"
34 #include "util/u_debug.h"
36 #include "instr-a3xx.h"
38 /* low level intermediate representation of an adreno shader program */
42 struct ir3_instruction
;
48 uint16_t instrs_count
; /* expanded to account for rpt's */
49 /* NOTE: max_reg, etc, does not include registers not touched
50 * by the shader (ie. vertex fetched via VFD_DECODE but not
53 int8_t max_reg
; /* highest GPR # used by shader */
57 /* number of sync bits: */
63 IR3_REG_CONST
= 0x001,
64 IR3_REG_IMMED
= 0x002,
66 /* high registers are used for some things in compute shaders,
67 * for example. Seems to be for things that are global to all
68 * threads in a wave, so possibly these are global/shared by
69 * all the threads in the wave?
72 IR3_REG_RELATIV
= 0x010,
74 /* Most instructions, it seems, can do float abs/neg but not
75 * integer. The CP pass needs to know what is intended (int or
76 * float) in order to do the right thing. For this reason the
77 * abs/neg flags are split out into float and int variants. In
78 * addition, .b (bitwise) operations, the negate is actually a
79 * bitwise not, so split that out into a new flag to make it
88 IR3_REG_POS_INF
= 0x1000,
89 /* (ei) flag, end-input? Set on last bary, presumably to signal
90 * that the shader needs no more input:
93 /* meta-flags, for intermediate stages of IR, ie.
94 * before register assignment is done:
96 IR3_REG_SSA
= 0x4000, /* 'instr' is ptr to assigning instr */
97 IR3_REG_ARRAY
= 0x8000,
102 * the component is in the low two bits of the reg #, so
103 * rN.x becomes: (N << 2) | x
118 /* For IR3_REG_SSA, src registers contain ptr back to assigning
121 * For IR3_REG_ARRAY, the pointer is back to the last dependent
122 * array access (although the net effect is the same, it points
123 * back to a previous instruction that we depend on).
125 struct ir3_instruction
*instr
;
128 /* used for cat5 instructions, but also for internal/IR level
129 * tracking of what registers are read/written by an instruction.
130 * wrmask may be a bad name since it is used to represent both
131 * src and dst that touch multiple adjacent registers.
134 /* for relative addressing, 32bits for array size is too small,
135 * but otoh we don't need to deal with disjoint sets, so instead
136 * use a simple size field (number of scalar components).
143 * Stupid/simple growable array implementation:
145 #define DECLARE_ARRAY(type, name) \
146 unsigned name ## _count, name ## _sz; \
149 #define array_insert(ctx, arr, val) do { \
150 if (arr ## _count == arr ## _sz) { \
151 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
152 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
154 arr[arr ##_count++] = val; \
157 struct ir3_instruction
{
158 struct ir3_block
*block
;
161 /* (sy) flag is set on first instruction, and after sample
162 * instructions (probably just on RAW hazard).
164 IR3_INSTR_SY
= 0x001,
165 /* (ss) flag is set on first instruction, and first instruction
166 * to depend on the result of "long" instructions (RAW hazard):
168 * rcp, rsq, log2, exp2, sin, cos, sqrt
170 * It seems to synchronize until all in-flight instructions are
171 * completed, for example:
174 * add.f hr2.z, (neg)hr2.z, hc0.y
175 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
178 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
180 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
181 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
182 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
184 * The last mul.f does not have (ss) set, presumably because the
185 * (ss) on the previous instruction does the job.
187 * The blob driver also seems to set it on WAR hazards, although
188 * not really clear if this is needed or just blob compiler being
189 * sloppy. So far I haven't found a case where removing the (ss)
190 * causes problems for WAR hazard, but I could just be getting
194 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
197 IR3_INSTR_SS
= 0x002,
198 /* (jp) flag is set on jump targets:
200 IR3_INSTR_JP
= 0x004,
201 IR3_INSTR_UL
= 0x008,
202 IR3_INSTR_3D
= 0x010,
207 IR3_INSTR_S2EN
= 0x200,
209 IR3_INSTR_SAT
= 0x800,
210 /* meta-flags, for intermediate stages of IR, ie.
