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 */
271 /* for sysvals, identifies the sysval type. Mostly so we can
272 * identify the special cases where a sysval should not be DCE'd
273 * (currently, just pre-fs texture fetch)
275 gl_system_value sysval
;
279 /* transient values used during various algorithms: */
281 /* The instruction depth is the max dependency distance to output.
283 * You can also think of it as the "cost", if we did any sort of
284 * optimization for register footprint. Ie. a value that is just
285 * result of moving a const to a reg would have a low cost, so to
286 * it could make sense to duplicate the instruction at various
287 * points where the result is needed to reduce register footprint.
290 /* When we get to the RA stage, we no longer need depth, but
291 * we do need instruction's position/name:
299 /* used for per-pass extra instruction data.
301 * TODO we should remove the per-pass data like this and 'use_count'
302 * and do something similar to what RA does w/ ir3_ra_instr_data..
303 * ie. use the ir3_count_instructions pass, and then use instr->ip
304 * to index into a table of pass-private data.
308 int sun
; /* Sethi–Ullman number, used by sched */
309 int use_count
; /* currently just updated/used by cp */
311 /* Used during CP and RA stages. For fanin and shader inputs/
312 * outputs where we need a sequence of consecutive registers,
313 * keep track of each src instructions left (ie 'n-1') and right
314 * (ie 'n+1') neighbor. The front-end must insert enough mov's
315 * to ensure that each instruction has at most one left and at
316 * most one right neighbor. During the copy-propagation pass,
317 * we only remove mov's when we can preserve this constraint.
318 * And during the RA stage, we use the neighbor information to
319 * allocate a block of registers in one shot.
321 * TODO: maybe just add something like:
322 * struct ir3_instruction_ref {
323 * struct ir3_instruction *instr;
327 * Or can we get away without the refcnt stuff? It seems like
328 * it should be overkill.. the problem is if, potentially after
329 * already eliminating some mov's, if you have a single mov that
330 * needs to be grouped with it's neighbors in two different
331 * places (ex. shader output and a fanin).
334 struct ir3_instruction
*left
, *right
;
335 uint16_t left_cnt
, right_cnt
;
338 /* an instruction can reference at most one address register amongst
339 * it's src/dst registers. Beyond that, you need to insert mov's.
341 * NOTE: do not write this directly, use ir3_instr_set_address()
343 struct ir3_instruction
*address
;
345 /* Tracking for additional dependent instructions. Used to handle
346 * barriers, WAR hazards for arrays/SSBOs/etc.
348 DECLARE_ARRAY(struct ir3_instruction
*, deps
);
351 * From PoV of instruction scheduling, not execution (ie. ignores global/
352 * local distinction):
353 * shared image atomic SSBO everything
354 * barrier()/ - R/W R/W R/W R/W X
355 * groupMemoryBarrier()
356 * memoryBarrier() - R/W R/W
357 * (but only images declared coherent?)
358 * memoryBarrierAtomic() - R/W
359 * memoryBarrierBuffer() - R/W
360 * memoryBarrierImage() - R/W
361 * memoryBarrierShared() - R/W
363 * TODO I think for SSBO/image/shared, in cases where we can determine
364 * which variable is accessed, we don't need to care about accesses to
365 * different variables (unless declared coherent??)
368 IR3_BARRIER_EVERYTHING
= 1 << 0,
369 IR3_BARRIER_SHARED_R
= 1 << 1,
370 IR3_BARRIER_SHARED_W
= 1 << 2,
371 IR3_BARRIER_IMAGE_R
= 1 << 3,
372 IR3_BARRIER_IMAGE_W
= 1 << 4,
373 IR3_BARRIER_BUFFER_R
= 1 << 5,
374 IR3_BARRIER_BUFFER_W
= 1 << 6,
375 IR3_BARRIER_ARRAY_R
= 1 << 7,
376 IR3_BARRIER_ARRAY_W
= 1 << 8,
377 } barrier_class
, barrier_conflict
;
379 /* Entry in ir3_block's instruction list: */
380 struct list_head node
;
387 static inline struct ir3_instruction
*
388 ir3_neighbor_first(struct ir3_instruction
*instr
)
391 while (instr
->cp
.left
) {
392 instr
= instr
->cp
.left
;
393 if (++cnt
> 0xffff) {
401 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
405 debug_assert(!instr
->cp
.left
);
407 while (instr
->cp
.right
) {
409 instr
= instr
->cp
.right
;
420 struct ir3_compiler
*compiler
;
421 gl_shader_stage type
;
423 unsigned ninputs
, noutputs
;
424 struct ir3_instruction
**inputs
;
425 struct ir3_instruction
**outputs
;
427 /* Track bary.f (and ldlv) instructions.. this is needed in
428 * scheduling to ensure that all varying fetches happen before
429 * any potential kill instructions. The hw gets grumpy if all
430 * threads in a group are killed before the last bary.f gets
431 * a chance to signal end of input (ei).
