90086294de8f013771943634f4ee3232e9ff35d7
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 uint16_t nops_count
; /* # of nop instructions, including nopN */
53 /* NOTE: max_reg, etc, does not include registers not touched
54 * by the shader (ie. vertex fetched via VFD_DECODE but not
57 int8_t max_reg
; /* highest GPR # used by shader */
61 /* number of sync bits: */
64 /* estimate of number of cycles stalled on (ss) */
67 uint16_t last_baryf
; /* instruction # of last varying fetch */
72 IR3_REG_CONST
= 0x001,
73 IR3_REG_IMMED
= 0x002,
75 /* high registers are used for some things in compute shaders,
76 * for example. Seems to be for things that are global to all
77 * threads in a wave, so possibly these are global/shared by
78 * all the threads in the wave?
81 IR3_REG_RELATIV
= 0x010,
83 /* Most instructions, it seems, can do float abs/neg but not
84 * integer. The CP pass needs to know what is intended (int or
85 * float) in order to do the right thing. For this reason the
86 * abs/neg flags are split out into float and int variants. In
87 * addition, .b (bitwise) operations, the negate is actually a
88 * bitwise not, so split that out into a new flag to make it
97 IR3_REG_POS_INF
= 0x1000,
98 /* (ei) flag, end-input? Set on last bary, presumably to signal
99 * that the shader needs no more input:
102 /* meta-flags, for intermediate stages of IR, ie.
103 * before register assignment is done:
105 IR3_REG_SSA
= 0x4000, /* 'instr' is ptr to assigning instr */
106 IR3_REG_ARRAY
= 0x8000,
110 /* used for cat5 instructions, but also for internal/IR level
111 * tracking of what registers are read/written by an instruction.
112 * wrmask may be a bad name since it is used to represent both
113 * src and dst that touch multiple adjacent registers.
115 unsigned wrmask
: 16; /* up to vec16 */
117 /* for relative addressing, 32bits for array size is too small,
118 * but otoh we don't need to deal with disjoint sets, so instead
119 * use a simple size field (number of scalar components).
121 * Note the size field isn't important for relative const (since
122 * we don't have to do register allocation for constants).
126 bool merged
: 1; /* half-regs conflict with full regs (ie >= a6xx) */
129 * the component is in the low two bits of the reg #, so
130 * rN.x becomes: (N << 2) | x
145 /* For IR3_REG_SSA, src registers contain ptr back to assigning
148 * For IR3_REG_ARRAY, the pointer is back to the last dependent
149 * array access (although the net effect is the same, it points
150 * back to a previous instruction that we depend on).
152 struct ir3_instruction
*instr
;
156 * Stupid/simple growable array implementation:
158 #define DECLARE_ARRAY(type, name) \
159 unsigned name ## _count, name ## _sz; \
162 #define array_insert(ctx, arr, val) do { \
163 if (arr ## _count == arr ## _sz) { \
164 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
165 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
167 arr[arr ##_count++] = val; \
170 struct ir3_instruction
{
171 struct ir3_block
*block
;
174 /* (sy) flag is set on first instruction, and after sample
175 * instructions (probably just on RAW hazard).
177 IR3_INSTR_SY
= 0x001,
178 /* (ss) flag is set on first instruction, and first instruction
179 * to depend on the result of "long" instructions (RAW hazard):
181 * rcp, rsq, log2, exp2, sin, cos, sqrt
183 * It seems to synchronize until all in-flight instructions are
184 * completed, for example:
187 * add.f hr2.z, (neg)hr2.z, hc0.y
188 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
191 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
193 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
194 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
195 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
197 * The last mul.f does not have (ss) set, presumably because the
198 * (ss) on the previous instruction does the job.
200 * The blob driver also seems to set it on WAR hazards, although
201 * not really clear if this is needed or just blob compiler being
202 * sloppy. So far I haven't found a case where removing the (ss)
203 * causes problems for WAR hazard, but I could just be getting
207 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
210 IR3_INSTR_SS
= 0x002,
211 /* (jp) flag is set on jump targets:
213 IR3_INSTR_JP
= 0x004,
214 IR3_INSTR_UL
= 0x008,
215 IR3_INSTR_3D
= 0x010,
220 IR3_INSTR_S2EN
= 0x200,
222 IR3_INSTR_SAT
= 0x800,
223 /* (cat5/cat6) Bindless */
224 IR3_INSTR_B
= 0x1000,
225 /* (cat5-only) Get some parts of the encoding from a1.x */
226 IR3_INSTR_A1EN
= 0x2000,
227 /* meta-flags, for intermediate stages of IR, ie.
228 * before register assignment is done:
230 IR3_INSTR_MARK
= 0x4000,
231 IR3_INSTR_UNUSED
= 0x8000,
239 struct ir3_register
**regs
;
245 struct ir3_block
*target
;
248 type_t src_type
, dst_type
;
262 unsigned tex_base
: 3;
269 int iim_val
: 3; /* for ldgb/stgb, # of components */
270 unsigned d
: 3; /* for ldc, component offset */
275 unsigned w
: 1; /* write */
276 unsigned r
: 1; /* read */
277 unsigned l
: 1; /* local */
278 unsigned g
: 1; /* global */
280 /* for meta-instructions, just used to hold extra data
281 * before instruction scheduling, etc
284 int off
; /* component/offset */
287 /* for output collects, this maps back to the entry in the
288 * ir3_shader_variant::outputs table.
294 unsigned input_offset
;
295 unsigned samp_base
: 3;
296 unsigned tex_base
: 3;
299 /* maps back to entry in ir3_shader_variant::inputs table: */
301 /* for sysvals, identifies the sysval type. Mostly so we can
302 * identify the special cases where a sysval should not be DCE'd
303 * (currently, just pre-fs texture fetch)
305 gl_system_value sysval
;
309 /* When we get to the RA stage, we need instruction's position/name: */
313 /* used for per-pass extra instruction data.
