2 * Copyright 2010 Christoph Bumiller
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 shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19 * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 #include "nv50_program.h"
28 /* returns TRUE if operands 0 and 1 can be swapped */
30 nv_op_commutative(uint opcode
)
48 /* return operand to which the address register applies */
50 nv50_indirect_opnd(struct nv_instruction
*i
)
65 nv50_nvi_can_use_imm(struct nv_instruction
*nvi
, int s
)
67 if (nvi
->flags_src
|| nvi
->flags_def
)
70 switch (nvi
->opcode
) {
78 return (s
== 1) && (nvi
->src
[0]->value
->reg
.file
== NV_FILE_GPR
) &&
79 (nvi
->def
[0]->reg
.file
== NV_FILE_GPR
);
82 return (nvi
->def
[0]->reg
.file
== NV_FILE_GPR
);
89 nv50_nvi_can_load(struct nv_instruction
*nvi
, int s
, struct nv_value
*value
)
93 for (i
= 0; i
< 3 && nvi
->src
[i
]; ++i
)
94 if (nvi
->src
[i
]->value
->reg
.file
== NV_FILE_IMM
)
97 switch (nvi
->opcode
) {
110 if (s
== 0 && (value
->reg
.file
== NV_FILE_MEM_S
||
111 value
->reg
.file
== NV_FILE_MEM_P
))
114 value
->reg
.file
>= NV_FILE_MEM_C(0) &&
115 value
->reg
.file
<= NV_FILE_MEM_C(15))
117 if (s
== 2 && nvi
->src
[1]->value
->reg
.file
== NV_FILE_GPR
)
128 /* Return whether this instruction can be executed conditionally. */
130 nv50_nvi_can_predicate(struct nv_instruction
*nvi
)
136 for (i
= 0; i
< 4 && nvi
->src
[i
]; ++i
)
137 if (nvi
->src
[i
]->value
->reg
.file
== NV_FILE_IMM
)
143 nv50_supported_src_mods(uint opcode
, int s
)
147 return NV_MOD_NEG
| NV_MOD_ABS
; /* obviously */
166 return NV_MOD_ABS
| NV_MOD_NEG
;
173 nv_nvi_refcount(struct nv_instruction
*nvi
)
177 rc
= nvi
->flags_def
? nvi
->flags_def
->refc
: 0;
179 for (i
= 0; i
< 4; ++i
) {
182 rc
+= nvi
->def
[i
]->refc
;
188 nvcg_replace_value(struct nv_pc
*pc
, struct nv_value
*old_val
,
189 struct nv_value
*new_val
)
193 if (old_val
== new_val
)
194 return old_val
->refc
;
196 for (i
= 0, n
= 0; i
< pc
->num_refs
; ++i
) {
197 if (pc
->refs
[i
]->value
== old_val
) {
199 nv_reference(pc
, &pc
->refs
[i
], new_val
);
206 nv_pc_free_refs(struct nv_pc
*pc
)
209 for (i
= 0; i
< pc
->num_refs
; i
+= 64)
214 edge_name(ubyte type
)
217 case CFG_EDGE_FORWARD
: return "forward";
218 case CFG_EDGE_BACK
: return "back";
219 case CFG_EDGE_LOOP_ENTER
: return "loop";
220 case CFG_EDGE_LOOP_LEAVE
: return "break";
227 nv_pc_pass_in_order(struct nv_basic_block
*root
, nv_pc_pass_func f
, void *priv
)
229 struct nv_basic_block
*bb
[64], *bbb
[16], *b
;
240 for (j
= 1; j
>= 0; --j
) {
244 switch (b
->out_kind
[j
]) {
247 case CFG_EDGE_FORWARD
:
248 if (++b
->out
[j
]->priv
== b
->out
[j
]->num_in
)
251 case CFG_EDGE_LOOP_ENTER
:
254 case CFG_EDGE_LOOP_LEAVE
:
255 bbb
[pp
++] = b
