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
23 /* #define NV50PC_DEBUG */
26 #include "nv50_program.h"
30 /* returns TRUE if operands 0 and 1 can be swapped */
32 nv_op_commutative(uint opcode
)
50 /* return operand to which the address register applies */
52 nv50_indirect_opnd(struct nv_instruction
*i
)
67 nv50_nvi_can_use_imm(struct nv_instruction
*nvi
, int s
)
69 if (nvi
->flags_src
|| nvi
->flags_def
)
72 switch (nvi
->opcode
) {
80 return (s
== 1) && (nvi
->src
[0]->value
->reg
.file
== NV_FILE_GPR
) &&
81 (nvi
->def
[0]->reg
.file
== NV_FILE_GPR
);
84 return (nvi
->def
[0]->reg
.file
== NV_FILE_GPR
);
91 nv50_nvi_can_load(struct nv_instruction
*nvi
, int s
, struct nv_value
*value
)
95 for (i
= 0; i
< 3 && nvi
->src
[i
]; ++i
)
96 if (nvi
->src
[i
]->value
->reg
.file
== NV_FILE_IMM
)
99 switch (nvi
->opcode
) {
112 if (s
== 0 && (value
->reg
.file
== NV_FILE_MEM_S
||
113 value
->reg
.file
== NV_FILE_MEM_P
))
115 if (value
->reg
.file
< NV_FILE_MEM_C(0) ||
116 value
->reg
.file
> NV_FILE_MEM_C(15))
119 ((s
== 2) && (nvi
->src
[1]->value
->reg
.file
== NV_FILE_GPR
));
122 return /* TRUE */ FALSE
; /* don't turn MOVs into loads */
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 nvcg_find_constant(struct nv_ref
*ref
)
208 struct nv_value
*src
;
214 while (src
->insn
&& src
->insn
->opcode
== NV_OP_MOV
) {
215 assert(!src
->insn
->src
[0]->mod
);
216 src
= src
->insn
->src
[0]->value
;
218 if ((src
->reg
.file
== NV_FILE_IMM
) ||
219 (src
->insn
&& src
->insn
->opcode
== NV_OP_LDA
&&
220 src
->insn
->src
[0]->value
->reg
.file
>= NV_FILE_MEM_C(0) &&
221 src
->insn
->src
[0]->value
->reg
.file
<= NV_FILE_MEM_C(15)))
227 nvcg_find_immediate(struct nv_ref
*ref
)
229 struct nv_value
*src
= nvcg_find_constant(ref
);
231 return (src
&& src
->reg
.file
== NV_FILE_IMM
) ? src
: NULL
;
235 nv_pc_free_refs(struct nv_pc
*pc
)
238 for (i
= 0; i
< pc
->num_refs
; i
+= 64)
244 edge_name(ubyte type
)
247 case CFG_EDGE_FORWARD
: return "forward";
248 case CFG_EDGE_BACK
: return "back";
249 case CFG_EDGE_LOOP_ENTER
: return "loop";
250 case CFG_EDGE_LOOP_LEAVE
: return "break";
251 case CFG_EDGE_FAKE
: return "fake";
258 nv_pc_pass_in_order(struct nv_basic_block
*root
, nv_pc_pass_func f
, void *priv
)
260 struct nv_basic_block
*bb
[64], *bbb
[16], *b
;
271 for (j
= 1; j
>= 0; --j
) {
275 switch (b
->out_kind
[j
]) {
278 case CFG_EDGE_FORWARD
:
280 if (++b
->out
[j
]->priv
== b
->out
[j
]->num_in
)
283 case CFG_EDGE_LOOP_ENTER
:
286 case CFG_EDGE_LOOP_LEAVE
:
287 bbb
[pp
++] = b
->out
[j
];
300 bb
[pp
- 1] = bbb
[pp
- 1];
306 nv_do_print_function(void *priv
, struct nv_basic_block
*b
)
308 struct nv_instruction
*i
= b
->phi
;
310 debug_printf("=== BB %i ", b
->id
);
312 debug_printf("[%s -> %i] ", edge_name(b
->out_kind
[0]), b
->out
[0]->id
);
314 debug_printf("[%s -> %i] ", edge_name(b
->out_kind
[1]), b
->out
[1]->id
);
315 debug_printf("===\n");
320 for (; i
; i
= i
->next
)
321 nv_print_instruction(i
);
325 nv_print_function(struct nv_basic_block
*root
)
327 if (root
->subroutine
)
328 debug_printf("SUBROUTINE %i\n", root
->subroutine
);
330 debug_printf("MAIN\n");
332 nv_pc_pass_in_order(root
, nv_do_print_function
, root
);
336 nv_print_program(struct nv_pc
*pc
)
339 for (i
= 0; i
< pc
->num_subroutines
+ 1; ++i
)