315 * TODO we should remove the per-pass data like this and 'use_count'
316 * and do something similar to what RA does w/ ir3_ra_instr_data..
317 * ie. use the ir3_count_instructions pass, and then use instr->ip
318 * to index into a table of pass-private data.
323 * Valid if pass calls ir3_find_ssa_uses().. see foreach_ssa_use()
327 int use_count
; /* currently just updated/used by cp */
329 /* Used during CP and RA stages. For collect and shader inputs/
330 * outputs where we need a sequence of consecutive registers,
331 * keep track of each src instructions left (ie 'n-1') and right
332 * (ie 'n+1') neighbor. The front-end must insert enough mov's
333 * to ensure that each instruction has at most one left and at
334 * most one right neighbor. During the copy-propagation pass,
335 * we only remove mov's when we can preserve this constraint.
336 * And during the RA stage, we use the neighbor information to
337 * allocate a block of registers in one shot.
339 * TODO: maybe just add something like:
340 * struct ir3_instruction_ref {
341 * struct ir3_instruction *instr;
345 * Or can we get away without the refcnt stuff? It seems like
346 * it should be overkill.. the problem is if, potentially after
347 * already eliminating some mov's, if you have a single mov that
348 * needs to be grouped with it's neighbors in two different
349 * places (ex. shader output and a collect).
352 struct ir3_instruction
*left
, *right
;
353 uint16_t left_cnt
, right_cnt
;
356 /* an instruction can reference at most one address register amongst
357 * it's src/dst registers. Beyond that, you need to insert mov's.
359 * NOTE: do not write this directly, use ir3_instr_set_address()
361 struct ir3_instruction
*address
;
363 /* Tracking for additional dependent instructions. Used to handle
364 * barriers, WAR hazards for arrays/SSBOs/etc.
366 DECLARE_ARRAY(struct ir3_instruction
*, deps
);
369 * From PoV of instruction scheduling, not execution (ie. ignores global/
370 * local distinction):
371 * shared image atomic SSBO everything
372 * barrier()/ - R/W R/W R/W R/W X
373 * groupMemoryBarrier()
374 * memoryBarrier() - R/W R/W
375 * (but only images declared coherent?)
376 * memoryBarrierAtomic() - R/W
377 * memoryBarrierBuffer() - R/W
378 * memoryBarrierImage() - R/W
379 * memoryBarrierShared() - R/W
381 * TODO I think for SSBO/image/shared, in cases where we can determine
382 * which variable is accessed, we don't need to care about accesses to
383 * different variables (unless declared coherent??)
386 IR3_BARRIER_EVERYTHING
= 1 << 0,
387 IR3_BARRIER_SHARED_R
= 1 << 1,
388 IR3_BARRIER_SHARED_W
= 1 << 2,
389 IR3_BARRIER_IMAGE_R
= 1 << 3,
390 IR3_BARRIER_IMAGE_W
= 1 << 4,
391 IR3_BARRIER_BUFFER_R
= 1 << 5,
392 IR3_BARRIER_BUFFER_W
= 1 << 6,
393 IR3_BARRIER_ARRAY_R
= 1 << 7,
394 IR3_BARRIER_ARRAY_W
= 1 << 8,
395 } barrier_class
, barrier_conflict
;
397 /* Entry in ir3_block's instruction list: */
398 struct list_head node
;
404 // TODO only computerator/assembler:
408 static inline struct ir3_instruction
*
409 ir3_neighbor_first(struct ir3_instruction
*instr
)
412 while (instr
->cp
.left
) {
413 instr
= instr
->cp
.left
;
414 if (++cnt
> 0xffff) {
422 static inline int ir3_neighbor_count(struct ir3_instruction
*instr
)
426 debug_assert(!instr
->cp
.left
);
428 while (instr
->cp
.right
) {
430 instr
= instr
->cp
.right
;
441 struct ir3_compiler
*compiler
;
442 gl_shader_stage type
;
444 DECLARE_ARRAY(struct ir3_instruction
*, inputs
);
445 DECLARE_ARRAY(struct ir3_instruction
*, outputs
);
447 /* Track bary.f (and ldlv) instructions.. this is needed in
448 * scheduling to ensure that all varying fetches happen before
449 * any potential kill instructions. The hw gets grumpy if all
450 * threads in a group are killed before the last bary.f gets
451 * a chance to signal end of input (ei).
453 DECLARE_ARRAY(struct ir3_instruction
*, baryfs
);
455 /* Track all indirect instructions (read and write). To avoid
456 * deadlock scenario where an address register gets scheduled,
457 * but other dependent src instructions cannot be scheduled due
458 * to dependency on a *different* address register value, the
459 * scheduler needs to ensure that all dependencies other than
460 * the instruction other than the address register are scheduled
461 * before the one that writes the address register. Having a
462 * convenient list of instructions that reference some address
463 * register simplifies this.
465 DECLARE_ARRAY(struct ir3_instruction
*, a0_users
);
468 DECLARE_ARRAY(struct ir3_instruction
*, a1_users
);
470 /* and same for instructions that consume predicate register: */
471 DECLARE_ARRAY(struct ir3_instruction
*, predicates
);
473 /* Track texture sample instructions which need texture state
474 * patched in (for astc-srgb workaround):
476 DECLARE_ARRAY(struct ir3_instruction
*, astc_srgb
);
478 /* List of blocks: */
479 struct list_head block_list
;
481 /* List of ir3_array's: */
482 struct list_head array_list
;
485 unsigned block_count
, instr_count
;
490 struct list_head node
;
494 struct nir_register
*r
;
496 /* To avoid array write's from getting DCE'd, keep track of the
497 * most recent write. Any array access depends on the most
498 * recent write. This way, nothing depends on writes after the
499 * last read. But all the writes that happen before that have
500 * something depending on them
502 struct ir3_instruction
*last_write
;
504 /* extra stuff used in RA pass: */
505 unsigned base
; /* base vreg name */
506 unsigned reg
; /* base physical reg */
507 uint16_t start_ip
, end_ip
;
509 /* Indicates if half-precision */
513 struct ir3_array
* ir3_lookup_array(struct ir3
*ir
, unsigned id
);
516 struct list_head node
;
519 const struct nir_block
*nblock
;
521 struct list_head instr_list
; /* list of ir3_instruction */
523 /* each block has either one or two successors.. in case of
524 * two successors, 'condition' decides which one to follow.
