1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
31 * TGSI to LLVM IR translation -- SoA.
33 * @author Jose Fonseca <jfonseca@vmware.com>
35 * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell,
36 * Brian Paul, and others.
39 #include "pipe/p_config.h"
40 #include "pipe/p_shader_tokens.h"
41 #include "util/u_debug.h"
42 #include "util/u_math.h"
43 #include "util/u_memory.h"
44 #include "tgsi/tgsi_dump.h"
45 #include "tgsi/tgsi_exec.h"
46 #include "tgsi/tgsi_info.h"
47 #include "tgsi/tgsi_parse.h"
48 #include "tgsi/tgsi_util.h"
49 #include "tgsi/tgsi_scan.h"
50 #include "lp_bld_tgsi_action.h"
51 #include "lp_bld_type.h"
52 #include "lp_bld_const.h"
53 #include "lp_bld_arit.h"
54 #include "lp_bld_bitarit.h"
55 #include "lp_bld_gather.h"
56 #include "lp_bld_init.h"
57 #include "lp_bld_logic.h"
58 #include "lp_bld_swizzle.h"
59 #include "lp_bld_flow.h"
60 #include "lp_bld_quad.h"
61 #include "lp_bld_tgsi.h"
62 #include "lp_bld_limits.h"
63 #include "lp_bld_debug.h"
64 #include "lp_bld_printf.h"
67 static void lp_exec_mask_init(struct lp_exec_mask
*mask
, struct lp_build_context
*bld
)
69 LLVMTypeRef int_type
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
70 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
73 mask
->has_mask
= FALSE
;
74 mask
->cond_stack_size
= 0;
75 mask
->loop_stack_size
= 0;
76 mask
->call_stack_size
= 0;
78 mask
->int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, mask
->bld
->type
);
79 mask
->exec_mask
= mask
->ret_mask
= mask
->break_mask
= mask
->cont_mask
= mask
->cond_mask
=
80 LLVMConstAllOnes(mask
->int_vec_type
);
82 mask
->loop_limiter
= lp_build_alloca(bld
->gallivm
, int_type
, "looplimiter");
86 LLVMConstInt(int_type
, LP_MAX_TGSI_LOOP_ITERATIONS
, false),
90 static void lp_exec_mask_update(struct lp_exec_mask
*mask
)
92 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
94 if (mask
->loop_stack_size
) {
95 /*for loops we need to update the entire mask at runtime */
97 assert(mask
->break_mask
);
98 tmp
= LLVMBuildAnd(builder
,
102 mask
->exec_mask
= LLVMBuildAnd(builder
,
107 mask
->exec_mask
= mask
->cond_mask
;
109 if (mask
->call_stack_size
) {
110 mask
->exec_mask
= LLVMBuildAnd(builder
,
116 mask
->has_mask
= (mask
->cond_stack_size
> 0 ||
117 mask
->loop_stack_size
> 0 ||
118 mask
->call_stack_size
> 0);
121 static void lp_exec_mask_cond_push(struct lp_exec_mask
*mask
,
124 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
126 assert(mask
->cond_stack_size
< LP_MAX_TGSI_NESTING
);
127 if (mask
->cond_stack_size
== 0) {
128 assert(mask
->cond_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
130 mask
->cond_stack
[mask
->cond_stack_size
++] = mask
->cond_mask
;
131 assert(LLVMTypeOf(val
) == mask
->int_vec_type
);
132 mask
->cond_mask
= LLVMBuildAnd(builder
,
136 lp_exec_mask_update(mask
);
139 static void lp_exec_mask_cond_invert(struct lp_exec_mask
*mask
)
141 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
142 LLVMValueRef prev_mask
;
143 LLVMValueRef inv_mask
;
145 assert(mask
->cond_stack_size
);
146 prev_mask
= mask
->cond_stack
[mask
->cond_stack_size
- 1];
147 if (mask
->cond_stack_size
== 1) {
148 assert(prev_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
151 inv_mask
= LLVMBuildNot(builder
, mask
->cond_mask
, "");
153 mask
->cond_mask
= LLVMBuildAnd(builder
,
156 lp_exec_mask_update(mask
);
159 static void lp_exec_mask_cond_pop(struct lp_exec_mask
*mask
)
161 assert(mask
->cond_stack_size
);
162 mask
->cond_mask
= mask
->cond_stack
[--mask
->cond_stack_size
];
163 lp_exec_mask_update(mask
);
166 static void lp_exec_bgnloop(struct lp_exec_mask
*mask
)
168 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
170 if (mask
->loop_stack_size
== 0) {
171 assert(mask
->loop_block
== NULL
);
172 assert(mask
->cont_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
173 assert(mask
->break_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
174 assert(mask
->break_var
== NULL
);
177 assert(mask
->loop_stack_size
< LP_MAX_TGSI_NESTING
);
179 mask
->loop_stack
[mask
->loop_stack_size
].loop_block
= mask
->loop_block
;
180 mask
->loop_stack
[mask
->loop_stack_size
].cont_mask
= mask
->cont_mask
;
181 mask
->loop_stack
[mask
->loop_stack_size
].break_mask
= mask
->break_mask
;
182 mask
->loop_stack
[mask
->loop_stack_size
].break_var
= mask
->break_var
;
183 ++mask
->loop_stack_size
;
185 mask
->break_var
= lp_build_alloca(mask
->bld
->gallivm
, mask
->int_vec_type
, "");
186 LLVMBuildStore(builder
, mask
->break_mask
, mask
->break_var
);
188 mask
->loop_block
= lp_build_insert_new_block(mask
->bld
->gallivm
, "bgnloop");
190 LLVMBuildBr(builder
, mask
->loop_block
);
191 LLVMPositionBuilderAtEnd(builder
, mask
->loop_block
);
193 mask
->break_mask
= LLVMBuildLoad(builder
, mask
->break_var
, "");
195 lp_exec_mask_update(mask
);
198 static void lp_exec_break(struct lp_exec_mask
*mask
)
200 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
201 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
205 mask
->break_mask
= LLVMBuildAnd(builder
,
207 exec_mask
, "break_full");
209 lp_exec_mask_update(mask
);
212 static void lp_exec_continue(struct lp_exec_mask
*mask
)
214 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
215 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
219 mask
->cont_mask
= LLVMBuildAnd(builder
,
223 lp_exec_mask_update(mask
);
227 static void lp_exec_endloop(struct gallivm_state
*gallivm
,
228 struct lp_exec_mask
*mask
)
230 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
231 LLVMBasicBlockRef endloop
;
232 LLVMTypeRef int_type
= LLVMInt32TypeInContext(mask
->bld
->gallivm
->context
);
233 LLVMTypeRef reg_type
= LLVMIntTypeInContext(gallivm
->context
,
234 mask
->bld
->type
.width
*
235 mask
->bld
->type
.length
);
236 LLVMValueRef i1cond
, i2cond
, icond
, limiter
;
238 assert(mask
->break_mask
);
241 * Restore the cont_mask, but don't pop
243 assert(mask
->loop_stack_size
);
244 mask
->cont_mask
= mask
->loop_stack
[mask
->loop_stack_size
- 1].cont_mask
;
245 lp_exec_mask_update(mask
);
248 * Unlike the continue mask, the break_mask must be preserved across loop
251 LLVMBuildStore(builder
, mask
->break_mask
, mask
->break_var
);
253 /* Decrement the loop limiter */
254 limiter
= LLVMBuildLoad(builder
, mask
->loop_limiter
, "");
256 limiter
= LLVMBuildSub(
259 LLVMConstInt(int_type
, 1, false),
262 LLVMBuildStore(builder
, limiter
, mask
->loop_limiter
);
264 /* i1cond = (mask != 0) */
265 i1cond
= LLVMBuildICmp(
268 LLVMBuildBitCast(builder
, mask
->exec_mask
, reg_type
, ""),
269 LLVMConstNull(reg_type
), "");
271 /* i2cond = (looplimiter > 0) */
272 i2cond
= LLVMBuildICmp(
276 LLVMConstNull(int_type
), "");
278 /* if( i1cond && i2cond ) */
279 icond
= LLVMBuildAnd(builder
, i1cond
, i2cond
, "");
281 endloop
= lp_build_insert_new_block(mask
->bld
->gallivm
, "endloop");
283 LLVMBuildCondBr(builder
,
284 icond
, mask
->loop_block
, endloop
);
286 LLVMPositionBuilderAtEnd(builder
, endloop
);
288 assert(mask
->loop_stack_size
);
289 --mask
->loop_stack_size
;
290 mask
->loop_block
= mask
->loop_stack
[mask
->loop_stack_size
].loop_block
;
291 mask
->cont_mask
= mask
->loop_stack
[mask
->loop_stack_size
].cont_mask
;
292 mask
->break_mask
= mask
->loop_stack
[mask
->loop_stack_size
].break_mask
;
293 mask
->break_var
= mask
->loop_stack
[mask
->loop_stack_size
].break_var
;
295 lp_exec_mask_update(mask
);
298 /* stores val into an address pointed to by dst.
