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
))
116 value
->reg
.file
>= NV_FILE_MEM_C(0) &&
117 value
->reg
.file
<= NV_FILE_MEM_C(15))
119 if (s
== 2 && nvi
->src
[1]->value
->reg
.file
== NV_FILE_GPR
)
124 return /* TRUE */ FALSE
; /* don't turn MOVs into loads */
130 /* Return whether this instruction can be executed conditionally. */
132 nv50_nvi_can_predicate(struct nv_instruction
*nvi
)
138 for (i
= 0; i
< 4 && nvi
->src
[i
]; ++i
)
139 if (nvi
->src
[i
]->value
->reg
.file
== NV_FILE_IMM
)
145 nv50_supported_src_mods(uint opcode
, int s
)
149 return NV_MOD_NEG
| NV_MOD_ABS
; /* obviously */
168 return NV_MOD_ABS
| NV_MOD_NEG
;
175 nv_nvi_refcount(struct nv_instruction
*nvi
)
179 rc
= nvi
->flags_def
? nvi
->flags_def
->refc
: 0;
181 for (i
= 0; i
< 4; ++i
) {
184 rc
+= nvi
->def
[i
]->refc
;
190 nvcg_replace_value(struct nv_pc
*pc
, struct nv_value
*old_val
,
191 struct nv_value
*new_val
)
195 if (old_val
== new_val
)
196 return old_val
->refc
;
198 for (i
= 0, n
= 0; i
< pc
->num_refs
; ++i
) {
199 if (pc
->refs
[i
]->value
== old_val
) {
201 nv_reference(pc
, &pc
->refs
[i
], new_val
);
208 nvcg_find_constant(struct nv_ref
*ref
)
210 struct nv_value
*src
;
216 while (src
->insn
&& src
->insn
->opcode
== NV_OP_MOV
) {
217 assert(!src
->insn
->src
[0]->mod
);
218 src
= src
->insn
->src
[0]->value
;
220 if ((src
->reg
.file
== NV_FILE_IMM
) ||
221 (src
->insn
&& src
->insn
->opcode
== NV_OP_LDA
&&
222 src
->insn
->src
[0]->value
->reg
.file
>= NV_FILE_MEM_C(0) &&
223 src
->insn
->src
[0]->value
->reg
.file
<= NV_FILE_MEM_C(15)))
229 nvcg_find_immediate(struct nv_ref
*ref
)
231 struct nv_value
*src
= nvcg_find_constant(ref
);
233 return (src
&& src
->reg
.file
== NV_FILE_IMM
) ? src
: NULL
;
237 nv_pc_free_refs(struct nv_pc
*pc
)
240 for (i
= 0; i
< pc
->num_refs
; i
+= 64)
245 edge_name(ubyte type
)
248 case CFG_EDGE_FORWARD
: return "forward";
249 case CFG_EDGE_BACK
: return "back";
250 case CFG_EDGE_LOOP_ENTER
: return "loop";
251 case CFG_EDGE_LOOP_LEAVE
: return "break";
252 case CFG_EDGE_FAKE
: return "fake";
259 nv_pc_pass_in_order(struct nv_basic_block
*root
, nv_pc_pass_func f
, void *priv
)
261 struct nv_basic_block
*bb
[64], *bbb
[16], *b
;
272 for (j
= 1; j
>= 0; --j
) {
276 switch (b
->out_kind
[j
]) {
279 case CFG_EDGE_FORWARD
:
281 if (++b
->out
[j
]->priv
== b
->out
[j
]->num_in
)
284 case CFG_EDGE_LOOP_ENTER
:
287 case CFG_EDGE_LOOP_LEAVE
:
288 bbb
[pp
++] = b
->out
[j
];
301 bb
[pp
- 1] = bbb
[pp
- 1];
307 nv_do_print_function(void *priv
, struct nv_basic_block
*b
)
309 struct nv_instruction
*i
= b
->phi
;
311 debug_printf("=== BB %i ", b
->id
);
313 debug_printf("[%s -> %i] ", edge_name(b
->out_kind
[0]), b
->out
[0]->id
);
315 debug_printf("[%s -> %i] ", edge_name(b
->out_kind
[1]), b
->out
[1]->id
);
316 debug_printf("===\n");
321 for (; i
; i
= i
->next
)
