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"
35 #include "util/u_debug.h"
37 #include "instr-a3xx.h"
39 /* low level intermediate representation of an adreno shader program */
43 struct ir3_instruction
;
49 uint16_t instrs_count
; /* expanded to account for rpt's */
50 /* NOTE: max_reg, etc, does not include registers not touched
51 * by the shader (ie. vertex fetched via VFD_DECODE but not
54 int8_t max_reg
; /* highest GPR # used by shader */
58 /* number of sync bits: */
64 IR3_REG_CONST
= 0x001,
65 IR3_REG_IMMED
= 0x002,
67 /* high registers are used for some things in compute shaders,
68 * for example. Seems to be for things that are global to all
69 * threads in a wave, so possibly these are global/shared by
70 * all the threads in the wave?
73 IR3_REG_RELATIV
= 0x010,
75 /* Most instructions, it seems, can do float abs/neg but not
76 * integer. The CP pass needs to know what is intended (int or
77 * float) in order to do the right thing. For this reason the
78 * abs/neg flags are split out into float and int variants. In
79 * addition, .b (bitwise) operations, the negate is actually a
80 * bitwise not, so split that out into a new flag to make it
89 IR3_REG_POS_INF
= 0x1000,
90 /* (ei) flag, end-input? Set on last bary, presumably to signal
91 * that the shader needs no more input:
94 /* meta-flags, for intermediate stages of IR, ie.
95 * before register assignment is done:
97 IR3_REG_SSA
= 0x4000, /* 'instr' is ptr to assigning instr */
98 IR3_REG_ARRAY
= 0x8000,
102 bool merged
: 1; /* half-regs conflict with full regs (ie >= a6xx) */
105 * the component is in the low two bits of the reg #, so
106 * rN.x becomes: (N << 2) | x
121 /* For IR3_REG_SSA, src registers contain ptr back to assigning
124 * For IR3_REG_ARRAY, the pointer is back to the last dependent
125 * array access (although the net effect is the same, it points
126 * back to a previous instruction that we depend on).
128 struct ir3_instruction
*instr
;
131 /* used for cat5 instructions, but also for internal/IR level
132 * tracking of what registers are read/written by an instruction.
133 * wrmask may be a bad name since it is used to represent both
134 * src and dst that touch multiple adjacent registers.
137 /* for relative addressing, 32bits for array size is too small,
138 * but otoh we don't need to deal with disjoint sets, so instead
139 * use a simple size field (number of scalar components).
146 * Stupid/simple growable array implementation:
148 #define DECLARE_ARRAY(type, name) \
149 unsigned name ## _count, name ## _sz; \
152 #define array_insert(ctx, arr, val) do { \
153 if (arr ## _count == arr ## _sz) { \
154 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
155 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
157 arr[arr ##_count++] = val; \
160 struct ir3_instruction
{
161 struct ir3_block
*block
;
164 /* (sy) flag is set on first instruction, and after sample
165 * instructions (probably just on RAW hazard).
167 IR3_INSTR_SY
= 0x001,
168 /* (ss) flag is set on first instruction, and first instruction
169 * to depend on the result of "long" instructions (RAW hazard):
171 * rcp, rsq, log2, exp2, sin, cos, sqrt
173 * It seems to synchronize until all in-flight instructions are
174 * completed, for example:
177 * add.f hr2.z, (neg)hr2.z, hc0.y
178 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
181 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
183 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
184 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
185 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
187 * The last mul.f does not have (ss) set, presumably because the
188 * (ss) on the previous instruction does the job.
190 * The blob driver also seems to set it on WAR hazards, although
191 * not really clear if this is needed or just blob compiler being
192 * sloppy. So far I haven't found a case where removing the (ss)
193 * causes problems for WAR hazard, but I could just be getting
197 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
200 IR3_INSTR_SS
= 0x002,
201 /* (jp) flag is set on jump targets:
203 IR3_INSTR_JP
= 0x004,
204 IR3_INSTR_UL
= 0x008,
205 IR3_INSTR_3D
= 0x010,
210 IR3_INSTR_S2EN
= 0x200,
212 IR3_INSTR_SAT
= 0x800,
213 /* meta-flags, for intermediate stages of IR, ie.