299 * mask->exec_mask is used to figure out which bits of val
300 * should be stored into the address
301 * (0 means don't store this bit, 1 means do store).
303 static void lp_exec_mask_store(struct lp_exec_mask
*mask
,
304 struct lp_build_context
*bld_store
,
309 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
311 /* Mix the predicate and execution mask */
312 if (mask
->has_mask
) {
314 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
316 pred
= mask
->exec_mask
;
321 LLVMValueRef real_val
, dst_val
;
323 dst_val
= LLVMBuildLoad(builder
, dst
, "");
324 real_val
= lp_build_select(bld_store
,
328 LLVMBuildStore(builder
, real_val
, dst
);
330 LLVMBuildStore(builder
, val
, dst
);
333 static void lp_exec_mask_call(struct lp_exec_mask
*mask
,
337 assert(mask
->call_stack_size
< LP_MAX_TGSI_NESTING
);
338 mask
->call_stack
[mask
->call_stack_size
].pc
= *pc
;
339 mask
->call_stack
[mask
->call_stack_size
].ret_mask
= mask
->ret_mask
;
340 mask
->call_stack_size
++;
344 static void lp_exec_mask_ret(struct lp_exec_mask
*mask
, int *pc
)
346 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
347 LLVMValueRef exec_mask
;
349 if (mask
->call_stack_size
== 0) {
350 /* returning from main() */
354 exec_mask
= LLVMBuildNot(builder
,
358 mask
->ret_mask
= LLVMBuildAnd(builder
,
360 exec_mask
, "ret_full");
362 lp_exec_mask_update(mask
);
365 static void lp_exec_mask_bgnsub(struct lp_exec_mask
*mask
)
369 static void lp_exec_mask_endsub(struct lp_exec_mask
*mask
, int *pc
)
371 assert(mask
->call_stack_size
);
372 mask
->call_stack_size
--;
373 *pc
= mask
->call_stack
[mask
->call_stack_size
].pc
;
374 mask
->ret_mask
= mask
->call_stack
[mask
->call_stack_size
].ret_mask
;
375 lp_exec_mask_update(mask
);
380 * Return pointer to a temporary register channel (src or dest).
381 * Note that indirect addressing cannot be handled here.
382 * \param index which temporary register
383 * \param chan which channel of the temp register.
386 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context
*bld
,
390 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
392 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
393 LLVMValueRef lindex
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
, index
* 4 + chan
);
394 return LLVMBuildGEP(builder
, bld
->temps_array
, &lindex
, 1, "");
397 return bld
->temps
[index
][chan
];
402 * Return pointer to a output register channel (src or dest).
403 * Note that indirect addressing cannot be handled here.
404 * \param index which output register
405 * \param chan which channel of the output register.
408 lp_get_output_ptr(struct lp_build_tgsi_soa_context
*bld
,
412 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
414 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
415 LLVMValueRef lindex
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
,
417 return LLVMBuildGEP(builder
, bld
->outputs_array
, &lindex
, 1, "");
420 return bld
->outputs
[index
][chan
];
426 * XXX the lp_build_gather() function should be capable of doing this
427 * with a little work.
430 build_gather(struct lp_build_context
*bld
,
431 LLVMValueRef base_ptr
,
432 LLVMValueRef indexes
)
434 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
435 LLVMValueRef res
= bld
->undef
;
439 * Loop over elements of index_vec, load scalar value, insert it into 'res'.
441 for (i
= 0; i
< bld
->type
.length
; i
++) {
442 LLVMValueRef ii
= lp_build_const_int32(bld
->gallivm
, i
);
443 LLVMValueRef index
= LLVMBuildExtractElement(builder
,
445 LLVMValueRef scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
,
446 &index
, 1, "gather_ptr");
447 LLVMValueRef scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
449 res
= LLVMBuildInsertElement(builder
, res
, scalar
, ii
, "");
457 * Scatter/store vector.
460 emit_mask_scatter(struct lp_build_tgsi_soa_context
*bld
,
461 LLVMValueRef base_ptr
,
462 LLVMValueRef indexes
,
464 struct lp_exec_mask
*mask
,
467 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
468 LLVMBuilderRef builder
= gallivm
->builder
;
471 /* Mix the predicate and execution mask */
472 if (mask
->has_mask
) {
474 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
477 pred
= mask
->exec_mask
;
482 * Loop over elements of index_vec, store scalar value.
484 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
485 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
486 LLVMValueRef index
= LLVMBuildExtractElement(builder
, indexes
, ii
, "");
487 LLVMValueRef scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
, &index
, 1, "scatter_ptr");
488 LLVMValueRef val
= LLVMBuildExtractElement(builder
, values
, ii
, "scatter_val");
489 LLVMValueRef scalar_pred
= pred
?
490 LLVMBuildExtractElement(builder
, pred
, ii
, "scatter_pred") : NULL
;
493 lp_build_printf(gallivm
, "scatter %d: val %f at %d %p\n",
494 ii
, val
, index
, scalar_ptr
);
497 LLVMValueRef real_val
, dst_val
;
498 dst_val
= LLVMBuildLoad(builder
, scalar_ptr
, "");
499 real_val
= lp_build_select(&bld
->elem_bld
, scalar_pred
, val
, dst_val
);
500 LLVMBuildStore(builder
, real_val
, scalar_ptr
);
503 LLVMBuildStore(builder
, val
, scalar_ptr
);
510 * Read the current value of the ADDR register, convert the floats to
511 * ints, add the base index and return the vector of offsets.
512 * The offsets will be used to index into the constant buffer or
513 * temporary register file.
