1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007-2008 VMware, Inc.
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 VMWARE 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 "tgsi/tgsi_strings.h"
51 #include "lp_bld_tgsi_action.h"
52 #include "lp_bld_type.h"
53 #include "lp_bld_const.h"
54 #include "lp_bld_arit.h"
55 #include "lp_bld_bitarit.h"
56 #include "lp_bld_gather.h"
57 #include "lp_bld_init.h"
58 #include "lp_bld_logic.h"
59 #include "lp_bld_swizzle.h"
60 #include "lp_bld_flow.h"
61 #include "lp_bld_quad.h"
62 #include "lp_bld_tgsi.h"
63 #include "lp_bld_limits.h"
64 #include "lp_bld_debug.h"
65 #include "lp_bld_printf.h"
66 #include "lp_bld_sample.h"
67 #include "lp_bld_struct.h"
69 /* SM 4.0 says that subroutines can nest 32 deep and
70 * we need one more for our main function */
71 #define LP_MAX_NUM_FUNCS 33
73 #define DUMP_GS_EMITS 0
76 * If non-zero, the generated LLVM IR will print intermediate results on every TGSI
80 * - take execution masks in consideration
81 * - debug control-flow instructions
83 #define DEBUG_EXECUTION 0
87 * Emit code to print a register value.
90 emit_dump_reg(struct gallivm_state
*gallivm
,
98 util_snprintf(buf
, sizeof buf
, " %s[%u].%c = ",
100 index
, "xyzw"[chan
]);
102 lp_build_print_value(gallivm
, buf
, value
);
106 * Return the context for the current function.
107 * (always 'main', if shader doesn't do any function calls)
109 static inline struct function_ctx
*
110 func_ctx(struct lp_exec_mask
*mask
)
112 assert(mask
->function_stack_size
> 0);
113 assert(mask
->function_stack_size
<= LP_MAX_NUM_FUNCS
);
114 return &mask
->function_stack
[mask
->function_stack_size
- 1];
118 * Returns true if we're in a loop.
119 * It's global, meaning that it returns true even if there's
120 * no loop inside the current function, but we were inside
121 * a loop inside another function, from which this one was called.
123 static inline boolean
124 mask_has_loop(struct lp_exec_mask
*mask
)
127 for (i
= mask
->function_stack_size
- 1; i
>= 0; --i
) {
128 const struct function_ctx
*ctx
= &mask
->function_stack
[i
];
129 if (ctx
->loop_stack_size
> 0)
136 * Returns true if we're inside a switch statement.
137 * It's global, meaning that it returns true even if there's
138 * no switch in the current function, but we were inside
139 * a switch inside another function, from which this one was called.
141 static inline boolean
142 mask_has_switch(struct lp_exec_mask
*mask
)
145 for (i
= mask
->function_stack_size
- 1; i
>= 0; --i
) {
146 const struct function_ctx
*ctx
= &mask
->function_stack
[i
];
147 if (ctx
->switch_stack_size
> 0)
154 * Returns true if we're inside a conditional.
155 * It's global, meaning that it returns true even if there's
156 * no conditional in the current function, but we were inside
157 * a conditional inside another function, from which this one was called.
159 static inline boolean
160 mask_has_cond(struct lp_exec_mask
*mask
)
163 for (i
= mask
->function_stack_size
- 1; i
>= 0; --i
) {
164 const struct function_ctx
*ctx
= &mask
->function_stack
[i
];
165 if (ctx
->cond_stack_size
> 0)
173 * Initialize a function context at the specified index.
176 lp_exec_mask_function_init(struct lp_exec_mask
*mask
, int function_idx
)
178 LLVMTypeRef int_type
= LLVMInt32TypeInContext(mask
->bld
->gallivm
->context
);
179 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
180 struct function_ctx
*ctx
= &mask
->function_stack
[function_idx
];
182 ctx
->cond_stack_size
= 0;
183 ctx
->loop_stack_size
= 0;
184 ctx
->switch_stack_size
= 0;
186 if (function_idx
== 0) {
187 ctx
->ret_mask
= mask
->ret_mask
;
190 ctx
->loop_limiter
= lp_build_alloca(mask
->bld
->gallivm
,
191 int_type
, "looplimiter");
194 LLVMConstInt(int_type
, LP_MAX_TGSI_LOOP_ITERATIONS
, false),
198 static void lp_exec_mask_init(struct lp_exec_mask
*mask
, struct lp_build_context
*bld
)
201 mask
->has_mask
= FALSE
;
202 mask
->ret_in_main
= FALSE
;
203 /* For the main function */
204 mask
->function_stack_size
= 1;
206 mask
->int_vec_type
= lp_build_int_vec_type(bld
->gallivm
, mask
->bld
->type
);
207 mask
->exec_mask
= mask
->ret_mask
= mask
->break_mask
= mask
->cont_mask
=
208 mask
->cond_mask
= mask
->switch_mask
=
209 LLVMConstAllOnes(mask
->int_vec_type
);
211 mask
->function_stack
= CALLOC(LP_MAX_NUM_FUNCS
,
212 sizeof(mask
->function_stack
[0]));
213 lp_exec_mask_function_init(mask
, 0);
217 lp_exec_mask_fini(struct lp_exec_mask
*mask
)
219 FREE(mask
->function_stack
);
222 static void lp_exec_mask_update(struct lp_exec_mask
*mask
)
224 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
225 boolean has_loop_mask
= mask_has_loop(mask
);
226 boolean has_cond_mask
= mask_has_cond(mask
);
227 boolean has_switch_mask
= mask_has_switch(mask
);
228 boolean has_ret_mask
= mask
->function_stack_size
> 1 ||
232 /*for loops we need to update the entire mask at runtime */
234 assert(mask
->break_mask
);
235 tmp
= LLVMBuildAnd(builder
,
239 mask
->exec_mask
= LLVMBuildAnd(builder
,
244 mask
->exec_mask
= mask
->cond_mask
;
246 if (has_switch_mask
) {
247 mask
->exec_mask
= LLVMBuildAnd(builder
,
254 mask
->exec_mask
= LLVMBuildAnd(builder
,
260 mask
->has_mask
= (has_cond_mask
||
266 static void lp_exec_mask_cond_push(struct lp_exec_mask
*mask
,
269 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
270 struct function_ctx
*ctx
= func_ctx(mask
);
272 if (ctx
->cond_stack_size
>= LP_MAX_TGSI_NESTING
) {
273 ctx
->cond_stack_size
++;
276 if (ctx
->cond_stack_size
== 0 && mask
->function_stack_size
== 1) {
277 assert(mask
->cond_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
279 ctx
->cond_stack
[ctx
->cond_stack_size
++] = mask
->cond_mask
;
280 assert(LLVMTypeOf(val
) == mask
->int_vec_type
);
281 mask
->cond_mask
= LLVMBuildAnd(builder
,
285 lp_exec_mask_update(mask
);
288 static void lp_exec_mask_cond_invert(struct lp_exec_mask
*mask
)
290 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
291 struct function_ctx
*ctx
= func_ctx(mask
);
292 LLVMValueRef prev_mask
;
293 LLVMValueRef inv_mask
;
295 assert(ctx
->cond_stack_size
);
296 if (ctx
->cond_stack_size
>= LP_MAX_TGSI_NESTING
)
298 prev_mask
= ctx
->cond_stack
[ctx
->cond_stack_size
- 1];
299 if (ctx
->cond_stack_size
== 1 && mask
->function_stack_size
== 1) {
300 assert(prev_mask
== LLVMConstAllOnes(mask
->int_vec_type
));
303 inv_mask
= LLVMBuildNot(builder
, mask
->cond_mask
, "");
305 mask
->cond_mask
= LLVMBuildAnd(builder
,
308 lp_exec_mask_update(mask
);
311 static void lp_exec_mask_cond_pop(struct lp_exec_mask
*mask
)
313 struct function_ctx
*ctx
= func_ctx(mask
);
314 assert(ctx
->cond_stack_size
);
315 --ctx
->cond_stack_size
;
316 if (ctx
->cond_stack_size
>= LP_MAX_TGSI_NESTING
)
318 mask
->cond_mask
= ctx
->cond_stack
[ctx
->cond_stack_size
];
319 lp_exec_mask_update(mask
);
322 static void lp_exec_bgnloop(struct lp_exec_mask
*mask
)
324 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
325 struct function_ctx
*ctx
= func_ctx(mask
);
327 if (ctx
->loop_stack_size
>= LP_MAX_TGSI_NESTING
) {
328 ++ctx
->loop_stack_size
;
332 ctx
->break_type_stack
[ctx
->loop_stack_size
+ ctx
->switch_stack_size
] =
334 ctx
->break_type
= LP_EXEC_MASK_BREAK_TYPE_LOOP
;
336 ctx
->loop_stack
[ctx
->loop_stack_size
].loop_block
= ctx
->loop_block
;
337 ctx
->loop_stack
[ctx
->loop_stack_size
].cont_mask
= mask
->cont_mask
;
338 ctx
->loop_stack
[ctx
->loop_stack_size
].break_mask
= mask
->break_mask
;
339 ctx
->loop_stack
[ctx
->loop_stack_size
].break_var
= ctx
->break_var
;
340 ++ctx
->loop_stack_size
;
342 ctx
->break_var
= lp_build_alloca(mask
->bld
->gallivm
, mask
->int_vec_type
, "");
343 LLVMBuildStore(builder
, mask
->break_mask
, ctx
->break_var
);
345 ctx
->loop_block
= lp_build_insert_new_block(mask
->bld
->gallivm
, "bgnloop");
347 LLVMBuildBr(builder
, ctx
->loop_block
);
348 LLVMPositionBuilderAtEnd(builder
, ctx
->loop_block
);
350 mask
->break_mask
= LLVMBuildLoad(builder
, ctx
->break_var
, "");
352 lp_exec_mask_update(mask
);
355 static void lp_exec_break(struct lp_exec_mask
*mask
,
356 struct lp_build_tgsi_context
* bld_base
)
358 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
359 struct function_ctx
*ctx
= func_ctx(mask
);
361 if (ctx
->break_type
== LP_EXEC_MASK_BREAK_TYPE_LOOP
) {
362 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
366 mask
->break_mask
= LLVMBuildAnd(builder
,
368 exec_mask
, "break_full");
371 unsigned opcode
= bld_base
->instructions
[bld_base
->pc
+ 1].Instruction
.Opcode
;
372 boolean break_always
= (opcode
== TGSI_OPCODE_ENDSWITCH
||
373 opcode
== TGSI_OPCODE_CASE
);
376 if (ctx
->switch_in_default
) {
378 * stop default execution but only if this is an unconditional switch.
379 * (The condition here is not perfect since dead code after break is
380 * allowed but should be sufficient since false negatives are just
381 * unoptimized - so we don't have to pre-evaluate that).
383 if(break_always
&& ctx
->switch_pc
) {
384 bld_base
->pc
= ctx
->switch_pc
;
390 mask
->switch_mask
= LLVMConstNull(mask
->bld
->int_vec_type
);
393 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
396 mask
->switch_mask
= LLVMBuildAnd(builder
,
398 exec_mask
, "break_switch");
402 lp_exec_mask_update(mask
);
405 static void lp_exec_break_condition(struct lp_exec_mask
*mask
,
408 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
409 struct function_ctx
*ctx
= func_ctx(mask
);
410 LLVMValueRef cond_mask
= LLVMBuildAnd(builder
,
413 cond_mask
= LLVMBuildNot(builder
, cond_mask
, "break_cond");
415 if (ctx
->break_type
== LP_EXEC_MASK_BREAK_TYPE_LOOP
) {
416 mask
->break_mask
= LLVMBuildAnd(builder
,
418 cond_mask
, "breakc_full");
421 mask
->switch_mask
= LLVMBuildAnd(builder
,
423 cond_mask
, "breakc_switch");
426 lp_exec_mask_update(mask
);
429 static void lp_exec_continue(struct lp_exec_mask
*mask
)
431 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
432 LLVMValueRef exec_mask
= LLVMBuildNot(builder
,
436 mask
->cont_mask
= LLVMBuildAnd(builder
,
440 lp_exec_mask_update(mask
);
444 static void lp_exec_endloop(struct gallivm_state
*gallivm
,
445 struct lp_exec_mask
*mask
)
447 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
448 struct function_ctx
*ctx
= func_ctx(mask
);
449 LLVMBasicBlockRef endloop
;
450 LLVMTypeRef int_type
= LLVMInt32TypeInContext(mask
->bld
->gallivm
->context
);
451 LLVMTypeRef reg_type
= LLVMIntTypeInContext(gallivm
->context
,
452 mask
->bld
->type
.width
*
453 mask
->bld
->type
.length
);
454 LLVMValueRef i1cond
, i2cond
, icond
, limiter
;
456 assert(mask
->break_mask
);
459 assert(ctx
->loop_stack_size
);
460 if (ctx
->loop_stack_size
> LP_MAX_TGSI_NESTING
) {
461 --ctx
->loop_stack_size
;
466 * Restore the cont_mask, but don't pop
468 mask
->cont_mask
= ctx
->loop_stack
[ctx
->loop_stack_size
- 1].cont_mask
;
469 lp_exec_mask_update(mask
);
472 * Unlike the continue mask, the break_mask must be preserved across loop
475 LLVMBuildStore(builder
, mask
->break_mask
, ctx
->break_var
);
477 /* Decrement the loop limiter */
478 limiter
= LLVMBuildLoad(builder
, ctx
->loop_limiter
, "");
480 limiter
= LLVMBuildSub(
483 LLVMConstInt(int_type
, 1, false),
486 LLVMBuildStore(builder
, limiter
, ctx
->loop_limiter
);
488 /* i1cond = (mask != 0) */
489 i1cond
= LLVMBuildICmp(
492 LLVMBuildBitCast(builder
, mask
->exec_mask
, reg_type
, ""),
493 LLVMConstNull(reg_type
), "i1cond");
495 /* i2cond = (looplimiter > 0) */
496 i2cond
= LLVMBuildICmp(
500 LLVMConstNull(int_type
), "i2cond");
502 /* if( i1cond && i2cond ) */
503 icond
= LLVMBuildAnd(builder
, i1cond
, i2cond
, "");
505 endloop
= lp_build_insert_new_block(mask
->bld
->gallivm
, "endloop");
507 LLVMBuildCondBr(builder
,
508 icond
, ctx
->loop_block
, endloop
);
510 LLVMPositionBuilderAtEnd(builder
, endloop
);
512 assert(ctx
->loop_stack_size
);
513 --ctx
->loop_stack_size
;
514 mask
->cont_mask
= ctx
->loop_stack
[ctx
->loop_stack_size
].cont_mask
;
515 mask
->break_mask
= ctx
->loop_stack
[ctx
->loop_stack_size
].break_mask
;
516 ctx
->loop_block
= ctx
->loop_stack
[ctx
->loop_stack_size
].loop_block
;
517 ctx
->break_var
= ctx
->loop_stack
[ctx
->loop_stack_size
].break_var
;
518 ctx
->break_type
= ctx
->break_type_stack
[ctx
->loop_stack_size
+
519 ctx
->switch_stack_size
];
521 lp_exec_mask_update(mask
);
524 static void lp_exec_switch(struct lp_exec_mask
*mask
,
525 LLVMValueRef switchval
)
527 struct function_ctx
*ctx
= func_ctx(mask
);
529 if (ctx
->switch_stack_size
>= LP_MAX_TGSI_NESTING
||
530 ctx
->loop_stack_size
> LP_MAX_TGSI_NESTING
) {
531 ctx
->switch_stack_size
++;
535 ctx
->break_type_stack
[ctx
->loop_stack_size
+ ctx
->switch_stack_size
] =
537 ctx
->break_type
= LP_EXEC_MASK_BREAK_TYPE_SWITCH
;
539 ctx
->switch_stack
[ctx
->switch_stack_size
].switch_mask
= mask
->switch_mask
;
540 ctx
->switch_stack
[ctx
->switch_stack_size
].switch_val
= ctx
->switch_val
;
541 ctx
->switch_stack
[ctx
->switch_stack_size
].switch_mask_default
= ctx
->switch_mask_default
;
542 ctx
->switch_stack
[ctx
->switch_stack_size
].switch_in_default
= ctx
->switch_in_default
;
543 ctx
->switch_stack
[ctx
->switch_stack_size
].switch_pc
= ctx
->switch_pc
;
544 ctx
->switch_stack_size
++;
546 mask
->switch_mask
= LLVMConstNull(mask
->int_vec_type
);
547 ctx
->switch_val
= switchval
;
548 ctx
->switch_mask_default
= LLVMConstNull(mask
->int_vec_type
);
549 ctx
->switch_in_default
= false;
552 lp_exec_mask_update(mask
);
555 static void lp_exec_endswitch(struct lp_exec_mask
*mask
,
556 struct lp_build_tgsi_context
* bld_base
)
558 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
559 struct function_ctx
*ctx
= func_ctx(mask
);
561 if (ctx
->switch_stack_size
> LP_MAX_TGSI_NESTING
) {
562 ctx
->switch_stack_size
--;
566 /* check if there's deferred default if so do it now */
567 if (ctx
->switch_pc
&& !ctx
->switch_in_default
) {
568 LLVMValueRef prevmask
, defaultmask
;
570 prevmask
= ctx
->switch_stack
[ctx
->switch_stack_size
- 1].switch_mask
;
571 defaultmask
= LLVMBuildNot(builder
, ctx
->switch_mask_default
, "sw_default_mask");
572 mask
->switch_mask
= LLVMBuildAnd(builder
, prevmask
, defaultmask
, "sw_mask");
573 ctx
->switch_in_default
= true;
575 lp_exec_mask_update(mask
);
577 assert(bld_base
->instructions
[ctx
->switch_pc
- 1].Instruction
.Opcode
==
578 TGSI_OPCODE_DEFAULT
);
580 tmp_pc
= bld_base
->pc
;
581 bld_base
->pc
= ctx
->switch_pc
;
583 * re-purpose switch_pc to point to here again, since we stop execution of
584 * the deferred default after next break.
