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"
65 #include "lp_bld_sample.h"
66 #include "lp_bld_struct.h"
68 #define DUMP_GS_EMITS 0
70 static void lp_exec_mask_init(struct lp_exec_mask
*mask
, struct lp_build_context
*bld
)
72 LLVMTypeRef int_type
= LLVMInt32TypeInContext(bld
->gallivm
->context
);
73 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
76 mask
->has_mask
= FALSE
;
77 mask
->ret_in_main
= FALSE
;
78 mask
->cond_stack_size
= 0;
79 mask
->loop_stack_size
= 0;
80 mask
->call_stack_size
= 0;
81 mask
->switch_stack_size
= 0;
83 mask
->int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, mask
->bld
->type
);
84 mask
->exec_mask
= mask
->ret_mask
= mask
->break_mask
= mask
->cont_mask
=
85 mask
->cond_mask
= mask
->switch_mask
=
86 LLVMConstAllOnes(mask
->int_vec_type
);
88 mask
->loop_limiter
= lp_build_alloca(bld
->gallivm
, int_type
, "looplimiter");
92 LLVMConstInt(int_type
, LP_MAX_TGSI_LOOP_ITERATIONS
, false),
96 static void lp_exec_mask_update(struct lp_exec_mask
*mask
)
98 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
100 if (mask
->loop_stack_size
) {
101 /*for loops we need to update the entire mask at runtime */
103 assert(mask
->break_mask
);
104 tmp
= LLVMBuildAnd(builder
,
108 mask
->exec_mask
= LLVMBuildAnd(builder
,
113 mask
->exec_mask
= mask
->cond_mask
;
115 if (mask
->switch_stack_size
) {
116 mask
->exec_mask
= LLVMBuildAnd(builder
,
122 if (mask
->call_stack_size
|| mask
->ret_in_main
) {
123 mask
->exec_mask
= LLVMBuildAnd(builder
,
129 mask
->has_mask
= (mask
->cond_stack_size
> 0 ||
130 mask
->loop_stack_size
> 0 ||
131 mask
->call_stack_size
> 0 ||
132 mask
->switch_stack_size
> 0 ||
136 static void lp_exec_mask_cond_push(struct lp_exec_mask
*mask
,
139 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
141 assert(mask
->cond_stack_size
< LP_MAX_TGSI_NESTING
);
142 if (mask
->cond_stack_size
== 0) {
143 assert(mask
->cond_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
145 mask
->cond_stack
[mask
->cond_stack_size
++] = mask
->cond_mask
;
146 assert(LLVMTypeOf(val
) == mask
->int_vec_type
);
147 mask
->cond_mask
= LLVMBuildAnd(builder
,
151 lp_exec_mask_update(mask
);
154 static void lp_exec_mask_cond_invert(struct lp_exec_mask
*mask
)
156 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
157 LLVMValueRef prev_mask
;
158 LLVMValueRef inv_mask
;
160 assert(mask
->cond_stack_size
);
161 prev_mask
= mask
->cond_stack
[mask
->cond_stack_size
- 1];
162 if (mask
->cond_stack_size
== 1) {
163 assert(prev_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
166 inv_mask
= LLVMBuildNot(builder
, mask
->cond_mask
, "");
168 mask
->cond_mask
= LLVMBuildAnd(builder
,
171 lp_exec_mask_update(mask
);
174 static void lp_exec_mask_cond_pop(struct lp_exec_mask
*mask
)
176 assert(mask
->cond_stack_size
);
177 mask
->cond_mask
= mask
->cond_stack
[--mask
->cond_stack_size
];
178 lp_exec_mask_update(mask
);
181 static void lp_exec_bgnloop(struct lp_exec_mask
*mask
)
183 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
185 if (mask
->loop_stack_size
== 0) {
186 assert(mask
->loop_block
== NULL
);
187 assert(mask
->cont_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
188 assert(mask
->break_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
189 assert(mask
->break_var
== NULL
);
192 assert(mask
->loop_stack_size
< LP_MAX_TGSI_NESTING
);
194 mask
->break_type_stack
[mask
->loop_stack_size
+ mask
->switch_stack_size
] =
196 mask
->break_type
= LP_EXEC_MASK_BREAK_TYPE_LOOP
;
198 mask
->loop_stack
[mask
->loop_stack_size
].loop_block
= mask
->loop_block
;
199 mask
->loop_stack
[mask
->loop_stack_size
].cont_mask
= mask
->cont_mask
;
200 mask
->loop_stack
[mask
->loop_stack_size
].break_mask
= mask
->break_mask
;
201 mask
->loop_stack
[mask
->loop_stack_size
].break_var
= mask
->break_var
;
202 ++mask
->loop_stack_size
;
204 mask
->break_var
= lp_build_alloca(mask
->bld
->gallivm
, mask
->int_vec_type
, "");
205 LLVMBuildStore(builder
, mask
->break_mask
, mask
->break_var
);
207 mask
->loop_block
= lp_build_insert_new_block(mask
->bld
->gallivm
, "bgnloop");
209 LLVMBuildBr(builder
, mask
->loop_block
);
210 LLVMPositionBuilderAtEnd(builder
, mask
->loop_block
);
212 mask
->break_mask
= LLVMBuildLoad(builder
, mask
->break_var
, "");
214 lp_exec_mask_update(mask
);
217 static void lp_exec_break(struct lp_exec_mask
*mask
,
218 struct lp_build_tgsi_context
* bld_base
)
220 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
222 if (mask
->break_type
== LP_EXEC_MASK_BREAK_TYPE_LOOP
) {
223 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
227 mask
->break_mask
= LLVMBuildAnd(builder
,
229 exec_mask
, "break_full");
232 unsigned opcode
= bld_base
->instructions
[bld_base
->pc
+ 1].Instruction
.Opcode
;
233 boolean break_always
= (opcode
== TGSI_OPCODE_ENDSWITCH
||
234 opcode
== TGSI_OPCODE_CASE
);
237 if (mask
->switch_in_default
) {
239 * stop default execution but only if this is an unconditional switch.
240 * (The condition here is not perfect since dead code after break is
241 * allowed but should be sufficient since false negatives are just
242 * unoptimized - so we don't have to pre-evaluate that).
244 if(break_always
&& mask
->switch_pc
) {
245 bld_base
->pc
= mask
->switch_pc
;
251 mask
->switch_mask
= LLVMConstNull(mask
->bld
->int_vec_type
);
254 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
257 mask
->switch_mask
= LLVMBuildAnd(builder
,
259 exec_mask
, "break_switch");
263 lp_exec_mask_update(mask
);
266 static void lp_exec_break_condition(struct lp_exec_mask
*mask
,
269 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
270 LLVMValueRef cond_mask
= LLVMBuildAnd(builder
,
273 cond_mask
= LLVMBuildNot(builder
, cond_mask
, "break_cond");
275 if (mask
->break_type
== LP_EXEC_MASK_BREAK_TYPE_LOOP
) {
276 mask
->break_mask
= LLVMBuildAnd(builder
,
278 cond_mask
, "breakc_full");
281 mask
->switch_mask
= LLVMBuildAnd(builder
,
283 cond_mask
, "breakc_switch");
286 lp_exec_mask_update(mask
);
289 static void lp_exec_continue(struct lp_exec_mask
*mask
)
291 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
292 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
296 mask
->cont_mask
= LLVMBuildAnd(builder
,
300 lp_exec_mask_update(mask
);
304 static void lp_exec_endloop(struct gallivm_state
*gallivm
,
305 struct lp_exec_mask
*mask
)
307 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
308 LLVMBasicBlockRef endloop
;
309 LLVMTypeRef int_type
= LLVMInt32TypeInContext(mask
->bld
->gallivm
->context
);
310 LLVMTypeRef reg_type
= LLVMIntTypeInContext(gallivm
->context
,
311 mask
->bld
->type
.width
*
312 mask
->bld
->type
.length
);
313 LLVMValueRef i1cond
, i2cond
, icond
, limiter
;
315 assert(mask
->break_mask
);
318 * Restore the cont_mask, but don't pop
320 assert(mask
->loop_stack_size
);
321 mask
->cont_mask
= mask
->loop_stack
[mask
->loop_stack_size
- 1].cont_mask
;
322 lp_exec_mask_update(mask
);
325 * Unlike the continue mask, the break_mask must be preserved across loop
328 LLVMBuildStore(builder
, mask
->break_mask
, mask
->break_var
);
330 /* Decrement the loop limiter */
331 limiter
= LLVMBuildLoad(builder
, mask
->loop_limiter
, "");
333 limiter
= LLVMBuildSub(
336 LLVMConstInt(int_type
, 1, false),
339 LLVMBuildStore(builder
, limiter
, mask
->loop_limiter
);
341 /* i1cond = (mask != 0) */
342 i1cond
= LLVMBuildICmp(
345 LLVMBuildBitCast(builder
, mask
->exec_mask
, reg_type
, ""),
346 LLVMConstNull(reg_type
), "i1cond");
348 /* i2cond = (looplimiter > 0) */
349 i2cond
= LLVMBuildICmp(
353 LLVMConstNull(int_type
), "i2cond");
355 /* if( i1cond && i2cond ) */
356 icond
= LLVMBuildAnd(builder
, i1cond
, i2cond
, "");
358 endloop
= lp_build_insert_new_block(mask
->bld
->gallivm
, "endloop");
360 LLVMBuildCondBr(builder
,
361 icond
, mask
->loop_block
, endloop
);
363 LLVMPositionBuilderAtEnd(builder
, endloop
);
365 assert(mask
->loop_stack_size
);
366 --mask
->loop_stack_size
;
367 mask
->loop_block
= mask
->loop_stack
[mask
->loop_stack_size
].loop_block
;
368 mask
->cont_mask
= mask
->loop_stack
[mask
->loop_stack_size
].cont_mask
;
369 mask
->break_mask
= mask
->loop_stack
[mask
->loop_stack_size
].break_mask
;
370 mask
->break_var
= mask
->loop_stack
[mask
->loop_stack_size
].break_var
;
371 mask
->break_type
= mask
->break_type_stack
[mask
->loop_stack_size
+ mask
->switch_stack_size
];
373 lp_exec_mask_update(mask
);
376 static void lp_exec_switch(struct lp_exec_mask
*mask
,
377 LLVMValueRef switchval
)
379 mask
->break_type_stack
[mask
->loop_stack_size
+ mask
->switch_stack_size
] =
381 mask
->break_type
= LP_EXEC_MASK_BREAK_TYPE_SWITCH
;
383 mask
->switch_stack
[mask
->switch_stack_size
].switch_val
= mask
->switch_val
;
384 mask
->switch_stack
[mask
->switch_stack_size
].switch_mask
= mask
->switch_mask
;
385 mask
->switch_stack
[mask
->switch_stack_size
].switch_mask_default
= mask
->switch_mask_default
;
386 mask
->switch_stack
[mask
->switch_stack_size
].switch_in_default
= mask
->switch_in_default
;
387 mask
->switch_stack
[mask
->switch_stack_size
].switch_pc
= mask
->switch_pc
;
388 mask
->switch_stack_size
++;
390 mask
->switch_val
= switchval
;
391 mask
->switch_mask
= LLVMConstNull(mask
->int_vec_type
);
392 mask
->switch_mask_default
= LLVMConstNull(mask
->int_vec_type
);
393 mask
->switch_in_default
= false;
396 lp_exec_mask_update(mask
);
399 static void lp_exec_endswitch(struct lp_exec_mask
*mask
,
400 struct lp_build_tgsi_context
* bld_base
)
402 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
404 /* check if there's deferred default if so do it now */
405 if (mask
->switch_pc
&& !mask
->switch_in_default
) {
406 LLVMValueRef prevmask
, defaultmask
;
408 prevmask
= mask
->switch_stack
[mask
->switch_stack_size
- 1].switch_mask
;
409 defaultmask
= LLVMBuildNot(builder
, mask
->switch_mask_default
, "sw_default_mask");
410 mask
->switch_mask
= LLVMBuildAnd(builder
, prevmask
, defaultmask
, "sw_mask");
411 mask
->switch_in_default
= true;
413 lp_exec_mask_update(mask
);
415 assert(bld_base
->instructions
[mask
->switch_pc
- 1].Instruction
.Opcode
==
416 TGSI_OPCODE_DEFAULT
);
418 tmp_pc
= bld_base
->pc
;
419 bld_base
->pc
= mask
->switch_pc
;
421 * re-purpose switch_pc to point to here again, since we stop execution of
422 * the deferred default after next break.
