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 unsigned 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
!= -1 && 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_context
*bld
,
948 LLVMValueRef base_ptr
,
949 LLVMValueRef indexes
,
950 LLVMValueRef
*overflow_mask
)
952 LLVMBuilderRef builder
= bld
->gallivm
->builder
;
953 LLVMValueRef res
= bld
->undef
;
955 LLVMValueRef temp_ptr
;
958 temp_ptr
= lp_build_alloca(
960 lp_build_vec_type(bld
->gallivm
, bld
->type
), "");
964 * Loop over elements of index_vec, load scalar value, insert it into 'res'.
966 for (i
= 0; i
< bld
->type
.length
; i
++) {
967 LLVMValueRef ii
= lp_build_const_int32(bld
->gallivm
, i
);
968 LLVMValueRef index
= LLVMBuildExtractElement(builder
,
970 LLVMValueRef scalar_ptr
, scalar
;
971 LLVMValueRef overflow
;
972 struct lp_build_if_state if_ctx
;
975 * overflow_mask is a boolean vector telling us which channels
976 * in the vector overflowed. We use the overflow behavior for
977 * constant buffers which is defined as:
978 * Out of bounds access to constant buffer returns 0 in all
979 * componenets. Out of bounds behavior is always with respect
980 * to the size of the buffer bound at that slot.
983 overflow
= LLVMBuildExtractElement(builder
, *overflow_mask
,
985 lp_build_if(&if_ctx
, bld
->gallivm
, overflow
);
987 LLVMValueRef val
= LLVMBuildLoad(builder
, temp_ptr
, "");
988 val
= LLVMBuildInsertElement(
990 LLVMConstNull(LLVMFloatTypeInContext(bld
->gallivm
->context
)),
992 LLVMBuildStore(builder
, val
, temp_ptr
);
994 lp_build_else(&if_ctx
);
996 LLVMValueRef val
= LLVMBuildLoad(builder
, temp_ptr
, "");
998 scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
,
999 &index
, 1, "gather_ptr");
1000 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
1002 val
= LLVMBuildInsertElement(builder
, val
, scalar
, ii
, "");
1004 LLVMBuildStore(builder
, val
, temp_ptr
);
1006 lp_build_endif(&if_ctx
);
1008 scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
,
1009 &index
, 1, "gather_ptr");
1010 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
1012 res
= LLVMBuildInsertElement(builder
, res
, scalar
, ii
, "");
1016 if (overflow_mask
) {
1017 res
= LLVMBuildLoad(builder
, temp_ptr
, "gather_val");
1025 * Scatter/store vector.
1028 emit_mask_scatter(struct lp_build_tgsi_soa_context
*bld
,
1029 LLVMValueRef base_ptr
,
1030 LLVMValueRef indexes
,
1031 LLVMValueRef values
,
1032 struct lp_exec_mask
*mask
,
1035 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1036 LLVMBuilderRef builder
= gallivm
->builder
;
1039 /* Mix the predicate and execution mask */
1040 if (mask
->has_mask
) {
1042 pred
= LLVMBuildAnd(builder
, pred
, mask
->exec_mask
, "");
1045 pred
= mask
->exec_mask
;
1050 * Loop over elements of index_vec, store scalar value.
1052 for (i
= 0; i
< bld
->bld_base
.base
.type
.length
; i
++) {
1053 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1054 LLVMValueRef index
= LLVMBuildExtractElement(builder
, indexes
, ii
, "");
1055 LLVMValueRef scalar_ptr
= LLVMBuildGEP(builder
, base_ptr
, &index
, 1, "scatter_ptr");
1056 LLVMValueRef val
= LLVMBuildExtractElement(builder
, values
, ii
, "scatter_val");
1057 LLVMValueRef scalar_pred
= pred
?
1058 LLVMBuildExtractElement(builder
, pred
, ii
, "scatter_pred") : NULL
;
1061 lp_build_printf(gallivm
, "scatter %d: val %f at %d %p\n",
1062 ii
, val
, index
, scalar_ptr
);
1065 LLVMValueRef real_val
, dst_val
;
1066 dst_val
= LLVMBuildLoad(builder
, scalar_ptr
, "");
1067 real_val
= lp_build_select(&bld
->elem_bld
, scalar_pred
, val
, dst_val
);
1068 LLVMBuildStore(builder
, real_val
, scalar_ptr
);
1071 LLVMBuildStore(builder
, val
, scalar_ptr
);
1078 * Read the current value of the ADDR register, convert the floats to
1079 * ints, add the base index and return the vector of offsets.
1080 * The offsets will be used to index into the constant buffer or
1081 * temporary register file.
1084 get_indirect_index(struct lp_build_tgsi_soa_context
*bld
,
1085 unsigned reg_file
, unsigned reg_index
,
1086 const struct tgsi_ind_register
*indirect_reg
)
1088 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1089 struct lp_build_context
*uint_bld
= &bld
->bld_base
.uint_bld
;
1090 /* always use X component of address register */
1091 unsigned swizzle
= indirect_reg
->Swizzle
;
1094 LLVMValueRef max_index
;
1097 assert(bld
->indirect_files
& (1 << reg_file
));
1099 base
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
, uint_bld
->type
, reg_index
);
1101 assert(swizzle
< 4);
1102 switch (indirect_reg
->File
) {
1103 case TGSI_FILE_ADDRESS
:
1104 rel
= LLVMBuildLoad(builder
,
1105 bld
->addr
[indirect_reg
->Index
][swizzle
],
1107 /* ADDR LLVM values already have LLVM integer type. */
1109 case TGSI_FILE_TEMPORARY
:
1110 rel
= lp_get_temp_ptr_soa(bld
, indirect_reg
->Index
, swizzle
);
1111 rel
= LLVMBuildLoad(builder
, rel
, "load temp reg");
1112 /* TEMP LLVM values always have LLVM float type, but for indirection, the
1113 * value actually stored is expected to be an integer */
1114 rel
= LLVMBuildBitCast(builder
, rel
, uint_bld
->vec_type
, "");
1118 rel
= uint_bld
->zero
;
1121 index
= lp_build_add(uint_bld
, base
, rel
);
1124 * emit_fetch_constant handles constant buffer overflow so this code
1125 * is pointless for them.
1126 * Furthermore the D3D10 spec in section 6.5 says:
1127 * If the constant buffer bound to a slot is larger than the size
1128 * declared in the shader for that slot, implementations are allowed
1129 * to return incorrect data (not necessarily 0) for indices that are
1130 * larger than the declared size but smaller than the buffer size.
1132 if (reg_file
!= TGSI_FILE_CONSTANT
) {
1133 max_index
= lp_build_const_int_vec(bld
->bld_base
.base
.gallivm
,
1135 bld
->bld_base
.info
->file_max
[reg_file
]);
1137 assert(!uint_bld
->type
.sign
);
1138 index
= lp_build_min(uint_bld
, index
, max_index
);
1144 static struct lp_build_context
*
1145 stype_to_fetch(struct lp_build_tgsi_context
* bld_base
,
1146 enum tgsi_opcode_type stype
)
1148 struct lp_build_context
*bld_fetch
;
1151 case TGSI_TYPE_FLOAT
:
1152 case TGSI_TYPE_UNTYPED
:
1153 bld_fetch
= &bld_base
->base
;
1155 case TGSI_TYPE_UNSIGNED
:
1156 bld_fetch
= &bld_base
->uint_bld
;
1158 case TGSI_TYPE_SIGNED
:
1159 bld_fetch
= &bld_base
->int_bld
;
1161 case TGSI_TYPE_VOID
:
1162 case TGSI_TYPE_DOUBLE
:
1172 get_soa_array_offsets(struct lp_build_context
*uint_bld
,
1173 LLVMValueRef indirect_index
,
1174 unsigned chan_index
,
1175 boolean need_perelement_offset
)
1177 struct gallivm_state
*gallivm
= uint_bld
->gallivm
;
1178 LLVMValueRef chan_vec
=
1179 lp_build_const_int_vec(uint_bld
->gallivm
, uint_bld
->type
, chan_index
);
1180 LLVMValueRef length_vec
=
1181 lp_build_const_int_vec(gallivm
, uint_bld
->type
, uint_bld
->type
.length
);
1182 LLVMValueRef index_vec
;
1184 /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
1185 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1186 index_vec
= lp_build_add(uint_bld
, index_vec
, chan_vec
);
1187 index_vec
= lp_build_mul(uint_bld
, index_vec
, length_vec
);
1189 if (need_perelement_offset
) {
1190 LLVMValueRef pixel_offsets
;
1192 /* build pixel offset vector: {0, 1, 2, 3, ...} */
1193 pixel_offsets
= uint_bld
->undef
;
1194 for (i
= 0; i
< uint_bld
->type
.length
; i
++) {
1195 LLVMValueRef ii
= lp_build_const_int32(gallivm
, i
);
1196 pixel_offsets
= LLVMBuildInsertElement(gallivm
->builder
, pixel_offsets
,
1199 index_vec
= lp_build_add(uint_bld
, index_vec
, pixel_offsets
);
1205 emit_fetch_constant(
1206 struct lp_build_tgsi_context
* bld_base
,
1207 const struct tgsi_full_src_register
* reg
,
1208 enum tgsi_opcode_type stype
,
1211 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1212 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1213 LLVMBuilderRef builder
= gallivm
->builder
;
1214 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
1215 unsigned dimension
= 0;
1216 LLVMValueRef consts_ptr
;
1217 LLVMValueRef num_consts
;
1220 /* XXX: Handle fetching xyzw components as a vector */
1221 assert(swizzle
!= ~0);
1223 if (reg
->Register
.Dimension
) {
1224 assert(!reg
->Dimension
.Indirect
);
1225 dimension
= reg
->Dimension
.Index
;
1226 assert(dimension
< LP_MAX_TGSI_CONST_BUFFERS
);
1229 consts_ptr
= bld
->consts
[dimension
];
1230 num_consts
= bld
->consts_sizes
[dimension
];
1232 if (reg
->Register
.Indirect
) {
1233 LLVMValueRef indirect_index
;
1234 LLVMValueRef swizzle_vec
=
1235 lp_build_const_int_vec(gallivm
, uint_bld
->type
, swizzle
);
1236 LLVMValueRef index_vec
; /* index into the const buffer */
1237 LLVMValueRef overflow_mask
;
1239 indirect_index
= get_indirect_index(bld
,
1241 reg
->Register
.Index
,
1244 /* All fetches are from the same constant buffer, so
1245 * we need to propagate the size to a vector to do a
1246 * vector comparison */
1247 num_consts
= lp_build_broadcast_scalar(uint_bld
, num_consts
);
1248 /* Construct a boolean vector telling us which channels
1249 * overflow the bound constant buffer */
1250 overflow_mask
= LLVMBuildICmp(builder
, LLVMIntUGE
,
1254 /* index_vec = indirect_index * 4 + swizzle */
1255 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1256 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
1258 /* Gather values from the constant buffer */
1259 res
= build_gather(&bld_base
->base
, consts_ptr
, index_vec