516 get_indirect_index(struct lp_build_tgsi_soa_context
*bld
,
517 unsigned reg_file
, unsigned reg_index
,
518 const struct tgsi_src_register
*indirect_reg
)
520 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
521 struct lp_build_context
*uint_bld
= &bld
->bld_base
.uint_bld
;
522 /* always use X component of address register */
523 unsigned swizzle
= indirect_reg
->SwizzleX
;
526 LLVMValueRef max_index
;
529 assert(bld
->indirect_files
& (1 << reg_file
));
531 base
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, reg_index
);
534 rel
= LLVMBuildLoad(builder
,
535 bld
->addr
[indirect_reg
->Index
][swizzle
],
538 index
= lp_build_add(uint_bld
, base
, rel
);
540 max_index
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
,
542 bld
->bld_base
.info
->file_max
[reg_file
]);
544 assert(!uint_bld
->type
.sign
);
545 index
= lp_build_min(uint_bld
, index
, max_index
);
550 static struct lp_build_context
*
551 stype_to_fetch(struct lp_build_tgsi_context
* bld_base
,
552 enum tgsi_opcode_type stype
)
554 struct lp_build_context
*bld_fetch
;
557 case TGSI_TYPE_FLOAT
:
558 case TGSI_TYPE_UNTYPED
:
559 bld_fetch
= &bld_base
->base
;
561 case TGSI_TYPE_UNSIGNED
:
562 bld_fetch
= &bld_base
->uint_bld
;
564 case TGSI_TYPE_SIGNED
:
565 bld_fetch
= &bld_base
->int_bld
;
568 case TGSI_TYPE_DOUBLE
:
579 struct lp_build_tgsi_context
* bld_base
,
580 const struct tgsi_full_src_register
* reg
,
581 enum tgsi_opcode_type stype
,
584 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
585 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
586 LLVMBuilderRef builder
= gallivm
->builder
;
587 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
588 LLVMValueRef indirect_index
= NULL
;
589 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
591 /* XXX: Handle fetching xyzw components as a vector */
592 assert(swizzle
!= ~0);
594 if (reg
->Register
.Indirect
) {
595 indirect_index
= get_indirect_index(bld
,
601 if (reg
->Register
.Indirect
) {
602 LLVMValueRef swizzle_vec
=
603 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, swizzle
);
604 LLVMValueRef index_vec
; /* index into the const buffer */
606 /* index_vec = indirect_index * 4 + swizzle */
607 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
608 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
610 /* Gather values from the constant buffer */
611 return build_gather(bld_fetch
, bld
->consts_ptr
, index_vec
);
614 LLVMValueRef index
; /* index into the const buffer */
615 LLVMValueRef scalar
, scalar_ptr
;
617 index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
*4 + swizzle
);
619 scalar_ptr
= LLVMBuildGEP(builder
, bld
->consts_ptr
,
622 if (stype
!= TGSI_TYPE_FLOAT
&& stype
!= TGSI_TYPE_UNTYPED
) {
623 LLVMTypeRef ivtype
= LLVMPointerType(LLVMInt32TypeInContext(gallivm
->context
), 0);
624 LLVMValueRef temp_ptr
;
625 temp_ptr
= LLVMBuildBitCast(builder
, scalar_ptr
, ivtype
, "");
626 scalar
= LLVMBuildLoad(builder
, temp_ptr
, "");
628 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
630 return lp_build_broadcast_scalar(bld_fetch
, scalar
);
635 emit_fetch_immediate(
636 struct lp_build_tgsi_context
* bld_base
,
637 const struct tgsi_full_src_register
* reg
,
638 enum tgsi_opcode_type stype
,
641 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
642 LLVMValueRef res
= bld
->immediates
[reg
->Register
.Index
][swizzle
];
645 if (stype
== TGSI_TYPE_UNSIGNED
) {
646 res
= LLVMConstBitCast(res
, bld_base
->uint_bld
.vec_type
);
647 } else if (stype
== TGSI_TYPE_SIGNED
) {
648 res
= LLVMConstBitCast(res
, bld_base
->int_bld
.vec_type
);
655 struct lp_build_tgsi_context
* bld_base
,
656 const struct tgsi_full_src_register
* reg
,
657 enum tgsi_opcode_type stype
,
660 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
661 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
662 LLVMBuilderRef builder
= gallivm
->builder
;
663 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
664 LLVMValueRef indirect_index
= NULL
;
667 if (reg
->Register
.Indirect
) {
668 indirect_index
= get_indirect_index(bld
,
674 if (reg
->Register
.Indirect
) {
675 LLVMValueRef swizzle_vec
=
676 lp_build_const_int_vec(gallivm
, uint_bld
->type
, swizzle
);
677 LLVMValueRef length_vec
=
678 lp_build_const_int_vec(gallivm
, uint_bld
->type
, bld
->bld_base
.base
.type
.length
);
679 LLVMValueRef index_vec
; /* index into the const buffer */
680 LLVMValueRef inputs_array
;
681 LLVMTypeRef float4_ptr_type
;
683 /* index_vec = (indirect_index * 4 + swizzle) * length */
684 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
685 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
686 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
688 /* cast inputs_array pointer to float* */
689 float4_ptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
690 inputs_array
= LLVMBuildBitCast(builder
, bld
->inputs_array
,
691 float4_ptr_type
, "");
693 /* Gather values from the temporary register array */
694 res
= build_gather(&bld_base
->base
, inputs_array
, index_vec
);
696 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
697 LLVMValueRef lindex
= lp_build_const_int32(gallivm
,
698 reg
->Register
.Index
* 4 + swizzle
);
699 LLVMValueRef input_ptr
= LLVMBuildGEP(builder
,
700 bld
->inputs_array
, &lindex
, 1, "");
701 res
= LLVMBuildLoad(builder
, input_ptr
, "");
704 res
= bld
->inputs
[reg
->Register
.Index
][swizzle
];
712 emit_fetch_temporary(
713 struct lp_build_tgsi_context
* bld_base
,
714 const struct tgsi_full_src_register
* reg
,
715 enum tgsi_opcode_type stype
,
718 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
719 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
720 LLVMBuilderRef builder
= gallivm
->builder
;
721 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
722 LLVMValueRef indirect_index
= NULL
;
725 if (reg
->Register
.Indirect
) {
726 indirect_index
= get_indirect_index(bld
,
732 if (reg
->Register
.Indirect
) {
733 LLVMValueRef swizzle_vec
=
734 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, swizzle
);
735 LLVMValueRef length_vec
=
736 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
,
737 bld
->bld_base
.base
.type
.length
);
738 LLVMValueRef index_vec
; /* index into the const buffer */
739 LLVMValueRef temps_array
;
740 LLVMTypeRef float4_ptr_type
;
742 /* index_vec = (indirect_index * 4 + swizzle) * length */
743 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
744 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
745 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
747 /* cast temps_array pointer to float* */
748 float4_ptr_type
= LLVMPointerType(LLVMFloatTypeInContext(bld
->bld_base
.base
.gallivm
->context
), 0);
749 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
,
750 float4_ptr_type
, "");
752 /* Gather values from the temporary register array */
753 res
= build_gather(&bld_base
->base
, temps_array
, index_vec
);
756 LLVMValueRef temp_ptr
;
757 if (stype
!= TGSI_TYPE_FLOAT
&& stype
!= TGSI_TYPE_UNTYPED
) {
758 LLVMTypeRef itype
= LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4), 0);
759 LLVMValueRef tint_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
,
761 temp_ptr
= LLVMBuildBitCast(builder
, tint_ptr
, itype
, "");
763 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
764 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
766 return bld
->bld_base
.base
.undef
;
773 emit_fetch_system_value(
774 struct lp_build_tgsi_context
* bld_base
,
775 const struct tgsi_full_src_register
* reg
,
776 enum tgsi_opcode_type stype
,
779 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
780 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
781 LLVMBuilderRef builder
= gallivm
->builder
;
782 LLVMValueRef index
; /* index into the system value array */
783 LLVMValueRef scalar
, scalar_ptr
;
785 assert(!reg
->Register
.Indirect
);
787 index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
* 4 + swizzle
);
789 scalar_ptr
= LLVMBuildGEP(builder
, bld
->system_values_array
, &index
, 1, "");
790 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
792 return lp_build_broadcast_scalar(&bld
->bld_base
.base
, scalar
);
796 * Register fetch with derivatives.