424 mask
->switch_pc
= tmp_pc
- 1;
429 else if (mask
->switch_pc
&& mask
->switch_in_default
) {
430 assert(bld_base
->pc
== mask
->switch_pc
+ 1);
433 mask
->switch_stack_size
--;
434 mask
->switch_val
= mask
->switch_stack
[mask
->switch_stack_size
].switch_val
;
435 mask
->switch_mask
= mask
->switch_stack
[mask
->switch_stack_size
].switch_mask
;
436 mask
->switch_mask_default
= mask
->switch_stack
[mask
->switch_stack_size
].switch_mask_default
;
437 mask
->switch_in_default
= mask
->switch_stack
[mask
->switch_stack_size
].switch_in_default
;
438 mask
->switch_pc
= mask
->switch_stack
[mask
->switch_stack_size
].switch_pc
;
440 mask
->break_type
= mask
->break_type_stack
[mask
->loop_stack_size
+ mask
->switch_stack_size
];
442 lp_exec_mask_update(mask
);
445 static void lp_exec_case(struct lp_exec_mask
*mask
,
446 LLVMValueRef caseval
)
448 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
450 LLVMValueRef casemask
, prevmask
;
452 /* skipping case mask evaluation here is NOT optional (not in all cases anyway). */
453 if (!mask
->switch_in_default
) {
454 prevmask
= mask
->switch_stack
[mask
->switch_stack_size
- 1].switch_mask
;
455 casemask
= lp_build_cmp(mask
->bld
, PIPE_FUNC_EQUAL
, caseval
, mask
->switch_val
);
456 mask
->switch_mask_default
= LLVMBuildOr(builder
, casemask
,
457 mask
->switch_mask_default
, "sw_default_mask");
458 casemask
= LLVMBuildOr(builder
, casemask
, mask
->switch_mask
, "");
459 mask
->switch_mask
= LLVMBuildAnd(builder
, casemask
, prevmask
, "sw_mask");
461 lp_exec_mask_update(mask
);
466 * Analyse default statement in a switch.
467 * \return true if default is last statement, false otherwise
468 * \param default_pc_start contains pc of instruction to jump to
469 * if default wasn't last but there's no
470 * fallthrough into default.
472 static boolean
default_analyse_is_last(struct lp_exec_mask
*mask
,
473 struct lp_build_tgsi_context
* bld_base
,
474 int *default_pc_start
)
476 unsigned pc
= bld_base
->pc
;
477 unsigned curr_switch_stack
= mask
->switch_stack_size
;
479 /* skip over case statements which are together with default */
480 while (bld_base
->instructions
[pc
].Instruction
.Opcode
== TGSI_OPCODE_CASE
) {
484 while (pc
!= -1 && pc
< bld_base
->num_instructions
) {
485 unsigned opcode
= bld_base
->instructions
[pc
].Instruction
.Opcode
;
487 case TGSI_OPCODE_CASE
:
488 if (curr_switch_stack
== mask
->switch_stack_size
) {
489 *default_pc_start
= pc
- 1;
493 case TGSI_OPCODE_SWITCH
:
496 case TGSI_OPCODE_ENDSWITCH
:
497 if (curr_switch_stack
== mask
->switch_stack_size
) {
498 *default_pc_start
= pc
- 1;
506 /* should never arrive here */
511 static void lp_exec_default(struct lp_exec_mask
*mask
,
512 struct lp_build_tgsi_context
* bld_base
)
514 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
517 boolean default_is_last
;
520 * This is a messy opcode, because it may not be always at the end and
521 * there can be fallthrough in and out of it.
524 default_is_last
= default_analyse_is_last(mask
, bld_base
, &default_exec_pc
);
526 * If it is last statement in switch (note that case statements appearing
527 * "at the same time" as default don't change that) everything is just fine,
528 * update switch mask and go on. This means we can handle default with
529 * fallthrough INTO it without overhead, if it is last.
531 if (default_is_last
) {
532 LLVMValueRef prevmask
, defaultmask
;
533 prevmask
= mask
->switch_stack
[mask
->switch_stack_size
- 1].switch_mask
;
534 defaultmask
= LLVMBuildNot(builder
, mask
->switch_mask_default
, "sw_default_mask");
535 defaultmask
= LLVMBuildOr(builder
, defaultmask
, mask
->switch_mask
, "");
536 mask
->switch_mask
= LLVMBuildAnd(builder
, prevmask
, defaultmask
, "sw_mask");
537 mask
->switch_in_default
= true;
539 lp_exec_mask_update(mask
);
543 * Technically, "case" immediately before default isn't really a
544 * fallthrough, however we still have to count them as such as we
545 * already have updated the masks.
546 * If that happens in practice could add a switch optimizer pass
547 * which just gets rid of all case statements appearing together with
548 * default (or could do switch analysis at switch start time instead).
550 unsigned opcode
= bld_base
->instructions
[bld_base
->pc
- 1].Instruction
.Opcode
;
551 boolean ft_into
= (opcode
!= TGSI_OPCODE_BRK
||
552 opcode
!= TGSI_OPCODE_SWITCH
);
554 * If it is not last statement and there was no fallthrough into it,
555 * we record the PC and continue execution at next case (again, those
556 * case encountered at the same time don't count). At endswitch
557 * time, we update switchmask, and go back executing the code we skipped
558 * until the next break (possibly re-executing some code with changed mask
559 * if there was a fallthrough out of default).
560 * Finally, if it is not last statement and there was a fallthrough into it,
561 * do the same as with the former case, except instead of skipping the code
562 * just execute it without updating the mask, then go back and re-execute.
564 mask
->switch_pc
= bld_base
->pc
;
566 bld_base
->pc
= default_exec_pc
;
572 /* stores val into an address pointed to by dst_ptr.
573 * mask->exec_mask is used to figure out which bits of val
574 * should be stored into the address
575 * (0 means don't store this bit, 1 means do store).
577 static void lp_exec_mask_store(struct lp_exec_mask
*mask
,
578 struct lp_build_context
*bld_store
,
581 LLVMValueRef dst_ptr
)
583 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
585 assert(lp_check_value(bld_store
->type
, val
));
586 assert(LLVMGetTypeKind(LLVMTypeOf(dst_ptr
)) == LLVMPointerTypeKind
);
587 assert(LLVMGetElementType(LLVMTypeOf(dst_ptr
)) == LLVMTypeOf(val
));
589 /* Mix the predicate and execution mask */
590 if (mask
->has_mask
) {
592 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
594 pred
= mask
->exec_mask
;
599 LLVMValueRef res
, dst
;
601 dst
= LLVMBuildLoad(builder
, dst_ptr
, "");
602 res
= lp_build_select(bld_store
, pred
, val
, dst
);
603 LLVMBuildStore(builder
, res
, dst_ptr
);
605 LLVMBuildStore(builder
, val
, dst_ptr
);
608 static void lp_exec_mask_call(struct lp_exec_mask
*mask
,
612 assert(mask
->call_stack_size
< LP_MAX_TGSI_NESTING
);
613 mask
->call_stack
[mask
->call_stack_size
].pc
= *pc
;
614 mask
->call_stack
[mask
->call_stack_size
].ret_mask
= mask
->ret_mask
;
615 mask
->call_stack_size
++;
619 static void lp_exec_mask_ret(struct lp_exec_mask
*mask
, int *pc
)
621 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
622 LLVMValueRef exec_mask
;
624 if (mask
->cond_stack_size
== 0 &&
625 mask
->loop_stack_size
== 0 &&
626 mask
->switch_stack_size
== 0 &&
627 mask
->call_stack_size
== 0) {
628 /* returning from main() */
633 if (mask
->call_stack_size
== 0) {
635 * This requires special handling since we need to ensure
636 * we don't drop the mask even if we have no call stack
637 * (e.g. after a ret in a if clause after the endif)
639 mask
->ret_in_main
= TRUE
;
642 exec_mask
= LLVMBuildNot(builder
,
646 mask
->ret_mask
= LLVMBuildAnd(builder
,
648 exec_mask
, "ret_full");
650 lp_exec_mask_update(mask
);
653 static void lp_exec_mask_bgnsub(struct lp_exec_mask
*mask
)
657 static void lp_exec_mask_endsub(struct lp_exec_mask
*mask
, int *pc
)
659 assert(mask
->call_stack_size
);
660 mask
->call_stack_size
--;
661 *pc
= mask
->call_stack
[mask
->call_stack_size
].pc
;
662 mask
->ret_mask
= mask
->call_stack
[mask
->call_stack_size
].ret_mask
;
663 lp_exec_mask_update(mask
);
668 * Return pointer to a temporary register channel (src or dest).
669 * Note that indirect addressing cannot be handled here.
670 * \param index which temporary register
671 * \param chan which channel of the temp register.
674 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context
*bld
,
678 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
680 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
681 LLVMValueRef lindex
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
, index
* 4 + chan
);
682 return LLVMBuildGEP(builder
, bld
->temps_array
, &lindex
, 1, "");
685 return bld
->temps
[index
][chan
];
690 * Return pointer to a output register channel (src or dest).
691 * Note that indirect addressing cannot be handled here.
692 * \param index which output register
693 * \param chan which channel of the output register.
696 lp_get_output_ptr(struct lp_build_tgsi_soa_context
*bld
,
700 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
702 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
703 LLVMValueRef lindex
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
,
705 return LLVMBuildGEP(builder
, bld
->outputs_array
, &lindex
, 1, "");
708 return bld
->outputs
[index
][chan
];
713 * If we have indirect addressing in outputs copy our alloca array
714 * to the outputs slots specified by the caller to make sure
715 * our outputs are delivered consistently via the same interface.
718 gather_outputs(struct lp_build_tgsi_soa_context
* bld
)
720 if ((bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
721 unsigned index
, chan
;
722 assert(bld
->bld_base
.info
->num_outputs
<=
723 bld
->bld_base
.info
->file_max
[TGSI_FILE_OUTPUT
] + 1);
724 for (index
= 0; index
< bld
->bld_base
.info
->num_outputs
; ++index
) {
725 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
726 bld
->outputs
[index
][chan
] = lp_get_output_ptr(bld
, index
, chan
);
734 * XXX the lp_build_gather() function should be capable of doing this
735 * with a little work.
738 build_gather(struct lp_build_context
*bld
,
739 LLVMValueRef base_ptr
,
740 LLVMValueRef indexes
)
742 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
743 LLVMValueRef res
= bld
->undef
;
747 * Loop over elements of index_vec, load scalar value, insert it into 'res'.
749 for (i
= 0; i
< bld
->type
.length
; i
++) {
750 LLVMValueRef ii
= lp_build_const_int32(bld
->gallivm
, i
);
751 LLVMValueRef index
= LLVMBuildExtractElement(builder
,
753 LLVMValueRef scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
,
754 &index
, 1, "gather_ptr");
755 LLVMValueRef scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
757 res
= LLVMBuildInsertElement(builder
, res
, scalar
, ii
, "");
765 * Scatter/store vector.
768 emit_mask_scatter(struct lp_build_tgsi_soa_context
*bld
,
769 LLVMValueRef base_ptr
,
770 LLVMValueRef indexes
,
772 struct lp_exec_mask
*mask
,
775 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
776 LLVMBuilderRef builder
= gallivm
->builder
;
779 /* Mix the predicate and execution mask */
780 if (mask
->has_mask
) {
782 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
785 pred
= mask
->exec_mask
;
790 * Loop over elements of index_vec, store scalar value.
792 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
793 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
794 LLVMValueRef index
= LLVMBuildExtractElement(builder
, indexes
, ii
, "");
795 LLVMValueRef scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
, &index
, 1, "scatter_ptr");
796 LLVMValueRef val
= LLVMBuildExtractElement(builder
, values
, ii
, "scatter_val");
797 LLVMValueRef scalar_pred
= pred
?