586 ctx
->switch_pc
= tmp_pc
- 1;
591 else if (ctx
->switch_pc
&& ctx
->switch_in_default
) {
592 assert(bld_base
->pc
== ctx
->switch_pc
+ 1);
595 ctx
->switch_stack_size
--;
596 mask
->switch_mask
= ctx
->switch_stack
[ctx
->switch_stack_size
].switch_mask
;
597 ctx
->switch_val
= ctx
->switch_stack
[ctx
->switch_stack_size
].switch_val
;
598 ctx
->switch_mask_default
= ctx
->switch_stack
[ctx
->switch_stack_size
].switch_mask_default
;
599 ctx
->switch_in_default
= ctx
->switch_stack
[ctx
->switch_stack_size
].switch_in_default
;
600 ctx
->switch_pc
= ctx
->switch_stack
[ctx
->switch_stack_size
].switch_pc
;
602 ctx
->break_type
= ctx
->break_type_stack
[ctx
->loop_stack_size
+ ctx
->switch_stack_size
];
604 lp_exec_mask_update(mask
);
607 static void lp_exec_case(struct lp_exec_mask
*mask
,
608 LLVMValueRef caseval
)
610 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
611 struct function_ctx
*ctx
= func_ctx(mask
);
613 LLVMValueRef casemask
, prevmask
;
615 if (ctx
->switch_stack_size
> LP_MAX_TGSI_NESTING
) {
619 /* skipping case mask evaluation here is NOT optional (not in all cases anyway). */
620 if (!ctx
->switch_in_default
) {
621 prevmask
= ctx
->switch_stack
[ctx
->switch_stack_size
- 1].switch_mask
;
622 casemask
= lp_build_cmp(mask
->bld
, PIPE_FUNC_EQUAL
, caseval
, ctx
->switch_val
);
623 ctx
->switch_mask_default
= LLVMBuildOr(builder
, casemask
,
624 ctx
->switch_mask_default
, "sw_default_mask");
625 casemask
= LLVMBuildOr(builder
, casemask
, mask
->switch_mask
, "");
626 mask
->switch_mask
= LLVMBuildAnd(builder
, casemask
, prevmask
, "sw_mask");
628 lp_exec_mask_update(mask
);
633 * Analyse default statement in a switch.
634 * \return true if default is last statement, false otherwise
635 * \param default_pc_start contains pc of instruction to jump to
636 * if default wasn't last but there's no
637 * fallthrough into default.
639 static boolean
default_analyse_is_last(struct lp_exec_mask
*mask
,
640 struct lp_build_tgsi_context
* bld_base
,
641 int *default_pc_start
)
643 unsigned pc
= bld_base
->pc
;
644 struct function_ctx
*ctx
= func_ctx(mask
);
645 int curr_switch_stack
= ctx
->switch_stack_size
;
647 if (ctx
->switch_stack_size
> LP_MAX_TGSI_NESTING
) {
651 /* skip over case statements which are together with default */
652 while (bld_base
->instructions
[pc
].Instruction
.Opcode
== TGSI_OPCODE_CASE
) {
656 while (pc
!= ~0u && pc
< bld_base
->num_instructions
) {
657 unsigned opcode
= bld_base
->instructions
[pc
].Instruction
.Opcode
;
659 case TGSI_OPCODE_CASE
:
660 if (curr_switch_stack
== ctx
->switch_stack_size
) {
661 *default_pc_start
= pc
- 1;
665 case TGSI_OPCODE_SWITCH
:
668 case TGSI_OPCODE_ENDSWITCH
:
669 if (curr_switch_stack
== ctx
->switch_stack_size
) {
670 *default_pc_start
= pc
- 1;
678 /* should never arrive here */
683 static void lp_exec_default(struct lp_exec_mask
*mask
,
684 struct lp_build_tgsi_context
* bld_base
)
686 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
687 struct function_ctx
*ctx
= func_ctx(mask
);
690 boolean default_is_last
;
692 if (ctx
->switch_stack_size
> LP_MAX_TGSI_NESTING
) {
697 * This is a messy opcode, because it may not be always at the end and
698 * there can be fallthrough in and out of it.
701 default_is_last
= default_analyse_is_last(mask
, bld_base
, &default_exec_pc
);
703 * If it is last statement in switch (note that case statements appearing
704 * "at the same time" as default don't change that) everything is just fine,
705 * update switch mask and go on. This means we can handle default with
706 * fallthrough INTO it without overhead, if it is last.
708 if (default_is_last
) {
709 LLVMValueRef prevmask
, defaultmask
;
710 prevmask
= ctx
->switch_stack
[ctx
->switch_stack_size
- 1].switch_mask
;
711 defaultmask
= LLVMBuildNot(builder
, ctx
->switch_mask_default
, "sw_default_mask");
712 defaultmask
= LLVMBuildOr(builder
, defaultmask
, mask
->switch_mask
, "");
713 mask
->switch_mask
= LLVMBuildAnd(builder
, prevmask
, defaultmask
, "sw_mask");
714 ctx
->switch_in_default
= true;
716 lp_exec_mask_update(mask
);
720 * Technically, "case" immediately before default isn't really a
721 * fallthrough, however we still have to count them as such as we
722 * already have updated the masks.
723 * If that happens in practice could add a switch optimizer pass
724 * which just gets rid of all case statements appearing together with
725 * default (or could do switch analysis at switch start time instead).
727 unsigned opcode
= bld_base
->instructions
[bld_base
->pc
- 1].Instruction
.Opcode
;
728 boolean ft_into
= (opcode
!= TGSI_OPCODE_BRK
&&
729 opcode
!= TGSI_OPCODE_SWITCH
);
731 * If it is not last statement and there was no fallthrough into it,
732 * we record the PC and continue execution at next case (again, those
733 * case encountered at the same time don't count). At endswitch
734 * time, we update switchmask, and go back executing the code we skipped
735 * until the next break (possibly re-executing some code with changed mask
736 * if there was a fallthrough out of default).
737 * Finally, if it is not last statement and there was a fallthrough into it,
738 * do the same as with the former case, except instead of skipping the code
739 * just execute it without updating the mask, then go back and re-execute.
741 ctx
->switch_pc
= bld_base
->pc
;
743 bld_base
->pc
= default_exec_pc
;
749 /* stores val into an address pointed to by dst_ptr.
750 * mask->exec_mask is used to figure out which bits of val
751 * should be stored into the address
752 * (0 means don't store this bit, 1 means do store).
754 static void lp_exec_mask_store(struct lp_exec_mask
*mask
,
755 struct lp_build_context
*bld_store
,
758 LLVMValueRef dst_ptr
)
760 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
762 assert(lp_check_value(bld_store
->type
, val
));
763 assert(LLVMGetTypeKind(LLVMTypeOf(dst_ptr
)) == LLVMPointerTypeKind
);
764 assert(LLVMGetElementType(LLVMTypeOf(dst_ptr
)) == LLVMTypeOf(val
));
766 /* Mix the predicate and execution mask */
767 if (mask
->has_mask
) {
769 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
771 pred
= mask
->exec_mask
;
776 LLVMValueRef res
, dst
;
778 dst
= LLVMBuildLoad(builder
, dst_ptr
, "");
779 res
= lp_build_select(bld_store
, pred
, val
, dst
);
780 LLVMBuildStore(builder
, res
, dst_ptr
);
782 LLVMBuildStore(builder
, val
, dst_ptr
);
785 static void lp_exec_mask_call(struct lp_exec_mask
*mask
,
789 if (mask
->function_stack_size
>= LP_MAX_NUM_FUNCS
) {
793 lp_exec_mask_function_init(mask
, mask
->function_stack_size
);
794 mask
->function_stack
[mask
->function_stack_size
].pc
= *pc
;
795 mask
->function_stack
[mask
->function_stack_size
].ret_mask
= mask
->ret_mask
;
796 mask
->function_stack_size
++;
800 static void lp_exec_mask_ret(struct lp_exec_mask
*mask
, int *pc
)
802 LLVMBuilderRef builder
= mask
->bld
->gallivm
->builder
;
803 struct function_ctx
*ctx
= func_ctx(mask
);
804 LLVMValueRef exec_mask
;
806 if (ctx
->cond_stack_size
== 0 &&
807 ctx
->loop_stack_size
== 0 &&
808 ctx
->switch_stack_size
== 0 &&
809 mask
->function_stack_size
== 1) {
810 /* returning from main() */
815 if (mask
->function_stack_size
== 1) {
817 * This requires special handling since we need to ensure
818 * we don't drop the mask even if we have no call stack
819 * (e.g. after a ret in a if clause after the endif)
821 mask
->ret_in_main
= TRUE
;
824 exec_mask
= LLVMBuildNot(builder
,
828 mask
->ret_mask
= LLVMBuildAnd(builder
,
830 exec_mask
, "ret_full");
832 lp_exec_mask_update(mask
);
835 static void lp_exec_mask_bgnsub(struct lp_exec_mask
*mask
)
839 static void lp_exec_mask_endsub(struct lp_exec_mask
*mask
, int *pc
)
841 struct function_ctx
*ctx
;
843 assert(mask
->function_stack_size
> 1);
844 assert(mask
->function_stack_size
<= LP_MAX_NUM_FUNCS
);
846 ctx
= func_ctx(mask
);
847 mask
->function_stack_size
--;
850 mask
->ret_mask
= ctx
->ret_mask
;
852 lp_exec_mask_update(mask
);
857 get_file_ptr(struct lp_build_tgsi_soa_context
*bld
,
862 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
863 LLVMValueRef (*array_of_vars
)[TGSI_NUM_CHANNELS
];
864 LLVMValueRef var_of_array
;
867 case TGSI_FILE_TEMPORARY
:
868 array_of_vars
= bld
->temps
;
869 var_of_array
= bld
->temps_array
;
871 case TGSI_FILE_OUTPUT
:
872 array_of_vars
= bld
->outputs
;
873 var_of_array
= bld
->outputs_array
;
882 if (bld
->indirect_files
& (1 << file
)) {
883 LLVMValueRef lindex
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
, index
* 4 + chan
);
884 return LLVMBuildGEP(builder
, var_of_array
, &lindex
, 1, "");
887 assert(index
<= bld
->bld_base
.info
->file_max
[file
]);
888 return array_of_vars
[index
][chan
];
894 * Return pointer to a temporary register channel (src or dest).
895 * Note that indirect addressing cannot be handled here.
896 * \param index which temporary register
897 * \param chan which channel of the temp register.
900 lp_get_temp_ptr_soa(struct lp_build_tgsi_soa_context
*bld
,
904 return get_file_ptr(bld
, TGSI_FILE_TEMPORARY
, index
, chan
);
908 * Return pointer to a output register channel (src or dest).
909 * Note that indirect addressing cannot be handled here.
910 * \param index which output register
911 * \param chan which channel of the output register.
914 lp_get_output_ptr(struct lp_build_tgsi_soa_context
*bld
,
918 return get_file_ptr(bld
, TGSI_FILE_OUTPUT
, index
, chan
);
922 * If we have indirect addressing in outputs copy our alloca array
923 * to the outputs slots specified by the caller to make sure
924 * our outputs are delivered consistently via the same interface.
927 gather_outputs(struct lp_build_tgsi_soa_context
* bld
)
929 if ((bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
930 unsigned index
, chan
;
931 assert(bld
->bld_base
.info
->num_outputs
<=
932 bld
->bld_base
.info
->file_max
[TGSI_FILE_OUTPUT
] + 1);
933 for (index
= 0; index
< bld
->bld_base
.info
->num_outputs
; ++index
) {
934 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
935 bld
->outputs
[index
][chan
] = lp_get_output_ptr(bld
, index
, chan
);
943 * XXX the lp_build_gather() function should be capable of doing this
944 * with a little work.
947 build_gather(struct lp_build_tgsi_context
*bld_base
,
948 LLVMValueRef base_ptr
,
949 LLVMValueRef indexes
,
950 LLVMValueRef overflow_mask
,
951 LLVMValueRef indexes2
)
953 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
954 LLVMBuilderRef builder
= gallivm
->builder
;
955 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
956 struct lp_build_context
*bld
= &bld_base
->base
;
961 res
= LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), bld_base
->base
.type
.length
* 2));
965 * overflow_mask is a vector telling us which channels
966 * in the vector overflowed. We use the overflow behavior for
967 * constant buffers which is defined as:
968 * Out of bounds access to constant buffer returns 0 in all
969 * components. Out of bounds behavior is always with respect
970 * to the size of the buffer bound at that slot.
975 * We avoid per-element control flow here (also due to llvm going crazy,
976 * though I suspect it's better anyway since overflow is likely rare).
977 * Note that since we still fetch from buffers even if num_elements was
978 * zero (in this case we'll fetch from index zero) the jit func callers
979 * MUST provide valid fake constant buffers of size 4x32 (the values do
980 * not matter), otherwise we'd still need (not per element though)
983 indexes
= lp_build_select(uint_bld
, overflow_mask
, uint_bld
->zero
, indexes
);
985 indexes2
= lp_build_select(uint_bld
, overflow_mask
, uint_bld
->zero
, indexes2
);
989 * Loop over elements of index_vec, load scalar value, insert it into 'res'.