525 * A block preceding an if/else has two successors.
527 struct ir3_instruction
*condition
;
528 struct ir3_block
*successors
[2];
530 struct set
*predecessors
; /* set of ir3_block */
532 uint16_t start_ip
, end_ip
;
534 /* Track instructions which do not write a register but other-
535 * wise must not be discarded (such as kill, stg, etc)
537 DECLARE_ARRAY(struct ir3_instruction
*, keeps
);
539 /* used for per-pass extra block data. Mainly used right
540 * now in RA step to track livein/liveout.
549 static inline uint32_t
550 block_id(struct ir3_block
*block
)
553 return block
->serialno
;
555 return (uint32_t)(unsigned long)block
;
559 struct ir3
* ir3_create(struct ir3_compiler
*compiler
, gl_shader_stage type
);
560 void ir3_destroy(struct ir3
*shader
);
561 void * ir3_assemble(struct ir3
*shader
,
562 struct ir3_info
*info
, uint32_t gpu_id
);
563 void * ir3_alloc(struct ir3
*shader
, int sz
);
565 struct ir3_block
* ir3_block_create(struct ir3
*shader
);
567 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
);
568 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
569 opc_t opc
, int nreg
);
570 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
571 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
);
572 const char *ir3_instr_name(struct ir3_instruction
*instr
);
574 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
576 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
577 struct ir3_register
*reg
);
579 void ir3_instr_set_address(struct ir3_instruction
*instr
,
580 struct ir3_instruction
*addr
);
582 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
584 if (instr
->flags
& IR3_INSTR_MARK
)
585 return true; /* already visited */
586 instr
->flags
|= IR3_INSTR_MARK
;
590 void ir3_block_clear_mark(struct ir3_block
*block
);
591 void ir3_clear_mark(struct ir3
*shader
);
593 unsigned ir3_count_instructions(struct ir3
*ir
);
594 unsigned ir3_count_instructions_ra(struct ir3
*ir
);
597 * Move 'instr' to just before 'after'
600 ir3_instr_move_before(struct ir3_instruction
*instr
,
601 struct ir3_instruction
*after
)
603 list_delinit(&instr
->node
);
604 list_addtail(&instr
->node
, &after
->node
);
608 * Move 'instr' to just after 'before':
611 ir3_instr_move_after(struct ir3_instruction
*instr
,
612 struct ir3_instruction
*before
)
614 list_delinit(&instr
->node
);
615 list_add(&instr
->node
, &before
->node
);
618 void ir3_find_ssa_uses(struct ir3
*ir
, void *mem_ctx
, bool falsedeps
);
620 void ir3_set_dst_type(struct ir3_instruction
*instr
, bool half
);
621 void ir3_fixup_src_type(struct ir3_instruction
*instr
);
623 #include "util/set.h"
624 #define foreach_ssa_use(__use, __instr) \
625 for (struct ir3_instruction *__use = (void *)~0; \
626 __use && (__instr)->uses; __use = NULL) \
627 set_foreach ((__instr)->uses, __entry) \
628 if ((__use = (void *)__entry->key))
630 #define MAX_ARRAYS 16
638 static inline uint32_t regid(int num
, int comp
)
640 return (num
<< 2) | (comp
& 0x3);
643 static inline uint32_t reg_num(struct ir3_register
*reg
)
645 return reg
->num
>> 2;
648 static inline uint32_t reg_comp(struct ir3_register
*reg
)
650 return reg
->num
& 0x3;
653 #define INVALID_REG regid(63, 0)
654 #define VALIDREG(r) ((r) != INVALID_REG)
655 #define CONDREG(r, val) COND(VALIDREG(r), (val))
657 static inline bool is_flow(struct ir3_instruction
*instr
)
659 return (opc_cat(instr
->opc
) == 0);
662 static inline bool is_kill(struct ir3_instruction
*instr
)
664 return instr
->opc
== OPC_KILL
;
667 static inline bool is_nop(struct ir3_instruction
*instr
)
669 return instr
->opc
== OPC_NOP
;
672 static inline bool is_same_type_reg(struct ir3_register
*reg1
,
673 struct ir3_register
*reg2
)
675 unsigned type_reg1
= (reg1
->flags
& (IR3_REG_HIGH
| IR3_REG_HALF
));
676 unsigned type_reg2
= (reg2
->flags
& (IR3_REG_HIGH
| IR3_REG_HALF
));
678 if (type_reg1
^ type_reg2
)
684 /* Is it a non-transformative (ie. not type changing) mov? This can
685 * also include absneg.s/absneg.f, which for the most part can be
686 * treated as a mov (single src argument).
688 static inline bool is_same_type_mov(struct ir3_instruction
*instr
)
690 struct ir3_register
*dst
;
692 switch (instr
->opc
) {
694 if (instr
->cat1
.src_type
!= instr
->cat1
.dst_type
)
696 /* If the type of dest reg and src reg are different,
697 * it shouldn't be considered as same type mov
699 if (!is_same_type_reg(instr
->regs
[0], instr
->regs
[1]))
704 if (instr
->flags
& IR3_INSTR_SAT
)
706 /* If the type of dest reg and src reg are different,
707 * it shouldn't be considered as same type mov
709 if (!is_same_type_reg(instr
->regs
[0], instr
->regs
[1]))
716 dst
= instr
->regs
[0];
718 /* mov's that write to a0 or p0.x are special: */
719 if (dst
->num
== regid(REG_P0
, 0))
721 if (reg_num(dst
) == REG_A0
)
724 if (dst
->flags
& (IR3_REG_RELATIV
| IR3_REG_ARRAY
))
730 /* A move from const, which changes size but not type, can also be
731 * folded into dest instruction in some cases.
