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
];
710 if (stype
== TGSI_TYPE_UNSIGNED
) {
711 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
712 } else if (stype
== TGSI_TYPE_SIGNED
) {
713 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
720 emit_fetch_temporary(
721 struct lp_build_tgsi_context
* bld_base
,
722 const struct tgsi_full_src_register
* reg
,
723 enum tgsi_opcode_type stype
,
726 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
727 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
728 LLVMBuilderRef builder
= gallivm
->builder
;
729 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
730 LLVMValueRef indirect_index
= NULL
;
733 if (reg
->Register
.Indirect
) {
734 indirect_index
= get_indirect_index(bld
,
740 if (reg
->Register
.Indirect
) {
741 LLVMValueRef swizzle_vec
=
742 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, swizzle
);
743 LLVMValueRef length_vec
=
744 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
,
745 bld
->bld_base
.base
.type
.length
);
746 LLVMValueRef index_vec
; /* index into the const buffer */
747 LLVMValueRef temps_array
;
748 LLVMTypeRef float4_ptr_type
;
750 /* index_vec = (indirect_index * 4 + swizzle) * length */
751 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
752 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
753 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
755 /* cast temps_array pointer to float* */
756 float4_ptr_type
= LLVMPointerType(LLVMFloatTypeInContext(bld
->bld_base
.base
.gallivm
->context
), 0);
757 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
,
758 float4_ptr_type
, "");
760 /* Gather values from the temporary register array */
761 res
= build_gather(&bld_base
->base
, temps_array
, index_vec
);
764 LLVMValueRef temp_ptr
;
765 if (stype
!= TGSI_TYPE_FLOAT
&& stype
!= TGSI_TYPE_UNTYPED
) {
766 LLVMTypeRef itype
= LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4), 0);
767 LLVMValueRef tint_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
,
769 temp_ptr
= LLVMBuildBitCast(builder
, tint_ptr
, itype
, "");
771 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
772 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
774 return bld
->bld_base
.base
.undef
;
781 emit_fetch_system_value(
782 struct lp_build_tgsi_context
* bld_base
,
783 const struct tgsi_full_src_register
* reg
,
784 enum tgsi_opcode_type stype
,
787 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
788 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
789 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
790 LLVMBuilderRef builder
= gallivm
->builder
;
792 enum tgsi_opcode_type atype
; // Actual type of the value
794 assert(!reg
->Register
.Indirect
);
796 switch (info
->system_value_semantic_name
[reg
->Register
.Index
]) {
797 case TGSI_SEMANTIC_INSTANCEID
:
798 res
= lp_build_broadcast_scalar(&bld_base
->uint_bld
, bld
->system_values
.instance_id
);
799 atype
= TGSI_TYPE_UNSIGNED
;
802 case TGSI_SEMANTIC_VERTEXID
:
803 res
= bld
->system_values
.vertex_id
;
804 atype
= TGSI_TYPE_UNSIGNED
;
808 assert(!"unexpected semantic in emit_fetch_system_value");
809 res
= bld_base
->base
.zero
;
810 atype
= TGSI_TYPE_FLOAT
;
814 if (atype
!= stype
) {
815 if (stype
== TGSI_TYPE_FLOAT
) {
816 res
= LLVMBuildBitCast(builder
, res
, bld_base
->base
.vec_type
, "");
817 } else if (stype
== TGSI_TYPE_UNSIGNED
) {
818 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
819 } else if (stype
== TGSI_TYPE_SIGNED
) {
820 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
828 * Register fetch with derivatives.
832 struct lp_build_tgsi_soa_context
*bld
,
841 /* TODO: use interpolation coeffs for inputs */
844 *ddx
= lp_build_ddx(&bld
->bld_base
.base
, src
);
847 *ddy
= lp_build_ddy(&bld
->bld_base
.base
, src
);
855 emit_fetch_predicate(
856 struct lp_build_tgsi_soa_context
*bld
,
857 const struct tgsi_full_instruction
*inst
,
860 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
862 unsigned char swizzles
[4];
863 LLVMValueRef unswizzled
[4] = {NULL
, NULL
, NULL
, NULL
};
867 if (!inst
->Instruction
.Predicate
) {
868 TGSI_FOR_EACH_CHANNEL( chan
) {
874 swizzles
[0] = inst
->Predicate
.SwizzleX
;
875 swizzles
[1] = inst
->Predicate
.SwizzleY
;
876 swizzles
[2] = inst
->Predicate
.SwizzleZ
;
877 swizzles
[3] = inst
->Predicate
.SwizzleW
;
879 index
= inst
->Predicate
.Index
;
880 assert(index
< LP_MAX_TGSI_PREDS
);
882 TGSI_FOR_EACH_CHANNEL( chan
) {
883 unsigned swizzle
= swizzles
[chan
];
886 * Only fetch the predicate register channels that are actually listed
889 if (!unswizzled
[swizzle
]) {
890 value
= LLVMBuildLoad(builder
,
891 bld
->preds
[index
][swizzle
], "");
894 * Convert the value to an integer mask.
