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,
100 * the component is in the low two bits of the reg #, so
101 * rN.x becomes: (N << 2) | x
115 /* For IR3_REG_SSA, src registers contain ptr back to assigning
118 * For IR3_REG_ARRAY, the pointer is back to the last dependent
119 * array access (although the net effect is the same, it points
120 * back to a previous instruction that we depend on).
122 struct ir3_instruction
*instr
;
125 /* used for cat5 instructions, but also for internal/IR level
126 * tracking of what registers are read/written by an instruction.
127 * wrmask may be a bad name since it is used to represent both
128 * src and dst that touch multiple adjacent registers.
131 /* for relative addressing, 32bits for array size is too small,
132 * but otoh we don't need to deal with disjoint sets, so instead
133 * use a simple size field (number of scalar components).
140 * Stupid/simple growable array implementation:
142 #define DECLARE_ARRAY(type, name) \
143 unsigned name ## _count, name ## _sz; \
146 #define array_insert(ctx, arr, val) do { \
147 if (arr ## _count == arr ## _sz) { \
148 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
149 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
151 arr[arr ##_count++] = val; \
154 struct ir3_instruction
{
155 struct ir3_block
*block
;
158 /* (sy) flag is set on first instruction, and after sample
159 * instructions (probably just on RAW hazard).
161 IR3_INSTR_SY
= 0x001,
162 /* (ss) flag is set on first instruction, and first instruction
163 * to depend on the result of "long" instructions (RAW hazard):
165 * rcp, rsq, log2, exp2, sin, cos, sqrt
167 * It seems to synchronize until all in-flight instructions are
168 * completed, for example:
171 * add.f hr2.z, (neg)hr2.z, hc0.y
172 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
175 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
177 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
178 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
179 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
181 * The last mul.f does not have (ss) set, presumably because the
182 * (ss) on the previous instruction does the job.
184 * The blob driver also seems to set it on WAR hazards, although
185 * not really clear if this is needed or just blob compiler being
186 * sloppy. So far I haven't found a case where removing the (ss)
187 * causes problems for WAR hazard, but I could just be getting
191 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
194 IR3_INSTR_SS
= 0x002,
195 /* (jp) flag is set on jump targets:
197 IR3_INSTR_JP
= 0x004,
198 IR3_INSTR_UL
= 0x008,
199 IR3_INSTR_3D
= 0x010,
204 IR3_INSTR_S2EN
= 0x200,
206 IR3_INSTR_SAT
= 0x800,
207 /* meta-flags, for intermediate stages of IR, ie.
208 * before register assignment is done:
210 IR3_INSTR_MARK
= 0x1000,
211 IR3_INSTR_UNUSED
= 0x2000,
218 struct ir3_register
**regs
;
224 struct ir3_block
*target
;
227 type_t src_type
, dst_type
;
247 int iim_val
: 3; /* for ldgb/stgb, # of components */
252 unsigned w
: 1; /* write */
253 unsigned r
: 1; /* read */
254 unsigned l
: 1; /* local */
255 unsigned g
: 1; /* global */
257 /* for meta-instructions, just used to hold extra data
258 * before instruction scheduling, etc
261 int off
; /* component/offset */
264 struct ir3_block
*block
;
268 /* transient values used during various algorithms: */
270 /* The instruction depth is the max dependency distance to output.
272 * You can also think of it as the "cost", if we did any sort of
273 * optimization for register footprint. Ie. a value that is just
274 * result of moving a const to a reg would have a low cost, so to
275 * it could make sense to duplicate the instruction at various
276 * points where the result is needed to reduce register footprint.
279 /* When we get to the RA stage, we no longer need depth, but
280 * we do need instruction's position/name:
288 /* used for per-pass extra instruction data.
292 /* Used during CP and RA stages. For fanin and shader inputs/
293 * outputs where we need a sequence of consecutive registers,
294 * keep track of each src instructions left (ie 'n-1') and right
295 * (ie 'n+1') neighbor. The front-end must insert enough mov's
296 * to ensure that each instruction has at most one left and at
297 * most one right neighbor. During the copy-propagation pass,
298 * we only remove mov's when we can preserve this constraint.
299 * And during the RA stage, we use the neighbor information to
300 * allocate a block of registers in one shot.