733 static inline bool is_const_mov(struct ir3_instruction
*instr
)
735 if (instr
->opc
!= OPC_MOV
)
738 if (!(instr
->regs
[1]->flags
& IR3_REG_CONST
))
741 type_t src_type
= instr
->cat1
.src_type
;
742 type_t dst_type
= instr
->cat1
.dst_type
;
744 return (type_float(src_type
) && type_float(dst_type
)) ||
745 (type_uint(src_type
) && type_uint(dst_type
)) ||
746 (type_sint(src_type
) && type_sint(dst_type
));
749 static inline bool is_alu(struct ir3_instruction
*instr
)
751 return (1 <= opc_cat(instr
->opc
)) && (opc_cat(instr
->opc
) <= 3);
754 static inline bool is_sfu(struct ir3_instruction
*instr
)
756 return (opc_cat(instr
->opc
) == 4);
759 static inline bool is_tex(struct ir3_instruction
*instr
)
761 return (opc_cat(instr
->opc
) == 5);
764 static inline bool is_tex_or_prefetch(struct ir3_instruction
*instr
)
766 return is_tex(instr
) || (instr
->opc
== OPC_META_TEX_PREFETCH
);
769 static inline bool is_mem(struct ir3_instruction
*instr
)
771 return (opc_cat(instr
->opc
) == 6);
774 static inline bool is_barrier(struct ir3_instruction
*instr
)
776 return (opc_cat(instr
->opc
) == 7);
780 is_half(struct ir3_instruction
*instr
)
782 return !!(instr
->regs
[0]->flags
& IR3_REG_HALF
);
786 is_high(struct ir3_instruction
*instr
)
788 return !!(instr
->regs
[0]->flags
& IR3_REG_HIGH
);
792 is_store(struct ir3_instruction
*instr
)
794 /* these instructions, the "destination" register is
795 * actually a source, the address to store to.
797 switch (instr
->opc
) {
812 static inline bool is_load(struct ir3_instruction
*instr
)
814 switch (instr
->opc
) {
824 /* probably some others too.. */
831 static inline bool is_input(struct ir3_instruction
*instr
)
833 /* in some cases, ldlv is used to fetch varying without
834 * interpolation.. fortunately inloc is the first src
835 * register in either case
837 switch (instr
->opc
) {
846 static inline bool is_bool(struct ir3_instruction
*instr
)
848 switch (instr
->opc
) {
859 cat3_half_opc(opc_t opc
)
862 case OPC_MAD_F32
: return OPC_MAD_F16
;
863 case OPC_SEL_B32
: return OPC_SEL_B16
;
864 case OPC_SEL_S32
: return OPC_SEL_S16
;
865 case OPC_SEL_F32
: return OPC_SEL_F16
;
866 case OPC_SAD_S32
: return OPC_SAD_S16
;
872 cat3_full_opc(opc_t opc
)
875 case OPC_MAD_F16
: return OPC_MAD_F32
;
876 case OPC_SEL_B16
: return OPC_SEL_B32
;
877 case OPC_SEL_S16
: return OPC_SEL_S32
;
878 case OPC_SEL_F16
: return OPC_SEL_F32
;
879 case OPC_SAD_S16
: return OPC_SAD_S32
;
885 cat4_half_opc(opc_t opc
)
888 case OPC_RSQ
: return OPC_HRSQ
;
889 case OPC_LOG2
: return OPC_HLOG2
;
890 case OPC_EXP2
: return OPC_HEXP2
;
896 cat4_full_opc(opc_t opc
)
899 case OPC_HRSQ
: return OPC_RSQ
;
900 case OPC_HLOG2
: return OPC_LOG2
;
901 case OPC_HEXP2
: return OPC_EXP2
;
906 static inline bool is_meta(struct ir3_instruction
*instr
)
908 return (opc_cat(instr
->opc
) == -1);
911 static inline unsigned dest_regs(struct ir3_instruction
*instr
)
913 if ((instr
->regs_count
== 0) || is_store(instr
) || is_flow(instr
))
916 return util_last_bit(instr
->regs
[0]->wrmask
);
920 writes_gpr(struct ir3_instruction
*instr
)
922 if (dest_regs(instr
) == 0)
924 /* is dest a normal temp register: */
925 struct ir3_register
*reg
= instr
->regs
[0];
926 debug_assert(!(reg
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
)));
927 if ((reg_num(reg
) == REG_A0
) ||
928 (reg
->num
== regid(REG_P0
, 0)))
933 static inline bool writes_addr0(struct ir3_instruction
*instr
)
935 if (instr
->regs_count
> 0) {
936 struct ir3_register
*dst
= instr
->regs
[0];
937 return dst
->num
== regid(REG_A0
, 0);
942 static inline bool writes_addr1(struct ir3_instruction
*instr
)
944 if (instr
->regs_count
> 0) {
945 struct ir3_register
*dst
= instr
->regs
[0];
946 return dst
->num
== regid(REG_A0
, 1);
951 static inline bool writes_pred(struct ir3_instruction
*instr
)
953 if (instr
->regs_count
> 0) {
954 struct ir3_register
*dst
= instr
->regs
[0];
955 return reg_num(dst
) == REG_P0
;
960 /* returns defining instruction for reg */
961 /* TODO better name */
962 static inline struct ir3_instruction
*ssa(struct ir3_register
*reg
)
964 if (reg
->flags
& (IR3_REG_SSA
| IR3_REG_ARRAY
)) {
970 static inline bool conflicts(struct ir3_instruction
*a
,
971 struct ir3_instruction
*b
)
973 return (a
&& b
) && (a
!= b
);
976 static inline bool reg_gpr(struct ir3_register
*r
)
978 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
))
980 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
985 static inline type_t
half_type(type_t type
)
988 case TYPE_F32
: return TYPE_F16
;
989 case TYPE_U32