798 LLVMBuildExtractElement(builder
, pred
, ii
, "scatter_pred") : NULL
;
801 lp_build_printf(gallivm
, "scatter %d: val %f at %d %p\n",
802 ii
, val
, index
, scalar_ptr
);
805 LLVMValueRef real_val
, dst_val
;
806 dst_val
= LLVMBuildLoad(builder
, scalar_ptr
, "");
807 real_val
= lp_build_select(&bld
->elem_bld
, scalar_pred
, val
, dst_val
);
808 LLVMBuildStore(builder
, real_val
, scalar_ptr
);
811 LLVMBuildStore(builder
, val
, scalar_ptr
);
818 * Read the current value of the ADDR register, convert the floats to
819 * ints, add the base index and return the vector of offsets.
820 * The offsets will be used to index into the constant buffer or
821 * temporary register file.
824 get_indirect_index(struct lp_build_tgsi_soa_context
*bld
,
825 unsigned reg_file
, unsigned reg_index
,
826 const struct tgsi_ind_register
*indirect_reg
)
828 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
829 struct lp_build_context
*uint_bld
= &bld
->bld_base
.uint_bld
;
830 /* always use X component of address register */
831 unsigned swizzle
= indirect_reg
->Swizzle
;
834 LLVMValueRef max_index
;
837 assert(bld
->indirect_files
& (1 << reg_file
));
839 base
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, reg_index
);
842 switch (indirect_reg
->File
) {
843 case TGSI_FILE_ADDRESS
:
844 rel
= LLVMBuildLoad(builder
,
845 bld
->addr
[indirect_reg
->Index
][swizzle
],
847 /* ADDR LLVM values already have LLVM integer type. */
849 case TGSI_FILE_TEMPORARY
:
850 rel
= lp_get_temp_ptr_soa(bld
, indirect_reg
->Index
, swizzle
);
851 rel
= LLVMBuildLoad(builder
, rel
, "load temp reg");
852 /* TEMP LLVM values always have LLVM float type, but for indirection, the
853 * value actually stored is expected to be an integer */
854 rel
= LLVMBuildBitCast(builder
, rel
, uint_bld
->vec_type
, "");
858 rel
= uint_bld
->zero
;
861 index
= lp_build_add(uint_bld
, base
, rel
);
863 max_index
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
,
865 bld
->bld_base
.info
->file_max
[reg_file
]);
867 assert(!uint_bld
->type
.sign
);
868 index
= lp_build_min(uint_bld
, index
, max_index
);
873 static struct lp_build_context
*
874 stype_to_fetch(struct lp_build_tgsi_context
* bld_base
,
875 enum tgsi_opcode_type stype
)
877 struct lp_build_context
*bld_fetch
;
880 case TGSI_TYPE_FLOAT
:
881 case TGSI_TYPE_UNTYPED
:
882 bld_fetch
= &bld_base
->base
;
884 case TGSI_TYPE_UNSIGNED
:
885 bld_fetch
= &bld_base
->uint_bld
;
887 case TGSI_TYPE_SIGNED
:
888 bld_fetch
= &bld_base
->int_bld
;
891 case TGSI_TYPE_DOUBLE
:
902 struct lp_build_tgsi_context
* bld_base
,
903 const struct tgsi_full_src_register
* reg
,
904 enum tgsi_opcode_type stype
,
907 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
908 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
909 LLVMBuilderRef builder
= gallivm
->builder
;
910 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
911 LLVMValueRef indirect_index
= NULL
;
912 unsigned dimension
= 0;
913 LLVMValueRef dimension_index
;
914 LLVMValueRef consts_ptr
;
917 /* XXX: Handle fetching xyzw components as a vector */
918 assert(swizzle
!= ~0);
920 if (reg
->Register
.Dimension
) {
921 assert(!reg
->Dimension
.Indirect
);
922 dimension
= reg
->Dimension
.Index
;
923 assert(dimension
< LP_MAX_TGSI_CONST_BUFFERS
);
926 dimension_index
= lp_build_const_int32(gallivm
, dimension
);
927 consts_ptr
= lp_build_array_get(gallivm
, bld
->consts_ptr
, dimension_index
);
929 if (reg
->Register
.Indirect
) {
930 indirect_index
= get_indirect_index(bld
,
936 if (reg
->Register
.Indirect
) {
937 LLVMValueRef swizzle_vec
=
938 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, swizzle
);
939 LLVMValueRef index_vec
; /* index into the const buffer */
941 /* index_vec = indirect_index * 4 + swizzle */
942 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
943 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
945 /* Gather values from the constant buffer */
946 res
= build_gather(&bld_base
->base
, consts_ptr
, index_vec
);
949 LLVMValueRef index
; /* index into the const buffer */
950 LLVMValueRef scalar
, scalar_ptr
;
952 index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
*4 + swizzle
);
954 scalar_ptr
= LLVMBuildGEP(builder
, consts_ptr
,
956 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
957 res
= lp_build_broadcast_scalar(&bld_base
->base
, scalar
);
960 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
) {
961 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
962 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
968 emit_fetch_immediate(
969 struct lp_build_tgsi_context
* bld_base
,
970 const struct tgsi_full_src_register
* reg
,
971 enum tgsi_opcode_type stype
,
974 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
975 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
976 LLVMBuilderRef builder
= gallivm
->builder
;
977 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
978 struct lp_build_context
*float_bld
= &bld_base
->base
;
979 LLVMValueRef res
= NULL
;
980 LLVMValueRef indirect_index
= NULL
;
982 if (reg
->Register
.Indirect
) {
983 indirect_index
= get_indirect_index(bld
,
989 if (reg
->Register
.Indirect
) {
990 LLVMValueRef swizzle_vec
=
991 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
,
992 uint_bld
->type
, swizzle
);
993 LLVMValueRef length_vec
=
994 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
,
995 bld
->bld_base
.base
.type
.length
);
996 LLVMValueRef index_vec
; /* index into the const buffer */
997 LLVMValueRef imms_array
;
998 LLVMValueRef pixel_offsets
;
999 LLVMValueRef offsets
[LP_MAX_VECTOR_LENGTH
];
1000 LLVMTypeRef float4_ptr_type
;
1003 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1004 for (i
= 0; i
< float_bld
->type
.length
; i
++) {
1005 offsets
[i
] = lp_build_const_int32(gallivm
, i
);
1007 pixel_offsets
= LLVMConstVector(offsets
, float_bld
->type
.length
);
1009 /* index_vec = (indirect_index * 4 + swizzle) * length */
1010 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1011 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
1012 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1013 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1015 /* cast imms_array pointer to float* */
1016 float4_ptr_type
= LLVMPointerType(
1017 LLVMFloatTypeInContext(bld
->bld_base
.base
.gallivm
->context
), 0);
1018 imms_array
= LLVMBuildBitCast(builder
, bld
->imms_array
,
1019 float4_ptr_type
, "");
1021 /* Gather values from the temporary register array */
1022 res
= build_gather(&bld_base
->base
, imms_array
, index_vec
);
1025 res
= bld
->immediates
[reg
->Register
.Index
][swizzle
];
1028 if (stype
== TGSI_TYPE_UNSIGNED
) {
1029 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1030 } else if (stype
== TGSI_TYPE_SIGNED
) {
1031 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1038 struct lp_build_tgsi_context
* bld_base
,
1039 const struct tgsi_full_src_register
* reg
,
1040 enum tgsi_opcode_type stype
,
1043 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1044 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1045 LLVMBuilderRef builder
= gallivm
->builder
;
1046 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
1047 LLVMValueRef indirect_index
= NULL
;
1050 if (reg
->Register
.Indirect
) {
1051 indirect_index
= get_indirect_index(bld
,
1053 reg
->Register
.Index
,
1057 if (reg
->Register
.Indirect
) {
1058 LLVMValueRef swizzle_vec
=
1059 lp_build_const_int_vec(gallivm
, uint_bld
->type
, swizzle
);
1060 LLVMValueRef length_vec
=
1061 lp_build_const_int_vec(gallivm
, uint_bld
->type
, bld
->bld_base
.base
.type
.length
);
1062 LLVMValueRef index_vec
; /* index into the const buffer */
1063 LLVMValueRef inputs_array
;
1064 LLVMTypeRef float4_ptr_type
;
1066 /* index_vec = (indirect_index * 4 + swizzle) * length */
1067 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1068 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
1069 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1071 /* cast inputs_array pointer to float* */
1072 float4_ptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1073 inputs_array
= LLVMBuildBitCast(builder
, bld
->inputs_array
,
1074 float4_ptr_type
, "");
1076 /* Gather values from the temporary register array */
1077 res
= build_gather(&bld_base
->base
, inputs_array
, index_vec
);
1079 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
1080 LLVMValueRef lindex
= lp_build_const_int32(gallivm
,
1081 reg
->Register
.Index
* 4 + swizzle
);
1082 LLVMValueRef input_ptr
= LLVMBuildGEP(builder
,
1083 bld
->inputs_array
, &lindex
, 1, "");
1084 res
= LLVMBuildLoad(builder
, input_ptr
, "");
1087 res
= bld
->inputs
[reg
->Register
.Index
][swizzle
];
1093 if (stype
== TGSI_TYPE_UNSIGNED
) {
1094 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1095 } else if (stype
== TGSI_TYPE_SIGNED
) {
1096 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1104 emit_fetch_gs_input(
1105 struct lp_build_tgsi_context
* bld_base
,
1106 const struct tgsi_full_src_register
* reg
,
1107 enum tgsi_opcode_type stype
,
1110 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1111 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1112 LLVMBuilderRef builder
= gallivm
->builder
;
1113 LLVMValueRef attrib_index
= NULL
;
1114 LLVMValueRef vertex_index
= NULL
;
1115 LLVMValueRef swizzle_index
= lp_build_const_int32(gallivm
, swizzle
);
1118 if (reg
->Register
.Indirect
) {
1119 attrib_index
= get_indirect_index(bld
,
1121 reg
->Register
.Index
,
1124 attrib_index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
);
1127 if (reg
->Dimension
.Indirect
) {
1128 vertex_index
= get_indirect_index(bld
,
1130 reg
->Dimension
.Index
,
1133 vertex_index
= lp_build_const_int32(gallivm
, reg
->Dimension
.Index
);
1136 res
= bld
->gs_iface
->fetch_input(bld
->gs_iface
, bld_base
,
1137 reg
->Dimension
.Indirect
,
1138 vertex_index
, attrib_index
,
1143 if (stype
== TGSI_TYPE_UNSIGNED
) {
1144 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1145 } else if (stype
== TGSI_TYPE_SIGNED
) {
1146 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1153 emit_fetch_temporary(
1154 struct lp_build_tgsi_context
* bld_base
,
1155 const struct tgsi_full_src_register
* reg
,
1156 enum tgsi_opcode_type stype
,
1159 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1160 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1161 LLVMBuilderRef builder
= gallivm
->builder
;
1162 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
1163 struct lp_build_context
*float_bld
= &bld_base
->base
;
1164 LLVMValueRef indirect_index
= NULL
;
1167 if (reg
->Register
.Indirect
) {
1168 indirect_index
= get_indirect_index(bld
,
1170 reg
->Register
.Index
,
1174 if (reg
->Register
.Indirect
) {
1175 LLVMValueRef swizzle_vec
=
1176 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, swizzle
);
1177 LLVMValueRef length_vec
=
1178 lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
,
1179 bld
->bld_base
.base
.type
.length
);
1180 LLVMValueRef index_vec
; /* index into the const buffer */
1181 LLVMValueRef temps_array
;
1182 LLVMValueRef pixel_offsets
;
1183 LLVMValueRef offsets
[LP_MAX_VECTOR_LENGTH
];
1184 LLVMTypeRef float4_ptr_type
;
1187 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1188 for (i
= 0; i
< float_bld
->type
.length
; i
++) {
1189 offsets
[i
] = lp_build_const_int32(gallivm
, i
);
1191 pixel_offsets
= LLVMConstVector(offsets
, float_bld
->type
.length
);
1193 /* index_vec = (indirect_index * 4 + swizzle) * length */
1194 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1195 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
1196 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1197 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1199 /* cast temps_array pointer to float* */
1200 float4_ptr_type
= LLVMPointerType(LLVMFloatTypeInContext(bld
->bld_base
.base
.gallivm
->context
), 0);
1201 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
,
1202 float4_ptr_type
, "");
1204 /* Gather values from the temporary register array */
1205 res
= build_gather(&bld_base
->base
, temps_array
, index_vec
);
1208 LLVMValueRef temp_ptr
;
1209 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
1210 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
1213 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
) {
1214 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1215 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1222 emit_fetch_system_value(
1223 struct lp_build_tgsi_context
* bld_base
,
1224 const struct tgsi_full_src_register
* reg
,
1225 enum tgsi_opcode_type stype
,
1228 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1229 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1230 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
1231 LLVMBuilderRef builder
= gallivm
->builder
;
1233 enum tgsi_opcode_type atype
; // Actual type of the value
1235 assert(!reg
->Register
.Indirect
);
1237 switch (info
->system_value_semantic_name
[reg
->Register
.Index
]) {
1238 case TGSI_SEMANTIC_INSTANCEID
:
1239 res
= lp_build_broadcast_scalar(&bld_base
->uint_bld
, bld
->system_values
.instance_id
);
1240 atype
= TGSI_TYPE_UNSIGNED
;
1243 case TGSI_SEMANTIC_VERTEXID
:
1244 res
= bld
->system_values
.vertex_id
;
1245 atype
= TGSI_TYPE_UNSIGNED
;
1248 case TGSI_SEMANTIC_PRIMID
:
1249 res
= bld
->system_values
.prim_id
;
1250 atype
= TGSI_TYPE_UNSIGNED
;
1254 assert(!"unexpected semantic in emit_fetch_system_value");
1255 res
= bld_base
->base
.zero
;
1256 atype
= TGSI_TYPE_FLOAT
;
1260 if (atype
!= stype
) {
1261 if (stype
== TGSI_TYPE_FLOAT
) {
1262 res
= LLVMBuildBitCast(builder
, res
, bld_base
->base
.vec_type
, "");
1263 } else if (stype
== TGSI_TYPE_UNSIGNED
) {
1264 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1265 } else if (stype
== TGSI_TYPE_SIGNED
) {
1266 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1274 * Register fetch with derivatives.
