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
)
70 mask
->has_mask
= FALSE
;
71 mask
->cond_stack_size
= 0;
72 mask
->loop_stack_size
= 0;
73 mask
->call_stack_size
= 0;
75 mask
->int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, mask
->bld
->type
);
76 mask
->exec_mask
= mask
->ret_mask
= mask
->break_mask
= mask
->cont_mask
= mask
->cond_mask
=
77 LLVMConstAllOnes(mask
->int_vec_type
);
80 static void lp_exec_mask_update(struct lp_exec_mask
*mask
)
82 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
84 if (mask
->loop_stack_size
) {
85 /*for loops we need to update the entire mask at runtime */
87 assert(mask
->break_mask
);
88 tmp
= LLVMBuildAnd(builder
,
92 mask
->exec_mask
= LLVMBuildAnd(builder
,
97 mask
->exec_mask
= mask
->cond_mask
;
99 if (mask
->call_stack_size
) {
100 mask
->exec_mask
= LLVMBuildAnd(builder
,
106 mask
->has_mask
= (mask
->cond_stack_size
> 0 ||
107 mask
->loop_stack_size
> 0 ||
108 mask
->call_stack_size
> 0);
111 static void lp_exec_mask_cond_push(struct lp_exec_mask
*mask
,
114 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
116 assert(mask
->cond_stack_size
< LP_MAX_TGSI_NESTING
);
117 if (mask
->cond_stack_size
== 0) {
118 assert(mask
->cond_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
120 mask
->cond_stack
[mask
->cond_stack_size
++] = mask
->cond_mask
;
121 assert(LLVMTypeOf(val
) == mask
->int_vec_type
);
122 mask
->cond_mask
= LLVMBuildAnd(builder
,
126 lp_exec_mask_update(mask
);
129 static void lp_exec_mask_cond_invert(struct lp_exec_mask
*mask
)
131 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
132 LLVMValueRef prev_mask
;
133 LLVMValueRef inv_mask
;
135 assert(mask
->cond_stack_size
);
136 prev_mask
= mask
->cond_stack
[mask
->cond_stack_size
- 1];
137 if (mask
->cond_stack_size
== 1) {
138 assert(prev_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
141 inv_mask
= LLVMBuildNot(builder
, mask
->cond_mask
, "");
143 mask
->cond_mask
= LLVMBuildAnd(builder
,
146 lp_exec_mask_update(mask
);
149 static void lp_exec_mask_cond_pop(struct lp_exec_mask
*mask
)
151 assert(mask
->cond_stack_size
);
152 mask
->cond_mask
= mask
->cond_stack
[--mask
->cond_stack_size
];
153 lp_exec_mask_update(mask
);
156 static void lp_exec_bgnloop(struct lp_exec_mask
*mask
)
158 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
160 if (mask
->loop_stack_size
== 0) {
161 assert(mask
->loop_block
== NULL
);
162 assert(mask
->cont_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
163 assert(mask
->break_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
164 assert(mask
->break_var
== NULL
);
167 assert(mask
->loop_stack_size
< LP_MAX_TGSI_NESTING
);
169 mask
->loop_stack
[mask
->loop_stack_size
].loop_block
= mask
->loop_block
;
170 mask
->loop_stack
[mask
->loop_stack_size
].cont_mask
= mask
->cont_mask
;
171 mask
->loop_stack
[mask
->loop_stack_size
].break_mask
= mask
->break_mask
;
172 mask
->loop_stack
[mask
->loop_stack_size
].break_var
= mask
->break_var
;
173 ++mask
->loop_stack_size
;
175 mask
->break_var
= lp_build_alloca(mask
->bld
->gallivm
, mask
->int_vec_type
, "");
176 LLVMBuildStore(builder
, mask
->break_mask
, mask
->break_var
);
178 mask
->loop_block
= lp_build_insert_new_block(mask
->bld
->gallivm
, "bgnloop");
179 LLVMBuildBr(builder
, mask
->loop_block
);
180 LLVMPositionBuilderAtEnd(builder
, mask
->loop_block
);
182 mask
->break_mask
= LLVMBuildLoad(builder
, mask
->break_var
, "");
184 lp_exec_mask_update(mask
);
187 static void lp_exec_break(struct lp_exec_mask
*mask
)
189 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
190 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
194 mask
->break_mask
= LLVMBuildAnd(builder
,
196 exec_mask
, "break_full");
198 lp_exec_mask_update(mask
);
201 static void lp_exec_continue(struct lp_exec_mask
*mask
)
203 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
204 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
208 mask
->cont_mask
= LLVMBuildAnd(builder
,
212 lp_exec_mask_update(mask
);
216 static void lp_exec_endloop(struct gallivm_state
*gallivm
,
217 struct lp_exec_mask
*mask
)
219 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
220 LLVMBasicBlockRef endloop
;
221 LLVMTypeRef reg_type
= LLVMIntTypeInContext(gallivm
->context
,
222 mask
->bld
->type
.width
*
223 mask
->bld
->type
.length
);
226 assert(mask
->break_mask
);
229 * Restore the cont_mask, but don't pop
231 assert(mask
->loop_stack_size
);
232 mask
->cont_mask
= mask
->loop_stack
[mask
->loop_stack_size
- 1].cont_mask
;
233 lp_exec_mask_update(mask
);
236 * Unlike the continue mask, the break_mask must be preserved across loop
239 LLVMBuildStore(builder
, mask
->break_mask
, mask
->break_var
);
241 /* i1cond = (mask == 0) */
242 i1cond
= LLVMBuildICmp(
245 LLVMBuildBitCast(builder
, mask
->exec_mask
, reg_type
, ""),
246 LLVMConstNull(reg_type
), "");
248 endloop
= lp_build_insert_new_block(mask
->bld
->gallivm
, "endloop");
250 LLVMBuildCondBr(builder
,
251 i1cond
, mask
->loop_block
, endloop
);
253 LLVMPositionBuilderAtEnd(builder
, endloop
);
255 assert(mask
->loop_stack_size
);
256 --mask
->loop_stack_size
;
257 mask
->loop_block
= mask
->loop_stack
[mask
->loop_stack_size
].loop_block
;
258 mask
->cont_mask
= mask
->loop_stack
[mask
->loop_stack_size
].cont_mask
;
259 mask
->break_mask
= mask
->loop_stack
[mask
->loop_stack_size
].break_mask
;
260 mask
->break_var
= mask
->loop_stack
[mask
->loop_stack_size
].break_var
;
262 lp_exec_mask_update(mask
);
265 /* stores val into an address pointed to by dst.
