98715a98c0b892ff603a9ec8994677b2aa7b00e0
2 * Copyright (c) 2013 Rob Clark <robdclark@gmail.com>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 #include "util/u_debug.h"
32 #include "instr-a3xx.h"
33 #include "disasm.h" /* TODO move 'enum shader_t' somewhere else.. */
35 /* low level intermediate representation of an adreno shader program */
38 struct ir3_instruction
;
41 struct ir3
* fd_asm_parse(const char *src
);
45 uint16_t instrs_count
; /* expanded to account for rpt's */
46 /* NOTE: max_reg, etc, does not include registers not touched
47 * by the shader (ie. vertex fetched via VFD_DECODE but not
50 int8_t max_reg
; /* highest GPR # used by shader */
57 IR3_REG_CONST
= 0x001,
58 IR3_REG_IMMED
= 0x002,
60 IR3_REG_RELATIV
= 0x008,
62 IR3_REG_NEGATE
= 0x020,
65 IR3_REG_POS_INF
= 0x100,
66 /* (ei) flag, end-input? Set on last bary, presumably to signal
67 * that the shader needs no more input:
70 /* meta-flags, for intermediate stages of IR, ie.
71 * before register assignment is done:
73 IR3_REG_SSA
= 0x1000, /* 'instr' is ptr to assigning instr */
74 IR3_REG_IA
= 0x2000, /* meta-input dst is "assigned" */
75 IR3_REG_ADDR
= 0x4000, /* register is a0.x */
79 * the component is in the low two bits of the reg #, so
80 * rN.x becomes: (N << 2) | x
88 /* for IR3_REG_SSA, src registers contain ptr back to
89 * assigning instruction.
91 struct ir3_instruction
*instr
;
94 /* used for cat5 instructions, but also for internal/IR level
95 * tracking of what registers are read/written by an instruction.
96 * wrmask may be a bad name since it is used to represent both
97 * src and dst that touch multiple adjacent registers.
102 #define IR3_INSTR_SRCS 10
104 struct ir3_instruction
{
105 struct ir3_block
*block
;
109 /* (sy) flag is set on first instruction, and after sample
110 * instructions (probably just on RAW hazard).
112 IR3_INSTR_SY
= 0x001,
113 /* (ss) flag is set on first instruction, and first instruction
114 * to depend on the result of "long" instructions (RAW hazard):
116 * rcp, rsq, log2, exp2, sin, cos, sqrt
118 * It seems to synchronize until all in-flight instructions are
119 * completed, for example:
122 * add.f hr2.z, (neg)hr2.z, hc0.y
123 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
126 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
128 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
129 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
130 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
132 * The last mul.f does not have (ss) set, presumably because the
133 * (ss) on the previous instruction does the job.
135 * The blob driver also seems to set it on WAR hazards, although
136 * not really clear if this is needed or just blob compiler being
137 * sloppy. So far I haven't found a case where removing the (ss)
138 * causes problems for WAR hazard, but I could just be getting
142 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
145 IR3_INSTR_SS
= 0x002,
146 /* (jp) flag is set on jump targets:
148 IR3_INSTR_JP
= 0x004,
149 IR3_INSTR_UL
= 0x008,
150 IR3_INSTR_3D
= 0x010,
155 IR3_INSTR_S2EN
= 0x200,
156 /* meta-flags, for intermediate stages of IR, ie.
157 * before register assignment is done:
159 IR3_INSTR_MARK
= 0x1000,
163 struct ir3_register
*regs
[1 + IR3_INSTR_SRCS
];
171 type_t src_type
, dst_type
;
192 /* for meta-instructions, just used to hold extra data
193 * before instruction scheduling, etc
196 int off
; /* component/offset */
199 struct ir3_block
*if_block
, *else_block
;
202 struct ir3_block
*block
;
206 /* transient values used during various algorithms: */
208 /* The instruction depth is the max dependency distance to output.
210 * You can also think of it as the "cost", if we did any sort of
211 * optimization for register footprint. Ie. a value that is just
212 * result of moving a const to a reg would have a low cost, so to
213 * it could make sense to duplicate the instruction at various
214 * points where the result is needed to reduce register footprint.
216 * DEPTH_UNUSED used to mark unused instructions after depth
219 #define DEPTH_UNUSED ~0
222 /* Used just during cp stage, which comes before depth pass.