,
1263 LLVMValueRef index
; /* index into the const buffer */
1264 LLVMValueRef scalar
, scalar_ptr
;
1266 index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
* 4 + swizzle
);
1268 scalar_ptr
= LLVMBuildGEP(builder
, consts_ptr
,
1270 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
1271 res
= lp_build_broadcast_scalar(&bld_base
->base
, scalar
);
1274 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
) {
1275 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1276 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1283 emit_fetch_immediate(
1284 struct lp_build_tgsi_context
* bld_base
,
1285 const struct tgsi_full_src_register
* reg
,
1286 enum tgsi_opcode_type stype
,
1289 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1290 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1291 LLVMBuilderRef builder
= gallivm
->builder
;
1292 LLVMValueRef res
= NULL
;
1294 if (bld
->use_immediates_array
|| reg
->Register
.Indirect
) {
1295 LLVMValueRef imms_array
;
1296 LLVMTypeRef fptr_type
;
1298 /* cast imms_array pointer to float* */
1299 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1300 imms_array
= LLVMBuildBitCast(builder
, bld
->imms_array
, fptr_type
, "");
1302 if (reg
->Register
.Indirect
) {
1303 LLVMValueRef indirect_index
;
1304 LLVMValueRef index_vec
; /* index into the immediate register array */
1306 indirect_index
= get_indirect_index(bld
,
1308 reg
->Register
.Index
,
1311 * Unlike for other reg classes, adding pixel offsets is unnecessary -
1312 * immediates are stored as full vectors (FIXME??? - might be better
1313 * to store them the same as constants) but all elements are the same
1316 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1321 /* Gather values from the immediate register array */
1322 res
= build_gather(&bld_base
->base
, imms_array
, index_vec
, NULL
);
1324 LLVMValueRef lindex
= lp_build_const_int32(gallivm
,
1325 reg
->Register
.Index
* 4 + swizzle
);
1326 LLVMValueRef imms_ptr
= LLVMBuildGEP(builder
,
1327 bld
->imms_array
, &lindex
, 1, "");
1328 res
= LLVMBuildLoad(builder
, imms_ptr
, "");
1332 res
= bld
->immediates
[reg
->Register
.Index
][swizzle
];
1335 if (stype
== TGSI_TYPE_UNSIGNED
) {
1336 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1337 } else if (stype
== TGSI_TYPE_SIGNED
) {
1338 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1345 struct lp_build_tgsi_context
* bld_base
,
1346 const struct tgsi_full_src_register
* reg
,
1347 enum tgsi_opcode_type stype
,
1350 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1351 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1352 LLVMBuilderRef builder
= gallivm
->builder
;
1355 if (reg
->Register
.Indirect
) {
1356 LLVMValueRef indirect_index
;
1357 LLVMValueRef index_vec
; /* index into the input reg array */
1358 LLVMValueRef inputs_array
;
1359 LLVMTypeRef fptr_type
;
1361 indirect_index
= get_indirect_index(bld
,
1363 reg
->Register
.Index
,
1366 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1371 /* cast inputs_array pointer to float* */
1372 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1373 inputs_array
= LLVMBuildBitCast(builder
, bld
->inputs_array
, fptr_type
, "");
1375 /* Gather values from the input register array */
1376 res
= build_gather(&bld_base
->base
, inputs_array
, index_vec
, NULL
);
1378 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
1379 LLVMValueRef lindex
= lp_build_const_int32(gallivm
,
1380 reg
->Register
.Index
* 4 + swizzle
);
1381 LLVMValueRef input_ptr
= LLVMBuildGEP(builder
,
1382 bld
->inputs_array
, &lindex
, 1, "");
1383 res
= LLVMBuildLoad(builder
, input_ptr
, "");
1386 res
= bld
->inputs
[reg
->Register
.Index
][swizzle
];
1392 if (stype
== TGSI_TYPE_UNSIGNED
) {
1393 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1394 } else if (stype
== TGSI_TYPE_SIGNED
) {
1395 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1403 emit_fetch_gs_input(
1404 struct lp_build_tgsi_context
* bld_base
,
1405 const struct tgsi_full_src_register
* reg
,
1406 enum tgsi_opcode_type stype
,
1409 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1410 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1411 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
1412 LLVMBuilderRef builder
= gallivm
->builder
;
1413 LLVMValueRef attrib_index
= NULL
;
1414 LLVMValueRef vertex_index
= NULL
;
1415 LLVMValueRef swizzle_index
= lp_build_const_int32(gallivm
, swizzle
);
1418 if (info
->input_semantic_name
[reg
->Register
.Index
] == TGSI_SEMANTIC_PRIMID
) {
1419 /* This is really a system value not a regular input */
1420 assert(!reg
->Register
.Indirect
);
1421 assert(!reg
->Dimension
.Indirect
);
1422 res
= bld
->system_values
.prim_id
;
1423 if (stype
!= TGSI_TYPE_UNSIGNED
&& stype
!= TGSI_TYPE_SIGNED
) {
1424 res
= LLVMBuildBitCast(builder
, res
, bld_base
->base
.vec_type
, "");
1429 if (reg
->Register
.Indirect
) {
1430 attrib_index
= get_indirect_index(bld
,
1432 reg
->Register
.Index
,
1435 attrib_index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
);
1438 if (reg
->Dimension
.Indirect
) {
1439 vertex_index
= get_indirect_index(bld
,
1441 reg
->Dimension
.Index
,
1444 vertex_index
= lp_build_const_int32(gallivm
, reg
->Dimension
.Index
);
1447 res
= bld
->gs_iface
->fetch_input(bld
->gs_iface
, bld_base
,
1448 reg
->Dimension
.Indirect
,
1450 reg
->Register
.Indirect
,
1456 if (stype
== TGSI_TYPE_UNSIGNED
) {
1457 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1458 } else if (stype
== TGSI_TYPE_SIGNED
) {
1459 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1466 emit_fetch_temporary(
1467 struct lp_build_tgsi_context
* bld_base
,
1468 const struct tgsi_full_src_register
* reg
,
1469 enum tgsi_opcode_type stype
,
1472 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1473 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1474 LLVMBuilderRef builder
= gallivm
->builder
;
1477 if (reg
->Register
.Indirect
) {
1478 LLVMValueRef indirect_index
;
1479 LLVMValueRef index_vec
; /* index into the temp reg array */
1480 LLVMValueRef temps_array
;
1481 LLVMTypeRef fptr_type
;
1483 indirect_index
= get_indirect_index(bld
,
1485 reg
->Register
.Index
,
1488 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1493 /* cast temps_array pointer to float* */
1494 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1495 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
, fptr_type
, "");
1497 /* Gather values from the temporary register array */
1498 res
= build_gather(&bld_base
->base
, temps_array
, index_vec
, NULL
);
1501 LLVMValueRef temp_ptr
;
1502 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
1503 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
1506 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
) {
1507 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1508 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1515 emit_fetch_system_value(
1516 struct lp_build_tgsi_context
* bld_base
,
1517 const struct tgsi_full_src_register
* reg
,
1518 enum tgsi_opcode_type stype
,
1521 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1522 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1523 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
1524 LLVMBuilderRef builder
= gallivm
->builder
;
1526 enum tgsi_opcode_type atype
; // Actual type of the value
1528 assert(!reg
->Register
.Indirect
);
1530 switch (info
->system_value_semantic_name
[reg
->Register
.Index
]) {
1531 case TGSI_SEMANTIC_INSTANCEID
:
1532 res
= lp_build_broadcast_scalar(&bld_base
->uint_bld
, bld
->system_values
.instance_id
);
1533 atype
= TGSI_TYPE_UNSIGNED
;
1536 case TGSI_SEMANTIC_VERTEXID
:
1537 res
= bld
->system_values
.vertex_id
;
1538 atype
= TGSI_TYPE_UNSIGNED
;
1541 case TGSI_SEMANTIC_PRIMID
:
1542 res
= bld
->system_values
.prim_id
;
1543 atype
= TGSI_TYPE_UNSIGNED
;
1547 assert(!"unexpected semantic in emit_fetch_system_value");
1548 res
= bld_base
->base
.zero
;
1549 atype
= TGSI_TYPE_FLOAT
;
1553 if (atype
!= stype
) {
1554 if (stype
== TGSI_TYPE_FLOAT
) {
1555 res
= LLVMBuildBitCast(builder
, res
, bld_base
->base
.vec_type
, "");
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
, "");
1567 * Register fetch with derivatives.
1571 struct lp_build_tgsi_soa_context
*bld
,
1580 /* TODO: use interpolation coeffs for inputs */
1583 *ddx
= lp_build_ddx(&bld
->bld_base
.base
, src
);
1586 *ddy
= lp_build_ddy(&bld
->bld_base
.base
, src
);
1594 emit_fetch_predicate(
1595 struct lp_build_tgsi_soa_context
*bld
,
1596 const struct tgsi_full_instruction
*inst
,
1599 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1601 unsigned char swizzles
[4];
1602 LLVMValueRef unswizzled
[4] = {NULL
, NULL
, NULL
, NULL
};
1606 if (!inst
->Instruction
.Predicate
) {
1607 TGSI_FOR_EACH_CHANNEL( chan
) {
1613 swizzles
[0] = inst
->Predicate
.SwizzleX
;
1614 swizzles
[1] = inst
->Predicate
.SwizzleY
;
1615 swizzles
[2] = inst
->Predicate
.SwizzleZ
;
1616 swizzles
[3] = inst
->Predicate
.SwizzleW
;
1618 index
= inst
->Predicate
.Index
;
1619 assert(index
< LP_MAX_TGSI_PREDS
);
1621 TGSI_FOR_EACH_CHANNEL( chan
) {
1622 unsigned swizzle
= swizzles
[chan
];
1625 * Only fetch the predicate register channels that are actually listed
1628 if (!unswizzled
[swizzle
]) {
1629 value
= LLVMBuildLoad(builder
,
1630 bld
->preds
[index
][swizzle
], "");
1633 * Convert the value to an integer mask.