->out
[j
];
272 nv_do_print_program(void *priv
, struct nv_basic_block
*b
)
274 struct nv_instruction
*i
= b
->phi
;
276 debug_printf("=== BB %i ", b
->id
);
278 debug_printf("[%s -> %i] ", edge_name(b
->out_kind
[0]), b
->out
[0]->id
);
280 debug_printf("[%s -> %i] ", edge_name(b
->out_kind
[1]), b
->out
[1]->id
);
281 debug_printf("===\n");
286 for (; i
; i
= i
->next
)
287 nv_print_instruction(i
);
291 nv_print_program(struct nv_basic_block
*root
)
293 nv_pc_pass_in_order(root
, nv_do_print_program
, root
);
295 debug_printf("END\n\n");
299 nvcg_show_bincode(struct nv_pc
*pc
)
303 for (i
= 0; i
< pc
->bin_size
/ 4; ++i
)
304 debug_printf("0x%08x ", pc
->emit
[i
]);
309 nv50_emit_program(struct nv_pc
*pc
)
311 uint32_t *code
= pc
->emit
;
314 debug_printf("emitting program: size = %u\n", pc
->bin_size
);
316 for (n
= 0; n
< pc
->num_blocks
; ++n
) {
317 struct nv_instruction
*i
;
318 struct nv_basic_block
*b
= pc
->bb_list
[n
];
320 for (i
= b
->entry
; i
; i
= i
->next
) {
321 nv50_emit_instruction(pc
, i
);
323 pc
->bin_pos
+= 1 + (pc
->emit
[0] & 1);
324 pc
->emit
+= 1 + (pc
->emit
[0] & 1);
327 assert(pc
->emit
== &code
[pc
->bin_size
/ 4]);
329 /* XXX: we can do better than this ... */
330 if (!(pc
->emit
[-2] & 1) || (pc
->emit
[-2] & 2) || (pc
->emit
[-1] & 3)) {
331 pc
->emit
[0] = 0xf0000001;
332 pc
->emit
[1] = 0xe0000000;
337 code
[pc
->bin_size
/ 4 - 1] |= 1;
339 nvcg_show_bincode(pc
);
345 nv50_generate_code(struct nv50_translation_info
*ti
)
350 pc
= CALLOC_STRUCT(nv_pc
);
354 ret
= nv50_tgsi_to_nc(pc
, ti
);
357 nv_print_program(pc
->root
);
360 ret
= nv_pc_exec_pass0(pc
);
364 /* register allocation */
365 ret
= nv_pc_exec_pass1(pc
);
369 /* prepare for emission */
370 ret
= nv_pc_exec_pass2(pc
);
374 pc
->emit
= CALLOC(pc
->bin_size
/ 4 + 2, 4);
379 ret
= nv50_emit_program(pc
);
383 ti
->p
->code_size
= pc
->bin_size
;
384 ti
->p
->code
= pc
->emit
;
386 ti
->p
->immd_size
= pc
->immd_count
* 4;
387 ti
->p
->immd
= pc
->immd_buf
;
389 /* highest 16 bit reg to num of 32 bit regs */
390 ti
->p
->max_gpr
= (pc
->max_reg
[NV_FILE_GPR
] >> 1) + 1;
392 ti
->p
->fixups
= pc
->fixups
;
393 ti
->p
->num_fixups
= pc
->num_fixups
;
395 debug_printf("SHADER TRANSLATION - %s\n", ret
? "failure" : "success");
413 nvbb_insert_phi(struct nv_basic_block
*b
, struct nv_instruction
*i
)
420 assert(!b
->entry
->prev
&& b
->exit
);
428 if (b
->entry
->opcode
== NV_OP_PHI
) { /* insert after entry */
429 assert(b
->entry
== b
->exit
);
434 } else { /* insert before entry */
435 assert(b
->entry
->prev
&& b
->exit
);
437 i
->prev
= b
->entry
->prev
;
445 nvbb_insert_tail(struct nv_basic_block
*b
, struct nv_instruction
*i
)
447 if (i
->opcode
== NV_OP_PHI
) {
448 nvbb_insert_phi(b
