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
25 #define DESCEND_ARBITRARY(j, f) \
27 b->pass_seq = ctx->pc->pass_seq; \
29 for (j = 0; j < 2; ++j) \
30 if (b->out[j] && b->out[j]->pass_seq < ctx->pc->pass_seq) \
34 extern unsigned nv50_inst_min_size(struct nv_instruction
*);
41 values_equal(struct nv_value
*a
, struct nv_value
*b
)
44 return (a
->reg
.file
== b
->reg
.file
&& a
->join
->reg
.id
== b
->join
->reg
.id
);
48 inst_commutation_check(struct nv_instruction
*a
,
49 struct nv_instruction
*b
)
53 for (di
= 0; di
< 4; ++di
) {
56 for (si
= 0; si
< 5; ++si
) {
59 if (values_equal(a
->def
[di
], b
->src
[si
]->value
))
64 if (b
->flags_src
&& b
->flags_src
->value
== a
->flags_def
)
70 /* Check whether we can swap the order of the instructions,
71 * where a & b may be either the earlier or the later one.
74 inst_commutation_legal(struct nv_instruction
*a
,
75 struct nv_instruction
*b
)
77 return inst_commutation_check(a
, b
) && inst_commutation_check(b
, a
);
81 inst_cullable(struct nv_instruction
*nvi
)
83 return (!(nvi
->is_terminator
||
86 nv_nvi_refcount(nvi
)));
90 nvi_isnop(struct nv_instruction
*nvi
)
92 if (nvi
->opcode
== NV_OP_EXPORT
)
101 if (nvi
->def
[0]->join
->reg
.id
< 0)
104 if (nvi
->opcode
!= NV_OP_MOV
&& nvi
->opcode
!= NV_OP_SELECT
)
107 if (nvi
->def
[0]->reg
.file
!= nvi
->src
[0]->value
->reg
.file
)
110 if (nvi
->src
[0]->value
->join
->reg
.id
< 0) {
111 debug_printf("nvi_isnop: orphaned value detected\n");
115 if (nvi
->opcode
== NV_OP_SELECT
)
116 if (!values_equal(nvi
->def
[0], nvi
->src
[1]->value
))
119 return values_equal(nvi
->def
[0], nvi
->src
[0]->value
);
123 nv_pc_pass_pre_emission(void *priv
, struct nv_basic_block
*b
)
125 struct nv_pc
*pc
= (struct nv_pc
*)priv
;
126 struct nv_basic_block
*in
;
127 struct nv_instruction
*nvi
, *next
;
131 for (j
= pc
->num_blocks
- 1; j
>= 0 && !pc
->bb_list
[j
]->bin_size
; --j
);
135 /* check for no-op branches (BRA $PC+8) */
136 if (in
->exit
&& in
->exit
->opcode
== NV_OP_BRA
&& in
->exit
->target
== b
) {
140 for (++j
; j
< pc
->num_blocks
; ++j
)
141 pc
->bb_list
[j
]->bin_pos
-= 8;
143 nv_nvi_delete(in
->exit
);
145 b
->bin_pos
= in
->bin_pos
+ in
->bin_size
;
148 pc
->bb_list
[pc
->num_blocks
++] = b
;
152 for (nvi
= b
->entry
; nvi
; nvi
= next
) {
158 for (nvi
= b
->entry
; nvi
; nvi
= next
) {
161 size
= nv50_inst_min_size(nvi
);
162 if (nvi
->next
&& size
< 8)
165 if ((n32
& 1) && nvi
->next
&&
166 nv50_inst_min_size(nvi
->next
) == 4 &&
167 inst_commutation_legal(nvi
