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 dimension_index
;
1217 LLVMValueRef consts_ptr
;
1218 LLVMValueRef num_consts
;
1221 /* XXX: Handle fetching xyzw components as a vector */
1222 assert(swizzle
!= ~0);
1224 if (reg
->Register
.Dimension
) {
1225 assert(!reg
->Dimension
.Indirect
);
1226 dimension
= reg
->Dimension
.Index
;
1227 assert(dimension
< LP_MAX_TGSI_CONST_BUFFERS
);
1230 dimension_index
= lp_build_const_int32(gallivm
, dimension
);
1232 lp_build_array_get(gallivm
, bld
->consts_ptr
, dimension_index
);
1234 lp_build_array_get(gallivm
, bld
->const_sizes_ptr
, dimension_index
);
1236 if (reg
->Register
.Indirect
) {
1237 LLVMValueRef indirect_index
;
1238 LLVMValueRef swizzle_vec
=
1239 lp_build_const_int_vec(gallivm
, uint_bld
->type
, swizzle
);
1240 LLVMValueRef index_vec
; /* index into the const buffer */
1241 LLVMValueRef overflow_mask
;
1243 indirect_index
= get_indirect_index(bld
,
1245 reg
->Register
.Index
,
1248 /* All fetches are from the same constant buffer, so
1249 * we need to propagate the size to a vector to do a
1250 * vector comparison */
1251 num_consts
= lp_build_broadcast_scalar(uint_bld
, num_consts
);
1252 /* Construct a boolean vector telling us which channels
1253 * overflow the bound constant buffer */
1254 overflow_mask
= LLVMBuildICmp(builder
, LLVMIntUGE
,
1258 /* index_vec = indirect_index * 4 + swizzle */
1259 index_vec
= lp_build_shl_imm(uint_bld
, indirect_index
, 2);
1260 index_vec
= lp_build_add(uint_bld
, index_vec
, swizzle_vec
);
1262 /* Gather values from the constant buffer */
1263 res
= build_gather(&bld_base
->base
, consts_ptr
, index_vec
,
1267 LLVMValueRef index
; /* index into the const buffer */
1268 LLVMValueRef scalar
, scalar_ptr
;
1270 index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
* 4 + swizzle
);
1272 scalar_ptr
= LLVMBuildGEP(builder
, consts_ptr
,
1274 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
1275 res
= lp_build_broadcast_scalar(&bld_base
->base
, scalar
);
1278 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
) {
1279 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1280 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1287 emit_fetch_immediate(
1288 struct lp_build_tgsi_context
* bld_base
,
1289 const struct tgsi_full_src_register
* reg
,
1290 enum tgsi_opcode_type stype
,
1293 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1294 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1295 LLVMBuilderRef builder
= gallivm
->builder
;
1296 LLVMValueRef res
= NULL
;
1298 if (reg
->Register
.Indirect
) {
1299 LLVMValueRef indirect_index
;
1300 LLVMValueRef index_vec
; /* index into the immediate register array */
1301 LLVMValueRef imms_array
;
1302 LLVMTypeRef fptr_type
;
1304 indirect_index
= get_indirect_index(bld
,
1306 reg
->Register
.Index
,
1309 * Unlike for other reg classes, adding pixel offsets is unnecessary -
1310 * immediates are stored as full vectors (FIXME??? - might be better
1311 * to store them the same as constants) but all elements are the same
1314 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1319 /* cast imms_array pointer to float* */
1320 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1321 imms_array
= LLVMBuildBitCast(builder
, bld
->imms_array
, fptr_type
, "");
1323 /* Gather values from the immediate register array */
1324 res
= build_gather(&bld_base
->base
, imms_array
, index_vec
, NULL
);
1327 res
= bld
->immediates
[reg
->Register
.Index
][swizzle
];
1330 if (stype
== TGSI_TYPE_UNSIGNED
) {
1331 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1332 } else if (stype
== TGSI_TYPE_SIGNED
) {
1333 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1340 struct lp_build_tgsi_context
* bld_base
,
1341 const struct tgsi_full_src_register
* reg
,
1342 enum tgsi_opcode_type stype
,
1345 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1346 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1347 LLVMBuilderRef builder
= gallivm
->builder
;
1350 if (reg
->Register
.Indirect
) {
1351 LLVMValueRef indirect_index
;
1352 LLVMValueRef index_vec
; /* index into the input reg array */
1353 LLVMValueRef inputs_array
;
1354 LLVMTypeRef fptr_type
;
1356 indirect_index
= get_indirect_index(bld
,
1358 reg
->Register
.Index
,
1361 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1366 /* cast inputs_array pointer to float* */
1367 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1368 inputs_array
= LLVMBuildBitCast(builder
, bld
->inputs_array
, fptr_type
, "");
1370 /* Gather values from the input register array */
1371 res
= build_gather(&bld_base
->base
, inputs_array
, index_vec
, NULL
);
1373 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
)) {
1374 LLVMValueRef lindex
= lp_build_const_int32(gallivm
,
1375 reg
->Register
.Index
* 4 + swizzle
);
1376 LLVMValueRef input_ptr
= LLVMBuildGEP(builder
,
1377 bld
->inputs_array
, &lindex
, 1, "");
1378 res
= LLVMBuildLoad(builder
, input_ptr
, "");
1381 res
= bld
->inputs
[reg
->Register
.Index
][swizzle
];
1387 if (stype
== TGSI_TYPE_UNSIGNED
) {
1388 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1389 } else if (stype
== TGSI_TYPE_SIGNED
) {
1390 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1398 emit_fetch_gs_input(
1399 struct lp_build_tgsi_context
* bld_base
,
1400 const struct tgsi_full_src_register
* reg
,
1401 enum tgsi_opcode_type stype
,
1404 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1405 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1406 LLVMBuilderRef builder
= gallivm
->builder
;
1407 LLVMValueRef attrib_index
= NULL
;
1408 LLVMValueRef vertex_index
= NULL
;
1409 LLVMValueRef swizzle_index
= lp_build_const_int32(gallivm
, swizzle
);
1412 if (reg
->Register
.Indirect
) {
1413 attrib_index
= get_indirect_index(bld
,
1415 reg