991 for (i
= 0; i
< bld
->type
.length
* (indexes2
? 2 : 1); i
++) {
994 LLVMValueRef scalar_ptr
, scalar
;
996 di
= lp_build_const_int32(bld
->gallivm
, i
);
998 si
= lp_build_const_int32(bld
->gallivm
, i
>> 1);
1002 if (indexes2
&& (i
& 1)) {
1003 index
= LLVMBuildExtractElement(builder
,
1006 index
= LLVMBuildExtractElement(builder
,
1009 scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
,
1010 &index
, 1, "gather_ptr");
1011 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
1013 res
= LLVMBuildInsertElement(builder
, res
, scalar
, di
, "");
1016 if (overflow_mask
) {
1018 res
= LLVMBuildBitCast(builder
, res
, bld_base
->dbl_bld
.vec_type
, "");
1019 overflow_mask
= LLVMBuildSExt(builder
, overflow_mask
,
1020 bld_base
->dbl_bld
.int_vec_type
, "");
1021 res
= lp_build_select(&bld_base
->dbl_bld
, overflow_mask
,
1022 bld_base
->dbl_bld
.zero
, res
);
1024 res
= lp_build_select(bld
, overflow_mask
, bld
->zero
, res
);
1032 * Scatter/store vector.
1035 emit_mask_scatter(struct lp_build_tgsi_soa_context
*bld
,
1036 LLVMValueRef base_ptr
,
1037 LLVMValueRef indexes
,
1038 LLVMValueRef values
,
1039 struct lp_exec_mask
*mask
,
1042 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1043 LLVMBuilderRef builder
= gallivm
->builder
;
1046 /* Mix the predicate and execution mask */
1047 if (mask
->has_mask
) {
1049 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
1052 pred
= mask
->exec_mask
;
1057 * Loop over elements of index_vec, store scalar value.
1059 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
1060 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1061 LLVMValueRef index
= LLVMBuildExtractElement(builder
, indexes
, ii
, "");
1062 LLVMValueRef scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
, &index
, 1, "scatter_ptr");
1063 LLVMValueRef val
= LLVMBuildExtractElement(builder
, values
, ii
, "scatter_val");
1064 LLVMValueRef scalar_pred
= pred
?
1065 LLVMBuildExtractElement(builder
, pred
, ii
, "scatter_pred") : NULL
;
1068 lp_build_printf(gallivm
, "scatter %d: val %f at %d %p\n",
1069 ii
, val
, index
, scalar_ptr
);
1072 LLVMValueRef real_val
, dst_val
;
1073 dst_val
= LLVMBuildLoad(builder
, scalar_ptr
, "");
1074 real_val
= lp_build_select(&bld
->elem_bld
, scalar_pred
, val
, dst_val
);
1075 LLVMBuildStore(builder
, real_val
, scalar_ptr
);
1078 LLVMBuildStore(builder
, val
, scalar_ptr
);
1085 * Read the current value of the ADDR register, convert the floats to
1086 * ints, add the base index and return the vector of offsets.
1087 * The offsets will be used to index into the constant buffer or
1088 * temporary register file.
1091 get_indirect_index(struct lp_build_tgsi_soa_context
*bld
,
1092 unsigned reg_file
, unsigned reg_index
,
1093 const struct tgsi_ind_register
*indirect_reg
)
1095 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1096 struct lp_build_context
*uint_bld
= &bld
->bld_base
.uint_bld
;
1097 /* always use X component of address register */
1098 unsigned swizzle
= indirect_reg
->Swizzle
;
1101 LLVMValueRef max_index
;
1104 assert(bld
->indirect_files
& (1 << reg_file
));
1106 base
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, reg_index
);
1108 assert(swizzle
< 4);
1109 switch (indirect_reg
->File
) {
1110 case TGSI_FILE_ADDRESS
:
1111 rel
= LLVMBuildLoad(builder
,
1112 bld
->addr
[indirect_reg
->Index
][swizzle
],
1114 /* ADDR LLVM values already have LLVM integer type. */
1116 case TGSI_FILE_TEMPORARY
:
1117 rel
= lp_get_temp_ptr_soa(bld
, indirect_reg
->Index
, swizzle
);
1118 rel
= LLVMBuildLoad(builder
, rel
, "load temp reg");
1119 /* TEMP LLVM values always have LLVM float type, but for indirection, the
1120 * value actually stored is expected to be an integer */
1121 rel
= LLVMBuildBitCast(builder
, rel
, uint_bld
->vec_type
, "");
1125 rel
= uint_bld
->zero
;
1128 index
= lp_build_add(uint_bld
, base
, rel
);
1131 * emit_fetch_constant handles constant buffer overflow so this code
1132 * is pointless for them.
1133 * Furthermore the D3D10 spec in section 6.5 says:
1134 * If the constant buffer bound to a slot is larger than the size
1135 * declared in the shader for that slot, implementations are allowed
1136 * to return incorrect data (not necessarily 0) for indices that are
1137 * larger than the declared size but smaller than the buffer size.
1139 if (reg_file
!= TGSI_FILE_CONSTANT
) {
1140 max_index
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
,
1142 bld
->bld_base
.info
->file_max
[reg_file
]);
1144 assert(!uint_bld
->type
.sign
);
1145 index
= lp_build_min(uint_bld
, index
, max_index
);
1151 static struct lp_build_context
*
1152 stype_to_fetch(struct lp_build_tgsi_context
* bld_base
,
1153 enum tgsi_opcode_type stype
)
1155 struct lp_build_context
*bld_fetch
;
1158 case TGSI_TYPE_FLOAT
:
1159 case TGSI_TYPE_UNTYPED
:
1160 bld_fetch
= &bld_base
->base
;
1162 case TGSI_TYPE_UNSIGNED
:
1163 bld_fetch
= &bld_base
->uint_bld
;
1165 case TGSI_TYPE_SIGNED
:
1166 bld_fetch
= &bld_base
->int_bld
;
1168 case TGSI_TYPE_DOUBLE
:
1169 bld_fetch
= &bld_base
->dbl_bld
;
1171 case TGSI_TYPE_VOID
:
1181 get_soa_array_offsets(struct lp_build_context
*uint_bld
,
1182 LLVMValueRef indirect_index
,
1183 unsigned chan_index
,
1184 boolean need_perelement_offset
)
1186 struct gallivm_state
*gallivm
= uint_bld
->gallivm
;
1187 LLVMValueRef chan_vec
=
1188 lp_build_const_int_vec(uint_bld
->gallivm
, uint_bld
->type
, chan_index
);
1189 LLVMValueRef length_vec
=
1190 lp_build_const_int_vec(gallivm
, uint_bld
->type
, uint_bld
->type
.length
);
1191 LLVMValueRef index_vec
;
1193 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1194 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1195 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
1196 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1198 if (need_perelement_offset
) {
1199 LLVMValueRef pixel_offsets
;
1201 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1202 pixel_offsets
= uint_bld
->undef
;
1203 for (i
= 0; i
< uint_bld
->type
.length
; i
++) {
1204 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1205 pixel_offsets
= LLVMBuildInsertElement(gallivm
->builder
, pixel_offsets
,
1208 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1214 emit_fetch_constant(
1215 struct lp_build_tgsi_context
* bld_base
,
1216 const struct tgsi_full_src_register
* reg
,
1217 enum tgsi_opcode_type stype
,
1220 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1221 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1222 LLVMBuilderRef builder
= gallivm
->builder
;
1223 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
1224 unsigned dimension
= 0;
1225 LLVMValueRef consts_ptr
;
1226 LLVMValueRef num_consts
;
1229 /* XXX: Handle fetching xyzw components as a vector */
1230 assert(swizzle
!= ~0u);
1232 if (reg
->Register
.Dimension
) {
1233 assert(!reg
->Dimension
.Indirect
);
1234 dimension
= reg
->Dimension
.Index
;
1235 assert(dimension
< LP_MAX_TGSI_CONST_BUFFERS
);
1238 consts_ptr
= bld
->consts
[dimension
];
1239 num_consts
= bld
->consts_sizes
[dimension
];
1241 if (reg
->Register
.Indirect
) {
1242 LLVMValueRef indirect_index
;
1243 LLVMValueRef swizzle_vec
=
1244 lp_build_const_int_vec(gallivm
, uint_bld
->type
, swizzle
);
1245 LLVMValueRef index_vec
; /* index into the const buffer */
1246 LLVMValueRef overflow_mask
;
1247 LLVMValueRef index_vec2
= NULL
;
1249 indirect_index
= get_indirect_index(bld
,
1251 reg
->Register
.Index
,
1254 /* All fetches are from the same constant buffer, so
1255 * we need to propagate the size to a vector to do a
1256 * vector comparison */
1257 num_consts
= lp_build_broadcast_scalar(uint_bld
, num_consts
);
1258 /* Construct a boolean vector telling us which channels
1259 * overflow the bound constant buffer */
1260 overflow_mask
= lp_build_compare(gallivm
, uint_bld
->type
, PIPE_FUNC_GEQUAL
,
1261 indirect_index
, num_consts
);
1263 /* index_vec = indirect_index * 4 + swizzle */
1264 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1265 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
1267 if (tgsi_type_is_64bit(stype
)) {
1268 LLVMValueRef swizzle_vec2
;
1269 swizzle_vec2
= lp_build_const_int_vec(gallivm
, uint_bld
->type
, swizzle
+ 1);
1270 index_vec2
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1271 index_vec2
= lp_build_add(uint_bld
, index_vec2
, swizzle_vec2
);
1273 /* Gather values from the constant buffer */
1274 res
= build_gather(bld_base
, consts_ptr
, index_vec
, overflow_mask
, index_vec2
);
1277 LLVMValueRef index
; /* index into the const buffer */
1278 LLVMValueRef scalar
, scalar_ptr
;
1279 struct lp_build_context
*bld_broad
= &bld_base
->base
;
1280 index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
* 4 + swizzle
);
1282 scalar_ptr
= LLVMBuildGEP(builder
, consts_ptr
,
1284 if (stype
== TGSI_TYPE_DOUBLE
) {
1285 LLVMTypeRef dptr_type
= LLVMPointerType(LLVMDoubleTypeInContext(gallivm
->context
), 0);
1286 scalar_ptr
= LLVMBuildBitCast(builder
, scalar_ptr
, dptr_type
, "");
1287 bld_broad
= &bld_base
->dbl_bld
;
1289 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
1290 res
= lp_build_broadcast_scalar(bld_broad
, scalar
);
1293 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
|| stype
== TGSI_TYPE_DOUBLE
) {
1294 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1295 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1302 * Fetch 64-bit values from two separate channels.
1303 * 64-bit values are stored split across two channels, like xy and zw.
1304 * This function creates a set of 16 floats,
1305 * extracts the values from the two channels,
1306 * puts them in the correct place, then casts to 8 64-bits.
1310 struct lp_build_tgsi_context
* bld_base
,
1311 enum tgsi_opcode_type stype
,
1313 LLVMValueRef input2
)
1315 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1316 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1317 LLVMBuilderRef builder
= gallivm
->builder
;
1319 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1321 LLVMValueRef shuffles
[16];
1322 int len
= bld_base
->base
.type
.length
* 2;
1325 for (i
= 0; i
< bld_base
->base
.type
.length
* 2; i
+=2) {
1326 shuffles
[i
] = lp_build_const_int32(gallivm
, i
/ 2);
1327 shuffles
[i
+ 1] = lp_build_const_int32(gallivm
, i
/ 2 + bld_base
->base
.type
.length
);
1329 res
= LLVMBuildShuffleVector(builder
, input
, input2
, LLVMConstVector(shuffles
, len
), "");
1331 return LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1335 emit_fetch_immediate(
1336 struct lp_build_tgsi_context
* bld_base
,
1337 const struct tgsi_full_src_register
* reg
,
1338 enum tgsi_opcode_type stype
,
1341 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1342 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1343 LLVMBuilderRef builder
= gallivm
->builder
;
1344 LLVMValueRef res
= NULL
;
1346 if (bld
->use_immediates_array
|| reg
->Register
.Indirect
) {
1347 LLVMValueRef imms_array
;
1348 LLVMTypeRef fptr_type
;
1350 /* cast imms_array pointer to float* */
1351 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1352 imms_array
= LLVMBuildBitCast(builder
, bld
->imms_array
, fptr_type
, "");
1354 if (reg
->Register
.Indirect
) {
1355 LLVMValueRef indirect_index
;
1356 LLVMValueRef index_vec
; /* index into the immediate register array */
1357 LLVMValueRef index_vec2
= NULL
;
1358 indirect_index
= get_indirect_index(bld
,
1360 reg
->Register
.Index
,
1363 * Unlike for other reg classes, adding pixel offsets is unnecessary -
1364 * immediates are stored as full vectors (FIXME??? - might be better
1365 * to store them the same as constants) but all elements are the same
1368 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1372 if (tgsi_type_is_64bit(stype
))
1373 index_vec2
= get_soa_array_offsets(&bld_base
->uint_bld
,
1377 /* Gather values from the immediate register array */
1378 res
= build_gather(bld_base
, imms_array
, index_vec
, NULL
, index_vec2
);
1380 LLVMValueRef lindex
= lp_build_const_int32(gallivm
,
1381 reg
->Register
.Index
* 4 + swizzle
);
1382 LLVMValueRef imms_ptr
= LLVMBuildGEP(builder
,
1383 bld
->imms_array
, &lindex
, 1, "");
1384 res
= LLVMBuildLoad(builder
, imms_ptr
, "");
1386 if (tgsi_type_is_64bit(stype
)) {
1387 LLVMValueRef lindex1
;
1388 LLVMValueRef imms_ptr2
;
1391 lindex1
= lp_build_const_int32(gallivm
,
1392 reg
->Register
.Index
* 4 + swizzle
+ 1);
1393 imms_ptr2
= LLVMBuildGEP(builder
,
1394 bld
->imms_array
, &lindex1
, 1, "");
1395 res2
= LLVMBuildLoad(builder
, imms_ptr2
, "");
1396 res
= emit_fetch_64bit(bld_base
, stype
, res
, res2
);
1401 res
= bld
->immediates
[reg
->Register
.Index
][swizzle
];
1402 if (tgsi_type_is_64bit(stype
))
1403 res
= emit_fetch_64bit(bld_base
, stype
, res
, bld
->immediates
[reg
->Register
.Index
][swizzle
+ 1]);
1406 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
|| stype
== TGSI_TYPE_DOUBLE
) {
1407 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1408 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1415 struct lp_build_tgsi_context
* bld_base
,
1416 const struct tgsi_full_src_register
* reg
,
1417 enum tgsi_opcode_type stype
,
1420 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1421 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1422 LLVMBuilderRef builder
= gallivm
->builder
;
1425 if (reg
->Register
.Indirect
) {
1426 LLVMValueRef indirect_index
;
1427 LLVMValueRef index_vec
; /* index into the input reg array */
1428 LLVMValueRef index_vec2
= NULL
;
1429 LLVMValueRef inputs_array
;
1430 LLVMTypeRef fptr_type
;
1432 indirect_index
= get_indirect_index(bld
,
1434 reg
->Register
.Index
,
1437 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1441 if (tgsi_type_is_64bit(stype
)) {
1442 index_vec2
= get_soa_array_offsets(&bld_base
->uint_bld
,
1447 /* cast inputs_array pointer to float* */
1448 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1449 inputs_array
= LLVMBuildBitCast(builder
, bld
->inputs_array
, fptr_type
, "");
1451 /* Gather values from the input register array */
1452 res
= build_gather(bld_base
, inputs_array
, index_vec
, NULL
, index_vec2
);
1454 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
1455 LLVMValueRef lindex
= lp_build_const_int32(gallivm
,
1456 reg
->Register
.Index
* 4 + swizzle
);
1457 LLVMValueRef input_ptr
= LLVMBuildGEP(builder
,
1458 bld
->inputs_array
, &lindex
, 1, "");