, nvi
->next
)) {
169 debug_printf("permuting: ");
170 nv_print_instruction(nvi
);
171 nv_print_instruction(nvi
->next
);
172 nv_nvi_permute(nvi
, nvi
->next
);
177 b
->bin_size
+= n32
& 1;
179 nvi
->prev
->is_long
= 1;
182 b
->bin_size
+= 1 + nvi
->is_long
;
186 debug_printf("block %p is now empty\n", b
);
188 if (!b
->exit
->is_long
) {
190 b
->exit
->is_long
= 1;
193 /* might have del'd a hole tail of instructions */
194 if (!b
->exit
->prev
->is_long
&& !(n32
& 1)) {
196 b
->exit
->prev
->is_long
= 1;
199 assert(!b
->entry
|| (b
->exit
&& b
->exit
->is_long
));
201 pc
->bin_size
+= b
->bin_size
*= 4;
205 nv_pc_exec_pass2(struct nv_pc
*pc
)
207 debug_printf("preparing %u blocks for emission\n", pc
->num_blocks
);
209 pc
->bb_list
= CALLOC(pc
->num_blocks
, sizeof(struct nv_basic_block
*));
212 nv_pc_pass_in_order(pc
->root
, nv_pc_pass_pre_emission
, pc
);
217 static INLINE boolean
218 is_cmem_load(struct nv_instruction
*nvi
)
220 return (nvi
->opcode
== NV_OP_LDA
&&
221 nvi
->src
[0]->value
->reg
.file
>= NV_FILE_MEM_C(0) &&
222 nvi
->src
[0]->value
->reg
.file
<= NV_FILE_MEM_C(15));
225 static INLINE boolean
226 is_smem_load(struct nv_instruction
*nvi
)
228 return (nvi
->opcode
== NV_OP_LDA
&&
229 (nvi
->src
[0]->value
->reg
.file
== NV_FILE_MEM_S
||
230 nvi
->src
[0]->value
->reg
.file
<= NV_FILE_MEM_P
));
233 static INLINE boolean
234 is_immd_move(struct nv_instruction
*nvi
)
236 return (nvi
->opcode
== NV_OP_MOV
&&
237 nvi
->src
[0]->value
->reg
.file
== NV_FILE_IMM
);
241 check_swap_src_0_1(struct nv_instruction
*nvi
)
243 static const ubyte cc_swapped
[8] = { 0, 4, 2, 6, 1, 5, 3, 7 };
245 struct nv_ref
*src0
= nvi
->src
[0], *src1
= nvi
->src
[1];
247 if (!nv_op_commutative(nvi
->opcode
))
249 assert(src0
&& src1
);
251 if (src1
->value
->reg
.file
== NV_FILE_IMM
) {
252 /* should only be present from folding a constant MUL part of a MAD */
253 assert(nvi
->opcode
== NV_OP_ADD
);
257 if (is_cmem_load(src0
->value
->insn
)) {
258 if (!is_cmem_load(src1
->value
->insn
)) {
261 /* debug_printf("swapping cmem load to 1\n"); */
264 if (is_smem_load(src1
->value
->insn
)) {
265 if (!is_smem_load(src0
->value
->insn
)) {
268 /* debug_printf("swapping smem load to 0\n"); */
272 if (nvi
->opcode
== NV_OP_SET
&& nvi
->src
[0] != src0
)
273 nvi
->set_cond
= cc_swapped
[nvi
->set_cond
];
283 nv_pass_fold_stores(struct nv_pass
*ctx
, struct nv_basic_block
*b
)
285 struct nv_instruction
*nvi
, *sti
;
288 for (sti
= b
->entry
; sti
; sti
= sti
->next
) {
289 if (!sti
->def
[0] || sti