->Register
.Index
,
1418 attrib_index
= lp_build_const_int32(gallivm
, reg
->Register
.Index
);
1421 if (reg
->Dimension
.Indirect
) {
1422 vertex_index
= get_indirect_index(bld
,
1424 reg
->Dimension
.Index
,
1427 vertex_index
= lp_build_const_int32(gallivm
, reg
->Dimension
.Index
);
1430 res
= bld
->gs_iface
->fetch_input(bld
->gs_iface
, bld_base
,
1431 reg
->Dimension
.Indirect
,
1433 reg
->Register
.Indirect
,
1439 if (stype
== TGSI_TYPE_UNSIGNED
) {
1440 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1441 } else if (stype
== TGSI_TYPE_SIGNED
) {
1442 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1449 emit_fetch_temporary(
1450 struct lp_build_tgsi_context
* bld_base
,
1451 const struct tgsi_full_src_register
* reg
,
1452 enum tgsi_opcode_type stype
,
1455 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1456 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1457 LLVMBuilderRef builder
= gallivm
->builder
;
1460 if (reg
->Register
.Indirect
) {
1461 LLVMValueRef indirect_index
;
1462 LLVMValueRef index_vec
; /* index into the temp reg array */
1463 LLVMValueRef temps_array
;
1464 LLVMTypeRef fptr_type
;
1466 indirect_index
= get_indirect_index(bld
,
1468 reg
->Register
.Index
,
1471 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1476 /* cast temps_array pointer to float* */
1477 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1478 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
, fptr_type
, "");
1480 /* Gather values from the temporary register array */
1481 res
= build_gather(&bld_base
->base
, temps_array
, index_vec
, NULL
);
1484 LLVMValueRef temp_ptr
;
1485 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
1486 res
= LLVMBuildLoad(builder
, temp_ptr
, "");
1489 if (stype
== TGSI_TYPE_SIGNED
|| stype
== TGSI_TYPE_UNSIGNED
) {
1490 struct lp_build_context
*bld_fetch
= stype_to_fetch(bld_base
, stype
);
1491 res
= LLVMBuildBitCast(builder
, res
, bld_fetch
->vec_type
, "");
1498 emit_fetch_system_value(
1499 struct lp_build_tgsi_context
* bld_base
,
1500 const struct tgsi_full_src_register
* reg
,
1501 enum tgsi_opcode_type stype
,
1504 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1505 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
1506 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
1507 LLVMBuilderRef builder
= gallivm
->builder
;
1509 enum tgsi_opcode_type atype
; // Actual type of the value
1511 assert(!reg
->Register
.Indirect
);
1513 switch (info
->system_value_semantic_name
[reg
->Register
.Index
]) {
1514 case TGSI_SEMANTIC_INSTANCEID
:
1515 res
= lp_build_broadcast_scalar(&bld_base
->uint_bld
, bld
->system_values
.instance_id
);
1516 atype
= TGSI_TYPE_UNSIGNED
;
1519 case TGSI_SEMANTIC_VERTEXID
:
1520 res
= bld
->system_values
.vertex_id
;
1521 atype
= TGSI_TYPE_UNSIGNED
;
1524 case TGSI_SEMANTIC_PRIMID
:
1525 res
= bld
->system_values
.prim_id
;
1526 atype
= TGSI_TYPE_UNSIGNED
;
1530 assert(!"unexpected semantic in emit_fetch_system_value");
1531 res
= bld_base
->base
.zero
;
1532 atype
= TGSI_TYPE_FLOAT
;
1536 if (atype
!= stype
) {
1537 if (stype
== TGSI_TYPE_FLOAT
) {
1538 res
= LLVMBuildBitCast(builder
, res
, bld_base
->base
.vec_type
, "");
1539 } else if (stype
== TGSI_TYPE_UNSIGNED
) {
1540 res
= LLVMBuildBitCast(builder
, res
, bld_base
->uint_bld
.vec_type
, "");
1541 } else if (stype
== TGSI_TYPE_SIGNED
) {
1542 res
= LLVMBuildBitCast(builder
, res
, bld_base
->int_bld
.vec_type
, "");
1550 * Register fetch with derivatives.
1554 struct lp_build_tgsi_soa_context
*bld
,
1563 /* TODO: use interpolation coeffs for inputs */
1566 *ddx
= lp_build_ddx(&bld
->bld_base
.base
, src
);
1569 *ddy
= lp_build_ddy(&bld
->bld_base
.base
, src
);
1577 emit_fetch_predicate(
1578 struct lp_build_tgsi_soa_context
*bld
,
1579 const struct tgsi_full_instruction
*inst
,
1582 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
1584 unsigned char swizzles
[4];
1585 LLVMValueRef unswizzled
[4] = {NULL
, NULL
, NULL
, NULL
};
1589 if (!inst
->Instruction
.Predicate
) {
1590 TGSI_FOR_EACH_CHANNEL( chan
) {
1596 swizzles
[0] = inst
->Predicate
.SwizzleX
;
1597 swizzles
[1] = inst
->Predicate
.SwizzleY
;
1598 swizzles
[2] = inst
->Predicate
.SwizzleZ
;
1599 swizzles
[3] = inst
->Predicate
.SwizzleW
;
1601 index
= inst
->Predicate
.Index
;
1602 assert(index
< LP_MAX_TGSI_PREDS
);
1604 TGSI_FOR_EACH_CHANNEL( chan
) {
1605 unsigned swizzle
= swizzles
[chan
];
1608 * Only fetch the predicate register channels that are actually listed
1611 if (!unswizzled
[swizzle
]) {
1612 value
= LLVMBuildLoad(builder
,
1613 bld
->preds
[index
][swizzle
], "");
1616 * Convert the value to an integer mask.
1618 * TODO: Short-circuit this comparison -- a D3D setp_xx instructions
1619 * is needlessly causing two comparisons due to storing the intermediate
1620 * result as float vector instead of an integer mask vector.
1622 value
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
1623 bld
->bld_base
.base
.type
,
1626 bld
->bld_base
.base
.zero
);
1627 if (inst
->Predicate
.Negate
) {
1628 value
= LLVMBuildNot(builder
, value
, "");
1631 unswizzled
[swizzle
] = value
;
1633 value
= unswizzled
[swizzle
];
1646 struct lp_build_tgsi_context
*bld_base
,
1647 const struct tgsi_full_instruction
*inst
,
1649 unsigned chan_index
,
1653 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1654 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1655 LLVMBuilderRef builder
= gallivm
->builder
;
1656 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
1657 struct lp_build_context
*float_bld
= &bld_base
->base
;
1658 struct lp_build_context
*int_bld
= &bld_base
->int_bld
;
1659 LLVMValueRef indirect_index
= NULL
;
1660 enum tgsi_opcode_type dtype
= tgsi_opcode_infer_dst_type(inst
->Instruction
.Opcode
);
1665 * It is always assumed to be float.