211 * before register assignment is done:
213 IR3_INSTR_MARK
= 0x1000,
214 IR3_INSTR_UNUSED
= 0x2000,
222 struct ir3_register
**regs
;
228 struct ir3_block
*target
;
231 type_t src_type
, dst_type
;
251 int iim_val
: 3; /* for ldgb/stgb, # of components */
256 unsigned w
: 1; /* write */
257 unsigned r
: 1; /* read */
258 unsigned l
: 1; /* local */
259 unsigned g
: 1; /* global */
261 /* for meta-instructions, just used to hold extra data
262 * before instruction scheduling, etc
265 int off
; /* component/offset */
268 struct ir3_block
*block
;
272 /* transient values used during various algorithms: */
274 /* The instruction depth is the max dependency distance to output.
276 * You can also think of it as the "cost", if we did any sort of
277 * optimization for register footprint. Ie. a value that is just
278 * result of moving a const to a reg would have a low cost, so to
279 * it could make sense to duplicate the instruction at various
280 * points where the result is needed to reduce register footprint.
283 /* When we get to the RA stage, we no longer need depth, but
284 * we do need instruction's position/name:
292 /* used for per-pass extra instruction data.
296 int sun
; /* Sethi–Ullman number, used by sched */
297 int use_count
; /* currently just updated/used by cp */
299 /* Used during CP and RA stages. For fanin and shader inputs/
300 * outputs where we need a sequence of consecutive registers,
301 * keep track of each src instructions left (ie 'n-1') and right
302 * (ie 'n+1') neighbor. The front-end must insert enough mov's
303 * to ensure that each instruction has at most one left and at
304 * most one right neighbor. During the copy-propagation pass,
305 * we only remove mov's when we can preserve this constraint.
306 * And during the RA stage, we use the neighbor information to
307 * allocate a block of registers in one shot.
309 * TODO: maybe just add something like:
310 * struct ir3_instruction_ref {
311 * struct ir3_instruction *instr;
315 * Or can we get away without the refcnt stuff? It seems like
316 * it should be overkill.. the problem is if, potentially after
317 * already eliminating some mov's, if you have a single mov that
318 * needs to be grouped with it's neighbors in two different
319 * places (ex. shader output and a fanin).
322 struct ir3_instruction
*left
, *right
;
323 uint16_t left_cnt
, right_cnt
;
326 /* an instruction can reference at most one address register amongst
327 * it's src/dst registers. Beyond that, you need to insert mov's.
329 * NOTE: do not write this directly, use ir3_instr_set_address()
331 struct ir3_instruction
*address
;
333 /* Tracking for additional dependent instructions. Used to handle
334 * barriers, WAR hazards for arrays/SSBOs/etc.
336 DECLARE_ARRAY(struct ir3_instruction
*, deps
);
339 * From PoV of instruction scheduling, not execution (ie. ignores global/
340 * local distinction):
341 * shared image atomic SSBO everything
342 * barrier()/ - R/W R/W R/W R/W X
343 * groupMemoryBarrier()
344 * memoryBarrier() - R/W R/W
345 * (but only images declared coherent?)
346 * memoryBarrierAtomic() - R/W
347 * memoryBarrierBuffer() - R/W
348 * memoryBarrierImage() - R/W
349 * memoryBarrierShared() - R/W
351 * TODO I think for SSBO/image/shared, in cases where we can determine
352 * which variable is accessed, we don't need to care about accesses to
353 * different variables (unless declared coherent??)