433 DECLARE_ARRAY(struct ir3_instruction
*, baryfs
);
435 /* Track all indirect instructions (read and write). To avoid
436 * deadlock scenario where an address register gets scheduled,
437 * but other dependent src instructions cannot be scheduled due
438 * to dependency on a *different* address register value, the
439 * scheduler needs to ensure that all dependencies other than
440 * the instruction other than the address register are scheduled
441 * before the one that writes the address register. Having a
442 * convenient list of instructions that reference some address
443 * register simplifies this.
445 DECLARE_ARRAY(struct ir3_instruction
*, indirects
);
447 /* and same for instructions that consume predicate register: */
448 DECLARE_ARRAY(struct ir3_instruction
*, predicates
);
450 /* Track texture sample instructions which need texture state
451 * patched in (for astc-srgb workaround):
453 DECLARE_ARRAY(struct ir3_instruction
*, astc_srgb
);
455 /* List of blocks: */
456 struct list_head block_list
;
458 /* List of ir3_array's: */
459 struct list_head array_list
;
461 unsigned max_sun
; /* max Sethi–Ullman number */
464 unsigned block_count
, instr_count
;
469 struct list_head node
;
473 struct nir_register
*r
;
475 /* To avoid array write's from getting DCE'd, keep track of the
476 * most recent write. Any array access depends on the most
477 * recent write. This way, nothing depends on writes after the
478 * last read. But all the writes that happen before that have
479 * something depending on them
481 struct ir3_instruction
*last_write
;
483 /* extra stuff used in RA pass: */
484 unsigned base
; /* base vreg name */
485 unsigned reg
; /* base physical reg */
486 uint16_t start_ip
, end_ip
;
489 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
492 struct list_head node
;
495 const struct nir_block
*nblock
;
497 struct list_head instr_list
; /* list of ir3_instruction */
499 /* each block has either one or two successors.. in case of
500 * two successors, 'condition' decides which one to follow.
501 * A block preceding an if/else has two successors.
503 struct ir3_instruction
*condition
;
504 struct ir3_block
*successors
[2];
506 struct set
*predecessors
; /* set of ir3_block */
508 uint16_t start_ip
, end_ip
;
510 /* Track instructions which do not write a register but other-
511 * wise must not be discarded (such as kill, stg, etc)
513 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
515 /* used for per-pass extra block data. Mainly used right
516 * now in RA step to track livein/liveout.