1667 switch( inst
->Instruction
.Saturate
) {
1671 case TGSI_SAT_ZERO_ONE
:
1672 assert(dtype
== TGSI_TYPE_FLOAT
||
1673 dtype
== TGSI_TYPE_UNTYPED
);
1674 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1675 value
= lp_build_clamp_zero_one_nanzero(float_bld
, value
);
1678 case TGSI_SAT_MINUS_PLUS_ONE
:
1679 assert(dtype
== TGSI_TYPE_FLOAT
||
1680 dtype
== TGSI_TYPE_UNTYPED
);
1681 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1682 /* This will give -1.0 for NaN which is probably not what we want. */
1683 value
= lp_build_max_ext(float_bld
, value
,
1684 lp_build_const_vec(gallivm
, float_bld
->type
, -1.0),
1685 GALLIVM_NAN_RETURN_OTHER_SECOND_NONNAN
);
1686 value
= lp_build_min(float_bld
, value
, float_bld
->one
);
1693 if (reg
->Register
.Indirect
) {
1694 indirect_index
= get_indirect_index(bld
,
1696 reg
->Register
.Index
,
1699 assert(reg
->Register
.Index
<=
1700 bld_base
->info
->file_max
[reg
->Register
.File
]);
1703 if (DEBUG_EXECUTION
) {
1704 emit_dump_reg(gallivm
, reg
->Register
.File
, reg
->Register
.Index
, chan_index
, value
);
1707 switch( reg
->Register
.File
) {
1708 case TGSI_FILE_OUTPUT
:
1709 /* Outputs are always stored as floats */
1710 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1712 if (reg
->Register
.Indirect
) {
1713 LLVMValueRef index_vec
; /* indexes into the output registers */
1714 LLVMValueRef outputs_array
;
1715 LLVMTypeRef fptr_type
;
1717 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1722 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1723 outputs_array
= LLVMBuildBitCast(builder
, bld
->outputs_array
, fptr_type
, "");
1725 /* Scatter store values into output registers */
1726 emit_mask_scatter(bld
, outputs_array
, index_vec
, value
,
1727 &bld
->exec_mask
, pred
);
1730 LLVMValueRef out_ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
,
1732 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
, out_ptr
);
1736 case TGSI_FILE_TEMPORARY
:
1737 /* Temporaries are always stored as floats */
1738 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1740 if (reg
->Register
.Indirect
) {
1741 LLVMValueRef index_vec
; /* indexes into the temp registers */
1742 LLVMValueRef temps_array
;
1743 LLVMTypeRef fptr_type
;
1745 index_vec
= get_soa_array_offsets(&bld_base
->uint_bld
,
1750 fptr_type
= LLVMPointerType(LLVMFloatTypeInContext(gallivm
->context
), 0);
1751 temps_array
= LLVMBuildBitCast(builder
, bld
->temps_array
, fptr_type
, "");
1753 /* Scatter store values into temp registers */
1754 emit_mask_scatter(bld
, temps_array
, index_vec
, value
,
1755 &bld
->exec_mask
, pred
);
1758 LLVMValueRef temp_ptr
;
1759 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, chan_index
);
1760 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
, temp_ptr
);
1764 case TGSI_FILE_ADDRESS
:
1765 assert(dtype
== TGSI_TYPE_SIGNED
);
1766 assert(LLVMTypeOf(value
) == int_bld
->vec_type
);
1767 value
= LLVMBuildBitCast(builder
, value
, int_bld
->vec_type
, "");
1768 lp_exec_mask_store(&bld
->exec_mask
, int_bld
, pred
, value
,
1769 bld
->addr
[reg
->Register
.Index
][chan_index
]);
1772 case TGSI_FILE_PREDICATE
:
1773 assert(LLVMTypeOf(value
) == float_bld
->vec_type
);
1774 value
= LLVMBuildBitCast(builder
, value
, float_bld
->vec_type
, "");
1775 lp_exec_mask_store(&bld
->exec_mask
, float_bld
, pred
, value
,
1776 bld
->preds
[reg
->Register
.Index
][chan_index
]);
1787 * Called at the beginning of the translation of each TGSI instruction, to
1788 * emit some debug code.
1792 struct lp_build_tgsi_context
* bld_base
,
1793 const struct tgsi_full_instruction
* inst
,
1794 const struct tgsi_opcode_info
* info
)
1797 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1799 if (DEBUG_EXECUTION
) {
1801 * Dump the TGSI instruction.
1804 struct gallivm_state
*gallivm
= bld_base
->base
.gallivm
;
1808 tgsi_dump_instruction_str(inst
, bld_base
->pc
, &buf
[2], sizeof buf
- 2);
1809 lp_build_printf(gallivm
, buf
);
1811 /* Dump the execution mask.
1813 if (bld
->exec_mask
.has_mask
) {
1814 lp_build_print_value(gallivm
, " mask = ", bld
->exec_mask
.exec_mask
);
1821 struct lp_build_tgsi_context
* bld_base
,
1822 const struct tgsi_full_instruction
* inst
,
1823 const struct tgsi_opcode_info
* info
,
1824 LLVMValueRef dst
[4])
1827 unsigned chan_index
;
1828 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
1831 LLVMValueRef pred
[TGSI_NUM_CHANNELS
];
1833 emit_fetch_predicate( bld
, inst
, pred
);
1835 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1836 emit_store_chan(bld_base
, inst
, 0, chan_index
, pred
[chan_index
], dst
[chan_index
]);
1842 tgsi_to_pipe_tex_target(unsigned tgsi_target
)
1844 switch (tgsi_target
) {
1845 case TGSI_TEXTURE_BUFFER
:
1847 case TGSI_TEXTURE_1D
:
1848 case TGSI_TEXTURE_SHADOW1D
:
1849 return PIPE_TEXTURE_1D
;
1850 case TGSI_TEXTURE_2D
:
1851 case TGSI_TEXTURE_SHADOW2D
:
1852 case TGSI_TEXTURE_2D_MSAA
:
1853 return PIPE_TEXTURE_2D
;
1854 case TGSI_TEXTURE_3D
:
1855 return PIPE_TEXTURE_3D
;
1856 case TGSI_TEXTURE_CUBE
:
1857 case TGSI_TEXTURE_SHADOWCUBE
:
1858 return PIPE_TEXTURE_CUBE
;
1859 case TGSI_TEXTURE_RECT
:
1860 case TGSI_TEXTURE_SHADOWRECT
:
1861 return PIPE_TEXTURE_RECT
;
1862 case TGSI_TEXTURE_1D_ARRAY
:
1863 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1864 return PIPE_TEXTURE_1D_ARRAY
;
1865 case TGSI_TEXTURE_2D_ARRAY
:
1866 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1867 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
1868 return PIPE_TEXTURE_2D_ARRAY
;
1869 case TGSI_TEXTURE_CUBE_ARRAY
:
1870 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
1871 return PIPE_TEXTURE_CUBE_ARRAY
;
1879 static enum lp_sampler_lod_property
1880 lp_build_lod_property(
1881 struct lp_build_tgsi_context
*bld_base
,
1882 const struct tgsi_full_instruction
*inst
,
1885 const struct tgsi_full_src_register
*reg
= &inst
->Src
[src_op
];
1886 enum lp_sampler_lod_property lod_property
;
1889 * Not much we can do here. We could try catching inputs declared
1890 * with constant interpolation but not sure it's worth it - since for
1891 * TEX opcodes as well as FETCH/LD the lod comes from same reg as
1892 * the coords, so it could only work for SAMPLE/TXQ/SVIEWINFO), just
1893 * like the constant/immediate recognition below.