356 IR3_BARRIER_EVERYTHING
= 1 << 0,
357 IR3_BARRIER_SHARED_R
= 1 << 1,
358 IR3_BARRIER_SHARED_W
= 1 << 2,
359 IR3_BARRIER_IMAGE_R
= 1 << 3,
360 IR3_BARRIER_IMAGE_W
= 1 << 4,
361 IR3_BARRIER_BUFFER_R
= 1 << 5,
362 IR3_BARRIER_BUFFER_W
= 1 << 6,
363 IR3_BARRIER_ARRAY_R
= 1 << 7,
364 IR3_BARRIER_ARRAY_W
= 1 << 8,
365 } barrier_class
, barrier_conflict
;
367 /* Entry in ir3_block's instruction list: */
368 struct list_head node
;
375 static inline struct ir3_instruction
*
376 ir3_neighbor_first(struct ir3_instruction
*instr
)
379 while (instr
->cp
.left
) {
380 instr
= instr
->cp
.left
;
381 if (++cnt
> 0xffff) {
389 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
393 debug_assert(!instr
->cp
.left
);
395 while (instr
->cp
.right
) {
397 instr
= instr
->cp
.right
;
408 struct ir3_compiler
*compiler
;
410 unsigned ninputs
, noutputs
;
411 struct ir3_instruction
**inputs
;
412 struct ir3_instruction
**outputs
;
414 /* Track bary.f (and ldlv) instructions.. this is needed in
415 * scheduling to ensure that all varying fetches happen before
416 * any potential kill instructions. The hw gets grumpy if all
417 * threads in a group are killed before the last bary.f gets
418 * a chance to signal end of input (ei).
420 DECLARE_ARRAY(struct ir3_instruction
*, baryfs
);
422 /* Track all indirect instructions (read and write). To avoid
423 * deadlock scenario where an address register gets scheduled,
424 * but other dependent src instructions cannot be scheduled due
425 * to dependency on a *different* address register value, the
426 * scheduler needs to ensure that all dependencies other than
427 * the instruction other than the address register are scheduled
428 * before the one that writes the address register. Having a
429 * convenient list of instructions that reference some address
430 * register simplifies this.
432 DECLARE_ARRAY(struct ir3_instruction
*, indirects
);
434 /* and same for instructions that consume predicate register: */
435 DECLARE_ARRAY(struct ir3_instruction
*, predicates
);
437 /* Track texture sample instructions which need texture state
438 * patched in (for astc-srgb workaround):
440 DECLARE_ARRAY(struct ir3_instruction
*, astc_srgb
);
442 /* List of blocks: */
443 struct list_head block_list
;
445 /* List of ir3_array's: */
446 struct list_head array_list
;
448 unsigned max_sun
; /* max Sethi–Ullman number */
451 unsigned block_count
, instr_count
;
456 struct list_head node
;
460 struct nir_register
*r
;
462 /* To avoid array write's from getting DCE'd, keep track of the
463 * most recent write. Any array access depends on the most
464 * recent write. This way, nothing depends on writes after the
465 * last read. But all the writes that happen before that have
466 * something depending on them
468 struct ir3_instruction
*last_write
;
470 /* extra stuff used in RA pass: */
471 unsigned base
; /* base vreg name */
472 unsigned reg
; /* base physical reg */
473 uint16_t start_ip
, end_ip
;
476 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
479 struct list_head node
;
482 const struct nir_block
*nblock
;
484 struct list_head instr_list
; /* list of ir3_instruction */
486 /* each block has either one or two successors.. in case of
487 * two successors, 'condition' decides which one to follow.
488 * A block preceding an if/else has two successors.
490 struct ir3_instruction
*condition
;
491 struct ir3_block
*successors
[2];
493 unsigned predecessors_count
;
494 struct ir3_block
**predecessors
;
496 uint16_t start_ip
, end_ip
;
498 /* Track instructions which do not write a register but other-
499 * wise must not be discarded (such as kill, stg, etc)
501 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
503 /* used for per-pass extra block data. Mainly used right
504 * now in RA step to track livein/liveout.