214 * before register assignment is done:
216 IR3_INSTR_MARK
= 0x1000,
217 IR3_INSTR_UNUSED
= 0x2000,
225 struct ir3_register
**regs
;
231 struct ir3_block
*target
;
234 type_t src_type
, dst_type
;
254 int iim_val
: 3; /* for ldgb/stgb, # of components */
259 unsigned w
: 1; /* write */
260 unsigned r
: 1; /* read */
261 unsigned l
: 1; /* local */
262 unsigned g
: 1; /* global */
264 /* for meta-instructions, just used to hold extra data
265 * before instruction scheduling, etc
268 int off
; /* component/offset */
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.
294 * TODO we should remove the per-pass data like this and 'use_count'
295 * and do something similar to what RA does w/ ir3_ra_instr_data..
296 * ie. use the ir3_count_instructions pass, and then use instr->ip
297 * to index into a table of pass-private data.
301 int sun
; /* Sethi–Ullman number, used by sched */
302 int use_count
; /* currently just updated/used by cp */
304 /* Used during CP and RA stages. For fanin and shader inputs/
305 * outputs where we need a sequence of consecutive registers,
306 * keep track of each src instructions left (ie 'n-1') and right
307 * (ie 'n+1') neighbor. The front-end must insert enough mov's
308 * to ensure that each instruction has at most one left and at
309 * most one right neighbor. During the copy-propagation pass,
310 * we only remove mov's when we can preserve this constraint.
311 * And during the RA stage, we use the neighbor information to
312 * allocate a block of registers in one shot.
314 * TODO: maybe just add something like:
315 * struct ir3_instruction_ref {
316 * struct ir3_instruction *instr;
320 * Or can we get away without the refcnt stuff? It seems like
321 * it should be overkill.. the problem is if, potentially after
322 * already eliminating some mov's, if you have a single mov that
323 * needs to be grouped with it's neighbors in two different
324 * places (ex. shader output and a fanin).
327 struct ir3_instruction
*left
, *right
;
328 uint16_t left_cnt
, right_cnt
;
331 /* an instruction can reference at most one address register amongst
332 * it's src/dst registers. Beyond that, you need to insert mov's.
334 * NOTE: do not write this directly, use ir3_instr_set_address()
336 struct ir3_instruction
*address
;
338 /* Tracking for additional dependent instructions. Used to handle
339 * barriers, WAR hazards for arrays/SSBOs/etc.
341 DECLARE_ARRAY(struct ir3_instruction
*, deps
);
344 * From PoV of instruction scheduling, not execution (ie. ignores global/
345 * local distinction):
346 * shared image atomic SSBO everything
347 * barrier()/ - R/W R/W R/W R/W X
348 * groupMemoryBarrier()
349 * memoryBarrier() - R/W R/W
350 * (but only images declared coherent?)
351 * memoryBarrierAtomic() - R/W
352 * memoryBarrierBuffer() - R/W
353 * memoryBarrierImage() - R/W
354 * memoryBarrierShared() - R/W
356 * TODO I think for SSBO/image/shared, in cases where we can determine
357 * which variable is accessed, we don't need to care about accesses to
358 * different variables (unless declared coherent??)
361 IR3_BARRIER_EVERYTHING
= 1 << 0,
362 IR3_BARRIER_SHARED_R
= 1 << 1,
363 IR3_BARRIER_SHARED_W
= 1 << 2,
364 IR3_BARRIER_IMAGE_R
= 1 << 3,
365 IR3_BARRIER_IMAGE_W
= 1 << 4,
366 IR3_BARRIER_BUFFER_R
= 1 << 5,
367 IR3_BARRIER_BUFFER_W
= 1 << 6,
368 IR3_BARRIER_ARRAY_R
= 1 << 7,
369 IR3_BARRIER_ARRAY_W
= 1 << 8,
370 } barrier_class
, barrier_conflict
;
372 /* Entry in ir3_block's instruction list: */
373 struct list_head node
;
380 static inline struct ir3_instruction
*
381 ir3_neighbor_first(struct ir3_instruction
*instr
)
384 while (instr
->cp
.left
) {
385 instr
= instr
->cp
.left
;
386 if (++cnt
> 0xffff) {
394 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
398 debug_assert(!instr
->cp
.left
);
400 while (instr
->cp
.right
) {
402 instr
= instr
->cp
.right
;
413 struct ir3_compiler
*compiler
;
414 gl_shader_stage type
;
416 unsigned ninputs
, noutputs
;
417 struct ir3_instruction
**inputs
;
418 struct ir3_instruction
**outputs
;
420 /* Track bary.f (and ldlv) instructions.. this is needed in
421 * scheduling to ensure that all varying fetches happen before
422 * any potential kill instructions. The hw gets grumpy if all
423 * threads in a group are killed before the last bary.f gets
424 * a chance to signal end of input (ei).