1635 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
1636 * is needlessly causing two comparisons due to storing the intermediate
1637 * result as float vector instead of an integer mask vector.
1639 value
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
1640 bld
->bld_base
.base
.type
,
1643 bld
->bld_base
.base
.zero
);
1644 if (inst
->Predicate
.Negate
) {
1645 value
= LLVMBuildNot(builder
, value
, "");
1648 unswizzled
[swizzle
] = value
;
1650 value
= unswizzled
[swizzle
];
1663 struct lp_build_tgsi_context
*bld_base
,
1664 const struct tgsi_full_instruction
*inst
,
1666 unsigned chan_index
,
1670 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1671 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1672 LLVMBuilderRef builder
= gallivm
->builder
;
1673 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
1674 struct lp_build_context
*float_bld
= &bld_base
->base
;
1675 struct lp_build_context
*int_bld
= &bld_base
->int_bld
;
1676 LLVMValueRef indirect_index
= NULL
;
1677 enum tgsi_opcode_type dtype
= tgsi_opcode_infer_dst_type(inst
->Instruction
.Opcode
);
1682 * It is always assumed to be float.
1684 switch( inst
->Instruction
.Saturate
) {
1688 case TGSI_SAT_ZERO_ONE
:
1689 assert(dtype
== TGSI_TYPE_FLOAT
||
1690 dtype
== TGSI_TYPE_UNTYPED
);
1691 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1692 value
= lp_build_clamp_zero_one_nanzero(float_bld
, value
);
1695 case TGSI_SAT_MINUS_PLUS_ONE
:
1696 assert(dtype
== TGSI_TYPE_FLOAT
||
1697 dtype
== TGSI_TYPE_UNTYPED
);
1698 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1699 /* This will give -1.0 for NaN which is probably not what we want. */
1700 value
= lp_build_max_ext(float_bld
, value
,
1701 lp_build_const_vec(gallivm
, float_bld
->type
, -1.0),
1702 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
1703 value
= lp_build_min(float_bld
, value
, float_bld
->one
);
1710 if (reg
->Register
.Indirect
) {
1711 indirect_index
= get_indirect_index(bld
,
1713 reg
->Register
.Index
,
1716 assert(reg
->Register
.Index
<=
1717 bld_base
->info
->file_max
[reg
->Register
.File
]);
1720 if (DEBUG_EXECUTION
) {
1721 emit_dump_reg(gallivm
, reg
->Register
.File
, reg
->Register
.Index
, chan_index
, value
);
1724 switch( reg
->Register
.File
) {
1725 case TGSI_FILE_OUTPUT
:
1726 /* Outputs are always stored as floats */
1727 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1729 if (reg
->Register
.Indirect
) {
1730 LLVMValueRef index_vec
; /* indexes into the output registers */
1731 LLVMValueRef outputs_array
;
1732 LLVMTypeRef fptr_type
;
1734 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1739 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1740 outputs_array
= LLVMBuildBitCast(builder
, bld
->outputs_array
, fptr_type
, "");
1742 /* Scatter store values into output registers */
1743 emit_mask_scatter(bld
, outputs_array
, index_vec
, value
,
1744 &bld
->exec_mask
, pred
);
1747 LLVMValueRef out_ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
,
1749 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
, out_ptr
);
1753 case TGSI_FILE_TEMPORARY
:
1754 /* Temporaries are always stored as floats */
1755 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1757 if (reg
->Register
.Indirect
) {
1758 LLVMValueRef index_vec
; /* indexes into the temp registers */
1759 LLVMValueRef temps_array
;
1760 LLVMTypeRef fptr_type
;
1762 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1767 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1768 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
, fptr_type
, "");
1770 /* Scatter store values into temp registers */
1771 emit_mask_scatter(bld
, temps_array
, index_vec
, value
,
1772 &bld
->exec_mask
, pred
);
1775 LLVMValueRef temp_ptr
;
1776 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, chan_index
);
1777 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
, temp_ptr
);
1781 case TGSI_FILE_ADDRESS
:
1782 assert(dtype
== TGSI_TYPE_SIGNED
);
1783 assert(LLVMTypeOf(value
) == int_bld
->vec_type
);
1784 value
= LLVMBuildBitCast(builder
, value
, int_bld
->vec_type
, "");
1785 lp_exec_mask_store(&bld
->exec_mask
, int_bld
, pred
, value
,
1786 bld
->addr
[reg
->Register
.Index
][chan_index
]);
1789 case TGSI_FILE_PREDICATE
:
1790 assert(LLVMTypeOf(value
) == float_bld
->vec_type
);
1791 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1792 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
,
1793 bld
->preds
[reg
->Register
.Index
][chan_index
]);
1804 * Called at the beginning of the translation of each TGSI instruction, to
1805 * emit some debug code.
1809 struct lp_build_tgsi_context
* bld_base
,
1810 const struct tgsi_full_instruction
* inst
,
1811 const struct tgsi_opcode_info
* info
)
1814 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1816 if (DEBUG_EXECUTION
) {
1818 * Dump the TGSI instruction.
1821 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1825 tgsi_dump_instruction_str(inst
, bld_base
->pc
, &buf
[2], sizeof buf
- 2);
1826 lp_build_printf(gallivm
, buf
);
1828 /* Dump the execution mask.
1830 if (bld
->exec_mask
.has_mask
) {
1831 lp_build_print_value(gallivm
, " mask = ", bld
->exec_mask
.exec_mask
);
1838 struct lp_build_tgsi_context
* bld_base
,
1839 const struct tgsi_full_instruction
* inst
,
1840 const struct tgsi_opcode_info
* info
,
1841 LLVMValueRef dst
[4])
1844 unsigned chan_index
;
1845 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1848 LLVMValueRef pred
[TGSI_NUM_CHANNELS
];
1850 emit_fetch_predicate( bld
, inst
, pred
);
1852 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1853 emit_store_chan(bld_base
, inst
, 0, chan_index
, pred
[chan_index
], dst
[chan_index
]);
1859 tgsi_to_pipe_tex_target(unsigned tgsi_target
)
1861 switch (tgsi_target
) {
1862 case TGSI_TEXTURE_BUFFER
:
1864 case TGSI_TEXTURE_1D
:
1865 case TGSI_TEXTURE_SHADOW1D
:
1866 return PIPE_TEXTURE_1D
;
1867 case TGSI_TEXTURE_2D
:
1868 case TGSI_TEXTURE_SHADOW2D
:
1869 case TGSI_TEXTURE_2D_MSAA
:
1870 return PIPE_TEXTURE_2D
;
1871 case TGSI_TEXTURE_3D
:
1872 return PIPE_TEXTURE_3D
;
1873 case TGSI_TEXTURE_CUBE
:
1874 case TGSI_TEXTURE_SHADOWCUBE
:
1875 return PIPE_TEXTURE_CUBE
;
1876 case TGSI_TEXTURE_RECT
:
1877 case TGSI_TEXTURE_SHADOWRECT
:
1878 return PIPE_TEXTURE_RECT
;
1879 case TGSI_TEXTURE_1D_ARRAY
:
1880 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1881 return PIPE_TEXTURE_1D_ARRAY
;
1882 case TGSI_TEXTURE_2D_ARRAY
:
1883 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1884 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
1885 return PIPE_TEXTURE_2D_ARRAY
;
1886 case TGSI_TEXTURE_CUBE_ARRAY
:
1887 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
1888 return PIPE_TEXTURE_CUBE_ARRAY
;
1896 static enum lp_sampler_lod_property
1897 lp_build_lod_property(
1898 struct lp_build_tgsi_context
*bld_base
,
1899 const struct tgsi_full_instruction
*inst
,
1902 const struct tgsi_full_src_register
*reg
= &inst
->Src
[src_op
];
1903 enum lp_sampler_lod_property lod_property
;
1906 * Not much we can do here. We could try catching inputs declared
1907 * with constant interpolation but not sure it's worth it - since for
1908 * TEX opcodes as well as FETCH/LD the lod comes from same reg as
1909 * the coords, so it could only work for SAMPLE/TXQ/SVIEWINFO), just
1910 * like the constant/immediate recognition below.
1911 * What seems to be of more value would be to recognize temps holding
1912 * broadcasted scalars but no way we can do it.
1913 * Tried asking llvm but without any success (using LLVMIsConstant
1914 * even though this isn't exactly what we'd need), even as simple as
1915 * IMM[0] UINT32 (0,-1,0,0)
1916 * MOV TEMP[0] IMM[0].yyyy
1917 * SVIEWINFO TEMP[1], TEMP[0].xxxx, SVIEWINFO[0]
1919 * This means there's ZERO chance this will ever catch a scalar lod
1920 * with traditional tex opcodes as well as texel fetches, since the lod
1921 * comes from the same reg as coords (except some test shaders using
1922 * constant coords maybe).
1923 * There's at least hope for sample opcodes as well as size queries.
1925 if (reg
->Register
.File
== TGSI_FILE_CONSTANT
||
1926 reg
->Register
.File
== TGSI_FILE_IMMEDIATE
) {
1927 lod_property
= LP_SAMPLER_LOD_SCALAR
;
1929 else if (bld_base
->info
->processor
== TGSI_PROCESSOR_FRAGMENT
) {
1930 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
1931 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
1934 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
1938 /* never use scalar (per-quad) lod the results are just too wrong. */
1939 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
1941 return lod_property
;
1946 * High-level instruction translators.