1894 * What seems to be of more value would be to recognize temps holding
1895 * broadcasted scalars but no way we can do it.
1896 * Tried asking llvm but without any success (using LLVMIsConstant
1897 * even though this isn't exactly what we'd need), even as simple as
1898 * IMM[0] UINT32 (0,-1,0,0)
1899 * MOV TEMP[0] IMM[0].yyyy
1900 * SVIEWINFO TEMP[1], TEMP[0].xxxx, SVIEWINFO[0]
1902 * This means there's ZERO chance this will ever catch a scalar lod
1903 * with traditional tex opcodes as well as texel fetches, since the lod
1904 * comes from the same reg as coords (except some test shaders using
1905 * constant coords maybe).
1906 * There's at least hope for sample opcodes as well as size queries.
1908 if (reg
->Register
.File
== TGSI_FILE_CONSTANT
||
1909 reg
->Register
.File
== TGSI_FILE_IMMEDIATE
) {
1910 lod_property
= LP_SAMPLER_LOD_SCALAR
;
1912 else if (bld_base
->info
->processor
== TGSI_PROCESSOR_FRAGMENT
) {
1913 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
1914 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
1917 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
1921 /* never use scalar (per-quad) lod the results are just too wrong. */
1922 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
1924 return lod_property
;
1929 * High-level instruction translators.
1933 emit_tex( struct lp_build_tgsi_soa_context
*bld
,
1934 const struct tgsi_full_instruction
*inst
,
1935 enum lp_build_tex_modifier modifier
,
1936 LLVMValueRef
*texel
)
1939 LLVMValueRef lod_bias
, explicit_lod
;
1940 LLVMValueRef oow
= NULL
;
1941 LLVMValueRef coords
[5];
1942 LLVMValueRef offsets
[3] = { NULL
};
1943 struct lp_derivatives derivs
;
1944 struct lp_derivatives
*deriv_ptr
= NULL
;
1945 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
1946 unsigned num_derivs
, num_offsets
, i
;
1947 unsigned shadow_coord
= 0;
1948 unsigned layer_coord
= 0;
1950 if (!bld
->sampler
) {
1951 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
1952 for (i
= 0; i
< 4; i
++) {
1953 texel
[i
] = bld
->bld_base
.base
.undef
;
1958 switch (inst
->Texture
.Texture
) {
1959 case TGSI_TEXTURE_1D_ARRAY
:
1962 case TGSI_TEXTURE_1D
:
1966 case TGSI_TEXTURE_2D_ARRAY
:
1969 case TGSI_TEXTURE_2D
:
1970 case TGSI_TEXTURE_RECT
:
1974 case TGSI_TEXTURE_SHADOW1D_ARRAY
:
1977 case TGSI_TEXTURE_SHADOW1D
:
1982 case TGSI_TEXTURE_SHADOW2D_ARRAY
:
1988 case TGSI_TEXTURE_SHADOW2D
:
1989 case TGSI_TEXTURE_SHADOWRECT
:
1994 case TGSI_TEXTURE_CUBE
:
1998 case TGSI_TEXTURE_3D
:
2002 case TGSI_TEXTURE_SHADOWCUBE
:
2007 case TGSI_TEXTURE_CUBE_ARRAY
:
2008 case TGSI_TEXTURE_SHADOWCUBE_ARRAY
:
2009 case TGSI_TEXTURE_2D_MSAA
:
2010 case TGSI_TEXTURE_2D_ARRAY_MSAA
:
2016 /* Note lod and especially projected are illegal in a LOT of cases */
2017 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
||
2018 modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2019 LLVMValueRef lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2020 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
2022 explicit_lod
= NULL
;
2024 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2028 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2032 explicit_lod
= NULL
;
2035 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
) {
2036 oow
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2037 oow
= lp_build_rcp(&bld
->bld_base
.base
, oow
);
2040 for (i
= 0; i
< num_derivs
; i
++) {
2041 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2042 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2043 coords
[i
] = lp_build_mul(&bld
->bld_base
.base
, coords
[i
], oow
);
2045 for (i
= num_derivs
; i
< 5; i
++) {
2046 coords
[i
] = bld
->bld_base
.base
.undef
;
2049 /* Layer coord always goes into 3rd slot, except for cube map arrays */
2051 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2052 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2053 coords
[2] = lp_build_mul(&bld
->bld_base
.base
, coords
[2], oow
);
2055 /* Shadow coord occupies always 5th slot. */
2057 coords
[4] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, shadow_coord
);
2058 if (modifier
== LP_BLD_TEX_MODIFIER_PROJECTED
)
2059 coords
[4] = lp_build_mul(&bld
->bld_base
.base
, coords
[4], oow
);
2062 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
2064 for (dim
= 0; dim
< num_derivs
; ++dim
) {
2065 derivs
.ddx
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, dim
);
2066 derivs
.ddy
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, dim
);
2068 deriv_ptr
= &derivs
;
2069 unit
= inst
->Src
[3].Register
.Index
;
2071 * could also check all src regs if constant but I doubt such
2072 * cases exist in practice.
2074 if (bld
->bld_base
.info
->processor
== TGSI_PROCESSOR_FRAGMENT
) {
2075 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
2076 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2079 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
2083 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2086 unit
= inst
->Src
[1].Register
.Index
;
2089 /* some advanced gather instructions (txgo) would require 4 offsets */
2090 if (inst
->Texture
.NumOffsets
== 1) {
2092 for (dim
= 0; dim
< num_offsets
; dim
++) {
2093 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2097 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
2098 bld
->bld_base
.base
.gallivm
,
2099 bld
->bld_base
.base
.type
,
2105 lod_bias
, explicit_lod
, lod_property
,
2110 emit_sample(struct lp_build_tgsi_soa_context
*bld
,
2111 const struct tgsi_full_instruction
*inst
,
2112 enum lp_build_tex_modifier modifier
,
2114 LLVMValueRef
*texel
)
2116 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2117 unsigned texture_unit
, sampler_unit
;
2118 LLVMValueRef lod_bias
, explicit_lod
;
2119 LLVMValueRef coords
[5];
2120 LLVMValueRef offsets
[3] = { NULL
};
2121 struct lp_derivatives derivs
;
2122 struct lp_derivatives
*deriv_ptr
= NULL
;
2123 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
2125 unsigned num_offsets
, num_derivs
, i
;
2126 unsigned layer_coord
= 0;
2128 if (!bld
->sampler
) {
2129 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2130 for (i
= 0; i
< 4; i
++) {
2131 texel
[i
] = bld
->bld_base
.base
.undef
;
2137 * unlike old-style tex opcodes the texture/sampler indices
2138 * always come from src1 and src2 respectively.
