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 */
55 /* number of sync bits: */
61 IR3_REG_CONST
= 0x001,
62 IR3_REG_IMMED
= 0x002,
64 /* high registers are used for some things in compute shaders,
65 * for example. Seems to be for things that are global to all
66 * threads in a wave, so possibly these are global/shared by
67 * all the threads in the wave?
70 IR3_REG_RELATIV
= 0x010,
72 /* Most instructions, it seems, can do float abs/neg but not
73 * integer. The CP pass needs to know what is intended (int or
74 * float) in order to do the right thing. For this reason the
75 * abs/neg flags are split out into float and int variants. In
76 * addition, .b (bitwise) operations, the negate is actually a
77 * bitwise not, so split that out into a new flag to make it
86 IR3_REG_POS_INF
= 0x1000,
87 /* (ei) flag, end-input? Set on last bary, presumably to signal
88 * that the shader needs no more input:
91 /* meta-flags, for intermediate stages of IR, ie.
92 * before register assignment is done:
94 IR3_REG_SSA
= 0x4000, /* 'instr' is ptr to assigning instr */
95 IR3_REG_ARRAY
= 0x8000,
96 IR3_REG_PHI_SRC
= 0x10000, /* phi src, regs[0]->instr points to phi */
101 * the component is in the low two bits of the reg #, so
102 * rN.x becomes: (N << 2) | x
116 /* For IR3_REG_SSA, src registers contain ptr back to assigning
119 * For IR3_REG_ARRAY, the pointer is back to the last dependent
120 * array access (although the net effect is the same, it points
121 * back to a previous instruction that we depend on).
123 struct ir3_instruction
*instr
;
126 /* used for cat5 instructions, but also for internal/IR level
127 * tracking of what registers are read/written by an instruction.
128 * wrmask may be a bad name since it is used to represent both
129 * src and dst that touch multiple adjacent registers.
132 /* for relative addressing, 32bits for array size is too small,
133 * but otoh we don't need to deal with disjoint sets, so instead
134 * use a simple size field (number of scalar components).
141 * Stupid/simple growable array implementation:
143 #define DECLARE_ARRAY(type, name) \
144 unsigned name ## _count, name ## _sz; \
147 #define array_insert(ctx, arr, val) do { \
148 if (arr ## _count == arr ## _sz) { \
149 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
150 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
152 arr[arr ##_count++] = val; \
155 struct ir3_instruction
{
156 struct ir3_block
*block
;
159 /* (sy) flag is set on first instruction, and after sample
160 * instructions (probably just on RAW hazard).
162 IR3_INSTR_SY
= 0x001,
163 /* (ss) flag is set on first instruction, and first instruction
164 * to depend on the result of "long" instructions (RAW hazard):
166 * rcp, rsq, log2, exp2, sin, cos, sqrt
168 * It seems to synchronize until all in-flight instructions are
169 * completed, for example:
172 * add.f hr2.z, (neg)hr2.z, hc0.y
173 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
176 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
178 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
179 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
180 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
182 * The last mul.f does not have (ss) set, presumably because the
183 * (ss) on the previous instruction does the job.
185 * The blob driver also seems to set it on WAR hazards, although
186 * not really clear if this is needed or just blob compiler being
187 * sloppy. So far I haven't found a case where removing the (ss)
188 * causes problems for WAR hazard, but I could just be getting
192 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
195 IR3_INSTR_SS
= 0x002,
196 /* (jp) flag is set on jump targets:
198 IR3_INSTR_JP
= 0x004,
199 IR3_INSTR_UL
= 0x008,
200 IR3_INSTR_3D
= 0x010,
205 IR3_INSTR_S2EN
= 0x200,
207 IR3_INSTR_SAT
= 0x800,
208 /* meta-flags, for intermediate stages of IR, ie.
209 * before register assignment is done:
211 IR3_INSTR_MARK
= 0x1000,
212 IR3_INSTR_UNUSED
= 0x2000,
219 struct ir3_register
**regs
;
225 struct ir3_block
*target
;
228 type_t src_type
, dst_type
;
248 int iim_val
: 3; /* for ldgb/stgb, # of components */
253 unsigned w
: 1; /* write */
254 unsigned r
: 1; /* read */
255 unsigned l
: 1; /* local */
256 unsigned g
: 1; /* global */
258 /* for meta-instructions, just used to hold extra data
259 * before instruction scheduling, etc
262 int off
; /* component/offset */
265 /* used to temporarily hold reference to nir_phi_instr
266 * until we resolve the phi srcs
271 struct ir3_block
*block
;
275 /* transient values used during various algorithms: */
277 /* The instruction depth is the max dependency distance to output.