426 DECLARE_ARRAY(struct ir3_instruction
*, baryfs
);
428 /* Track all indirect instructions (read and write). To avoid
429 * deadlock scenario where an address register gets scheduled,
430 * but other dependent src instructions cannot be scheduled due
431 * to dependency on a *different* address register value, the
432 * scheduler needs to ensure that all dependencies other than
433 * the instruction other than the address register are scheduled
434 * before the one that writes the address register. Having a
435 * convenient list of instructions that reference some address
436 * register simplifies this.
438 DECLARE_ARRAY(struct ir3_instruction
*, indirects
);
440 /* and same for instructions that consume predicate register: */
441 DECLARE_ARRAY(struct ir3_instruction
*, predicates
);
443 /* Track texture sample instructions which need texture state
444 * patched in (for astc-srgb workaround):
446 DECLARE_ARRAY(struct ir3_instruction
*, astc_srgb
);
448 /* List of blocks: */
449 struct list_head block_list
;
451 /* List of ir3_array's: */
452 struct list_head array_list
;
454 unsigned max_sun
; /* max Sethi–Ullman number */
457 unsigned block_count
, instr_count
;
462 struct list_head node
;
466 struct nir_register
*r
;
468 /* To avoid array write's from getting DCE'd, keep track of the
469 * most recent write. Any array access depends on the most
470 * recent write. This way, nothing depends on writes after the
471 * last read. But all the writes that happen before that have
472 * something depending on them
474 struct ir3_instruction
*last_write
;
476 /* extra stuff used in RA pass: */
477 unsigned base
; /* base vreg name */
478 unsigned reg
; /* base physical reg */
479 uint16_t start_ip
, end_ip
;
482 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
485 struct list_head node
;
488 const struct nir_block
*nblock
;
490 struct list_head instr_list
; /* list of ir3_instruction */
492 /* each block has either one or two successors.. in case of
493 * two successors, 'condition' decides which one to follow.
494 * A block preceding an if/else has two successors.
496 struct ir3_instruction
*condition
;
497 struct ir3_block
*successors
[2];
499 struct set
*predecessors
; /* set of ir3_block */
501 uint16_t start_ip
, end_ip
;
503 /* Track instructions which do not write a register but other-
504 * wise must not be discarded (such as kill, stg, etc)
506 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
508 /* used for per-pass extra block data. Mainly used right
509 * now in RA step to track livein/liveout.