896 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
897 * is needlessly causing two comparisons due to storing the intermediate
898 * result as float vector instead of an integer mask vector.
900 value
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
901 bld
->bld_base
.base
.type
,
904 bld
->bld_base
.base
.zero
);
905 if (inst
->Predicate
.Negate
) {
906 value
= LLVMBuildNot(builder
, value
, "");
909 unswizzled
[swizzle
] = value
;
911 value
= unswizzled
[swizzle
];
923 struct lp_build_tgsi_context
*bld_base
,
924 const struct tgsi_full_instruction
*inst
,
930 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
931 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
932 LLVMBuilderRef builder
= gallivm
->builder
;
933 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
934 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
935 LLVMValueRef indirect_index
= NULL
;
936 struct lp_build_context
*bld_store
;
937 enum tgsi_opcode_type dtype
= tgsi_opcode_infer_dst_type(inst
->Instruction
.Opcode
);
941 case TGSI_TYPE_FLOAT
:
942 case TGSI_TYPE_UNTYPED
:
943 bld_store
= &bld_base
->base
;
945 case TGSI_TYPE_UNSIGNED
:
946 bld_store
= &bld_base
->uint_bld
;
948 case TGSI_TYPE_SIGNED
:
949 bld_store
= &bld_base
->int_bld
;
951 case TGSI_TYPE_DOUBLE
:
958 switch( inst
->Instruction
.Saturate
) {
962 case TGSI_SAT_ZERO_ONE
:
963 value
= lp_build_max(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.zero
);
964 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
967 case TGSI_SAT_MINUS_PLUS_ONE
:
968 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));
969 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
976 if (reg
->Register
.Indirect
) {
977 indirect_index
= get_indirect_index(bld
,
982 assert(reg
->Register
.Index
<=
983 bld
->bld_base
.info
->file_max
[reg
->Register
.File
]);
986 switch( reg
->Register
.File
) {
987 case TGSI_FILE_OUTPUT
:
988 if (reg
->Register
.Indirect
) {
989 LLVMValueRef chan_vec
=
990 lp_build_const_int_vec(gallivm
, uint_bld
->type
, chan_index
);
991 LLVMValueRef length_vec
=
992 lp_build_const_int_vec(gallivm
, uint_bld
->type
, bld
->bld_base
.base
.type
.length
);
993 LLVMValueRef index_vec
; /* indexes into the temp registers */
994 LLVMValueRef outputs_array
;
995 LLVMValueRef pixel_offsets
;
996 LLVMTypeRef float_ptr_type
;
999 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1000 pixel_offsets
= uint_bld
->undef
;
1001 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
1002 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1003 pixel_offsets
= LLVMBuildInsertElement(builder
, pixel_offsets
,
1007 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1008 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1009 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
1010 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1011 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1014 LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1015 outputs_array
= LLVMBuildBitCast(builder
, bld
->outputs_array
,
1016 float_ptr_type
, "");
1018 /* Scatter store values into temp registers */
1019 emit_mask_scatter(bld
, outputs_array
, index_vec
, value
,
1020 &bld
->exec_mask
, pred
);
1023 LLVMValueRef out_ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
,
1025 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
, out_ptr
);
1029 case TGSI_FILE_TEMPORARY
:
1030 if (reg
->Register
.Indirect
) {
1031 LLVMValueRef chan_vec
=
1032 lp_build_const_int_vec(gallivm
, uint_bld
->type
, chan_index
);
1033 LLVMValueRef length_vec
=
1034 lp_build_const_int_vec(gallivm
, uint_bld
->type
,
1035 bld
->bld_base
.base
.type
.length
);
1036 LLVMValueRef index_vec
; /* indexes into the temp registers */
1037 LLVMValueRef temps_array
;
1038 LLVMValueRef pixel_offsets
;
1039 LLVMTypeRef float_ptr_type
;
1042 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1043 pixel_offsets
= uint_bld
->undef
;
1044 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
1045 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1046 pixel_offsets
= LLVMBuildInsertElement(builder
, pixel_offsets
,
1050 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1051 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1052 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
1053 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1054 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1057 LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1058 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
,
1059 float_ptr_type
, "");
1061 /* Scatter store values into temp registers */
1062 emit_mask_scatter(bld
, temps_array
, index_vec
, value
,
1063 &bld
->exec_mask
, pred
);
1066 LLVMValueRef temp_ptr
;
1069 case TGSI_TYPE_UNSIGNED
:
1070 case TGSI_TYPE_SIGNED
: {
1071 LLVMTypeRef itype
= LLVMVectorType(LLVMInt32TypeInContext(gallivm
->context
), 4);
1072 LLVMTypeRef ivtype
= LLVMPointerType(itype
, 0);
1073 LLVMValueRef tint_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
,
1075 LLVMValueRef temp_value_ptr
;
1077 temp_ptr
= LLVMBuildBitCast(builder
, tint_ptr
, ivtype
, "");
1078 temp_value_ptr
= LLVMBuildBitCast(builder
, value
, itype
, "");
1079 value
= temp_value_ptr
;
1083 case TGSI_TYPE_FLOAT
:
1084 case TGSI_TYPE_UNTYPED
:
1085 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
,
1090 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
, temp_ptr
);
1094 case TGSI_FILE_ADDRESS
:
1095 assert(dtype
== TGSI_TYPE_SIGNED
);
1096 assert(LLVMTypeOf(value
) == bld_base
->base
.int_vec_type
);
1097 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
,
1098 bld
->addr
[reg
->Register
.Index
][chan_index
]);
1101 case TGSI_FILE_PREDICATE
:
1102 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
,
1103 bld
->preds
[reg
->Register
.Index
][chan_index
]);
1113 struct lp_build_tgsi_context
* bld_base
,
1114 const struct tgsi_full_instruction
* inst
,
1115 const struct tgsi_opcode_info
* info
,
1116 LLVMValueRef dst
[4])
1119 unsigned chan_index
;
1120 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1123 LLVMValueRef pred
[TGSI_NUM_CHANNELS
];
1125 emit_fetch_predicate( bld
, inst
, pred
);
1127 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1128 emit_store_chan(bld_base
, inst
, 0, chan_index
, pred
[chan_index
], dst
[chan_index
]);
1134 * High-level instruction translators.