2140 texture_unit
= inst
->Src
[1].Register
.Index
;
2141 sampler_unit
= inst
->Src
[2].Register
.Index
;
2144 * Note inst->Texture.Texture will contain the number of offsets,
2145 * however the target information is NOT there and comes from the
2146 * declared sampler views instead.
2148 switch (bld
->sv
[texture_unit
].Resource
) {
2149 case TGSI_TEXTURE_1D
:
2153 case TGSI_TEXTURE_1D_ARRAY
:
2158 case TGSI_TEXTURE_2D
:
2159 case TGSI_TEXTURE_RECT
:
2163 case TGSI_TEXTURE_2D_ARRAY
:
2168 case TGSI_TEXTURE_CUBE
:
2172 case TGSI_TEXTURE_3D
:
2176 case TGSI_TEXTURE_CUBE_ARRAY
:
2186 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
||
2187 modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2188 LLVMValueRef lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, 0);
2189 if (modifier
== LP_BLD_TEX_MODIFIER_LOD_BIAS
) {
2191 explicit_lod
= NULL
;
2193 else if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
) {
2197 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2199 else if (modifier
== LP_BLD_TEX_MODIFIER_LOD_ZERO
) {
2201 /* XXX might be better to explicitly pass the level zero information */
2202 explicit_lod
= lp_build_const_vec(gallivm
, bld
->bld_base
.base
.type
, 0.0F
);
2206 explicit_lod
= NULL
;
2209 for (i
= 0; i
< num_derivs
; i
++) {
2210 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2212 for (i
= num_derivs
; i
< 5; i
++) {
2213 coords
[i
] = bld
->bld_base
.base
.undef
;
2216 /* Layer coord always goes into 3rd slot, except for cube map arrays */
2218 if (layer_coord
== 3)
2219 coords
[3] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2221 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2223 /* Shadow coord occupies always 5th slot. */
2225 coords
[4] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, 0);
2228 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
2230 for (dim
= 0; dim
< num_derivs
; ++dim
) {
2231 derivs
.ddx
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 3, dim
);
2232 derivs
.ddy
[dim
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 4, dim
);
2234 deriv_ptr
= &derivs
;
2236 * could also check all src regs if constant but I doubt such
2237 * cases exist in practice.
2239 if (bld
->bld_base
.info
->processor
== TGSI_PROCESSOR_FRAGMENT
) {
2240 if (gallivm_debug
& GALLIVM_DEBUG_NO_QUAD_LOD
) {
2241 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2244 lod_property
= LP_SAMPLER_LOD_PER_QUAD
;
2248 lod_property
= LP_SAMPLER_LOD_PER_ELEMENT
;
2252 /* some advanced gather instructions (txgo) would require 4 offsets */
2253 if (inst
->Texture
.NumOffsets
== 1) {
2255 for (dim
= 0; dim
< num_offsets
; dim
++) {
2256 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2260 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
2261 bld
->bld_base
.base
.gallivm
,
2262 bld
->bld_base
.base
.type
,
2264 texture_unit
, sampler_unit
,
2268 lod_bias
, explicit_lod
, lod_property
,
2271 if (inst
->Src
[1].Register
.SwizzleX
!= PIPE_SWIZZLE_RED
||
2272 inst
->Src
[1].Register
.SwizzleY
!= PIPE_SWIZZLE_GREEN
||
2273 inst
->Src
[1].Register
.SwizzleZ
!= PIPE_SWIZZLE_BLUE
||
2274 inst
->Src
[1].Register
.SwizzleW
!= PIPE_SWIZZLE_ALPHA
) {
2275 unsigned char swizzles
[4];
2276 swizzles
[0] = inst
->Src
[1].Register
.SwizzleX
;
2277 swizzles
[1] = inst
->Src
[1].Register
.SwizzleY
;
2278 swizzles
[2] = inst
->Src
[1].Register
.SwizzleZ
;
2279 swizzles
[3] = inst
->Src
[1].Register
.SwizzleW
;
2281 lp_build_swizzle_soa_inplace(&bld
->bld_base
.base
, texel
, swizzles
);
2286 emit_fetch_texels( struct lp_build_tgsi_soa_context
*bld
,
2287 const struct tgsi_full_instruction
*inst
,
2288 LLVMValueRef
*texel
,
2291 unsigned unit
, target
;
2292 LLVMValueRef coord_undef
= LLVMGetUndef(bld
->bld_base
.base
.int_vec_type
);
2293 LLVMValueRef explicit_lod
= NULL
;
2294 LLVMValueRef coords
[3];
2295 LLVMValueRef offsets
[3] = { NULL
};
2296 enum lp_sampler_lod_property lod_property
= LP_SAMPLER_LOD_SCALAR
;
2298 unsigned layer_coord
= 0;
2300 if (!bld
->sampler
) {
2301 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
2302 for (i
= 0; i
< 4; i
++) {
2303 texel
[i
] = coord_undef
;
2308 unit
= inst
->Src
[1].Register
.Index
;
2311 target
= bld
->sv
[unit
].Resource
;
2314 target
= inst
->Texture
.Texture
;
2318 case TGSI_TEXTURE_1D
:
2319 case TGSI_TEXTURE_BUFFER
:
2322 case TGSI_TEXTURE_1D_ARRAY
:
2326 case TGSI_TEXTURE_2D
:
2327 case TGSI_TEXTURE_RECT
:
2330 case TGSI_TEXTURE_2D_ARRAY
:
2334 case TGSI_TEXTURE_3D
:
2342 /* always have lod except for buffers ? */
2343 if (target
!= TGSI_TEXTURE_BUFFER
) {
2344 explicit_lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 3);
2345 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2348 for (i
= 0; i
< dims
; i
++) {
2349 coords
[i
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, i
);
2351 for (i
= dims
; i
< 3; i
++) {
2352 coords
[i
] = coord_undef
;
2355 coords
[2] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, layer_coord
);
2357 if (inst
->Texture
.NumOffsets
== 1) {
2359 for (dim
= 0; dim
< dims
; dim
++) {
2360 offsets
[dim
] = lp_build_emit_fetch_texoffset(&bld
->bld_base
, inst
, 0, dim
);
2364 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
2365 bld