800 struct lp_build_tgsi_soa_context
*bld
,
809 /* TODO: use interpolation coeffs for inputs */
812 *ddx
= lp_build_ddx(&bld
->bld_base
.base
, src
);
815 *ddy
= lp_build_ddy(&bld
->bld_base
.base
, src
);
823 emit_fetch_predicate(
824 struct lp_build_tgsi_soa_context
*bld
,
825 const struct tgsi_full_instruction
*inst
,
828 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
830 unsigned char swizzles
[4];
831 LLVMValueRef unswizzled
[4] = {NULL
, NULL
, NULL
, NULL
};
835 if (!inst
->Instruction
.Predicate
) {
836 TGSI_FOR_EACH_CHANNEL( chan
) {
842 swizzles
[0] = inst
->Predicate
.SwizzleX
;
843 swizzles
[1] = inst
->Predicate
.SwizzleY
;
844 swizzles
[2] = inst
->Predicate
.SwizzleZ
;
845 swizzles
[3] = inst
->Predicate
.SwizzleW
;
847 index
= inst
->Predicate
.Index
;
848 assert(index
< LP_MAX_TGSI_PREDS
);
850 TGSI_FOR_EACH_CHANNEL( chan
) {
851 unsigned swizzle
= swizzles
[chan
];
854 * Only fetch the predicate register channels that are actually listed
857 if (!unswizzled
[swizzle
]) {
858 value
= LLVMBuildLoad(builder
,
859 bld
->preds
[index
][swizzle
], "");
862 * Convert the value to an integer mask.
864 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
865 * is needlessly causing two comparisons due to storing the intermediate
866 * result as float vector instead of an integer mask vector.
868 value
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
869 bld
->bld_base
.base
.type
,
872 bld
->bld_base
.base
.zero
);
873 if (inst
->Predicate
.Negate
) {
874 value
= LLVMBuildNot(builder
, value
, "");
877 unswizzled
[swizzle
] = value
;
879 value
= unswizzled
[swizzle
];
891 struct lp_build_tgsi_context
*bld_base
,
892 const struct tgsi_full_instruction
*inst
,
898 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
899 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
900 LLVMBuilderRef builder
= gallivm
->builder
;
901 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
902 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
903 LLVMValueRef indirect_index
= NULL
;
904 struct lp_build_context
*bld_store
;
905 enum tgsi_opcode_type dtype
= tgsi_opcode_infer_dst_type(inst
->Instruction
.Opcode
);
909 case TGSI_TYPE_FLOAT
:
910 case TGSI_TYPE_UNTYPED
:
911 bld_store
= &bld_base
->base
;
913 case TGSI_TYPE_UNSIGNED
:
914 bld_store
= &bld_base
->uint_bld
;
916 case TGSI_TYPE_SIGNED
:
917 bld_store
= &bld_base
->int_bld
;
919 case TGSI_TYPE_DOUBLE
:
926 switch( inst
->Instruction
.Saturate
) {
930 case TGSI_SAT_ZERO_ONE
:
931 value
= lp_build_max(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.zero
);
932 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
935 case TGSI_SAT_MINUS_PLUS_ONE
:
936 value
= lp_build_max(&bld
->bld_base
.base
, value
, lp_build_const_vec(bld
->bld_base
.base
.gallivm
, bld
->bld_base
.base
.type
, -1.0));
937 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
944 if (reg
->Register
.Indirect
) {
945 indirect_index
= get_indirect_index(bld
,
950 assert(reg
->Register
.Index
<=
951 bld
->bld_base
.info
->file_max
[reg
->Register
.File
]);
954 switch( reg
->Register
.File
) {
955 case TGSI_FILE_OUTPUT
:
956 if (reg
->Register
.Indirect
) {
957 LLVMValueRef chan_vec
=
958 lp_build_const_int_vec(gallivm
, uint_bld
->type
, chan_index
);
959 LLVMValueRef length_vec
=
960 lp_build_const_int_vec(gallivm
, uint_bld
->type
, bld
->bld_base
.base
.type
.length
);
961 LLVMValueRef index_vec
; /* indexes into the temp registers */
962 LLVMValueRef outputs_array
;
963 LLVMValueRef pixel_offsets
;
964 LLVMTypeRef float_ptr_type
;
967 /* build pixel offset vector: {0, 1, 2, 3, ...} */
968 pixel_offsets
= uint_bld
->undef
;
969 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
970 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
971 pixel_offsets
= LLVMBuildInsertElement(builder
, pixel_offsets
,
975 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
976 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
977 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
978 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
979 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
982 LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
983 outputs_array
= LLVMBuildBitCast(builder
, bld
->outputs_array
,
986 /* Scatter store values into temp registers */
987 emit_mask_scatter(bld
, outputs_array
, index_vec
, value
,
988 &bld
->exec_mask
, pred
);
991 LLVMValueRef out_ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
,
993 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
, out_ptr
);
997 case TGSI_FILE_TEMPORARY
:
998 if (reg
->Register
.Indirect
) {
999 LLVMValueRef chan_vec
=
1000 lp_build_const_int_vec(gallivm
, uint_bld
->type
, chan_index
);
1001 LLVMValueRef length_vec
=
1002 lp_build_const_int_vec(gallivm
, uint_bld
->type
,
1003 bld
->bld_base
.base
.type
.length
);
1004 LLVMValueRef index_vec
; /* indexes into the temp registers */
1005 LLVMValueRef temps_array
;
1006 LLVMValueRef pixel_offsets
;
1007 LLVMTypeRef float_ptr_type
;
1010 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1011 pixel_offsets
= uint_bld
->undef
;
1012 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
1013 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1014 pixel_offsets
= LLVMBuildInsertElement(builder
, pixel_offsets
,
1018 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1019 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1020 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
1021 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1022 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1025 LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1026 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
,
1027 float_ptr_type
, "");
1029 /* Scatter store values into temp registers */
1030 emit_mask_scatter(bld
, temps_array
, index_vec
, value
,
1031 &bld
->exec_mask
, pred
);
1034 LLVMValueRef temp_ptr
;
1037 case TGSI_TYPE_UNSIGNED
:
1038 case TGSI_TYPE_SIGNED
: {
1039 LLVMTypeRef itype
= LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4);
1040 LLVMTypeRef ivtype
= LLVMPointerType(itype
, 0);
1041 LLVMValueRef tint_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
,
1043 LLVMValueRef temp_value_ptr
;
1045 temp_ptr
= LLVMBuildBitCast(builder
, tint_ptr
, ivtype
, "");
1046 temp_value_ptr
= LLVMBuildBitCast(builder
, value
, itype
, "");
1047 value
= temp_value_ptr
;
1051 case TGSI_TYPE_FLOAT
:
1052 case TGSI_TYPE_UNTYPED
:
1053 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
,
1058 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
, temp_ptr
);
1062 case TGSI_FILE_ADDRESS
:
1063 assert(dtype
== TGSI_TYPE_SIGNED
);
1064 assert(LLVMTypeOf(value
) == bld_base
->base
.int_vec_type
);
1065 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
,
1066 bld
->addr
[reg
->Register
.Index
][chan_index
]);
1069 case TGSI_FILE_PREDICATE
:
1070 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
,
1071 bld
->preds
[reg
->Register
.Index
][chan_index
]);
1081 struct lp_build_tgsi_context
* bld_base
,
1082 const struct tgsi_full_instruction
* inst
,
1083 const struct tgsi_opcode_info
* info
,
1084 LLVMValueRef dst
[4])
1087 unsigned chan_index
;
1088 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1091 LLVMValueRef pred
[TGSI_NUM_CHANNELS
];
1093 emit_fetch_predicate( bld
, inst
, pred
);
1095 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1096 emit_store_chan(bld_base
, inst
, 0, chan_index
, pred
[chan_index
], dst
[chan_index
]);
1102 * High-level instruction translators.