223 * For fanin, where we need a sequence of consecutive registers,
224 * keep track of each src instructions left (ie 'n-1') and right
225 * (ie 'n+1') neighbor. The front-end must insert enough mov's
226 * to ensure that each instruction has at most one left and at
227 * most one right neighbor. During the copy-propagation pass,
228 * we only remove mov's when we can preserve this constraint.
231 struct ir3_instruction
*left
, *right
;
232 uint16_t left_cnt
, right_cnt
;
235 struct ir3_instruction
*next
;
241 struct ir3_heap_chunk
;
244 unsigned instrs_count
, instrs_sz
;
245 struct ir3_instruction
**instrs
;
246 unsigned baryfs_count
, baryfs_sz
;
247 struct ir3_instruction
**baryfs
;
248 struct ir3_block
*block
;
250 struct ir3_heap_chunk
*chunk
;
255 unsigned ntemporaries
, ninputs
, noutputs
;
256 /* maps TGSI_FILE_TEMPORARY index back to the assigning instruction: */
257 struct ir3_instruction
**temporaries
;
258 struct ir3_instruction
**inputs
;
259 struct ir3_instruction
**outputs
;
260 /* only a single address register: */
261 struct ir3_instruction
*address
;
262 struct ir3_block
*parent
;
263 struct ir3_instruction
*head
;
266 struct ir3
* ir3_create(void);
267 void ir3_destroy(struct ir3
*shader
);
268 void * ir3_assemble(struct ir3
*shader
,
269 struct ir3_info
*info
, uint32_t gpu_id
);
270 void * ir3_alloc(struct ir3
*shader
, int sz
);
272 struct ir3_block
* ir3_block_create(struct ir3
*shader
,
273 unsigned ntmp
, unsigned nin
, unsigned nout
);
275 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
,
276 int category
, opc_t opc
);
277 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
);
278 const char *ir3_instr_name(struct ir3_instruction
*instr
);
280 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
284 static inline bool ir3_instr_check_mark(struct ir3_instruction
*instr
)
286 if (instr
->flags
& IR3_INSTR_MARK
)
287 return true; /* already visited */
288 instr
->flags
|= IR3_INSTR_MARK
;
292 static inline void ir3_clear_mark(struct ir3
*shader
)
294 /* TODO would be nice to drop the instruction array.. for
295 * new compiler, _clear_mark() is all we use it for, and
296 * we could probably manage a linked list instead..
298 * Also, we'll probably want to mark instructions within
299 * a block, so tracking the list of instrs globally is
300 * unlikely to be what we want.
303 for (i
= 0; i
< shader
->instrs_count
; i
++) {
304 struct ir3_instruction
*instr
= shader
->instrs
[i
];
305 instr
->flags
&= ~IR3_INSTR_MARK
;
309 static inline int ir3_instr_regno(struct ir3_instruction
*instr
,
310 struct ir3_register
*reg
)
313 for (i
= 0; i
< instr
->regs_count
; i
++)
314 if (reg
== instr
->regs
[i
])
326 static inline uint32_t regid(int num
, int comp
)
328 return (num
<< 2) | (comp
& 0x3);
331 static inline uint32_t reg_num(struct ir3_register
*reg
)
333 return reg
->num
>> 2;
336 static inline uint32_t reg_comp(struct ir3_register
*reg
)
338 return reg
->num
& 0x3;
341 static inline bool is_flow(struct ir3_instruction
*instr
)
343 return (instr
->category
== 0);
346 static inline bool is_kill(struct ir3_instruction
*instr
)
348 return is_flow(instr
) && (instr
->opc
== OPC_KILL
);
351 static inline bool is_nop(struct ir3_instruction
*instr
)
353 return is_flow(instr
) && (instr
->opc
== OPC_NOP
);
356 static inline bool is_alu(struct ir3_instruction
*instr
)
358 return (1 <= instr
->category
) && (instr
->category
<= 3);
361 static inline bool is_sfu(struct ir3_instruction
*instr
)
363 return (instr
->category
== 4);