1278 struct lp_build_tgsi_soa_context
*bld
,
1287 /* TODO: use interpolation coeffs for inputs */
1290 *ddx
= lp_build_ddx(&bld
->bld_base
.base
, src
);
1293 *ddy
= lp_build_ddy(&bld
->bld_base
.base
, src
);
1301 emit_fetch_predicate(
1302 struct lp_build_tgsi_soa_context
*bld
,
1303 const struct tgsi_full_instruction
*inst
,
1306 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1308 unsigned char swizzles
[4];
1309 LLVMValueRef unswizzled
[4] = {NULL
, NULL
, NULL
, NULL
};
1313 if (!inst
->Instruction
.Predicate
) {
1314 TGSI_FOR_EACH_CHANNEL( chan
) {
1320 swizzles
[0] = inst
->Predicate
.SwizzleX
;
1321 swizzles
[1] = inst
->Predicate
.SwizzleY
;
1322 swizzles
[2] = inst
->Predicate
.SwizzleZ
;
1323 swizzles
[3] = inst
->Predicate
.SwizzleW
;
1325 index
= inst
->Predicate
.Index
;
1326 assert(index
< LP_MAX_TGSI_PREDS
);
1328 TGSI_FOR_EACH_CHANNEL( chan
) {
1329 unsigned swizzle
= swizzles
[chan
];
1332 * Only fetch the predicate register channels that are actually listed
1335 if (!unswizzled
[swizzle
]) {
1336 value
= LLVMBuildLoad(builder
,
1337 bld
->preds
[index
][swizzle
], "");
1340 * Convert the value to an integer mask.
1342 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
1343 * is needlessly causing two comparisons due to storing the intermediate
1344 * result as float vector instead of an integer mask vector.
1346 value
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
1347 bld
->bld_base
.base
.type
,
1350 bld
->bld_base
.base
.zero
);
1351 if (inst
->Predicate
.Negate
) {
1352 value
= LLVMBuildNot(builder
, value
, "");
1355 unswizzled
[swizzle
] = value
;
1357 value
= unswizzled
[swizzle
];
1369 struct lp_build_tgsi_context
*bld_base
,
1370 const struct tgsi_full_instruction
*inst
,
1372 unsigned chan_index
,
1376 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1377 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1378 LLVMBuilderRef builder
= gallivm
->builder
;
1379 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
1380 struct lp_build_context
*float_bld
= &bld_base
->base
;
1381 struct lp_build_context
*int_bld
= &bld_base
->int_bld
;
1382 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
1383 LLVMValueRef indirect_index
= NULL
;
1384 enum tgsi_opcode_type dtype
= tgsi_opcode_infer_dst_type(inst
->Instruction
.Opcode
);
1389 * It is always assumed to be float.
1391 switch( inst
->Instruction
.Saturate
) {
1395 case TGSI_SAT_ZERO_ONE
:
1396 assert(dtype
== TGSI_TYPE_FLOAT
||
1397 dtype
== TGSI_TYPE_UNTYPED
);
1398 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1399 value
= lp_build_max_ext(float_bld
, value
, float_bld
->zero
,
1400 GALLIVM_NAN_RETURN_SECOND
);
1401 value
= lp_build_min_ext(float_bld
, value
, float_bld
->one
,
1402 GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1405 case TGSI_SAT_MINUS_PLUS_ONE
:
1406 assert(dtype
== TGSI_TYPE_FLOAT
||
1407 dtype
== TGSI_TYPE_UNTYPED
);
1408 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1409 value
= lp_build_max_ext(float_bld
, value
,
1410 lp_build_const_vec(gallivm
, float_bld
->type
, -1.0),
1411 GALLIVM_NAN_RETURN_SECOND
);
1412 value
= lp_build_min_ext(float_bld
, value
, float_bld
->one
,
1413 GALLIVM_NAN_BEHAVIOR_UNDEFINED
);
1420 if (reg
->Register
.Indirect
) {
1421 indirect_index
= get_indirect_index(bld
,
1423 reg
->Register
.Index
,
1426 assert(reg
->Register
.Index
<=
1427 bld_base
->info
->file_max
[reg
->Register
.File
]);
1430 switch( reg
->Register
.File
) {
1431 case TGSI_FILE_OUTPUT
:
1432 /* Outputs are always stored as floats */
1433 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1435 if (reg
->Register
.Indirect
) {
1436 LLVMValueRef chan_vec
=
1437 lp_build_const_int_vec(gallivm
, uint_bld
->type
, chan_index
);
1438 LLVMValueRef length_vec
=
1439 lp_build_const_int_vec(gallivm
, uint_bld
->type
, float_bld
->type
.length
);
1440 LLVMValueRef index_vec
; /* indexes into the temp registers */
1441 LLVMValueRef outputs_array
;
1442 LLVMValueRef pixel_offsets
;
1443 LLVMTypeRef float_ptr_type
;
1446 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1447 pixel_offsets
= uint_bld
->undef
;
1448 for (i
= 0; i
< float_bld
->type
.length
; i
++) {
1449 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1450 pixel_offsets
= LLVMBuildInsertElement(builder
, pixel_offsets
,
1454 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1455 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1456 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
1457 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1458 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1461 LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1462 outputs_array
= LLVMBuildBitCast(builder
, bld
->outputs_array
,
1463 float_ptr_type
, "");
1465 /* Scatter store values into temp registers */
1466 emit_mask_scatter(bld
, outputs_array
, index_vec
, value
,
1467 &bld
->exec_mask
, pred
);
1470 LLVMValueRef out_ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
,
1472 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
, out_ptr
);
1476 case TGSI_FILE_TEMPORARY
:
1477 /* Temporaries are always stored as floats */
1478 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1480 if (reg
->Register
.Indirect
) {
1481 LLVMValueRef chan_vec
=
1482 lp_build_const_int_vec(gallivm
, uint_bld
->type
, chan_index
);
1483 LLVMValueRef length_vec
=
1484 lp_build_const_int_vec(gallivm
, uint_bld
->type
,
1485 float_bld
->type
.length
);
1486 LLVMValueRef index_vec
; /* indexes into the temp registers */
1487 LLVMValueRef temps_array
;
1488 LLVMValueRef pixel_offsets
;
1489 LLVMTypeRef float_ptr_type
;
1492 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1493 pixel_offsets
= uint_bld
->undef
;
1494 for (i
= 0; i
< float_bld
->type
.length
; i
++) {
1495 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1496 pixel_offsets
= LLVMBuildInsertElement(builder
, pixel_offsets
,
1500 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1501 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1502 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
1503 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1504 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1507 LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1508 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
,
1509 float_ptr_type
, "");
1511 /* Scatter store values into temp registers */
1512 emit_mask_scatter(bld
, temps_array
, index_vec
, value
,
1513 &bld
->exec_mask
, pred
);
1516 LLVMValueRef temp_ptr
;
1517 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
,
1519 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
, temp_ptr
);
1523 case TGSI_FILE_ADDRESS
:
1524 assert(dtype
== TGSI_TYPE_SIGNED
);
1525 assert(LLVMTypeOf(value
) == int_bld
->vec_type
);
1526 value
= LLVMBuildBitCast(builder
, value
, int_bld
->vec_type
, "");
1527 lp_exec_mask_store(&bld
->exec_mask
, int_bld
, pred
, value
,
1528 bld
->addr
[reg
->Register
.Index
][chan_index
]);
1531 case TGSI_FILE_PREDICATE
:
1532 assert(LLVMTypeOf(value
) == float_bld
->vec_type
);
1533 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1534 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
,
1535 bld
->preds
[reg
->Register
.Index
][chan_index
]);
1547 struct lp_build_tgsi_context
* bld_base
,
1548 const struct tgsi_full_instruction
* inst
,
1549 const struct tgsi_opcode_info
* info
,
1550 LLVMValueRef dst
[4])
1553 unsigned chan_index
;
1554 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1557 LLVMValueRef pred
[TGSI_NUM_CHANNELS
];
1559 emit_fetch_predicate( bld
, inst
, pred
);
1561 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1562 emit_store_chan(bld_base
, inst
, 0, chan_index
, pred
[chan_index
], dst
[chan_index
]);
1568 tgsi_to_pipe_tex_target(unsigned tgsi_target
)
1570 switch (tgsi_target
) {
1571 case TGSI_TEXTURE_BUFFER
:
1573 case TGSI_TEXTURE_1D
:
1574 case TGSI_TEXTURE_SHADOW1D
:
1575 return PIPE_TEXTURE_1D
;
1576 case TGSI_TEXTURE_2D
:
1577 case TGSI_TEXTURE_SHADOW2D
:
1578 case TGSI_TEXTURE_2D_MSAA
:
1579 return PIPE_TEXTURE_2D
;
1580 case TGSI_TEXTURE_3D
:
1581 return PIPE_TEXTURE_3D
;
1582 case TGSI_TEXTURE_CUBE
:
1583 case TGSI_TEXTURE_SHADOWCUBE
:
1584 return PIPE_TEXTURE_CUBE
;
1585 case TGSI_TEXTURE_RECT
:
1586 case TGSI_TEXTURE_SHADOWRECT
:
1587 return PIPE_TEXTURE_RECT
;
1588 case TGSI_TEXTURE_1D_ARRAY
:
1589 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1590 return PIPE_TEXTURE_1D_ARRAY
;
1591 case TGSI_TEXTURE_2D_ARRAY
:
1592 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1593 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
1594 return PIPE_TEXTURE_2D_ARRAY
;
1595 case TGSI_TEXTURE_CUBE_ARRAY
:
1596 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
1597 return PIPE_TEXTURE_CUBE_ARRAY
;
1605 * High-level instruction translators.