302 * TODO: maybe just add something like:
303 * struct ir3_instruction_ref {
304 * struct ir3_instruction *instr;
308 * Or can we get away without the refcnt stuff? It seems like
309 * it should be overkill.. the problem is if, potentially after
310 * already eliminating some mov's, if you have a single mov that
311 * needs to be grouped with it's neighbors in two different
312 * places (ex. shader output and a fanin).
315 struct ir3_instruction
*left
, *right
;
316 uint16_t left_cnt
, right_cnt
;
319 /* an instruction can reference at most one address register amongst
320 * it's src/dst registers. Beyond that, you need to insert mov's.
322 * NOTE: do not write this directly, use ir3_instr_set_address()
324 struct ir3_instruction
*address
;
326 /* Tracking for additional dependent instructions. Used to handle
327 * barriers, WAR hazards for arrays/SSBOs/etc.
329 DECLARE_ARRAY(struct ir3_instruction
*, deps
);
332 * From PoV of instruction scheduling, not execution (ie. ignores global/
333 * local distinction):
334 * shared image atomic SSBO everything
335 * barrier()/ - R/W R/W R/W R/W X
336 * groupMemoryBarrier()
337 * memoryBarrier() - R/W R/W
338 * (but only images declared coherent?)
339 * memoryBarrierAtomic() - R/W
340 * memoryBarrierBuffer() - R/W
341 * memoryBarrierImage() - R/W
342 * memoryBarrierShared() - R/W
344 * TODO I think for SSBO/image/shared, in cases where we can determine
345 * which variable is accessed, we don't need to care about accesses to
346 * different variables (unless declared coherent??)
349 IR3_BARRIER_EVERYTHING
= 1 << 0,
350 IR3_BARRIER_SHARED_R
= 1 << 1,
351 IR3_BARRIER_SHARED_W
= 1 << 2,
352 IR3_BARRIER_IMAGE_R
= 1 << 3,
353 IR3_BARRIER_IMAGE_W
= 1 << 4,
354 IR3_BARRIER_BUFFER_R
= 1 << 5,
355 IR3_BARRIER_BUFFER_W
= 1 << 6,
356 IR3_BARRIER_ARRAY_R
= 1 << 7,
357 IR3_BARRIER_ARRAY_W
= 1 << 8,
358 } barrier_class
, barrier_conflict
;
360 /* Entry in ir3_block's instruction list: */
361 struct list_head node
;
363 int use_count
; /* currently just updated/used by cp */
370 static inline struct ir3_instruction
*
371 ir3_neighbor_first(struct ir3_instruction
*instr
)
374 while (instr
->cp
.left
) {
375 instr
= instr
->cp
.left
;
376 if (++cnt
> 0xffff) {
384 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
388 debug_assert(!instr
->cp
.left
);
390 while (instr
->cp
.right
) {
392 instr
= instr
->cp
.right
;
403 struct ir3_compiler
*compiler
;
405 unsigned ninputs
, noutputs
;
406 struct ir3_instruction
**inputs
;
407 struct ir3_instruction
**outputs
;
409 /* Track bary.f (and ldlv) instructions.. this is needed in
410 * scheduling to ensure that all varying fetches happen before
411 * any potential kill instructions. The hw gets grumpy if all
412 * threads in a group are killed before the last bary.f gets
413 * a chance to signal end of input (ei).
415 DECLARE_ARRAY(struct ir3_instruction
*, baryfs
);
417 /* Track all indirect instructions (read and write). To avoid
418 * deadlock scenario where an address register gets scheduled,
419 * but other dependent src instructions cannot be scheduled due
420 * to dependency on a *different* address register value, the
421 * scheduler needs to ensure that all dependencies other than
422 * the instruction other than the address register are scheduled
423 * before the one that writes the address register. Having a
424 * convenient list of instructions that reference some address
425 * register simplifies this.