1950 emit_tex( struct lp_build_tgsi_soa_context
*bld
,
1951 const struct tgsi_full_instruction
*inst
,
1952 enum lp_build_tex_modifier modifier
,
1953 LLVMValueRef
*texel
,
1954 unsigned sampler_reg
)
1956 unsigned unit
= inst
->Src
[sampler_reg
].Register
.Index
;
1957 LLVMValueRef lod_bias
, explicit_lod
;
1958 LLVMValueRef oow
= NULL
;
1959 LLVMValueRef coords
[5];
1960 LLVMValueRef offsets
[3] = { NULL
};
1961 struct lp_derivatives derivs
;
1962 struct lp_derivatives
*deriv_ptr
= NULL
;
1963 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
1964 unsigned num_derivs
, num_offsets
, i
;
1965 unsigned shadow_coord
= 0;
1966 unsigned layer_coord
= 0;
1968 if (!bld
->sampler
) {
1969 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1970 for (i
= 0; i
< 4; i
++) {
1971 texel
[i
] = bld
->bld_base
.base
.undef
;
1976 switch (inst
->Texture
.Texture
) {
1977 case TGSI_TEXTURE_1D_ARRAY
:
1980 case TGSI_TEXTURE_1D
:
1984 case TGSI_TEXTURE_2D_ARRAY
:
1987 case TGSI_TEXTURE_2D
:
1988 case TGSI_TEXTURE_RECT
:
1992 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1995 case TGSI_TEXTURE_SHADOW1D
:
2000 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
2006 case TGSI_TEXTURE_SHADOW2D
:
2007 case TGSI_TEXTURE_SHADOWRECT
:
2012 case TGSI_TEXTURE_CUBE
:
2016 case TGSI_TEXTURE_3D
:
2020 case TGSI_TEXTURE_SHADOWCUBE
:
2025 case TGSI_TEXTURE_CUBE_ARRAY
:
2030 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
2034 shadow_coord
= 4; /* shadow coord special different reg */
2036 case TGSI_TEXTURE_2D_MSAA
:
2037 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
2043 /* Note lod and especially projected are illegal in a LOT of cases */
2044 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
||
2045 modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2047 if (inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE
||
2048 inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
) {
2049 /* note that shadow cube array with bias/explicit lod does not exist */
2050 lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, 0);
2053 lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2055 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
2057 explicit_lod
= NULL
;
2059 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2063 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2067 explicit_lod
= NULL
;
2070 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
) {
2071 oow
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2072 oow
= lp_build_rcp(&bld
->bld_base
.base
, oow
);
2075 for (i
= 0; i
< num_derivs
; i
++) {
2076 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2077 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2078 coords
[i
] = lp_build_mul(&bld
->bld_base
.base
, coords
[i
], oow
);
2080 for (i
= num_derivs
; i
< 5; i
++) {
2081 coords
[i
] = bld
->bld_base
.base
.undef
;
2084 /* Layer coord always goes into 3rd slot, except for cube map arrays */
2086 if (layer_coord
== 3) {
2087 coords
[3] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2090 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2092 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2093 coords
[2] = lp_build_mul(&bld
->bld_base
.base
, coords
[2], oow
);
2095 /* Shadow coord occupies always 5th slot. */
2097 if (shadow_coord
== 4) {
2098 coords
[4] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, 0);
2101 coords
[4] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, shadow_coord
);
2103 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2104 coords
[4] = lp_build_mul(&bld
->bld_base
.base
, coords
[4], oow
);
2107 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
2109 for (dim
= 0; dim
< num_derivs
; ++dim
) {
2110 derivs
.ddx
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, dim
);
2111 derivs
.ddy
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, dim
);
2113 deriv_ptr
= &derivs
;
2115 * could also check all src regs if constant but I doubt such
2116 * cases exist in practice.
2118 if (bld
->bld_base
.info
->processor
== TGSI_PROCESSOR_FRAGMENT
) {
2119 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
2120 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2123 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
2127 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2131 /* some advanced gather instructions (txgo) would require 4 offsets */
2132 if (inst
->Texture
.NumOffsets
== 1) {
2134 for (dim
= 0; dim
< num_offsets
; dim
++) {
2135 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2139 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
2140 bld
->bld_base
.base
.gallivm
,
2141 bld
->bld_base
.base
.type
,
2147 lod_bias
, explicit_lod
, lod_property
,
2152 emit_sample(struct lp_build_tgsi_soa_context
*bld
,
2153 const struct tgsi_full_instruction
*inst
,
2154 enum lp_build_tex_modifier modifier
,
2156 LLVMValueRef
*texel
)
2158 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2159 unsigned texture_unit
, sampler_unit
;
2160 LLVMValueRef lod_bias
, explicit_lod
;
2161 LLVMValueRef coords
[5];
2162 LLVMValueRef offsets
[3] = { NULL
};
2163 struct lp_derivatives derivs
;
2164 struct lp_derivatives
*deriv_ptr
= NULL
;
2165 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
2167 unsigned num_offsets
, num_derivs
, i
;
2168 unsigned layer_coord
= 0;
2170 if (!bld
->sampler
) {
2171 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2172 for (i
= 0; i
< 4; i
++) {
2173 texel
[i
] = bld
->bld_base
.base
.undef
;
2179 * unlike old-style tex opcodes the texture/sampler indices
2180 * always come from src1 and src2 respectively.
2182 texture_unit
= inst
->Src
[1].Register
.Index
;
2183 sampler_unit
= inst
->Src
[2].Register
.Index
;
2186 * Note inst->Texture.Texture will contain the number of offsets,
2187 * however the target information is NOT there and comes from the
2188 * declared sampler views instead.
2190 switch (bld
->sv
[texture_unit
].Resource
) {
2191 case TGSI_TEXTURE_1D
:
2195 case TGSI_TEXTURE_1D_ARRAY
:
2200 case TGSI_TEXTURE_2D
:
2201 case TGSI_TEXTURE_RECT
:
2205 case TGSI_TEXTURE_2D_ARRAY
:
2210 case TGSI_TEXTURE_CUBE
:
2214 case TGSI_TEXTURE_3D
:
2218 case TGSI_TEXTURE_CUBE_ARRAY
:
2228 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
||
2229 modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2230 LLVMValueRef lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, 0);
2231 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
2233 explicit_lod
= NULL
;
2235 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2239 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2241 else if (modifier
== LP_BLD_TEX_MODIFIER_LOD_ZERO
) {
2243 /* XXX might be better to explicitly pass the level zero information */
2244 explicit_lod
= lp_build_const_vec(gallivm
, bld
->bld_base
.base
.type
, 0.0F
);
2248 explicit_lod
= NULL
;
2251 for (i
= 0; i
< num_derivs
; i
++) {
2252 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2254 for (i
= num_derivs
; i
< 5; i
++) {
2255 coords
[i
] = bld
->bld_base
.base
.undef
;
2258 /* Layer coord always goes into 3rd slot, except for cube map arrays */
2260 if (layer_coord
== 3)
2261 coords
[3] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2263 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2265 /* Shadow coord occupies always 5th slot. */
2267 coords
[4] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, 0);
2270 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
2272 for (dim
= 0; dim
< num_derivs
; ++dim
) {
2273 derivs
.ddx
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, dim
);
2274 derivs
.ddy
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 4, dim
);
2276 deriv_ptr
= &derivs
;
2278 * could also check all src regs if constant but I doubt such
2279 * cases exist in practice.
2281 if (bld
->bld_base
.info
->processor
== TGSI_PROCESSOR_FRAGMENT
) {
2282 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
2283 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2286 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
2290 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2294 /* some advanced gather instructions (txgo) would require 4 offsets */
2295 if (inst
->Texture
.NumOffsets
== 1) {
2297 for (dim
= 0; dim
< num_offsets
; dim
++) {
2298 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2302 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
2303 bld
->bld_base
.base
.gallivm
,
2304 bld
->bld_base
.base
.type
,
2306 texture_unit
, sampler_unit
,
2310 lod_bias
, explicit_lod
, lod_property
,
2313 if (inst
->Src
[1].Register
.SwizzleX
!= PIPE_SWIZZLE_RED
||
2314 inst
->Src
[1].Register
.SwizzleY
!= PIPE_SWIZZLE_GREEN
||
2315 inst
->Src
[1].Register
.SwizzleZ
!= PIPE_SWIZZLE_BLUE
||
2316 inst
->Src
[1].Register
.SwizzleW
!= PIPE_SWIZZLE_ALPHA
) {
2317 unsigned char swizzles
[4];
2318 swizzles
[0] = inst
->Src
[1].Register
.SwizzleX
;
2319 swizzles
[1] = inst
->Src
[1].Register
.SwizzleY
;
2320 swizzles
[2] = inst
->Src
[1].Register
.SwizzleZ
;
2321 swizzles
[3] = inst
->Src
[1].Register
.SwizzleW
;
2323 lp_build_swizzle_soa_inplace(&bld
->bld_base
.base
, texel
, swizzles
);
2328 emit_fetch_texels( struct lp_build_tgsi_soa_context
*bld
,
2329 const struct tgsi_full_instruction
*inst
,
2330 LLVMValueRef
*texel
,
2333 unsigned unit
, target
;
2334 LLVMValueRef coord_undef
= LLVMGetUndef(bld
->bld_base
.base
.int_vec_type
);
2335 LLVMValueRef explicit_lod
= NULL
;
2336 LLVMValueRef coords
[5];
2337 LLVMValueRef offsets
[3] = { NULL
};
2338 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
2340 unsigned layer_coord
= 0;
2342 if (!bld
->sampler
) {
2343 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2344 for (i
= 0; i
< 4; i