525 static inline uint32_t
526 block_id(struct ir3_block
*block
)
529 return block
->serialno
;
531 return (uint32_t)(unsigned long)block
;
535 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
536 gl_shader_stage type
, unsigned nin
, unsigned nout
);
537 void ir3_destroy(struct ir3
*shader
);
538 void * ir3_assemble(struct ir3
*shader
,
539 struct ir3_info
*info
, uint32_t gpu_id
);
540 void * ir3_alloc(struct ir3
*shader
, int sz
);
542 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
544 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
545 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
546 opc_t opc
, int nreg
);
547 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
548 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
);
549 const char *ir3_instr_name(struct ir3_instruction
*instr
);
551 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
553 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
554 struct ir3_register
*reg
);
556 void ir3_instr_set_address(struct ir3_instruction
*instr
,
557 struct ir3_instruction
*addr
);
559 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
561 if (instr
->flags
& IR3_INSTR_MARK
)
562 return true; /* already visited */
563 instr
->flags
|= IR3_INSTR_MARK
;
567 void ir3_block_clear_mark(struct ir3_block
*block
);
568 void ir3_clear_mark(struct ir3
*shader
);
570 unsigned ir3_count_instructions(struct ir3
*ir
);
572 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
573 struct ir3_register
*reg
)
576 for (i
= 0; i
< instr
->regs_count
; i
++)
577 if (reg
== instr
->regs
[i
])
583 #define MAX_ARRAYS 16
591 static inline uint32_t regid(int num
, int comp
)
593 return (num
<< 2) | (comp
& 0x3);
596 static inline uint32_t reg_num(struct ir3_register
*reg
)
598 return reg
->num
>> 2;
601 static inline uint32_t reg_comp(struct ir3_register
*reg
)
603 return reg
->num
& 0x3;
606 static inline bool is_flow(struct ir3_instruction
*instr
)
608 return (opc_cat(instr
->opc
) == 0);
611 static inline bool is_kill(struct ir3_instruction
*instr
)
613 return instr
->opc
== OPC_KILL
;
616 static inline bool is_nop(struct ir3_instruction
*instr
)
618 return instr
->opc
== OPC_NOP
;
621 static inline bool is_same_type_reg(struct ir3_register
*reg1
,
622 struct ir3_register
*reg2
)
624 unsigned type_reg1
= (reg1
->flags
& (IR3_REG_HIGH
| IR3_REG_HALF
));
625 unsigned type_reg2
= (reg2
->flags
& (IR3_REG_HIGH
| IR3_REG_HALF
));
627 if (type_reg1
^ type_reg2
)
633 /* Is it a non-transformative (ie. not type changing) mov? This can
634 * also include absneg.s/absneg.f, which for the most part can be
635 * treated as a mov (single src argument).
637 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
639 struct ir3_register
*dst
;
641 switch (instr
->opc
) {
643 if (instr
->cat1
.src_type
!= instr
->cat1
.dst_type
)
648 if (instr
->flags
& IR3_INSTR_SAT
)
650 /* If the type of dest reg and src reg are different,
651 * it shouldn't be considered as same type mov */
652 if (!is_same_type_reg(instr
->regs
[0], instr
->regs
[1]))
659 dst
= instr
->regs
[0];
661 /* mov's that write to a0.x or p0.x are special: */
662 if (dst
->num
== regid(REG_P0
, 0))
664 if (dst
->num
== regid(REG_A0
, 0))
667 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
673 static inline bool is_alu(struct ir3_instruction
*instr
)
675 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
678 static inline bool is_sfu(struct ir3_instruction
*instr
)
680 return (opc_cat(instr
->opc
) == 4);
683 static inline bool is_tex(struct ir3_instruction
*instr
)
685 return (opc_cat(instr
->opc
) == 5);
688 static inline bool is_mem(struct ir3_instruction
*instr
)
690 return (opc_cat(instr
->opc
) == 6);
693 static inline bool is_barrier(struct ir3_instruction
*instr
)
695 return (opc_cat(instr
->opc
) == 7);
699 is_store(struct ir3_instruction
*instr
)
701 /* these instructions, the "destination" register is
702 * actually a source, the address to store to.