518 static inline uint32_t
519 block_id(struct ir3_block
*block
)
522 return block
->serialno
;
524 return (uint32_t)(unsigned long)block
;
528 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
529 gl_shader_stage type
, unsigned nin
, unsigned nout
);
530 void ir3_destroy(struct ir3
*shader
);
531 void * ir3_assemble(struct ir3
*shader
,
532 struct ir3_info
*info
, uint32_t gpu_id
);
533 void * ir3_alloc(struct ir3
*shader
, int sz
);
535 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
537 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
538 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
539 opc_t opc
, int nreg
);
540 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
541 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
);
542 const char *ir3_instr_name(struct ir3_instruction
*instr
);
544 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
546 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
547 struct ir3_register
*reg
);
549 void ir3_instr_set_address(struct ir3_instruction
*instr
,
550 struct ir3_instruction
*addr
);
552 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
554 if (instr
->flags
& IR3_INSTR_MARK
)
555 return true; /* already visited */
556 instr
->flags
|= IR3_INSTR_MARK
;
560 void ir3_block_clear_mark(struct ir3_block
*block
);
561 void ir3_clear_mark(struct ir3
*shader
);
563 unsigned ir3_count_instructions(struct ir3
*ir
);
565 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
566 struct ir3_register
*reg
)
569 for (i
= 0; i
< instr
->regs_count
; i
++)
570 if (reg
== instr
->regs
[i
])
576 #define MAX_ARRAYS 16
584 static inline uint32_t regid(int num
, int comp
)
586 return (num
<< 2) | (comp
& 0x3);
589 static inline uint32_t reg_num(struct ir3_register
*reg
)
591 return reg
->num
>> 2;
594 static inline uint32_t reg_comp(struct ir3_register
*reg
)
596 return reg
->num
& 0x3;
599 static inline bool is_flow(struct ir3_instruction
*instr
)
601 return (opc_cat(instr
->opc
) == 0);
604 static inline bool is_kill(struct ir3_instruction
*instr
)
606 return instr
->opc
== OPC_KILL
;
609 static inline bool is_nop(struct ir3_instruction
*instr
)
611 return instr
->opc
== OPC_NOP
;
614 static inline bool is_same_type_reg(struct ir3_register
*reg1
,
615 struct ir3_register
*reg2
)
617 unsigned type_reg1
= (reg1
->flags
& (IR3_REG_HIGH
| IR3_REG_HALF
));
618 unsigned type_reg2
= (reg2
->flags
& (IR3_REG_HIGH
| IR3_REG_HALF
));
620 if (type_reg1
^ type_reg2
)
626 /* Is it a non-transformative (ie. not type changing) mov? This can
627 * also include absneg.s/absneg.f, which for the most part can be
628 * treated as a mov (single src argument).
630 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
632 struct ir3_register
*dst
;
634 switch (instr
->opc
) {
636 if (instr
->cat1
.src_type
!= instr
->cat1
.dst_type
)
641 if (instr
->flags
& IR3_INSTR_SAT
)
643 /* If the type of dest reg and src reg are different,
644 * it shouldn't be considered as same type mov */
645 if (!is_same_type_reg(instr
->regs
[0], instr
->regs
[1]))
652 dst
= instr
->regs
[0];
654 /* mov's that write to a0.x or p0.x are special: */
655 if (dst
->num
== regid(REG_P0
, 0))
657 if (dst
->num
== regid(REG_A0
, 0))
660 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
666 static inline bool is_alu(struct ir3_instruction
*instr
)
668 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
671 static inline bool is_sfu(struct ir3_instruction
*instr
)
673 return (opc_cat(instr
->opc
) == 4);
676 static inline bool is_tex(struct ir3_instruction
*instr
)
678 return (opc_cat(instr
->opc
) == 5);
681 static inline bool is_mem(struct ir3_instruction
*instr
)
683 return (opc_cat(instr
->opc
) == 6);
686 static inline bool is_barrier(struct ir3_instruction
*instr
)
688 return (opc_cat(instr
->opc
) == 7);
692 is_store(struct ir3_instruction
*instr
)
694 /* these instructions, the "destination" register is
695 * actually a source, the address to store to.