266 * mask->exec_mask is used to figure out which bits of val
267 * should be stored into the address
268 * (0 means don't store this bit, 1 means do store).
270 static void lp_exec_mask_store(struct lp_exec_mask
*mask
,
271 struct lp_build_context
*bld_store
,
276 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
278 /* Mix the predicate and execution mask */
279 if (mask
->has_mask
) {
281 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
283 pred
= mask
->exec_mask
;
288 LLVMValueRef real_val
, dst_val
;
290 dst_val
= LLVMBuildLoad(builder
, dst
, "");
291 real_val
= lp_build_select(bld_store
,
295 LLVMBuildStore(builder
, real_val
, dst
);
297 LLVMBuildStore(builder
, val
, dst
);
300 static void lp_exec_mask_call(struct lp_exec_mask
*mask
,
304 assert(mask
->call_stack_size
< LP_MAX_TGSI_NESTING
);
305 mask
->call_stack
[mask
->call_stack_size
].pc
= *pc
;
306 mask
->call_stack
[mask
->call_stack_size
].ret_mask
= mask
->ret_mask
;
307 mask
->call_stack_size
++;
311 static void lp_exec_mask_ret(struct lp_exec_mask
*mask
, int *pc
)
313 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
314 LLVMValueRef exec_mask
;
316 if (mask
->call_stack_size
== 0) {
317 /* returning from main() */
321 exec_mask
= LLVMBuildNot(builder
,
325 mask
->ret_mask
= LLVMBuildAnd(builder
,
327 exec_mask
, "ret_full");
329 lp_exec_mask_update(mask
);
332 static void lp_exec_mask_bgnsub(struct lp_exec_mask
*mask
)
336 static void lp_exec_mask_endsub(struct lp_exec_mask
*mask
, int *pc
)
338 assert(mask
->call_stack_size
);
339 mask
->call_stack_size
--;
340 *pc
= mask
->call_stack
[mask
->call_stack_size
].pc
;
341 mask
->ret_mask
= mask
->call_stack
[mask
->call_stack_size
].ret_mask
;
342 lp_exec_mask_update(mask
);
347 * Return pointer to a temporary register channel (src or dest).
348 * Note that indirect addressing cannot be handled here.
349 * \param index which temporary register
350 * \param chan which channel of the temp register.
353 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context
*bld
,
357 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
359 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
360 LLVMValueRef lindex
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
, index
* 4 + chan
);
361 return LLVMBuildGEP(builder
, bld
->temps_array
, &lindex
, 1, "");
364 return bld
->temps
[index
][chan
];
369 * Return pointer to a output register channel (src or dest).
370 * Note that indirect addressing cannot be handled here.
371 * \param index which output register
372 * \param chan which channel of the output register.
375 lp_get_output_ptr(struct lp_build_tgsi_soa_context
*bld
,
379 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
381 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
382 LLVMValueRef lindex
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
,
384 return LLVMBuildGEP(builder
, bld
->outputs_array
, &lindex
, 1, "");
387 return bld
->outputs
[index
][chan
];
393 * XXX the lp_build_gather() function should be capable of doing this
394 * with a little work.
397 build_gather(struct lp_build_tgsi_soa_context
*bld
,
398 LLVMValueRef base_ptr
,
399 LLVMValueRef indexes
)
401 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
402 LLVMValueRef res
= bld
->bld_base
.base
.undef
;
406 * Loop over elements of index_vec, load scalar value, insert it into 'res'.
408 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
409 LLVMValueRef ii
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
, i
);
410 LLVMValueRef index
= LLVMBuildExtractElement(builder
,
412 LLVMValueRef scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
,
413 &index
, 1, "gather_ptr");
414 LLVMValueRef scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
416 res
= LLVMBuildInsertElement(builder
, res
, scalar
, ii
, "");
424 * Scatter/store vector.
427 emit_mask_scatter(struct lp_build_tgsi_soa_context
*bld
,
428 LLVMValueRef base_ptr
,
429 LLVMValueRef indexes
,
431 struct lp_exec_mask
*mask
,
434 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
435 LLVMBuilderRef builder
= gallivm
->builder
;
438 /* Mix the predicate and execution mask */
439 if (mask
->has_mask
) {
441 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
444 pred
= mask
->exec_mask
;
449 * Loop over elements of index_vec, store scalar value.
451 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
452 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
453 LLVMValueRef index
= LLVMBuildExtractElement(builder
, indexes
, ii
, "");
454 LLVMValueRef scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
, &index
, 1, "scatter_ptr");
455 LLVMValueRef val
= LLVMBuildExtractElement(builder
, values
, ii
, "scatter_val");
456 LLVMValueRef scalar_pred
= pred
?
457 LLVMBuildExtractElement(builder
, pred
, ii
, "scatter_pred") : NULL
;
460 lp_build_printf(gallivm
, "scatter %d: val %f at %d %p\n",
461 ii
, val
, index
, scalar_ptr
);
464 LLVMValueRef real_val
, dst_val
;
465 dst_val
= LLVMBuildLoad(builder
, scalar_ptr
, "");
466 real_val
= lp_build_select(&bld
->elem_bld
, scalar_pred
, val
, dst_val
);
467 LLVMBuildStore(builder
, real_val
, scalar_ptr
);
470 LLVMBuildStore(builder
, val
, scalar_ptr
);
477 * Read the current value of the ADDR register, convert the floats to
478 * ints, add the base index and return the vector of offsets.
479 * The offsets will be used to index into the constant buffer or
480 * temporary register file.