427 DECLARE_ARRAY(struct ir3_instruction
*, indirects
);
429 /* and same for instructions that consume predicate register: */
430 DECLARE_ARRAY(struct ir3_instruction
*, predicates
);
432 /* Track texture sample instructions which need texture state
433 * patched in (for astc-srgb workaround):
435 DECLARE_ARRAY(struct ir3_instruction
*, astc_srgb
);
437 /* List of blocks: */
438 struct list_head block_list
;
440 /* List of ir3_array's: */
441 struct list_head array_list
;
444 unsigned block_count
, instr_count
;
448 typedef struct nir_register nir_register
;
451 struct list_head node
;
457 /* To avoid array write's from getting DCE'd, keep track of the
458 * most recent write. Any array access depends on the most
459 * recent write. This way, nothing depends on writes after the
460 * last read. But all the writes that happen before that have
461 * something depending on them
463 struct ir3_instruction
*last_write
;
465 /* extra stuff used in RA pass: */
466 unsigned base
; /* base vreg name */
467 unsigned reg
; /* base physical reg */
468 uint16_t start_ip
, end_ip
;
471 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
473 typedef struct nir_block nir_block
;
476 struct list_head node
;
479 const nir_block
*nblock
;
481 struct list_head instr_list
; /* list of ir3_instruction */
483 /* each block has either one or two successors.. in case of
484 * two successors, 'condition' decides which one to follow.
485 * A block preceding an if/else has two successors.
487 struct ir3_instruction
*condition
;
488 struct ir3_block
*successors
[2];
490 unsigned predecessors_count
;
491 struct ir3_block
**predecessors
;
493 uint16_t start_ip
, end_ip
;
495 /* Track instructions which do not write a register but other-
496 * wise must not be discarded (such as kill, stg, etc)
498 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
500 /* used for per-pass extra block data. Mainly used right
501 * now in RA step to track livein/liveout.
510 static inline uint32_t
511 block_id(struct ir3_block
*block
)
514 return block
->serialno
;
516 return (uint32_t)(unsigned long)block
;
520 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
521 unsigned nin
, unsigned nout
);
522 void ir3_destroy(struct ir3
*shader
);
523 void * ir3_assemble(struct ir3
*shader
,
524 struct ir3_info
*info
, uint32_t gpu_id
);
525 void * ir3_alloc(struct ir3
*shader
, int sz
);
527 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
529 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
530 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
531 opc_t opc
, int nreg
);
532 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
533 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
);
534 const char *ir3_instr_name(struct ir3_instruction
*instr
);
536 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
538 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
539 struct ir3_register
*reg
);
541 void ir3_instr_set_address(struct ir3_instruction
*instr
,
542 struct ir3_instruction
*addr
);
544 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
546 if (instr
->flags
& IR3_INSTR_MARK
)
547 return true; /* already visited */
548 instr
->flags
|= IR3_INSTR_MARK
;
552 void ir3_block_clear_mark(struct ir3_block
*block
);
553 void ir3_clear_mark(struct ir3
*shader
);
555 unsigned ir3_count_instructions(struct ir3
*ir
);
557 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
558 struct ir3_register
*reg
)
561 for (i
= 0; i
< instr
->regs_count
; i
++)
562 if (reg
== instr
->regs
[i
])
568 #define MAX_ARRAYS 16
576 static inline uint32_t regid(int num
, int comp
)
578 return (num
<< 2) | (comp
& 0x3);
581 static inline uint32_t reg_num(struct ir3_register
*reg
)
583 return reg
->num
>> 2;
586 static inline uint32_t reg_comp(struct ir3_register
*reg
)
588 return reg
->num
& 0x3;
591 static inline bool is_flow(struct ir3_instruction
*instr
)
593 return (opc_cat(instr
->opc
) == 0);
596 static inline bool is_kill(struct ir3_instruction
*instr
)
598 return instr
->opc
== OPC_KILL
;
601 static inline bool is_nop(struct ir3_instruction
*instr
)
603 return instr
->opc
== OPC_NOP
;
606 /* Is it a non-transformative (ie. not type changing) mov? This can
607 * also include absneg.s/absneg.f, which for the most part can be
608 * treated as a mov (single src argument).