697 switch (instr
->opc
) {
712 static inline bool is_load(struct ir3_instruction
*instr
)
714 switch (instr
->opc
) {
724 /* probably some others too.. */
731 static inline bool is_input(struct ir3_instruction
*instr
)
733 /* in some cases, ldlv is used to fetch varying without
734 * interpolation.. fortunately inloc is the first src
735 * register in either case
737 switch (instr
->opc
) {
746 static inline bool is_bool(struct ir3_instruction
*instr
)
748 switch (instr
->opc
) {
758 static inline bool is_meta(struct ir3_instruction
*instr
)
760 /* TODO how should we count PHI (and maybe fan-in/out) which
761 * might actually contribute some instructions to the final
764 return (opc_cat(instr
->opc
) == -1);
767 static inline unsigned dest_regs(struct ir3_instruction
*instr
)
769 if ((instr
->regs_count
== 0) || is_store(instr
))
772 return util_last_bit(instr
->regs
[0]->wrmask
);
775 static inline bool writes_addr(struct ir3_instruction
*instr
)
777 if (instr
->regs_count
> 0) {
778 struct ir3_register
*dst
= instr
->regs
[0];
779 return reg_num(dst
) == REG_A0
;
784 static inline bool writes_pred(struct ir3_instruction
*instr
)
786 if (instr
->regs_count
> 0) {
787 struct ir3_register
*dst
= instr
->regs
[0];
788 return reg_num(dst
) == REG_P0
;
793 /* returns defining instruction for reg */
794 /* TODO better name */
795 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
797 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
803 static inline bool conflicts(struct ir3_instruction
*a
,
804 struct ir3_instruction
*b
)
806 return (a
&& b
) && (a
!= b
);
809 static inline bool reg_gpr(struct ir3_register
*r
)
811 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
813 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
818 static inline type_t
half_type(type_t type
)
821 case TYPE_F32
: return TYPE_F16
;
822 case TYPE_U32
: return TYPE_U16
;
823 case TYPE_S32
: return TYPE_S16
;
834 /* some cat2 instructions (ie. those which are not float) can embed an
837 static inline bool ir3_cat2_int(opc_t opc
)
877 static inline bool ir3_cat2_float(opc_t opc
)
900 static inline bool ir3_cat3_float(opc_t opc
)
913 /* map cat2 instruction to valid abs/neg flags: */
914 static inline unsigned ir3_cat2_absneg(opc_t opc
)
931 return IR3_REG_FABS
| IR3_REG_FNEG
;
952 return IR3_REG_SABS
| IR3_REG_SNEG
;
973 /* map cat3 instructions to valid abs/neg flags: */
974 static inline unsigned ir3_cat3_absneg(opc_t opc
)
993 /* neg *may* work on 3rd src.. */
1003 #define MASK(n) ((1 << (n)) - 1)
1005 /* iterator for an instructions's sources (reg), also returns src #: */
1006 #define foreach_src_n(__srcreg, __n, __instr) \
1007 if ((__instr)->regs_count) \
1008 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
1009 if ((__srcreg = (__instr)->regs[__n + 1]))
1011 /* iterator for an instructions's sources (reg): */
1012 #define foreach_src(__srcreg, __instr) \
1013 foreach_src_n(__srcreg, __i, __instr)
1015 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
1017 unsigned cnt
= instr
->regs_count
+ instr
->deps_count
;
1023 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
1025 if (n
== (instr
->regs_count
+ instr
->deps_count
))
1026 return instr
->address
;
1027 if (n
>= instr
->regs_count
)
1028 return instr
->deps
[n
- instr
->regs_count
];
1029 return ssa(instr
->regs
[n
]);
1032 static inline bool __is_false_dep(struct ir3_instruction
*instr
, unsigned n
)
1034 if (n
== (instr
->regs_count
+ instr
->deps_count
))
1036 if (n
>= instr
->regs_count
)
1041 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
1043 /* iterator for an instruction's SSA sources (instr), also returns src #: */
1044 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
1045 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
1046 if ((__srcinst = __ssa_src_n(__instr, __n)))
1048 /* iterator for an instruction's SSA sources (instr): */