1106 emit_tex( struct lp_build_tgsi_soa_context
*bld
,
1107 const struct tgsi_full_instruction
*inst
,
1108 enum lp_build_tex_modifier modifier
,
1109 LLVMValueRef
*texel
)
1111 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1113 LLVMValueRef lod_bias
, explicit_lod
;
1114 LLVMValueRef oow
= NULL
;
1115 LLVMValueRef coords
[3];
1116 LLVMValueRef ddx
[3];
1117 LLVMValueRef ddy
[3];
1118 unsigned num_coords
;
1121 if (!bld
->sampler
) {
1122 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1123 for (i
= 0; i
< 4; i
++) {
1124 texel
[i
] = bld
->bld_base
.base
.undef
;
1129 switch (inst
->Texture
.Texture
) {
1130 case TGSI_TEXTURE_1D
:
1133 case TGSI_TEXTURE_1D_ARRAY
:
1134 case TGSI_TEXTURE_2D
:
1135 case TGSI_TEXTURE_RECT
:
1138 case TGSI_TEXTURE_SHADOW1D
:
1139 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1140 case TGSI_TEXTURE_SHADOW2D
:
1141 case TGSI_TEXTURE_SHADOWRECT
:
1142 case TGSI_TEXTURE_2D_ARRAY
:
1143 case TGSI_TEXTURE_3D
:
1144 case TGSI_TEXTURE_CUBE
:
1147 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1155 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
1156 lod_bias
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1157 explicit_lod
= NULL
;
1159 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
1161 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1165 explicit_lod
= NULL
;
1168 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
) {
1169 oow
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1170 oow
= lp_build_rcp(&bld
->bld_base
.base
, oow
);
1173 for (i
= 0; i
< num_coords
; i
++) {
1174 coords
[i
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, i
);
1175 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
1176 coords
[i
] = lp_build_mul(&bld
->bld_base
.base
, coords
[i
], oow
);
1178 for (i
= num_coords
; i
< 3; i
++) {
1179 coords
[i
] = bld
->bld_base
.base
.undef
;
1182 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
1183 LLVMValueRef index0
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
, 0);
1184 for (i
= 0; i
< num_coords
; i
++) {
1185 LLVMValueRef src1
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 1, i
);
1186 LLVMValueRef src2
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 2, i
);
1187 ddx
[i
] = LLVMBuildExtractElement(builder
, src1
, index0
, "");
1188 ddy
[i
] = LLVMBuildExtractElement(builder
, src2
, index0
, "");
1190 unit
= inst
->Src
[3].Register
.Index
;
1192 for (i
= 0; i
< num_coords
; i
++) {
1193 ddx
[i
] = lp_build_scalar_ddx( &bld
->bld_base
.base
, coords
[i
] );
1194 ddy
[i
] = lp_build_scalar_ddy( &bld
->bld_base
.base
, coords
[i
] );
1196 unit
= inst
->Src
[1].Register
.Index
;
1198 for (i
= num_coords
; i
< 3; i
++) {
1199 ddx
[i
] = LLVMGetUndef(bld
->bld_base
.base
.elem_type
);
1200 ddy
[i
] = LLVMGetUndef(bld
->bld_base
.base
.elem_type
);
1203 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
1204 bld
->bld_base
.base
.gallivm
,
1205 bld
->bld_base
.base
.type
,
1206 unit
, num_coords
, coords
,
1208 lod_bias
, explicit_lod
,
1213 emit_txq( struct lp_build_tgsi_soa_context
*bld
,
1214 const struct tgsi_full_instruction
*inst
,
1215 LLVMValueRef
*sizes_out
)
1217 LLVMValueRef explicit_lod
;
1218 unsigned num_coords
, has_lod
;
1221 switch (inst
->Texture
.Texture
) {
1222 case TGSI_TEXTURE_1D
:
1223 case TGSI_TEXTURE_SHADOW1D
:
1224 case TGSI_TEXTURE_SHADOW2D
:
1225 case TGSI_TEXTURE_SHADOWCUBE
:
1229 case TGSI_TEXTURE_2D
:
1230 case TGSI_TEXTURE_CUBE
:
1231 case TGSI_TEXTURE_1D_ARRAY
:
1232 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1236 case TGSI_TEXTURE_3D
:
1237 // case TGSI_TEXTURE_CUBE_ARRAY:
1238 // case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
1239 case TGSI_TEXTURE_2D_ARRAY
:
1240 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1245 case TGSI_TEXTURE_BUFFER
:
1250 case TGSI_TEXTURE_RECT
:
1251 case TGSI_TEXTURE_SHADOWRECT
:
1252 // case TGSI_TEXTURE_2D_MS:
1257 // case TGSI_TEXTURE_2D_MS_ARRAY:
1267 if (!bld
->sampler
) {
1268 _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1269 for (i
= 0; i
< num_coords
; i
++)
1270 sizes_out
[i
] = bld
->bld_base
.base
.undef
;
1275 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 2 );
1277 explicit_lod
= NULL
;
1279 bld
->sampler
->emit_size_query(bld
->sampler
,
1280 bld
->bld_base
.base
.gallivm
,
1281 inst
->Src
[1].Register
.Index
,
1287 near_end_of_shader(struct lp_build_tgsi_soa_context
*bld
,
1292 for (i
= 0; i
< 5; i
++) {
1295 if (pc
+ i
>= bld
->bld_base
.info
->num_instructions
)
1298 opcode
= bld
->bld_base
.instructions
[pc
+ i
].Instruction
.Opcode
;
1300 if (opcode
== TGSI_OPCODE_END
)
1303 if (opcode
== TGSI_OPCODE_TEX
||
1304 opcode
== TGSI_OPCODE_TXP
||
1305 opcode
== TGSI_OPCODE_TXD
||
1306 opcode
== TGSI_OPCODE_TXB
||
1307 opcode
== TGSI_OPCODE_TXL
||
1308 opcode
== TGSI_OPCODE_TXF
||
1309 opcode
== TGSI_OPCODE_TXQ
||
1310 opcode
== TGSI_OPCODE_CAL
||
1311 opcode
== TGSI_OPCODE_CALLNZ
||
1312 opcode
== TGSI_OPCODE_IF
||
1313 opcode
== TGSI_OPCODE_IFC
||
1314 opcode
== TGSI_OPCODE_BGNLOOP
||
1315 opcode
== TGSI_OPCODE_SWITCH
)
1325 * Kill fragment if any of the src register values are negative.
1329 struct lp_build_tgsi_soa_context
*bld
,
1330 const struct tgsi_full_instruction
*inst
,
1333 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1334 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
1335 LLVMValueRef terms
[TGSI_NUM_CHANNELS
];
1337 unsigned chan_index
;
1339 memset(&terms
, 0, sizeof terms
);
1341 TGSI_FOR_EACH_CHANNEL( chan_index
) {
1344 /* Unswizzle channel */
1345 swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
1347 /* Check if the component has not been already tested. */
1348 assert(swizzle
< TGSI_NUM_CHANNELS
);
1349 if( !terms
[swizzle
] )
1350 /* TODO: change the comparison operator instead of setting the sign */
1351 terms
[swizzle
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, chan_index
);
1355 TGSI_FOR_EACH_CHANNEL( chan_index
) {
1356 if(terms
[chan_index
]) {
1357 LLVMValueRef chan_mask
;
1360 * If term < 0 then mask = 0 else mask = ~0.