1138 emit_tex( struct lp_build_tgsi_soa_context
*bld
,
1139 const struct tgsi_full_instruction
*inst
,
1140 enum lp_build_tex_modifier modifier
,
1141 LLVMValueRef
*texel
)
1143 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1145 LLVMValueRef lod_bias
, explicit_lod
;
1146 LLVMValueRef oow
= NULL
;
1147 LLVMValueRef coords
[3];
1148 LLVMValueRef ddx
[3];
1149 LLVMValueRef ddy
[3];
1150 unsigned num_coords
;
1153 if (!bld
->sampler
) {
1154 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1155 for (i
= 0; i
< 4; i
++) {
1156 texel
[i
] = bld
->bld_base
.base
.undef
;
1161 switch (inst
->Texture
.Texture
) {
1162 case TGSI_TEXTURE_1D
:
1165 case TGSI_TEXTURE_1D_ARRAY
:
1166 case TGSI_TEXTURE_2D
:
1167 case TGSI_TEXTURE_RECT
:
1170 case TGSI_TEXTURE_SHADOW1D
:
1171 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1172 case TGSI_TEXTURE_SHADOW2D
:
1173 case TGSI_TEXTURE_SHADOWRECT
:
1174 case TGSI_TEXTURE_2D_ARRAY
:
1175 case TGSI_TEXTURE_3D
:
1176 case TGSI_TEXTURE_CUBE
:
1179 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1187 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
1188 lod_bias
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1189 explicit_lod
= NULL
;
1191 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
1193 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1197 explicit_lod
= NULL
;
1200 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
) {
1201 oow
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1202 oow
= lp_build_rcp(&bld
->bld_base
.base
, oow
);
1205 for (i
= 0; i
< num_coords
; i
++) {
1206 coords
[i
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, i
);
1207 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
1208 coords
[i
] = lp_build_mul(&bld
->bld_base
.base
, coords
[i
], oow
);
1210 for (i
= num_coords
; i
< 3; i
++) {
1211 coords
[i
] = bld
->bld_base
.base
.undef
;
1214 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
1215 LLVMValueRef index0
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
, 0);
1216 for (i
= 0; i
< num_coords
; i
++) {
1217 LLVMValueRef src1
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 1, i
);
1218 LLVMValueRef src2
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 2, i
);
1219 ddx
[i
] = LLVMBuildExtractElement(builder
, src1
, index0
, "");
1220 ddy
[i
] = LLVMBuildExtractElement(builder
, src2
, index0
, "");
1222 unit
= inst
->Src
[3].Register
.Index
;
1224 for (i
= 0; i
< num_coords
; i
++) {
1225 ddx
[i
] = lp_build_scalar_ddx( &bld
->bld_base
.base
, coords
[i
] );
1226 ddy
[i
] = lp_build_scalar_ddy( &bld
->bld_base
.base
, coords
[i
] );
1228 unit
= inst
->Src
[1].Register
.Index
;
1230 for (i
= num_coords
; i
< 3; i
++) {
1231 ddx
[i
] = LLVMGetUndef(bld
->bld_base
.base
.elem_type
);
1232 ddy
[i
] = LLVMGetUndef(bld
->bld_base
.base
.elem_type
);
1235 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
1236 bld
->bld_base
.base
.gallivm
,
1237 bld
->bld_base
.base
.type
,
1238 unit
, num_coords
, coords
,
1240 lod_bias
, explicit_lod
,
1245 emit_txq( struct lp_build_tgsi_soa_context
*bld
,
1246 const struct tgsi_full_instruction
*inst
,
1247 LLVMValueRef
*sizes_out
)
1249 LLVMValueRef explicit_lod
;
1250 unsigned num_coords
, has_lod
;
1253 switch (inst
->Texture
.Texture
) {