483 get_indirect_index(struct lp_build_tgsi_soa_context
*bld
,
484 unsigned reg_file
, unsigned reg_index
,
485 const struct tgsi_src_register
*indirect_reg
)
487 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
488 struct lp_build_context
*uint_bld
= &bld
->uint_bld
;
489 /* always use X component of address register */
490 unsigned swizzle
= indirect_reg
->SwizzleX
;
493 LLVMValueRef max_index
;
496 assert(bld
->indirect_files
& (1 << reg_file
));
498 base
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, reg_index
);
501 rel
= LLVMBuildLoad(builder
,
502 bld
->addr
[indirect_reg
->Index
][swizzle
],
505 /* for indexing we want integers */
506 rel
= LLVMBuildFPToSI(builder
,
508 uint_bld
->vec_type
, "");
510 index
= lp_build_add(uint_bld
, base
, rel
);
512 max_index
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
,
514 bld
->bld_base
.info
->file_max
[reg_file
]);
516 assert(!uint_bld
->type
.sign
);
517 index
= lp_build_min(uint_bld
, index
, max_index
);
524 struct lp_build_tgsi_context
* bld_base
,
525 const struct tgsi_full_src_register
* reg
,
526 const unsigned swizzle
)
528 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
529 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
530 LLVMBuilderRef builder
= gallivm
->builder
;
531 struct lp_build_context
*uint_bld
= &bld
->uint_bld
;
532 LLVMValueRef indirect_index
= NULL
;
534 /* XXX: Handle fetching xyzw components as a vector */
535 assert(swizzle
!= ~0);
537 if (reg
->Register
.Indirect
) {
538 indirect_index
= get_indirect_index(bld
,
544 if (reg
->Register
.Indirect
) {
545 LLVMValueRef swizzle_vec
=
546 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, swizzle
);
547 LLVMValueRef index_vec
; /* index into the const buffer */
549 /* index_vec = indirect_index * 4 + swizzle */
550 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
551 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
553 /* Gather values from the constant buffer */
554 return build_gather(bld
, bld
->consts_ptr
, index_vec
);
557 LLVMValueRef index
; /* index into the const buffer */
558 LLVMValueRef scalar
, scalar_ptr
;
560 index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
*4 + swizzle
);
562 scalar_ptr
= LLVMBuildGEP(builder
, bld
->consts_ptr
,
564 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
566 return lp_build_broadcast_scalar(&bld
->bld_base
.base
, scalar
);
571 emit_fetch_immediate(
572 struct lp_build_tgsi_context
* bld_base
,
573 const struct tgsi_full_src_register
* reg
,
574 const unsigned swizzle
)
576 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
577 LLVMValueRef res
= bld
->immediates
[reg
->Register
.Index
][swizzle
];
584 struct lp_build_tgsi_context
* bld_base
,
585 const struct tgsi_full_src_register
* reg
,
586 const unsigned swizzle
)
588 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
589 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
590 LLVMBuilderRef builder
= gallivm
->builder
;
591 struct lp_build_context
*uint_bld
= &bld
->uint_bld
;
592 LLVMValueRef indirect_index
= NULL
;
595 if (reg
->Register
.Indirect
) {
596 indirect_index
= get_indirect_index(bld
,
602 if (reg
->Register
.Indirect
) {
603 LLVMValueRef swizzle_vec
=
604 lp_build_const_int_vec(gallivm
, uint_bld
->type
, swizzle
);
605 LLVMValueRef length_vec
=
606 lp_build_const_int_vec(gallivm
, uint_bld
->type
, bld
->bld_base
.base
.type
.length
);
607 LLVMValueRef index_vec
; /* index into the const buffer */
608 LLVMValueRef inputs_array
;
609 LLVMTypeRef float4_ptr_type
;
611 /* index_vec = (indirect_index * 4 + swizzle) * length */
612 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
613 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
614 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
616 /* cast inputs_array pointer to float* */
617 float4_ptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
618 inputs_array
= LLVMBuildBitCast(builder
, bld
->inputs_array
,
619 float4_ptr_type
, "");
621 /* Gather values from the temporary register array */
622 res
= build_gather(bld
, inputs_array
, index_vec
);
624 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
625 LLVMValueRef lindex
= lp_build_const_int32(gallivm
,
626 reg
->Register
.Index
* 4 + swizzle
);
627 LLVMValueRef input_ptr
= LLVMBuildGEP(builder
,
628 bld
->inputs_array
, &lindex
, 1, "");
629 res
= LLVMBuildLoad(builder
, input_ptr
, "");
632 res
= bld
->inputs
[reg
->Register
.Index
][swizzle
];
640 emit_fetch_temporary(
641 struct lp_build_tgsi_context
* bld_base
,
642 const struct tgsi_full_src_register
* reg
,
643 const unsigned swizzle
)
645 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
646 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
647 LLVMBuilderRef builder
= gallivm
->builder
;
648 struct lp_build_context
*uint_bld
= &bld
->uint_bld
;
649 LLVMValueRef indirect_index
= NULL
;
652 if (reg
->Register
.Indirect
) {
653 indirect_index
= get_indirect_index(bld
,
659 if (reg
->Register
.Indirect
) {
660 LLVMValueRef swizzle_vec
=
661 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, swizzle
);
662 LLVMValueRef length_vec
=
663 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
,
664 bld
->bld_base
.base
.type
.length
);
665 LLVMValueRef index_vec
; /* index into the const buffer */
666 LLVMValueRef temps_array
;
667 LLVMTypeRef float4_ptr_type
;
669 /* index_vec = (indirect_index * 4 + swizzle) * length */
670 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
671 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
672 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
674 /* cast temps_array pointer to float* */
675 float4_ptr_type
= LLVMPointerType(LLVMFloatTypeInContext(bld
->bld_base
.base
.gallivm
->context
), 0);
676 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
,
677 float4_ptr_type
, "");
679 /* Gather values from the temporary register array */
680 res
= build_gather(bld
, temps_array
, index_vec
);
683 LLVMValueRef temp_ptr
;
684 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
685 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
687 return bld
->bld_base
.base
.undef
;
694 emit_fetch_system_value(
695 struct lp_build_tgsi_context
* bld_base
,
696 const struct tgsi_full_src_register
* reg
,
697 const unsigned swizzle
)
699 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
700 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
701 LLVMBuilderRef builder
= gallivm
->builder
;
702 LLVMValueRef index
; /* index into the system value array */
703 LLVMValueRef scalar
, scalar_ptr
;
705 assert(!reg
->Register
.Indirect
);
707 index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
* 4 + swizzle
);
709 scalar_ptr
= LLVMBuildGEP(builder
, bld
->system_values_array
, &index
, 1, "");
710 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
712 return lp_build_broadcast_scalar(&bld
->bld_base
.base
, scalar
);
716 * Register fetch with derivatives.