1460 res
= LLVMBuildLoad(builder
, input_ptr
, "");
1461 if (tgsi_type_is_64bit(stype
)) {
1462 LLVMValueRef lindex1
;
1463 LLVMValueRef input_ptr2
;
1466 lindex1
= lp_build_const_int32(gallivm
,
1467 reg
->Register
.Index
* 4 + swizzle
+ 1);
1468 input_ptr2
= LLVMBuildGEP(builder
,
1469 bld
->inputs_array
, &lindex1
, 1, "");
1470 res2
= LLVMBuildLoad(builder
, input_ptr2
, "");
1471 res
= emit_fetch_64bit(bld_base
, stype
, res
, res2
);
1475 res
= bld
->inputs
[reg
->Register
.Index
][swizzle
];
1476 if (tgsi_type_is_64bit(stype
))
1477 res
= emit_fetch_64bit(bld_base
, stype
, res
, bld
->inputs
[reg
->Register
.Index
][swizzle
+ 1]);
1483 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
|| stype
== TGSI_TYPE_DOUBLE
) {
1484 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1485 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1493 emit_fetch_gs_input(
1494 struct lp_build_tgsi_context
* bld_base
,
1495 const struct tgsi_full_src_register
* reg
,
1496 enum tgsi_opcode_type stype
,
1499 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1500 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1501 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
1502 LLVMBuilderRef builder
= gallivm
->builder
;
1503 LLVMValueRef attrib_index
= NULL
;
1504 LLVMValueRef vertex_index
= NULL
;
1505 LLVMValueRef swizzle_index
= lp_build_const_int32(gallivm
, swizzle
);
1508 if (info
->input_semantic_name
[reg
->Register
.Index
] == TGSI_SEMANTIC_PRIMID
) {
1509 /* This is really a system value not a regular input */
1510 assert(!reg
->Register
.Indirect
);
1511 assert(!reg
->Dimension
.Indirect
);
1512 res
= bld
->system_values
.prim_id
;
1513 if (stype
!= TGSI_TYPE_UNSIGNED
&& stype
!= TGSI_TYPE_SIGNED
) {
1514 res
= LLVMBuildBitCast(builder
, res
, bld_base
->base
.vec_type
, "");
1519 if (reg
->Register
.Indirect
) {
1520 attrib_index
= get_indirect_index(bld
,
1522 reg
->Register
.Index
,
1525 attrib_index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
);
1528 if (reg
->Dimension
.Indirect
) {
1529 vertex_index
= get_indirect_index(bld
,
1531 reg
->Dimension
.Index
,
1534 vertex_index
= lp_build_const_int32(gallivm
, reg
->Dimension
.Index
);
1537 res
= bld
->gs_iface
->fetch_input(bld
->gs_iface
, bld_base
,
1538 reg
->Dimension
.Indirect
,
1540 reg
->Register
.Indirect
,
1545 if (tgsi_type_is_64bit(stype
)) {
1546 LLVMValueRef swizzle_index
= lp_build_const_int32(gallivm
, swizzle
+ 1);
1548 res2
= bld
->gs_iface
->fetch_input(bld
->gs_iface
, bld_base
,
1549 reg
->Dimension
.Indirect
,
1551 reg
->Register
.Indirect
,
1555 res
= emit_fetch_64bit(bld_base
, stype
, res
, res2
);
1556 } else if (stype
== TGSI_TYPE_UNSIGNED
) {
1557 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1558 } else if (stype
== TGSI_TYPE_SIGNED
) {
1559 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1566 emit_fetch_temporary(
1567 struct lp_build_tgsi_context
* bld_base
,
1568 const struct tgsi_full_src_register
* reg
,
1569 enum tgsi_opcode_type stype
,
1572 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1573 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1574 LLVMBuilderRef builder
= gallivm
->builder
;
1577 if (reg
->Register
.Indirect
) {
1578 LLVMValueRef indirect_index
;
1579 LLVMValueRef index_vec
, index_vec2
= NULL
; /* index into the temp reg array */
1580 LLVMValueRef temps_array
;
1581 LLVMTypeRef fptr_type
;
1583 indirect_index
= get_indirect_index(bld
,
1585 reg
->Register
.Index
,
1588 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1592 if (tgsi_type_is_64bit(stype
)) {
1593 index_vec2
= get_soa_array_offsets(&bld_base
->uint_bld
,
1599 /* cast temps_array pointer to float* */
1600 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1601 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
, fptr_type
, "");
1603 /* Gather values from the temporary register array */
1604 res
= build_gather(bld_base
, temps_array
, index_vec
, NULL
, index_vec2
);
1607 LLVMValueRef temp_ptr
;
1608 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
1609 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
1611 if (tgsi_type_is_64bit(stype
)) {
1612 LLVMValueRef temp_ptr2
, res2
;
1614 temp_ptr2
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
+ 1);
1615 res2
= LLVMBuildLoad(builder
, temp_ptr2
, "");
1616 res
= emit_fetch_64bit(bld_base
, stype
, res
, res2
);
1620 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
|| stype
== TGSI_TYPE_DOUBLE
) {
1621 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1622 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1629 emit_fetch_system_value(
1630 struct lp_build_tgsi_context
* bld_base
,
1631 const struct tgsi_full_src_register
* reg
,
1632 enum tgsi_opcode_type stype
,
1635 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1636 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1637 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
1638 LLVMBuilderRef builder
= gallivm
->builder
;
1640 enum tgsi_opcode_type atype
; // Actual type of the value
1642 assert(!reg
->Register
.Indirect
);
1644 switch (info
->system_value_semantic_name
[reg
->Register
.Index
]) {
1645 case TGSI_SEMANTIC_INSTANCEID
:
1646 res
= lp_build_broadcast_scalar(&bld_base
->uint_bld
, bld
->system_values
.instance_id
);
1647 atype
= TGSI_TYPE_UNSIGNED
;
1650 case TGSI_SEMANTIC_VERTEXID
:
1651 res
= bld
->system_values
.vertex_id
;
1652 atype
= TGSI_TYPE_UNSIGNED
;
1655 case TGSI_SEMANTIC_VERTEXID_NOBASE
:
1656 res
= bld
->system_values
.vertex_id_nobase
;
1657 atype
= TGSI_TYPE_UNSIGNED
;
1660 case TGSI_SEMANTIC_BASEVERTEX
:
1661 res
= bld
->system_values
.basevertex
;
1662 atype
= TGSI_TYPE_UNSIGNED
;
1665 case TGSI_SEMANTIC_PRIMID
:
1666 res
= bld
->system_values
.prim_id
;
1667 atype
= TGSI_TYPE_UNSIGNED
;
1670 case TGSI_SEMANTIC_INVOCATIONID
:
1671 res
= lp_build_broadcast_scalar(&bld_base
->uint_bld
, bld
->system_values
.invocation_id
);
1672 atype
= TGSI_TYPE_UNSIGNED
;
1676 assert(!"unexpected semantic in emit_fetch_system_value");
1677 res
= bld_base
->base
.zero
;
1678 atype
= TGSI_TYPE_FLOAT
;
1682 if (atype
!= stype
) {
1683 if (stype
== TGSI_TYPE_FLOAT
) {
1684 res
= LLVMBuildBitCast(builder
, res
, bld_base
->base
.vec_type
, "");
1685 } else if (stype
== TGSI_TYPE_UNSIGNED
) {
1686 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1687 } else if (stype
== TGSI_TYPE_SIGNED
) {
1688 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1696 * Register fetch with derivatives.
1700 struct lp_build_tgsi_soa_context
*bld
,
1709 /* TODO: use interpolation coeffs for inputs */
1712 *ddx
= lp_build_ddx(&bld
->bld_base
.base
, src
);
1715 *ddy
= lp_build_ddy(&bld
->bld_base
.base
, src
);
1723 emit_fetch_predicate(
1724 struct lp_build_tgsi_soa_context
*bld
,
1725 const struct tgsi_full_instruction
*inst
,
1728 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1730 unsigned char swizzles
[4];
1731 LLVMValueRef unswizzled
[4] = {NULL
, NULL
, NULL
, NULL
};
1735 if (!inst
->Instruction
.Predicate
) {
1736 TGSI_FOR_EACH_CHANNEL( chan
) {
1742 swizzles
[0] = inst
->Predicate
.SwizzleX
;
1743 swizzles
[1] = inst
->Predicate
.SwizzleY
;
1744 swizzles
[2] = inst
->Predicate
.SwizzleZ
;
1745 swizzles
[3] = inst
->Predicate
.SwizzleW
;
1747 index
= inst
->Predicate
.Index
;
1748 assert(index
< LP_MAX_TGSI_PREDS
);
1750 TGSI_FOR_EACH_CHANNEL( chan
) {
1751 unsigned swizzle
= swizzles
[chan
];
1754 * Only fetch the predicate register channels that are actually listed
1757 if (!unswizzled
[swizzle
]) {
1758 value
= LLVMBuildLoad(builder
,
1759 bld
->preds
[index
][swizzle
], "");
1762 * Convert the value to an integer mask.
1764 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
1765 * is needlessly causing two comparisons due to storing the intermediate
1766 * result as float vector instead of an integer mask vector.
1768 value
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
1769 bld
->bld_base
.base
.type
,
1772 bld
->bld_base
.base
.zero
);
1773 if (inst
->Predicate
.Negate
) {
1774 value
= LLVMBuildNot(builder
, value
, "");
1777 unswizzled
[swizzle
] = value
;
1779 value
= unswizzled
[swizzle
];
1787 * store an array of 8 64-bit into two arrays of 8 floats
1789 * value is d0, d1, d2, d3 etc.
1790 * each 64-bit has high and low pieces x, y
1791 * so gets stored into the separate channels as:
1792 * chan_ptr = d0.x, d1.x, d2.x, d3.x
1793 * chan_ptr2 = d0.y, d1.y, d2.y, d3.y
1796 emit_store_64bit_chan(struct lp_build_tgsi_context
*bld_base
,
1797 LLVMValueRef chan_ptr
, LLVMValueRef chan_ptr2
,
1801 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1802 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1803 LLVMBuilderRef builder
= gallivm
->builder
;
1804 struct lp_build_context
*float_bld
= &bld_base
->base
;
1806 LLVMValueRef temp
, temp2
;
1807 LLVMValueRef shuffles
[8];
1808 LLVMValueRef shuffles2
[8];
1810 for (i
= 0; i
< bld_base
->base
.type
.length
; i
++) {
1811 shuffles
[i
] = lp_build_const_int32(gallivm
, i
* 2);
1812 shuffles2
[i
] = lp_build_const_int32(gallivm
, (i
* 2) + 1);
1815 temp
= LLVMBuildShuffleVector(builder
, value
,
1816 LLVMGetUndef(LLVMTypeOf(value
)),
1817 LLVMConstVector(shuffles
,
1818 bld_base
->base
.type
.length
),
1820 temp2
= LLVMBuildShuffleVector(builder
, value
,
1821 LLVMGetUndef(LLVMTypeOf(value
)),
1822 LLVMConstVector(shuffles2
,
1823 bld_base
->base
.type
.length
),
1826 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, temp
, chan_ptr
);
1827 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, temp2
, chan_ptr2
);
1835 struct lp_build_tgsi_context
*bld_base
,
1836 const struct tgsi_full_instruction
*inst
,
1838 unsigned chan_index
,
1842 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1843 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1844 LLVMBuilderRef builder
= gallivm
->builder
;
1845 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
1846 struct lp_build_context
*float_bld
= &bld_base
->base
;
1847 struct lp_build_context
*int_bld
= &bld_base
->int_bld
;
1848 LLVMValueRef indirect_index
= NULL
;
1849 enum tgsi_opcode_type dtype
= tgsi_opcode_infer_dst_type(inst
->Instruction
.Opcode
);
1854 * It is always assumed to be float.
1856 if (inst
->Instruction
.Saturate
) {
1857 assert(dtype
== TGSI_TYPE_FLOAT
||
1858 dtype
== TGSI_TYPE_UNTYPED
);
1859 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1860 value
= lp_build_clamp_zero_one_nanzero(float_bld
, value
);
1863 if (reg
->Register
.Indirect
) {
1865 * Currently the mesa/st doesn't generate indirect stores
1866 * to 64-bit values, it normally uses MOV to do indirect stores.
1868 assert(!tgsi_type_is_64bit(dtype
));
1869 indirect_index
= get_indirect_index(bld
,
1871 reg
->Register
.Index
,
1874 assert(reg
->Register
.Index
<=
1875 bld_base
->info
->file_max
[reg
->Register
.File
]);
1878 if (DEBUG_EXECUTION
) {
1879 emit_dump_reg(gallivm
, reg
->Register
.File
, reg
->Register
.Index
, chan_index
, value
);
1882 switch( reg
->Register
.File
) {
1883 case TGSI_FILE_OUTPUT
:
1884 /* Outputs are always stored as floats */
1885 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1887 if (reg
->Register
.Indirect
) {
1888 LLVMValueRef index_vec
; /* indexes into the output registers */
1889 LLVMValueRef outputs_array
;
1890 LLVMTypeRef fptr_type
;
1892 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1897 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1898 outputs_array
= LLVMBuildBitCast(builder
, bld
->outputs_array
, fptr_type
, "");
1900 /* Scatter store values into output registers */
1901 emit_mask_scatter(bld
, outputs_array
, index_vec
, value
,
1902 &bld
->exec_mask
, pred
);
1905 LLVMValueRef out_ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
,
1908 if (tgsi_type_is_64bit(dtype
)) {
1909 LLVMValueRef out_ptr2
= lp_get_output_ptr(bld
, reg
->Register
.Index
,
1911 emit_store_64bit_chan(bld_base
, out_ptr
, out_ptr2
,
1914 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
, out_ptr
);
1918 case TGSI_FILE_TEMPORARY
:
1919 /* Temporaries are always stored as floats */
1920 if (!tgsi_type_is_64bit(dtype
))
1921 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1923 value
= LLVMBuildBitCast(builder
, value
, LLVMVectorType(LLVMFloatTypeInContext(gallivm
->context
), bld_base
->base
.type
.length
* 2), "");
1925 if (reg
->Register
.Indirect
) {
1926 LLVMValueRef index_vec
; /* indexes into the temp registers */
1927 LLVMValueRef temps_array
;
1928 LLVMTypeRef fptr_type
;
1930 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1935 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1936 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
, fptr_type
, "");
1938 /* Scatter store values into temp registers */
1939 emit_mask_scatter(bld
, temps_array
, index_vec
, value
,
1940 &bld
->exec_mask
, pred
);
1943 LLVMValueRef temp_ptr
;
1944 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, chan_index
);
1946 if (tgsi_type_is_64bit(dtype
)) {
1947 LLVMValueRef temp_ptr2
= lp_get_temp_ptr_soa(bld
,
1948 reg
->Register
.Index
,
1950 emit_store_64bit_chan(bld_base
, temp_ptr
, temp_ptr2
,
1954 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
, temp_ptr
);
1958 case TGSI_FILE_ADDRESS
:
1959 assert(dtype
== TGSI_TYPE_SIGNED
);
1960 assert(LLVMTypeOf(value
) == int_bld
->vec_type
);
1961 value
= LLVMBuildBitCast(builder
, value
, int_bld
->vec_type
, "");
1962 lp_exec_mask_store(&bld
->exec_mask
, int_bld
, pred
, value
,
1963 bld
->addr
[reg
->Register
.Index
][chan_index
]);
1966 case TGSI_FILE_PREDICATE
:
1967 assert(LLVMTypeOf(value
) == float_bld
->vec_type
);
1968 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1969 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
,
1970 bld
->preds
[reg
->Register
.Index
][chan_index
]);
1981 * Called at the beginning of the translation of each TGSI instruction, to
1982 * emit some debug code.
1986 struct lp_build_tgsi_context
* bld_base
,
1987 const struct tgsi_full_instruction
* inst
,
1988 const struct tgsi_opcode_info
* info
)
1991 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1993 if (DEBUG_EXECUTION
) {
1995 * Dump the TGSI instruction.
1998 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
2002 tgsi_dump_instruction_str(inst
, bld_base
->pc
, &buf
[2], sizeof buf
- 2);
2003 lp_build_printf(gallivm
, buf
);
2005 /* Dump the execution mask.