1254 case TGSI_TEXTURE_1D
:
1255 case TGSI_TEXTURE_SHADOW1D
:
1256 case TGSI_TEXTURE_SHADOW2D
:
1257 case TGSI_TEXTURE_SHADOWCUBE
:
1261 case TGSI_TEXTURE_2D
:
1262 case TGSI_TEXTURE_CUBE
:
1263 case TGSI_TEXTURE_1D_ARRAY
:
1264 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1268 case TGSI_TEXTURE_3D
:
1269 // case TGSI_TEXTURE_CUBE_ARRAY:
1270 // case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
1271 case TGSI_TEXTURE_2D_ARRAY
:
1272 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1277 case TGSI_TEXTURE_BUFFER
:
1282 case TGSI_TEXTURE_RECT
:
1283 case TGSI_TEXTURE_SHADOWRECT
:
1284 // case TGSI_TEXTURE_2D_MS:
1289 // case TGSI_TEXTURE_2D_MS_ARRAY:
1299 if (!bld
->sampler
) {
1300 _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
1301 for (i
= 0; i
< num_coords
; i
++)
1302 sizes_out
[i
] = bld
->bld_base
.base
.undef
;
1307 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 2 );
1309 explicit_lod
= NULL
;
1311 bld
->sampler
->emit_size_query(bld
->sampler
,
1312 bld
->bld_base
.base
.gallivm
,
1313 inst
->Src
[1].Register
.Index
,
1319 near_end_of_shader(struct lp_build_tgsi_soa_context
*bld
,
1324 for (i
= 0; i
< 5; i
++) {
1327 if (pc
+ i
>= bld
->bld_base
.info
->num_instructions
)
1330 opcode
= bld
->bld_base
.instructions
[pc
+ i
].Instruction
.Opcode
;
1332 if (opcode
== TGSI_OPCODE_END
)
1335 if (opcode
== TGSI_OPCODE_TEX
||
1336 opcode
== TGSI_OPCODE_TXP
||
1337 opcode
== TGSI_OPCODE_TXD
||
1338 opcode
== TGSI_OPCODE_TXB
||
1339 opcode
== TGSI_OPCODE_TXL
||
1340 opcode
== TGSI_OPCODE_TXF
||
1341 opcode
== TGSI_OPCODE_TXQ
||
1342 opcode
== TGSI_OPCODE_CAL
||
1343 opcode
== TGSI_OPCODE_CALLNZ
||
1344 opcode
== TGSI_OPCODE_IF
||
1345 opcode
== TGSI_OPCODE_IFC
||
1346 opcode
== TGSI_OPCODE_BGNLOOP
||
1347 opcode
== TGSI_OPCODE_SWITCH
)
1357 * Kill fragment if any of the src register values are negative.
1361 struct lp_build_tgsi_soa_context
*bld
,
1362 const struct tgsi_full_instruction
*inst
,
1365 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1366 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
1367 LLVMValueRef terms
[TGSI_NUM_CHANNELS
];
1369 unsigned chan_index
;
1371 memset(&terms
, 0, sizeof terms
);
1373 TGSI_FOR_EACH_CHANNEL( chan_index
) {
1376 /* Unswizzle channel */
1377 swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
1379 /* Check if the component has not been already tested. */
1380 assert(swizzle
< TGSI_NUM_CHANNELS
);
1381 if( !terms
[swizzle
] )
1382 /* TODO: change the comparison operator instead of setting the sign */
1383 terms
[swizzle
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, chan_index
);
1387 TGSI_FOR_EACH_CHANNEL( chan_index
) {
1388 if(terms
[chan_index
]) {
1389 LLVMValueRef chan_mask
;
1392 * If term < 0 then mask = 0 else mask = ~0.
1394 chan_mask
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, terms
[chan_index
], bld
->bld_base
.base
.zero
);
1397 mask
= LLVMBuildAnd(builder
, mask
, chan_mask
, "");
1404 lp_build_mask_update(bld
->mask
, mask
);
1406 if (!near_end_of_shader(bld
, pc
))
1407 lp_build_mask_check(bld
->mask
);
1413 * Predicated fragment kill.
1414 * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1415 * The only predication is the execution mask which will apply if
1416 * we're inside a loop or conditional.