++) {
2345 texel
[i
] = coord_undef
;
2350 unit
= inst
->Src
[1].Register
.Index
;
2353 target
= bld
->sv
[unit
].Resource
;
2356 target
= inst
->Texture
.Texture
;
2360 case TGSI_TEXTURE_1D
:
2361 case TGSI_TEXTURE_BUFFER
:
2364 case TGSI_TEXTURE_1D_ARRAY
:
2368 case TGSI_TEXTURE_2D
:
2369 case TGSI_TEXTURE_RECT
:
2370 case TGSI_TEXTURE_2D_MSAA
:
2373 case TGSI_TEXTURE_2D_ARRAY
:
2374 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
2378 case TGSI_TEXTURE_3D
:
2386 /* always have lod except for buffers and msaa targets ? */
2387 if (target
!= TGSI_TEXTURE_BUFFER
&&
2388 target
!= TGSI_TEXTURE_2D_MSAA
&&
2389 target
!= TGSI_TEXTURE_2D_ARRAY_MSAA
) {
2390 explicit_lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2391 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2393 /* XXX: for real msaa support, the w component would be the sample index. */
2395 for (i
= 0; i
< dims
; i
++) {
2396 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2398 /* never use more than 3 coords here but emit_fetch_texel copies all 5 anyway */
2399 for (i
= dims
; i
< 5; i
++) {
2400 coords
[i
] = coord_undef
;
2403 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2405 if (inst
->Texture
.NumOffsets
== 1) {
2407 for (dim
= 0; dim
< dims
; dim
++) {
2408 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2412 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
2413 bld
->bld_base
.base
.gallivm
,
2414 bld
->bld_base
.base
.type
,
2420 NULL
, explicit_lod
, lod_property
,
2424 (inst
->Src
[1].Register
.SwizzleX
!= PIPE_SWIZZLE_RED
||
2425 inst
->Src
[1].Register
.SwizzleY
!= PIPE_SWIZZLE_GREEN
||
2426 inst
->Src
[1].Register
.SwizzleZ
!= PIPE_SWIZZLE_BLUE
||
2427 inst
->Src
[1].Register
.SwizzleW
!= PIPE_SWIZZLE_ALPHA
)) {
2428 unsigned char swizzles
[4];
2429 swizzles
[0] = inst
->Src
[1].Register
.SwizzleX
;
2430 swizzles
[1] = inst
->Src
[1].Register
.SwizzleY
;
2431 swizzles
[2] = inst
->Src
[1].Register
.SwizzleZ
;
2432 swizzles
[3] = inst
->Src
[1].Register
.SwizzleW
;
2434 lp_build_swizzle_soa_inplace(&bld
->bld_base
.base
, texel
, swizzles
);
2439 emit_size_query( struct lp_build_tgsi_soa_context
*bld
,
2440 const struct tgsi_full_instruction
*inst
,
2441 LLVMValueRef
*sizes_out
,
2442 boolean is_sviewinfo
)
2444 LLVMValueRef explicit_lod
;
2445 enum lp_sampler_lod_property lod_property
;
2448 unsigned unit
= inst
->Src
[1].Register
.Index
;
2449 unsigned target
, pipe_target
;
2452 target
= bld
->sv
[unit
].Resource
;
2455 target
= inst
->Texture
.Texture
;
2458 case TGSI_TEXTURE_BUFFER
:
2459 case TGSI_TEXTURE_RECT
:
2460 case TGSI_TEXTURE_SHADOWRECT
:
2468 if (!bld
->sampler
) {
2469 _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
2470 for (i
= 0; i
< 4; i
++)
2471 sizes_out
[i
] = bld
->bld_base
.int_bld
.undef
;
2476 explicit_lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 0);
2477 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2480 explicit_lod
= NULL
;
2481 lod_property
= LP_SAMPLER_LOD_SCALAR
;
2485 pipe_target
= tgsi_to_pipe_tex_target(target
);
2487 bld
->sampler
->emit_size_query(bld
->sampler
,
2488 bld
->bld_base
.base
.gallivm
,
2489 bld
->bld_base
.int_bld
.type
,
2498 near_end_of_shader(struct lp_build_tgsi_soa_context
*bld
,
2503 for (i
= 0; i
< 5; i
++) {
2506 if (pc
+ i
>= bld
->bld_base
.info
->num_instructions
)
2509 opcode
= bld
->bld_base
.instructions
[pc
+ i
].Instruction
.Opcode
;
2511 if (opcode
== TGSI_OPCODE_END
)
2514 if (opcode
== TGSI_OPCODE_TEX
||
2515 opcode
== TGSI_OPCODE_TXP
||
2516 opcode
== TGSI_OPCODE_TXD
||
2517 opcode
== TGSI_OPCODE_TXB
||
2518 opcode
== TGSI_OPCODE_TXL
||
2519 opcode
== TGSI_OPCODE_TXF
||
2520 opcode
== TGSI_OPCODE_TXQ
||
2521 opcode
== TGSI_OPCODE_TEX2
||
2522 opcode
== TGSI_OPCODE_TXB2
||
2523 opcode
== TGSI_OPCODE_TXL2
||
2524 opcode
== TGSI_OPCODE_SAMPLE
||
2525 opcode
== TGSI_OPCODE_SAMPLE_B
||
2526 opcode
== TGSI_OPCODE_SAMPLE_C
||
2527 opcode
== TGSI_OPCODE_SAMPLE_C_LZ
||
2528 opcode
== TGSI_OPCODE_SAMPLE_D
||
2529 opcode
== TGSI_OPCODE_SAMPLE_I
||
2530 opcode
== TGSI_OPCODE_SAMPLE_L
||
2531 opcode
== TGSI_OPCODE_SVIEWINFO
||
2532 opcode
== TGSI_OPCODE_CAL
||
2533 opcode
== TGSI_OPCODE_CALLNZ
||
2534 opcode
== TGSI_OPCODE_IF
||
2535 opcode
== TGSI_OPCODE_UIF
||
2536 opcode
== TGSI_OPCODE_BGNLOOP
||
2537 opcode
== TGSI_OPCODE_SWITCH
)
2547 * Kill fragment if any of the src register values are negative.
2551 struct lp_build_tgsi_soa_context
*bld
,
2552 const struct tgsi_full_instruction
*inst
,
2555 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2556 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
2557 LLVMValueRef terms
[TGSI_NUM_CHANNELS
];
2559 unsigned chan_index
;
2561 memset(&terms
, 0, sizeof terms
);
2563 TGSI_FOR_EACH_CHANNEL( chan_index
) {
2566 /* Unswizzle channel */
2567 swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
2569 /* Check if the component has not been already tested. */
2570 assert(swizzle
< TGSI_NUM_CHANNELS
);
2571 if( !terms
[swizzle
] )
2572 /* TODO: change the comparison operator instead of setting the sign */
2573 terms
[swizzle
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, chan_index
);
2577 TGSI_FOR_EACH_CHANNEL( chan_index
) {
2578 if(terms
[chan_index
]) {
2579 LLVMValueRef chan_mask
;
2582 * If term < 0 then mask = 0 else mask = ~0.
2584 chan_mask
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, terms
[chan_index
], bld
->bld_base
.base
.zero
);
2587 mask
= LLVMBuildAnd(builder
, mask
, chan_mask
, "");
2593 if (bld
->exec_mask
.has_mask
) {
2594 LLVMValueRef invmask
;
2595 invmask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
2596 mask
= LLVMBuildOr(builder
, mask
, invmask
, "");
2599 lp_build_mask_update(bld
->mask
, mask
);
2600 if (!near_end_of_shader(bld
, pc
))
2601 lp_build_mask_check(bld
->mask
);
2606 * Unconditional fragment kill.
2607 * The only predication is the execution mask which will apply if
2608 * we're inside a loop or conditional.
2611 emit_kill(struct lp_build_tgsi_soa_context
*bld
,
2614 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2617 /* For those channels which are "alive", disable fragment shader
2620 if (bld
->exec_mask
.has_mask
) {
2621 mask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
2624 LLVMValueRef zero
= LLVMConstNull(bld
->bld_base
.base
.int_vec_type
);
2628 lp_build_mask_update(bld
->mask
, mask
);
2630 if (!near_end_of_shader(bld
, pc
))
2631 lp_build_mask_check(bld
->mask
);
2636 * Emit code which will dump the value of all the temporary registers
2640 emit_dump_file(struct lp_build_tgsi_soa_context
*bld
,
2643 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
2644 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2645 LLVMBuilderRef builder
= gallivm
->builder
;
2646 LLVMValueRef reg_ptr
;
2648 int max_index
= info
->file_max
[file
];
2651 * Some register files, particularly constants, can be very large,
2652 * and dumping everything could make this unusably slow.
2654 max_index
= MIN2(max_index
, 32);
2656 for (index
= 0; index
<= max_index
; index
++) {
2661 if (index
< 8 * sizeof(unsigned) &&
2662 (info
->file_mask
[file
] & (1 << index
)) == 0) {
2663 /* This was not declared.*/
2667 if (file
== TGSI_FILE_INPUT
) {
2668 mask
= info
->input_usage_mask
[index
];
2670 mask
= TGSI_WRITEMASK_XYZW
;
2673 for (chan
= 0; chan
< 4; chan
++) {
2674 if ((mask
& (1 << chan
)) == 0) {
2675 /* This channel is not used.*/
2679 if (file
== TGSI_FILE_CONSTANT
) {
2680 struct tgsi_full_src_register reg
;
2681 memset(®
, 0, sizeof reg
);
2682 reg
.Register
.File
= file
;
2683 reg
.Register
.Index
= index
;
2684 reg
.Register
.SwizzleX
= 0;
2685 reg
.Register
.SwizzleY
= 1;
2686 reg
.Register
.SwizzleZ
= 2;
2687 reg
.Register
.SwizzleW
= 3;
2689 res
= bld
->bld_base
.emit_fetch_funcs
[file
](&bld
->bld_base
, ®
, TGSI_TYPE_FLOAT
, chan
);
2693 } else if (file
== TGSI_FILE_INPUT
) {
2694 res
= bld
->inputs
[index
][chan
];
2698 } else if (file
== TGSI_FILE_TEMPORARY
) {
2699 reg_ptr
= lp_get_temp_ptr_soa(bld
, index
, chan
);
2701 res
= LLVMBuildLoad(builder
, reg_ptr
, "");
2702 } else if (file
== TGSI_FILE_OUTPUT
) {
2703 reg_ptr
= lp_get_output_ptr(bld
, index
, chan
);
2705 res
= LLVMBuildLoad(builder
, reg_ptr
, "");
2711 emit_dump_reg(gallivm
, file
, index
, chan
, res
);
2719 lp_emit_declaration_soa(
2720 struct lp_build_tgsi_context
*bld_base
,
2721 const struct tgsi_full_declaration
*decl
)
2723 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
2724 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2725 LLVMTypeRef vec_type
= bld
->bld_base
.base
.vec_type
;
2726 const unsigned first
= decl
->Range
.First
;
2727 const unsigned last
= decl
->Range
.Last
;
2730 assert(last
<= bld
->bld_base
.info
->file_max
[decl
->Declaration
.File
]);
2732 switch (decl
->Declaration
.File
) {
2733 case TGSI_FILE_TEMPORARY
:
2734 if (!(bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
))) {
2735 assert(last
< LP_MAX_INLINED_TEMPS
);
2736 for (idx
= first
; idx
<= last
; ++idx
) {
2737 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2738 bld
->temps
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
, "temp");
2743 case TGSI_FILE_OUTPUT
:
2744 if (!(bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
2745 for (idx
= first
; idx
<= last
; ++idx
) {
2746 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2747 bld
->outputs
[idx
][i
] = lp_build_alloca(gallivm
,
2748 vec_type
, "output");
2753 case TGSI_FILE_ADDRESS
:
2754 /* ADDR registers are only allocated with an integer LLVM IR type,
2755 * as they are guaranteed to always have integers.
2756 * XXX: Not sure if this exception is worthwhile (or the whole idea of
2757 * an ADDR register for that matter).