2007 if (bld
->exec_mask
.has_mask
) {
2008 lp_build_print_value(gallivm
, " mask = ", bld
->exec_mask
.exec_mask
);
2015 struct lp_build_tgsi_context
* bld_base
,
2016 const struct tgsi_full_instruction
* inst
,
2017 const struct tgsi_opcode_info
* info
,
2018 LLVMValueRef dst
[4])
2021 unsigned chan_index
;
2022 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2023 enum tgsi_opcode_type dtype
= tgsi_opcode_infer_dst_type(inst
->Instruction
.Opcode
);
2025 LLVMValueRef pred
[TGSI_NUM_CHANNELS
];
2027 emit_fetch_predicate( bld
, inst
, pred
);
2029 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
2031 if (tgsi_type_is_64bit(dtype
) && (chan_index
== 1 || chan_index
== 3))
2033 emit_store_chan(bld_base
, inst
, 0, chan_index
, pred
[chan_index
], dst
[chan_index
]);
2039 tgsi_to_pipe_tex_target(unsigned tgsi_target
)
2041 switch (tgsi_target
) {
2042 case TGSI_TEXTURE_BUFFER
:
2044 case TGSI_TEXTURE_1D
:
2045 case TGSI_TEXTURE_SHADOW1D
:
2046 return PIPE_TEXTURE_1D
;
2047 case TGSI_TEXTURE_2D
:
2048 case TGSI_TEXTURE_SHADOW2D
:
2049 case TGSI_TEXTURE_2D_MSAA
:
2050 return PIPE_TEXTURE_2D
;
2051 case TGSI_TEXTURE_3D
:
2052 return PIPE_TEXTURE_3D
;
2053 case TGSI_TEXTURE_CUBE
:
2054 case TGSI_TEXTURE_SHADOWCUBE
:
2055 return PIPE_TEXTURE_CUBE
;
2056 case TGSI_TEXTURE_RECT
:
2057 case TGSI_TEXTURE_SHADOWRECT
:
2058 return PIPE_TEXTURE_RECT
;
2059 case TGSI_TEXTURE_1D_ARRAY
:
2060 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
2061 return PIPE_TEXTURE_1D_ARRAY
;
2062 case TGSI_TEXTURE_2D_ARRAY
:
2063 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
2064 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
2065 return PIPE_TEXTURE_2D_ARRAY
;
2066 case TGSI_TEXTURE_CUBE_ARRAY
:
2067 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
2068 return PIPE_TEXTURE_CUBE_ARRAY
;
2076 static enum lp_sampler_lod_property
2077 lp_build_lod_property(
2078 struct lp_build_tgsi_context
*bld_base
,
2079 const struct tgsi_full_instruction
*inst
,
2082 const struct tgsi_full_src_register
*reg
= &inst
->Src
[src_op
];
2083 enum lp_sampler_lod_property lod_property
;
2086 * Not much we can do here. We could try catching inputs declared
2087 * with constant interpolation but not sure it's worth it - since for
2088 * TEX opcodes as well as FETCH/LD the lod comes from same reg as
2089 * the coords, so it could only work for SAMPLE/TXQ/SVIEWINFO), just
2090 * like the constant/immediate recognition below.
2091 * What seems to be of more value would be to recognize temps holding
2092 * broadcasted scalars but no way we can do it.
2093 * Tried asking llvm but without any success (using LLVMIsConstant
2094 * even though this isn't exactly what we'd need), even as simple as
2095 * IMM[0] UINT32 (0,-1,0,0)
2096 * MOV TEMP[0] IMM[0].yyyy
2097 * SVIEWINFO TEMP[1], TEMP[0].xxxx, SVIEWINFO[0]
2099 * This means there's ZERO chance this will ever catch a scalar lod
2100 * with traditional tex opcodes as well as texel fetches, since the lod
2101 * comes from the same reg as coords (except some test shaders using
2102 * constant coords maybe).
2103 * There's at least hope for sample opcodes as well as size queries.
2105 if (reg
->Register
.File
== TGSI_FILE_CONSTANT
||
2106 reg
->Register
.File
== TGSI_FILE_IMMEDIATE
) {
2107 lod_property
= LP_SAMPLER_LOD_SCALAR
;
2109 else if (bld_base
->info
->processor
== PIPE_SHADER_FRAGMENT
) {
2110 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
2111 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2114 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
2118 /* never use scalar (per-quad) lod the results are just too wrong. */
2119 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2121 return lod_property
;
2126 * High-level instruction translators.
2130 emit_tex( struct lp_build_tgsi_soa_context
*bld
,
2131 const struct tgsi_full_instruction
*inst
,
2132 enum lp_build_tex_modifier modifier
,
2133 LLVMValueRef
*texel
,
2134 unsigned sampler_reg
,
2135 enum lp_sampler_op_type sampler_op
)
2137 unsigned unit
= inst
->Src
[sampler_reg
].Register
.Index
;
2138 LLVMValueRef oow
= NULL
;
2139 LLVMValueRef lod
= NULL
;
2140 LLVMValueRef coords
[5];
2141 LLVMValueRef offsets
[3] = { NULL
};
2142 struct lp_derivatives derivs
;
2143 struct lp_sampler_params params
;
2144 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
2145 unsigned num_derivs
, num_offsets
, i
;
2146 unsigned shadow_coord
= 0;
2147 unsigned layer_coord
= 0;
2148 unsigned sample_key
= sampler_op
<< LP_SAMPLER_OP_TYPE_SHIFT
;
2150 memset(¶ms
, 0, sizeof(params
));
2152 if (!bld
->sampler
) {
2153 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2154 for (i
= 0; i
< 4; i
++) {
2155 texel
[i
] = bld
->bld_base
.base
.undef
;
2160 switch (inst
->Texture
.Texture
) {
2161 case TGSI_TEXTURE_1D_ARRAY
:
2164 case TGSI_TEXTURE_1D
:
2168 case TGSI_TEXTURE_2D_ARRAY
:
2171 case TGSI_TEXTURE_2D
:
2172 case TGSI_TEXTURE_RECT
:
2176 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
2179 case TGSI_TEXTURE_SHADOW1D
:
2184 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
2190 case TGSI_TEXTURE_SHADOW2D
:
2191 case TGSI_TEXTURE_SHADOWRECT
:
2196 case TGSI_TEXTURE_CUBE
:
2200 case TGSI_TEXTURE_3D
:
2204 case TGSI_TEXTURE_SHADOWCUBE
:
2209 case TGSI_TEXTURE_CUBE_ARRAY
:
2214 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
2218 shadow_coord
= 4; /* shadow coord special different reg */
2220 case TGSI_TEXTURE_2D_MSAA
:
2221 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
2227 /* Note lod and especially projected are illegal in a LOT of cases */
2228 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
||
2229 modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2230 if (inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE
||
2231 inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
) {
2232 /* note that shadow cube array with bias/explicit lod does not exist */
2233 lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, 0);
2236 lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2238 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
2239 sample_key
|= LP_SAMPLER_LOD_BIAS
<< LP_SAMPLER_LOD_CONTROL_SHIFT
;
2241 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2242 sample_key
|= LP_SAMPLER_LOD_EXPLICIT
<< LP_SAMPLER_LOD_CONTROL_SHIFT
;
2244 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2247 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
) {
2248 oow
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2249 oow
= lp_build_rcp(&bld
->bld_base
.base
, oow
);
2252 for (i
= 0; i
< num_derivs
; i
++) {
2253 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2254 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2255 coords
[i
] = lp_build_mul(&bld
->bld_base
.base
, coords
[i
], oow
);
2257 for (i
= num_derivs
; i
< 5; i
++) {
2258 coords
[i
] = bld
->bld_base
.base
.undef
;
2261 /* Layer coord always goes into 3rd slot, except for cube map arrays */
2263 if (layer_coord
== 3) {
2264 coords
[3] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2267 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2269 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2270 coords
[2] = lp_build_mul(&bld
->bld_base
.base
, coords
[2], oow
);
2272 /* Shadow coord occupies always 5th slot. */
2274 sample_key
|= LP_SAMPLER_SHADOW
;
2275 if (shadow_coord
== 4) {
2276 coords
[4] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, 0);
2279 coords
[4] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, shadow_coord
);
2281 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2282 coords
[4] = lp_build_mul(&bld
->bld_base
.base
, coords
[4], oow
);
2285 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
2287 sample_key
|= LP_SAMPLER_LOD_DERIVATIVES
<< LP_SAMPLER_LOD_CONTROL_SHIFT
;
2288 for (dim
= 0; dim
< num_derivs
; ++dim
) {
2289 derivs
.ddx
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, dim
);
2290 derivs
.ddy
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, dim
);
2292 params
.derivs
= &derivs
;
2294 * could also check all src regs if constant but I doubt such
2295 * cases exist in practice.
2297 if (bld
->bld_base
.info
->processor
== PIPE_SHADER_FRAGMENT
) {
2298 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
2299 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2302 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
2306 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2309 sample_key
|= lod_property
<< LP_SAMPLER_LOD_PROPERTY_SHIFT
;
2311 /* we don't handle the 4 offset version of tg4 */
2312 if (inst
->Texture
.NumOffsets
== 1) {
2314 sample_key
|= LP_SAMPLER_OFFSETS
;
2315 for (dim
= 0; dim
< num_offsets
; dim
++) {
2316 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2320 params
.type
= bld
->bld_base
.base
.type
;
2321 params
.sample_key
= sample_key
;
2322 params
.texture_index
= unit
;
2323 params
.sampler_index
= unit
;
2324 params
.context_ptr
= bld
->context_ptr
;
2325 params
.thread_data_ptr
= bld
->thread_data_ptr
;
2326 params
.coords
= coords
;
2327 params
.offsets
= offsets
;
2329 params
.texel
= texel
;
2331 bld
->sampler
->emit_tex_sample(bld
->sampler
,
2332 bld
->bld_base
.base
.gallivm
,
2337 emit_sample(struct lp_build_tgsi_soa_context
*bld
,
2338 const struct tgsi_full_instruction
*inst
,
2339 enum lp_build_tex_modifier modifier
,
2341 LLVMValueRef
*texel
)
2343 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2344 unsigned texture_unit
, sampler_unit
;
2345 LLVMValueRef lod
= NULL
;
2346 LLVMValueRef coords
[5];
2347 LLVMValueRef offsets
[3] = { NULL
};
2348 struct lp_derivatives derivs
;
2349 struct lp_sampler_params params
;
2350 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
2352 unsigned num_offsets
, num_derivs
, i
;
2353 unsigned layer_coord
= 0;
2354 unsigned sample_key
= LP_SAMPLER_OP_TEXTURE
<< LP_SAMPLER_OP_TYPE_SHIFT
;
2356 memset(¶ms
, 0, sizeof(params
));
2358 if (!bld
->sampler
) {
2359 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2360 for (i
= 0; i
< 4; i
++) {
2361 texel
[i
] = bld
->bld_base
.base
.undef
;
2367 * unlike old-style tex opcodes the texture/sampler indices
2368 * always come from src1 and src2 respectively.
2370 texture_unit
= inst
->Src
[1].Register
.Index
;
2371 sampler_unit
= inst
->Src
[2].Register
.Index
;
2374 * Note inst->Texture.Texture will contain the number of offsets,
2375 * however the target information is NOT there and comes from the
2376 * declared sampler views instead.
2378 switch (bld
->sv
[texture_unit
].Resource
) {
2379 case TGSI_TEXTURE_1D
:
2383 case TGSI_TEXTURE_1D_ARRAY
:
2388 case TGSI_TEXTURE_2D
:
2389 case TGSI_TEXTURE_RECT
:
2393 case TGSI_TEXTURE_2D_ARRAY
:
2398 case TGSI_TEXTURE_CUBE
:
2402 case TGSI_TEXTURE_3D
:
2406 case TGSI_TEXTURE_CUBE_ARRAY
:
2416 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
||
2417 modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2418 lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, 0);
2419 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
2420 sample_key
|= LP_SAMPLER_LOD_BIAS
<< LP_SAMPLER_LOD_CONTROL_SHIFT
;
2422 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2423 sample_key
|= LP_SAMPLER_LOD_EXPLICIT
<< LP_SAMPLER_LOD_CONTROL_SHIFT
;
2425 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2427 else if (modifier
== LP_BLD_TEX_MODIFIER_LOD_ZERO
) {
2428 /* XXX might be better to explicitly pass the level zero information */
2429 sample_key
|= LP_SAMPLER_LOD_EXPLICIT
<< LP_SAMPLER_LOD_CONTROL_SHIFT
;
2430 lod
= lp_build_const_vec(gallivm
, bld
->bld_base
.base
.type
, 0.0F
);
2433 for (i
= 0; i
< num_derivs
; i
++) {
2434 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2436 for (i
= num_derivs
; i
< 5; i
++) {
2437 coords
[i
] = bld
->bld_base
.base
.undef
;
2440 /* Layer coord always goes into 3rd slot, except for cube map arrays */
2442 if (layer_coord
== 3)
2443 coords
[3] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2445 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2447 /* Shadow coord occupies always 5th slot. */
2449 sample_key
|= LP_SAMPLER_SHADOW
;
2450 coords
[4] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, 0);
2453 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
2455 sample_key
|= LP_SAMPLER_LOD_DERIVATIVES
<< LP_SAMPLER_LOD_CONTROL_SHIFT
;
2456 for (dim
= 0; dim
< num_derivs
; ++dim
) {
2457 derivs
.ddx
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, dim
);
2458 derivs
.ddy
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 4, dim
);
2460 params
.derivs
= &derivs
;
2462 * could also check all src regs if constant but I doubt such
2463 * cases exist in practice.
2465 if (bld
->bld_base
.info
->processor
== PIPE_SHADER_FRAGMENT
) {
2466 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
2467 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2470 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
2474 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2478 /* some advanced gather instructions (txgo) would require 4 offsets */
2479 if (inst
->Texture
.NumOffsets
== 1) {
2481 sample_key
|= LP_SAMPLER_OFFSETS
;
2482 for (dim
= 0; dim
< num_offsets
; dim
++) {
2483 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2486 sample_key
|= lod_property
<< LP_SAMPLER_LOD_PROPERTY_SHIFT
;
2488 params
.type
= bld
->bld_base
.base
.type
;
2489 params
.sample_key
= sample_key
;
2490 params
.texture_index
= texture_unit
;
2491 params
.sampler_index
= sampler_unit
;
2492 params
.context_ptr
= bld
->context_ptr
;
2493 params
.thread_data_ptr
= bld
->thread_data_ptr
;
2494 params
.coords
= coords
;
2495 params
.offsets
= offsets
;
2497 params
.texel
= texel
;
2499 bld
->sampler
->emit_tex_sample(bld
->sampler
,
2500 bld
->bld_base
.base
.gallivm
,
2503 if (inst
->Src
[1].Register
.SwizzleX
!= PIPE_SWIZZLE_X
||
2504 inst
->Src
[1].Register
.SwizzleY
!= PIPE_SWIZZLE_Y
||
2505 inst
->Src
[1].Register
.SwizzleZ
!= PIPE_SWIZZLE_Z
||
2506 inst
->Src
[1].Register
.SwizzleW
!= PIPE_SWIZZLE_W
) {
2507 unsigned char swizzles
[4];
2508 swizzles
[0] = inst
->Src
[1].Register
.SwizzleX
;
2509 swizzles
[1] = inst
->Src
[1].Register
.SwizzleY
;
2510 swizzles
[2] = inst
->Src
[1].Register
.SwizzleZ
;
2511 swizzles
[3] = inst
->Src
[1].Register
.SwizzleW
;
2513 lp_build_swizzle_soa_inplace(&bld
->bld_base
.base
, texel
, swizzles
);
2518 emit_fetch_texels( struct lp_build_tgsi_soa_context
*bld
,
2519 const struct tgsi_full_instruction
*inst
,
2520 LLVMValueRef
*texel
,
2523 unsigned unit
, target
;
2524 LLVMValueRef coord_undef
= LLVMGetUndef(bld
->bld_base
.base
.int_vec_type
);
2525 LLVMValueRef explicit_lod
= NULL
;
2526 LLVMValueRef coords
[5];
2527 LLVMValueRef offsets
[3] = { NULL
};
2528 struct lp_sampler_params params
;
2529 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
2531 unsigned layer_coord
= 0;
2532 unsigned sample_key
= LP_SAMPLER_OP_FETCH
<< LP_SAMPLER_OP_TYPE_SHIFT
;
2534 memset(¶ms
, 0, sizeof(params
));
2536 if (!bld
->sampler
) {
2537 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2538 for (i
= 0; i
< 4; i
++) {
2539 texel
[i
] = coord_undef
;
2544 unit
= inst
->Src
[1].Register
.Index
;
2547 target
= bld
->sv
[unit
].Resource
;
2550 target
= inst
->Texture
.Texture
;
2554 case TGSI_TEXTURE_1D
:
2555 case TGSI_TEXTURE_BUFFER
:
2558 case TGSI_TEXTURE_1D_ARRAY
:
2562 case TGSI_TEXTURE_2D
:
2563 case TGSI_TEXTURE_RECT
:
2564 case TGSI_TEXTURE_2D_MSAA
:
2567 case TGSI_TEXTURE_2D_ARRAY
:
2568 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
2572 case TGSI_TEXTURE_3D
:
2580 /* always have lod except for buffers and msaa targets ? */
2581 if (target
!= TGSI_TEXTURE_BUFFER
&&
2582 target
!= TGSI_TEXTURE_2D_MSAA
&&
2583 target
!= TGSI_TEXTURE_2D_ARRAY_MSAA
) {
2584 sample_key
|= LP_SAMPLER_LOD_EXPLICIT
<< LP_SAMPLER_LOD_CONTROL_SHIFT
;
2585 explicit_lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2586 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2589 * XXX: for real msaa support, the w component (or src2.x for sample_i_ms)
2590 * would be the sample index.