610 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
612 struct ir3_register
*dst
;
614 switch (instr
->opc
) {
616 if (instr
->cat1
.src_type
!= instr
->cat1
.dst_type
)
621 if (instr
->flags
& IR3_INSTR_SAT
)
628 dst
= instr
->regs
[0];
630 /* mov's that write to a0.x or p0.x are special: */
631 if (dst
->num
== regid(REG_P0
, 0))
633 if (dst
->num
== regid(REG_A0
, 0))
636 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
642 static inline bool is_alu(struct ir3_instruction
*instr
)
644 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
647 static inline bool is_sfu(struct ir3_instruction
*instr
)
649 return (opc_cat(instr
->opc
) == 4);
652 static inline bool is_tex(struct ir3_instruction
*instr
)
654 return (opc_cat(instr
->opc
) == 5);
657 static inline bool is_mem(struct ir3_instruction
*instr
)
659 return (opc_cat(instr
->opc
) == 6);
662 static inline bool is_barrier(struct ir3_instruction
*instr
)
664 return (opc_cat(instr
->opc
) == 7);
668 is_store(struct ir3_instruction
*instr
)
670 /* these instructions, the "destination" register is
671 * actually a source, the address to store to.
673 switch (instr
->opc
) {
688 static inline bool is_load(struct ir3_instruction
*instr
)
690 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
)) {
765 debug_assert(!(reg
->instr
&& (reg
->instr
->flags
& IR3_INSTR_UNUSED
)));
771 static inline bool conflicts(struct ir3_instruction
*a
,
772 struct ir3_instruction
*b
)
774 return (a
&& b
) && (a
!= b
);
777 static inline bool reg_gpr(struct ir3_register
*r
)
779 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
781 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
786 static inline type_t
half_type(type_t type
)
789 case TYPE_F32
: return TYPE_F16
;
790 case TYPE_U32
: return TYPE_U16
;
791 case TYPE_S32
: return TYPE_S16
;
802 /* some cat2 instructions (ie. those which are not float) can embed an
805 static inline bool ir3_cat2_int(opc_t opc
)
846 /* map cat2 instruction to valid abs/neg flags: */
847 static inline unsigned ir3_cat2_absneg(opc_t opc
)
864 return IR3_REG_FABS
| IR3_REG_FNEG
;
885 return IR3_REG_SABS
| IR3_REG_SNEG
;
906 /* map cat3 instructions to valid abs/neg flags: */
907 static inline unsigned ir3_cat3_absneg(opc_t opc
)
926 /* neg *may* work on 3rd src.. */
936 #define MASK(n) ((1 << (n)) - 1)
938 /* iterator for an instructions's sources (reg), also returns src #: */
939 #define foreach_src_n(__srcreg, __n, __instr) \
940 if ((__instr)->regs_count) \
941 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
942 if ((__srcreg = (__instr)->regs[__n + 1]))
944 /* iterator for an instructions's sources (reg): */
945 #define foreach_src(__srcreg, __instr) \
946 foreach_src_n(__srcreg, __i, __instr)
948 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
950 unsigned cnt
= instr
->regs_count
+ instr
->deps_count
;
956 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
958 if (n
== (instr
->regs_count
+ instr
->deps_count
))
959 return instr
->address
;
960 if (n
>= instr
->regs_count
)
961 return instr
->deps
[n
- instr
->regs_count
];
962 return ssa(instr
->regs
[n
]);
965 static inline bool __is_false_dep(struct ir3_instruction
*instr
, unsigned n
)
967 if (n
== (instr
->regs_count
+ instr
->deps_count
))
969 if (n
>= instr
->regs_count
)
974 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
976 /* iterator for an instruction's SSA sources (instr), also returns src #: */
977 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
978 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
979 if ((__srcinst = __ssa_src_n(__instr, __n)))
981 /* iterator for an instruction's SSA sources (instr): */
982 #define foreach_ssa_src(__srcinst, __instr) \
983 foreach_ssa_src_n(__srcinst, __i, __instr)
987 void ir3_print(struct ir3
*ir
);
988 void ir3_print_instr(struct ir3_instruction
*instr
);
990 /* depth calculation: */
991 int ir3_delayslots(struct ir3_instruction
*assigner
,
992 struct ir3_instruction
*consumer
, unsigned n
);
993 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
994 void ir3_depth(struct ir3
*ir
);
996 /* copy-propagate: */
997 struct ir3_shader_variant
;