1049 #define foreach_ssa_src(__srcinst, __instr) \
1050 foreach_ssa_src_n(__srcinst, __i, __instr)
1054 void ir3_print(struct ir3
*ir
);
1055 void ir3_print_instr(struct ir3_instruction
*instr
);
1057 /* depth calculation: */
1058 int ir3_delayslots(struct ir3_instruction
*assigner
,
1059 struct ir3_instruction
*consumer
, unsigned n
);
1060 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
1061 void ir3_depth(struct ir3
*ir
);
1063 /* copy-propagate: */
1064 struct ir3_shader_variant
;
1065 void ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1067 /* group neighbors and insert mov's to resolve conflicts: */
1068 void ir3_group(struct ir3
*ir
);
1070 /* Sethi–Ullman numbering: */
1071 void ir3_sun(struct ir3
*ir
);
1074 void ir3_sched_add_deps(struct ir3
*ir
);
1075 int ir3_sched(struct ir3
*ir
);
1077 void ir3_a6xx_fixup_atomic_dests(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1079 /* register assignment: */
1080 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(struct ir3_compiler
*compiler
);
1081 int ir3_ra(struct ir3_shader_variant
*v
, struct ir3_instruction
**precolor
, unsigned nprecolor
);
1084 void ir3_legalize(struct ir3
*ir
, bool *has_ssbo
, bool *need_pixlod
, int *max_bary
);
1086 /* ************************************************************************* */
1087 /* instruction helpers */
1089 static inline struct ir3_instruction
*
1090 create_immed_typed(struct ir3_block
*block
, uint32_t val
, type_t type
)
1092 struct ir3_instruction
*mov
;
1093 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1095 mov
= ir3_instr_create(block
, OPC_MOV
);
1096 mov
->cat1
.src_type
= type
;
1097 mov
->cat1
.dst_type
= type
;
1098 ir3_reg_create(mov
, 0, flags
);
1099 ir3_reg_create(mov
, 0, IR3_REG_IMMED
)->uim_val
= val
;
1104 static inline struct ir3_instruction
*
1105 create_immed(struct ir3_block
*block
, uint32_t val
)
1107 return create_immed_typed(block
, val
, TYPE_U32
);
1110 static inline struct ir3_instruction
*
1111 create_uniform_typed(struct ir3_block
*block
, unsigned n
, type_t type
)
1113 struct ir3_instruction
*mov
;
1114 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1116 mov
= ir3_instr_create(block
, OPC_MOV
);
1117 mov
->cat1
.src_type
= type
;
1118 mov
->cat1
.dst_type
= type
;
1119 ir3_reg_create(mov
, 0, flags
);
1120 ir3_reg_create(mov
, n
, IR3_REG_CONST
| flags
);
1125 static inline struct ir3_instruction
*
1126 create_uniform(struct ir3_block
*block
, unsigned n
)
1128 return create_uniform_typed(block
, n
, TYPE_F32
);
1131 static inline struct ir3_instruction
*
1132 create_uniform_indirect(struct ir3_block
*block
, int n
,
1133 struct ir3_instruction
*address
)
1135 struct ir3_instruction
*mov
;
1137 mov
= ir3_instr_create(block
, OPC_MOV
);
1138 mov
->cat1
.src_type
= TYPE_U32
;
1139 mov
->cat1
.dst_type
= TYPE_U32
;
1140 ir3_reg_create(mov
, 0, 0);
1141 ir3_reg_create(mov
, 0, IR3_REG_CONST
| IR3_REG_RELATIV
)->array
.offset
= n
;
1143 ir3_instr_set_address(mov
, address
);
1148 /* creates SSA src of correct type (ie. half vs full precision) */
1149 static inline struct ir3_register
* __ssa_src(struct ir3_instruction
*instr
,
1150 struct ir3_instruction
*src
, unsigned flags
)
1152 struct ir3_register
*reg
;
1153 if (src
->regs
[0]->flags
& IR3_REG_HALF
)
1154 flags
|= IR3_REG_HALF
;
1155 reg
= ir3_reg_create(instr
, 0, IR3_REG_SSA
| flags
);
1157 reg
->wrmask
= src
->regs
[0]->wrmask
;
1161 static inline struct ir3_instruction
*
1162 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
1164 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1165 ir3_reg_create(instr
, 0, 0); /* dst */
1166 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
1167 struct ir3_register
*src_reg
= __ssa_src(instr
, src
, IR3_REG_ARRAY
);
1168 src_reg
->array
= src
->regs
[0]->array
;
1170 __ssa_src(instr
, src
, src