720 struct lp_build_tgsi_soa_context
*bld
,
729 /* TODO: use interpolation coeffs for inputs */
732 *ddx
= lp_build_ddx(&bld
->bld_base
.base
, src
);
735 *ddy
= lp_build_ddy(&bld
->bld_base
.base
, src
);
743 emit_fetch_predicate(
744 struct lp_build_tgsi_soa_context
*bld
,
745 const struct tgsi_full_instruction
*inst
,
748 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
750 unsigned char swizzles
[4];
751 LLVMValueRef unswizzled
[4] = {NULL
, NULL
, NULL
, NULL
};
755 if (!inst
->Instruction
.Predicate
) {
756 TGSI_FOR_EACH_CHANNEL( chan
) {
762 swizzles
[0] = inst
->Predicate
.SwizzleX
;
763 swizzles
[1] = inst
->Predicate
.SwizzleY
;
764 swizzles
[2] = inst
->Predicate
.SwizzleZ
;
765 swizzles
[3] = inst
->Predicate
.SwizzleW
;
767 index
= inst
->Predicate
.Index
;
768 assert(index
< LP_MAX_TGSI_PREDS
);
770 TGSI_FOR_EACH_CHANNEL( chan
) {
771 unsigned swizzle
= swizzles
[chan
];
774 * Only fetch the predicate register channels that are actually listed
777 if (!unswizzled
[swizzle
]) {
778 value
= LLVMBuildLoad(builder
,
779 bld
->preds
[index
][swizzle
], "");
782 * Convert the value to an integer mask.
784 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
785 * is needlessly causing two comparisons due to storing the intermediate
786 * result as float vector instead of an integer mask vector.
788 value
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
789 bld
->bld_base
.base
.type
,
792 bld
->bld_base
.base
.zero
);
793 if (inst
->Predicate
.Negate
) {
794 value
= LLVMBuildNot(builder
, value
, "");
797 unswizzled
[swizzle
] = value
;
799 value
= unswizzled
[swizzle
];
812 struct lp_build_tgsi_context
*bld_base
,
813 const struct tgsi_full_instruction
*inst
,
819 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
820 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
821 LLVMBuilderRef builder
= gallivm
->builder
;
822 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
823 struct lp_build_context
*uint_bld
= &bld
->uint_bld
;
824 LLVMValueRef indirect_index
= NULL
;
825 struct lp_build_context
*bld_store
;
827 bld_store
= &bld
->bld_base
.base
;
829 switch( inst
->Instruction
.Saturate
) {
833 case TGSI_SAT_ZERO_ONE
:
834 value
= lp_build_max(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.zero
);
835 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
838 case TGSI_SAT_MINUS_PLUS_ONE
:
839 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));
840 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
847 if (reg
->Register
.Indirect
) {
848 indirect_index
= get_indirect_index(bld
,
853 assert(reg
->Register
.Index
<=
854 bld
->bld_base
.info
->file_max
[reg
->Register
.File
]);
857 switch( reg
->Register
.File
) {
858 case TGSI_FILE_OUTPUT
:
859 if (reg
->Register
.Indirect
) {
860 LLVMValueRef chan_vec
=
861 lp_build_const_int_vec(gallivm
, uint_bld
->type
, chan_index
);
862 LLVMValueRef length_vec
=
863 lp_build_const_int_vec(gallivm
, uint_bld
->type
, bld
->bld_base
.base
.type
.length
);
864 LLVMValueRef index_vec
; /* indexes into the temp registers */
865 LLVMValueRef outputs_array
;
866 LLVMValueRef pixel_offsets
;
867 LLVMTypeRef float_ptr_type
;
870 /* build pixel offset vector: {0, 1, 2, 3, ...} */
871 pixel_offsets
= uint_bld
->undef
;
872 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
873 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
874 pixel_offsets
= LLVMBuildInsertElement(builder
, pixel_offsets
,
878 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
879 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
880 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
881 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
882 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
885 LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
886 outputs_array
= LLVMBuildBitCast(builder
, bld
->outputs_array
,
889 /* Scatter store values into temp registers */
890 emit_mask_scatter(bld
, outputs_array
, index_vec
, value
,
891 &bld
->exec_mask
, pred
);
894 LLVMValueRef out_ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
,
896 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
, out_ptr
);
900 case TGSI_FILE_TEMPORARY
:
901 if (reg
->Register
.Indirect
) {
902 LLVMValueRef chan_vec
=
903 lp_build_const_int_vec(gallivm
, uint_bld
->type
, chan_index
);
904 LLVMValueRef length_vec
=
905 lp_build_const_int_vec(gallivm
, uint_bld
->type
,
906 bld
->bld_base
.base
.type
.length
);
907 LLVMValueRef index_vec
; /* indexes into the temp registers */
908 LLVMValueRef temps_array
;
909 LLVMValueRef pixel_offsets
;
910 LLVMTypeRef float_ptr_type
;
913 /* build pixel offset vector: {0, 1, 2, 3, ...} */
914 pixel_offsets
= uint_bld
->undef
;
915 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
916 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
917 pixel_offsets
= LLVMBuildInsertElement(builder
, pixel_offsets
,
921 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
922 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
923 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
924 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
925 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
928 LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
929 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
,
932 /* Scatter store values into temp registers */
933 emit_mask_scatter(bld
, temps_array
, index_vec
, value
,
934 &bld
->exec_mask
, pred
);
937 LLVMValueRef temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
,
939 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
, temp_ptr
);
943 case TGSI_FILE_ADDRESS
:
944 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
,
945 bld
->addr
[reg
->Register
.Index
][chan_index
]);
948 case TGSI_FILE_PREDICATE
:
949 lp_exec_mask_store(&bld
->exec_mask
, bld_store
, pred
, value
,
950 bld
->preds
[reg
->Register
.Index
][chan_index
]);
960 struct lp_build_tgsi_context
* bld_base
,
961 const struct tgsi_full_instruction
* inst
,
962 const struct tgsi_opcode_info
* info
,
967 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
970 LLVMValueRef pred
[TGSI_NUM_CHANNELS
];
972 emit_fetch_predicate( bld
, inst
, pred
);
974 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
975 emit_store_chan(bld_base
, inst
, 0, chan_index
, pred
[chan_index
], dst
[chan_index
]);
981 * High-level instruction translators.