998 void ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1000 /* group neighbors and insert mov's to resolve conflicts: */
1001 void ir3_group(struct ir3
*ir
);
1004 void ir3_sched_add_deps(struct ir3
*ir
);
1005 int ir3_sched(struct ir3
*ir
);
1007 /* register assignment: */
1008 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(void *memctx
);
1009 int ir3_ra(struct ir3
*ir3
, enum shader_t type
,
1010 bool frag_coord
, bool frag_face
);
1013 void ir3_legalize(struct ir3
*ir
, bool *has_samp
, bool *has_ssbo
, int *max_bary
);
1015 /* ************************************************************************* */
1016 /* instruction helpers */
1018 /* creates SSA src of correct type (ie. half vs full precision) */
1019 static inline struct ir3_register
* __ssa_src(struct ir3_instruction
*instr
,
1020 struct ir3_instruction
*src
, unsigned flags
)
1022 struct ir3_register
*reg
;
1023 if (src
->regs
[0]->flags
& IR3_REG_HALF
)
1024 flags
|= IR3_REG_HALF
;
1025 reg
= ir3_reg_create(instr
, 0, IR3_REG_SSA
| flags
);
1030 static inline struct ir3_instruction
*
1031 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
1033 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1034 ir3_reg_create(instr
, 0, 0); /* dst */
1035 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
1036 struct ir3_register
*src_reg
= __ssa_src(instr
, src
, IR3_REG_ARRAY
);
1037 src_reg
->array
= src
->regs
[0]->array
;
1039 __ssa_src(instr
, src
, 0);
1041 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
1042 instr
->cat1
.src_type
= type
;
1043 instr
->cat1
.dst_type
= type
;
1047 static inline struct ir3_instruction
*
1048 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
1049 type_t src_type
, type_t dst_type
)
1051 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1052 ir3_reg_create(instr
, 0, 0); /* dst */
1053 __ssa_src(instr
, src
, 0);
1054 instr
->cat1
.src_type
= src_type
;
1055 instr
->cat1
.dst_type
= dst_type
;
1056 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
1060 static inline struct ir3_instruction
*
1061 ir3_NOP(struct ir3_block
*block
)
1063 return ir3_instr_create(block
, OPC_NOP
);
1066 #define INSTR0(name) \
1067 static inline struct ir3_instruction * \
1068 ir3_##name(struct ir3_block *block) \
1070 struct ir3_instruction *instr = \
1071 ir3_instr_create(block, OPC_##name); \
1075 #define INSTR1(name) \
1076 static inline struct ir3_instruction * \
1077 ir3_##name(struct ir3_block *block, \
1078 struct ir3_instruction *a, unsigned aflags) \
1080 struct ir3_instruction *instr = \
1081 ir3_instr_create(block, OPC_##name); \
1082 ir3_reg_create(instr, 0, 0); /* dst */ \
1083 __ssa_src(instr, a, aflags); \
1087 #define INSTR2(name) \
1088 static inline struct ir3_instruction * \
1089 ir3_##name(struct ir3_block *block, \
1090 struct ir3_instruction *a, unsigned aflags, \
1091 struct ir3_instruction *b, unsigned bflags) \
1093 struct ir3_instruction *instr = \
1094 ir3_instr_create(block, OPC_##name); \
1095 ir3_reg_create(instr, 0, 0); /* dst */ \
1096 __ssa_src(instr, a, aflags); \
1097 __ssa_src(instr, b, bflags); \
1101 #define INSTR3(name) \
1102 static inline struct ir3_instruction * \
1103 ir3_##name(struct ir3_block *block, \
1104 struct ir3_instruction *a, unsigned aflags, \
1105 struct ir3_instruction *b, unsigned bflags, \
1106 struct ir3_instruction *c, unsigned cflags) \
1108 struct ir3_instruction *instr = \
1109 ir3_instr_create(block, OPC_##name); \
1110 ir3_reg_create(instr, 0, 0); /* dst */ \
1111 __ssa_src(instr, a, aflags); \
1112 __ssa_src(instr, b, bflags); \
1113 __ssa_src(instr, c, cflags); \
1117 #define INSTR4(name) \
1118 static inline struct ir3_instruction * \
1119 ir3_##name(struct ir3_block *block, \
1120 struct ir3_instruction *a, unsigned aflags, \
1121 struct ir3_instruction *b, unsigned bflags, \
1122 struct ir3_instruction *c, unsigned cflags, \
1123 struct ir3_instruction *d, unsigned dflags) \
1125 struct ir3_instruction *instr = \
1126 ir3_instr_create2(block, OPC_##name, 5); \
1127 ir3_reg_create(instr, 0, 0); /* dst */ \
1128 __ssa_src(instr, a, aflags); \
1129 __ssa_src(instr, b, bflags); \
1130 __ssa_src(instr, c, cflags); \