1419 emit_kilp(struct lp_build_tgsi_soa_context
*bld
,
1422 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1425 /* For those channels which are "alive", disable fragment shader
1428 if (bld
->exec_mask
.has_mask
) {
1429 mask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
1432 LLVMValueRef zero
= LLVMConstNull(bld
->bld_base
.base
.int_vec_type
);
1436 lp_build_mask_update(bld
->mask
, mask
);
1438 if (!near_end_of_shader(bld
, pc
))
1439 lp_build_mask_check(bld
->mask
);
1444 * Emit code which will dump the value of all the temporary registers
1448 emit_dump_temps(struct lp_build_tgsi_soa_context
*bld
)
1450 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1451 LLVMBuilderRef builder
= gallivm
->builder
;
1452 LLVMValueRef temp_ptr
;
1453 LLVMValueRef i0
= lp_build_const_int32(gallivm
, 0);
1454 LLVMValueRef i1
= lp_build_const_int32(gallivm
, 1);
1455 LLVMValueRef i2
= lp_build_const_int32(gallivm
, 2);
1456 LLVMValueRef i3
= lp_build_const_int32(gallivm
, 3);
1458 int n
= bld
->bld_base
.info
->file_max
[TGSI_FILE_TEMPORARY
];
1460 for (index
= 0; index
< n
; index
++) {
1461 LLVMValueRef idx
= lp_build_const_int32(gallivm
, index
);
1462 LLVMValueRef v
[4][4], res
;
1465 lp_build_printf(gallivm
, "TEMP[%d]:\n", idx
);
1467 for (chan
= 0; chan
< 4; chan
++) {
1468 temp_ptr
= lp_get_temp_ptr_soa(bld
, index
, chan
);
1469 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
1470 v
[chan
][0] = LLVMBuildExtractElement(builder
, res
, i0
, "");
1471 v
[chan
][1] = LLVMBuildExtractElement(builder
, res
, i1
, "");
1472 v
[chan
][2] = LLVMBuildExtractElement(builder
, res
, i2
, "");
1473 v
[chan
][3] = LLVMBuildExtractElement(builder
, res
, i3
, "");
1476 lp_build_printf(gallivm
, " X: %f %f %f %f\n",
1477 v
[0][0], v
[0][1], v
[0][2], v
[0][3]);
1478 lp_build_printf(gallivm
, " Y: %f %f %f %f\n",
1479 v
[1][0], v
[1][1], v
[1][2], v
[1][3]);
1480 lp_build_printf(gallivm
, " Z: %f %f %f %f\n",
1481 v
[2][0], v
[2][1], v
[2][2], v
[2][3]);
1482 lp_build_printf(gallivm
, " W: %f %f %f %f\n",
1483 v
[3][0], v
[3][1], v
[3][2], v
[3][3]);
1490 lp_emit_declaration_soa(
1491 struct lp_build_tgsi_context
*bld_base
,
1492 const struct tgsi_full_declaration
*decl
)
1494 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
1495 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1496 LLVMTypeRef vec_type
= bld
->bld_base
.base
.vec_type
;
1497 const unsigned first
= decl
->Range
.First
;
1498 const unsigned last
= decl
->Range
.Last
;
1501 for (idx
= first
; idx
<= last
; ++idx
) {
1502 assert(last
<= bld
->bld_base
.info
->file_max
[decl
->Declaration
.File
]);
1503 switch (decl
->Declaration
.File
) {
1504 case TGSI_FILE_TEMPORARY
:
1505 assert(idx
< LP_MAX_TGSI_TEMPS
);
1506 if (!(bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
))) {
1507 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1508 bld
->temps
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
, "temp");
1512 case TGSI_FILE_OUTPUT
:
1513 if (!(bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
1514 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1515 bld
->outputs
[idx
][i
] = lp_build_alloca(gallivm
,
1516 vec_type
, "output");
1520 case TGSI_FILE_ADDRESS
:
1521 /* ADDR registers are the only allocated with an integer LLVM IR type,
1522 * as they are guaranteed to always have integers.
1523 * XXX: Not sure if this exception is worthwhile (or the whole idea of
1524 * an ADDR register for that matter).