: return TYPE_U16
;
990 case TYPE_S32
: return TYPE_S16
;
1001 static inline type_t
full_type(type_t type
)
1004 case TYPE_F16
: return TYPE_F32
;
1005 case TYPE_U16
: return TYPE_U32
;
1006 case TYPE_S16
: return TYPE_S32
;
1017 /* some cat2 instructions (ie. those which are not float) can embed an
1020 static inline bool ir3_cat2_int(opc_t opc
)
1060 /* map cat2 instruction to valid abs/neg flags: */
1061 static inline unsigned ir3_cat2_absneg(opc_t opc
)
1078 return IR3_REG_FABS
| IR3_REG_FNEG
;
1099 return IR3_REG_SABS
| IR3_REG_SNEG
;
1113 return IR3_REG_BNOT
;
1120 /* map cat3 instructions to valid abs/neg flags: */
1121 static inline unsigned ir3_cat3_absneg(opc_t opc
)
1128 return IR3_REG_FNEG
;
1140 /* neg *may* work on 3rd src.. */
1150 #define MASK(n) ((1 << (n)) - 1)
1152 /* iterator for an instructions's sources (reg), also returns src #: */
1153 #define foreach_src_n(__srcreg, __n, __instr) \
1154 if ((__instr)->regs_count) \
1155 for (struct ir3_register *__srcreg = (void *)~0; __srcreg; __srcreg = NULL) \
1156 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
1157 if ((__srcreg = (__instr)->regs[__n + 1]))
1159 /* iterator for an instructions's sources (reg): */
1160 #define foreach_src(__srcreg, __instr) \
1161 foreach_src_n(__srcreg, __i, __instr)
1163 static inline unsigned __ssa_src_cnt(struct ir3_instruction
*instr
)
1165 unsigned cnt
= instr
->regs_count
+ instr
->deps_count
;
1171 static inline struct ir3_instruction
**
1172 __ssa_srcp_n(struct ir3_instruction
*instr
, unsigned n
)
1174 if (n
== (instr
->regs_count
+ instr
->deps_count
))
1175 return &instr
->address
;
1176 if (n
>= instr
->regs_count
)
1177 return &instr
->deps
[n
- instr
->regs_count
];
1178 if (ssa(instr
->regs
[n
]))
1179 return &instr
->regs
[n
]->instr
;
1183 static inline bool __is_false_dep(struct ir3_instruction
*instr
, unsigned n
)
1185 if (n
== (instr
->regs_count
+ instr
->deps_count
))
1187 if (n
>= instr
->regs_count
)
1192 #define foreach_ssa_srcp_n(__srcp, __n, __instr) \
1193 for (struct ir3_instruction **__srcp = (void *)~0; __srcp; __srcp = NULL) \
1194 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
1195 if ((__srcp = __ssa_srcp_n(__instr, __n)))
1197 #define foreach_ssa_srcp(__srcp, __instr) \
1198 foreach_ssa_srcp_n(__srcp, __i, __instr)
1200 /* iterator for an instruction's SSA sources (instr), also returns src #: */
1201 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
1202 for (struct ir3_instruction *__srcinst = (void *)~0; __srcinst; __srcinst = NULL) \
1203 foreach_ssa_srcp_n(__srcp, __n, __instr) \
1204 if ((__srcinst = *__srcp))
1206 /* iterator for an instruction's SSA sources (instr): */
1207 #define foreach_ssa_src(__srcinst, __instr) \
1208 foreach_ssa_src_n(__srcinst, __i, __instr)
1210 /* iterators for shader inputs: */
1211 #define foreach_input_n(__ininstr, __cnt, __ir) \
1212 for (struct ir3_instruction *__ininstr = (void *)~0; __ininstr; __ininstr = NULL) \
1213 for (unsigned __cnt = 0; __cnt < (__ir)->inputs_count; __cnt++) \
1214 if ((__ininstr = (__ir)->inputs[__cnt]))
1215 #define foreach_input(__ininstr, __ir) \
1216 foreach_input_n(__ininstr, __i, __ir)
1218 /* iterators for shader outputs: */
1219 #define foreach_output_n(__outinstr, __cnt, __ir) \
1220 for (struct ir3_instruction *__outinstr = (void *)~0; __outinstr; __outinstr = NULL) \
1221 for (unsigned __cnt = 0; __cnt < (__ir)->outputs_count; __cnt++) \
1222 if ((__outinstr = (__ir)->outputs[__cnt]))
1223 #define foreach_output(__outinstr, __ir) \
1224 foreach_output_n(__outinstr, __i, __ir)
1226 /* iterators for instructions: */
1227 #define foreach_instr(__instr, __list) \
1228 list_for_each_entry(struct ir3_instruction, __instr, __list, node)
1229 #define foreach_instr_rev(__instr, __list) \
1230 list_for_each_entry_rev(struct ir3_instruction, __instr, __list, node)
1231 #define foreach_instr_safe(__instr, __list) \
1232 list_for_each_entry_safe(struct ir3_instruction, __instr, __list, node)
1234 /* iterators for blocks: */
1235 #define foreach_block(__block, __list) \
1236 list_for_each_entry(struct ir3_block, __block, __list, node)
1237 #define foreach_block_safe(__block, __list) \
1238 list_for_each_entry_safe(struct ir3_block, __block, __list, node)
1239 #define foreach_block_rev(__block, __list) \
1240 list_for_each_entry_rev(struct ir3_block, __block, __list, node)
1242 /* iterators for arrays: */
1243 #define foreach_array(__array, __list) \
1244 list_for_each_entry(struct ir3_array, __array, __list, node)
1246 /* Check if condition is true for any src instruction.