1609 emit_tex( struct lp_build_tgsi_soa_context
*bld
,
1610 const struct tgsi_full_instruction
*inst
,
1611 enum lp_build_tex_modifier modifier
,
1612 LLVMValueRef
*texel
)
1615 LLVMValueRef lod_bias
, explicit_lod
;
1616 LLVMValueRef oow
= NULL
;
1617 LLVMValueRef coords
[4];
1618 LLVMValueRef offsets
[3] = { NULL
};
1619 struct lp_derivatives derivs
;
1620 struct lp_derivatives
*deriv_ptr
= NULL
;
1622 unsigned num_coords
, num_derivs
, num_offsets
;
1625 if (!bld
->sampler
) {
1626 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1627 for (i
= 0; i
< 4; i
++) {
1628 texel
[i
] = bld
->bld_base
.base
.undef
;
1633 switch (inst
->Texture
.Texture
) {
1634 case TGSI_TEXTURE_1D
:
1639 case TGSI_TEXTURE_1D_ARRAY
:
1644 case TGSI_TEXTURE_2D
:
1645 case TGSI_TEXTURE_RECT
:
1650 case TGSI_TEXTURE_SHADOW1D
:
1651 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1656 case TGSI_TEXTURE_SHADOW2D
:
1657 case TGSI_TEXTURE_SHADOWRECT
:
1658 case TGSI_TEXTURE_2D_ARRAY
:
1663 case TGSI_TEXTURE_CUBE
:
1668 case TGSI_TEXTURE_3D
:
1673 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1678 case TGSI_TEXTURE_SHADOWCUBE
:
1688 /* Note lod and especially projected are illegal in a LOT of cases */
1689 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
1690 assert(num_coords
< 4);
1691 lod_bias
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1692 explicit_lod
= NULL
;
1694 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
1695 assert(num_coords
< 4);
1697 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1701 explicit_lod
= NULL
;
1704 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
) {
1705 assert(num_coords
< 4);
1706 oow
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1707 oow
= lp_build_rcp(&bld
->bld_base
.base
, oow
);
1710 for (i
= 0; i
< num_coords
; i
++) {
1711 coords
[i
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, i
);
1712 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
1713 coords
[i
] = lp_build_mul(&bld
->bld_base
.base
, coords
[i
], oow
);
1715 for (i
= num_coords
; i
< 4; i
++) {
1716 coords
[i
] = bld
->bld_base
.base
.undef
;
1719 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
1721 for (dim
= 0; dim
< num_derivs
; ++dim
) {
1722 derivs
.ddx
[dim
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 1, dim
);
1723 derivs
.ddy
[dim
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 2, dim
);
1725 deriv_ptr
= &derivs
;
1726 unit
= inst
->Src
[3].Register
.Index
;
1728 unit
= inst
->Src
[1].Register
.Index
;
1731 /* some advanced gather instructions (txgo) would require 4 offsets */
1732 if (inst
->Texture
.NumOffsets
== 1) {
1734 for (dim
= 0; dim
< num_offsets
; dim
++) {
1735 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
1739 /* TODO: use scalar lod if explicit_lod, lod_bias or derivs are broadcasted scalars */
1740 scalar_lod
= bld
->bld_base
.info
->processor
== TGSI_PROCESSOR_FRAGMENT
;
1742 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
1743 bld
->bld_base
.base
.gallivm
,
1744 bld
->bld_base
.base
.type
,
1750 lod_bias
, explicit_lod
, scalar_lod
,
1755 emit_sample(struct lp_build_tgsi_soa_context
*bld
,
1756 const struct tgsi_full_instruction
*inst
,
1757 enum lp_build_tex_modifier modifier
,
1759 LLVMValueRef
*texel
)
1761 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1762 unsigned texture_unit
, sampler_unit
;
1763 LLVMValueRef lod_bias
, explicit_lod
;
1764 LLVMValueRef coords
[4];
1765 LLVMValueRef offsets
[3] = { NULL
};
1766 struct lp_derivatives derivs
;
1767 struct lp_derivatives
*deriv_ptr
= NULL
;
1769 unsigned num_coords
, num_offsets
, num_derivs
;
1772 if (!bld
->sampler
) {
1773 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1774 for (i
= 0; i
< 4; i
++) {
1775 texel
[i
] = bld
->bld_base
.base
.undef
;
1781 * unlike old-style tex opcodes the texture/sampler indices
1782 * always come from src1 and src2 respectively.
1784 texture_unit
= inst
->Src
[1].Register
.Index
;
1785 sampler_unit
= inst
->Src
[2].Register
.Index
;
1788 * Note inst->Texture.Texture will contain the number of offsets,
1789 * however the target information is NOT there and comes from the
1790 * declared sampler views instead.
1792 switch (bld
->sv
[texture_unit
].Resource
) {
1793 case TGSI_TEXTURE_1D
:
1798 case TGSI_TEXTURE_1D_ARRAY
:
1803 case TGSI_TEXTURE_2D
:
1804 case TGSI_TEXTURE_RECT
:
1809 case TGSI_TEXTURE_2D_ARRAY
:
1814 case TGSI_TEXTURE_CUBE
:
1819 case TGSI_TEXTURE_3D
:
1824 case TGSI_TEXTURE_CUBE_ARRAY
:
1834 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
1835 lod_bias
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 3, 0 );
1836 explicit_lod
= NULL
;
1838 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
1840 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 3, 0 );
1842 else if (modifier
== LP_BLD_TEX_MODIFIER_LOD_ZERO
) {
1844 /* XXX might be better to explicitly pass the level zero information */
1845 explicit_lod
= lp_build_const_vec(gallivm
, bld
->bld_base
.base
.type
, 0.0F
);
1849 explicit_lod
= NULL
;
1852 for (i
= 0; i
< num_coords
; i
++) {
1853 coords
[i
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, i
);
1855 for (i
= num_coords
; i
< 4; i
++) {
1856 coords
[i
] = bld
->bld_base
.base
.undef
;
1859 * XXX: whack shadow comparison value into place.
1860 * Should probably fix the interface for separate value
1861 * (it will not work for cube arrays if it is part of coords).
1864 unsigned c_coord
= num_coords
> 2 ? 3 : 2;
1865 assert(num_coords
< 4);
1866 coords
[c_coord
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 3, 0 );
1869 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
1871 for (dim
= 0; dim
< num_derivs
; ++dim
) {
1872 derivs
.ddx
[dim
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 3, dim
);
1873 derivs
.ddy
[dim
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 4, dim
);
1875 deriv_ptr
= &derivs
;
1878 /* some advanced gather instructions (txgo) would require 4 offsets */
1879 if (inst
->Texture
.NumOffsets
== 1) {
1881 for (dim
= 0; dim
< num_offsets
; dim
++) {
1882 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
1886 /* TODO: use scalar lod if explicit_lod, lod_bias or derivs are broadcasted scalars */
1887 scalar_lod
= bld
->bld_base
.info
->processor
== TGSI_PROCESSOR_FRAGMENT
;
1889 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
1890 bld
->bld_base
.base
.gallivm
,
1891 bld
->bld_base
.base
.type
,
1893 texture_unit
, sampler_unit
,
1897 lod_bias
, explicit_lod
, scalar_lod
,
1900 if (inst
->Src
[1].Register
.SwizzleX
!= PIPE_SWIZZLE_RED
||
1901 inst
->Src
[1].Register
.SwizzleY
!= PIPE_SWIZZLE_GREEN
||
1902 inst
->Src
[1].Register
.SwizzleZ
!= PIPE_SWIZZLE_BLUE
||
1903 inst
->Src
[1].Register
.SwizzleW
!= PIPE_SWIZZLE_ALPHA
) {
1904 unsigned char swizzles
[4];
1905 swizzles
[0] = inst
->Src
[1].Register
.SwizzleX
;
1906 swizzles
[1] = inst
->Src
[1].Register
.SwizzleY
;
1907 swizzles
[2] = inst
->Src
[1].Register
.SwizzleZ
;
1908 swizzles
[3] = inst
->Src
[1].Register
.SwizzleW
;
1910 lp_build_swizzle_soa_inplace(&bld
->bld_base
.base
, texel
, swizzles
);
1915 emit_fetch_texels( struct lp_build_tgsi_soa_context
*bld
,
1916 const struct tgsi_full_instruction
*inst
,
1917 LLVMValueRef
*texel
,
1920 unsigned unit
, target
;
1921 LLVMValueRef coord_undef
= LLVMGetUndef(bld
->bld_base
.base
.int_vec_type
);
1922 LLVMValueRef explicit_lod
= NULL
;
1923 LLVMValueRef coords
[3];
1924 LLVMValueRef offsets
[3] = { NULL
};
1926 unsigned num_coords
;
1930 if (!bld
->sampler
) {
1931 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1932 for (i
= 0; i
< 4; i
++) {
1933 texel
[i
] = coord_undef
;
1938 unit
= inst
->Src
[1].Register
.Index
;
1941 target
= bld
->sv
[unit
].Resource
;
1944 target
= inst
->Texture
.Texture
;
1948 case TGSI_TEXTURE_1D
:
1949 case TGSI_TEXTURE_BUFFER
:
1953 case TGSI_TEXTURE_1D_ARRAY
:
1957 case TGSI_TEXTURE_2D
:
1958 case TGSI_TEXTURE_RECT
:
1962 case TGSI_TEXTURE_2D_ARRAY
:
1966 case TGSI_TEXTURE_3D
:
1975 /* always have lod except for buffers ? */
1976 if (target
!= TGSI_TEXTURE_BUFFER
) {
1977 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 3 );
1980 for (i
= 0; i
< num_coords
; i
++) {
1981 coords
[i
] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, i
);
1983 for (i
= num_coords
; i
< 3; i
++) {
1984 coords
[i
] = coord_undef
;
1987 if (inst
->Texture
.NumOffsets
== 1) {
1989 for (dim
= 0; dim
< dims
; dim
++) {
1990 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
1994 /* TODO: use scalar lod if explicit_lod is broadcasted scalar */
1995 scalar_lod
= bld
->bld_base
.info
->processor
== TGSI_PROCESSOR_FRAGMENT
;
1997 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
1998 bld
->bld_base
.base
.gallivm
,
1999 bld
->bld_base
.base
.type
,
2005 NULL
, explicit_lod
, scalar_lod
,
2009 (inst
->Src
[1].Register
.SwizzleX
!= PIPE_SWIZZLE_RED
||
2010 inst
->Src
[1].Register
.SwizzleY
!= PIPE_SWIZZLE_GREEN
||
2011 inst
->Src
[1].Register
.SwizzleZ
!= PIPE_SWIZZLE_BLUE
||
2012 inst
->Src
[1].Register
.SwizzleW
!= PIPE_SWIZZLE_ALPHA
)) {
2013 unsigned char swizzles
[4];
2014 swizzles
[0] = inst
->Src
[1].Register
.SwizzleX
;
2015 swizzles
[1] = inst
->Src
[1].Register
.SwizzleY
;
2016 swizzles
[2] = inst
->Src
[1].Register
.SwizzleZ
;
2017 swizzles
[3] = inst
->Src
[1].Register
.SwizzleW
;
2019 lp_build_swizzle_soa_inplace(&bld
->bld_base
.base
, texel
, swizzles
);
2024 emit_size_query( struct lp_build_tgsi_soa_context
*bld
,
2025 const struct tgsi_full_instruction
*inst
,
2026 LLVMValueRef
*sizes_out
,
2027 boolean is_sviewinfo
)
2029 LLVMValueRef explicit_lod
;
2033 unsigned unit
= inst
->Src
[1].Register
.Index
;
2034 unsigned target
, pipe_target
;
2037 target
= bld
->sv
[unit
].Resource
;
2040 target
= inst
->Texture
.Texture
;
2043 case TGSI_TEXTURE_BUFFER
:
2044 case TGSI_TEXTURE_RECT
:
2045 case TGSI_TEXTURE_SHADOWRECT
:
2053 if (!bld
->sampler
) {
2054 _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
2055 for (i
= 0; i
< 4; i
++)
2056 sizes_out
[i
] = bld
->bld_base
.int_bld
.undef
;
2061 explicit_lod
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0, 0 );
2063 explicit_lod
= NULL
;
2065 pipe_target
= tgsi_to_pipe_tex_target(target
);
2067 /* TODO: use scalar lod if explicit_lod is broadcasted scalar */
2068 scalar_lod
= bld
->bld_base
.info
->processor
== TGSI_PROCESSOR_FRAGMENT
;
2070 bld
->sampler
->emit_size_query(bld
->sampler
,
2071 bld
->bld_base
.base
.gallivm
,
2072 bld
->bld_base
.int_bld
.type
,
2081 near_end_of_shader(struct lp_build_tgsi_soa_context
*bld
,
2086 for (i
= 0; i
< 5; i
++) {
2089 if (pc
+ i
>= bld
->bld_base
.info
->num_instructions
)
2092 opcode
= bld
->bld_base
.instructions
[pc
+ i
].Instruction
.Opcode
;
2094 if (opcode
== TGSI_OPCODE_END
)
2097 if (opcode
== TGSI_OPCODE_TEX
||
2098 opcode
== TGSI_OPCODE_TXP
||
2099 opcode
== TGSI_OPCODE_TXD
||
2100 opcode
== TGSI_OPCODE_TXB
||
2101 opcode
== TGSI_OPCODE_TXL
||
2102 opcode
== TGSI_OPCODE_TXF
||
2103 opcode
== TGSI_OPCODE_TXQ
||
2104 opcode
== TGSI_OPCODE_CAL
||
2105 opcode
== TGSI_OPCODE_CALLNZ
||
2106 opcode
== TGSI_OPCODE_IF
||
2107 opcode
== TGSI_OPCODE_UIF
||
2108 opcode
== TGSI_OPCODE_BGNLOOP
||
2109 opcode
== TGSI_OPCODE_SWITCH
)
2119 * Kill fragment if any of the src register values are negative.