->def
[0]->reg
.file
!= NV_FILE_OUT
)
292 /* only handling MOV to $oX here */
293 if (sti
->opcode
!= NV_OP_MOV
&& sti
->opcode
!= NV_OP_STA
)
296 nvi
= sti
->src
[0]->value
->insn
;
297 if (!nvi
|| nvi
->opcode
== NV_OP_PHI
)
299 assert(nvi
->def
[0] == sti
->src
[0]->value
);
301 if (nvi
->def
[0]->refc
> 1)
304 /* cannot write to $oX when using immediate */
305 for (j
= 0; j
< 4 && nvi
->src
[j
]; ++j
)
306 if (nvi
->src
[j
]->value
->reg
.file
== NV_FILE_IMM
)
311 nvi
->def
[0] = sti
->def
[0];
313 nvi
->fixed
= sti
->fixed
;
316 DESCEND_ARBITRARY(j
, nv_pass_fold_stores
);
322 nv_pass_fold_loads(struct nv_pass
*ctx
, struct nv_basic_block
*b
)
324 struct nv_instruction
*nvi
, *ld
;
327 for (nvi
= b
->entry
; nvi
; nvi
= nvi
->next
) {
328 check_swap_src_0_1(nvi
);
330 for (j
= 0; j
< 3; ++j
) {
333 ld
= nvi
->src
[j
]->value
->insn
;
337 if (is_immd_move(ld
) && nv50_nvi_can_use_imm(nvi
, j
)) {
338 nv_reference(ctx
->pc
, &nvi
->src
[j
], ld
->src
[0]->value
);
342 if (ld
->opcode
!= NV_OP_LDA
)
344 if (!nv50_nvi_can_load(nvi
, j
, ld
->src
[0]->value
))
347 if (j
== 0 && ld
->src
[4]) /* can't load shared mem */
350 /* fold it ! */ /* XXX: ref->insn */
351 nv_reference(ctx
->pc
, &nvi
->src
[j
], ld
->src
[0]->value
);
353 nv_reference(ctx
->pc
, &nvi
->src
[4], ld
->src
[4]->value
);
356 DESCEND_ARBITRARY(j
, nv_pass_fold_loads
);
362 nv_pass_lower_mods(struct nv_pass
*ctx
, struct nv_basic_block
*b
)
365 struct nv_instruction
*nvi
, *mi
, *next
;
368 for (nvi
= b
->entry
; nvi
; nvi
= next
) {
370 if (nvi
->opcode
== NV_OP_SUB
) {
371 nvi
->opcode
= NV_OP_ADD
;
372 nvi
->src
[1]->mod
^= NV_MOD_NEG
;
375 /* should not put any modifiers on NEG and ABS */
376 assert(nvi
->opcode
!= NV_MOD_NEG
|| !nvi
->src
[0]->mod
);
377 assert(nvi
->opcode
!= NV_MOD_ABS
|| !nvi
->src
[0]->mod
);
379 for (j
= 0; j
< 4; ++j
) {
383 mi
= nvi
->src
[j
]->value
->insn
;
386 if (mi
->def
[0]->refc
> 1)
389 if (mi
->opcode
== NV_OP_NEG
) mod
= NV_MOD_NEG
;
391 if (mi
->opcode
== NV_OP_ABS
) mod
= NV_MOD_ABS
;
395 if (nvi
->opcode
== NV_OP_ABS
)
396 mod
&= ~(NV_MOD_NEG
| NV_MOD_ABS
);
398 if (nvi
->opcode
== NV_OP_NEG
&& mod
== NV_MOD_NEG
) {
399 nvi
->opcode
= NV_OP_MOV
;
403 if (!(nv50_supported_src_mods(nvi
->opcode
, j
) & mod
))
406 nv_reference(ctx
->pc
, &nvi
->src
[j
], mi
->src
[0]->value
);
408 nvi
->src
[j
]->mod
^= mod
;
411 if (nvi
->opcode
== NV_OP_SAT
) {
412 mi
= nvi
->src
[0]->value
->insn
;
414 if ((mi
->opcode
== NV_OP_MAD
) && !mi
->flags_def
) {
416 mi
->def