1526 assert(idx
< LP_MAX_TGSI_ADDRS
);
1527 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1528 bld
->addr
[idx
][i
] = lp_build_alloca(gallivm
, bld_base
->base
.int_vec_type
, "addr");
1531 case TGSI_FILE_PREDICATE
:
1532 assert(idx
< LP_MAX_TGSI_PREDS
);
1533 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1534 bld
->preds
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
,
1539 /* don't need to declare other vars */
1546 void lp_emit_immediate_soa(
1547 struct lp_build_tgsi_context
*bld_base
,
1548 const struct tgsi_full_immediate
*imm
)
1550 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
1551 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
1553 /* simply copy the immediate values into the next immediates[] slot */
1555 const uint size
= imm
->Immediate
.NrTokens
- 1;
1557 assert(bld
->num_immediates
< LP_MAX_TGSI_IMMEDIATES
);
1558 switch (imm
->Immediate
.DataType
) {
1559 case TGSI_IMM_FLOAT32
:
1560 for( i
= 0; i
< size
; ++i
)
1561 bld
->immediates
[bld
->num_immediates
][i
] =
1562 lp_build_const_vec(gallivm
, bld_base
->base
.type
, imm
->u
[i
].Float
);
1565 case TGSI_IMM_UINT32
:
1566 for( i
= 0; i
< size
; ++i
) {
1567 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->uint_bld
.type
, imm
->u
[i
].Uint
);
1568 bld
->immediates
[bld
->num_immediates
][i
] =
1569 LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
1573 case TGSI_IMM_INT32
:
1574 for( i
= 0; i
< size
; ++i
) {
1575 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->int_bld
.type
, imm
->u
[i
].Int
);
1576 bld
->immediates
[bld
->num_immediates
][i
] =
1577 LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
1582 for( i
= size
; i
< 4; ++i
)
1583 bld
->immediates
[bld
->num_immediates
][i
] = bld_base
->base
.undef
;
1585 bld
->num_immediates
++;
1590 const struct lp_build_tgsi_action
* action
,
1591 struct lp_build_tgsi_context
* bld_base
,
1592 struct lp_build_emit_data
* emit_data
)
1594 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1596 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
,
1597 &emit_data
->output
[emit_data
->chan
], NULL
);
1602 const struct lp_build_tgsi_action
* action
,
1603 struct lp_build_tgsi_context
* bld_base
,
1604 struct lp_build_emit_data
* emit_data
)
1606 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1608 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
, NULL
,
1609 &emit_data
->output
[emit_data
->chan
]);
1614 const struct lp_build_tgsi_action
* action
,
1615 struct lp_build_tgsi_context
* bld_base
,
1616 struct lp_build_emit_data
* emit_data
)
1618 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1620 emit_kilp(bld
, bld_base
->pc
- 1);
1625 const struct lp_build_tgsi_action
* action
,
1626 struct lp_build_tgsi_context
* bld_base
,
1627 struct lp_build_emit_data
* emit_data
)
1629 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1631 emit_kil(bld
, emit_data
->inst
, bld_base
->pc
- 1);
1636 const struct lp_build_tgsi_action
* action
,
1637 struct lp_build_tgsi_context
* bld_base
,
1638 struct lp_build_emit_data
* emit_data
)
1640 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1642 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
, emit_data
->output
);
1647 const struct lp_build_tgsi_action
* action
,
1648 struct lp_build_tgsi_context
* bld_base
,
1649 struct lp_build_emit_data
* emit_data
)
1651 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1653 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
1659 const struct lp_build_tgsi_action
* action
,
1660 struct lp_build_tgsi_context
* bld_base
,
1661 struct lp_build_emit_data
* emit_data
)
1663 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1665 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
1671 const struct lp_build_tgsi_action
* action
,
1672 struct lp_build_tgsi_context
* bld_base
,
1673 struct lp_build_emit_data
* emit_data
)
1675 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1677 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
1683 const struct lp_build_tgsi_action
* action
,
1684 struct lp_build_tgsi_context
* bld_base
,
1685 struct lp_build_emit_data
* emit_data
)
1687 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1689 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_PROJECTED
,
1695 const struct lp_build_tgsi_action
* action
,
1696 struct lp_build_tgsi_context
* bld_base
,
1697 struct lp_build_emit_data
* emit_data
)
1699 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1701 emit_txq(bld
, emit_data
->inst
, emit_data
->output
);
1706 const struct lp_build_tgsi_action
* action
,
1707 struct lp_build_tgsi_context
* bld_base
,
1708 struct lp_build_emit_data
* emit_data
)
1710 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1712 lp_exec_mask_call(&bld
->exec_mask
, emit_data
->inst
->Label
.Label
,
1718 const struct lp_build_tgsi_action
* action
,
1719 struct lp_build_tgsi_context
* bld_base
,
1720 struct lp_build_emit_data
* emit_data
)
1722 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1724 lp_exec_mask_ret(&bld
->exec_mask
, &bld_base
->pc
);
1729 const struct lp_build_tgsi_action
* action
,
1730 struct lp_build_tgsi_context
* bld_base
,
1731 struct lp_build_emit_data
* emit_data
)
1733 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1735 lp_exec_break(&bld
->exec_mask
);
1740 const struct lp_build_tgsi_action
* action
,
1741 struct lp_build_tgsi_context
* bld_base
,
1742 struct lp_build_emit_data
* emit_data
)
1745 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1747 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_NOTEQUAL
,
1748 emit_data
->args
[0], bld
->bld_base
.base
.zero
);
1749 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
1754 const struct lp_build_tgsi_action
* action
,
1755 struct lp_build_tgsi_context
* bld_base
,
1756 struct lp_build_emit_data