704 switch (instr
->opc
) {
719 static inline bool is_load(struct ir3_instruction
*instr
)
721 switch (instr
->opc
) {
731 /* probably some others too.. */
738 static inline bool is_input(struct ir3_instruction
*instr
)
740 /* in some cases, ldlv is used to fetch varying without
741 * interpolation.. fortunately inloc is the first src
742 * register in either case
744 switch (instr
->opc
) {
753 static inline bool is_bool(struct ir3_instruction
*instr
)
755 switch (instr
->opc
) {
765 static inline bool is_meta(struct ir3_instruction
*instr
)
767 /* TODO how should we count PHI (and maybe fan-in/out) which
768 * might actually contribute some instructions to the final
771 return (opc_cat(instr
->opc
) == -1);
774 static inline unsigned dest_regs(struct ir3_instruction
*instr
)
776 if ((instr
->regs_count
== 0) || is_store(instr
))
779 return util_last_bit(instr
->regs
[0]->wrmask
);
782 static inline bool writes_addr(struct ir3_instruction
*instr
)
784 if (instr
->regs_count
> 0) {
785 struct ir3_register
*dst
= instr
->regs
[0];
786 return reg_num(dst
) == REG_A0
;
791 static inline bool writes_pred(struct ir3_instruction
*instr
)
793 if (instr
->regs_count
> 0) {
794 struct ir3_register
*dst
= instr
->regs
[0];
795 return reg_num(dst
) == REG_P0
;
800 /* returns defining instruction for reg */
801 /* TODO better name */
802 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
804 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
810 static inline bool conflicts(struct ir3_instruction
*a
,
811 struct ir3_instruction
*b
)
813 return (a
&& b
) && (a
!= b
);
816 static inline bool reg_gpr(struct ir3_register
*r
)
818 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
820 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
825 static inline type_t
half_type(type_t type
)
828 case TYPE_F32
: return TYPE_F16
;
829 case TYPE_U32
: return TYPE_U16
;
830 case TYPE_S32
: return TYPE_S16
;
841 /* some cat2 instructions (ie. those which are not float) can embed an
844 static inline bool ir3_cat2_int(opc_t opc
)
884 static inline bool ir3_cat2_float(opc_t opc
)
907 static inline bool ir3_cat3_float(opc_t opc
)
920 /* map cat2 instruction to valid abs/neg flags: */
921 static inline unsigned ir3_cat2_absneg(opc_t opc
)
938 return IR3_REG_FABS
| IR3_REG_FNEG
;
959 return IR3_REG_SABS
| IR3_REG_SNEG
;
980 /* map cat3 instructions to valid abs/neg flags: */
981 static inline unsigned ir3_cat3_absneg(opc_t opc
)
1000 /* neg *may* work on 3rd src.. */
1010 #define MASK(n) ((1 << (n)) - 1)
1012 /* iterator for an instructions's sources (reg), also returns src #: */
1013 #define foreach_src_n(__srcreg, __n, __instr) \
1014 if ((__instr)->regs_count) \
1015 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
1016 if ((__srcreg = (__instr)->regs[__n + 1]))
1018 /* iterator for an instructions's sources (reg): */
1019 #define foreach_src(__srcreg, __instr) \
1020 foreach_src_n(__srcreg, __i, __instr)
1022 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
1024 unsigned cnt
= instr
->regs_count
+ instr
->deps_count
;
1030 static inline struct ir3_instruction
* __ssa_src_n(struct ir3_instruction
*instr
, unsigned n
)
1032 if (n
== (instr
->regs_count
+ instr
->deps_count
))
1033 return instr
->address
;
1034 if (n
>= instr
->regs_count
)
1035 return instr
->deps
[n
- instr
->regs_count
];
1036 return ssa(instr
->regs
[n
]);
1039 static inline bool __is_false_dep(struct ir3_instruction
*instr
, unsigned n
)
1041 if (n
== (instr
->regs_count
+ instr
->deps_count
))
1043 if (n
>= instr
->regs_count
)
1048 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
1050 /* iterator for an instruction's SSA sources (instr), also returns src #: */
1051 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
1052 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
1053 if ((__srcinst = __ssa_src_n(__instr, __n)))