985 emit_tex( struct lp_build_tgsi_soa_context
*bld
,
986 const struct tgsi_full_instruction
*inst
,
987 enum lp_build_tex_modifier modifier
,
990 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
992 LLVMValueRef lod_bias
, explicit_lod
;
993 LLVMValueRef oow
= NULL
;
994 LLVMValueRef coords
[3];
1000 if (!bld
->sampler
) {
1001 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1002 for (i
= 0; i
< 4; i
++) {
1003 texel
[i
] = bld
->bld_base
.base
.undef
;
1008 switch (inst
->Texture
.Texture
) {
1009 case TGSI_TEXTURE_1D
:
1012 case TGSI_TEXTURE_1D_ARRAY
:
1013 case TGSI_TEXTURE_2D
:
1014 case TGSI_TEXTURE_RECT
:
1017 case TGSI_TEXTURE_SHADOW1D
:
1018 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1019 case TGSI_TEXTURE_SHADOW2D
:
1020 case TGSI_TEXTURE_SHADOWRECT
:
1021 case TGSI_TEXTURE_2D_ARRAY
:
1022 case TGSI_TEXTURE_3D
:
1023 case TGSI_TEXTURE_CUBE
:
1026 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1034 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
1035 lod_bias
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1036 explicit_lod
= NULL
;
1038 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
1040 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1044 explicit_lod
= NULL
;
1047 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
) {
1048 oow
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1049 oow
= lp_build_rcp(&bld
->bld_base
.base
, oow
);
1052 for (i
= 0; i
< num_coords
; i
++) {
1053 coords
[i
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, i
);
1054 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
1055 coords
[i
] = lp_build_mul(&bld
->bld_base
.base
, coords
[i
], oow
);
1057 for (i
= num_coords
; i
< 3; i
++) {
1058 coords
[i
] = bld
->bld_base
.base
.undef
;
1061 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
1062 LLVMValueRef index0
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
, 0);
1063 for (i
= 0; i
< num_coords
; i
++) {
1064 LLVMValueRef src1
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 1, i
);
1065 LLVMValueRef src2
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 2, i
);
1066 ddx
[i
] = LLVMBuildExtractElement(builder
, src1
, index0
, "");
1067 ddy
[i
] = LLVMBuildExtractElement(builder
, src2
, index0
, "");
1069 unit
= inst
->Src
[3].Register
.Index
;
1071 for (i
= 0; i
< num_coords
; i
++) {
1072 ddx
[i
] = lp_build_scalar_ddx( &bld
->bld_base
.base
, coords
[i
] );
1073 ddy
[i
] = lp_build_scalar_ddy( &bld
->bld_base
.base
, coords
[i
] );
1075 unit
= inst
->Src
[1].Register
.Index
;
1077 for (i
= num_coords
; i
< 3; i
++) {
1078 ddx
[i
] = LLVMGetUndef(bld
->bld_base
.base
.elem_type
);
1079 ddy
[i
] = LLVMGetUndef(bld
->bld_base
.base
.elem_type
);
1082 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
1083 bld
->bld_base
.base
.gallivm
,
1084 bld
->bld_base
.base
.type
,
1085 unit
, num_coords
, coords
,
1087 lod_bias
, explicit_lod
,
1092 near_end_of_shader(struct lp_build_tgsi_soa_context
*bld
,
1097 for (i
= 0; i
< 5; i
++) {
1100 if (pc
+ i
>= bld
->bld_base
.info
->num_instructions
)
1103 opcode
= bld
->bld_base
.instructions
[pc
+ i
].Instruction
.Opcode
;
1105 if (opcode
== TGSI_OPCODE_END
)
1108 if (opcode
== TGSI_OPCODE_TEX
||
1109 opcode
== TGSI_OPCODE_TXP
||
1110 opcode
== TGSI_OPCODE_TXD
||
1111 opcode
== TGSI_OPCODE_TXB
||
1112 opcode
== TGSI_OPCODE_TXL
||
1113 opcode
== TGSI_OPCODE_TXF
||
1114 opcode
== TGSI_OPCODE_TXQ
||
1115 opcode
== TGSI_OPCODE_CAL
||
1116 opcode
== TGSI_OPCODE_CALLNZ
||
1117 opcode
== TGSI_OPCODE_IF
||
1118 opcode
== TGSI_OPCODE_IFC
||
1119 opcode
== TGSI_OPCODE_BGNLOOP
||
1120 opcode
== TGSI_OPCODE_SWITCH
)
1130 * Kill fragment if any of the src register values are negative.
1134 struct lp_build_tgsi_soa_context
*bld
,
1135 const struct tgsi_full_instruction
*inst
,
1138 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1139 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
1140 LLVMValueRef terms
[TGSI_NUM_CHANNELS
];
1142 unsigned chan_index
;
1144 memset(&terms
, 0, sizeof terms
);
1146 TGSI_FOR_EACH_CHANNEL( chan_index
) {
1149 /* Unswizzle channel */
1150 swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
1152 /* Check if the component has not been already tested. */
1153 assert(swizzle
< TGSI_NUM_CHANNELS
);
1154 if( !terms
[swizzle
] )
1155 /* TODO: change the comparison operator instead of setting the sign */
1156 terms
[swizzle
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, chan_index
);
1160 TGSI_FOR_EACH_CHANNEL( chan_index
) {
1161 if(terms
[chan_index
]) {
1162 LLVMValueRef chan_mask
;
1165 * If term < 0 then mask = 0 else mask = ~0.
1167 chan_mask
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, terms
[chan_index
], bld
->bld_base
.base
.zero
);
1170 mask
= LLVMBuildAnd(builder
, mask
, chan_mask
, "");
1177 lp_build_mask_update(bld
->mask
, mask
);
1179 if (!near_end_of_shader(bld
, pc
))
1180 lp_build_mask_check(bld
->mask
);
1186 * Predicated fragment kill.
1187 * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
1188 * The only predication is the execution mask which will apply if
1189 * we're inside a loop or conditional.