366 static inline bool is_tex(struct ir3_instruction
*instr
)
368 return (instr
->category
== 5);
371 static inline bool is_input(struct ir3_instruction
*instr
)
373 return (instr
->category
== 2) && (instr
->opc
== OPC_BARY_F
);
376 static inline bool is_meta(struct ir3_instruction
*instr
)
378 /* TODO how should we count PHI (and maybe fan-in/out) which
379 * might actually contribute some instructions to the final
382 return (instr
->category
== -1);
385 static inline bool is_addr(struct ir3_instruction
*instr
)
387 return is_meta(instr
) && (instr
->opc
== OPC_META_DEREF
);
390 static inline bool writes_addr(struct ir3_instruction
*instr
)
392 if (instr
->regs_count
> 0) {
393 struct ir3_register
*dst
= instr
->regs
[0];
394 return !!(dst
->flags
& IR3_REG_ADDR
);
399 static inline bool writes_pred(struct ir3_instruction
*instr
)
401 if (instr
->regs_count
> 0) {
402 struct ir3_register
*dst
= instr
->regs
[0];
403 return reg_num(dst
) == REG_P0
;
408 static inline bool reg_gpr(struct ir3_register
*r
)
410 if (r
->flags
& (IR3_REG_CONST
| IR3_REG_IMMED
| IR3_REG_RELATIV
| IR3_REG_ADDR
))
412 if ((reg_num(r
) == REG_A0
) || (reg_num(r
) == REG_P0
))
419 void ir3_dump(struct ir3
*shader
, const char *name
,
420 struct ir3_block
*block
/* XXX maybe 'block' ptr should move to ir3? */,
422 void ir3_dump_instr_single(struct ir3_instruction
*instr
);
423 void ir3_dump_instr_list(struct ir3_instruction
*instr
);
425 /* flatten if/else: */
426 int ir3_block_flatten(struct ir3_block
*block
);
428 /* depth calculation: */
429 int ir3_delayslots(struct ir3_instruction
*assigner
,
430 struct ir3_instruction
*consumer
, unsigned n
);
431 void ir3_block_depth(struct ir3_block
*block
);
433 /* copy-propagate: */
434 void ir3_block_cp(struct ir3_block
*block
);
437 int ir3_block_sched(struct ir3_block
*block
);
439 /* register assignment: */
440 int ir3_block_ra(struct ir3_block
*block
, enum shader_t type
,
441 bool half_precision
, bool frag_coord
, bool frag_face
,
442 bool *has_samp
, int *max_bary
);
445 # define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]))
448 /* ************************************************************************* */
449 /* split this out or find some helper to use.. like main/bitset.h.. */
455 typedef uint8_t regmask_t
[2 * MAX_REG
/ 8];
457 static inline unsigned regmask_idx(struct ir3_register
*reg
)
459 unsigned num
= reg
->num
;
460 debug_assert(num
< MAX_REG
);
461 if (reg
->flags
& IR3_REG_HALF
)
466 static inline void regmask_init(regmask_t
*regmask
)
468 memset(regmask
, 0, sizeof(*regmask
));
471 static inline void regmask_set(regmask_t
*regmask
, struct ir3_register
*reg
)
473 unsigned idx
= regmask_idx(reg
);
475 for (i
= 0; i
< IR3_INSTR_SRCS
; i
++, idx
++)
476 if (reg
->wrmask
& (1 << i
))
477 (*regmask
)[idx
/ 8] |= 1 << (idx
% 8);
480 /* set bits in a if not set in b, conceptually:
483 static inline void regmask_set_if_not(regmask_t
*a
,
484 struct ir3_register
*reg
, regmask_t
*b
)
486 unsigned idx
= regmask_idx(reg
);
488 for (i
= 0; i
< IR3_INSTR_SRCS
; i
++, idx
++)
489 if (reg
->wrmask
& (1 << i
))
490 if (!((*b
)[idx
/ 8] & (1 << (idx
% 8))))
491 (*a
)[idx
/ 8] |= 1 << (idx
% 8);
494 static inline unsigned regmask_get(regmask_t
*regmask
,
495 struct ir3_register
*reg
)
497 unsigned idx
= regmask_idx(reg
);
499 for (i
= 0; i
< IR3_INSTR_SRCS
; i
++, idx
++)
500 if (reg
->wrmask
& (1 << i
))
501 if ((*regmask
)[idx
/ 8] & (1 << (idx
% 8)))
506 /* ************************************************************************* */