2593 for (i
= 0; i
< dims
; i
++) {
2594 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2596 /* never use more than 3 coords here but emit_fetch_texel copies all 5 anyway */
2597 for (i
= dims
; i
< 5; i
++) {
2598 coords
[i
] = coord_undef
;
2601 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2603 if (inst
->Texture
.NumOffsets
== 1) {
2605 sample_key
|= LP_SAMPLER_OFFSETS
;
2606 for (dim
= 0; dim
< dims
; dim
++) {
2607 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2610 sample_key
|= lod_property
<< LP_SAMPLER_LOD_PROPERTY_SHIFT
;
2612 params
.type
= bld
->bld_base
.base
.type
;
2613 params
.sample_key
= sample_key
;
2614 params
.texture_index
= unit
;
2616 * sampler not actually used, set to 0 so it won't exceed PIPE_MAX_SAMPLERS
2617 * and trigger some assertions with d3d10 where the sampler view number
2620 params
.sampler_index
= 0;
2621 params
.context_ptr
= bld
->context_ptr
;
2622 params
.thread_data_ptr
= bld
->thread_data_ptr
;
2623 params
.coords
= coords
;
2624 params
.offsets
= offsets
;
2625 params
.derivs
= NULL
;
2626 params
.lod
= explicit_lod
;
2627 params
.texel
= texel
;
2629 bld
->sampler
->emit_tex_sample(bld
->sampler
,
2630 bld
->bld_base
.base
.gallivm
,
2634 (inst
->Src
[1].Register
.SwizzleX
!= PIPE_SWIZZLE_X
||
2635 inst
->Src
[1].Register
.SwizzleY
!= PIPE_SWIZZLE_Y
||
2636 inst
->Src
[1].Register
.SwizzleZ
!= PIPE_SWIZZLE_Z
||
2637 inst
->Src
[1].Register
.SwizzleW
!= PIPE_SWIZZLE_W
)) {
2638 unsigned char swizzles
[4];
2639 swizzles
[0] = inst
->Src
[1].Register
.SwizzleX
;
2640 swizzles
[1] = inst
->Src
[1].Register
.SwizzleY
;
2641 swizzles
[2] = inst
->Src
[1].Register
.SwizzleZ
;
2642 swizzles
[3] = inst
->Src
[1].Register
.SwizzleW
;
2644 lp_build_swizzle_soa_inplace(&bld
->bld_base
.base
, texel
, swizzles
);
2649 emit_size_query( struct lp_build_tgsi_soa_context
*bld
,
2650 const struct tgsi_full_instruction
*inst
,
2651 LLVMValueRef
*sizes_out
,
2652 boolean is_sviewinfo
)
2654 LLVMValueRef explicit_lod
;
2655 enum lp_sampler_lod_property lod_property
;
2658 unsigned unit
= inst
->Src
[1].Register
.Index
;
2659 unsigned target
, pipe_target
;
2660 struct lp_sampler_size_query_params params
;
2663 target
= bld
->sv
[unit
].Resource
;
2666 target
= inst
->Texture
.Texture
;
2669 case TGSI_TEXTURE_BUFFER
:
2670 case TGSI_TEXTURE_RECT
:
2671 case TGSI_TEXTURE_SHADOWRECT
:
2679 if (!bld
->sampler
) {
2680 _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
2681 for (i
= 0; i
< 4; i
++)
2682 sizes_out
[i
] = bld
->bld_base
.int_bld
.undef
;
2687 explicit_lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 0);
2688 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2691 explicit_lod
= NULL
;
2692 lod_property
= LP_SAMPLER_LOD_SCALAR
;
2696 pipe_target
= tgsi_to_pipe_tex_target(target
);
2698 params
.int_type
= bld
->bld_base
.int_bld
.type
;
2699 params
.texture_unit
= unit
;
2700 params
.target
= pipe_target
;
2701 params
.context_ptr
= bld
->context_ptr
;
2702 params
.is_sviewinfo
= TRUE
;
2703 params
.lod_property
= lod_property
;
2704 params
.explicit_lod
= explicit_lod
;
2705 params
.sizes_out
= sizes_out
;
2707 bld
->sampler
->emit_size_query(bld
->sampler
,
2708 bld
->bld_base
.base
.gallivm
,
2713 near_end_of_shader(struct lp_build_tgsi_soa_context
*bld
,
2718 for (i
= 0; i
< 5; i
++) {
2721 if (pc
+ i
>= bld
->bld_base
.info
->num_instructions
)
2724 opcode
= bld
->bld_base
.instructions
[pc
+ i
].Instruction
.Opcode
;
2726 if (opcode
== TGSI_OPCODE_END
)
2729 if (opcode
== TGSI_OPCODE_TEX
||
2730 opcode
== TGSI_OPCODE_TXP
||
2731 opcode
== TGSI_OPCODE_TXD
||
2732 opcode
== TGSI_OPCODE_TXB
||
2733 opcode
== TGSI_OPCODE_TXL
||
2734 opcode
== TGSI_OPCODE_TXF
||
2735 opcode
== TGSI_OPCODE_TXQ
||
2736 opcode
== TGSI_OPCODE_TEX2
||
2737 opcode
== TGSI_OPCODE_TXB2
||
2738 opcode
== TGSI_OPCODE_TXL2
||
2739 opcode
== TGSI_OPCODE_SAMPLE
||
2740 opcode
== TGSI_OPCODE_SAMPLE_B
||
2741 opcode
== TGSI_OPCODE_SAMPLE_C
||
2742 opcode
== TGSI_OPCODE_SAMPLE_C_LZ
||
2743 opcode
== TGSI_OPCODE_SAMPLE_D
||
2744 opcode
== TGSI_OPCODE_SAMPLE_I
||
2745 opcode
== TGSI_OPCODE_SAMPLE_I_MS
||
2746 opcode
== TGSI_OPCODE_SAMPLE_L
||
2747 opcode
== TGSI_OPCODE_SVIEWINFO
||
2748 opcode
== TGSI_OPCODE_CAL
||
2749 opcode
== TGSI_OPCODE_CALLNZ
||
2750 opcode
== TGSI_OPCODE_IF
||
2751 opcode
== TGSI_OPCODE_UIF
||
2752 opcode
== TGSI_OPCODE_BGNLOOP
||
2753 opcode
== TGSI_OPCODE_SWITCH
)
2763 * Kill fragment if any of the src register values are negative.
2767 struct lp_build_tgsi_soa_context
*bld
,
2768 const struct tgsi_full_instruction
*inst
,
2771 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2772 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
2773 LLVMValueRef terms
[TGSI_NUM_CHANNELS
];
2775 unsigned chan_index
;
2777 memset(&terms
, 0, sizeof terms
);
2779 TGSI_FOR_EACH_CHANNEL( chan_index
) {
2782 /* Unswizzle channel */
2783 swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
2785 /* Check if the component has not been already tested. */
2786 assert(swizzle
< TGSI_NUM_CHANNELS
);
2787 if( !terms
[swizzle
] )
2788 /* TODO: change the comparison operator instead of setting the sign */
2789 terms
[swizzle
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, chan_index
);
2793 TGSI_FOR_EACH_CHANNEL( chan_index
) {
2794 if(terms
[chan_index
]) {
2795 LLVMValueRef chan_mask
;
2798 * If term < 0 then mask = 0 else mask = ~0.
2800 chan_mask
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, terms
[chan_index
], bld
->bld_base
.base
.zero
);
2803 mask
= LLVMBuildAnd(builder
, mask
, chan_mask
, "");
2809 if (bld
->exec_mask
.has_mask
) {
2810 LLVMValueRef invmask
;
2811 invmask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
2812 mask
= LLVMBuildOr(builder
, mask
, invmask
, "");
2815 lp_build_mask_update(bld
->mask
, mask
);
2816 if (!near_end_of_shader(bld
, pc
))
2817 lp_build_mask_check(bld
->mask
);
2822 * Unconditional fragment kill.
2823 * The only predication is the execution mask which will apply if
2824 * we're inside a loop or conditional.
2827 emit_kill(struct lp_build_tgsi_soa_context
*bld
,
2830 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2833 /* For those channels which are "alive", disable fragment shader
2836 if (bld
->exec_mask
.has_mask
) {
2837 mask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
2840 LLVMValueRef zero
= LLVMConstNull(bld
->bld_base
.base
.int_vec_type
);
2844 lp_build_mask_update(bld
->mask
, mask
);
2846 if (!near_end_of_shader(bld
, pc
))
2847 lp_build_mask_check(bld
->mask
);
2852 * Emit code which will dump the value of all the temporary registers
2856 emit_dump_file(struct lp_build_tgsi_soa_context
*bld
,
2859 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
2860 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2861 LLVMBuilderRef builder
= gallivm
->builder
;
2862 LLVMValueRef reg_ptr
;
2864 int max_index
= info
->file_max
[file
];
2867 * Some register files, particularly constants, can be very large,
2868 * and dumping everything could make this unusably slow.
2870 max_index
= MIN2(max_index
, 32);
2872 for (index
= 0; index
<= max_index
; index
++) {
2877 if (index
< 8 * sizeof(unsigned) &&
2878 (info
->file_mask
[file
] & (1u << index
)) == 0) {
2879 /* This was not declared.*/
2883 if (file
== TGSI_FILE_INPUT
) {
2884 mask
= info
->input_usage_mask
[index
];
2886 mask
= TGSI_WRITEMASK_XYZW
;
2889 for (chan
= 0; chan
< 4; chan
++) {
2890 if ((mask
& (1 << chan
)) == 0) {
2891 /* This channel is not used.*/
2895 if (file
== TGSI_FILE_CONSTANT
) {
2896 struct tgsi_full_src_register reg
;
2897 memset(®
, 0, sizeof reg
);
2898 reg
.Register
.File
= file
;
2899 reg
.Register
.Index
= index
;
2900 reg
.Register
.SwizzleX
= 0;
2901 reg
.Register
.SwizzleY
= 1;
2902 reg
.Register
.SwizzleZ
= 2;
2903 reg
.Register
.SwizzleW
= 3;
2905 res
= bld
->bld_base
.emit_fetch_funcs
[file
](&bld
->bld_base
, ®
, TGSI_TYPE_FLOAT
, chan
);
2909 } else if (file
== TGSI_FILE_INPUT
) {
2910 res
= bld
->inputs
[index
][chan
];
2914 } else if (file
== TGSI_FILE_TEMPORARY
) {
2915 reg_ptr
= lp_get_temp_ptr_soa(bld
, index
, chan
);
2917 res
= LLVMBuildLoad(builder
, reg_ptr
, "");
2918 } else if (file
== TGSI_FILE_OUTPUT
) {
2919 reg_ptr
= lp_get_output_ptr(bld
, index
, chan
);
2921 res
= LLVMBuildLoad(builder
, reg_ptr
, "");
2927 emit_dump_reg(gallivm
, file
, index
, chan
, res
);
2935 lp_emit_declaration_soa(
2936 struct lp_build_tgsi_context
*bld_base
,
2937 const struct tgsi_full_declaration
*decl
)
2939 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
2940 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2941 LLVMTypeRef vec_type
= bld
->bld_base
.base
.vec_type
;
2942 const unsigned first
= decl
->Range
.First
;
2943 const unsigned last
= decl
->Range
.Last
;
2946 assert(last
<= bld
->bld_base
.info
->file_max
[decl
->Declaration
.File
]);
2948 switch (decl
->Declaration
.File
) {
2949 case TGSI_FILE_TEMPORARY
:
2950 if (!(bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
))) {
2951 assert(last
< LP_MAX_INLINED_TEMPS
);
2952 for (idx
= first
; idx
<= last
; ++idx
) {
2953 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2954 bld
->temps
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
, "temp");
2959 case TGSI_FILE_OUTPUT
:
2960 if (!(bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
2961 for (idx
= first
; idx
<= last
; ++idx
) {
2962 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2963 bld
->outputs
[idx
][i
] = lp_build_alloca(gallivm
,
2964 vec_type
, "output");
2969 case TGSI_FILE_ADDRESS
:
2970 /* ADDR registers are only allocated with an integer LLVM IR type,
2971 * as they are guaranteed to always have integers.
2972 * XXX: Not sure if this exception is worthwhile (or the whole idea of
2973 * an ADDR register for that matter).
2975 assert(last
< LP_MAX_TGSI_ADDRS
);
2976 for (idx
= first
; idx
<= last
; ++idx
) {
2977 assert(idx
< LP_MAX_TGSI_ADDRS
);
2978 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2979 bld
->addr
[idx
][i
] = lp_build_alloca(gallivm
, bld_base
->base
.int_vec_type
, "addr");
2983 case TGSI_FILE_PREDICATE
:
2984 assert(last
< LP_MAX_TGSI_PREDS
);
2985 for (idx
= first
; idx
<= last
; ++idx
) {
2986 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2987 bld
->preds
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
,
2992 case TGSI_FILE_SAMPLER_VIEW
:
2994 * The target stored here MUST match whatever there actually
2995 * is in the set sampler views (what about return type?).
2997 assert(last
< PIPE_MAX_SHADER_SAMPLER_VIEWS
);
2998 for (idx
= first
; idx
<= last
; ++idx
) {
2999 bld
->sv
[idx
] = decl
->SamplerView
;
3003 case TGSI_FILE_CONSTANT
:
3006 * We could trivially fetch the per-buffer pointer when fetching the
3007 * constant, relying on llvm to figure out it's always the same pointer
3008 * anyway. However, doing so results in a huge (more than factor of 10)
3009 * slowdown in llvm compilation times for some (but not all) shaders
3010 * (more specifically, the IR optimization spends way more time in
3011 * DominatorTree::dominates). At least with llvm versions 3.1, 3.3.