279 * You can also think of it as the "cost", if we did any sort of
280 * optimization for register footprint. Ie. a value that is just
281 * result of moving a const to a reg would have a low cost, so to
282 * it could make sense to duplicate the instruction at various
283 * points where the result is needed to reduce register footprint.
286 /* When we get to the RA stage, we no longer need depth, but
287 * we do need instruction's position/name:
295 /* used for per-pass extra instruction data.
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
;
449 unsigned block_count
, instr_count
;
453 typedef struct nir_register nir_register
;
456 struct list_head node
;
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
);
478 typedef struct nir_block nir_block
;
481 struct list_head node
;
486 struct list_head instr_list
; /* list of ir3_instruction */
488 /* each block has either one or two successors.. in case of
489 * two successors, 'condition' decides which one to follow.
490 * A block preceding an if/else has two successors.
492 struct ir3_instruction
*condition
;
493 struct ir3_block
*successors
[2];
495 uint16_t start_ip
, end_ip
;
497 /* Track instructions which do not write a register but other-
498 * wise must not be discarded (such as kill, stg, etc)
500 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
502 /* used for per-pass extra block data. Mainly used right
503 * now in RA step to track livein/liveout.
512 static inline uint32_t
513 block_id(struct ir3_block
*block
)
516 return block
->serialno
;
518 return (uint32_t)(unsigned long)block
;
522 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
523 unsigned nin
, unsigned nout
);
524 void ir3_destroy(struct ir3
*shader
);
525 void * ir3_assemble(struct ir3
*shader
,
526 struct ir3_info
*info
, uint32_t gpu_id
);
527 void * ir3_alloc(struct ir3
*shader
, int sz
);
529 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
531 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
532 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
533 opc_t opc
, int nreg
);
534 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
535 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
);
536 const char *ir3_instr_name(struct ir3_instruction
*instr
);
538 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
540 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
541 struct ir3_register
*reg
);
543 void ir3_instr_set_address(struct ir3_instruction
*instr
,
544 struct ir3_instruction
*addr
);
546 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
548 if (instr
->flags
& IR3_INSTR_MARK
)
549 return true; /* already visited */
550 instr
->flags
|= IR3_INSTR_MARK
;
554 void ir3_block_clear_mark(struct ir3_block
*block
);
555 void ir3_clear_mark(struct ir3
*shader
);
557 unsigned ir3_count_instructions(struct ir3
*ir
);
559 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
560 struct ir3_register
*reg
)
563 for (i
= 0; i
< instr
->regs_count
; i
++)
564 if (reg
== instr
->regs
[i
])
570 #define MAX_ARRAYS 16
578 static inline uint32_t regid(int num
, int comp
)
580 return (num
<< 2) | (comp
& 0x3);
583 static inline uint32_t reg_num(struct ir3_register
*reg
)
585 return reg
->num
>> 2;
588 static inline uint32_t reg_comp(struct ir3_register
*reg
)
590 return reg
->num
& 0x3;
593 static inline bool is_flow(struct ir3_instruction
*instr
)
595 return (opc_cat(instr
->opc
) == 0);
598 static inline bool is_kill(struct ir3_instruction
*instr
)
600 return instr
->opc
== OPC_KILL
;
603 static inline bool is_nop(struct ir3_instruction
*instr
)
605 return instr
->opc
== OPC_NOP
;
608 /* Is it a non-transformative (ie. not type changing) mov? This can
609 * also include absneg.s/absneg.f, which for the most part can be
610 * treated as a mov (single src argument).