2759 assert(last
< LP_MAX_TGSI_ADDRS
);
2760 for (idx
= first
; idx
<= last
; ++idx
) {
2761 assert(idx
< LP_MAX_TGSI_ADDRS
);
2762 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2763 bld
->addr
[idx
][i
] = lp_build_alloca(gallivm
, bld_base
->base
.int_vec_type
, "addr");
2767 case TGSI_FILE_PREDICATE
:
2768 assert(last
< LP_MAX_TGSI_PREDS
);
2769 for (idx
= first
; idx
<= last
; ++idx
) {
2770 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2771 bld
->preds
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
,
2776 case TGSI_FILE_SAMPLER_VIEW
:
2778 * The target stored here MUST match whatever there actually
2779 * is in the set sampler views (what about return type?).
2781 assert(last
< PIPE_MAX_SHADER_SAMPLER_VIEWS
);
2782 for (idx
= first
; idx
<= last
; ++idx
) {
2783 bld
->sv
[idx
] = decl
->SamplerView
;
2787 case TGSI_FILE_CONSTANT
:
2790 * We could trivially fetch the per-buffer pointer when fetching the
2791 * constant, relying on llvm to figure out it's always the same pointer
2792 * anyway. However, doing so results in a huge (more than factor of 10)
2793 * slowdown in llvm compilation times for some (but not all) shaders
2794 * (more specifically, the IR optimization spends way more time in
2795 * DominatorTree::dominates). At least with llvm versions 3.1, 3.3.
2797 unsigned idx2D
= decl
->Dim
.Index2D
;
2798 LLVMValueRef index2D
= lp_build_const_int32(gallivm
, idx2D
);
2799 assert(idx2D
< LP_MAX_TGSI_CONST_BUFFERS
);
2800 bld
->consts
[idx2D
] =
2801 lp_build_array_get(gallivm
, bld
->consts_ptr
, index2D
);
2802 bld
->consts_sizes
[idx2D
] =
2803 lp_build_array_get(gallivm
, bld
->const_sizes_ptr
, index2D
);
2808 /* don't need to declare other vars */
2814 void lp_emit_immediate_soa(
2815 struct lp_build_tgsi_context
*bld_base
,
2816 const struct tgsi_full_immediate
*imm
)
2818 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
2819 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
2820 LLVMValueRef imms
[4];
2822 const uint size
= imm
->Immediate
.NrTokens
- 1;
2824 switch (imm
->Immediate
.DataType
) {
2825 case TGSI_IMM_FLOAT32
:
2826 for( i
= 0; i
< size
; ++i
)
2828 lp_build_const_vec(gallivm
, bld_base
->base
.type
, imm
->u
[i
].Float
);
2831 case TGSI_IMM_UINT32
:
2832 for( i
= 0; i
< size
; ++i
) {
2833 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->uint_bld
.type
, imm
->u
[i
].Uint
);
2834 imms
[i
] = LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
2838 case TGSI_IMM_INT32
:
2839 for( i
= 0; i
< size
; ++i
) {
2840 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->int_bld
.type
, imm
->u
[i
].Int
);
2841 imms
[i
] = LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
2846 for( i
= size
; i
< 4; ++i
)
2847 imms
[i
] = bld_base
->base
.undef
;
2849 if (bld
->use_immediates_array
) {
2850 unsigned index
= bld
->num_immediates
;
2851 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2852 LLVMBuilderRef builder
= gallivm
->builder
;
2854 assert(bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
));
2855 for (i
= 0; i
< 4; ++i
) {
2856 LLVMValueRef lindex
= lp_build_const_int32(
2857 bld
->bld_base
.base
.gallivm
, index
* 4 + i
);
2858 LLVMValueRef imm_ptr
= LLVMBuildGEP(builder
,
2859 bld
->imms_array
, &lindex
, 1, "");
2860 LLVMBuildStore(builder
, imms
[i
], imm_ptr
);
2863 /* simply copy the immediate values into the next immediates[] slot */
2865 const uint size
= imm
->Immediate
.NrTokens
- 1;
2867 assert(bld
->num_immediates
< LP_MAX_INLINED_IMMEDIATES
);
2869 for(i
= 0; i
< 4; ++i
)
2870 bld
->immediates
[bld
->num_immediates
][i
] = imms
[i
];
2872 if (bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
)) {
2873 unsigned index
= bld
->num_immediates
;
2874 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2875 LLVMBuilderRef builder
= gallivm
->builder
;
2876 for (i
= 0; i
< 4; ++i
) {
2877 LLVMValueRef lindex
= lp_build_const_int32(
2878 bld
->bld_base
.base
.gallivm
, index
* 4 + i
);
2879 LLVMValueRef imm_ptr
= LLVMBuildGEP(builder
,
2880 bld
->imms_array
, &lindex
, 1, "");
2881 LLVMBuildStore(builder
,
2882 bld
->immediates
[index
][i
],
2888 bld
->num_immediates
++;
2893 const struct lp_build_tgsi_action
* action
,
2894 struct lp_build_tgsi_context
* bld_base
,
2895 struct lp_build_emit_data
* emit_data
)
2897 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2899 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
,
2900 &emit_data
->output
[emit_data
->chan
], NULL
);
2905 const struct lp_build_tgsi_action
* action
,
2906 struct lp_build_tgsi_context
* bld_base
,
2907 struct lp_build_emit_data
* emit_data
)
2909 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2911 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
, NULL
,
2912 &emit_data
->output
[emit_data
->chan
]);
2917 const struct lp_build_tgsi_action
* action
,
2918 struct lp_build_tgsi_context
* bld_base
,
2919 struct lp_build_emit_data
* emit_data
)
2921 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2923 emit_kill(bld
, bld_base
->pc
- 1);
2928 const struct lp_build_tgsi_action
* action
,
2929 struct lp_build_tgsi_context
* bld_base
,
2930 struct lp_build_emit_data
* emit_data
)
2932 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2934 emit_kill_if(bld
, emit_data
->inst
, bld_base
->pc
- 1);
2939 const struct lp_build_tgsi_action
* action
,
2940 struct lp_build_tgsi_context
* bld_base
,
2941 struct lp_build_emit_data
* emit_data
)
2943 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2945 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
2946 emit_data
->output
, 1);
2951 const struct lp_build_tgsi_action
* action
,
2952 struct lp_build_tgsi_context
* bld_base
,
2953 struct lp_build_emit_data
* emit_data
)
2955 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2957 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
2958 emit_data
->output
, 2);
2963 const struct lp_build_tgsi_action
* action
,
2964 struct lp_build_tgsi_context
* bld_base
,
2965 struct lp_build_emit_data
* emit_data
)
2967 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2969 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
2970 emit_data
->output
, 1);
2975 const struct lp_build_tgsi_action
* action
,
2976 struct lp_build_tgsi_context
* bld_base
,
2977 struct lp_build_emit_data
* emit_data
)
2979 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2981 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
2982 emit_data
->output
, 2);
2987 const struct lp_build_tgsi_action
* action
,
2988 struct lp_build_tgsi_context
* bld_base
,
2989 struct lp_build_emit_data
* emit_data
)
2991 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2993 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
2994 emit_data
->output
, 3);
2999 const struct lp_build_tgsi_action
* action
,
3000 struct lp_build_tgsi_context
* bld_base
,
3001 struct lp_build_emit_data
* emit_data
)
3003 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3005 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
3006 emit_data
->output
, 1);
3011 const struct lp_build_tgsi_action
* action
,
3012 struct lp_build_tgsi_context
* bld_base
,
3013 struct lp_build_emit_data
* emit_data
)
3015 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3017 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
3018 emit_data
->output
, 2);
3023 const struct lp_build_tgsi_action
* action
,
3024 struct lp_build_tgsi_context
* bld_base
,
3025 struct lp_build_emit_data
* emit_data
)
3027 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3029 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_PROJECTED
,
3030 emit_data
->output
, 1);
3035 const struct lp_build_tgsi_action
* action
,
3036 struct lp_build_tgsi_context
* bld_base
,
3037 struct lp_build_emit_data
* emit_data
)
3039 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3041 emit_size_query(bld
, emit_data
->inst
, emit_data
->output
, FALSE
);
3046 const struct lp_build_tgsi_action
* action
,
3047 struct lp_build_tgsi_context
* bld_base
,
3048 struct lp_build_emit_data
* emit_data
)
3050 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3052 emit_fetch_texels(bld
, emit_data
->inst
, emit_data
->output
, FALSE
);
3057 const struct lp_build_tgsi_action
* action
,
3058 struct lp_build_tgsi_context
* bld_base
,
3059 struct lp_build_emit_data
* emit_data
)
3061 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3063 emit_fetch_texels(bld
, emit_data
->inst
, emit_data
->output
, TRUE
);
3068 const struct lp_build_tgsi_action
* action
,
3069 struct lp_build_tgsi_context
* bld_base
,
3070 struct lp_build_emit_data
* emit_data
)
3072 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3074 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
3075 FALSE
, emit_data
->output
);
3080 const struct lp_build_tgsi_action
* action
,
3081 struct lp_build_tgsi_context
* bld_base
,
3082 struct lp_build_emit_data
* emit_data
)
3084 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3086 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
3087 FALSE
, emit_data
->output
);
3092 const struct lp_build_tgsi_action
* action
,
3093 struct lp_build_tgsi_context
* bld_base
,
3094 struct lp_build_emit_data
* emit_data
)
3096 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3098 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
3099 TRUE
, emit_data
->output
);
3104 const struct lp_build_tgsi_action
* action
,
3105 struct lp_build_tgsi_context
* bld_base
,
3106 struct lp_build_emit_data
* emit_data
)
3108 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3110 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_ZERO
,
3111 TRUE
, emit_data
->output
);
3116 const struct lp_build_tgsi_action
* action
,
3117 struct lp_build_tgsi_context
* bld_base
,
3118 struct lp_build_emit_data
* emit_data
)
3120 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3122 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
3123 FALSE
, emit_data
->output
);
3128 const struct lp_build_tgsi_action
* action
,
3129 struct lp_build_tgsi_context
* bld_base
,
3130 struct lp_build_emit_data
* emit_data
)
3132 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3134 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
3135 FALSE
, emit_data
->output
);
3140 const struct lp_build_tgsi_action
* action
,
3141 struct lp_build_tgsi_context
* bld_base
,
3142 struct lp_build_emit_data
* emit_data
)
3144 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3146 emit_size_query(bld
, emit_data
->inst
, emit_data
->output
, TRUE
);
3150 mask_vec(struct lp_build_tgsi_context