[0] = nvi
->def
[0];
421 DESCEND_ARBITRARY(j
, nv_pass_lower_mods
);
426 #define SRC_IS_MUL(s) ((s)->insn && (s)->insn->opcode == NV_OP_MUL)
428 static struct nv_value
*
429 find_immediate(struct nv_ref
*ref
)
431 struct nv_value
*src
;
437 while (src
->insn
&& src
->insn
->opcode
== NV_OP_MOV
) {
438 assert(!src
->insn
->src
[0]->mod
);
439 src
= src
->insn
->src
[0]->value
;
441 return (src
->reg
.file
== NV_FILE_IMM
) ? src
: NULL
;
445 modifiers_apply(uint32_t *val
, ubyte type
, ubyte mod
)
447 if (mod
& NV_MOD_ABS
) {
448 if (type
== NV_TYPE_F32
)
451 if ((*val
) & (1 << 31))
454 if (mod
& NV_MOD_NEG
) {
455 if (type
== NV_TYPE_F32
)
463 modifiers_opcode(ubyte mod
)
466 case NV_MOD_NEG
: return NV_OP_NEG
;
467 case NV_MOD_ABS
: return NV_OP_ABS
;
476 constant_expression(struct nv_pc
*pc
, struct nv_instruction
*nvi
,
477 struct nv_value
*src0
, struct nv_value
*src1
)
479 struct nv_value
*val
;
489 type
= nvi
->def
[0]->reg
.type
;
492 u0
.u32
= src0
->reg
.imm
.u32
;
493 u1
.u32
= src1
->reg
.imm
.u32
;
495 modifiers_apply(&u0
.u32
, type
, nvi
->src
[0]->mod
);
496 modifiers_apply(&u0
.u32
, type
, nvi
->src
[1]->mod
);
498 switch (nvi
->opcode
) {
500 if (nvi
->src
[2]->value
->reg
.file
!= NV_FILE_GPR
)
505 case NV_TYPE_F32
: u
.f32
= u0
.f32
* u1
.f32
; break;
506 case NV_TYPE_U32
: u
.u32
= u0
.u32
* u1
.u32
; break;
507 case NV_TYPE_S32
: u
.s32
= u0
.s32
* u1
.s32
; break;
515 case NV_TYPE_F32
: u
.f32
= u0
.f32
+ u1
.f32
; break;
516 case NV_TYPE_U32
: u
.u32
= u0
.u32
+ u1
.u32
; break;
517 case NV_TYPE_S32
: u
.s32
= u0
.s32
+ u1
.s32
; break;
525 case NV_TYPE_F32
: u
.f32
= u0
.f32
- u1
.f32
;
526 case NV_TYPE_U32
: u
.u32
= u0
.u32
- u1
.u32
;
527 case NV_TYPE_S32
: u
.s32
= u0
.s32
- u1
.s32
;
537 nvi
->opcode
= NV_OP_MOV
;
539 val
= new_value(pc
, NV_FILE_IMM
, type
);
541 val
->reg
.imm
.u32
= u
.u32
;
543 nv_reference(pc
, &nvi
->src
[1], NULL
);
544 nv_reference(pc
, &nvi
->src
[0], val
);
546 if (nvi
->src
[2]) { /* from MAD */
547 nvi
->src
[1] = nvi
->src
[0];
548 nvi
->src
[0] = nvi
->src
[2];
550 nvi
->opcode
= NV_OP_ADD
;
555 constant_operand(struct nv_pc
*pc
,
556 struct nv_instruction
*nvi
, struct nv_value
*val
, int s
)
569 type
= nvi
->def
[0]->reg
.type
;
571 u
.u32
= val
->reg
.imm
.u32
;
572 modifiers_apply(&u
.u32
, type
, nvi
->src
[s
]->mod
);
574 switch (nvi
->opcode
) {
576 if ((type
== NV_TYPE_F32
&& u
.f32
== 1.0f
) ||
577 (NV_TYPE_ISINT(type
) && u
.u32
== 1)) {
578 if ((op
= modifiers_opcode(nvi
->src
[t
]->mod
)) == NV_OP_NOP
)