612 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
614 struct ir3_register
*dst
;
616 switch (instr
->opc
) {
618 if (instr
->cat1
.src_type
!= instr
->cat1
.dst_type
)
623 if (instr
->flags
& IR3_INSTR_SAT
)
630 dst
= instr
->regs
[0];
632 /* mov's that write to a0.x or p0.x are special: */
633 if (dst
->num
== regid(REG_P0
, 0))
635 if (dst
->num
== regid(REG_A0
, 0))
638 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
644 static inline bool is_alu(struct ir3_instruction
*instr
)
646 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
649 static inline bool is_sfu(struct ir3_instruction
*instr
)
651 return (opc_cat(instr
->opc
) == 4);
654 static inline bool is_tex(struct ir3_instruction
*instr
)
656 return (opc_cat(instr
->opc
) == 5);
659 static inline bool is_mem(struct ir3_instruction
*instr
)
661 return (opc_cat(instr
->opc
) == 6);
664 static inline bool is_barrier(struct ir3_instruction
*instr
)
666 return (opc_cat(instr
->opc
) == 7);
670 is_store(struct ir3_instruction
*instr
)
672 /* these instructions, the "destination" register is
673 * actually a source, the address to store to.
675 switch (instr
->opc
) {
690 static inline bool is_load(struct ir3_instruction
*instr
)
692 switch (instr
->opc
) {
701 /* probably some others too.. */
708 static inline bool is_input(struct ir3_instruction
*instr
)
710 /* in some cases, ldlv is used to fetch varying without
711 * interpolation.. fortunately inloc is the first src
712 * register in either case
714 switch (instr
->opc
) {
723 static inline bool is_bool(struct ir3_instruction
*instr
)
725 switch (instr
->opc
) {
735 static inline bool is_meta(struct ir3_instruction
*instr
)
737 /* TODO how should we count PHI (and maybe fan-in/out) which
738 * might actually contribute some instructions to the final
741 return (opc_cat(instr
->opc
) == -1);
744 static inline bool writes_addr(struct ir3_instruction
*instr
)
746 if (instr
->regs_count
> 0) {
747 struct ir3_register
*dst
= instr
->regs
[0];
748 return reg_num(dst
) == REG_A0
;
753 static inline bool writes_pred(struct ir3_instruction
*instr
)
755 if (instr
->regs_count
> 0) {
756 struct ir3_register
*dst
= instr
->regs
[0];
757 return reg_num(dst
) == REG_P0
;
762 /* returns defining instruction for reg */
763 /* TODO better name */
764 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
766 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
767 debug_assert(!(reg
->instr
&& (reg
->instr
->flags
& IR3_INSTR_UNUSED
)));
773 static inline bool conflicts(struct ir3_instruction
*a
,
774 struct ir3_instruction
*b
)
776 return (a
&& b
) && (a
!= b
);
779 static inline bool reg_gpr(struct ir3_register
*r
)
781 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
783 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
788 static inline type_t
half_type(type_t type
)
791 case TYPE_F32
: return TYPE_F16
;
792 case TYPE_U32
: return TYPE_U16
;
793 case TYPE_S32
: return TYPE_S16
;
804 /* some cat2 instructions (ie. those which are not float) can embed an
807 static inline bool ir3_cat2_int(opc_t opc
)
848 /* map cat2 instruction to valid abs/neg flags: */
849 static inline unsigned ir3_cat2_absneg(opc_t opc
)
866 return IR3_REG_FABS
| IR3_REG_FNEG
;
887 return IR3_REG_SABS
| IR3_REG_SNEG
;
908 /* map cat3 instructions to valid abs/neg flags: */
909 static inline unsigned ir3_cat3_absneg(opc_t opc
)
928 /* neg *may* work on 3rd src.. */
938 #define MASK(n) ((1 << (n)) - 1)
940 /* iterator for an instructions's sources (reg), also returns src #: */
941 #define foreach_src_n(__srcreg, __n, __instr) \
942 if ((__instr)->regs_count) \
943 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
944 if ((__srcreg = (__instr)->regs[__n + 1]))
946 /* iterator for an instructions's sources (reg): */
947 #define foreach_src(__srcreg, __instr) \
948 foreach_src_n(__srcreg, __i, __instr)
950 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
952 unsigned cnt
= instr
->regs_count
+ instr
->deps_count
;
958 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
960 if (n
== (instr
->regs_count
+ instr
->deps_count
))
961 return instr
->address
;
962 if (n
>= instr
->regs_count
)
963 return instr
->deps
[n
- instr
->regs_count
];
964 return ssa(instr
->regs
[n
]);
967 static inline bool __is_false_dep(struct ir3_instruction
*instr
, unsigned n
)
969 if (n
== (instr
->regs_count
+ instr
->deps_count
))
971 if (n
>= instr
->regs_count
)
976 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
978 /* iterator for an instruction's SSA sources (instr), also returns src #: */