513 static inline uint32_t
514 block_id(struct ir3_block
*block
)
517 return block
->serialno
;
519 return (uint32_t)(unsigned long)block
;
523 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
524 unsigned nin
, unsigned nout
);
525 void ir3_destroy(struct ir3
*shader
);
526 void * ir3_assemble(struct ir3
*shader
,
527 struct ir3_info
*info
, uint32_t gpu_id
);
528 void * ir3_alloc(struct ir3
*shader
, int sz
);
530 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
532 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
533 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
534 opc_t opc
, int nreg
);
535 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
536 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
);
537 const char *ir3_instr_name(struct ir3_instruction
*instr
);
539 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
541 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
542 struct ir3_register
*reg
);
544 void ir3_instr_set_address(struct ir3_instruction
*instr
,
545 struct ir3_instruction
*addr
);
547 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
549 if (instr
->flags
& IR3_INSTR_MARK
)
550 return true; /* already visited */
551 instr
->flags
|= IR3_INSTR_MARK
;
555 void ir3_block_clear_mark(struct ir3_block
*block
);
556 void ir3_clear_mark(struct ir3
*shader
);
558 unsigned ir3_count_instructions(struct ir3
*ir
);
560 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
561 struct ir3_register
*reg
)
564 for (i
= 0; i
< instr
->regs_count
; i
++)
565 if (reg
== instr
->regs
[i
])
571 #define MAX_ARRAYS 16
579 static inline uint32_t regid(int num
, int comp
)
581 return (num
<< 2) | (comp
& 0x3);
584 static inline uint32_t reg_num(struct ir3_register
*reg
)
586 return reg
->num
>> 2;
589 static inline uint32_t reg_comp(struct ir3_register
*reg
)
591 return reg
->num
& 0x3;
594 static inline bool is_flow(struct ir3_instruction
*instr
)
596 return (opc_cat(instr
->opc
) == 0);
599 static inline bool is_kill(struct ir3_instruction
*instr
)
601 return instr
->opc
== OPC_KILL
;
604 static inline bool is_nop(struct ir3_instruction
*instr
)
606 return instr
->opc
== OPC_NOP
;
609 /* Is it a non-transformative (ie. not type changing) mov? This can
610 * also include absneg.s/absneg.f, which for the most part can be
611 * treated as a mov (single src argument).
613 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
615 struct ir3_register
*dst
;
617 switch (instr
->opc
) {
619 if (instr
->cat1
.src_type
!= instr
->cat1
.dst_type
)
624 if (instr
->flags
& IR3_INSTR_SAT
)
631 dst
= instr
->regs
[0];
633 /* mov's that write to a0.x or p0.x are special: */
634 if (dst
->num
== regid(REG_P0
, 0))
636 if (dst
->num
== regid(REG_A0
, 0))
639 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
645 static inline bool is_alu(struct ir3_instruction
*instr
)
647 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
650 static inline bool is_sfu(struct ir3_instruction
*instr
)
652 return (opc_cat(instr
->opc
) == 4);
655 static inline bool is_tex(struct ir3_instruction
*instr
)
657 return (opc_cat(instr
->opc
) == 5);
660 static inline bool is_mem(struct ir3_instruction
*instr
)
662 return (opc_cat(instr
->opc
) == 6);
665 static inline bool is_barrier(struct ir3_instruction
*instr
)
667 return (opc_cat(instr
->opc
) == 7);
671 is_store(struct ir3_instruction
*instr
)
673 /* these instructions, the "destination" register is
674 * actually a source, the address to store to.