->bld_base
.base
.gallivm
,
2366 bld
->bld_base
.base
.type
,
2372 NULL
, explicit_lod
, lod_property
,
2376 (inst
->Src
[1].Register
.SwizzleX
!= PIPE_SWIZZLE_RED
||
2377 inst
->Src
[1].Register
.SwizzleY
!= PIPE_SWIZZLE_GREEN
||
2378 inst
->Src
[1].Register
.SwizzleZ
!= PIPE_SWIZZLE_BLUE
||
2379 inst
->Src
[1].Register
.SwizzleW
!= PIPE_SWIZZLE_ALPHA
)) {
2380 unsigned char swizzles
[4];
2381 swizzles
[0] = inst
->Src
[1].Register
.SwizzleX
;
2382 swizzles
[1] = inst
->Src
[1].Register
.SwizzleY
;
2383 swizzles
[2] = inst
->Src
[1].Register
.SwizzleZ
;
2384 swizzles
[3] = inst
->Src
[1].Register
.SwizzleW
;
2386 lp_build_swizzle_soa_inplace(&bld
->bld_base
.base
, texel
, swizzles
);
2391 emit_size_query( struct lp_build_tgsi_soa_context
*bld
,
2392 const struct tgsi_full_instruction
*inst
,
2393 LLVMValueRef
*sizes_out
,
2394 boolean is_sviewinfo
)
2396 LLVMValueRef explicit_lod
;
2397 enum lp_sampler_lod_property lod_property
;
2400 unsigned unit
= inst
->Src
[1].Register
.Index
;
2401 unsigned target
, pipe_target
;
2404 target
= bld
->sv
[unit
].Resource
;
2407 target
= inst
->Texture
.Texture
;
2410 case TGSI_TEXTURE_BUFFER
:
2411 case TGSI_TEXTURE_RECT
:
2412 case TGSI_TEXTURE_SHADOWRECT
:
2420 if (!bld
->sampler
) {
2421 _debug_printf("warning: found texture query instruction but no sampler generator supplied\n");
2422 for (i
= 0; i
< 4; i
++)
2423 sizes_out
[i
] = bld
->bld_base
.int_bld
.undef
;
2428 explicit_lod
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, 0);
2429 lod_property
= lp_build_lod_property(&bld
->bld_base
, inst
, 0);
2432 explicit_lod
= NULL
;
2433 lod_property
= LP_SAMPLER_LOD_SCALAR
;
2437 pipe_target
= tgsi_to_pipe_tex_target(target
);
2439 bld
->sampler
->emit_size_query(bld
->sampler
,
2440 bld
->bld_base
.base
.gallivm
,
2441 bld
->bld_base
.int_bld
.type
,
2450 near_end_of_shader(struct lp_build_tgsi_soa_context
*bld
,
2455 for (i
= 0; i
< 5; i
++) {
2458 if (pc
+ i
>= bld
->bld_base
.info
->num_instructions
)
2461 opcode
= bld
->bld_base
.instructions
[pc
+ i
].Instruction
.Opcode
;
2463 if (opcode
== TGSI_OPCODE_END
)
2466 if (opcode
== TGSI_OPCODE_TEX
||
2467 opcode
== TGSI_OPCODE_TXP
||
2468 opcode
== TGSI_OPCODE_TXD
||
2469 opcode
== TGSI_OPCODE_TXB
||
2470 opcode
== TGSI_OPCODE_TXL
||
2471 opcode
== TGSI_OPCODE_TXF
||
2472 opcode
== TGSI_OPCODE_TXQ
||
2473 opcode
== TGSI_OPCODE_CAL
||
2474 opcode
== TGSI_OPCODE_CALLNZ
||
2475 opcode
== TGSI_OPCODE_IF
||
2476 opcode
== TGSI_OPCODE_UIF
||
2477 opcode
== TGSI_OPCODE_BGNLOOP
||
2478 opcode
== TGSI_OPCODE_SWITCH
)
2488 * Kill fragment if any of the src register values are negative.
2492 struct lp_build_tgsi_soa_context
*bld
,
2493 const struct tgsi_full_instruction
*inst
,
2496 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2497 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
2498 LLVMValueRef terms
[TGSI_NUM_CHANNELS
];
2500 unsigned chan_index
;
2502 memset(&terms
, 0, sizeof terms
);
2504 TGSI_FOR_EACH_CHANNEL( chan_index
) {
2507 /* Unswizzle channel */
2508 swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
2510 /* Check if the component has not been already tested. */
2511 assert(swizzle
< TGSI_NUM_CHANNELS
);
2512 if( !terms
[swizzle
] )
2513 /* TODO: change the comparison operator instead of setting the sign */
2514 terms
[swizzle
] = lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, chan_index
);
2518 TGSI_FOR_EACH_CHANNEL( chan_index
) {
2519 if(terms
[chan_index
]) {
2520 LLVMValueRef chan_mask
;
2523 * If term < 0 then mask = 0 else mask = ~0.
2525 chan_mask
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, terms
[chan_index
], bld
->bld_base
.base
.zero
);
2528 mask
= LLVMBuildAnd(builder
, mask
, chan_mask
, "");
2534 if (bld
->exec_mask
.has_mask
) {
2535 LLVMValueRef invmask
;
2536 invmask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
2537 mask
= LLVMBuildOr(builder
, mask
, invmask
, "");
2540 lp_build_mask_update(bld
->mask
, mask
);
2541 if (!near_end_of_shader(bld
, pc
))
2542 lp_build_mask_check(bld
->mask
);
2547 * Unconditional fragment kill.
2548 * The only predication is the execution mask which will apply if
2549 * we're inside a loop or conditional.
2552 emit_kill(struct lp_build_tgsi_soa_context
*bld
,
2555 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
2558 /* For those channels which are "alive", disable fragment shader
2561 if (bld
->exec_mask
.has_mask
) {
2562 mask
= LLVMBuildNot(builder
, bld
->exec_mask
.exec_mask
, "kilp");
2565 LLVMValueRef zero
= LLVMConstNull(bld
->bld_base
.base
.int_vec_type
);
2569 lp_build_mask_update(bld
->mask
, mask
);
2571 if (!near_end_of_shader(bld
, pc
))
2572 lp_build_mask_check(bld
->mask
);
2577 * Emit code which will dump the value of all the temporary registers
2581 emit_dump_file(struct lp_build_tgsi_soa_context
*bld
,
2584 const struct tgsi_shader_info
*info
= bld
->bld_base
.info
;
2585 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2586 LLVMBuilderRef builder
= gallivm
->builder
;
2587 LLVMValueRef reg_ptr
;
2589 int max_index
= info
->file_max
[file
];
2592 * Some register files, particularly constants, can be very large,
2593 * and dumping everything could make this unusably slow.