, i
);
457 if (i
->prev
&& i
->prev
->opcode
== NV_OP_PHI
)
462 b
->num_instructions
++;
466 nv_nvi_delete(struct nv_instruction
*nvi
)
468 struct nv_basic_block
*b
= nvi
->bb
;
471 /* debug_printf("REM: "); nv_print_instruction(nvi); */
473 for (j
= 0; j
< 5; ++j
)
474 nv_reference(NULL
, &nvi
->src
[j
], NULL
);
475 nv_reference(NULL
, &nvi
->flags_src
, NULL
);
478 nvi
->next
->prev
= nvi
->prev
;
480 assert(nvi
== b
->exit
);
485 nvi
->prev
->next
= nvi
->next
;
487 if (nvi
== b
->entry
) {
488 /* PHIs don't get hooked to b->entry */
489 b
->entry
= nvi
->next
;
490 assert(!nvi
->prev
|| nvi
->prev
->opcode
== NV_OP_PHI
);
494 if (nvi
->opcode
!= NV_OP_PHI
)
495 debug_printf("NOTE: b->phi points to non-PHI instruction\n");
498 if (!nvi
->next
|| nvi
->next
->opcode
!= NV_OP_PHI
)
506 nv_nvi_permute(struct nv_instruction
*i1
, struct nv_instruction
*i2
)
508 struct nv_basic_block
*b
= i1
->bb
;
510 assert(i1
->opcode
!= NV_OP_PHI
&&
511 i2
->opcode
!= NV_OP_PHI
);
512 assert(i1
->next
== i2
);
532 nvbb_attach_block(struct nv_basic_block
*parent
,
533 struct nv_basic_block
*b
, ubyte edge_kind
)
535 assert(b
->num_in
< 8);
537 if (parent
->out
[0]) {
538 assert(!parent
->out
[1]);
540 parent
->out_kind
[1] = edge_kind
;
543 parent
->out_kind
[0] = edge_kind
;
546 b
->in
[b
->num_in
] = parent
;
547 b
->in_kind
[b
->num_in
++] = edge_kind
;
550 /* NOTE: all BRKs are treated as conditional, so there are 2 outgoing BBs */
553 nvbb_dominated_by(struct nv_basic_block
*b
, struct nv_basic_block
*d
)
560 for (j
= 0; j
< b
->num_in
; ++j
)
561 if ((b
->in_kind
[j
] != CFG_EDGE_BACK
) && !nvbb_dominated_by(b
->in
[j
], d
))
564 return j
? TRUE
: FALSE
;
567 /* check if bf (future) can be reached from bp (past) */
569 nvbb_reachable_by(struct nv_basic_block
*bf
, struct nv_basic_block
*bp
,
570 struct nv_basic_block
*bt
)
577 if (bp
->out
[0] && bp
->out_kind
[0] != CFG_EDGE_BACK
&&
578 nvbb_reachable_by(bf
, bp
->out
[0], bt
))
580 if (bp
->out
[1] && bp
->out_kind
[1] != CFG_EDGE_BACK
&&
581 nvbb_reachable_by(bf
, bp
->out
[1], bt
))
586 static struct nv_basic_block
*
587 nvbb_find_dom_frontier(struct nv_basic_block
*b
, struct nv_basic_block
*df
)
591 if (!nvbb_dominated_by(df
, b
)) {
592 for (i
= 0; i
< df
->num_in
; ++i
) {
593 if (df
->in_kind
[i
] == CFG_EDGE_BACK
)
595 if (nvbb_dominated_by(df
->in
[i
], b
))
599 for (i
= 0; i
< 2 && b
->out
[i
]; ++i
) {
600 if (b
->out_kind
[i
] == CFG_EDGE_BACK
)
602 if ((df
= nvbb_find_dom_frontier(b
, b
->out
[i
])))
608 struct nv_basic_block
*
609 nvbb_dom_frontier(struct nv_basic_block
*b
)
611 struct nv_basic_block
*df
;
614 for (i
= 0; i
< 2 && b
->out
[i
]; ++i
)
615 if ((df
= nvbb_find_dom_frontier(b
, b
->out
[i
])))