676 switch (instr
->opc
) {
691 static inline bool is_load(struct ir3_instruction
*instr
)
693 switch (instr
->opc
) {
703 /* probably some others too.. */
710 static inline bool is_input(struct ir3_instruction
*instr
)
712 /* in some cases, ldlv is used to fetch varying without
713 * interpolation.. fortunately inloc is the first src
714 * register in either case
716 switch (instr
->opc
) {
725 static inline bool is_bool(struct ir3_instruction
*instr
)
727 switch (instr
->opc
) {
737 static inline bool is_meta(struct ir3_instruction
*instr
)
739 /* TODO how should we count PHI (and maybe fan-in/out) which
740 * might actually contribute some instructions to the final
743 return (opc_cat(instr
->opc
) == -1);
746 static inline unsigned dest_regs(struct ir3_instruction
*instr
)
748 if ((instr
->regs_count
== 0) || is_store(instr
))
751 return util_last_bit(instr
->regs
[0]->wrmask
);
754 static inline bool writes_addr(struct ir3_instruction
*instr
)
756 if (instr
->regs_count
> 0) {
757 struct ir3_register
*dst
= instr
->regs
[0];
758 return reg_num(dst
) == REG_A0
;
763 static inline bool writes_pred(struct ir3_instruction
*instr
)
765 if (instr
->regs_count
> 0) {
766 struct ir3_register
*dst
= instr
->regs
[0];
767 return reg_num(dst
) == REG_P0
;
772 /* returns defining instruction for reg */
773 /* TODO better name */
774 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
776 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
782 static inline bool conflicts(struct ir3_instruction
*a
,
783 struct ir3_instruction
*b
)
785 return (a
&& b
) && (a
!= b
);
788 static inline bool reg_gpr(struct ir3_register
*r
)
790 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
792 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
797 static inline type_t
half_type(type_t type
)
800 case TYPE_F32
: return TYPE_F16
;
801 case TYPE_U32
: return TYPE_U16
;
802 case TYPE_S32
: return TYPE_S16
;
813 /* some cat2 instructions (ie. those which are not float) can embed an
816 static inline bool ir3_cat2_int(opc_t opc
)
857 /* map cat2 instruction to valid abs/neg flags: */
858 static inline unsigned ir3_cat2_absneg(opc_t opc
)
875 return IR3_REG_FABS
| IR3_REG_FNEG
;
896 return IR3_REG_SABS
| IR3_REG_SNEG
;
917 /* map cat3 instructions to valid abs/neg flags: */
918 static inline unsigned ir3_cat3_absneg(opc_t opc
)
937 /* neg *may* work on 3rd src.. */
947 #define MASK(n) ((1 << (n)) - 1)
949 /* iterator for an instructions's sources (reg), also returns src #: */
950 #define foreach_src_n(__srcreg, __n, __instr) \
951 if ((__instr)->regs_count) \
952 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
953 if ((__srcreg = (__instr)->regs[__n + 1]))
955 /* iterator for an instructions's sources (reg): */
956 #define foreach_src(__srcreg, __instr) \
957 foreach_src_n(__srcreg, __i, __instr)
959 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
961 unsigned cnt
= instr
->regs_count
+ instr
->deps_count
;
967 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
969 if (n
== (instr
->regs_count
+ instr
->deps_count
))
970 return instr
->address
;
971 if (n
>= instr
->regs_count
)
972 return instr
->deps
[n
- instr
->regs_count
];
973 return ssa(instr
->regs
[n
]);
976 static inline bool __is_false_dep(struct ir3_instruction
*instr
, unsigned n
)
978 if (n
== (instr
->regs_count
+ instr
->deps_count
))
980 if (n
>= instr
->regs_count
)
985 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
987 /* iterator for an instruction's SSA sources (instr), also returns src #: */
988 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
989 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
990 if ((__srcinst = __ssa_src_n(__instr, __n)))
992 /* iterator for an instruction's SSA sources (instr): */
993 #define foreach_ssa_src(__srcinst, __instr) \
994 foreach_ssa_src_n(__srcinst, __i, __instr)