1362 chan_mask
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, terms
[chan_index
], bld
->bld_base
.base
.zero
);
1365 mask
= LLVMBuildAnd(builder
, mask
, chan_mask
, "");
1372 lp_build_mask_update(bld
->mask
, mask
);
1374 if (!near_end_of_shader(bld
, pc
))
1375 lp_build_mask_check(bld
->mask
);
1381 * Predicated fragment kill.
1382 * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1383 * The only predication is the execution mask which will apply if
1384 * we're inside a loop or conditional.
1387 emit_kilp(struct lp_build_tgsi_soa_context
*bld
,
1390 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1393 /* For those channels which are "alive", disable fragment shader
1396 if (bld
->exec_mask
.has_mask
) {
1397 mask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
1400 LLVMValueRef zero
= LLVMConstNull(bld
->bld_base
.base
.int_vec_type
);
1404 lp_build_mask_update(bld
->mask
, mask
);
1406 if (!near_end_of_shader(bld
, pc
))
1407 lp_build_mask_check(bld
->mask
);
1412 * Emit code which will dump the value of all the temporary registers
1416 emit_dump_temps(struct lp_build_tgsi_soa_context
*bld
)
1418 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1419 LLVMBuilderRef builder
= gallivm
->builder
;
1420 LLVMValueRef temp_ptr
;
1421 LLVMValueRef i0
= lp_build_const_int32(gallivm
, 0);
1422 LLVMValueRef i1
= lp_build_const_int32(gallivm
, 1);
1423 LLVMValueRef i2
= lp_build_const_int32(gallivm
, 2);
1424 LLVMValueRef i3
= lp_build_const_int32(gallivm
, 3);
1426 int n
= bld
->bld_base
.info
->file_max
[TGSI_FILE_TEMPORARY
];
1428 for (index
= 0; index
< n
; index
++) {
1429 LLVMValueRef idx
= lp_build_const_int32(gallivm
, index
);
1430 LLVMValueRef v
[4][4], res
;
1433 lp_build_printf(gallivm
, "TEMP[%d]:\n", idx
);
1435 for (chan
= 0; chan
< 4; chan
++) {
1436 temp_ptr
= lp_get_temp_ptr_soa(bld
, index
, chan
);
1437 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
1438 v
[chan
][0] = LLVMBuildExtractElement(builder
, res
, i0
, "");
1439 v
[chan
][1] = LLVMBuildExtractElement(builder
, res
, i1
, "");
1440 v
[chan
][2] = LLVMBuildExtractElement(builder
, res
, i2
, "");
1441 v
[chan
][3] = LLVMBuildExtractElement(builder
, res
, i3
, "");
1444 lp_build_printf(gallivm
, " X: %f %f %f %f\n",
1445 v
[0][0], v
[0][1], v
[0][2], v
[0][3]);
1446 lp_build_printf(gallivm
, " Y: %f %f %f %f\n",
1447 v
[1][0], v
[1][1], v
[1][2], v
[1][3]);
1448 lp_build_printf(gallivm
, " Z: %f %f %f %f\n",
1449 v
[2][0], v
[2][1], v
[2][2], v
[2][3]);
1450 lp_build_printf(gallivm
, " W: %f %f %f %f\n",
1451 v
[3][0], v
[3][1], v
[3][2], v
[3][3]);
1458 lp_emit_declaration_soa(
1459 struct lp_build_tgsi_context
*bld_base
,
1460 const struct tgsi_full_declaration
*decl
)
1462 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
1463 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1464 LLVMTypeRef vec_type
= bld
->bld_base
.base
.vec_type
;
1465 const unsigned first
= decl
->Range
.First
;
1466 const unsigned last
= decl
->Range
.Last
;
1469 for (idx
= first
; idx
<= last
; ++idx
) {
1470 assert(last
<= bld
->bld_base
.info
->file_max
[decl
->Declaration
.File
]);
1471 switch (decl
->Declaration
.File
) {
1472 case TGSI_FILE_TEMPORARY
:
1473 assert(idx
< LP_MAX_TGSI_TEMPS
);
1474 if (!(bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
))) {
1475 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1476 bld
->temps
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
, "temp");
1480 case TGSI_FILE_OUTPUT
:
1481 if (!(bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
1482 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1483 bld
->outputs
[idx
][i
] = lp_build_alloca(gallivm
,
1484 vec_type
, "output");
1488 case TGSI_FILE_ADDRESS
:
1489 /* ADDR registers are the only allocated with an integer LLVM IR type,
1490 * as they are guaranteed to always have integers.
1491 * XXX: Not sure if this exception is worthwhile (or the whole idea of
1492 * an ADDR register for that matter).
1494 assert(idx
< LP_MAX_TGSI_ADDRS
);
1495 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1496 bld
->addr
[idx
][i
] = lp_build_alloca(gallivm
, bld_base
->base
.int_vec_type
, "addr");
1499 case TGSI_FILE_PREDICATE
:
1500 assert(idx
< LP_MAX_TGSI_PREDS
);
1501 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1502 bld
->preds
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
,
1507 /* don't need to declare other vars */
1514 void lp_emit_immediate_soa(
1515 struct lp_build_tgsi_context
*bld_base
,
1516 const struct tgsi_full_immediate
*imm
)
1518 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
1519 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
1521 /* simply copy the immediate values into the next immediates[] slot */
1523 const uint size
= imm
->Immediate
.NrTokens
- 1;
1525 assert(bld
->num_immediates
< LP_MAX_TGSI_IMMEDIATES
);
1526 switch (imm
->Immediate
.DataType
) {
1527 case TGSI_IMM_FLOAT32
:
1528 for( i
= 0; i
< size
; ++i
)
1529 bld
->immediates
[bld
->num_immediates
][i
] =
1530 lp_build_const_vec(gallivm
, bld_base
->base
.type
, imm
->u
[i
].Float
);
1533 case TGSI_IMM_UINT32
:
1534 for( i
= 0; i
< size
; ++i
) {
1535 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->uint_bld
.type
, imm
->u
[i
].Uint
);
1536 bld
->immediates
[bld
->num_immediates
][i
] =
1537 LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
1541 case TGSI_IMM_INT32
:
1542 for( i
= 0; i
< size
; ++i
) {
1543 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->int_bld
.type
, imm
->u
[i
].Int
);
1544 bld
->immediates
[bld
->num_immediates
][i
] =
1545 LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
1550 for( i
= size
; i
< 4; ++i
)
1551 bld
->immediates
[bld
->num_immediates
][i
] = bld_base
->base
.undef
;
1553 bld
->num_immediates
++;
1558 const struct lp_build_tgsi_action
* action
,
1559 struct lp_build_tgsi_context
* bld_base
,
1560 struct lp_build_emit_data
* emit_data
)
1562 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1564 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
,
1565 &emit_data
->output
[emit_data
->chan
], NULL
);
1570 const struct lp_build_tgsi_action
* action
,
1571 struct lp_build_tgsi_context
* bld_base
,
1572 struct lp_build_emit_data
* emit_data
)
1574 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1576 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
, NULL
,
1577 &emit_data
->output
[emit_data
->chan
]);
1582 const struct lp_build_tgsi_action
* action
,
1583 struct lp_build_tgsi_context
* bld_base
,
1584 struct lp_build_emit_data
* emit_data
)
1586 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1588 emit_kilp(bld
, bld_base
->pc
- 1);
1593 const struct lp_build_tgsi_action
* action
,
1594 struct lp_build_tgsi_context
* bld_base
,
1595 struct lp_build_emit_data
* emit_data
)
1597 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1599 emit_kil(bld
, emit_data
->inst
, bld_base
->pc
- 1);
1604 const struct lp_build_tgsi_action
* action
,
1605 struct lp_build_tgsi_context
* bld_base
,
1606 struct lp_build_emit_data
* emit_data
)
1608 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1610 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
, emit_data
->output
);
1615 const struct lp_build_tgsi_action
* action
,
1616 struct lp_build_tgsi_context
* bld_base
,
1617 struct lp_build_emit_data
* emit_data
)
1619 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1621 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
1627 const struct lp_build_tgsi_action
* action
,
1628 struct lp_build_tgsi_context
* bld_base
,
1629 struct lp_build_emit_data
* emit_data
)
1631 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1633 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