*bld_base
)
3152 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3153 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3154 struct lp_exec_mask
*exec_mask
= &bld
->exec_mask
;
3156 if (!exec_mask
->has_mask
) {
3157 return lp_build_mask_value(bld
->mask
);
3159 return LLVMBuildAnd(builder
, lp_build_mask_value(bld
->mask
),
3160 exec_mask
->exec_mask
, "");
3164 increment_vec_ptr_by_mask(struct lp_build_tgsi_context
* bld_base
,
3168 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3169 LLVMValueRef current_vec
= LLVMBuildLoad(builder
, ptr
, "");
3171 current_vec
= LLVMBuildSub(builder
, current_vec
, mask
, "");
3173 LLVMBuildStore(builder
, current_vec
, ptr
);
3177 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context
* bld_base
,
3181 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3182 LLVMValueRef current_vec
= LLVMBuildLoad(builder
, ptr
, "");
3184 current_vec
= lp_build_select(&bld_base
->uint_bld
,
3186 bld_base
->uint_bld
.zero
,
3189 LLVMBuildStore(builder
, current_vec
, ptr
);
3193 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context
* bld
,
3194 LLVMValueRef current_mask_vec
,
3195 LLVMValueRef total_emitted_vertices_vec
)
3197 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3198 struct lp_build_context
*int_bld
= &bld
->bld_base
.int_bld
;
3199 LLVMValueRef max_mask
= lp_build_cmp(int_bld
, PIPE_FUNC_LESS
,
3200 total_emitted_vertices_vec
,
3201 bld
->max_output_vertices_vec
);
3203 return LLVMBuildAnd(builder
, current_mask_vec
, max_mask
, "");
3208 const struct lp_build_tgsi_action
* action
,
3209 struct lp_build_tgsi_context
* bld_base
,
3210 struct lp_build_emit_data
* emit_data
)
3212 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3213 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3215 if (bld
->gs_iface
->emit_vertex
) {
3216 LLVMValueRef mask
= mask_vec(bld_base
);
3217 LLVMValueRef total_emitted_vertices_vec
=
3218 LLVMBuildLoad(builder
, bld
->total_emitted_vertices_vec_ptr
, "");
3219 mask
= clamp_mask_to_max_output_vertices(bld
, mask
,
3220 total_emitted_vertices_vec
);
3221 gather_outputs(bld
);
3222 bld
->gs_iface
->emit_vertex(bld
->gs_iface
, &bld
->bld_base
,
3224 total_emitted_vertices_vec
);
3225 increment_vec_ptr_by_mask(bld_base
, bld
->emitted_vertices_vec_ptr
,
3227 increment_vec_ptr_by_mask(bld_base
, bld
->total_emitted_vertices_vec_ptr
,
3230 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3231 " +++ emit vertex masked ones = ",
3233 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3234 " +++ emit vertex emitted = ",
3235 total_emitted_vertices_vec
);
3242 end_primitive_masked(struct lp_build_tgsi_context
* bld_base
,
3245 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3246 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3248 if (bld
->gs_iface
->end_primitive
) {
3249 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3250 LLVMValueRef emitted_vertices_vec
=
3251 LLVMBuildLoad(builder
, bld
->emitted_vertices_vec_ptr
, "");
3252 LLVMValueRef emitted_prims_vec
=
3253 LLVMBuildLoad(builder
, bld
->emitted_prims_vec_ptr
, "");
3255 LLVMValueRef emitted_mask
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3256 emitted_vertices_vec
,
3258 /* We need to combine the current execution mask with the mask
3259 telling us which, if any, execution slots actually have
3260 unemitted primitives, this way we make sure that end_primitives
3261 executes only on the paths that have unflushed vertices */
3262 mask
= LLVMBuildAnd(builder
, mask
, emitted_mask
, "");
3264 bld
->gs_iface
->end_primitive(bld
->gs_iface
, &bld
->bld_base
,
3265 emitted_vertices_vec
,
3269 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3270 " +++ end prim masked ones = ",
3272 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3273 " +++ end prim emitted verts1 = ",
3274 emitted_vertices_vec
);
3275 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3276 " +++ end prim emitted prims1 = ",
3277 LLVMBuildLoad(builder
,
3278 bld
->emitted_prims_vec_ptr
, ""));
3280 increment_vec_ptr_by_mask(bld_base
, bld
->emitted_prims_vec_ptr
,
3282 clear_uint_vec_ptr_from_mask(bld_base
, bld
->emitted_vertices_vec_ptr
,
3285 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3286 " +++ end prim emitted verts2 = ",
3287 LLVMBuildLoad(builder
,
3288 bld
->emitted_vertices_vec_ptr
, ""));
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 if (bld
->gs_iface
->end_primitive
) {
3303 LLVMValueRef mask
= mask_vec(bld_base
);
3304 end_primitive_masked(bld_base
, mask
);
3310 const struct lp_build_tgsi_action
* action
,
3311 struct lp_build_tgsi_context
* bld_base
,
3312 struct lp_build_emit_data
* emit_data
)
3314 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3316 lp_exec_mask_call(&bld
->exec_mask
, emit_data
->inst
->Label
.Label
,
3322 const struct lp_build_tgsi_action
* action
,
3323 struct lp_build_tgsi_context
* bld_base
,
3324 struct lp_build_emit_data
* emit_data
)
3326 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3328 lp_exec_mask_ret(&bld
->exec_mask
, &bld_base
->pc
);
3333 const struct lp_build_tgsi_action
* action
,
3334 struct lp_build_tgsi_context
* bld_base
,
3335 struct lp_build_emit_data
* emit_data
)
3337 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3339 lp_exec_break(&bld
->exec_mask
, bld_base
);
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
);
3349 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3350 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3351 LLVMValueRef unsigned_cond
=
3352 LLVMBuildBitCast(builder
, emit_data
->args
[0], uint_bld
->vec_type
, "");
3353 LLVMValueRef cond
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3357 lp_exec_break_condition(&bld
->exec_mask
, cond
);
3362 const struct lp_build_tgsi_action
* action
,
3363 struct lp_build_tgsi_context
* bld_base
,
3364 struct lp_build_emit_data
* emit_data
)
3367 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3369 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_NOTEQUAL
,
3370 emit_data
->args
[0], bld
->bld_base
.base
.zero
);
3371 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
3376 const struct lp_build_tgsi_action
* action
,
3377 struct lp_build_tgsi_context
* bld_base
,
3378 struct lp_build_emit_data
* emit_data
)
3381 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3382 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3384 tmp
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3385 emit_data
->args
[0], uint_bld
->zero
);
3386 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
3391 const struct lp_build_tgsi_action
* action
,
3392 struct lp_build_tgsi_context
* bld_base
,
3393 struct lp_build_emit_data
* emit_data
)
3395 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3397 lp_exec_case(&bld
->exec_mask
, emit_data
->args
[0]);
3402 const struct lp_build_tgsi_action
* action
,
3403 struct lp_build_tgsi_context
* bld_base
,
3404 struct lp_build_emit_data
* emit_data
)
3406 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3408 lp_exec_default(&bld
->exec_mask
, bld_base
);
3413 const struct lp_build_tgsi_action
* action
,
3414 struct lp_build_tgsi_context
* bld_base
,
3415 struct lp_build_emit_data
* emit_data
)
3417 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3419 lp_exec_switch(&bld
->exec_mask
, emit_data
->args
[0]);
3424 const struct lp_build_tgsi_action
* action
,
3425 struct lp_build_tgsi_context
* bld_base
,
3426 struct lp_build_emit_data
* emit_data
)
3428 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3430 lp_exec_endswitch(&bld
->exec_mask
, bld_base
);
3435 const struct lp_build_tgsi_action
* action
,
3436 struct lp_build_tgsi_context
* bld_base
,
3437 struct lp_build_emit_data
* emit_data
)
3439 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3441 lp_exec_bgnloop(&bld
->exec_mask
);
3446 const struct lp_build_tgsi_action
* action
,
3447 struct lp_build_tgsi_context
* bld_base
,
3448 struct lp_build_emit_data
* emit_data
)
3450 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3452 lp_exec_mask_bgnsub(&bld
->exec_mask
);
3457 const struct lp_build_tgsi_action
* action
,
3458 struct lp_build_tgsi_context
* bld_base
,
3459 struct lp_build_emit_data
* emit_data
)
3461 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3463 lp_exec_mask_cond_invert(&bld
->exec_mask
);
3468 const struct lp_build_tgsi_action
* action
,
3469 struct lp_build_tgsi_context
* bld_base
,
3470 struct lp_build_emit_data
* emit_data
)
3472 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3474 lp_exec_mask_cond_pop(&bld
->exec_mask
);
3479 const struct lp_build_tgsi_action
* action
,
3480 struct lp_build_tgsi_context
* bld_base
,
3481 struct lp_build_emit_data
* emit_data
)
3483 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3485 lp_exec_endloop(bld_base
->base
.gallivm
, &bld
->exec_mask
);
3490 const struct lp_build_tgsi_action
* action
,
3491 struct lp_build_tgsi_context
* bld_base
,
3492 struct lp_build_emit_data
* emit_data
)
3494 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3496 lp_exec_mask_endsub(&bld
->exec_mask
, &bld_base
->pc
);
3501 const struct lp_build_tgsi_action
* action
,
3502 struct lp_build_tgsi_context
* bld_base
,
3503 struct lp_build_emit_data
* emit_data
)
3505 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3507 lp_exec_continue(&bld
->exec_mask
);
3510 static void emit_prologue(struct lp_build_tgsi_context
* bld_base
)
3512 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3513 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
3515 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
3516 LLVMValueRef array_size
=
3517 lp_build_const_int32(gallivm
,
3518 bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] * 4 + 4);
3519 bld
->temps_array
= lp_build_array_alloca(gallivm
,
3520 bld_base
->base
.vec_type
, array_size
,
3524 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
3525 LLVMValueRef array_size
=
3526 lp_build_const_int32(gallivm
,
3527 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] * 4 + 4);
3528 bld
->outputs_array
= lp_build_array_alloca(gallivm
,
3529 bld_base
->base
.vec_type
, array_size
,
3533 if (bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
)) {
3534 LLVMValueRef array_size
=
3535 lp_build_const_int32(gallivm
,
3536 bld_base
->info
->file_max
[TGSI_FILE_IMMEDIATE
] * 4 + 4);
3537 bld
->imms_array
= lp_build_array_alloca(gallivm
,
3538 bld_base
->base
.vec_type
, array_size
,
3542 /* If we have indirect addressing in inputs we need to copy them into