998 void ir3_print(struct ir3
*ir
);
999 void ir3_print_instr(struct ir3_instruction
*instr
);
1001 /* depth calculation: */
1002 int ir3_delayslots(struct ir3_instruction
*assigner
,
1003 struct ir3_instruction
*consumer
, unsigned n
);
1004 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
1005 void ir3_depth(struct ir3
*ir
);
1007 /* copy-propagate: */
1008 struct ir3_shader_variant
;
1009 void ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1011 /* group neighbors and insert mov's to resolve conflicts: */
1012 void ir3_group(struct ir3
*ir
);
1014 /* Sethi–Ullman numbering: */
1015 void ir3_sun(struct ir3
*ir
);
1018 void ir3_sched_add_deps(struct ir3
*ir
);
1019 int ir3_sched(struct ir3
*ir
);
1021 void ir3_a6xx_fixup_atomic_dests(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1023 /* register assignment: */
1024 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(struct ir3_compiler
*compiler
);
1025 int ir3_ra(struct ir3
*ir3
, gl_shader_stage type
,
1026 bool frag_coord
, bool frag_face
);
1029 void ir3_legalize(struct ir3
*ir
, int *num_samp
, bool *has_ssbo
, int *max_bary
);
1031 /* ************************************************************************* */
1032 /* instruction helpers */
1034 static inline struct ir3_instruction
*
1035 create_immed_typed(struct ir3_block
*block
, uint32_t val
, type_t type
)
1037 struct ir3_instruction
*mov
;
1038 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1040 mov
= ir3_instr_create(block
, OPC_MOV
);
1041 mov
->cat1
.src_type
= type
;
1042 mov
->cat1
.dst_type
= type
;
1043 ir3_reg_create(mov
, 0, flags
);
1044 ir3_reg_create(mov
, 0, IR3_REG_IMMED
)->uim_val
= val
;
1049 static inline struct ir3_instruction
*
1050 create_immed(struct ir3_block
*block
, uint32_t val
)
1052 return create_immed_typed(block
, val
, TYPE_U32
);
1055 static inline struct ir3_instruction
*
1056 create_uniform(struct ir3_block
*block
, unsigned n
)
1058 struct ir3_instruction
*mov
;
1060 mov
= ir3_instr_create(block
, OPC_MOV
);
1061 /* TODO get types right? */
1062 mov
->cat1
.src_type
= TYPE_F32
;
1063 mov
->cat1
.dst_type
= TYPE_F32
;
1064 ir3_reg_create(mov
, 0, 0);
1065 ir3_reg_create(mov
, n
, IR3_REG_CONST
);
1070 static inline struct ir3_instruction
*
1071 create_uniform_indirect(struct ir3_block
*block
, int n
,
1072 struct ir3_instruction
*address
)
1074 struct ir3_instruction
*mov
;
1076 mov
= ir3_instr_create(block
, OPC_MOV
);
1077 mov
->cat1
.src_type
= TYPE_U32
;
1078 mov
->cat1
.dst_type
= TYPE_U32
;
1079 ir3_reg_create(mov
, 0, 0);
1080 ir3_reg_create(mov
, 0, IR3_REG_CONST
| IR3_REG_RELATIV
)->array
.offset
= n
;
1082 ir3_instr_set_address(mov
, address
);
1087 /* creates SSA src of correct type (ie. half vs full precision) */
1088 static inline struct ir3_register
* __ssa_src(struct ir3_instruction
*instr
,
1089 struct ir3_instruction
*src
, unsigned flags
)
1091 struct ir3_register
*reg
;
1092 if (src
->regs
[0]->flags
& IR3_REG_HALF
)
1093 flags
|= IR3_REG_HALF
;
1094 reg
= ir3_reg_create(instr
, 0, IR3_REG_SSA
| flags
);
1096 reg
->wrmask
= src
->regs
[0]->wrmask
;
1100 static inline struct ir3_instruction
*
1101 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
1103 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1104 ir3_reg_create(instr
, 0, 0); /* dst */
1105 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
1106 struct ir3_register
*src_reg
= __ssa_src(instr
, src
, IR3_REG_ARRAY
);
1107 src_reg
->array
= src
->regs
[0]->array
;
1109 __ssa_src(instr
, src
, src
->regs
[0]->flags
& IR3_REG_HIGH
);
1111 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
1112 instr
->cat1
.src_type
= type
;
1113 instr
->cat1
.dst_type
= type
;
1117 static inline struct ir3_instruction
*
1118 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
1119 type_t src_type
, type_t dst_type
)