581 nv_reference(pc
, &nvi
->src
[s
], NULL
);
582 nvi
->src
[0] = nvi
->src
[t
];
585 if ((type
== NV_TYPE_F32
&& u
.f32
== 2.0f
) ||
586 (NV_TYPE_ISINT(type
) && u
.u32
== 2)) {
587 nvi
->opcode
= NV_OP_ADD
;
588 nv_reference(pc
, &nvi
->src
[s
], nvi
->src
[t
]->value
);
589 nvi
->src
[s
]->mod
= nvi
->src
[t
]->mod
;
591 if (type
== NV_TYPE_F32
&& u
.f32
== -1.0f
) {
592 if (nvi
->src
[t
]->mod
& NV_MOD_NEG
)
593 nvi
->opcode
= NV_OP_MOV
;
595 nvi
->opcode
= NV_OP_NEG
;
596 nv_reference(pc
, &nvi
->src
[s
], NULL
);
597 nvi
->src
[0] = nvi
->src
[t
];
600 if (type
== NV_TYPE_F32
&& u
.f32
== -2.0f
) {
601 nvi
->opcode
= NV_OP_ADD
;
602 nv_reference(pc
, &nvi
->src
[s
], nvi
->src
[t
]->value
);
603 nvi
->src
[s
]->mod
= (nvi
->src
[t
]->mod
^= NV_MOD_NEG
);
606 nvi
->opcode
= NV_OP_MOV
;
607 nv_reference(pc
, &nvi
->src
[t
], NULL
);
609 nvi
->src
[0] = nvi
->src
[1];
616 if ((op
= modifiers_opcode(nvi
->src
[t
]->mod
)) == NV_OP_NOP
)
619 nv_reference(pc
, &nvi
->src
[s
], NULL
);
620 nvi
->src
[0] = nvi
->src
[t
];
625 u
.f32
= 1.0f
/ u
.f32
;
626 (val
= new_value(pc
, NV_FILE_IMM
, NV_TYPE_F32
))->reg
.imm
.f32
= u
.f32
;
627 nvi
->opcode
= NV_OP_MOV
;
629 nv_reference(pc
, &nvi
->src
[0], val
);
632 u
.f32
= 1.0f
/ sqrtf(u
.f32
);
633 (val
= new_value(pc
, NV_FILE_IMM
, NV_TYPE_F32
))->reg
.imm
.f32
= u
.f32
;
634 nvi
->opcode
= NV_OP_MOV
;
636 nv_reference(pc
, &nvi
->src
[0], val
);
644 nv_pass_lower_arith(struct nv_pass
*ctx
, struct nv_basic_block
*b
)
646 struct nv_instruction
*nvi
, *next
;
649 for (nvi
= b
->entry
; nvi
; nvi
= next
) {
650 struct nv_value
*src0
, *src1
, *src
;
655 src0
= find_immediate(nvi
->src
[0]);
656 src1
= find_immediate(nvi
->src
[1]);
659 constant_expression(ctx
->pc
, nvi
, src0
, src1
);
662 constant_operand(ctx
->pc
, nvi
, src0
, 0);
665 constant_operand(ctx
->pc
, nvi
, src1
, 1);
668 /* try to combine MUL, ADD into MAD */
669 if (nvi
->opcode
!= NV_OP_ADD
)
672 src0
= nvi
->src
[0]->value
;
673 src1
= nvi
->src
[1]->value
;
675 if (SRC_IS_MUL(src0
) && src0
->refc
== 1)
678 if (SRC_IS_MUL(src1
) && src1
->refc
== 1)
683 nvi
->opcode
= NV_OP_MAD
;
684 mod
= nvi
->src
[(src
== src0
) ? 0 : 1]->mod
;
685 nv_reference(ctx
->pc
, &nvi
->src
[(src
== src0
) ? 0 : 1], NULL
);
686 nvi
->src
[2] = nvi
->src
[(src
== src0
) ? 1 : 0];
688 assert(!(mod
& ~NV_MOD_NEG
));
689 nvi
->src
[0] = new_ref(ctx
->pc
, src
->insn
->src
[0]->value
);
690 nvi
->src
[1] = new_ref(ctx
->pc
, src
->insn
->src
[1]->value
);
691 nvi
->src
[0]->mod
= src