2595 max_index
= MIN2(max_index
, 32);
2597 for (index
= 0; index
<= max_index
; index
++) {
2602 if (index
< 8 * sizeof(unsigned) &&
2603 (info
->file_mask
[file
] & (1 << index
)) == 0) {
2604 /* This was not declared.*/
2608 if (file
== TGSI_FILE_INPUT
) {
2609 mask
= info
->input_usage_mask
[index
];
2611 mask
= TGSI_WRITEMASK_XYZW
;
2614 for (chan
= 0; chan
< 4; chan
++) {
2615 if ((mask
& (1 << chan
)) == 0) {
2616 /* This channel is not used.*/
2620 if (file
== TGSI_FILE_CONSTANT
) {
2621 struct tgsi_full_src_register reg
;
2622 memset(®
, 0, sizeof reg
);
2623 reg
.Register
.File
= file
;
2624 reg
.Register
.Index
= index
;
2625 reg
.Register
.SwizzleX
= 0;
2626 reg
.Register
.SwizzleY
= 1;
2627 reg
.Register
.SwizzleZ
= 2;
2628 reg
.Register
.SwizzleW
= 3;
2630 res
= bld
->bld_base
.emit_fetch_funcs
[file
](&bld
->bld_base
, ®
, TGSI_TYPE_FLOAT
, chan
);
2634 } else if (file
== TGSI_FILE_INPUT
) {
2635 res
= bld
->inputs
[index
][chan
];
2639 } else if (file
== TGSI_FILE_TEMPORARY
) {
2640 reg_ptr
= lp_get_temp_ptr_soa(bld
, index
, chan
);
2642 res
= LLVMBuildLoad(builder
, reg_ptr
, "");
2643 } else if (file
== TGSI_FILE_OUTPUT
) {
2644 reg_ptr
= lp_get_output_ptr(bld
, index
, chan
);
2646 res
= LLVMBuildLoad(builder
, reg_ptr
, "");
2652 emit_dump_reg(gallivm
, file
, index
, chan
, res
);
2660 lp_emit_declaration_soa(
2661 struct lp_build_tgsi_context
*bld_base
,
2662 const struct tgsi_full_declaration
*decl
)
2664 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
2665 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2666 LLVMTypeRef vec_type
= bld
->bld_base
.base
.vec_type
;
2667 const unsigned first
= decl
->Range
.First
;
2668 const unsigned last
= decl
->Range
.Last
;
2671 for (idx
= first
; idx
<= last
; ++idx
) {
2672 assert(last
<= bld
->bld_base
.info
->file_max
[decl
->Declaration
.File
]);
2673 switch (decl
->Declaration
.File
) {
2674 case TGSI_FILE_TEMPORARY
:
2675 if (!(bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
))) {
2676 assert(idx
< LP_MAX_INLINED_TEMPS
);
2677 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2678 bld
->temps
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
, "temp");
2682 case TGSI_FILE_OUTPUT
:
2683 if (!(bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
))) {
2684 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2685 bld
->outputs
[idx
][i
] = lp_build_alloca(gallivm
,
2686 vec_type
, "output");
2690 case TGSI_FILE_ADDRESS
:
2691 /* ADDR registers are only allocated with an integer LLVM IR type,
2692 * as they are guaranteed to always have integers.
2693 * XXX: Not sure if this exception is worthwhile (or the whole idea of
2694 * an ADDR register for that matter).
2696 assert(idx
< LP_MAX_TGSI_ADDRS
);
2697 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2698 bld
->addr
[idx
][i
] = lp_build_alloca(gallivm
, bld_base
->base
.int_vec_type
, "addr");
2701 case TGSI_FILE_PREDICATE
:
2702 assert(idx
< LP_MAX_TGSI_PREDS
);
2703 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++)
2704 bld
->preds
[idx
][i
] = lp_build_alloca(gallivm
, vec_type
,
2708 case TGSI_FILE_SAMPLER_VIEW
:
2710 * The target stored here MUST match whatever there actually
2711 * is in the set sampler views (what about return type?).
2713 assert(idx
< PIPE_MAX_SHADER_SAMPLER_VIEWS
);
2714 bld
->sv
[idx
] = decl
->SamplerView
;
2718 /* don't need to declare other vars */
2725 void lp_emit_immediate_soa(
2726 struct lp_build_tgsi_context
*bld_base
,
2727 const struct tgsi_full_immediate
*imm
)
2729 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
2730 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
2732 /* simply copy the immediate values into the next immediates[] slot */
2734 const uint size
= imm
->Immediate
.NrTokens
- 1;
2736 assert(bld
->num_immediates
< LP_MAX_TGSI_IMMEDIATES
);
2737 switch (imm
->Immediate
.DataType
) {
2738 case TGSI_IMM_FLOAT32
:
2739 for( i
= 0; i
< size
; ++i
)
2740 bld
->immediates
[bld
->num_immediates
][i
] =
2741 lp_build_const_vec(gallivm
, bld_base
->base
.type
, imm
->u
[i
].Float
);
2744 case TGSI_IMM_UINT32
:
2745 for( i
= 0; i
< size
; ++i
) {
2746 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->uint_bld
.type
, imm
->u
[i
].Uint
);
2747 bld
->immediates
[bld
->num_immediates
][i
] =
2748 LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
2752 case TGSI_IMM_INT32
:
2753 for( i
= 0; i
< size
; ++i
) {
2754 LLVMValueRef tmp
= lp_build_const_vec(gallivm
, bld_base
->int_bld
.type
, imm
->u
[i
].Int
);
2755 bld
->immediates
[bld
->num_immediates
][i
] =
2756 LLVMConstBitCast(tmp
, bld_base
->base
.vec_type
);
2761 for( i
= size
; i
< 4; ++i
)
2762 bld
->immediates
[bld
->num_immediates
][i
] = bld_base
->base
.undef
;
2764 if (bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
)) {
2765 unsigned index
= bld
->num_immediates
;
2766 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
2767 LLVMBuilderRef builder
= gallivm
->builder
;
2768 for (i
= 0; i
< 4; ++i
) {
2769 LLVMValueRef lindex
= lp_build_const_int32(
2770 bld
->bld_base
.base
.gallivm
, index
* 4 + i
);
2771 LLVMValueRef imm_ptr
= LLVMBuildGEP(builder
,
2772 bld
->imms_array
, &lindex
, 1, "");
2773 LLVMBuildStore(builder
,
2774 bld
->immediates
[index
][i
],
2779 bld
->num_immediates
++;
2784 const struct lp_build_tgsi_action
* action
,
2785 struct lp_build_tgsi_context
* bld_base
,
2786 struct lp_build_emit_data
* emit_data
)
2788 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2790 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
,
2791 &emit_data
->output
[emit_data
->chan
], NULL
);
2796 const struct lp_build_tgsi_action
* action
,
2797 struct lp_build_tgsi_context
* bld_base
,
2798 struct lp_build_emit_data
* emit_data
)
2800 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2802 emit_fetch_deriv(bld
, emit_data
->args
[0], NULL
, NULL
,
2803 &emit_data
->output
[emit_data
->chan
]);
2808 const struct lp_build_tgsi_action
* action
,
2809 struct lp_build_tgsi_context
* bld_base
,
2810 struct lp_build_emit_data
* emit_data
)
2812 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2814 emit_kill(bld
, bld_base
->pc
- 1);
2819 const struct lp_build_tgsi_action
* action
,
2820 struct lp_build_tgsi_context
* bld_base
,
2821 struct lp_build_emit_data
* emit_data
)
2823 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2825 emit_kill_if(bld
, emit_data
->inst
, bld_base
->pc
- 1);