2123 struct lp_build_tgsi_soa_context
*bld
,
2124 const struct tgsi_full_instruction
*inst
,
2127 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2128 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
2129 LLVMValueRef terms
[TGSI_NUM_CHANNELS
];
2131 unsigned chan_index
;
2133 memset(&terms
, 0, sizeof terms
);
2135 TGSI_FOR_EACH_CHANNEL( chan_index
) {
2138 /* Unswizzle channel */
2139 swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
2141 /* Check if the component has not been already tested. */
2142 assert(swizzle
< TGSI_NUM_CHANNELS
);
2143 if( !terms
[swizzle
] )
2144 /* TODO: change the comparison operator instead of setting the sign */
2145 terms
[swizzle
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, chan_index
);
2149 TGSI_FOR_EACH_CHANNEL( chan_index
) {
2150 if(terms
[chan_index
]) {
2151 LLVMValueRef chan_mask
;
2154 * If term < 0 then mask = 0 else mask = ~0.
2156 chan_mask
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, terms
[chan_index
], bld
->bld_base
.base
.zero
);
2159 mask
= LLVMBuildAnd(builder
, mask
, chan_mask
, "");
2166 lp_build_mask_update(bld
->mask
, mask
);
2168 if (!near_end_of_shader(bld
, pc
))
2169 lp_build_mask_check(bld
->mask
);
2175 * Unconditional fragment kill.
2176 * The only predication is the execution mask which will apply if
2177 * we're inside a loop or conditional.
2180 emit_kill(struct lp_build_tgsi_soa_context
*bld
,
2183 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2186 /* For those channels which are "alive", disable fragment shader
2189 if (bld
->exec_mask
.has_mask
) {
2190 mask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
2193 LLVMValueRef zero
= LLVMConstNull(bld
->bld_base
.base
.int_vec_type
);
2197 lp_build_mask_update(bld
->mask
, mask
);
2199 if (!near_end_of_shader(bld
, pc
))
2200 lp_build_mask_check(bld
->mask
);
2205 * Emit code which will dump the value of all the temporary registers
2209 emit_dump_temps(struct lp_build_tgsi_soa_context
*bld
)
2211 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2212 LLVMBuilderRef builder
= gallivm
->builder
;
2213 LLVMValueRef temp_ptr
;
2214 LLVMValueRef i0
= lp_build_const_int32(gallivm
, 0);
2215 LLVMValueRef i1
= lp_build_const_int32(gallivm
, 1);
2216 LLVMValueRef i2
= lp_build_const_int32(gallivm
, 2);
2217 LLVMValueRef i3
= lp_build_const_int32(gallivm
, 3);
2219 int n
= bld
->bld_base
.info
->file_max
[TGSI_FILE_TEMPORARY
];
2221 for (index
= 0; index
< n
; index
++) {
2222 LLVMValueRef idx
= lp_build_const_int32(gallivm
, index
);
2223 LLVMValueRef v
[4][4], res
;
2226 lp_build_printf(gallivm
, "TEMP[%d]:\n", idx
);
2228 for (chan
= 0; chan
< 4; chan
++) {
2229 temp_ptr
= lp_get_temp_ptr_soa(bld
, index
, chan
);
2230 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
2231 v
[chan
][0] = LLVMBuildExtractElement(builder
, res
, i0
, "");
2232 v
[chan
][1] = LLVMBuildExtractElement(builder
, res
, i1
, "");
2233 v
[chan
][2] = LLVMBuildExtractElement(builder
, res
, i2
, "");
2234 v
[chan
][3] = LLVMBuildExtractElement(builder
, res
, i3
, "");
2237 lp_build_printf(gallivm
, " X: %f %f %f %f\n",
2238 v
[0][0], v
[0][1], v
[0][2], v
[0][3]);
2239 lp_build_printf(gallivm
, " Y: %f %f %f %f\n",
2240 v
[1][0], v
[1][1], v
[1][2], v
[1][3]);
2241 lp_build_printf(gallivm
, " Z: %f %f %f %f\n",
2242 v
[2][0], v
[2][1], v
[2][2], v
[2][3]);
2243 lp_build_printf(gallivm
, " W: %f %f %f %f\n",
2244 v
[3][0], v
[3][1], v
[3][2], v
[3][3]);
2251 lp_emit_declaration_soa(
2252 struct lp_build_tgsi_context
*bld_base
,
2253 const struct tgsi_full_declaration
*decl
)
2255 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
2256 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2257 LLVMTypeRef vec_type
= bld
->bld_base
.base
.vec_type
;
2258 const unsigned first
= decl
->Range
.First
;
2259 const unsigned last
= decl
->Range
.Last
;
2262 for (idx
= first
; idx
<= last
; ++idx
) {
2263 assert(last
<= bld
->bld_base
.info
->file_max
[decl
->Declaration
.File
]);
2264 switch (decl
->Declaration
.File
) {
2265 case TGSI_FILE_TEMPORARY
:
2266 assert(idx
< LP_MAX_TGSI_TEMPS
);
2267 if (!(bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
))) {
2268 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2269 bld
->temps
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
, "temp");
2273 case TGSI_FILE_OUTPUT
:
2274 if (!(bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
2275 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2276 bld
->outputs
[idx
][i
] = lp_build_alloca(gallivm
,
2277 vec_type
, "output");
2281 case TGSI_FILE_ADDRESS
:
2282 /* ADDR registers are only allocated with an integer LLVM IR type,
2283 * as they are guaranteed to always have integers.
2284 * XXX: Not sure if this exception is worthwhile (or the whole idea of
2285 * an ADDR register for that matter).
2287 assert(idx
< LP_MAX_TGSI_ADDRS
);
2288 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2289 bld
->addr
[idx
][i
] = lp_build_alloca(gallivm
, bld_base
->base
.int_vec_type
, "addr");
2292 case TGSI_FILE_PREDICATE
:
2293 assert(idx
< LP_MAX_TGSI_PREDS
);
2294 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2295 bld
->preds
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
,
2299 case TGSI_FILE_SAMPLER_VIEW
:
2301 * The target stored here MUST match whatever there actually
2302 * is in the set sampler views (what about return type?).
2304 assert(idx
< PIPE_MAX_SHADER_SAMPLER_VIEWS
);
2305 bld
->sv
[idx
] = decl
->SamplerView
;
2309 /* don't need to declare other vars */
2316 void lp_emit_immediate_soa(
2317 struct lp_build_tgsi_context
*bld_base
,
2318 const struct tgsi_full_immediate
*imm
)
2320 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
2321 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
2323 /* simply copy the immediate values into the next immediates[] slot */
2325 const uint size
= imm
->Immediate
.NrTokens
- 1;
2327 assert(bld
->num_immediates
< LP_MAX_TGSI_IMMEDIATES
);
2328 switch (imm
->Immediate
.DataType
) {
2329 case TGSI_IMM_FLOAT32
:
2330 for( i
= 0; i
< size
; ++i
)
2331 bld
->immediates
[bld
->num_immediates
][i
] =
2332 lp_build_const_vec(gallivm
, bld_base
->base
.type
, imm
->u
[i
].Float
);
2335 case TGSI_IMM_UINT32
:
2336 for( i
= 0; i
< size
; ++i
) {
2337 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->uint_bld
.type
, imm
->u
[i
].Uint
);
2338 bld
->immediates
[bld
->num_immediates
][i
] =
2339 LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
2343 case TGSI_IMM_INT32
:
2344 for( i
= 0; i
< size
; ++i
) {
2345 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->int_bld
.type
, imm
->u
[i
].Int
);
2346 bld
->immediates
[bld
->num_immediates
][i
] =
2347 LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
2352 for( i
= size
; i
< 4; ++i
)
2353 bld
->immediates
[bld
->num_immediates
][i
] = bld_base
->base
.undef
;
2355 if (bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
)) {
2356 unsigned index
= bld
->num_immediates
;
2357 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2358 LLVMBuilderRef builder
= gallivm
->builder
;
2359 for (i
= 0; i
< 4; ++i
) {
2360 LLVMValueRef lindex
= lp_build_const_int32(
2361 bld
->bld_base
.base
.gallivm
, index
* 4 + i
);
2362 LLVMValueRef imm_ptr
= LLVMBuildGEP(builder
,
2363 bld
->imms_array
, &lindex
, 1, "");
2364 LLVMBuildStore(builder
,
2365 bld
->immediates
[index
][i
],
2370 bld
->num_immediates
++;
2375 const struct lp_build_tgsi_action
* action
,
2376 struct lp_build_tgsi_context
* bld_base
,
2377 struct lp_build_emit_data
* emit_data
)
2379 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2381 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
,
2382 &emit_data
->output
[emit_data
->chan
], NULL
);
2387 const struct lp_build_tgsi_action
* action
,
2388 struct lp_build_tgsi_context
* bld_base
,
2389 struct lp_build_emit_data
* emit_data
)
2391 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2393 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
, NULL
,
2394 &emit_data
->output
[emit_data
->chan
]);
2399 const struct lp_build_tgsi_action
* action
,
2400 struct lp_build_tgsi_context
* bld_base
,
2401 struct lp_build_emit_data
* emit_data
)
2403 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2405 emit_kill(bld
, bld_base
->pc
- 1);
2410 const struct lp_build_tgsi_action
* action
,
2411 struct lp_build_tgsi_context
* bld_base
,
2412 struct lp_build_emit_data
* emit_data
)
2414 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2416 emit_kill_if(bld
, emit_data
->inst
, bld_base
->pc
- 1);
2421 const struct lp_build_tgsi_action
* action
,
2422 struct lp_build_tgsi_context
* bld_base
,
2423 struct lp_build_emit_data
* emit_data
)
2425 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2427 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
, emit_data
->output
);
2432 const struct lp_build_tgsi_action
* action
,
2433 struct lp_build_tgsi_context
* bld_base
,
2434 struct lp_build_emit_data
* emit_data
)
2436 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2438 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
2444 const struct lp_build_tgsi_action
* action
,
2445 struct lp_build_tgsi_context
* bld_base
,
2446 struct lp_build_emit_data
* emit_data
)
2448 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2450 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
2456 const struct lp_build_tgsi_action
* action
,
2457 struct lp_build_tgsi_context
* bld_base
,
2458 struct lp_build_emit_data
* emit_data
)
2460 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2462 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
2468 const struct lp_build_tgsi_action
* action
,
2469 struct lp_build_tgsi_context
* bld_base
,
2470 struct lp_build_emit_data
* emit_data
)
2472 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2474 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_PROJECTED
,
2480 const struct lp_build_tgsi_action
* action
,
2481 struct lp_build_tgsi_context
* bld_base
,
2482 struct lp_build_emit_data
* emit_data
)
2484 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2486 emit_size_query(bld
, emit_data
->inst
, emit_data
->output
, FALSE
);
2491 const struct lp_build_tgsi_action
* action
,
2492 struct lp_build_tgsi_context
* bld_base
,
2493 struct lp_build_emit_data
* emit_data
)
2495 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2497 emit_fetch_texels(bld
, emit_data
->inst
, emit_data
->output
, FALSE
);
2502 const struct lp_build_tgsi_action
* action
,
2503 struct lp_build_tgsi_context
* bld_base
,
2504 struct lp_build_emit_data
* emit_data
)
2506 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2508 emit_fetch_texels(bld
, emit_data
->inst
, emit_data
->output
, TRUE
);
2513 const struct lp_build_tgsi_action
* action
,
2514 struct lp_build_tgsi_context
* bld_base
,
2515 struct lp_build_emit_data
* emit_data
)
2517 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2519 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
2520 FALSE
, emit_data
->output
);
2525 const struct lp_build_tgsi_action
* action
,
2526 struct lp_build_tgsi_context
* bld_base
,
2527 struct lp_build_emit_data
* emit_data
)
2529 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2531 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
2532 FALSE
, emit_data
->output
);
2537 const struct lp_build_tgsi_action
* action
,
2538 struct lp_build_tgsi_context
* bld_base
,
2539 struct lp_build_emit_data
* emit_data
)
2541 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2543 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
2544 TRUE
, emit_data
->output
);
2549 const struct lp_build_tgsi_action
* action
,
2550 struct lp_build_tgsi_context
* bld_base
,
2551 struct lp_build_emit_data
* emit_data
)
2553 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2555 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_ZERO
,
2556 TRUE
, emit_data
->output
);
2561 const struct lp_build_tgsi_action
* action
,
2562 struct lp_build_tgsi_context
* bld_base
,
2563 struct lp_build_emit_data
* emit_data
)