1249 check_src_cond(struct ir3_instruction
*instr
, bool (*cond
)(struct ir3_instruction
*))
1251 /* Note that this is also used post-RA so skip the ssa iterator: */
1252 foreach_src (reg
, instr
) {
1253 struct ir3_instruction
*src
= reg
->instr
;
1258 /* meta:split/collect aren't real instructions, the thing that
1259 * we actually care about is *their* srcs
1261 if ((src
->opc
== OPC_META_SPLIT
) || (src
->opc
== OPC_META_COLLECT
)) {
1262 if (check_src_cond(src
, cond
))
1273 #define IR3_PASS(ir, pass, ...) ({ \
1274 bool progress = pass(ir, ##__VA_ARGS__); \
1276 ir3_debug_print(ir, "AFTER: " #pass); \
1283 void ir3_validate(struct ir3
*ir
);
1286 void ir3_print(struct ir3
*ir
);
1287 void ir3_print_instr(struct ir3_instruction
*instr
);
1289 /* delay calculation: */
1290 int ir3_delayslots(struct ir3_instruction
*assigner
,
1291 struct ir3_instruction
*consumer
, unsigned n
, bool soft
);
1292 unsigned ir3_delay_calc(struct ir3_block
*block
, struct ir3_instruction
*instr
,
1293 bool soft
, bool pred
);
1294 void ir3_remove_nops(struct ir3
*ir
);
1296 /* dead code elimination: */
1297 struct ir3_shader_variant
;
1298 bool ir3_dce(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1300 /* fp16 conversion folding */
1301 bool ir3_cf(struct ir3
*ir
);
1303 /* copy-propagate: */
1304 bool ir3_cp(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1306 /* group neighbors and insert mov's to resolve conflicts: */
1307 bool ir3_group(struct ir3
*ir
);
1310 bool ir3_sched_add_deps(struct ir3
*ir
);
1311 int ir3_sched(struct ir3
*ir
);
1314 bool ir3_postsched(struct ir3
*ir
);
1316 bool ir3_a6xx_fixup_atomic_dests(struct ir3
*ir
, struct ir3_shader_variant
*so
);
1318 /* register assignment: */
1319 struct ir3_ra_reg_set
* ir3_ra_alloc_reg_set(struct ir3_compiler
*compiler
);
1320 int ir3_ra(struct ir3_shader_variant
*v
, struct ir3_instruction
**precolor
, unsigned nprecolor
);
1323 bool ir3_legalize(struct ir3
*ir
, struct ir3_shader_variant
*so
, int *max_bary
);
1326 ir3_has_latency_to_hide(struct ir3
*ir
)
1328 /* VS/GS/TCS/TESS co-exist with frag shader invocations, but we don't
1329 * know the nature of the fragment shader. Just assume it will have
1332 if (ir
->type
!= MESA_SHADER_FRAGMENT
)
1335 foreach_block (block
, &ir
->block_list
) {
1336 foreach_instr (instr
, &block
->instr_list
) {
1337 if (is_tex_or_prefetch(instr
))
1340 if (is_load(instr
)) {
1341 switch (instr
->opc
) {
1356 /* ************************************************************************* */
1357 /* instruction helpers */
1359 /* creates SSA src of correct type (ie. half vs full precision) */
1360 static inline struct ir3_register
* __ssa_src(struct ir3_instruction
*instr
,
1361 struct ir3_instruction
*src
, unsigned flags
)
1363 struct ir3_register
*reg
;
1364 if (src
->regs
[0]->flags
& IR3_REG_HALF
)
1365 flags
|= IR3_REG_HALF
;
1366 reg
= ir3_reg_create(instr
, 0, IR3_REG_SSA
| flags
);
1368 reg
->wrmask
= src
->regs
[0]->wrmask
;
1372 static inline struct ir3_register
* __ssa_dst(struct ir3_instruction
*instr
)
1374 struct ir3_register
*reg
= ir3_reg_create(instr
, 0, 0);
1375 reg
->flags
|= IR3_REG_SSA
;
1379 static inline struct ir3_instruction
*
1380 create_immed_typed(struct ir3_block
*block
, uint32_t val
, type_t type
)
1382 struct ir3_instruction
*mov
;
1383 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1385 mov
= ir3_instr_create(block
, OPC_MOV
);
1386 mov
->cat1
.src_type
= type
;
1387 mov
->cat1
.dst_type
= type
;
1388 __ssa_dst(mov
)->flags
|= flags
;
1389 ir3_reg_create(mov
, 0, IR3_REG_IMMED
| flags
)->uim_val
= val
;
1394 static inline struct ir3_instruction
*
1395 create_immed(struct ir3_block
*block
, uint32_t val
)
1397 return create_immed_typed(block
, val
, TYPE_U32
);
1400 static inline struct ir3_instruction
*
1401 create_uniform_typed(struct ir3_block
*block
, unsigned n
, type_t type
)
1403 struct ir3_instruction
*mov
;
1404 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1406 mov
= ir3_instr_create(block
, OPC_MOV
);
1407 mov
->cat1
.src_type
= type
;
1408 mov
->cat1
.dst_type
= type
;
1409 __ssa_dst(mov
)->flags
|= flags
;
1410 ir3_reg_create(mov
, n
, IR3_REG_CONST
| flags
);
1415 static inline struct ir3_instruction
*
1416 create_uniform(struct ir3_block
*block
, unsigned n
)
1418 return create_uniform_typed(block
, n
, TYPE_F32
);