322 nv_print_instruction(i
);
326 nv_print_function(struct nv_basic_block
*root
)
328 if (root
->subroutine
)
329 debug_printf("SUBROUTINE %i\n", root
->subroutine
);
331 debug_printf("MAIN\n");
333 nv_pc_pass_in_order(root
, nv_do_print_function
, root
);
337 nv_print_program(struct nv_pc
*pc
)
340 for (i
= 0; i
< pc
->num_subroutines
+ 1; ++i
)
342 nv_print_function(pc
->root
[i
]);
346 nvcg_show_bincode(struct nv_pc
*pc
)
350 for (i
= 0; i
< pc
->bin_size
/ 4; ++i
)
351 debug_printf("0x%08x ", pc
->emit
[i
]);
356 nv50_emit_program(struct nv_pc
*pc
)
358 uint32_t *code
= pc
->emit
;
361 NV50_DBGMSG("emitting program: size = %u\n", pc
->bin_size
);
363 for (n
= 0; n
< pc
->num_blocks
; ++n
) {
364 struct nv_instruction
*i
;
365 struct nv_basic_block
*b
= pc
->bb_list
[n
];
367 for (i
= b
->entry
; i
; i
= i
->next
) {
368 nv50_emit_instruction(pc
, i
);
370 pc
->bin_pos
+= 1 + (pc
->emit
[0] & 1);
371 pc
->emit
+= 1 + (pc
->emit
[0] & 1);
374 assert(pc
->emit
== &code
[pc
->bin_size
/ 4]);
376 /* XXX: we can do better than this ... */
377 if (!(pc
->emit
[-2] & 1) || (pc
->emit
[-2] & 2) || (pc
->emit
[-1] & 3)) {
378 pc
->emit
[0] = 0xf0000001;
379 pc
->emit
[1] = 0xe0000000;
384 code
[pc
->bin_size
/ 4 - 1] |= 1;
387 nvcg_show_bincode(pc
);
394 nv50_generate_code(struct nv50_translation_info
*ti
)
399 pc
= CALLOC_STRUCT(nv_pc
);
403 pc
->root
= CALLOC(ti
->subr_nr
+ 1, sizeof(pc
->root
[0]));
408 pc
->num_subroutines
= ti
->subr_nr
;
410 ret
= nv50_tgsi_to_nc(pc
, ti
);
414 nv_print_program(pc
);
417 pc
->opt_reload_elim
= ti
->store_to_memory
? FALSE
: TRUE
;
420 ret
= nv_pc_exec_pass0(pc
);
424 nv_print_program(pc
);
427 /* register allocation */
428 ret
= nv_pc_exec_pass1(pc
);
432 nv_print_program(pc
);
435 /* prepare for emission */
436 ret
= nv_pc_exec_pass2(pc
);
440 pc
->emit
= CALLOC(pc
->bin_size
/ 4 + 2, 4);
445 ret
= nv50_emit_program(pc
);
449 ti
->p
->code_size
= pc
->bin_size
;
450 ti
->p
->code
= pc
->emit
;
452 ti
->p
->immd_size
= pc
->immd_count
* 4;
453 ti
->p
->immd
= pc
->immd_buf
;
455 /* highest 16 bit reg to num of 32 bit regs */
456 ti
->p
->max_gpr
= (pc
->max_reg
[NV_FILE_GPR
] >> 1) + 1;
458 ti
->p
->fixups
= pc
->fixups
;
459 ti
->p
->num_fixups
= pc
->num_fixups
;
461 NV50_DBGMSG("SHADER TRANSLATION - %s\n", ret
? "failure" : "success");
469 if (ret
) { /* on success, these will be referenced by nv50_program */
482 nvbb_insert_phi(struct nv_basic_block
*b
, struct nv_instruction
*i
)
489 assert(!b
->entry
->prev
&& b
->exit
);
497 if (b
->entry
->opcode
== NV_OP_PHI
) { /* insert after entry */
498 assert(b
->entry
== b
->exit
);
503 } else { /* insert before entry */
504 assert(b
->entry
->prev
&& b
->exit
);
506 i
->prev
= b
->entry
->prev
;
514 nvbb_insert_tail(struct nv_basic_block
*b
, struct nv_instruction
*i
)
516 if (i
->opcode
== NV_OP_PHI
) {
517 nvbb_insert_phi(b
, i