* emit_data
)
1758 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1760 lp_exec_bgnloop(&bld
->exec_mask
);
1765 const struct lp_build_tgsi_action
* action
,
1766 struct lp_build_tgsi_context
* bld_base
,
1767 struct lp_build_emit_data
* emit_data
)
1769 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1771 lp_exec_mask_bgnsub(&bld
->exec_mask
);
1776 const struct lp_build_tgsi_action
* action
,
1777 struct lp_build_tgsi_context
* bld_base
,
1778 struct lp_build_emit_data
* emit_data
)
1780 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1782 lp_exec_mask_cond_invert(&bld
->exec_mask
);
1787 const struct lp_build_tgsi_action
* action
,
1788 struct lp_build_tgsi_context
* bld_base
,
1789 struct lp_build_emit_data
* emit_data
)
1791 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1793 lp_exec_mask_cond_pop(&bld
->exec_mask
);
1798 const struct lp_build_tgsi_action
* action
,
1799 struct lp_build_tgsi_context
* bld_base
,
1800 struct lp_build_emit_data
* emit_data
)
1802 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1804 lp_exec_endloop(bld_base
->base
.gallivm
, &bld
->exec_mask
);
1809 const struct lp_build_tgsi_action
* action
,
1810 struct lp_build_tgsi_context
* bld_base
,
1811 struct lp_build_emit_data
* emit_data
)
1813 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1815 lp_exec_mask_endsub(&bld
->exec_mask
, &bld_base
->pc
);
1820 const struct lp_build_tgsi_action
* action
,
1821 struct lp_build_tgsi_context
* bld_base
,
1822 struct lp_build_emit_data
* emit_data
)
1824 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1826 lp_exec_continue(&bld
->exec_mask
);
1829 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1831 * XXX: What do the comments about xmm registers mean? Maybe they are left over
1832 * from old code, but there is no garauntee that LLVM will use those registers
1835 * XXX: There should be no calls to lp_build_emit_fetch in this function. This
1836 * should be handled by the emit_data->fetch_args function. */
1839 const struct lp_build_tgsi_action
* action
,
1840 struct lp_build_tgsi_context
* bld_base
,
1841 struct lp_build_emit_data
* emit_data
)
1843 LLVMValueRef tmp0
, tmp1
;
1844 LLVMValueRef tmp4
= NULL
;
1845 LLVMValueRef tmp5
= NULL
;
1846 LLVMValueRef tmp6
= NULL
;
1847 LLVMValueRef tmp7
= NULL
;
1848 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1850 uint dims
= (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_NRM
) ? 3 : 4;
1852 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) ||
1853 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
) ||
1854 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
) ||
1855 (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 4)) {
1857 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1860 /* xmm0 = src.x * src.x */
1861 tmp0
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
1862 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
1865 tmp0
= lp_build_mul( &bld
->bld_base
.base
, tmp0
, tmp0
);
1868 /* xmm0 = xmm0 + src.y * src.y */
1869 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Y
);
1870 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
1873 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1874 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1877 /* xmm0 = xmm0 + src.z * src.z */
1878 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Z
);
1879 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
1882 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1883 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1887 /* xmm0 = xmm0 + src.w * src.w */
1888 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
1889 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
)) {
1892 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1893 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1895 /* xmm1 = 1 / sqrt(xmm0) */
1896 tmp1
= lp_build_rsqrt( &bld
->bld_base
.base
, tmp0
);
1897 /* dst.x = xmm1 * src.x */
1898 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
1899 emit_data
->output
[TGSI_CHAN_X
] = lp_build_mul( &bld
->bld_base
.base
, tmp4
, tmp1
);
1901 /* dst.y = xmm1 * src.y */
1902 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
1903 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_mul( &bld
->bld_base
.base
, tmp5
, tmp1
);
1906 /* dst.z = xmm1 * src.z */
1907 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
1908 emit_data
->output
[TGSI_CHAN_Z
] = lp_build_mul( &bld
->bld_base
.base
, tmp6
, tmp1
);
1910 /* dst.w = xmm1 * src.w */
1911 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) && dims
== 4) {
1912 emit_data
->output
[TGSI_CHAN_W
] = lp_build_mul( &bld
->bld_base
.base
, tmp7
, tmp1
);
1917 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 3) {
1918 emit_data
->output
[TGSI_CHAN_W
] = bld
->bld_base
.base
.one
;
1922 static void emit_prologue(struct lp_build_tgsi_context
* bld_base
)
1924 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1925 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
1927 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
1928 LLVMValueRef array_size
=
1929 lp_build_const_int32(gallivm
,
1930 bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] * 4 + 4);
1931 bld
->temps_array
= lp_build_array_alloca(gallivm
,
1932 bld_base
->base
.vec_type
, array_size
,
1936 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
1937 LLVMValueRef array_size
=
1938 lp_build_const_int32(gallivm
,
1939 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] * 4 + 4);
1940 bld
->outputs_array
= lp_build_array_alloca(gallivm
,
1941 bld_base
->base
.vec_type
, array_size
,
1945 /* If we have indirect addressing in inputs we need to copy them into
1946 * our alloca array to be able to iterate over them */
1947 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
1948 unsigned index
, chan