1055 /* iterator for an instruction's SSA sources (instr): */
1056 #define foreach_ssa_src(__srcinst, __instr) \
1057 foreach_ssa_src_n(__srcinst, __i, __instr)
1061 void ir3_print(struct ir3
*ir
);
1062 void ir3_print_instr(struct ir3_instruction
*instr
);
1064 /* depth calculation: */
1065 struct ir3_shader_variant
;
1066 int ir3_delayslots(struct ir3_instruction
*assigner
,
1067 struct ir3_instruction
*consumer
, unsigned n
);
1068 void ir3_insert_by_depth(struct ir3_instruction
*instr
, struct list_head
*list
);
1069 void ir3_depth(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1071 /* copy-propagate: */
1072 void ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1074 /* group neighbors and insert mov's to resolve conflicts: */
1075 void ir3_group(struct ir3
*ir
);
1077 /* Sethi–Ullman numbering: */
1078 void ir3_sun(struct ir3
*ir
);
1081 void ir3_sched_add_deps(struct ir3
*ir
);
1082 int ir3_sched(struct ir3
*ir
);
1084 void ir3_a6xx_fixup_atomic_dests(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1086 /* register assignment: */
1087 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(struct ir3_compiler
*compiler
);
1088 int ir3_ra(struct ir3_shader_variant
*v
, struct ir3_instruction
**precolor
, unsigned nprecolor
);
1091 void ir3_legalize(struct ir3
*ir
, bool *has_ssbo
, bool *need_pixlod
, int *max_bary
);
1093 /* ************************************************************************* */
1094 /* instruction helpers */
1096 static inline struct ir3_instruction
*
1097 create_immed_typed(struct ir3_block
*block
, uint32_t val
, type_t type
)
1099 struct ir3_instruction
*mov
;
1100 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1102 mov
= ir3_instr_create(block
, OPC_MOV
);
1103 mov
->cat1
.src_type
= type
;
1104 mov
->cat1
.dst_type
= type
;
1105 ir3_reg_create(mov
, 0, flags
);
1106 ir3_reg_create(mov
, 0, IR3_REG_IMMED
)->uim_val
= val
;
1111 static inline struct ir3_instruction
*
1112 create_immed(struct ir3_block
*block
, uint32_t val
)
1114 return create_immed_typed(block
, val
, TYPE_U32
);
1117 static inline struct ir3_instruction
*
1118 create_uniform_typed(struct ir3_block
*block
, unsigned n
, type_t type
)
1120 struct ir3_instruction
*mov
;
1121 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1123 mov
= ir3_instr_create(block
, OPC_MOV
);
1124 mov
->cat1
.src_type
= type
;
1125 mov
->cat1
.dst_type
= type
;
1126 ir3_reg_create(mov
, 0, flags
);
1127 ir3_reg_create(mov
, n
, IR3_REG_CONST
| flags
);
1132 static inline struct ir3_instruction
*
1133 create_uniform(struct ir3_block
*block
, unsigned n
)
1135 return create_uniform_typed(block
, n
, TYPE_F32
);
1138 static inline struct ir3_instruction
*
1139 create_uniform_indirect(struct ir3_block
*block
, int n
,
1140 struct ir3_instruction
*address
)
1142 struct ir3_instruction
*mov
;
1144 mov
= ir3_instr_create(block
, OPC_MOV
);
1145 mov
->cat1
.src_type
= TYPE_U32
;
1146 mov
->cat1
.dst_type
= TYPE_U32
;
1147 ir3_reg_create(mov
, 0, 0);
1148 ir3_reg_create(mov
, 0, IR3_REG_CONST
| IR3_REG_RELATIV
)->array
.offset
= n
;
1150 ir3_instr_set_address(mov
, address
);
1155 /* creates SSA src of correct type (ie. half vs full precision) */
1156 static inline struct ir3_register
* __ssa_src(struct ir3_instruction
*instr
,
1157 struct ir3_instruction
*src
, unsigned flags
)
1159 struct ir3_register
*reg
;
1160 if (src
->regs
[0]->flags
& IR3_REG_HALF
)
1161 flags
|= IR3_REG_HALF
;
1162 reg
= ir3_reg_create(instr
, 0, IR3_REG_SSA
| flags
);
1164 reg
->wrmask
= src
->regs
[0]->wrmask
;
1168 static inline struct ir3_instruction
*
1169 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
1171 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1172 ir3_reg_create(instr
, 0, 0); /* dst */
1173 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
1174 struct ir3_register
*src_reg
= __ssa_src(instr
, src
, IR3_REG_ARRAY
);
1175 src_reg
->array
= src
->regs