3013 unsigned idx2D
= decl
->Dim
.Index2D
;
3014 LLVMValueRef index2D
= lp_build_const_int32(gallivm
, idx2D
);
3015 assert(idx2D
< LP_MAX_TGSI_CONST_BUFFERS
);
3016 bld
->consts
[idx2D
] =
3017 lp_build_array_get(gallivm
, bld
->consts_ptr
, index2D
);
3018 bld
->consts_sizes
[idx2D
] =
3019 lp_build_array_get(gallivm
, bld
->const_sizes_ptr
, index2D
);
3024 /* don't need to declare other vars */
3030 void lp_emit_immediate_soa(
3031 struct lp_build_tgsi_context
*bld_base
,
3032 const struct tgsi_full_immediate
*imm
)
3034 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
3035 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
3036 LLVMValueRef imms
[4];
3038 const uint size
= imm
->Immediate
.NrTokens
- 1;
3040 switch (imm
->Immediate
.DataType
) {
3041 case TGSI_IMM_FLOAT32
:
3042 for( i
= 0; i
< size
; ++i
)
3044 lp_build_const_vec(gallivm
, bld_base
->base
.type
, imm
->u
[i
].Float
);
3047 case TGSI_IMM_FLOAT64
:
3048 case TGSI_IMM_UINT32
:
3049 for( i
= 0; i
< size
; ++i
) {
3050 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->uint_bld
.type
, imm
->u
[i
].Uint
);
3051 imms
[i
] = LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
3055 case TGSI_IMM_INT32
:
3056 for( i
= 0; i
< size
; ++i
) {
3057 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->int_bld
.type
, imm
->u
[i
].Int
);
3058 imms
[i
] = LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
3063 for( i
= size
; i
< 4; ++i
)
3064 imms
[i
] = bld_base
->base
.undef
;
3066 if (bld
->use_immediates_array
) {
3067 unsigned index
= bld
->num_immediates
;
3068 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
3069 LLVMBuilderRef builder
= gallivm
->builder
;
3071 assert(bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
));
3072 for (i
= 0; i
< 4; ++i
) {
3073 LLVMValueRef lindex
= lp_build_const_int32(
3074 bld
->bld_base
.base
.gallivm
, index
* 4 + i
);
3075 LLVMValueRef imm_ptr
= LLVMBuildGEP(builder
,
3076 bld
->imms_array
, &lindex
, 1, "");
3077 LLVMBuildStore(builder
, imms
[i
], imm_ptr
);
3080 /* simply copy the immediate values into the next immediates[] slot */
3082 assert(imm
->Immediate
.NrTokens
- 1 <= 4);
3083 assert(bld
->num_immediates
< LP_MAX_INLINED_IMMEDIATES
);
3085 for(i
= 0; i
< 4; ++i
)
3086 bld
->immediates
[bld
->num_immediates
][i
] = imms
[i
];
3088 if (bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
)) {
3089 unsigned index
= bld
->num_immediates
;
3090 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
3091 LLVMBuilderRef builder
= gallivm
->builder
;
3092 for (i
= 0; i
< 4; ++i
) {
3093 LLVMValueRef lindex
= lp_build_const_int32(
3094 bld
->bld_base
.base
.gallivm
, index
* 4 + i
);
3095 LLVMValueRef imm_ptr
= LLVMBuildGEP(builder
,
3096 bld
->imms_array
, &lindex
, 1, "");
3097 LLVMBuildStore(builder
,
3098 bld
->immediates
[index
][i
],
3104 bld
->num_immediates
++;
3109 const struct lp_build_tgsi_action
* action
,
3110 struct lp_build_tgsi_context
* bld_base
,
3111 struct lp_build_emit_data
* emit_data
)
3113 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3115 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
,
3116 &emit_data
->output
[emit_data
->chan
], NULL
);
3121 const struct lp_build_tgsi_action
* action
,
3122 struct lp_build_tgsi_context
* bld_base
,
3123 struct lp_build_emit_data
* emit_data
)
3125 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3127 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
, NULL
,
3128 &emit_data
->output
[emit_data
->chan
]);
3133 const struct lp_build_tgsi_action
* action
,
3134 struct lp_build_tgsi_context
* bld_base
,
3135 struct lp_build_emit_data
* emit_data
)
3137 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3139 emit_kill(bld
, bld_base
->pc
- 1);
3144 const struct lp_build_tgsi_action
* action
,
3145 struct lp_build_tgsi_context
* bld_base
,
3146 struct lp_build_emit_data
* emit_data
)
3148 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3150 emit_kill_if(bld
, emit_data
->inst
, bld_base
->pc
- 1);
3155 const struct lp_build_tgsi_action
* action
,
3156 struct lp_build_tgsi_context
* bld_base
,
3157 struct lp_build_emit_data
* emit_data
)
3159 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3161 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
3162 emit_data
->output
, 1, LP_SAMPLER_OP_TEXTURE
);
3167 const struct lp_build_tgsi_action
* action
,
3168 struct lp_build_tgsi_context
* bld_base
,
3169 struct lp_build_emit_data
* emit_data
)
3171 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3173 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
3174 emit_data
->output
, 2, LP_SAMPLER_OP_TEXTURE
);
3179 const struct lp_build_tgsi_action
* action
,
3180 struct lp_build_tgsi_context
* bld_base
,
3181 struct lp_build_emit_data
* emit_data
)
3183 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3185 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
3186 emit_data
->output
, 1, LP_SAMPLER_OP_TEXTURE
);
3191 const struct lp_build_tgsi_action
* action
,
3192 struct lp_build_tgsi_context
* bld_base
,
3193 struct lp_build_emit_data
* emit_data
)
3195 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3197 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
3198 emit_data
->output
, 2, LP_SAMPLER_OP_TEXTURE
);
3203 const struct lp_build_tgsi_action
* action
,
3204 struct lp_build_tgsi_context
* bld_base
,
3205 struct lp_build_emit_data
* emit_data
)
3207 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3209 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
3210 emit_data
->output
, 3, LP_SAMPLER_OP_TEXTURE
);
3215 const struct lp_build_tgsi_action
* action
,
3216 struct lp_build_tgsi_context
* bld_base
,
3217 struct lp_build_emit_data
* emit_data
)
3219 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3221 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
3222 emit_data
->output
, 1, LP_SAMPLER_OP_TEXTURE
);
3227 const struct lp_build_tgsi_action
* action
,
3228 struct lp_build_tgsi_context
* bld_base
,
3229 struct lp_build_emit_data
* emit_data
)
3231 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3233 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
3234 emit_data
->output
, 2, LP_SAMPLER_OP_TEXTURE
);
3239 const struct lp_build_tgsi_action
* action
,
3240 struct lp_build_tgsi_context
* bld_base
,
3241 struct lp_build_emit_data
* emit_data
)
3243 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3245 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_PROJECTED
,
3246 emit_data
->output
, 1, LP_SAMPLER_OP_TEXTURE
);
3251 const struct lp_build_tgsi_action
* action
,
3252 struct lp_build_tgsi_context
* bld_base
,
3253 struct lp_build_emit_data
* emit_data
)
3255 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3257 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
3258 emit_data
->output
, 2, LP_SAMPLER_OP_GATHER
);
3263 const struct lp_build_tgsi_action
* action
,
3264 struct lp_build_tgsi_context
* bld_base
,
3265 struct lp_build_emit_data
* emit_data
)
3267 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3269 emit_size_query(bld
, emit_data
->inst
, emit_data
->output
, FALSE
);
3274 const struct lp_build_tgsi_action
* action
,
3275 struct lp_build_tgsi_context
* bld_base
,
3276 struct lp_build_emit_data
* emit_data
)
3278 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3280 emit_fetch_texels(bld
, emit_data
->inst
, emit_data
->output
, FALSE
);
3285 const struct lp_build_tgsi_action
* action
,
3286 struct lp_build_tgsi_context
* bld_base
,
3287 struct lp_build_emit_data
* emit_data
)
3289 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3291 emit_fetch_texels(bld
, emit_data
->inst
, emit_data
->output
, TRUE
);
3296 const struct lp_build_tgsi_action
* action
,
3297 struct lp_build_tgsi_context
* bld_base
,
3298 struct lp_build_emit_data
* emit_data
)
3300 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3302 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
3303 FALSE
, emit_data
->output
);
3308 const struct lp_build_tgsi_action
* action
,
3309 struct lp_build_tgsi_context
* bld_base
,
3310 struct lp_build_emit_data
* emit_data
)
3312 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3314 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
3315 FALSE
, emit_data
->output
);
3320 const struct lp_build_tgsi_action
* action
,
3321 struct lp_build_tgsi_context
* bld_base
,
3322 struct lp_build_emit_data
* emit_data
)
3324 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3326 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
3327 TRUE
, emit_data
->output
);
3332 const struct lp_build_tgsi_action
* action
,
3333 struct lp_build_tgsi_context
* bld_base
,
3334 struct lp_build_emit_data
* emit_data
)
3336 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3338 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_ZERO
,
3339 TRUE
, emit_data
->output
);
3344 const struct lp_build_tgsi_action
* action
,
3345 struct lp_build_tgsi_context
* bld_base
,
3346 struct lp_build_emit_data
* emit_data
)
3348 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3350 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
3351 FALSE
, emit_data
->output
);
3356 const struct lp_build_tgsi_action
* action
,
3357 struct lp_build_tgsi_context
* bld_base
,
3358 struct lp_build_emit_data
* emit_data
)
3360 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3362 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
3363 FALSE
, emit_data
->output
);
3368 const struct lp_build_tgsi_action
* action
,
3369 struct lp_build_tgsi_context
* bld_base
,
3370 struct lp_build_emit_data
* emit_data
)
3372 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3374 emit_size_query(bld
, emit_data
->inst
, emit_data
->output
, TRUE
);
3378 mask_vec(struct lp_build_tgsi_context
*bld_base
)
3380 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3381 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3382 struct lp_exec_mask
*exec_mask
= &bld
->exec_mask
;
3384 if (!exec_mask
->has_mask
) {
3385 return lp_build_mask_value(bld
->mask
);
3387 return LLVMBuildAnd(builder
, lp_build_mask_value(bld
->mask
),
3388 exec_mask
->exec_mask
, "");
3392 increment_vec_ptr_by_mask(struct lp_build_tgsi_context
* bld_base
,
3396 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3397 LLVMValueRef current_vec
= LLVMBuildLoad(builder
, ptr
, "");
3399 current_vec
= LLVMBuildSub(builder
, current_vec
, mask
, "");
3401 LLVMBuildStore(builder
, current_vec
, ptr
);
3405 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context
* bld_base
,
3409 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3410 LLVMValueRef current_vec
= LLVMBuildLoad(builder
, ptr
, "");
3412 current_vec
= lp_build_select(&bld_base
->uint_bld
,
3414 bld_base
->uint_bld
.zero
,
3417 LLVMBuildStore(builder
, current_vec
, ptr
);
3421 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context
* bld
,
3422 LLVMValueRef current_mask_vec
,
3423 LLVMValueRef total_emitted_vertices_vec
)
3425 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3426 struct lp_build_context
*int_bld
= &bld
->bld_base
.int_bld
;
3427 LLVMValueRef max_mask
= lp_build_cmp(int_bld
, PIPE_FUNC_LESS
,
3428 total_emitted_vertices_vec
,
3429 bld
->max_output_vertices_vec
);
3431 return LLVMBuildAnd(builder
, current_mask_vec
, max_mask
, "");
3436 const struct lp_build_tgsi_action
* action
,
3437 struct lp_build_tgsi_context
* bld_base
,
3438 struct lp_build_emit_data
* emit_data
)
3440 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3441 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3443 if (bld
->gs_iface
->emit_vertex
) {
3444 LLVMValueRef mask
= mask_vec(bld_base
);
3445 LLVMValueRef total_emitted_vertices_vec
=
3446 LLVMBuildLoad(builder
, bld
->total_emitted_vertices_vec_ptr
, "");
3447 mask
= clamp_mask_to_max_output_vertices(bld
, mask
,
3448 total_emitted_vertices_vec
);
3449 gather_outputs(bld
);
3450 bld
->gs_iface
->emit_vertex(bld
->gs_iface
, &bld
->bld_base
,
3452 total_emitted_vertices_vec
);
3453 increment_vec_ptr_by_mask(bld_base
, bld
->emitted_vertices_vec_ptr
,
3455 increment_vec_ptr_by_mask(bld_base
, bld
->total_emitted_vertices_vec_ptr
,
3458 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3459 " +++ emit vertex masked ones = ",
3461 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3462 " +++ emit vertex emitted = ",
3463 total_emitted_vertices_vec
);
3470 end_primitive_masked(struct lp_build_tgsi_context
* bld_base
,
3473 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3474 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3476 if (bld
->gs_iface
->end_primitive
) {
3477 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3478 LLVMValueRef emitted_vertices_vec
=
3479 LLVMBuildLoad(builder
, bld
->emitted_vertices_vec_ptr
, "");
3480 LLVMValueRef emitted_prims_vec
=
3481 LLVMBuildLoad(builder
, bld
->emitted_prims_vec_ptr
, "");
3483 LLVMValueRef emitted_mask
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3484 emitted_vertices_vec
,
3486 /* We need to combine the current execution mask with the mask
3487 telling us which, if any, execution slots actually have
3488 unemitted primitives, this way we make sure that end_primitives
3489 executes only on the paths that have unflushed vertices */
3490 mask
= LLVMBuildAnd(builder
, mask
, emitted_mask
, "");
3492 bld
->gs_iface
->end_primitive(bld
->gs_iface
, &bld
->bld_base
,
3493 emitted_vertices_vec
,
3497 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3498 " +++ end prim masked ones = ",
3500 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3501 " +++ end prim emitted verts1 = ",
3502 emitted_vertices_vec
);
3503 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3504 " +++ end prim emitted prims1 = ",
3505 LLVMBuildLoad(builder
,
3506 bld
->emitted_prims_vec_ptr
, ""));
3508 increment_vec_ptr_by_mask(bld_base
, bld
->emitted_prims_vec_ptr
,
3510 clear_uint_vec_ptr_from_mask(bld_base
, bld
->emitted_vertices_vec_ptr
,
3513 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3514 " +++ end prim emitted verts2 = ",
3515 LLVMBuildLoad(builder
,
3516 bld
->emitted_vertices_vec_ptr
, ""));
3524 const struct lp_build_tgsi_action
* action
,
3525 struct lp_build_tgsi_context
* bld_base
,
3526 struct lp_build_emit_data
* emit_data
)
3528 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3530 if (bld
->gs_iface
->end_primitive
) {
3531 LLVMValueRef mask
= mask_vec(bld_base
);
3532 end_primitive_masked(bld_base
, mask
);
3538 const struct lp_build_tgsi_action
* action
,
3539 struct lp_build_tgsi_context
* bld_base
,
3540 struct lp_build_emit_data
* emit_data
)
3542 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3544 lp_exec_mask_call(&bld
->exec_mask
, emit_data
->inst
->Label
.Label
,
3550 const struct lp_build_tgsi_action
* action
,
3551 struct lp_build_tgsi_context
* bld_base
,
3552 struct lp_build_emit_data
* emit_data
)
3554 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3556 lp_exec_mask_ret(&bld
->exec_mask
, &bld_base
->pc
);
3561 const struct lp_build_tgsi_action
* action
,
3562 struct lp_build_tgsi_context
* bld_base
,
3563 struct lp_build_emit_data
* emit_data
)
3565 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3567 lp_exec_break(&bld
->exec_mask
, bld_base
);
3572 const struct lp_build_tgsi_action
* action
,
3573 struct lp_build_tgsi_context
* bld_base
,
3574 struct lp_build_emit_data
* emit_data
)
3576 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3577 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3578 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3579 LLVMValueRef unsigned_cond
=
3580 LLVMBuildBitCast(builder
, emit_data
->args
[0], uint_bld
->vec_type
, "");
3581 LLVMValueRef cond
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3585 lp_exec_break_condition(&bld