;
1949 LLVMTypeRef vec_type
= bld_base
->base
.vec_type
;
1950 LLVMValueRef array_size
= lp_build_const_int32(gallivm
,
1951 bld_base
->info
->file_max
[TGSI_FILE_INPUT
]*4 + 4);
1952 bld
->inputs_array
= lp_build_array_alloca(gallivm
,
1953 vec_type
, array_size
,
1956 assert(bld_base
->info
->num_inputs
1957 <= bld_base
->info
->file_max
[TGSI_FILE_INPUT
] + 1);
1959 for (index
= 0; index
< bld_base
->info
->num_inputs
; ++index
) {
1960 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1961 LLVMValueRef lindex
=
1962 lp_build_const_int32(gallivm
, index
* 4 + chan
);
1963 LLVMValueRef input_ptr
=
1964 LLVMBuildGEP(gallivm
->builder
, bld
->inputs_array
,
1966 LLVMValueRef value
= bld
->inputs
[index
][chan
];
1968 LLVMBuildStore(gallivm
->builder
, value
, input_ptr
);
1974 static void emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
1976 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1980 emit_dump_temps(bld
);
1983 /* If we have indirect addressing in outputs we need to copy our alloca array
1984 * to the outputs slots specified by the called */
1985 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
1986 unsigned index
, chan
;
1987 assert(bld_base
->info
->num_outputs
<=
1988 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] + 1);
1989 for (index
= 0; index
< bld_base
->info
->num_outputs
; ++index
) {
1990 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1991 bld
->outputs
[index
][chan
] = lp_get_output_ptr(bld
, index
, chan
);
1998 lp_build_tgsi_soa(struct gallivm_state
*gallivm
,
1999 const struct tgsi_token
*tokens
,
2000 struct lp_type type
,
2001 struct lp_build_mask_context
*mask
,
2002 LLVMValueRef consts_ptr
,
2003 const struct lp_bld_tgsi_system_values
*system_values
,
2004 const LLVMValueRef
*pos
,
2005 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
2006 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
2007 struct lp_build_sampler_soa
*sampler
,
2008 const struct tgsi_shader_info
*info
)
2010 struct lp_build_tgsi_soa_context bld
;
2012 struct lp_type res_type
;
2014 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
2015 memset(&res_type
, 0, sizeof res_type
);
2016 res_type
.width
= type
.width
;
2017 res_type
.length
= type
.length
;
2020 /* Setup build context */
2021 memset(&bld
, 0, sizeof bld
);
2022 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
2023 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
2024 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
2025 lp_build_context_init(&bld
.elem_bld
, gallivm
, lp_elem_type(type
));
2028 bld
.inputs
= inputs
;
2029 bld
.outputs
= outputs
;
2030 bld
.consts_ptr
= consts_ptr
;
2031 bld
.sampler
= sampler
;
2032 bld
.bld_base
.info
= info
;
2033 bld
.indirect_files
= info
->indirect_files
;
2035 bld
.bld_base
.soa
= TRUE
;
2036 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
2037 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
2038 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
2039 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
2040 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = emit_fetch_system_value
;
2041 bld
.bld_base
.emit_store
= emit_store
;
2043 bld
.bld_base
.emit_declaration
= lp_emit_declaration_soa
;
2044 bld
.bld_base
.emit_immediate
= lp_emit_immediate_soa
;
2046 bld
.bld_base
.emit_prologue
= emit_prologue
;
2047 bld
.bld_base
.emit_epilogue
= emit_epilogue
;
2049 /* Set opcode actions */
2050 lp_set_default_actions_cpu(&bld
.bld_base
);
2052 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
2053 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNSUB
].emit
= bgnsub_emit
;
2054 bld
.bld_base
.op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
2055 bld
.bld_base
.op_actions
[TGSI_OPCODE_CAL
].emit
= cal_emit
;
2056 bld
.bld_base
.op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
2057 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDX
].emit
= ddx_emit
;
2058 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDY
].emit
= ddy_emit
;
2059 bld
.bld_base
.op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
2060 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
2061 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
2062 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSUB
].emit
= endsub_emit
;
2063 bld
.bld_base
.op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
2064 bld
.bld_base
.op_actions
[TGSI_OPCODE_KIL
].emit
= kil_emit
;
2065 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILP
].emit
= kilp_emit
;
2066 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM
].emit
= nrm_emit
;
2067 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM4
].emit
= nrm_emit
;
2068 bld
.bld_base
.op_actions
[TGSI_OPCODE_RET
].emit
= ret_emit
;
2069 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX
].emit
= tex_emit
;
2070 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB
].emit
= txb_emit
;
2071 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXD
].emit
= txd_emit
;
2072 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL
].emit
= txl_emit
;
2073 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXP
].emit
= txp_emit
;
2074 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXQ
].emit
= txq_emit
;
2076 lp_exec_mask_init(&bld
.exec_mask
, &bld
.bld_base
.base
);
2078 bld
.system_values
= *system_values
;
2080 lp_build_tgsi_llvm(&bld
.bld_base
, tokens
);
2083 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
2084 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
2085 debug_printf("11111111111111111111111111111 \n");
2086 tgsi_dump(tokens
, 0);
2087 lp_debug_dump_value(function
);
2088 debug_printf("2222222222222222222222222222 \n");
2092 LLVMModuleRef module
= LLVMGetGlobalParent(
2093 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
2094 LLVMDumpModule(module
);