1131 __ssa_src(instr, d, dflags); \
1135 #define INSTR4F(f, name) \
1136 static inline struct ir3_instruction * \
1137 ir3_##name##_##f(struct ir3_block *block, \
1138 struct ir3_instruction *a, unsigned aflags, \
1139 struct ir3_instruction *b, unsigned bflags, \
1140 struct ir3_instruction *c, unsigned cflags, \
1141 struct ir3_instruction *d, unsigned dflags) \
1143 struct ir3_instruction *instr = \
1144 ir3_instr_create2(block, OPC_##name, 5); \
1145 ir3_reg_create(instr, 0, 0); /* dst */ \
1146 __ssa_src(instr, a, aflags); \
1147 __ssa_src(instr, b, bflags); \
1148 __ssa_src(instr, c, cflags); \
1149 __ssa_src(instr, d, dflags); \
1150 instr->flags |= IR3_INSTR_##f; \
1154 /* cat0 instructions: */
1160 /* cat2 instructions, most 2 src but some 1 src: */
1208 /* cat3 instructions: */
1226 /* cat4 instructions: */
1235 /* cat5 instructions: */
1239 static inline struct ir3_instruction
*
1240 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1241 unsigned wrmask
, unsigned flags
, unsigned samp
, unsigned tex
,
1242 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1244 struct ir3_instruction
*sam
;
1245 struct ir3_register
*reg
;
1247 sam
= ir3_instr_create(block
, opc
);
1248 sam
->flags
|= flags
;
1249 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1251 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1252 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1256 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1258 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1260 sam
->cat5
.samp
= samp
;
1261 sam
->cat5
.tex
= tex
;
1262 sam
->cat5
.type
= type
;
1267 /* cat6 instructions: */
1283 INSTR2(ATOMIC_CMPXCHG
)
1289 INSTR4F(G
, ATOMIC_ADD
)
1290 INSTR4F(G
, ATOMIC_SUB
)
1291 INSTR4F(G
, ATOMIC_XCHG
)
1292 INSTR4F(G
, ATOMIC_INC
)
1293 INSTR4F(G
, ATOMIC_DEC
)
1294 INSTR4F(G
, ATOMIC_CMPXCHG
)
1295 INSTR4F(G
, ATOMIC_MIN
)
1296 INSTR4F(G
, ATOMIC_MAX
)
1297 INSTR4F(G
, ATOMIC_AND
)
1298 INSTR4F(G
, ATOMIC_OR
)
1299 INSTR4F(G
, ATOMIC_XOR
)
1301 /* cat7 instructions: */
1305 /* ************************************************************************* */
1306 /* split this out or find some helper to use.. like main/bitset.h.. */
1312 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1314 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1316 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1317 debug_assert(num
< MAX_REG
);
1318 if (reg
->flags
& IR3_REG_HALF
)
1323 static inline void regmask_init(regmask_t
*regmask
)
1325 memset(regmask
, 0, sizeof(*regmask
));
1328 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1330 unsigned idx
= regmask_idx(reg
);
1331 if (reg
->flags
& IR3_REG_RELATIV
) {
1333 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1334 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1337 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1339 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1343 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1346 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1347 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1350 /* set bits in a if not set in b, conceptually:
1353 static inline void regmask_set_if_not(regmask_t
*a
,
1354 struct ir3_register
*reg
, regmask_t
*b
)
1356 unsigned idx
= regmask_idx(reg
);
1357 if (reg
->flags
& IR3_REG_RELATIV
) {
1359 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1360 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1361 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1364 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1366 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1367 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1371 static inline bool regmask_get(regmask_t
*regmask
,
1372 struct ir3_register
*reg
)
1374 unsigned idx
= regmask_idx(reg
);
1375 if (reg
->flags
& IR3_REG_RELATIV
) {
1377 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1378 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1382 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1384 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1390 /* ************************************************************************* */