341 nv_print_function(pc
->root
[i
]);
346 nv_do_print_cfgraph(struct nv_pc
*pc
, FILE *f
, struct nv_basic_block
*b
)
350 b
->pass_seq
= pc
->pass_seq
;
352 fprintf(f
, "\t%i [shape=box]\n", b
->id
);
354 for (i
= 0; i
< 2; ++i
) {
357 switch (b
->out_kind
[i
]) {
358 case CFG_EDGE_FORWARD
:
359 fprintf(f
, "\t%i -> %i;\n", b
->id
, b
->out
[i
]->id
);
361 case CFG_EDGE_LOOP_ENTER
:
362 fprintf(f
, "\t%i -> %i [color=green];\n", b
->id
, b
->out
[i
]->id
);
364 case CFG_EDGE_LOOP_LEAVE
:
365 fprintf(f
, "\t%i -> %i [color=red];\n", b
->id
, b
->out
[i
]->id
);
368 fprintf(f
, "\t%i -> %i;\n", b
->id
, b
->out
[i
]->id
);
371 fprintf(f
, "\t%i -> %i [style=dotted];\n", b
->id
, b
->out
[i
]->id
);
377 if (b
->out
[i
]->pass_seq
< pc
->pass_seq
)
378 nv_do_print_cfgraph(pc
, f
, b
->out
[i
]);
382 /* Print the control flow graph of subroutine @subr (0 == MAIN) to a file. */
384 nv_print_cfgraph(struct nv_pc
*pc
, const char *filepath
, int subr
)
388 f
= fopen(filepath
, "a");
392 fprintf(f
, "digraph G {\n");
396 nv_do_print_cfgraph(pc
, f
, pc
->root
[subr
]);
405 nvcg_show_bincode(struct nv_pc
*pc
)
409 for (i
= 0; i
< pc
->bin_size
/ 4; ++i
)
410 debug_printf("0x%08x ", pc
->emit
[i
]);
415 nv50_emit_program(struct nv_pc
*pc
)
417 uint32_t *code
= pc
->emit
;
420 NV50_DBGMSG("emitting program: size = %u\n", pc
->bin_size
);
422 for (n
= 0; n
< pc
->num_blocks
; ++n
) {
423 struct nv_instruction
*i
;
424 struct nv_basic_block
*b
= pc
->bb_list
[n
];
426 for (i
= b
->entry
; i
; i
= i
->next
) {
427 nv50_emit_instruction(pc
, i
);
429 pc
->bin_pos
+= 1 + (pc
->emit
[0] & 1);
430 pc
->emit
+= 1 + (pc
->emit
[0] & 1);
433 assert(pc
->emit
== &code
[pc
->bin_size
/ 4]);
435 /* XXX: we can do better than this ... */
436 if (!(pc
->emit
[-2] & 1) || (pc
->emit
[-2] & 2) || (pc
->emit
[-1] & 3)) {
437 pc
->emit
[0] = 0xf0000001;
438 pc
->emit
[1] = 0xe0000000;
443 code
[pc
->bin_size
/ 4 - 1] |= 1;
446 nvcg_show_bincode(pc
);
453 nv50_generate_code(struct nv50_translation_info
*ti
)
459 pc
= CALLOC_STRUCT(nv_pc
);
463 pc
->root
= CALLOC(ti
->subr_nr
+ 1, sizeof(pc
->root
[0]));
468 pc
->num_subroutines
= ti
->subr_nr
;
470 ret
= nv50_tgsi_to_nc(pc
, ti
);
474 nv_print_program(pc
);
477 pc
->opt_reload_elim
= ti
->store_to_memory
? FALSE
: TRUE
;
480 ret
= nv_pc_exec_pass0(pc
);
484 nv_print_program(pc
);
487 /* register allocation */
488 ret
= nv_pc_exec_pass1(pc
);
492 nv_print_program(pc
);
493 nv_print_cfgraph(pc
, "nv50_shader_cfgraph.dot", 0);
496 /* prepare for emission */
497 ret
= nv_pc_exec_pass2(pc
);
501 pc
->emit
= CALLOC(pc
->bin_size
/ 4 + 2, 4);
506 ret
= nv50_emit_program(pc
);
510 ti
->p
->code_size
= pc
->bin_size
;
511 ti
->p
->code
= pc
->emit
;
513 ti
->p
->immd_size
= pc
->immd_count
* 4;
514 ti
->p
->immd
= pc
->immd_buf
;
516 /* highest 16 bit reg to num of 32 bit regs */
517 ti
->p
->max_gpr
= (pc
->max_reg
[NV_FILE_GPR
] >> 1) + 1;
519 ti
->p
->fixups
= pc
->fixups
;
520 ti
->p
->num_fixups
= pc
->num_fixups
;
522 NV50_DBGMSG("SHADER TRANSLATION - %s\n", ret
? "failure" : "success");
527 for (i
= 0; i
< pc
->num_blocks
; ++i
)
528 FREE(pc
->bb_list
[i
]);
531 if (ret
) { /* on success, these will be referenced by nv50_program */
544 nvbb_insert_phi(struct nv_basic_block
*b
, struct nv_instruction
*i
)