979 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
980 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
981 if ((__srcinst = __ssa_src_n(__instr, __n)))
983 /* iterator for an instruction's SSA sources (instr): */
984 #define foreach_ssa_src(__srcinst, __instr) \
985 foreach_ssa_src_n(__srcinst, __i, __instr)
989 void ir3_print(struct ir3
*ir
);
990 void ir3_print_instr(struct ir3_instruction
*instr
);
992 /* depth calculation: */
993 int ir3_delayslots(struct ir3_instruction
*assigner
,
994 struct ir3_instruction
*consumer
, unsigned n
);
995 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
996 void ir3_depth(struct ir3
*ir
);
998 /* copy-propagate: */
999 struct ir3_shader_variant
;
1000 void ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1002 /* group neighbors and insert mov's to resolve conflicts: */
1003 void ir3_group(struct ir3
*ir
);
1006 void ir3_sched_add_deps(struct ir3
*ir
);
1007 int ir3_sched(struct ir3
*ir
);
1009 /* register assignment: */
1010 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(void *memctx
);
1011 int ir3_ra(struct ir3
*ir3
, enum shader_t type
,
1012 bool frag_coord
, bool frag_face
);
1015 void ir3_legalize(struct ir3
*ir
, bool *has_samp
, bool *has_ssbo
, int *max_bary
);
1017 /* ************************************************************************* */
1018 /* instruction helpers */
1020 static inline struct ir3_instruction
*
1021 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
1023 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1024 ir3_reg_create(instr
, 0, 0); /* dst */
1025 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
1026 struct ir3_register
*src_reg
=
1027 ir3_reg_create(instr
, 0, IR3_REG_ARRAY
);
1028 src_reg
->array
= src
->regs
[0]->array
;
1029 src_reg
->instr
= src
;
1031 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
1033 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
1034 instr
->cat1
.src_type
= type
;
1035 instr
->cat1
.dst_type
= type
;
1039 static inline struct ir3_instruction
*
1040 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
1041 type_t src_type
, type_t dst_type
)
1043 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1044 ir3_reg_create(instr
, 0, 0); /* dst */
1045 ir3_reg_create(instr
, 0, IR3_REG_SSA
)->instr
= src
;
1046 instr
->cat1
.src_type
= src_type
;
1047 instr
->cat1
.dst_type
= dst_type
;
1048 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
1052 static inline struct ir3_instruction
*
1053 ir3_NOP(struct ir3_block
*block
)
1055 return ir3_instr_create(block
, OPC_NOP
);
1058 #define INSTR0(name) \
1059 static inline struct ir3_instruction * \
1060 ir3_##name(struct ir3_block *block) \
1062 struct ir3_instruction *instr = \
1063 ir3_instr_create(block, OPC_##name); \
1067 #define INSTR1(name) \
1068 static inline struct ir3_instruction * \
1069 ir3_##name(struct ir3_block *block, \
1070 struct ir3_instruction *a, unsigned aflags) \
1072 struct ir3_instruction *instr = \
1073 ir3_instr_create(block, OPC_##name); \
1074 ir3_reg_create(instr, 0, 0); /* dst */ \
1075 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1079 #define INSTR2(name) \
1080 static inline struct ir3_instruction * \
1081 ir3_##name(struct ir3_block *block, \
1082 struct ir3_instruction *a, unsigned aflags, \
1083 struct ir3_instruction *b, unsigned bflags) \
1085 struct ir3_instruction *instr = \
1086 ir3_instr_create(block, OPC_##name); \
1087 ir3_reg_create(instr, 0, 0); /* dst */ \
1088 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1089 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1093 #define INSTR3(name) \
1094 static inline struct ir3_instruction * \
1095 ir3_##name(struct ir3_block *block, \
1096 struct ir3_instruction *a, unsigned aflags, \
1097 struct ir3_instruction *b, unsigned bflags, \
1098 struct ir3_instruction *c, unsigned cflags) \
1100 struct ir3_instruction *instr = \
1101 ir3_instr_create(block, OPC_##name); \
1102 ir3_reg_create(instr, 0, 0); /* dst */ \
1103 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1104 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1105 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1109 #define INSTR4(name) \
1110 static inline struct ir3_instruction * \
1111 ir3_##name(struct ir3_block *block, \
1112 struct ir3_instruction *a, unsigned aflags, \
1113 struct ir3_instruction *b, unsigned bflags, \
1114 struct ir3_instruction *c, unsigned cflags, \
1115 struct ir3_instruction *d, unsigned dflags) \
1117 struct ir3_instruction *instr = \