->insn
->src
[0]->mod
^ mod
;
692 nvi
->src
[1]->mod
= src
->insn
->src
[1]->mod
;
694 DESCEND_ARBITRARY(j
, nv_pass_lower_arith
);
700 set $r2 g f32 $r2 $r3
701 cvt abs rn f32 $r2 s32 $r2
702 cvt f32 $c0 # f32 $r2
707 nv_pass_lower_cond(struct nv_pass
*ctx
, struct nv_basic_block
*b
)
709 /* XXX: easier in IR builder for now */
714 /* TODO: redundant store elimination */
717 struct load_record
*next
;
719 struct nv_value
*value
;
722 #define LOAD_RECORD_POOL_SIZE 1024
724 struct nv_pass_reld_elim
{
727 struct load_record
*imm
;
728 struct load_record
*mem_s
;
729 struct load_record
*mem_v
;
730 struct load_record
*mem_c
[16];
731 struct load_record
*mem_l
;
733 struct load_record pool
[LOAD_RECORD_POOL_SIZE
];
738 nv_pass_reload_elim(struct nv_pass_reld_elim
*ctx
, struct nv_basic_block
*b
)
740 struct load_record
**rec
, *it
;
741 struct nv_instruction
*ld
, *next
;
743 struct nv_value
*val
;
746 for (ld
= b
->entry
; ld
; ld
= next
) {
750 val
= ld
->src
[0]->value
;
753 if (ld
->opcode
== NV_OP_LINTERP
|| ld
->opcode
== NV_OP_PINTERP
) {
757 if (ld
->opcode
== NV_OP_LDA
) {
759 if (val
->reg
.file
>= NV_FILE_MEM_C(0) &&
760 val
->reg
.file
<= NV_FILE_MEM_C(15))
761 rec
= &ctx
->mem_c
[val
->reg
.file
- NV_FILE_MEM_C(0)];
763 if (val
->reg
.file
== NV_FILE_MEM_S
)
766 if (val
->reg
.file
== NV_FILE_MEM_L
)
769 if ((ld
->opcode
== NV_OP_MOV
) && (val
->reg
.file
== NV_FILE_IMM
)) {
770 data
= val
->reg
.imm
.u32
;
774 if (!rec
|| !ld
->def
[0]->refc
)
777 for (it
= *rec
; it
; it
= it
->next
)
778 if (it
->data
== data
)
782 if (ld
->def
[0]->reg
.id
>= 0)
783 it
->value
= ld
->def
[0];
785 nvcg_replace_value(ctx
->pc
, ld
->def
[0], it
->value
);
787 if (ctx
->alloc
== LOAD_RECORD_POOL_SIZE
)
789 it
= &ctx
->pool
[ctx
->alloc
++];
792 it
->value
= ld
->def
[0];
800 for (j
= 0; j
< 16; ++j
)
801 ctx
->mem_c
[j
] = NULL
;
805 DESCEND_ARBITRARY(j
, nv_pass_reload_elim
);
811 nv_pass_tex_mask(struct nv_pass
*ctx
, struct nv_basic_block
*b
)
815 for (i
= 0; i
< ctx
->pc
->num_instructions
; ++i
) {
816 struct nv_instruction
*nvi
= &ctx
->pc
->instructions
[i
];
817 struct nv_value
*def
[4];
819 if (!nv_is_vector_op(nvi
->opcode
))
823 for (c
= 0; c
< 4; ++c
) {
824 if (nvi
->def
[c
]->refc
)
825 nvi
->tex_mask
|= 1 << c
;
826 def
[c
] = nvi
->def
[c
];
830 for (c
= 0; c
< 4; ++c
)
831 if (nvi
->tex_mask
& (1 << c
))
832 nvi
->def
[j
++] = def
[c
];
833 for (c
= 0; c
< 4; ++c
)
834 if (!(nvi
->tex_mask
& (1 << c
)))
835 nvi
->def
[j
++] = def
[c
];
847 nv_pass_dce(struct nv_pass_dce