2565 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2567 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
2568 FALSE
, emit_data
->output
);
2573 const struct lp_build_tgsi_action
* action
,
2574 struct lp_build_tgsi_context
* bld_base
,
2575 struct lp_build_emit_data
* emit_data
)
2577 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2579 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
2580 FALSE
, emit_data
->output
);
2585 const struct lp_build_tgsi_action
* action
,
2586 struct lp_build_tgsi_context
* bld_base
,
2587 struct lp_build_emit_data
* emit_data
)
2589 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2591 emit_size_query(bld
, emit_data
->inst
, emit_data
->output
, TRUE
);
2595 mask_vec(struct lp_build_tgsi_context
*bld_base
)
2597 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2598 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2599 struct lp_exec_mask
*exec_mask
= &bld
->exec_mask
;
2601 if (!exec_mask
->has_mask
) {
2602 return lp_build_mask_value(bld
->mask
);
2604 return LLVMBuildAnd(builder
, lp_build_mask_value(bld
->mask
),
2605 exec_mask
->exec_mask
, "");
2609 increment_vec_ptr_by_mask(struct lp_build_tgsi_context
* bld_base
,
2613 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2614 LLVMValueRef current_vec
= LLVMBuildLoad(builder
, ptr
, "");
2616 current_vec
= LLVMBuildSub(builder
, current_vec
, mask
, "");
2618 LLVMBuildStore(builder
, current_vec
, ptr
);
2622 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context
* bld_base
,
2626 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2627 LLVMValueRef current_vec
= LLVMBuildLoad(builder
, ptr
, "");
2629 current_vec
= lp_build_select(&bld_base
->uint_bld
,
2631 bld_base
->uint_bld
.zero
,
2634 LLVMBuildStore(builder
, current_vec
, ptr
);
2638 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context
* bld
,
2639 LLVMValueRef current_mask_vec
,
2640 LLVMValueRef total_emitted_vertices_vec
)
2642 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2643 struct lp_build_context
*int_bld
= &bld
->bld_base
.int_bld
;
2644 LLVMValueRef max_mask
= lp_build_cmp(int_bld
, PIPE_FUNC_LESS
,
2645 total_emitted_vertices_vec
,
2646 bld
->max_output_vertices_vec
);
2648 return LLVMBuildAnd(builder
, current_mask_vec
, max_mask
, "");
2653 const struct lp_build_tgsi_action
* action
,
2654 struct lp_build_tgsi_context
* bld_base
,
2655 struct lp_build_emit_data
* emit_data
)
2657 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2658 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2660 if (bld
->gs_iface
->emit_vertex
) {
2661 LLVMValueRef mask
= mask_vec(bld_base
);
2662 LLVMValueRef total_emitted_vertices_vec
=
2663 LLVMBuildLoad(builder
, bld
->total_emitted_vertices_vec_ptr
, "");
2664 mask
= clamp_mask_to_max_output_vertices(bld
, mask
,
2665 total_emitted_vertices_vec
);
2666 gather_outputs(bld
);
2667 bld
->gs_iface
->emit_vertex(bld
->gs_iface
, &bld
->bld_base
,
2669 total_emitted_vertices_vec
);
2670 increment_vec_ptr_by_mask(bld_base
, bld
->emitted_vertices_vec_ptr
,
2672 increment_vec_ptr_by_mask(bld_base
, bld
->total_emitted_vertices_vec_ptr
,
2675 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
2676 " +++ emit vertex masked ones = ",
2678 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
2679 " +++ emit vertex emitted = ",
2680 total_emitted_vertices_vec
);
2687 end_primitive_masked(struct lp_build_tgsi_context
* bld_base
,
2690 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2691 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2693 if (bld
->gs_iface
->end_primitive
) {
2694 LLVMValueRef emitted_vertices_vec
=
2695 LLVMBuildLoad(builder
, bld
->emitted_vertices_vec_ptr
, "");
2696 LLVMValueRef emitted_prims_vec
=
2697 LLVMBuildLoad(builder
, bld
->emitted_prims_vec_ptr
, "");
2699 bld
->gs_iface
->end_primitive(bld
->gs_iface
, &bld
->bld_base
,
2700 emitted_vertices_vec
,
2704 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
2705 " +++ end prim masked ones = ",
2707 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
2708 " +++ end prim emitted verts1 = ",
2709 emitted_vertices_vec
);
2710 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
2711 " +++ end prim emitted prims1 = ",
2712 LLVMBuildLoad(builder
,
2713 bld
->emitted_prims_vec_ptr
, ""));
2715 increment_vec_ptr_by_mask(bld_base
, bld
->emitted_prims_vec_ptr
,
2717 clear_uint_vec_ptr_from_mask(bld_base
, bld
->emitted_vertices_vec_ptr
,
2720 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
2721 " +++ end prim emitted verts2 = ",
2722 LLVMBuildLoad(builder
,
2723 bld
->emitted_vertices_vec_ptr
, ""));
2731 const struct lp_build_tgsi_action
* action
,
2732 struct lp_build_tgsi_context
* bld_base
,
2733 struct lp_build_emit_data
* emit_data
)
2735 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2737 if (bld
->gs_iface
->end_primitive
) {
2738 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2739 LLVMValueRef mask
= mask_vec(bld_base
);
2740 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
2741 LLVMValueRef emitted_verts
= LLVMBuildLoad(
2742 builder
, bld
->emitted_vertices_vec_ptr
, "");
2743 LLVMValueRef emitted_mask
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
2746 /* We need to combine the current execution mask with the mask
2747 telling us which, if any, execution slots actually have
2748 unemitted primitives, this way we make sure that end_primitives
2749 executes only on the paths that have unflushed vertices */
2750 mask
= LLVMBuildAnd(builder
, mask
, emitted_mask
, "");
2752 end_primitive_masked(bld_base
, mask
);
2758 const struct lp_build_tgsi_action
* action
,
2759 struct lp_build_tgsi_context
* bld_base
,
2760 struct lp_build_emit_data
* emit_data
)
2762 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2764 lp_exec_mask_call(&bld
->exec_mask
, emit_data
->inst
->Label
.Label
,
2770 const struct lp_build_tgsi_action
* action
,
2771 struct lp_build_tgsi_context
* bld_base
,
2772 struct lp_build_emit_data
* emit_data
)
2774 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2776 lp_exec_mask_ret(&bld
->exec_mask
, &bld_base
->pc
);
2781 const struct lp_build_tgsi_action
* action
,
2782 struct lp_build_tgsi_context
* bld_base
,
2783 struct lp_build_emit_data
* emit_data
)
2785 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2787 lp_exec_break(&bld
->exec_mask
, bld_base
);
2792 const struct lp_build_tgsi_action
* action
,
2793 struct lp_build_tgsi_context
* bld_base
,
2794 struct lp_build_emit_data
* emit_data
)
2796 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2797 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
2798 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
2799 LLVMValueRef unsigned_cond
=
2800 LLVMBuildBitCast(builder
, emit_data
->args
[0], uint_bld
->vec_type
, "");
2801 LLVMValueRef cond
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
2805 lp_exec_break_condition(&bld
->exec_mask
, cond
);
2810 const struct lp_build_tgsi_action
* action
,
2811 struct lp_build_tgsi_context
* bld_base
,
2812 struct lp_build_emit_data
* emit_data
)
2815 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2817 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_NOTEQUAL
,
2818 emit_data
->args
[0], bld
->bld_base
.base
.zero
);
2819 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
2824 const struct lp_build_tgsi_action
* action
,
2825 struct lp_build_tgsi_context
* bld_base
,
2826 struct lp_build_emit_data
* emit_data
)
2829 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2830 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
2832 tmp
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
2833 emit_data
->args
[0], uint_bld
->zero
);
2834 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
2839 const struct lp_build_tgsi_action
* action
,
2840 struct lp_build_tgsi_context
* bld_base
,
2841 struct lp_build_emit_data
* emit_data
)
2843 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2845 lp_exec_case(&bld
->exec_mask
, emit_data
->args
[0]);
2850 const struct lp_build_tgsi_action
* action
,
2851 struct lp_build_tgsi_context
* bld_base
,
2852 struct lp_build_emit_data
* emit_data
)
2854 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2856 lp_exec_default(&bld
->exec_mask
, bld_base
);
2861 const struct lp_build_tgsi_action
* action
,
2862 struct lp_build_tgsi_context
* bld_base
,
2863 struct lp_build_emit_data
* emit_data
)
2865 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2867 lp_exec_switch(&bld
->exec_mask
, emit_data
->args
[0]);
2872 const struct lp_build_tgsi_action
* action
,
2873 struct lp_build_tgsi_context
* bld_base
,
2874 struct lp_build_emit_data
* emit_data
)
2876 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2878 lp_exec_endswitch(&bld
->exec_mask
, bld_base
);
2883 const struct lp_build_tgsi_action
* action
,
2884 struct lp_build_tgsi_context
* bld_base
,
2885 struct lp_build_emit_data
* emit_data
)
2887 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2889 lp_exec_bgnloop(&bld
->exec_mask
);
2894 const struct lp_build_tgsi_action
* action
,
2895 struct lp_build_tgsi_context
* bld_base
,
2896 struct lp_build_emit_data
* emit_data
)
2898 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2900 lp_exec_mask_bgnsub(&bld
->exec_mask
);
2905 const struct lp_build_tgsi_action
* action
,
2906 struct lp_build_tgsi_context
* bld_base
,
2907 struct lp_build_emit_data
* emit_data
)
2909 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2911 lp_exec_mask_cond_invert(&bld
->exec_mask
);
2916 const struct lp_build_tgsi_action
* action
,
2917 struct lp_build_tgsi_context
* bld_base
,
2918 struct lp_build_emit_data
* emit_data
)
2920 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2922 lp_exec_mask_cond_pop(&bld
->exec_mask
);
2927 const struct lp_build_tgsi_action
* action
,
2928 struct lp_build_tgsi_context
* bld_base
,
2929 struct lp_build_emit_data
* emit_data
)
2931 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2933 lp_exec_endloop(bld_base
->base
.gallivm
, &bld
->exec_mask
);
2938 const struct lp_build_tgsi_action
* action
,
2939 struct lp_build_tgsi_context
* bld_base
,
2940 struct lp_build_emit_data
* emit_data
)
2942 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2944 lp_exec_mask_endsub(&bld
->exec_mask
, &bld_base
->pc
);
2949 const struct lp_build_tgsi_action
* action
,
2950 struct lp_build_tgsi_context
* bld_base
,
2951 struct lp_build_emit_data
* emit_data
)
2953 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2955 lp_exec_continue(&bld
->exec_mask
);
2958 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
2960 * XXX: What do the comments about xmm registers mean? Maybe they are left over
2961 * from old code, but there is no garauntee that LLVM will use those registers
2964 * XXX: There should be no calls to lp_build_emit_fetch in this function. This
2965 * should be handled by the emit_data->fetch_args function. */
2968 const struct lp_build_tgsi_action
* action
,
2969 struct lp_build_tgsi_context
* bld_base
,
2970 struct lp_build_emit_data
* emit_data
)
2972 LLVMValueRef tmp0
, tmp1
;
2973 LLVMValueRef tmp4
= NULL
;
2974 LLVMValueRef tmp5
= NULL
;
2975 LLVMValueRef tmp6
= NULL
;
2976 LLVMValueRef tmp7
= NULL
;
2977 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2979 uint dims
= (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_NRM
) ? 3 : 4;
2981 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) ||
2982 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
) ||
2983 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
) ||
2984 (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 4)) {
2986 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
2989 /* xmm0 = src.x * src.x */
2990 tmp0
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
2991 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
2994 tmp0
= lp_build_mul( &bld
->bld_base
.base
, tmp0
, tmp0
);
2997 /* xmm0 = xmm0 + src.y * src.y */
2998 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Y
);
2999 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