->exec_mask
, cond
);
3590 const struct lp_build_tgsi_action
* action
,
3591 struct lp_build_tgsi_context
* bld_base
,
3592 struct lp_build_emit_data
* emit_data
)
3595 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3597 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_NOTEQUAL
,
3598 emit_data
->args
[0], bld
->bld_base
.base
.zero
);
3599 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
3604 const struct lp_build_tgsi_action
* action
,
3605 struct lp_build_tgsi_context
* bld_base
,
3606 struct lp_build_emit_data
* emit_data
)
3609 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3610 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3612 tmp
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3613 emit_data
->args
[0], uint_bld
->zero
);
3614 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
3619 const struct lp_build_tgsi_action
* action
,
3620 struct lp_build_tgsi_context
* bld_base
,
3621 struct lp_build_emit_data
* emit_data
)
3623 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3625 lp_exec_case(&bld
->exec_mask
, emit_data
->args
[0]);
3630 const struct lp_build_tgsi_action
* action
,
3631 struct lp_build_tgsi_context
* bld_base
,
3632 struct lp_build_emit_data
* emit_data
)
3634 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3636 lp_exec_default(&bld
->exec_mask
, bld_base
);
3641 const struct lp_build_tgsi_action
* action
,
3642 struct lp_build_tgsi_context
* bld_base
,
3643 struct lp_build_emit_data
* emit_data
)
3645 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3647 lp_exec_switch(&bld
->exec_mask
, emit_data
->args
[0]);
3652 const struct lp_build_tgsi_action
* action
,
3653 struct lp_build_tgsi_context
* bld_base
,
3654 struct lp_build_emit_data
* emit_data
)
3656 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3658 lp_exec_endswitch(&bld
->exec_mask
, bld_base
);
3663 const struct lp_build_tgsi_action
* action
,
3664 struct lp_build_tgsi_context
* bld_base
,
3665 struct lp_build_emit_data
* emit_data
)
3667 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3669 lp_exec_bgnloop(&bld
->exec_mask
);
3674 const struct lp_build_tgsi_action
* action
,
3675 struct lp_build_tgsi_context
* bld_base
,
3676 struct lp_build_emit_data
* emit_data
)
3678 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3680 lp_exec_mask_bgnsub(&bld
->exec_mask
);
3685 const struct lp_build_tgsi_action
* action
,
3686 struct lp_build_tgsi_context
* bld_base
,
3687 struct lp_build_emit_data
* emit_data
)
3689 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3691 lp_exec_mask_cond_invert(&bld
->exec_mask
);
3696 const struct lp_build_tgsi_action
* action
,
3697 struct lp_build_tgsi_context
* bld_base
,
3698 struct lp_build_emit_data
* emit_data
)
3700 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3702 lp_exec_mask_cond_pop(&bld
->exec_mask
);
3707 const struct lp_build_tgsi_action
* action
,
3708 struct lp_build_tgsi_context
* bld_base
,
3709 struct lp_build_emit_data
* emit_data
)
3711 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3713 lp_exec_endloop(bld_base
->base
.gallivm
, &bld
->exec_mask
);
3718 const struct lp_build_tgsi_action
* action
,
3719 struct lp_build_tgsi_context
* bld_base
,
3720 struct lp_build_emit_data
* emit_data
)
3722 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3724 lp_exec_mask_endsub(&bld
->exec_mask
, &bld_base
->pc
);
3729 const struct lp_build_tgsi_action
* action
,
3730 struct lp_build_tgsi_context
* bld_base
,
3731 struct lp_build_emit_data
* emit_data
)
3733 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3735 lp_exec_continue(&bld
->exec_mask
);
3738 static void emit_prologue(struct lp_build_tgsi_context
* bld_base
)
3740 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3741 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
3743 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
3744 LLVMValueRef array_size
=
3745 lp_build_const_int32(gallivm
,
3746 bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] * 4 + 4);
3747 bld
->temps_array
= lp_build_array_alloca(gallivm
,
3748 bld_base
->base
.vec_type
, array_size
,
3752 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
3753 LLVMValueRef array_size
=
3754 lp_build_const_int32(gallivm
,
3755 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] * 4 + 4);
3756 bld
->outputs_array
= lp_build_array_alloca(gallivm
,
3757 bld_base
->base
.vec_type
, array_size
,
3761 if (bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
)) {
3762 LLVMValueRef array_size
=
3763 lp_build_const_int32(gallivm
,
3764 bld_base
->info
->file_max
[TGSI_FILE_IMMEDIATE
] * 4 + 4);
3765 bld
->imms_array
= lp_build_array_alloca(gallivm
,
3766 bld_base
->base
.vec_type
, array_size
,
3770 /* If we have indirect addressing in inputs we need to copy them into
3771 * our alloca array to be able to iterate over them */
3772 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
) && !bld
->gs_iface
) {
3773 unsigned index
, chan
;
3774 LLVMTypeRef vec_type
= bld_base
->base
.vec_type
;
3775 LLVMValueRef array_size
= lp_build_const_int32(gallivm
,
3776 bld_base
->info
->file_max
[TGSI_FILE_INPUT
]*4 + 4);
3777 bld
->inputs_array
= lp_build_array_alloca(gallivm
,
3778 vec_type
, array_size
,
3781 assert(bld_base
->info
->num_inputs
3782 <= bld_base
->info
->file_max
[TGSI_FILE_INPUT
] + 1);
3784 for (index
= 0; index
< bld_base
->info
->num_inputs
; ++index
) {
3785 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3786 LLVMValueRef lindex
=
3787 lp_build_const_int32(gallivm
, index
* 4 + chan
);
3788 LLVMValueRef input_ptr
=
3789 LLVMBuildGEP(gallivm
->builder
, bld
->inputs_array
,
3791 LLVMValueRef value
= bld
->inputs
[index
][chan
];
3793 LLVMBuildStore(gallivm
->builder
, value
, input_ptr
);
3798 if (bld
->gs_iface
) {
3799 struct lp_build_context
*uint_bld
= &bld
->bld_base
.uint_bld
;
3800 bld
->emitted_prims_vec_ptr
=
3801 lp_build_alloca(gallivm
,
3803 "emitted_prims_ptr");
3804 bld
->emitted_vertices_vec_ptr
=
3805 lp_build_alloca(gallivm
,
3807 "emitted_vertices_ptr");
3808 bld
->total_emitted_vertices_vec_ptr
=
3809 lp_build_alloca(gallivm
,
3811 "total_emitted_vertices_ptr");
3813 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3814 bld
->emitted_prims_vec_ptr
);
3815 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3816 bld
->emitted_vertices_vec_ptr
);
3817 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3818 bld
->total_emitted_vertices_vec_ptr
);
3821 if (DEBUG_EXECUTION
) {
3822 lp_build_printf(gallivm
, "\n");
3823 emit_dump_file(bld
, TGSI_FILE_CONSTANT
);
3825 emit_dump_file(bld
, TGSI_FILE_INPUT
);
3829 static void emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
3831 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3832 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3834 if (DEBUG_EXECUTION
) {
3837 emit_dump_file(bld
, TGSI_FILE_TEMPORARY
);
3839 emit_dump_file(bld
, TGSI_FILE_OUTPUT
);
3840 lp_build_printf(bld_base
->base
.gallivm
, "\n");
3843 /* If we have indirect addressing in outputs we need to copy our alloca array
3844 * to the outputs slots specified by the caller */
3845 if (bld
->gs_iface
) {
3846 LLVMValueRef total_emitted_vertices_vec
;
3847 LLVMValueRef emitted_prims_vec
;
3848 /* implicit end_primitives, needed in case there are any unflushed
3849 vertices in the cache. Note must not call end_primitive here
3850 since the exec_mask is not valid at this point. */
3851 end_primitive_masked(bld_base
, lp_build_mask_value(bld
->mask
));
3853 total_emitted_vertices_vec
=
3854 LLVMBuildLoad(builder
, bld
->total_emitted_vertices_vec_ptr
, "");
3856 LLVMBuildLoad(builder
, bld
->emitted_prims_vec_ptr
, "");
3858 bld
->gs_iface
->gs_epilogue(bld
->gs_iface
,
3860 total_emitted_vertices_vec
,
3863 gather_outputs(bld
);
3868 lp_build_tgsi_soa(struct gallivm_state
*gallivm
,
3869 const struct tgsi_token
*tokens
,
3870 struct lp_type type
,
3871 struct lp_build_mask_context
*mask
,
3872 LLVMValueRef consts_ptr
,
3873 LLVMValueRef const_sizes_ptr
,
3874 const struct lp_bld_tgsi_system_values
*system_values
,
3875 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
3876 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
3877 LLVMValueRef context_ptr
,
3878 LLVMValueRef thread_data_ptr
,
3879 struct lp_build_sampler_soa
*sampler
,
3880 const struct tgsi_shader_info
*info
,
3881 const struct lp_build_tgsi_gs_iface
*gs_iface
)
3883 struct lp_build_tgsi_soa_context bld
;
3885 struct lp_type res_type
;
3887 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
3888 memset(&res_type
, 0, sizeof res_type
);
3889 res_type
.width
= type
.width
;
3890 res_type
.length
= type
.length
;
3893 /* Setup build context */
3894 memset(&bld
, 0, sizeof bld
);
3895 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
3896 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
3897 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
3898 lp_build_context_init(&bld
.elem_bld
, gallivm
, lp_elem_type(type
));
3900 struct lp_type dbl_type
;
3902 dbl_type
.width
*= 2;
3903 lp_build_context_init(&bld
.bld_base
.dbl_bld
, gallivm
, dbl_type
);
3906 bld
.inputs
= inputs
;
3907 bld
.outputs
= outputs
;
3908 bld
.consts_ptr
= consts_ptr
;
3909 bld
.const_sizes_ptr
= const_sizes_ptr
;
3910 bld
.sampler
= sampler
;
3911 bld
.bld_base
.info
= info
;
3912 bld
.indirect_files
= info
->indirect_files
;
3913 bld
.context_ptr
= context_ptr
;
3914 bld
.thread_data_ptr
= thread_data_ptr
;
3917 * If the number of temporaries is rather large then we just
3918 * allocate them as an array right from the start and treat
3919 * like indirect temporaries.
3921 if (info
->file_max
[TGSI_FILE_TEMPORARY
] >= LP_MAX_INLINED_TEMPS
) {
3922 bld
.indirect_files
|= (1 << TGSI_FILE_TEMPORARY
);
3925 * For performance reason immediates are always backed in a static
3926 * array, but if their number is too great, we have to use just
3927 * a dynamically allocated array.
3929 bld
.use_immediates_array
=
3930 (info
->file_max
[TGSI_FILE_IMMEDIATE
] >= LP_MAX_INLINED_IMMEDIATES
);
3931 if (bld
.use_immediates_array
) {
3932 bld
.indirect_files
|= (1 << TGSI_FILE_IMMEDIATE
);
3936 bld
.bld_base
.soa
= TRUE
;
3937 bld
.bld_base
.emit_debug
= emit_debug
;
3938 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
3939 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
3940 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
3941 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
3942 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = emit_fetch_system_value
;
3943 bld
.bld_base
.emit_store
= emit_store
;
3945 bld
.bld_base
.emit_declaration
= lp_emit_declaration_soa
;
3946 bld
.bld_base
.emit_immediate
= lp_emit_immediate_soa
;
3948 bld
.bld_base
.emit_prologue
= emit_prologue
;
3949 bld
.bld_base
.emit_epilogue
= emit_epilogue
;
3951 /* Set opcode actions */
3952 lp_set_default_actions_cpu(&bld
.bld_base
);
3954 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
3955 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNSUB
].emit
= bgnsub_emit
;
3956 bld
.bld_base
.op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
3957 bld
.bld_base
.op_actions
[TGSI_OPCODE_BREAKC
].emit
= breakc_emit
;
3958 bld
.bld_base
.op_actions
[TGSI_OPCODE_CAL
].emit
= cal_emit
;
3959 bld
.bld_base
.op_actions
[TGSI_OPCODE_CASE
].emit
= case_emit
;
3960 bld
.bld_base
.op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
3961 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDX
].emit
= ddx_emit
;
3962 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDY
].emit
= ddy_emit
;
3963 bld
.bld_base
.op_actions
[TGSI_OPCODE_DEFAULT
].emit
= default_emit
;
3964 bld
.bld_base
.op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
3965 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
3966 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
3967 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSUB
].emit
= endsub_emit
;
3968 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSWITCH
].emit
= endswitch_emit
;
3969 bld
.bld_base
.op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
3970 bld
.bld_base
.op_actions
[TGSI_OPCODE_UIF
].emit
= uif_emit
;
3971 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILL_IF
].emit
= kill_if_emit
;
3972 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILL
].emit
= kill_emit
;
3973 bld
.bld_base
.op_actions
[TGSI_OPCODE_RET
].emit
= ret_emit
;
3974 bld
.bld_base
.op_actions
[TGSI_OPCODE_SWITCH
].emit
= switch_emit
;
3975 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX
].emit
= tex_emit
;
3976 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB
].emit
= txb_emit
;
3977 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXD
].emit
= txd_emit
;
3978 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL
].emit
= txl_emit
;
3979 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXP
].emit
= txp_emit
;
3980 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXQ
].emit
= txq_emit
;
3981 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXF
].emit
= txf_emit
;
3982 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX2
].emit
= tex2_emit
;
3983 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB2
].emit
= txb2_emit
;
3984 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL2
].emit
= txl2_emit
;
3985 bld
.bld_base
.op_actions
[TGSI_OPCODE_TG4
].emit
= tg4_emit
;
3986 /* DX10 sampling ops */
3987 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE
].emit
= sample_emit
;
3988 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_B
].emit
= sample_b_emit
;
3989 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_C
].emit
= sample_c_emit
;
3990 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_C_LZ
].emit
= sample_c_lz_emit
;
3991 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_D
].emit
= sample_d_emit
;
3992 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_I
].emit
= sample_i_emit
;
3993 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_I_MS
].emit
= sample_i_emit
;
3994 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_L
].emit
= sample_l_emit
;
3995 bld
.bld_base
.op_actions
[TGSI_OPCODE_SVIEWINFO
].emit
= sviewinfo_emit
;
3998 /* There's no specific value for this because it should always
3999 * be set, but apps using ext_geometry_shader4 quite often
4000 * were forgetting so we're using MAX_VERTEX_VARYING from
4001 * that spec even though we could debug_assert if it's not
4002 * set, but that's a lot uglier. */
4003 uint max_output_vertices
;
4005 /* inputs are always indirect with gs */
4006 bld
.indirect_files
|= (1 << TGSI_FILE_INPUT
);
4007 bld
.gs_iface
= gs_iface
;
4008 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_gs_input
;
4009 bld
.bld_base
.op_actions
[TGSI_OPCODE_EMIT
].emit
= emit_vertex
;
4010 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= end_primitive
;
4012 max_output_vertices
=
4013 info
->properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
];
4014 if (!max_output_vertices
)
4015 max_output_vertices
= 32;
4017 bld
.max_output_vertices_vec
=
4018 lp_build_const_int_vec(gallivm
, bld
.bld_base
.int_bld
.type
,
4019 max_output_vertices
);
4022 lp_exec_mask_init(&bld
.exec_mask
, &bld
.bld_base
.int_bld
);
4024 bld
.system_values
= *system_values
;
4026 lp_build_tgsi_llvm(&bld
.bld_base
, tokens
);
4029 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
4030 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
4031 debug_printf("11111111111111111111111111111 \n");
4032 tgsi_dump(tokens
, 0);
4033 lp_debug_dump_value(function
);
4034 debug_printf("2222222222222222222222222222 \n");
4038 LLVMModuleRef module
= LLVMGetGlobalParent(
4039 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
4040 LLVMDumpModule(module
);
4043 lp_exec_mask_fini(&bld
.exec_mask
);