*ctx
, struct nv_basic_block
*b
)
850 struct nv_instruction
*nvi
, *next
;
852 for (nvi
= b
->entry
; nvi
; nvi
= next
) {
855 if (inst_cullable(nvi
)) {
861 DESCEND_ARBITRARY(j
, nv_pass_dce
);
866 static INLINE boolean
867 bb_simple_if_endif(struct nv_basic_block
*bb
)
869 return (bb
->out
[0] && bb
->out
[1] &&
870 bb
->out
[0]->out
[0] == bb
->out
[1] &&
871 !bb
->out
[0]->out
[1]);
875 nv_pass_flatten(struct nv_pass
*ctx
, struct nv_basic_block
*b
)
879 if (bb_simple_if_endif(b
)) {
881 debug_printf("nv_pass_flatten: total IF/ENDIF constructs: %i\n", ctx
->n
);
883 DESCEND_ARBITRARY(j
, nv_pass_flatten
);
888 /* local common subexpression elimination, stupid O(n^2) implementation */
890 nv_pass_cse(struct nv_pass
*ctx
, struct nv_basic_block
*b
)
892 struct nv_instruction
*ir
, *ik
, *next
;
893 struct nv_instruction
*entry
= b
->phi
? b
->phi
: b
->entry
;
899 for (ir
= entry
; ir
; ir
= next
) {
901 for (ik
= entry
; ik
!= ir
; ik
= ik
->next
) {
902 if (ir
->opcode
!= ik
->opcode
)
905 if (ik
->opcode
== NV_OP_LDA
||
906 ik
->opcode
== NV_OP_STA
||
907 ik
->opcode
== NV_OP_MOV
||
908 nv_is_vector_op(ik
->opcode
))
909 continue; /* ignore loads, stores & moves */
911 if (ik
->src
[4] || ir
->src
[4])
912 continue; /* don't mess with address registers */
914 if (ik
->flags_src
|| ir
->flags_src
||
915 ik
->flags_def
|| ir
->flags_def
)
916 continue; /* and also not with flags, for now */
918 for (s
= 0; s
< 3; ++s
) {
919 struct nv_value
*a
, *b
;
926 if (ik
->src
[s
]->mod
!= ir
->src
[s
]->mod
)
928 a
= ik
->src
[s
]->value
;
929 b
= ir
->src
[s
]->value
;
932 if (a
->reg
.file
!= b
->reg
.file
||
934 a
->reg
.id
!= b
->reg
.id
)
940 nvcg_replace_value(ctx
->pc
, ir
->def
[0], ik
->def
[0]);
947 DESCEND_ARBITRARY(s
, nv_pass_cse
);
953 nv_pc_exec_pass0(struct nv_pc
*pc
)
955 struct nv_pass_reld_elim
*reldelim
;
957 struct nv_pass_dce dce
;
964 ret
= nv_pass_flatten(&pass
, pc
->root
);
968 /* Do this first, so we don't have to pay attention
969 * to whether sources are supported memory loads.
972 ret
= nv_pass_lower_arith(&pass
, pc
->root
);
977 ret
= nv_pass_fold_loads(&pass
, pc
->root
);
982 ret
= nv_pass_fold_stores(&pass
, pc
->root
);
986 reldelim
= CALLOC_STRUCT(nv_pass_reld_elim
);
989 ret
= nv_pass_reload_elim(reldelim
, pc
->root
);
995 ret
= nv_pass_cse(&pass
, pc
->root
);
1000 ret
= nv_pass_lower_mods(&pass
, pc
->root
);
1008 ret
= nv_pass_dce(&dce
, pc
->root
);
1011 } while (dce
.removed
);
1013 ret
= nv_pass_tex_mask(&pass
, pc
->root
);