->regs
[0]->flags
& IR3_REG_HIGH
);
1172 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
1173 instr
->cat1
.src_type
= type
;
1174 instr
->cat1
.dst_type
= type
;
1178 static inline struct ir3_instruction
*
1179 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
1180 type_t src_type
, type_t dst_type
)
1182 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1183 unsigned dst_flags
= (type_size(dst_type
) < 32) ? IR3_REG_HALF
: 0;
1184 unsigned src_flags
= (type_size(src_type
) < 32) ? IR3_REG_HALF
: 0;
1186 debug_assert((src
->regs
[0]->flags
& IR3_REG_HALF
) == src_flags
);
1188 ir3_reg_create(instr
, 0, dst_flags
); /* dst */
1189 __ssa_src(instr
, src
, 0);
1190 instr
->cat1
.src_type
= src_type
;
1191 instr
->cat1
.dst_type
= dst_type
;
1192 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
1196 static inline struct ir3_instruction
*
1197 ir3_NOP(struct ir3_block
*block
)
1199 return ir3_instr_create(block
, OPC_NOP
);
1202 #define IR3_INSTR_0 0
1204 #define __INSTR0(flag, name, opc) \
1205 static inline struct ir3_instruction * \
1206 ir3_##name(struct ir3_block *block) \
1208 struct ir3_instruction *instr = \
1209 ir3_instr_create(block, opc); \
1210 instr->flags |= flag; \
1213 #define INSTR0F(f, name) __INSTR0(IR3_INSTR_##f, name##_##f, OPC_##name)
1214 #define INSTR0(name) __INSTR0(0, name, OPC_##name)
1216 #define __INSTR1(flag, name, opc) \
1217 static inline struct ir3_instruction * \
1218 ir3_##name(struct ir3_block *block, \
1219 struct ir3_instruction *a, unsigned aflags) \
1221 struct ir3_instruction *instr = \
1222 ir3_instr_create(block, opc); \
1223 ir3_reg_create(instr, 0, 0); /* dst */ \
1224 __ssa_src(instr, a, aflags); \
1225 instr->flags |= flag; \
1228 #define INSTR1F(f, name) __INSTR1(IR3_INSTR_##f, name##_##f, OPC_##name)
1229 #define INSTR1(name) __INSTR1(0, name, OPC_##name)
1231 #define __INSTR2(flag, name, opc) \
1232 static inline struct ir3_instruction * \
1233 ir3_##name(struct ir3_block *block, \
1234 struct ir3_instruction *a, unsigned aflags, \
1235 struct ir3_instruction *b, unsigned bflags) \
1237 struct ir3_instruction *instr = \
1238 ir3_instr_create(block, opc); \
1239 ir3_reg_create(instr, 0, 0); /* dst */ \
1240 __ssa_src(instr, a, aflags); \
1241 __ssa_src(instr, b, bflags); \
1242 instr->flags |= flag; \
1245 #define INSTR2F(f, name) __INSTR2(IR3_INSTR_##f, name##_##f, OPC_##name)
1246 #define INSTR2(name) __INSTR2(0, name, OPC_##name)
1248 #define __INSTR3(flag, name, opc) \
1249 static inline struct ir3_instruction * \
1250 ir3_##name(struct ir3_block *block, \
1251 struct ir3_instruction *a, unsigned aflags, \
1252 struct ir3_instruction *b, unsigned bflags, \
1253 struct ir3_instruction *c, unsigned cflags) \
1255 struct ir3_instruction *instr = \
1256 ir3_instr_create2(block, opc, 4); \
1257 ir3_reg_create(instr, 0, 0); /* dst */ \
1258 __ssa_src(instr, a, aflags); \
1259 __ssa_src(instr, b, bflags); \
1260 __ssa_src(instr, c, cflags); \
1261 instr->flags |= flag; \
1264 #define INSTR3F(f, name) __INSTR3(IR3_INSTR_##f, name##_##f, OPC_##name)
1265 #define INSTR3(name) __INSTR3(0, name, OPC_##name)
1267 #define __INSTR4(flag, name, opc) \
1268 static inline struct ir3_instruction * \
1269 ir3_##name(struct ir3_block *block, \
1270 struct ir3_instruction *a, unsigned aflags, \
1271 struct ir3_instruction *b, unsigned bflags, \
1272 struct ir3_instruction *c, unsigned cflags, \
1273 struct ir3_instruction *d, unsigned dflags) \
1275 struct ir3_instruction *instr = \
1276 ir3_instr_create2(block, opc, 5); \
1277 ir3_reg_create(instr, 0, 0); /* dst */ \
1278 __ssa_src(instr, a, aflags); \
1279 __ssa_src(instr, b, bflags); \
1280 __ssa_src(instr, c, cflags); \
1281 __ssa_src(instr, d, dflags); \
1282 instr->flags |= flag; \
1285 #define INSTR4F(f, name) __INSTR4(IR3_INSTR_##f, name##_##f, OPC_##name)