1639 const struct lp_build_tgsi_action
* action
,
1640 struct lp_build_tgsi_context
* bld_base
,
1641 struct lp_build_emit_data
* emit_data
)
1643 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1645 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
1651 const struct lp_build_tgsi_action
* action
,
1652 struct lp_build_tgsi_context
* bld_base
,
1653 struct lp_build_emit_data
* emit_data
)
1655 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1657 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_PROJECTED
,
1663 const struct lp_build_tgsi_action
* action
,
1664 struct lp_build_tgsi_context
* bld_base
,
1665 struct lp_build_emit_data
* emit_data
)
1667 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1669 emit_txq(bld
, emit_data
->inst
, emit_data
->output
);
1674 const struct lp_build_tgsi_action
* action
,
1675 struct lp_build_tgsi_context
* bld_base
,
1676 struct lp_build_emit_data
* emit_data
)
1678 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1680 lp_exec_mask_call(&bld
->exec_mask
, emit_data
->inst
->Label
.Label
,
1686 const struct lp_build_tgsi_action
* action
,
1687 struct lp_build_tgsi_context
* bld_base
,
1688 struct lp_build_emit_data
* emit_data
)
1690 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1692 lp_exec_mask_ret(&bld
->exec_mask
, &bld_base
->pc
);
1697 const struct lp_build_tgsi_action
* action
,
1698 struct lp_build_tgsi_context
* bld_base
,
1699 struct lp_build_emit_data
* emit_data
)
1701 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1703 lp_exec_break(&bld
->exec_mask
);
1708 const struct lp_build_tgsi_action
* action
,
1709 struct lp_build_tgsi_context
* bld_base
,
1710 struct lp_build_emit_data
* emit_data
)
1713 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1715 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_NOTEQUAL
,
1716 emit_data
->args
[0], bld
->bld_base
.base
.zero
);
1717 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
1722 const struct lp_build_tgsi_action
* action
,
1723 struct lp_build_tgsi_context
* bld_base
,
1724 struct lp_build_emit_data
* emit_data
)
1726 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1728 lp_exec_bgnloop(&bld
->exec_mask
);
1733 const struct lp_build_tgsi_action
* action
,
1734 struct lp_build_tgsi_context
* bld_base
,
1735 struct lp_build_emit_data
* emit_data
)
1737 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1739 lp_exec_mask_bgnsub(&bld
->exec_mask
);
1744 const struct lp_build_tgsi_action
* action
,
1745 struct lp_build_tgsi_context
* bld_base
,
1746 struct lp_build_emit_data
* emit_data
)
1748 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1750 lp_exec_mask_cond_invert(&bld
->exec_mask
);
1755 const struct lp_build_tgsi_action
* action
,
1756 struct lp_build_tgsi_context
* bld_base
,
1757 struct lp_build_emit_data
* emit_data
)
1759 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1761 lp_exec_mask_cond_pop(&bld
->exec_mask
);
1766 const struct lp_build_tgsi_action
* action
,
1767 struct lp_build_tgsi_context
* bld_base
,
1768 struct lp_build_emit_data
* emit_data
)
1770 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1772 lp_exec_endloop(bld_base
->base
.gallivm
, &bld
->exec_mask
);
1777 const struct lp_build_tgsi_action
* action
,
1778 struct lp_build_tgsi_context
* bld_base
,
1779 struct lp_build_emit_data
* emit_data
)
1781 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1783 lp_exec_mask_endsub(&bld
->exec_mask
, &bld_base
->pc
);
1788 const struct lp_build_tgsi_action
* action
,
1789 struct lp_build_tgsi_context
* bld_base
,
1790 struct lp_build_emit_data
* emit_data
)
1792 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1794 lp_exec_continue(&bld
->exec_mask
);
1797 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1799 * XXX: What do the comments about xmm registers mean? Maybe they are left over
1800 * from old code, but there is no garauntee that LLVM will use those registers
1803 * XXX: There should be no calls to lp_build_emit_fetch in this function. This
1804 * should be handled by the emit_data->fetch_args function. */
1807 const struct lp_build_tgsi_action
* action
,
1808 struct lp_build_tgsi_context
* bld_base
,
1809 struct lp_build_emit_data
* emit_data
)
1811 LLVMValueRef tmp0
, tmp1
;
1812 LLVMValueRef tmp4
= NULL
;
1813 LLVMValueRef tmp5
= NULL
;
1814 LLVMValueRef tmp6
= NULL
;
1815 LLVMValueRef tmp7
= NULL
;
1816 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1818 uint dims
= (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_NRM
) ? 3 : 4;
1820 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) ||
1821 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
) ||
1822 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
) ||
1823 (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 4)) {
1825 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1828 /* xmm0 = src.x * src.x */
1829 tmp0
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
1830 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
1833 tmp0
= lp_build_mul( &bld
->bld_base
.base
, tmp0
, tmp0
);
1836 /* xmm0 = xmm0 + src.y * src.y */
1837 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Y
);
1838 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
1841 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1842 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1845 /* xmm0 = xmm0 + src.z * src.z */
1846 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Z
);
1847 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
1850 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1851 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1855 /* xmm0 = xmm0 + src.w * src.w */
1856 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
1857 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
)) {
1860 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1861 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1863 /* xmm1 = 1 / sqrt(xmm0) */
1864 tmp1
= lp_build_rsqrt( &bld
->bld_base
.base
, tmp0
);
1865 /* dst.x = xmm1 * src.x */
1866 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
1867 emit_data
->output
[TGSI_CHAN_X
] = lp_build_mul( &bld
->bld_base
.base
, tmp4
, tmp1
);
1869 /* dst.y = xmm1 * src.y */
1870 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
1871 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_mul( &bld
->bld_base
.base
, tmp5
, tmp1
);
1874 /* dst.z = xmm1 * src.z */
1875 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
1876 emit_data
->output
[TGSI_CHAN_Z
] = lp_build_mul( &bld
->bld_base
.base
, tmp6
, tmp1
);
1878 /* dst.w = xmm1 * src.w */
1879 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) && dims
== 4) {
1880 emit_data
->output
[TGSI_CHAN_W
] = lp_build_mul( &bld
->bld_base
.base
, tmp7
, tmp1
);
1885 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 3) {
1886 emit_data
->output
[TGSI_CHAN_W
] = bld
->bld_base
.base
.one
;
1890 static void emit_prologue(struct lp_build_tgsi_context
* bld_base
)
1892 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1893 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
1895 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
1896 LLVMValueRef array_size
=
1897 lp_build_const_int32(gallivm
,
1898 bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] * 4 + 4);
1899 bld
->temps_array
= lp_build_array_alloca(gallivm
,
1900 bld_base
->base
.vec_type
, array_size
,
1904 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
1905 LLVMValueRef array_size
=
1906 lp_build_const_int32(gallivm
,
1907 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] * 4 + 4);
1908 bld
->outputs_array
= lp_build_array_alloca(gallivm
,
1909 bld_base