3543 * our alloca array to be able to iterate over them */
3544 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
) && !bld
->gs_iface
) {
3545 unsigned index
, chan
;
3546 LLVMTypeRef vec_type
= bld_base
->base
.vec_type
;
3547 LLVMValueRef array_size
= lp_build_const_int32(gallivm
,
3548 bld_base
->info
->file_max
[TGSI_FILE_INPUT
]*4 + 4);
3549 bld
->inputs_array
= lp_build_array_alloca(gallivm
,
3550 vec_type
, array_size
,
3553 assert(bld_base
->info
->num_inputs
3554 <= bld_base
->info
->file_max
[TGSI_FILE_INPUT
] + 1);
3556 for (index
= 0; index
< bld_base
->info
->num_inputs
; ++index
) {
3557 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3558 LLVMValueRef lindex
=
3559 lp_build_const_int32(gallivm
, index
* 4 + chan
);
3560 LLVMValueRef input_ptr
=
3561 LLVMBuildGEP(gallivm
->builder
, bld
->inputs_array
,
3563 LLVMValueRef value
= bld
->inputs
[index
][chan
];
3565 LLVMBuildStore(gallivm
->builder
, value
, input_ptr
);
3570 if (bld
->gs_iface
) {
3571 struct lp_build_context
*uint_bld
= &bld
->bld_base
.uint_bld
;
3572 bld
->emitted_prims_vec_ptr
=
3573 lp_build_alloca(gallivm
,
3575 "emitted_prims_ptr");
3576 bld
->emitted_vertices_vec_ptr
=
3577 lp_build_alloca(gallivm
,
3579 "emitted_vertices_ptr");
3580 bld
->total_emitted_vertices_vec_ptr
=
3581 lp_build_alloca(gallivm
,
3583 "total_emitted_vertices_ptr");
3585 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3586 bld
->emitted_prims_vec_ptr
);
3587 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3588 bld
->emitted_vertices_vec_ptr
);
3589 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3590 bld
->total_emitted_vertices_vec_ptr
);
3593 if (DEBUG_EXECUTION
) {
3594 lp_build_printf(gallivm
, "\n");
3595 emit_dump_file(bld
, TGSI_FILE_CONSTANT
);
3597 emit_dump_file(bld
, TGSI_FILE_INPUT
);
3601 static void emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
3603 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3604 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3606 if (DEBUG_EXECUTION
) {
3609 emit_dump_file(bld
, TGSI_FILE_TEMPORARY
);
3611 emit_dump_file(bld
, TGSI_FILE_OUTPUT
);
3612 lp_build_printf(bld_base
->base
.gallivm
, "\n");
3615 /* If we have indirect addressing in outputs we need to copy our alloca array
3616 * to the outputs slots specified by the caller */
3617 if (bld
->gs_iface
) {
3618 LLVMValueRef total_emitted_vertices_vec
;
3619 LLVMValueRef emitted_prims_vec
;
3620 /* implicit end_primitives, needed in case there are any unflushed
3621 vertices in the cache. Note must not call end_primitive here
3622 since the exec_mask is not valid at this point. */
3623 end_primitive_masked(bld_base
, lp_build_mask_value(bld
->mask
));
3625 total_emitted_vertices_vec
=
3626 LLVMBuildLoad(builder
, bld
->total_emitted_vertices_vec_ptr
, "");
3628 LLVMBuildLoad(builder
, bld
->emitted_prims_vec_ptr
, "");
3630 bld
->gs_iface
->gs_epilogue(bld
->gs_iface
,
3632 total_emitted_vertices_vec
,
3635 gather_outputs(bld
);
3640 lp_build_tgsi_soa(struct gallivm_state
*gallivm
,
3641 const struct tgsi_token
*tokens
,
3642 struct lp_type type
,
3643 struct lp_build_mask_context
*mask
,
3644 LLVMValueRef consts_ptr
,
3645 LLVMValueRef const_sizes_ptr
,
3646 const struct lp_bld_tgsi_system_values
*system_values
,
3647 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
3648 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
3649 struct lp_build_sampler_soa
*sampler
,
3650 const struct tgsi_shader_info
*info
,
3651 const struct lp_build_tgsi_gs_iface
*gs_iface
)
3653 struct lp_build_tgsi_soa_context bld
;
3655 struct lp_type res_type
;
3657 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
3658 memset(&res_type
, 0, sizeof res_type
);
3659 res_type
.width
= type
.width
;
3660 res_type
.length
= type
.length
;
3663 /* Setup build context */
3664 memset(&bld
, 0, sizeof bld
);
3665 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
3666 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
3667 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
3668 lp_build_context_init(&bld
.elem_bld
, gallivm
, lp_elem_type(type
));
3670 bld
.inputs
= inputs
;
3671 bld
.outputs
= outputs
;
3672 bld
.consts_ptr
= consts_ptr
;
3673 bld
.const_sizes_ptr
= const_sizes_ptr
;
3674 bld
.sampler
= sampler
;
3675 bld
.bld_base
.info
= info
;
3676 bld
.indirect_files
= info
->indirect_files
;
3679 * If the number of temporaries is rather large then we just
3680 * allocate them as an array right from the start and treat
3681 * like indirect temporaries.
3683 if (info
->file_max
[TGSI_FILE_TEMPORARY
] >= LP_MAX_INLINED_TEMPS
) {
3684 bld
.indirect_files
|= (1 << TGSI_FILE_TEMPORARY
);
3687 * For performance reason immediates are always backed in a static
3688 * array, but if their number is too great, we have to use just
3689 * a dynamically allocated array.
3691 bld
.use_immediates_array
=
3692 (info
->file_max
[TGSI_FILE_IMMEDIATE
] >= LP_MAX_INLINED_IMMEDIATES
);
3693 if (bld
.use_immediates_array
) {
3694 bld
.indirect_files
|= (1 << TGSI_FILE_IMMEDIATE
);
3698 bld
.bld_base
.soa
= TRUE
;
3699 bld
.bld_base
.emit_debug
= emit_debug
;
3700 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
3701 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
3702 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
3703 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
3704 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = emit_fetch_system_value
;
3705 bld
.bld_base
.emit_store
= emit_store
;
3707 bld
.bld_base
.emit_declaration
= lp_emit_declaration_soa
;
3708 bld
.bld_base
.emit_immediate
= lp_emit_immediate_soa
;
3710 bld
.bld_base
.emit_prologue
= emit_prologue
;
3711 bld
.bld_base
.emit_epilogue
= emit_epilogue
;
3713 /* Set opcode actions */
3714 lp_set_default_actions_cpu(&bld
.bld_base
);
3716 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
3717 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNSUB
].emit
= bgnsub_emit
;
3718 bld
.bld_base
.op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
3719 bld
.bld_base
.op_actions
[TGSI_OPCODE_BREAKC
].emit
= breakc_emit
;
3720 bld
.bld_base
.op_actions
[TGSI_OPCODE_CAL
].emit
= cal_emit
;
3721 bld
.bld_base
.op_actions
[TGSI_OPCODE_CASE
].emit
= case_emit
;
3722 bld
.bld_base
.op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
3723 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDX
].emit
= ddx_emit
;
3724 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDY
].emit
= ddy_emit
;
3725 bld
.bld_base
.op_actions
[TGSI_OPCODE_DEFAULT
].emit
= default_emit
;
3726 bld
.bld_base
.op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
3727 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
3728 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
3729 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSUB
].emit
= endsub_emit
;
3730 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSWITCH
].emit
= endswitch_emit
;
3731 bld
.bld_base
.op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
3732 bld
.bld_base
.op_actions
[TGSI_OPCODE_UIF
].emit
= uif_emit
;
3733 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILL_IF
].emit
= kill_if_emit
;
3734 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILL
].emit
= kill_emit
;
3735 bld
.bld_base
.op_actions
[TGSI_OPCODE_RET
].emit
= ret_emit
;
3736 bld
.bld_base
.op_actions
[TGSI_OPCODE_SWITCH
].emit
= switch_emit
;
3737 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX
].emit
= tex_emit
;
3738 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB
].emit
= txb_emit
;
3739 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXD
].emit
= txd_emit
;
3740 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL
].emit
= txl_emit
;
3741 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXP
].emit
= txp_emit
;
3742 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXQ
].emit
= txq_emit
;
3743 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXF
].emit
= txf_emit
;
3744 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX2
].emit
= tex2_emit
;
3745 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB2
].emit
= txb2_emit
;
3746 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL2
].emit
= txl2_emit
;
3747 /* DX10 sampling ops */
3748 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE
].emit
= sample_emit
;
3749 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_B
].emit
= sample_b_emit
;
3750 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_C
].emit
= sample_c_emit
;
3751 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_C_LZ
].emit
= sample_c_lz_emit
;
3752 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_D
].emit
= sample_d_emit
;
3753 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_I
].emit
= sample_i_emit
;
3754 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_L
].emit
= sample_l_emit
;
3755 bld
.bld_base
.op_actions
[TGSI_OPCODE_SVIEWINFO
].emit
= sviewinfo_emit
;
3758 /* There's no specific value for this because it should always
3759 * be set, but apps using ext_geometry_shader4 quite often
3760 * were forgetting so we're using MAX_VERTEX_VARYING from
3761 * that spec even though we could debug_assert if it's not
3762 * set, but that's a lot uglier. */
3763 uint max_output_vertices
;
3765 /* inputs are always indirect with gs */
3766 bld
.indirect_files
|= (1 << TGSI_FILE_INPUT
);
3767 bld
.gs_iface
= gs_iface
;
3768 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_gs_input
;
3769 bld
.bld_base
.op_actions
[TGSI_OPCODE_EMIT
].emit
= emit_vertex
;
3770 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= end_primitive
;
3772 max_output_vertices
=
3773 info
->properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
];
3774 if (!max_output_vertices
)
3775 max_output_vertices
= 32;
3777 bld
.max_output_vertices_vec
=
3778 lp_build_const_int_vec(gallivm
, bld
.bld_base
.int_bld
.type
,
3779 max_output_vertices
);
3782 lp_exec_mask_init(&bld
.exec_mask
, &bld
.bld_base
.int_bld
);
3784 bld
.system_values
= *system_values
;
3786 lp_build_tgsi_llvm(&bld
.bld_base
, tokens
);
3789 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
3790 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
3791 debug_printf("11111111111111111111111111111 \n");
3792 tgsi_dump(tokens
, 0);
3793 lp_debug_dump_value(function
);
3794 debug_printf("2222222222222222222222222222 \n");
3798 LLVMModuleRef module
= LLVMGetGlobalParent(
3799 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
3800 LLVMDumpModule(module
);
3803 lp_exec_mask_fini(&bld
.exec_mask
);