1286 #define INSTR4(name) __INSTR4(0, name, OPC_##name)
1288 /* cat0 instructions: */
1296 /* cat2 instructions, most 2 src but some 1 src: */
1344 /* cat3 instructions: */
1362 /* cat4 instructions: */
1371 /* cat5 instructions: */
1378 static inline struct ir3_instruction
*
1379 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1380 unsigned wrmask
, unsigned flags
, struct ir3_instruction
*samp_tex
,
1381 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1383 struct ir3_instruction
*sam
;
1384 struct ir3_register
*reg
;
1386 sam
= ir3_instr_create(block
, opc
);
1387 sam
->flags
|= flags
| IR3_INSTR_S2EN
;
1388 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1389 __ssa_src(sam
, samp_tex
, IR3_REG_HALF
);
1391 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1392 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1396 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1398 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1400 sam
->cat5
.type
= type
;
1405 /* cat6 instructions: */
1420 INSTR2(ATOMIC_CMPXCHG
)
1429 INSTR3F(G
, ATOMIC_ADD
)
1430 INSTR3F(G
, ATOMIC_SUB
)
1431 INSTR3F(G
, ATOMIC_XCHG
)
1432 INSTR3F(G
, ATOMIC_INC
)
1433 INSTR3F(G
, ATOMIC_DEC
)
1434 INSTR3F(G
, ATOMIC_CMPXCHG
)
1435 INSTR3F(G
, ATOMIC_MIN
)
1436 INSTR3F(G
, ATOMIC_MAX
)
1437 INSTR3F(G
, ATOMIC_AND
)
1438 INSTR3F(G
, ATOMIC_OR
)
1439 INSTR3F(G
, ATOMIC_XOR
)
1444 INSTR4F(G
, ATOMIC_ADD
)
1445 INSTR4F(G
, ATOMIC_SUB
)
1446 INSTR4F(G
, ATOMIC_XCHG
)
1447 INSTR4F(G
, ATOMIC_INC
)
1448 INSTR4F(G
, ATOMIC_DEC
)
1449 INSTR4F(G
, ATOMIC_CMPXCHG
)
1450 INSTR4F(G
, ATOMIC_MIN
)
1451 INSTR4F(G
, ATOMIC_MAX
)
1452 INSTR4F(G
, ATOMIC_AND
)
1453 INSTR4F(G
, ATOMIC_OR
)
1454 INSTR4F(G
, ATOMIC_XOR
)
1457 /* cat7 instructions: */
1461 /* ************************************************************************* */
1462 /* split this out or find some helper to use.. like main/bitset.h.. */
1468 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1470 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1472 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1473 debug_assert(num
< MAX_REG
);
1474 if (reg
->flags
& IR3_REG_HALF
) {
1484 static inline void regmask_init(regmask_t
*regmask
)
1486 memset(regmask
, 0, sizeof(*regmask
));
1489 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1491 unsigned idx
= regmask_idx(reg
);
1492 if (reg
->flags
& IR3_REG_RELATIV
) {
1494 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1495 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1498 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1500 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1504 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1507 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1508 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1511 /* set bits in a if not set in b, conceptually:
1514 static inline void regmask_set_if_not(regmask_t
*a
,
1515 struct ir3_register
*reg
, regmask_t
*b
)
1517 unsigned idx
= regmask_idx(reg
);
1518 if (reg
->flags
& IR3_REG_RELATIV
) {
1520 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1521 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1522 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1525 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1527 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1528 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1532 static inline bool regmask_get(regmask_t
*regmask
,
1533 struct ir3_register
*reg
)
1535 unsigned idx
= regmask_idx(reg
);
1536 if (reg
->flags
& IR3_REG_RELATIV
) {
1538 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1539 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1543 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1545 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1551 /* ************************************************************************* */