1121 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1122 unsigned dst_flags
= (type_size(dst_type
) < 32) ? IR3_REG_HALF
: 0;
1123 unsigned src_flags
= (type_size(src_type
) < 32) ? IR3_REG_HALF
: 0;
1125 debug_assert((src
->regs
[0]->flags
& IR3_REG_HALF
) == src_flags
);
1127 ir3_reg_create(instr
, 0, dst_flags
); /* dst */
1128 __ssa_src(instr
, src
, 0);
1129 instr
->cat1
.src_type
= src_type
;
1130 instr
->cat1
.dst_type
= dst_type
;
1131 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
1135 static inline struct ir3_instruction
*
1136 ir3_NOP(struct ir3_block
*block
)
1138 return ir3_instr_create(block
, OPC_NOP
);
1141 #define INSTR0(name) \
1142 static inline struct ir3_instruction * \
1143 ir3_##name(struct ir3_block *block) \
1145 struct ir3_instruction *instr = \
1146 ir3_instr_create(block, OPC_##name); \
1150 #define INSTR1(name) \
1151 static inline struct ir3_instruction * \
1152 ir3_##name(struct ir3_block *block, \
1153 struct ir3_instruction *a, unsigned aflags) \
1155 struct ir3_instruction *instr = \
1156 ir3_instr_create(block, OPC_##name); \
1157 ir3_reg_create(instr, 0, 0); /* dst */ \
1158 __ssa_src(instr, a, aflags); \
1162 #define INSTR2(name) \
1163 static inline struct ir3_instruction * \
1164 ir3_##name(struct ir3_block *block, \
1165 struct ir3_instruction *a, unsigned aflags, \
1166 struct ir3_instruction *b, unsigned bflags) \
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); \
1176 #define INSTR3(name) \
1177 static inline struct ir3_instruction * \
1178 ir3_##name(struct ir3_block *block, \
1179 struct ir3_instruction *a, unsigned aflags, \
1180 struct ir3_instruction *b, unsigned bflags, \
1181 struct ir3_instruction *c, unsigned cflags) \
1183 struct ir3_instruction *instr = \
1184 ir3_instr_create(block, OPC_##name); \
1185 ir3_reg_create(instr, 0, 0); /* dst */ \
1186 __ssa_src(instr, a, aflags); \
1187 __ssa_src(instr, b, bflags); \
1188 __ssa_src(instr, c, cflags); \
1192 #define INSTR3F(f, name) \
1193 static inline struct ir3_instruction * \
1194 ir3_##name##_##f(struct ir3_block *block, \
1195 struct ir3_instruction *a, unsigned aflags, \
1196 struct ir3_instruction *b, unsigned bflags, \
1197 struct ir3_instruction *c, unsigned cflags) \
1199 struct ir3_instruction *instr = \
1200 ir3_instr_create2(block, OPC_##name, 5); \
1201 ir3_reg_create(instr, 0, 0); /* dst */ \
1202 __ssa_src(instr, a, aflags); \
1203 __ssa_src(instr, b, bflags); \
1204 __ssa_src(instr, c, cflags); \
1205 instr->flags |= IR3_INSTR_##f; \
1209 #define INSTR4(name) \
1210 static inline struct ir3_instruction * \
1211 ir3_##name(struct ir3_block *block, \
1212 struct ir3_instruction *a, unsigned aflags, \
1213 struct ir3_instruction *b, unsigned bflags, \
1214 struct ir3_instruction *c, unsigned cflags, \
1215 struct ir3_instruction *d, unsigned dflags) \
1217 struct ir3_instruction *instr = \
1218 ir3_instr_create2(block, OPC_##name, 5); \
1219 ir3_reg_create(instr, 0, 0); /* dst */ \
1220 __ssa_src(instr, a, aflags); \
1221 __ssa_src(instr, b, bflags); \
1222 __ssa_src(instr, c, cflags); \
1223 __ssa_src(instr, d, dflags); \
1227 #define INSTR4F(f, name) \
1228 static inline struct ir3_instruction * \
1229 ir3_##name##_##f(struct ir3_block *block, \
1230 struct ir3_instruction *a, unsigned aflags, \
1231 struct ir3_instruction *b, unsigned bflags, \
1232 struct ir3_instruction *c, unsigned cflags, \
1233 struct ir3_instruction *d, unsigned dflags) \
1235 struct ir3_instruction *instr = \
1236 ir3_instr_create2(block, OPC_##name, 5); \
1237 ir3_reg_create(instr, 0, 0); /* dst */ \
1238 __ssa_src(instr, a, aflags); \
1239 __ssa_src(instr, b, bflags); \
1240 __ssa_src(instr, c, cflags); \
1241 __ssa_src(instr, d, dflags); \
1242 instr->flags |= IR3_INSTR_##f; \
1246 /* cat0 instructions: */
1252 /* cat2 instructions, most 2 src but some 1 src: */