1421 static inline struct ir3_instruction
*
1422 create_uniform_indirect(struct ir3_block
*block
, int n
,
1423 struct ir3_instruction
*address
)
1425 struct ir3_instruction
*mov
;
1427 mov
= ir3_instr_create(block
, OPC_MOV
);
1428 mov
->cat1
.src_type
= TYPE_U32
;
1429 mov
->cat1
.dst_type
= TYPE_U32
;
1431 ir3_reg_create(mov
, 0, IR3_REG_CONST
| IR3_REG_RELATIV
)->array
.offset
= n
;
1433 ir3_instr_set_address(mov
, address
);
1438 static inline struct ir3_instruction
*
1439 ir3_MOV(struct ir3_block
*block
, struct ir3_instruction
*src
, type_t type
)
1441 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1442 unsigned flags
= (type_size(type
) < 32) ? IR3_REG_HALF
: 0;
1444 __ssa_dst(instr
)->flags
|= flags
;
1445 if (src
->regs
[0]->flags
& IR3_REG_ARRAY
) {
1446 struct ir3_register
*src_reg
= __ssa_src(instr
, src
, IR3_REG_ARRAY
);
1447 src_reg
->array
= src
->regs
[0]->array
;
1449 __ssa_src(instr
, src
, src
->regs
[0]->flags
& IR3_REG_HIGH
);
1451 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_RELATIV
));
1452 instr
->cat1
.src_type
= type
;
1453 instr
->cat1
.dst_type
= type
;
1457 static inline struct ir3_instruction
*
1458 ir3_COV(struct ir3_block
*block
, struct ir3_instruction
*src
,
1459 type_t src_type
, type_t dst_type
)
1461 struct ir3_instruction
*instr
= ir3_instr_create(block
, OPC_MOV
);
1462 unsigned dst_flags
= (type_size(dst_type
) < 32) ? IR3_REG_HALF
: 0;
1463 unsigned src_flags
= (type_size(src_type
) < 32) ? IR3_REG_HALF
: 0;
1465 debug_assert((src
->regs
[0]->flags
& IR3_REG_HALF
) == src_flags
);
1467 __ssa_dst(instr
)->flags
|= dst_flags
;
1468 __ssa_src(instr
, src
, 0);
1469 instr
->cat1
.src_type
= src_type
;
1470 instr
->cat1
.dst_type
= dst_type
;
1471 debug_assert(!(src
->regs
[0]->flags
& IR3_REG_ARRAY
));
1475 static inline struct ir3_instruction
*
1476 ir3_NOP(struct ir3_block
*block
)
1478 return ir3_instr_create(block
, OPC_NOP
);
1481 #define IR3_INSTR_0 0
1483 #define __INSTR0(flag, name, opc) \
1484 static inline struct ir3_instruction * \
1485 ir3_##name(struct ir3_block *block) \
1487 struct ir3_instruction *instr = \
1488 ir3_instr_create(block, opc); \
1489 instr->flags |= flag; \
1492 #define INSTR0F(f, name) __INSTR0(IR3_INSTR_##f, name##_##f, OPC_##name)
1493 #define INSTR0(name) __INSTR0(0, name, OPC_##name)
1495 #define __INSTR1(flag, name, opc) \
1496 static inline struct ir3_instruction * \
1497 ir3_##name(struct ir3_block *block, \
1498 struct ir3_instruction *a, unsigned aflags) \
1500 struct ir3_instruction *instr = \
1501 ir3_instr_create(block, opc); \
1503 __ssa_src(instr, a, aflags); \
1504 instr->flags |= flag; \
1507 #define INSTR1F(f, name) __INSTR1(IR3_INSTR_##f, name##_##f, OPC_##name)
1508 #define INSTR1(name) __INSTR1(0, name, OPC_##name)
1510 #define __INSTR2(flag, name, opc) \
1511 static inline struct ir3_instruction * \
1512 ir3_##name(struct ir3_block *block, \
1513 struct ir3_instruction *a, unsigned aflags, \
1514 struct ir3_instruction *b, unsigned bflags) \
1516 struct ir3_instruction *instr = \
1517 ir3_instr_create(block, opc); \
1519 __ssa_src(instr, a, aflags); \
1520 __ssa_src(instr, b, bflags); \
1521 instr->flags |= flag; \
1524 #define INSTR2F(f, name) __INSTR2(IR3_INSTR_##f, name##_##f, OPC_##name)
1525 #define INSTR2(name) __INSTR2(0, name, OPC_##name)
1527 #define __INSTR3(flag, name, opc) \
1528 static inline struct ir3_instruction * \
1529 ir3_##name(struct ir3_block *block, \
1530 struct ir3_instruction *a, unsigned aflags, \
1531 struct ir3_instruction *b, unsigned bflags, \
1532 struct ir3_instruction *c, unsigned cflags) \
1534 struct ir3_instruction *instr = \
1535 ir3_instr_create2(block, opc, 4); \
1537 __ssa_src(instr, a, aflags); \
1538 __ssa_src(instr, b, bflags); \
1539 __ssa_src(instr, c, cflags); \
1540 instr->flags |= flag; \
1543 #define INSTR3F(f, name) __INSTR3(IR3_INSTR_##f, name##_##f, OPC_##name)
1544 #define INSTR3(name) __INSTR3(0, name, OPC_##name)
1546 #define __INSTR4(flag, name, opc) \
1547 static inline struct ir3_instruction * \
1548 ir3_##name(struct ir3_block *block, \
1549 struct ir3_instruction *a, unsigned aflags, \
1550 struct ir3_instruction *b, unsigned bflags, \
1551 struct ir3_instruction *c, unsigned cflags, \
1552 struct ir3_instruction *d, unsigned dflags) \
1554 struct ir3_instruction *instr = \
1555 ir3_instr_create2(block, opc, 5); \