1192 emit_kilp(struct lp_build_tgsi_soa_context
*bld
,
1195 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1198 /* For those channels which are "alive", disable fragment shader
1201 if (bld
->exec_mask
.has_mask
) {
1202 mask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
1205 LLVMValueRef zero
= LLVMConstNull(bld
->bld_base
.base
.int_vec_type
);
1209 lp_build_mask_update(bld
->mask
, mask
);
1211 if (!near_end_of_shader(bld
, pc
))
1212 lp_build_mask_check(bld
->mask
);
1217 * Emit code which will dump the value of all the temporary registers
1221 emit_dump_temps(struct lp_build_tgsi_soa_context
*bld
)
1223 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1224 LLVMBuilderRef builder
= gallivm
->builder
;
1225 LLVMValueRef temp_ptr
;
1226 LLVMValueRef i0
= lp_build_const_int32(gallivm
, 0);
1227 LLVMValueRef i1
= lp_build_const_int32(gallivm
, 1);
1228 LLVMValueRef i2
= lp_build_const_int32(gallivm
, 2);
1229 LLVMValueRef i3
= lp_build_const_int32(gallivm
, 3);
1231 int n
= bld
->bld_base
.info
->file_max
[TGSI_FILE_TEMPORARY
];
1233 for (index
= 0; index
< n
; index
++) {
1234 LLVMValueRef idx
= lp_build_const_int32(gallivm
, index
);
1235 LLVMValueRef v
[4][4], res
;
1238 lp_build_printf(gallivm
, "TEMP[%d]:\n", idx
);
1240 for (chan
= 0; chan
< 4; chan
++) {
1241 temp_ptr
= lp_get_temp_ptr_soa(bld
, index
, chan
);
1242 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
1243 v
[chan
][0] = LLVMBuildExtractElement(builder
, res
, i0
, "");
1244 v
[chan
][1] = LLVMBuildExtractElement(builder
, res
, i1
, "");
1245 v
[chan
][2] = LLVMBuildExtractElement(builder
, res
, i2
, "");
1246 v
[chan
][3] = LLVMBuildExtractElement(builder
, res
, i3
, "");
1249 lp_build_printf(gallivm
, " X: %f %f %f %f\n",
1250 v
[0][0], v
[0][1], v
[0][2], v
[0][3]);
1251 lp_build_printf(gallivm
, " Y: %f %f %f %f\n",
1252 v
[1][0], v
[1][1], v
[1][2], v
[1][3]);
1253 lp_build_printf(gallivm
, " Z: %f %f %f %f\n",
1254 v
[2][0], v
[2][1], v
[2][2], v
[2][3]);
1255 lp_build_printf(gallivm
, " W: %f %f %f %f\n",
1256 v
[3][0], v
[3][1], v
[3][2], v
[3][3]);
1263 lp_emit_declaration_soa(
1264 struct lp_build_tgsi_context
*bld_base
,
1265 const struct tgsi_full_declaration
*decl
)
1267 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
1268 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1269 LLVMTypeRef vec_type
= bld
->bld_base
.base
.vec_type
;
1270 const unsigned first
= decl
->Range
.First
;
1271 const unsigned last
= decl
->Range
.Last
;
1274 for (idx
= first
; idx
<= last
; ++idx
) {
1275 assert(last
<= bld
->bld_base
.info
->file_max
[decl
->Declaration
.File
]);
1276 switch (decl
->Declaration
.File
) {
1277 case TGSI_FILE_TEMPORARY
:
1278 assert(idx
< LP_MAX_TGSI_TEMPS
);
1279 if (!(bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
))) {
1280 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1281 bld
->temps
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
, "temp");
1285 case TGSI_FILE_OUTPUT
:
1286 if (!(bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
1287 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1288 bld
->outputs
[idx
][i
] = lp_build_alloca(gallivm
,
1289 vec_type
, "output");
1293 case TGSI_FILE_ADDRESS
:
1294 assert(idx
< LP_MAX_TGSI_ADDRS
);
1295 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1296 bld
->addr
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
, "addr");
1299 case TGSI_FILE_PREDICATE
:
1300 assert(idx
< LP_MAX_TGSI_PREDS
);
1301 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
1302 bld
->preds
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
,
1307 /* don't need to declare other vars */
1314 void lp_emit_immediate_soa(
1315 struct lp_build_tgsi_context
*bld_base
,
1316 const struct tgsi_full_immediate
*imm
)
1318 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
1319 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
1321 /* simply copy the immediate values into the next immediates[] slot */
1323 const uint size
= imm
->Immediate
.NrTokens
- 1;
1325 assert(bld
->num_immediates
< LP_MAX_TGSI_IMMEDIATES
);
1327 for( i
= 0; i
< size
; ++i
)
1328 bld
->immediates
[bld
->num_immediates
][i
] =
1329 lp_build_const_vec(gallivm
, bld_base
->base
.type
, imm
->u
[i
].Float
);
1331 for( i
= size
; i
< 4; ++i
)
1332 bld
->immediates
[bld
->num_immediates
][i
] = bld_base
->base
.undef
;
1334 bld
->num_immediates
++;
1339 const struct lp_build_tgsi_action
* action
,
1340 struct lp_build_tgsi_context
* bld_base
,
1341 struct lp_build_emit_data
* emit_data
)
1343 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1345 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
,
1346 &emit_data
->output
[emit_data
->chan
], NULL
);
1351 const struct lp_build_tgsi_action
* action
,
1352 struct lp_build_tgsi_context
* bld_base
,
1353 struct lp_build_emit_data
* emit_data
)
1355 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1357 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
, NULL
,
1358 &emit_data
->output
[emit_data
->chan
]);
1363 const struct lp_build_tgsi_action
* action
,
1364 struct lp_build_tgsi_context
* bld_base
,
1365 struct lp_build_emit_data
* emit_data
)
1367 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1369 emit_kilp(bld
, bld_base
->pc
- 1);
1374 const struct lp_build_tgsi_action
* action
,
1375 struct lp_build_tgsi_context
* bld_base
,
1376 struct lp_build_emit_data
* emit_data
)
1378 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1380 emit_kil(bld
, emit_data
->inst
, bld_base
->pc
- 1);
1385 const struct lp_build_tgsi_action
* action
,
1386 struct lp_build_tgsi_context
* bld_base
,
1387 struct lp_build_emit_data
* emit_data
)
1389 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1391 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
, emit_data
->output
);
1396 const struct lp_build_tgsi_action
* action
,
1397 struct lp_build_tgsi_context
* bld_base
,
1398 struct lp_build_emit_data
* emit_data
)
1400 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1402 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
1408 const struct lp_build_tgsi_action
* action
,
1409 struct lp_build_tgsi_context
* bld_base
,
1410 struct lp_build_emit_data
* emit_data
)
1412 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1414 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
1420 const struct lp_build_tgsi_action
* action
,
1421 struct lp_build_tgsi_context
* bld_base
,
1422 struct lp_build_emit_data
* emit_data
)
1424 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1426 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
1432 const struct lp_build_tgsi_action
* action
,
1433 struct lp_build_tgsi_context
* bld_base
,
1434 struct lp_build_emit_data
* emit_data
)
1436 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1438 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_PROJECTED
,
1444 const struct lp_build_tgsi_action
* action
,
1445 struct lp_build_tgsi_context
* bld_base
,
1446 struct lp_build_emit_data
* emit_data
)
1448 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1450 lp_exec_mask_call(&bld