3002 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
3003 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
3006 /* xmm0 = xmm0 + src.z * src.z */
3007 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Z
);
3008 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
3011 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
3012 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
3016 /* xmm0 = xmm0 + src.w * src.w */
3017 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
3018 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
)) {
3021 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
3022 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
3024 /* xmm1 = 1 / sqrt(xmm0) */
3025 tmp1
= lp_build_rsqrt( &bld
->bld_base
.base
, tmp0
);
3026 /* dst.x = xmm1 * src.x */
3027 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
3028 emit_data
->output
[TGSI_CHAN_X
] = lp_build_mul( &bld
->bld_base
.base
, tmp4
, tmp1
);
3030 /* dst.y = xmm1 * src.y */
3031 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
3032 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_mul( &bld
->bld_base
.base
, tmp5
, tmp1
);
3035 /* dst.z = xmm1 * src.z */
3036 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
3037 emit_data
->output
[TGSI_CHAN_Z
] = lp_build_mul( &bld
->bld_base
.base
, tmp6
, tmp1
);
3039 /* dst.w = xmm1 * src.w */
3040 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) && dims
== 4) {
3041 emit_data
->output
[TGSI_CHAN_W
] = lp_build_mul( &bld
->bld_base
.base
, tmp7
, tmp1
);
3046 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 3) {
3047 emit_data
->output
[TGSI_CHAN_W
] = bld
->bld_base
.base
.one
;
3051 static void emit_prologue(struct lp_build_tgsi_context
* bld_base
)
3053 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3054 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
3056 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
3057 LLVMValueRef array_size
=
3058 lp_build_const_int32(gallivm
,
3059 bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] * 4 + 4);
3060 bld
->temps_array
= lp_build_array_alloca(gallivm
,
3061 bld_base
->base
.vec_type
, array_size
,
3065 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
3066 LLVMValueRef array_size
=
3067 lp_build_const_int32(gallivm
,
3068 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] * 4 + 4);
3069 bld
->outputs_array
= lp_build_array_alloca(gallivm
,
3070 bld_base
->base
.vec_type
, array_size
,
3074 if (bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
)) {
3075 LLVMValueRef array_size
=
3076 lp_build_const_int32(gallivm
,
3077 bld_base
->info
->file_max
[TGSI_FILE_IMMEDIATE
] * 4 + 4);
3078 bld
->imms_array
= lp_build_array_alloca(gallivm
,
3079 bld_base
->base
.vec_type
, array_size
,
3083 /* If we have indirect addressing in inputs we need to copy them into
3084 * our alloca array to be able to iterate over them */
3085 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
) && !bld
->gs_iface
) {
3086 unsigned index
, chan
;
3087 LLVMTypeRef vec_type
= bld_base
->base
.vec_type
;
3088 LLVMValueRef array_size
= lp_build_const_int32(gallivm
,
3089 bld_base
->info
->file_max
[TGSI_FILE_INPUT
]*4 + 4);
3090 bld
->inputs_array
= lp_build_array_alloca(gallivm
,
3091 vec_type
, array_size
,
3094 assert(bld_base
->info
->num_inputs
3095 <= bld_base
->info
->file_max
[TGSI_FILE_INPUT
] + 1);
3097 for (index
= 0; index
< bld_base
->info
->num_inputs
; ++index
) {
3098 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3099 LLVMValueRef lindex
=
3100 lp_build_const_int32(gallivm
, index
* 4 + chan
);
3101 LLVMValueRef input_ptr
=
3102 LLVMBuildGEP(gallivm
->builder
, bld
->inputs_array
,
3104 LLVMValueRef value
= bld
->inputs
[index
][chan
];
3106 LLVMBuildStore(gallivm
->builder
, value
, input_ptr
);
3111 if (bld
->gs_iface
) {
3112 struct lp_build_context
*uint_bld
= &bld
->bld_base
.uint_bld
;
3113 bld
->emitted_prims_vec_ptr
=
3114 lp_build_alloca(gallivm
,
3116 "emitted_prims_ptr");
3117 bld
->emitted_vertices_vec_ptr
=
3118 lp_build_alloca(gallivm
,
3120 "emitted_vertices_ptr");
3121 bld
->total_emitted_vertices_vec_ptr
=
3122 lp_build_alloca(gallivm
,
3124 "total_emitted_vertices_ptr");
3126 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3127 bld
->emitted_prims_vec_ptr
);
3128 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3129 bld
->emitted_vertices_vec_ptr
);
3130 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3131 bld
->total_emitted_vertices_vec_ptr
);
3135 static void emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
3137 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3138 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3142 emit_dump_temps(bld
);
3145 /* If we have indirect addressing in outputs we need to copy our alloca array
3146 * to the outputs slots specified by the caller */
3147 if (bld
->gs_iface
) {
3148 LLVMValueRef total_emitted_vertices_vec
;
3149 LLVMValueRef emitted_prims_vec
;
3150 /* implicit end_primitives, needed in case there are any unflushed
3151 vertices in the cache */
3152 end_primitive(NULL
, bld_base
, NULL
);
3154 total_emitted_vertices_vec
=
3155 LLVMBuildLoad(builder
, bld
->total_emitted_vertices_vec_ptr
, "");
3157 LLVMBuildLoad(builder
, bld
->emitted_prims_vec_ptr
, "");
3159 bld
->gs_iface
->gs_epilogue(bld
->gs_iface
,
3161 total_emitted_vertices_vec
,
3164 gather_outputs(bld
);
3169 lp_build_tgsi_soa(struct gallivm_state
*gallivm
,
3170 const struct tgsi_token
*tokens
,
3171 struct lp_type type
,
3172 struct lp_build_mask_context
*mask
,
3173 LLVMValueRef consts_ptr
,
3174 const struct lp_bld_tgsi_system_values
*system_values
,
3175 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
3176 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
3177 struct lp_build_sampler_soa
*sampler
,
3178 const struct tgsi_shader_info
*info
,
3179 const struct lp_build_tgsi_gs_iface
*gs_iface
)
3181 struct lp_build_tgsi_soa_context bld
;
3183 struct lp_type res_type
;
3185 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
3186 memset(&res_type
, 0, sizeof res_type
);
3187 res_type
.width
= type
.width
;
3188 res_type
.length
= type
.length
;
3191 /* Setup build context */
3192 memset(&bld
, 0, sizeof bld
);
3193 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
3194 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
3195 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
3196 lp_build_context_init(&bld
.elem_bld
, gallivm
, lp_elem_type(type
));
3198 bld
.inputs
= inputs
;
3199 bld
.outputs
= outputs
;
3200 bld
.consts_ptr
= consts_ptr
;
3201 bld
.sampler
= sampler
;
3202 bld
.bld_base
.info
= info
;
3203 bld
.indirect_files
= info
->indirect_files
;
3205 bld
.bld_base
.soa
= TRUE
;
3206 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
3207 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
3208 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
3209 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
3210 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = emit_fetch_system_value
;
3211 bld
.bld_base
.emit_store
= emit_store
;
3213 bld
.bld_base
.emit_declaration
= lp_emit_declaration_soa
;
3214 bld
.bld_base
.emit_immediate
= lp_emit_immediate_soa
;
3216 bld
.bld_base
.emit_prologue
= emit_prologue
;
3217 bld
.bld_base
.emit_epilogue
= emit_epilogue
;
3219 /* Set opcode actions */
3220 lp_set_default_actions_cpu(&bld
.bld_base
);
3222 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
3223 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNSUB
].emit
= bgnsub_emit
;
3224 bld
.bld_base
.op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
3225 bld
.bld_base
.op_actions
[TGSI_OPCODE_BREAKC
].emit
= breakc_emit
;
3226 bld
.bld_base
.op_actions
[TGSI_OPCODE_CAL
].emit
= cal_emit
;
3227 bld
.bld_base
.op_actions
[TGSI_OPCODE_CASE
].emit
= case_emit
;
3228 bld
.bld_base
.op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
3229 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDX
].emit
= ddx_emit
;
3230 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDY
].emit
= ddy_emit
;
3231 bld
.bld_base
.op_actions
[TGSI_OPCODE_DEFAULT
].emit
= default_emit
;
3232 bld
.bld_base
.op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
3233 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
3234 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
3235 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSUB
].emit
= endsub_emit
;
3236 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSWITCH
].emit
= endswitch_emit
;
3237 bld
.bld_base
.op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
3238 bld
.bld_base
.op_actions
[TGSI_OPCODE_UIF
].emit
= uif_emit
;
3239 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILL_IF
].emit
= kill_if_emit
;
3240 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILL
].emit
= kill_emit
;
3241 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM
].emit
= nrm_emit
;
3242 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM4
].emit
= nrm_emit
;
3243 bld
.bld_base
.op_actions
[TGSI_OPCODE_RET
].emit
= ret_emit
;
3244 bld
.bld_base
.op_actions
[TGSI_OPCODE_SWITCH
].emit
= switch_emit
;
3245 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX
].emit
= tex_emit
;
3246 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB
].emit
= txb_emit
;
3247 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXD
].emit
= txd_emit
;
3248 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL
].emit
= txl_emit
;
3249 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXP
].emit
= txp_emit
;
3250 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXQ
].emit
= txq_emit
;
3251 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXF
].emit
= txf_emit
;
3252 /* DX10 sampling ops */
3253 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE
].emit
= sample_emit
;
3254 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_B
].emit
= sample_b_emit
;
3255 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_C
].emit
= sample_c_emit
;
3256 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_C_LZ
].emit
= sample_c_lz_emit
;
3257 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_D
].emit
= sample_d_emit
;
3258 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_I
].emit
= sample_i_emit
;
3259 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_L
].emit
= sample_l_emit
;
3260 bld
.bld_base
.op_actions
[TGSI_OPCODE_SVIEWINFO
].emit
= sviewinfo_emit
;
3263 /* There's no specific value for this because it should always
3264 * be set, but apps using ext_geometry_shader4 quite often
3265 * were forgetting so we're using MAX_VERTEX_VARYING from
3266 * that spec even though we could debug_assert if it's not
3267 * set, but that's a lot uglier. */
3268 uint max_output_vertices
= 32;
3270 /* inputs are always indirect with gs */
3271 bld
.indirect_files
|= (1 << TGSI_FILE_INPUT
);
3272 bld
.gs_iface
= gs_iface
;
3273 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_gs_input
;
3274 bld
.bld_base
.op_actions
[TGSI_OPCODE_EMIT
].emit
= emit_vertex
;
3275 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= end_primitive
;
3277 for (i
= 0; i
< info
->num_properties
; ++i
) {
3278 if (info
->properties
[i
].name
==
3279 TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
) {
3280 max_output_vertices
= info
->properties
[i
].data
[0];
3283 bld
.max_output_vertices_vec
=
3284 lp_build_const_int_vec(gallivm
, bld
.bld_base
.int_bld
.type
,
3285 max_output_vertices
);
3288 lp_exec_mask_init(&bld
.exec_mask
, &bld
.bld_base
.int_bld
);
3290 bld
.system_values
= *system_values
;
3292 lp_build_tgsi_llvm(&bld
.bld_base
, tokens
);
3295 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
3296 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
3297 debug_printf("11111111111111111111111111111 \n");
3298 tgsi_dump(tokens
, 0);
3299 lp_debug_dump_value(function
);
3300 debug_printf("2222222222222222222222222222 \n");
3304 LLVMModuleRef module
= LLVMGetGlobalParent(
3305 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
3306 LLVMDumpModule(module
);