1557 __ssa_src(instr, a, aflags); \
1558 __ssa_src(instr, b, bflags); \
1559 __ssa_src(instr, c, cflags); \
1560 __ssa_src(instr, d, dflags); \
1561 instr->flags |= flag; \
1564 #define INSTR4F(f, name) __INSTR4(IR3_INSTR_##f, name##_##f, OPC_##name)
1565 #define INSTR4(name) __INSTR4(0, name, OPC_##name)
1567 /* cat0 instructions: */
1578 /* cat2 instructions, most 2 src but some 1 src: */
1626 /* cat3 instructions: */
1635 /* NOTE: SEL_B32 checks for zero vs nonzero */
1645 /* cat4 instructions: */
1657 /* cat5 instructions: */
1666 static inline struct ir3_instruction
*
1667 ir3_SAM(struct ir3_block
*block
, opc_t opc
, type_t type
,
1668 unsigned wrmask
, unsigned flags
, struct ir3_instruction
*samp_tex
,
1669 struct ir3_instruction
*src0
, struct ir3_instruction
*src1
)
1671 struct ir3_instruction
*sam
;
1673 sam
= ir3_instr_create(block
, opc
);
1674 sam
->flags
|= flags
;
1675 __ssa_dst(sam
)->wrmask
= wrmask
;
1676 if (flags
& IR3_INSTR_S2EN
) {
1677 __ssa_src(sam
, samp_tex
, IR3_REG_HALF
);
1680 __ssa_src(sam
, src0
, 0);
1683 __ssa_src(sam
, src1
, 0);
1685 sam
->cat5
.type
= type
;
1690 /* cat6 instructions: */
1705 INSTR2(ATOMIC_CMPXCHG
)
1715 INSTR3F(G
, ATOMIC_ADD
)
1716 INSTR3F(G
, ATOMIC_SUB
)
1717 INSTR3F(G
, ATOMIC_XCHG
)
1718 INSTR3F(G
, ATOMIC_INC
)
1719 INSTR3F(G
, ATOMIC_DEC
)
1720 INSTR3F(G
, ATOMIC_CMPXCHG
)
1721 INSTR3F(G
, ATOMIC_MIN
)
1722 INSTR3F(G
, ATOMIC_MAX
)
1723 INSTR3F(G
, ATOMIC_AND
)
1724 INSTR3F(G
, ATOMIC_OR
)
1725 INSTR3F(G
, ATOMIC_XOR
)
1730 INSTR4F(G
, ATOMIC_ADD
)
1731 INSTR4F(G
, ATOMIC_SUB
)
1732 INSTR4F(G
, ATOMIC_XCHG
)
1733 INSTR4F(G
, ATOMIC_INC
)
1734 INSTR4F(G
, ATOMIC_DEC
)
1735 INSTR4F(G
, ATOMIC_CMPXCHG
)
1736 INSTR4F(G
, ATOMIC_MIN
)
1737 INSTR4F(G
, ATOMIC_MAX
)
1738 INSTR4F(G
, ATOMIC_AND
)
1739 INSTR4F(G
, ATOMIC_OR
)
1740 INSTR4F(G
, ATOMIC_XOR
)
1745 /* cat7 instructions: */
1749 /* meta instructions: */
1750 INSTR0(META_TEX_PREFETCH
);
1752 /* ************************************************************************* */
1753 /* split this out or find some helper to use.. like main/bitset.h.. */
1756 #include "util/bitset.h"
1760 typedef BITSET_DECLARE(regmask_t
, 2 * MAX_REG
);
1763 __regmask_get(regmask_t
*regmask
, struct ir3_register
*reg
, unsigned n
)
1766 /* a6xx+ case, with merged register file, we track things in terms
1767 * of half-precision registers, with a full precisions register
1768 * using two half-precision slots:
1770 if (reg
->flags
& IR3_REG_HALF
) {
1771 return BITSET_TEST(*regmask
, n
);
1774 return BITSET_TEST(*regmask
, n
) || BITSET_TEST(*regmask
, n
+1);
1777 /* pre a6xx case, with separate register file for half and full
1780 if (reg
->flags
& IR3_REG_HALF
)
1782 return BITSET_TEST(*regmask
, n
);
1787 __regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
, unsigned n
)
1790 /* a6xx+ case, with merged register file, we track things in terms
1791 * of half-precision registers, with a full precisions register
1792 * using two half-precision slots:
1794 if (reg
->flags
& IR3_REG_HALF
) {
1795 BITSET_SET(*regmask
, n
);
1798 BITSET_SET(*regmask
, n
);
1799 BITSET_SET(*regmask
, n
+1);
1802 /* pre a6xx case, with separate register file for half and full
1805 if (reg
->flags
& IR3_REG_HALF
)
1807 BITSET_SET(*regmask
, n
);
1811 static inline void regmask_init(regmask_t
*regmask
)
1813 memset(regmask
, 0, sizeof(*regmask
));
1816 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
1818 if (reg
->flags
& IR3_REG_RELATIV
) {
1819 for (unsigned i
= 0; i
< reg
->size
; i
++)
1820 __regmask_set(regmask
, reg
, reg
->array
.offset
+ i
);
1822 for (unsigned mask
= reg
->wrmask
, n
= reg
->num
; mask
; mask
>>= 1, n
++)
1824 __regmask_set(regmask
, reg
, n
);
1828 static inline void regmask_or(regmask_t
*dst
, regmask_t
*a
, regmask_t
*b
)
1831 for (i
= 0; i
< ARRAY_SIZE(*dst
); i
++)
1832 (*dst
)[i
] = (*a
)[i
] | (*b
)[i
];
1835 static inline bool regmask_get(regmask_t
*regmask
,
1836 struct ir3_register
*reg
)
1838 if (reg
->flags
& IR3_REG_RELATIV
) {
1839 for (unsigned i
= 0; i
< reg
->size
; i
++)
1840 if (__regmask_get(regmask
, reg
, reg
->array
.offset
+ i
))
1843 for (unsigned mask
= reg
->wrmask
, n
= reg
->num
; mask
; mask
>>= 1, n
++)
1845 if (__regmask_get(regmask
, reg
, n
))
1851 /* ************************************************************************* */