->exec_mask
, emit_data
->inst
->Label
.Label
,
1456 const struct lp_build_tgsi_action
* action
,
1457 struct lp_build_tgsi_context
* bld_base
,
1458 struct lp_build_emit_data
* emit_data
)
1460 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1462 lp_exec_mask_ret(&bld
->exec_mask
, &bld_base
->pc
);
1467 const struct lp_build_tgsi_action
* action
,
1468 struct lp_build_tgsi_context
* bld_base
,
1469 struct lp_build_emit_data
* emit_data
)
1471 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1473 lp_exec_break(&bld
->exec_mask
);
1478 const struct lp_build_tgsi_action
* action
,
1479 struct lp_build_tgsi_context
* bld_base
,
1480 struct lp_build_emit_data
* emit_data
)
1483 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1485 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_NOTEQUAL
,
1486 emit_data
->args
[0], bld
->bld_base
.base
.zero
);
1487 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
1492 const struct lp_build_tgsi_action
* action
,
1493 struct lp_build_tgsi_context
* bld_base
,
1494 struct lp_build_emit_data
* emit_data
)
1496 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1498 lp_exec_bgnloop(&bld
->exec_mask
);
1503 const struct lp_build_tgsi_action
* action
,
1504 struct lp_build_tgsi_context
* bld_base
,
1505 struct lp_build_emit_data
* emit_data
)
1507 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1509 lp_exec_mask_bgnsub(&bld
->exec_mask
);
1514 const struct lp_build_tgsi_action
* action
,
1515 struct lp_build_tgsi_context
* bld_base
,
1516 struct lp_build_emit_data
* emit_data
)
1518 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1520 lp_exec_mask_cond_invert(&bld
->exec_mask
);
1525 const struct lp_build_tgsi_action
* action
,
1526 struct lp_build_tgsi_context
* bld_base
,
1527 struct lp_build_emit_data
* emit_data
)
1529 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1531 lp_exec_mask_cond_pop(&bld
->exec_mask
);
1536 const struct lp_build_tgsi_action
* action
,
1537 struct lp_build_tgsi_context
* bld_base
,
1538 struct lp_build_emit_data
* emit_data
)
1540 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1542 lp_exec_endloop(bld_base
->base
.gallivm
, &bld
->exec_mask
);
1547 const struct lp_build_tgsi_action
* action
,
1548 struct lp_build_tgsi_context
* bld_base
,
1549 struct lp_build_emit_data
* emit_data
)
1551 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1553 lp_exec_mask_endsub(&bld
->exec_mask
, &bld_base
->pc
);
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 lp_exec_continue(&bld
->exec_mask
);
1567 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
1569 * XXX: What do the comments about xmm registers mean? Maybe they are left over
1570 * from old code, but there is no garauntee that LLVM will use those registers
1573 * XXX: There should be no calls to lp_build_emit_fetch in this function. This
1574 * should be handled by the emit_data->fetch_args function. */
1577 const struct lp_build_tgsi_action
* action
,
1578 struct lp_build_tgsi_context
* bld_base
,
1579 struct lp_build_emit_data
* emit_data
)
1581 LLVMValueRef tmp0
, tmp1
;
1582 LLVMValueRef tmp4
= NULL
;
1583 LLVMValueRef tmp5
= NULL
;
1584 LLVMValueRef tmp6
= NULL
;
1585 LLVMValueRef tmp7
= NULL
;
1586 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1588 uint dims
= (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_NRM
) ? 3 : 4;
1590 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) ||
1591 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
) ||
1592 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
) ||
1593 (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 4)) {
1595 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1598 /* xmm0 = src.x * src.x */
1599 tmp0
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
1600 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
1603 tmp0
= lp_build_mul( &bld
->bld_base
.base
, tmp0
, tmp0
);
1606 /* xmm0 = xmm0 + src.y * src.y */
1607 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Y
);
1608 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
1611 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1612 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1615 /* xmm0 = xmm0 + src.z * src.z */
1616 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Z
);
1617 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
1620 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1621 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1625 /* xmm0 = xmm0 + src.w * src.w */
1626 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
1627 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
)) {
1630 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
1631 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
1633 /* xmm1 = 1 / sqrt(xmm0) */
1634 tmp1
= lp_build_rsqrt( &bld
->bld_base
.base
, tmp0
);
1635 /* dst.x = xmm1 * src.x */
1636 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
1637 emit_data
->output
[TGSI_CHAN_X
] = lp_build_mul( &bld
->bld_base
.base
, tmp4
, tmp1
);
1639 /* dst.y = xmm1 * src.y */
1640 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
1641 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_mul( &bld
->bld_base
.base
, tmp5
, tmp1
);
1644 /* dst.z = xmm1 * src.z */
1645 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
1646 emit_data
->output
[TGSI_CHAN_Z
] = lp_build_mul( &bld
->bld_base
.base
, tmp6
, tmp1
);
1648 /* dst.w = xmm1 * src.w */
1649 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) && dims
== 4) {
1650 emit_data
->output
[TGSI_CHAN_W
] = lp_build_mul( &bld
->bld_base
.base
, tmp7
, tmp1
);
1655 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 3) {
1656 emit_data
->output
[TGSI_CHAN_W
] = bld
->bld_base
.base
.one
;
1660 static void emit_prologue(struct lp_build_tgsi_context
* bld_base
)
1662 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1663 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
1665 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
1666 LLVMValueRef array_size
=
1667 lp_build_const_int32(gallivm
,
1668 bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] * 4 + 4);
1669 bld
->temps_array
= lp_build_array_alloca(gallivm
,
1670 bld_base
->base
.vec_type
, array_size
,
1674 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
1675 LLVMValueRef array_size
=
1676 lp_build_const_int32(gallivm
,
1677 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] * 4 + 4);
1678 bld
->outputs_array
= lp_build_array_alloca(gallivm
,
1679 bld_base
->base
.vec_type
, array_size
,
1683 /* If we have indirect addressing in inputs we need to copy them into
1684 * our alloca array to be able to iterate over them */
1685 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
1686 unsigned index
, chan
;
1687 LLVMTypeRef vec_type
= bld_base
->base
.vec_type
;
1688 LLVMValueRef array_size
= lp_build_const_int32(gallivm
,
1689 bld_base
->info
->file_max
[TGSI_FILE_INPUT
]*4 + 4);
1690 bld
->inputs_array
= lp_build_array_alloca(gallivm
,
1691 vec_type
, array_size