[0]->array
;
1177 __ssa_src(instr
, src
, src
->regs
[0]->flags
& IR3_REG_HIGH
);
1179 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
1180 instr
->cat1
.src_type
= type
;
1181 instr
->cat1
.dst_type
= type
;
1185 static inline struct ir3_instruction
*
1186 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
1187 type_t src_type
, type_t dst_type
)
1189 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1190 unsigned dst_flags
= (type_size(dst_type
) < 32) ? IR3_REG_HALF
: 0;
1191 unsigned src_flags
= (type_size(src_type
) < 32) ? IR3_REG_HALF
: 0;
1193 debug_assert((src
->regs
[0]->flags
& IR3_REG_HALF
) == src_flags
);
1195 ir3_reg_create(instr
, 0, dst_flags
); /* dst */
1196 __ssa_src(instr
, src
, 0);
1197 instr
->cat1
.src_type
= src_type
;
1198 instr
->cat1
.dst_type
= dst_type
;
1199 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
1203 static inline struct ir3_instruction
*
1204 ir3_NOP(struct ir3_block
*block
)
1206 return ir3_instr_create(block
, OPC_NOP
);
1209 #define IR3_INSTR_0 0
1211 #define __INSTR0(flag, name, opc) \
1212 static inline struct ir3_instruction * \
1213 ir3_##name(struct ir3_block *block) \
1215 struct ir3_instruction *instr = \
1216 ir3_instr_create(block, opc); \
1217 instr->flags |= flag; \
1220 #define INSTR0F(f, name) __INSTR0(IR3_INSTR_##f, name##_##f, OPC_##name)
1221 #define INSTR0(name) __INSTR0(0, name, OPC_##name)
1223 #define __INSTR1(flag, name, opc) \
1224 static inline struct ir3_instruction * \
1225 ir3_##name(struct ir3_block *block, \
1226 struct ir3_instruction *a, unsigned aflags) \
1228 struct ir3_instruction *instr = \
1229 ir3_instr_create(block, opc); \
1230 ir3_reg_create(instr, 0, 0); /* dst */ \
1231 __ssa_src(instr, a, aflags); \
1232 instr->flags |= flag; \
1235 #define INSTR1F(f, name) __INSTR1(IR3_INSTR_##f, name##_##f, OPC_##name)
1236 #define INSTR1(name) __INSTR1(0, name, OPC_##name)
1238 #define __INSTR2(flag, name, opc) \
1239 static inline struct ir3_instruction * \
1240 ir3_##name(struct ir3_block *block, \
1241 struct ir3_instruction *a, unsigned aflags, \
1242 struct ir3_instruction *b, unsigned bflags) \
1244 struct ir3_instruction *instr = \
1245 ir3_instr_create(block, opc); \
1246 ir3_reg_create(instr, 0, 0); /* dst */ \
1247 __ssa_src(instr, a, aflags); \
1248 __ssa_src(instr, b, bflags); \
1249 instr->flags |= flag; \
1252 #define INSTR2F(f, name) __INSTR2(IR3_INSTR_##f, name##_##f, OPC_##name)
1253 #define INSTR2(name) __INSTR2(0, name, OPC_##name)
1255 #define __INSTR3(flag, name, opc) \
1256 static inline struct ir3_instruction * \
1257 ir3_##name(struct ir3_block *block, \
1258 struct ir3_instruction *a, unsigned aflags, \
1259 struct ir3_instruction *b, unsigned bflags, \
1260 struct ir3_instruction *c, unsigned cflags) \
1262 struct ir3_instruction *instr = \
1263 ir3_instr_create2(block, opc, 4); \
1264 ir3_reg_create(instr, 0, 0); /* dst */ \
1265 __ssa_src(instr, a, aflags); \
1266 __ssa_src(instr, b, bflags); \
1267 __ssa_src(instr, c, cflags); \
1268 instr->flags |= flag; \
1271 #define INSTR3F(f, name) __INSTR3(IR3_INSTR_##f, name##_##f, OPC_##name)
1272 #define INSTR3(name) __INSTR3(0, name, OPC_##name)
1274 #define __INSTR4(flag, name, opc) \
1275 static inline struct ir3_instruction * \
1276 ir3_##name(struct ir3_block *block, \
1277 struct ir3_instruction *a, unsigned aflags, \
1278 struct ir3_instruction *b, unsigned bflags, \
1279 struct ir3_instruction *c, unsigned cflags, \
1280 struct ir3_instruction *d, unsigned dflags) \
1282 struct ir3_instruction *instr = \
1283 ir3_instr_create2(block, opc, 5); \
1284 ir3_reg_create(instr, 0, 0); /* dst */ \
1285 __ssa_src(instr, a, aflags); \
1286 __ssa_src(instr, b, bflags); \
1287 __ssa_src(instr, c, cflags); \
1288 __ssa_src(instr, d, dflags); \
1289 instr->flags |= flag; \
1292 #define INSTR4F(f, name) __INSTR4(IR3_INSTR_##f, name##_##f, OPC_##name)