1300 /* cat3 instructions: */
1318 /* cat4 instructions: */
1327 /* cat5 instructions: */
1331 static inline struct ir3_instruction
*
1332 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1333 unsigned wrmask
, unsigned flags
, unsigned samp
, unsigned tex
,
1334 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1336 struct ir3_instruction
*sam
;
1337 struct ir3_register
*reg
;
1339 sam
= ir3_instr_create(block
, opc
);
1340 sam
->flags
|= flags
;
1341 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1342 // temporary step, extra dummy src which will become the
1343 // hvec2(samp, tex) argument:
1344 ir3_reg_create(sam
, 0, 0);
1346 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1347 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1351 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1353 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1355 sam
->cat5
.samp
= samp
;
1356 sam
->cat5
.tex
= tex
;
1357 sam
->cat5
.type
= type
;
1362 /* cat6 instructions: */
1375 INSTR2(ATOMIC_CMPXCHG
)
1384 INSTR3F(G
, ATOMIC_ADD
)
1385 INSTR3F(G
, ATOMIC_SUB
)
1386 INSTR3F(G
, ATOMIC_XCHG
)
1387 INSTR3F(G
, ATOMIC_INC
)
1388 INSTR3F(G
, ATOMIC_DEC
)
1389 INSTR3F(G
, ATOMIC_CMPXCHG
)
1390 INSTR3F(G
, ATOMIC_MIN
)
1391 INSTR3F(G
, ATOMIC_MAX
)
1392 INSTR3F(G
, ATOMIC_AND
)
1393 INSTR3F(G
, ATOMIC_OR
)
1394 INSTR3F(G
, ATOMIC_XOR
)
1399 INSTR4F(G
, ATOMIC_ADD
)
1400 INSTR4F(G
, ATOMIC_SUB
)
1401 INSTR4F(G
, ATOMIC_XCHG
)
1402 INSTR4F(G
, ATOMIC_INC
)
1403 INSTR4F(G
, ATOMIC_DEC
)
1404 INSTR4F(G
, ATOMIC_CMPXCHG
)
1405 INSTR4F(G
, ATOMIC_MIN
)
1406 INSTR4F(G
, ATOMIC_MAX
)
1407 INSTR4F(G
, ATOMIC_AND
)
1408 INSTR4F(G
, ATOMIC_OR
)
1409 INSTR4F(G
, ATOMIC_XOR
)
1412 /* cat7 instructions: */
1416 /* ************************************************************************* */
1417 /* split this out or find some helper to use.. like main/bitset.h.. */
1423 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1425 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1427 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1428 debug_assert(num
< MAX_REG
);
1429 if (reg
->flags
& IR3_REG_HALF
)
1434 static inline void regmask_init(regmask_t
*regmask
)
1436 memset(regmask
, 0, sizeof(*regmask
));
1439 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1441 unsigned idx
= regmask_idx(reg
);
1442 if (reg
->flags
& IR3_REG_RELATIV
) {
1444 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1445 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1448 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1450 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1454 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1457 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1458 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1461 /* set bits in a if not set in b, conceptually:
1464 static inline void regmask_set_if_not(regmask_t
*a
,
1465 struct ir3_register
*reg
, regmask_t
*b
)
1467 unsigned idx
= regmask_idx(reg
);
1468 if (reg
->flags
& IR3_REG_RELATIV
) {
1470 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1471 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1472 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1475 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1477 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1478 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1482 static inline bool regmask_get(regmask_t
*regmask
,
1483 struct ir3_register
*reg
)
1485 unsigned idx
= regmask_idx(reg
);
1486 if (reg
->flags
& IR3_REG_RELATIV
) {
1488 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1489 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1493 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1495 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1501 /* ************************************************************************* */