);
526 if (i
->prev
&& i
->prev
->opcode
== NV_OP_PHI
)
531 b
->num_instructions
++;
535 nvi_insert_after(struct nv_instruction
*at
, struct nv_instruction
*ni
)
538 nvbb_insert_tail(at
->bb
, ni
);
548 nv_nvi_delete(struct nv_instruction
*nvi
)
550 struct nv_basic_block
*b
= nvi
->bb
;
553 /* debug_printf("REM: "); nv_print_instruction(nvi); */
555 for (j
= 0; j
< 5; ++j
)
556 nv_reference(NULL
, &nvi
->src
[j
], NULL
);
557 nv_reference(NULL
, &nvi
->flags_src
, NULL
);
560 nvi
->next
->prev
= nvi
->prev
;
562 assert(nvi
== b
->exit
);
567 nvi
->prev
->next
= nvi
->next
;
569 if (nvi
== b
->entry
) {
570 /* PHIs don't get hooked to b->entry */
571 b
->entry
= nvi
->next
;
572 assert(!nvi
->prev
|| nvi
->prev
->opcode
== NV_OP_PHI
);
576 if (nvi
->opcode
!= NV_OP_PHI
)
577 NV50_DBGMSG("NOTE: b->phi points to non-PHI instruction\n");
580 if (!nvi
->next
|| nvi
->next
->opcode
!= NV_OP_PHI
)
588 nv_nvi_permute(struct nv_instruction
*i1
, struct nv_instruction
*i2
)
590 struct nv_basic_block
*b
= i1
->bb
;
592 assert(i1
->opcode
!= NV_OP_PHI
&&
593 i2
->opcode
!= NV_OP_PHI
);
594 assert(i1
->next
== i2
);
614 nvbb_attach_block(struct nv_basic_block
*parent
,
615 struct nv_basic_block
*b
, ubyte edge_kind
)
617 assert(b
->num_in
< 8);
619 if (parent
->out
[0]) {
620 assert(!parent
->out
[1]);
622 parent
->out_kind
[1] = edge_kind
;
625 parent
->out_kind
[0] = edge_kind
;
628 b
->in
[b
->num_in
] = parent
;
629 b
->in_kind
[b
->num_in
++] = edge_kind
;
632 /* NOTE: all BRKs are treated as conditional, so there are 2 outgoing BBs */
635 nvbb_dominated_by(struct nv_basic_block
*b
, struct nv_basic_block
*d
)
642 for (j
= 0; j
< b
->num_in
; ++j
)
643 if ((b
->in_kind
[j
] != CFG_EDGE_BACK
) && !nvbb_dominated_by(b
->in
[j
], d
))
646 return j
? TRUE
: FALSE
;
649 /* check if bf (future) can be reached from bp (past) */
651 nvbb_reachable_by(struct nv_basic_block
*bf
, struct nv_basic_block
*bp
,
652 struct nv_basic_block
*bt
)
659 if (bp
->out
[0] && !IS_WALL_EDGE(bp
->out_kind
[0]) &&
660 nvbb_reachable_by(bf
, bp
->out
[0], bt
))
662 if (bp
->out
[1] && !IS_WALL_EDGE(bp
->out_kind
[1]) &&
663 nvbb_reachable_by(bf
, bp
->out
[1], bt
))
668 static struct nv_basic_block
*
669 nvbb_find_dom_frontier(struct nv_basic_block
*b
, struct nv_basic_block
*df
)
671 struct nv_basic_block
*out
;
674 if (!nvbb_dominated_by(df
, b
)) {
675 for (i
= 0; i
< df
->num_in
; ++i
) {
676 if (df
->in_kind
[i
] == CFG_EDGE_BACK
)
678 if (nvbb_dominated_by(df
->in
[i
], b
))
682 for (i
= 0; i
< 2 && df
->out
[i
]; ++i
) {
683 if (df
->out_kind
[i
] == CFG_EDGE_BACK
)
685 if ((out
= nvbb_find_dom_frontier(b
, df
->out
[i
])))
691 struct nv_basic_block
*
692 nvbb_dom_frontier(struct nv_basic_block
*b
)
694 struct nv_basic_block
*df
;
697 for (i
= 0; i
< 2 && b
->out
[i
]; ++i
)
698 if ((df
= nvbb_find_dom_frontier(b
, b
->out
[i
])))