->base
.vec_type
, array_size
,
1913 /* If we have indirect addressing in inputs we need to copy them into
1914 * our alloca array to be able to iterate over them */
1915 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
1916 unsigned index
, chan
;
1917 LLVMTypeRef vec_type
= bld_base
->base
.vec_type
;
1918 LLVMValueRef array_size
= lp_build_const_int32(gallivm
,
1919 bld_base
->info
->file_max
[TGSI_FILE_INPUT
]*4 + 4);
1920 bld
->inputs_array
= lp_build_array_alloca(gallivm
,
1921 vec_type
, array_size
,
1924 assert(bld_base
->info
->num_inputs
1925 <= bld_base
->info
->file_max
[TGSI_FILE_INPUT
] + 1);
1927 for (index
= 0; index
< bld_base
->info
->num_inputs
; ++index
) {
1928 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1929 LLVMValueRef lindex
=
1930 lp_build_const_int32(gallivm
, index
* 4 + chan
);
1931 LLVMValueRef input_ptr
=
1932 LLVMBuildGEP(gallivm
->builder
, bld
->inputs_array
,
1934 LLVMValueRef value
= bld
->inputs
[index
][chan
];
1936 LLVMBuildStore(gallivm
->builder
, value
, input_ptr
);
1942 static void emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
1944 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1948 emit_dump_temps(bld
);
1951 /* If we have indirect addressing in outputs we need to copy our alloca array
1952 * to the outputs slots specified by the called */
1953 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
1954 unsigned index
, chan
;
1955 assert(bld_base
->info
->num_outputs
<=
1956 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] + 1);
1957 for (index
= 0; index
< bld_base
->info
->num_outputs
; ++index
) {
1958 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1959 bld
->outputs
[index
][chan
] = lp_get_output_ptr(bld
, index
, chan
);
1966 lp_build_tgsi_soa(struct gallivm_state
*gallivm
,
1967 const struct tgsi_token
*tokens
,
1968 struct lp_type type
,
1969 struct lp_build_mask_context
*mask
,
1970 LLVMValueRef consts_ptr
,
1971 LLVMValueRef system_values_array
,
1972 const LLVMValueRef
*pos
,
1973 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
1974 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1975 struct lp_build_sampler_soa
*sampler
,
1976 const struct tgsi_shader_info
*info
)
1978 struct lp_build_tgsi_soa_context bld
;
1980 struct lp_type res_type
;
1982 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
1983 memset(&res_type
, 0, sizeof res_type
);
1984 res_type
.width
= type
.width
;
1985 res_type
.length
= type
.length
;
1988 /* Setup build context */
1989 memset(&bld
, 0, sizeof bld
);
1990 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
1991 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
1992 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
1993 lp_build_context_init(&bld
.elem_bld
, gallivm
, lp_elem_type(type
));
1996 bld
.inputs
= inputs
;
1997 bld
.outputs
= outputs
;
1998 bld
.consts_ptr
= consts_ptr
;
1999 bld
.sampler
= sampler
;
2000 bld
.bld_base
.info
= info
;
2001 bld
.indirect_files
= info
->indirect_files
;
2003 bld
.bld_base
.soa
= TRUE
;
2004 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
2005 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
2006 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
2007 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
2008 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = emit_fetch_system_value
;
2009 bld
.bld_base
.emit_store
= emit_store
;
2011 bld
.bld_base
.emit_declaration
= lp_emit_declaration_soa
;
2012 bld
.bld_base
.emit_immediate
= lp_emit_immediate_soa
;
2014 bld
.bld_base
.emit_prologue
= emit_prologue
;
2015 bld
.bld_base
.emit_epilogue
= emit_epilogue
;
2017 /* Set opcode actions */
2018 lp_set_default_actions_cpu(&bld
.bld_base
);
2020 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
2021 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNSUB
].emit
= bgnsub_emit
;
2022 bld
.bld_base
.op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
2023 bld
.bld_base
.op_actions
[TGSI_OPCODE_CAL
].emit
= cal_emit
;
2024 bld
.bld_base
.op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
2025 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDX
].emit
= ddx_emit
;
2026 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDY
].emit
= ddy_emit
;
2027 bld
.bld_base
.op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
2028 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
2029 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
2030 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSUB
].emit
= endsub_emit
;
2031 bld
.bld_base
.op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
2032 bld
.bld_base
.op_actions
[TGSI_OPCODE_KIL
].emit
= kil_emit
;
2033 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILP
].emit
= kilp_emit
;
2034 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM
].emit
= nrm_emit
;
2035 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM4
].emit
= nrm_emit
;
2036 bld
.bld_base
.op_actions
[TGSI_OPCODE_RET
].emit
= ret_emit
;
2037 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX
].emit
= tex_emit
;
2038 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB
].emit
= txb_emit
;
2039 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXD
].emit
= txd_emit
;
2040 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL
].emit
= txl_emit
;
2041 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXP
].emit
= txp_emit
;
2042 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXQ
].emit
= txq_emit
;
2044 lp_exec_mask_init(&bld
.exec_mask
, &bld
.bld_base
.base
);
2047 bld
.system_values_array
= system_values_array
;
2049 lp_build_tgsi_llvm(&bld
.bld_base
, tokens
);
2052 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
2053 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
2054 debug_printf("11111111111111111111111111111 \n");
2055 tgsi_dump(tokens
, 0);
2056 lp_debug_dump_value(function
);
2057 debug_printf("2222222222222222222222222222 \n");
2061 LLVMModuleRef module
= LLVMGetGlobalParent(
2062 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
2063 LLVMDumpModule(module
);
2070 * Build up the system values array out of individual values such as
2071 * the instance ID, front-face, primitive ID, etc. The shader info is
2072 * used to determine which system values are needed and where to put
2073 * them in the system values array.
2075 * XXX only instance ID is implemented at this time.
2077 * The system values register file is similar to the constants buffer.
2078 * Example declaration:
2079 * DCL SV[0], INSTANCEID
2080 * Example instruction:
2081 * MOVE foo, SV[0].xxxx;
2083 * \return LLVM float array (interpreted as float [][4])
2086 lp_build_system_values_array(struct gallivm_state
*gallivm
,
2087 const struct tgsi_shader_info
*info
,
2088 LLVMValueRef instance_id
,
2089 LLVMValueRef facing
)
2091 LLVMValueRef size
= lp_build_const_int32(gallivm
, 4 * info
->num_system_values
);
2092 LLVMTypeRef float_t
= LLVMFloatTypeInContext(gallivm
->context
);
2093 LLVMValueRef array
= lp_build_array_alloca(gallivm
, float_t
,
2094 size
, "sysvals_array");
2097 for (i
= 0; i
< info
->num_system_values
; i
++) {
2098 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
* 4);
2099 LLVMValueRef ptr
, value
= 0;
2101 switch (info
->system_value_semantic_name
[i
]) {
2102 case TGSI_SEMANTIC_INSTANCEID
:
2103 /* convert instance ID from int to float */
2104 value
= LLVMBuildSIToFP(gallivm
->builder
, instance_id
, float_t
,
2105 "sysval_instanceid");
2107 case TGSI_SEMANTIC_FACE
:
2110 assert(0 && "unexpected semantic in build_system_values_array()");
2113 ptr
= LLVMBuildGEP(gallivm
->builder
, array
, &index
, 1, "");
2114 LLVMBuildStore(gallivm
->builder
, value
, ptr
);