1118 ir3_instr_create2(block, OPC_##name, 5); \
1119 ir3_reg_create(instr, 0, 0); /* dst */ \
1120 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1121 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1122 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1123 ir3_reg_create(instr, 0, IR3_REG_SSA | dflags)->instr = d; \
1127 #define INSTR4F(f, name) \
1128 static inline struct ir3_instruction * \
1129 ir3_##name##_##f(struct ir3_block *block, \
1130 struct ir3_instruction *a, unsigned aflags, \
1131 struct ir3_instruction *b, unsigned bflags, \
1132 struct ir3_instruction *c, unsigned cflags, \
1133 struct ir3_instruction *d, unsigned dflags) \
1135 struct ir3_instruction *instr = \
1136 ir3_instr_create2(block, OPC_##name, 5); \
1137 ir3_reg_create(instr, 0, 0); /* dst */ \
1138 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1139 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1140 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1141 ir3_reg_create(instr, 0, IR3_REG_SSA | dflags)->instr = d; \
1142 instr->flags |= IR3_INSTR_##f; \
1146 /* cat0 instructions: */
1152 /* cat2 instructions, most 2 src but some 1 src: */
1200 /* cat3 instructions: */
1218 /* cat4 instructions: */
1227 /* cat5 instructions: */
1231 static inline struct ir3_instruction
*
1232 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1233 unsigned wrmask
, unsigned flags
, unsigned samp
, unsigned tex
,
1234 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1236 struct ir3_instruction
*sam
;
1237 struct ir3_register
*reg
;
1239 sam
= ir3_instr_create(block
, opc
);
1240 sam
->flags
|= flags
;
1241 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1243 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1244 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1248 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1250 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1252 sam
->cat5
.samp
= samp
;
1253 sam
->cat5
.tex
= tex
;
1254 sam
->cat5
.type
= type
;
1259 /* cat6 instructions: */
1275 INSTR2(ATOMIC_CMPXCHG
)
1281 INSTR4F(G
, ATOMIC_ADD
)
1282 INSTR4F(G
, ATOMIC_SUB
)
1283 INSTR4F(G
, ATOMIC_XCHG
)
1284 INSTR4F(G
, ATOMIC_INC
)
1285 INSTR4F(G
, ATOMIC_DEC
)
1286 INSTR4F(G
, ATOMIC_CMPXCHG
)
1287 INSTR4F(G
, ATOMIC_MIN
)
1288 INSTR4F(G
, ATOMIC_MAX
)
1289 INSTR4F(G
, ATOMIC_AND
)
1290 INSTR4F(G
, ATOMIC_OR
)
1291 INSTR4F(G
, ATOMIC_XOR
)
1293 /* cat7 instructions: */
1297 /* ************************************************************************* */
1298 /* split this out or find some helper to use.. like main/bitset.h.. */
1304 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1306 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1308 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1309 debug_assert(num
< MAX_REG
);
1310 if (reg
->flags
& IR3_REG_HALF
)
1315 static inline void regmask_init(regmask_t
*regmask
)
1317 memset(regmask
, 0, sizeof(*regmask
));
1320 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1322 unsigned idx
= regmask_idx(reg
);
1323 if (reg
->flags
& IR3_REG_RELATIV
) {
1325 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1326 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1329 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1331 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1335 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1338 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1339 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1342 /* set bits in a if not set in b, conceptually:
1345 static inline void regmask_set_if_not(regmask_t
*a
,
1346 struct ir3_register
*reg
, regmask_t
*b
)
1348 unsigned idx
= regmask_idx(reg
);
1349 if (reg
->flags
& IR3_REG_RELATIV
) {
1351 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1352 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1353 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1356 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1358 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1359 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1363 static inline bool regmask_get(regmask_t
*regmask
,
1364 struct ir3_register
*reg
)
1366 unsigned idx
= regmask_idx(reg
);
1367 if (reg
->flags
& IR3_REG_RELATIV
) {
1369 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1370 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1374 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1376 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1382 /* ************************************************************************* */