551 assert(!b
->entry
->prev
&& b
->exit
);
559 if (b
->entry
->opcode
== NV_OP_PHI
) { /* insert after entry */
560 assert(b
->entry
== b
->exit
);
565 } else { /* insert before entry */
566 assert(b
->entry
->prev
&& b
->exit
);
568 i
->prev
= b
->entry
->prev
;
576 nvbb_insert_tail(struct nv_basic_block
*b
, struct nv_instruction
*i
)
578 if (i
->opcode
== NV_OP_PHI
) {
579 nvbb_insert_phi(b
, i
);
588 if (i
->prev
&& i
->prev
->opcode
== NV_OP_PHI
)
593 b
->num_instructions
++;
597 nvi_insert_after(struct nv_instruction
*at
, struct nv_instruction
*ni
)
600 nvbb_insert_tail(at
->bb
, ni
);
610 nv_nvi_delete(struct nv_instruction
*nvi
)
612 struct nv_basic_block
*b
= nvi
->bb
;
615 /* debug_printf("REM: "); nv_print_instruction(nvi); */
617 for (j
= 0; j
< 5; ++j
)
618 nv_reference(NULL
, &nvi
->src
[j
], NULL
);
619 nv_reference(NULL
, &nvi
->flags_src
, NULL
);
622 nvi
->next
->prev
= nvi
->prev
;
624 assert(nvi
== b
->exit
);
629 nvi
->prev
->next
= nvi
->next
;
631 if (nvi
== b
->entry
) {
632 /* PHIs don't get hooked to b->entry */
633 b
->entry
= nvi
->next
;
634 assert(!nvi
->prev
|| nvi
->prev
->opcode
== NV_OP_PHI
);
638 if (nvi
->opcode
!= NV_OP_PHI
)
639 NV50_DBGMSG("NOTE: b->phi points to non-PHI instruction\n");
642 if (!nvi
->next
|| nvi
->next
->opcode
!= NV_OP_PHI
)
650 nv_nvi_permute(struct nv_instruction
*i1
, struct nv_instruction
*i2
)
652 struct nv_basic_block
*b
= i1
->bb
;
654 assert(i1
->opcode
!= NV_OP_PHI
&&
655 i2
->opcode
!= NV_OP_PHI
);
656 assert(i1
->next
== i2
);
676 nvbb_attach_block(struct nv_basic_block
*parent
,
677 struct nv_basic_block
*b
, ubyte edge_kind
)
679 assert(b
->num_in
< 8);
681 if (parent
->out
[0]) {
682 assert(!parent
->out
[1]);
684 parent
->out_kind
[1] = edge_kind
;
687 parent
->out_kind
[0] = edge_kind
;
690 b
->in
[b
->num_in
] = parent
;
691 b
->in_kind
[b
->num_in
++] = edge_kind
;
694 /* NOTE: all BRKs are treated as conditional, so there are 2 outgoing BBs */
697 nvbb_dominated_by(struct nv_basic_block
*b
, struct nv_basic_block
*d
)
704 for (j
= 0; j
< b
->num_in
; ++j
)
705 if ((b
->in_kind
[j
] != CFG_EDGE_BACK
) && !nvbb_dominated_by(b
->in
[j
], d
))
708 return j
? TRUE
: FALSE
;
711 /* check if @bf (future) can be reached from @bp (past), stop at @bt */
713 nvbb_reachable_by(struct nv_basic_block
*bf
, struct nv_basic_block
*bp
,
714 struct nv_basic_block
*bt
)
716 struct nv_basic_block
*q
[NV_PC_MAX_BASIC_BLOCKS
], *b
;
730 assert(n
<= (1024 - 2));
732 for (i
= 0; i
< 2; ++i
) {
733 if (b
->out
[i
] && !IS_WALL_EDGE(b
->out_kind
[i
]) && !b
->out
[i
]->priv
) {
739 for (--n
; n
>= 0; --n
)
745 static struct nv_basic_block
*
746 nvbb_find_dom_frontier(struct nv_basic_block
*b
, struct nv_basic_block
*df
)
748 struct nv_basic_block
*out
;
751 if (!nvbb_dominated_by(df
, b
)) {
752 for (i
= 0; i
< df
->num_in
; ++i
) {
753 if (df
->in_kind
[i
] == CFG_EDGE_BACK
)
755 if (nvbb_dominated_by(df
->in
[i
], b
))
759 for (i
= 0; i
< 2 && df
->out
[i
]; ++i
) {
760 if (df
->out_kind
[i
] == CFG_EDGE_BACK
)
762 if ((out
= nvbb_find_dom_frontier(b
, df
->out
[i
])))
768 struct nv_basic_block
*
769 nvbb_dom_frontier(struct nv_basic_block
*b
)
771 struct nv_basic_block
*df
;
774 for (i
= 0; i
< 2 && b
->out
[i
]; ++i
)
775 if ((df
= nvbb_find_dom_frontier(b
, b
->out
[i
])))