,
1694 assert(bld_base
->info
->num_inputs
1695 <= bld_base
->info
->file_max
[TGSI_FILE_INPUT
] + 1);
1697 for (index
= 0; index
< bld_base
->info
->num_inputs
; ++index
) {
1698 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1699 LLVMValueRef lindex
=
1700 lp_build_const_int32(gallivm
, index
* 4 + chan
);
1701 LLVMValueRef input_ptr
=
1702 LLVMBuildGEP(gallivm
->builder
, bld
->inputs_array
,
1704 LLVMValueRef value
= bld
->inputs
[index
][chan
];
1706 LLVMBuildStore(gallivm
->builder
, value
, input_ptr
);
1712 static void emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
1714 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1718 emit_dump_temps(bld
);
1721 /* If we have indirect addressing in outputs we need to copy our alloca array
1722 * to the outputs slots specified by the called */
1723 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
1724 unsigned index
, chan
;
1725 assert(bld_base
->info
->num_outputs
<=
1726 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] + 1);
1727 for (index
= 0; index
< bld_base
->info
->num_outputs
; ++index
) {
1728 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
1729 bld
->outputs
[index
][chan
] = lp_get_output_ptr(bld
, index
, chan
);
1736 lp_build_tgsi_soa(struct gallivm_state
*gallivm
,
1737 const struct tgsi_token
*tokens
,
1738 struct lp_type type
,
1739 struct lp_build_mask_context
*mask
,
1740 LLVMValueRef consts_ptr
,
1741 LLVMValueRef system_values_array
,
1742 const LLVMValueRef
*pos
,
1743 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
1744 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
1745 struct lp_build_sampler_soa
*sampler
,
1746 const struct tgsi_shader_info
*info
)
1748 struct lp_build_tgsi_soa_context bld
;
1750 struct lp_type res_type
;
1752 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
1753 memset(&res_type
, 0, sizeof res_type
);
1754 res_type
.width
= type
.width
;
1755 res_type
.length
= type
.length
;
1758 /* Setup build context */
1759 memset(&bld
, 0, sizeof bld
);
1760 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
1761 lp_build_context_init(&bld
.uint_bld
, gallivm
, lp_uint_type(type
));
1762 lp_build_context_init(&bld
.elem_bld
, gallivm
, lp_elem_type(type
));
1765 bld
.inputs
= inputs
;
1766 bld
.outputs
= outputs
;
1767 bld
.consts_ptr
= consts_ptr
;
1768 bld
.sampler
= sampler
;
1769 bld
.bld_base
.info
= info
;
1770 bld
.indirect_files
= info
->indirect_files
;
1772 bld
.bld_base
.soa
= TRUE
;
1773 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
1774 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
1775 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
1776 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
1777 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = emit_fetch_system_value
;
1778 bld
.bld_base
.emit_store
= emit_store
;
1780 bld
.bld_base
.emit_declaration
= lp_emit_declaration_soa
;
1781 bld
.bld_base
.emit_immediate
= lp_emit_immediate_soa
;
1783 bld
.bld_base
.emit_prologue
= emit_prologue
;
1784 bld
.bld_base
.emit_epilogue
= emit_epilogue
;
1786 /* Set opcode actions */
1787 lp_set_default_actions_cpu(&bld
.bld_base
);
1789 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
1790 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNSUB
].emit
= bgnsub_emit
;
1791 bld
.bld_base
.op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
1792 bld
.bld_base
.op_actions
[TGSI_OPCODE_CAL
].emit
= cal_emit
;
1793 bld
.bld_base
.op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
1794 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDX
].emit
= ddx_emit
;
1795 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDY
].emit
= ddy_emit
;
1796 bld
.bld_base
.op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
1797 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
1798 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
1799 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSUB
].emit
= endsub_emit
;
1800 bld
.bld_base
.op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
1801 bld
.bld_base
.op_actions
[TGSI_OPCODE_KIL
].emit
= kil_emit
;
1802 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILP
].emit
= kilp_emit
;
1803 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM
].emit
= nrm_emit
;
1804 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM4
].emit
= nrm_emit
;
1805 bld
.bld_base
.op_actions
[TGSI_OPCODE_RET
].emit
= ret_emit
;
1806 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX
].emit
= tex_emit
;
1807 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB
].emit
= txb_emit
;
1808 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXD
].emit
= txd_emit
;
1809 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL
].emit
= txl_emit
;
1810 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXP
].emit
= txp_emit
;
1812 lp_exec_mask_init(&bld
.exec_mask
, &bld
.bld_base
.base
);
1815 bld
.system_values_array
= system_values_array
;
1817 lp_build_tgsi_llvm(&bld
.bld_base
, tokens
);
1820 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
1821 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
1822 debug_printf("11111111111111111111111111111 \n");
1823 tgsi_dump(tokens
, 0);
1824 lp_debug_dump_value(function
);
1825 debug_printf("2222222222222222222222222222 \n");
1829 LLVMModuleRef module
= LLVMGetGlobalParent(
1830 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
1831 LLVMDumpModule(module
);
1838 * Build up the system values array out of individual values such as
1839 * the instance ID, front-face, primitive ID, etc. The shader info is
1840 * used to determine which system values are needed and where to put
1841 * them in the system values array.
1843 * XXX only instance ID is implemented at this time.
1845 * The system values register file is similar to the constants buffer.
1846 * Example declaration:
1847 * DCL SV[0], INSTANCEID
1848 * Example instruction:
1849 * MOVE foo, SV[0].xxxx;
1851 * \return LLVM float array (interpreted as float [][4])
1854 lp_build_system_values_array(struct gallivm_state
*gallivm
,
1855 const struct tgsi_shader_info
*info
,
1856 LLVMValueRef instance_id
,
1857 LLVMValueRef facing
)
1859 LLVMValueRef size
= lp_build_const_int32(gallivm
, 4 * info
->num_system_values
);
1860 LLVMTypeRef float_t
= LLVMFloatTypeInContext(gallivm
->context
);
1861 LLVMValueRef array
= lp_build_array_alloca(gallivm
, float_t
,
1862 size
, "sysvals_array");
1865 for (i
= 0; i
< info
->num_system_values
; i
++) {
1866 LLVMValueRef index
= lp_build_const_int32(gallivm
, i
* 4);
1867 LLVMValueRef ptr
, value
= 0;
1869 switch (info
->system_value_semantic_name
[i
]) {
1870 case TGSI_SEMANTIC_INSTANCEID
:
1871 /* convert instance ID from int to float */
1872 value
= LLVMBuildSIToFP(gallivm
->builder
, instance_id
, float_t
,
1873 "sysval_instanceid");
1875 case TGSI_SEMANTIC_FACE
:
1878 assert(0 && "unexpected semantic in build_system_values_array()");
1881 ptr
= LLVMBuildGEP(gallivm
->builder
, array
, &index
, 1, "");
1882 LLVMBuildStore(gallivm
->builder
, value
, ptr
);