1293 #define INSTR4(name) __INSTR4(0, name, OPC_##name)
1295 /* cat0 instructions: */
1303 /* cat2 instructions, most 2 src but some 1 src: */
1351 /* cat3 instructions: */
1369 /* cat4 instructions: */
1378 /* cat5 instructions: */
1385 static inline struct ir3_instruction
*
1386 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1387 unsigned wrmask
, unsigned flags
, struct ir3_instruction
*samp_tex
,
1388 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1390 struct ir3_instruction
*sam
;
1391 struct ir3_register
*reg
;
1393 sam
= ir3_instr_create(block
, opc
);
1394 sam
->flags
|= flags
| IR3_INSTR_S2EN
;
1395 ir3_reg_create(sam
, 0, 0)->wrmask
= wrmask
;
1396 __ssa_src(sam
, samp_tex
, IR3_REG_HALF
);
1398 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1399 reg
->wrmask
= (1 << (src0
->regs_count
- 1)) - 1;
1403 reg
= ir3_reg_create(sam
, 0, IR3_REG_SSA
);
1405 reg
->wrmask
= (1 << (src1
->regs_count
- 1)) - 1;
1407 sam
->cat5
.type
= type
;
1412 /* cat6 instructions: */
1427 INSTR2(ATOMIC_CMPXCHG
)
1436 INSTR3F(G
, ATOMIC_ADD
)
1437 INSTR3F(G
, ATOMIC_SUB
)
1438 INSTR3F(G
, ATOMIC_XCHG
)
1439 INSTR3F(G
, ATOMIC_INC
)
1440 INSTR3F(G
, ATOMIC_DEC
)
1441 INSTR3F(G
, ATOMIC_CMPXCHG
)
1442 INSTR3F(G
, ATOMIC_MIN
)
1443 INSTR3F(G
, ATOMIC_MAX
)
1444 INSTR3F(G
, ATOMIC_AND
)
1445 INSTR3F(G
, ATOMIC_OR
)
1446 INSTR3F(G
, ATOMIC_XOR
)
1451 INSTR4F(G
, ATOMIC_ADD
)
1452 INSTR4F(G
, ATOMIC_SUB
)
1453 INSTR4F(G
, ATOMIC_XCHG
)
1454 INSTR4F(G
, ATOMIC_INC
)
1455 INSTR4F(G
, ATOMIC_DEC
)
1456 INSTR4F(G
, ATOMIC_CMPXCHG
)
1457 INSTR4F(G
, ATOMIC_MIN
)
1458 INSTR4F(G
, ATOMIC_MAX
)
1459 INSTR4F(G
, ATOMIC_AND
)
1460 INSTR4F(G
, ATOMIC_OR
)
1461 INSTR4F(G
, ATOMIC_XOR
)
1464 /* cat7 instructions: */
1468 /* ************************************************************************* */
1469 /* split this out or find some helper to use.. like main/bitset.h.. */
1475 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
1477 static inline unsigned regmask_idx(struct ir3_register
*reg
)
1479 unsigned num
= (reg
->flags
& IR3_REG_RELATIV
) ? reg
->array
.offset
: reg
->num
;
1480 debug_assert(num
< MAX_REG
);
1481 if (reg
->flags
& IR3_REG_HALF
) {
1491 static inline void regmask_init(regmask_t
*regmask
)
1493 memset(regmask
, 0, sizeof(*regmask
));
1496 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1498 unsigned idx
= regmask_idx(reg
);
1499 if (reg
->flags
& IR3_REG_RELATIV
) {
1501 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1502 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1505 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1507 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
1511 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1514 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1515 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1518 /* set bits in a if not set in b, conceptually:
1521 static inline void regmask_set_if_not(regmask_t
*a
,
1522 struct ir3_register
*reg
, regmask_t
*b
)
1524 unsigned idx
= regmask_idx(reg
);
1525 if (reg
->flags
& IR3_REG_RELATIV
) {
1527 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1528 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1529 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1532 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1534 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
1535 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
1539 static inline bool regmask_get(regmask_t
*regmask
,
1540 struct ir3_register
*reg
)
1542 unsigned idx
= regmask_idx(reg
);
1543 if (reg
->flags
& IR3_REG_RELATIV
) {
1545 for (i
= 0; i
< reg
->size
; i
++, idx
++)
1546 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1550 for (mask
= reg
->wrmask
; mask
; mask
>>= 1, idx
++)
1552 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
1558 /* ************************************************************************* */