2830 const struct lp_build_tgsi_action
* action
,
2831 struct lp_build_tgsi_context
* bld_base
,
2832 struct lp_build_emit_data
* emit_data
)
2834 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2836 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
, emit_data
->output
);
2841 const struct lp_build_tgsi_action
* action
,
2842 struct lp_build_tgsi_context
* bld_base
,
2843 struct lp_build_emit_data
* emit_data
)
2845 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2847 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
2853 const struct lp_build_tgsi_action
* action
,
2854 struct lp_build_tgsi_context
* bld_base
,
2855 struct lp_build_emit_data
* emit_data
)
2857 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2859 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
2865 const struct lp_build_tgsi_action
* action
,
2866 struct lp_build_tgsi_context
* bld_base
,
2867 struct lp_build_emit_data
* emit_data
)
2869 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2871 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
2877 const struct lp_build_tgsi_action
* action
,
2878 struct lp_build_tgsi_context
* bld_base
,
2879 struct lp_build_emit_data
* emit_data
)
2881 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2883 emit_tex(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_PROJECTED
,
2889 const struct lp_build_tgsi_action
* action
,
2890 struct lp_build_tgsi_context
* bld_base
,
2891 struct lp_build_emit_data
* emit_data
)
2893 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2895 emit_size_query(bld
, emit_data
->inst
, emit_data
->output
, FALSE
);
2900 const struct lp_build_tgsi_action
* action
,
2901 struct lp_build_tgsi_context
* bld_base
,
2902 struct lp_build_emit_data
* emit_data
)
2904 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2906 emit_fetch_texels(bld
, emit_data
->inst
, emit_data
->output
, FALSE
);
2911 const struct lp_build_tgsi_action
* action
,
2912 struct lp_build_tgsi_context
* bld_base
,
2913 struct lp_build_emit_data
* emit_data
)
2915 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2917 emit_fetch_texels(bld
, emit_data
->inst
, emit_data
->output
, TRUE
);
2922 const struct lp_build_tgsi_action
* action
,
2923 struct lp_build_tgsi_context
* bld_base
,
2924 struct lp_build_emit_data
* emit_data
)
2926 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2928 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
2929 FALSE
, emit_data
->output
);
2934 const struct lp_build_tgsi_action
* action
,
2935 struct lp_build_tgsi_context
* bld_base
,
2936 struct lp_build_emit_data
* emit_data
)
2938 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2940 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
,
2941 FALSE
, emit_data
->output
);
2946 const struct lp_build_tgsi_action
* action
,
2947 struct lp_build_tgsi_context
* bld_base
,
2948 struct lp_build_emit_data
* emit_data
)
2950 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2952 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_NONE
,
2953 TRUE
, emit_data
->output
);
2958 const struct lp_build_tgsi_action
* action
,
2959 struct lp_build_tgsi_context
* bld_base
,
2960 struct lp_build_emit_data
* emit_data
)
2962 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2964 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_LOD_ZERO
,
2965 TRUE
, emit_data
->output
);
2970 const struct lp_build_tgsi_action
* action
,
2971 struct lp_build_tgsi_context
* bld_base
,
2972 struct lp_build_emit_data
* emit_data
)
2974 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2976 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
,
2977 FALSE
, emit_data
->output
);
2982 const struct lp_build_tgsi_action
* action
,
2983 struct lp_build_tgsi_context
* bld_base
,
2984 struct lp_build_emit_data
* emit_data
)
2986 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
2988 emit_sample(bld
, emit_data
->inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
,
2989 FALSE
, emit_data
->output
);
2994 const struct lp_build_tgsi_action
* action
,
2995 struct lp_build_tgsi_context
* bld_base
,
2996 struct lp_build_emit_data
* emit_data
)
2998 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3000 emit_size_query(bld
, emit_data
->inst
, emit_data
->output
, TRUE
);
3004 mask_vec(struct lp_build_tgsi_context
*bld_base
)
3006 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3007 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3008 struct lp_exec_mask
*exec_mask
= &bld
->exec_mask
;
3010 if (!exec_mask
->has_mask
) {
3011 return lp_build_mask_value(bld
->mask
);
3013 return LLVMBuildAnd(builder
, lp_build_mask_value(bld
->mask
),
3014 exec_mask
->exec_mask
, "");
3018 increment_vec_ptr_by_mask(struct lp_build_tgsi_context
* bld_base
,
3022 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3023 LLVMValueRef current_vec
= LLVMBuildLoad(builder
, ptr
, "");
3025 current_vec
= LLVMBuildSub(builder
, current_vec
, mask
, "");
3027 LLVMBuildStore(builder
, current_vec
, ptr
);
3031 clear_uint_vec_ptr_from_mask(struct lp_build_tgsi_context
* bld_base
,
3035 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3036 LLVMValueRef current_vec
= LLVMBuildLoad(builder
, ptr
, "");
3038 current_vec
= lp_build_select(&bld_base
->uint_bld
,
3040 bld_base
->uint_bld
.zero
,
3043 LLVMBuildStore(builder
, current_vec
, ptr
);
3047 clamp_mask_to_max_output_vertices(struct lp_build_tgsi_soa_context
* bld
,
3048 LLVMValueRef current_mask_vec
,
3049 LLVMValueRef total_emitted_vertices_vec
)
3051 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3052 struct lp_build_context
*int_bld
= &bld
->bld_base
.int_bld
;
3053 LLVMValueRef max_mask
= lp_build_cmp(int_bld
, PIPE_FUNC_LESS
,
3054 total_emitted_vertices_vec
,
3055 bld
->max_output_vertices_vec
);
3057 return LLVMBuildAnd(builder
, current_mask_vec
, max_mask
, "");
3062 const struct lp_build_tgsi_action
* action
,
3063 struct lp_build_tgsi_context
* bld_base
,
3064 struct lp_build_emit_data
* emit_data
)
3066 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3067 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3069 if (bld
->gs_iface
->emit_vertex
) {
3070 LLVMValueRef mask
= mask_vec(bld_base
);
3071 LLVMValueRef total_emitted_vertices_vec
=
3072 LLVMBuildLoad(builder
, bld
->total_emitted_vertices_vec_ptr
, "");
3073 mask
= clamp_mask_to_max_output_vertices(bld
, mask
,
3074 total_emitted_vertices_vec
);
3075 gather_outputs(bld
);
3076 bld
->gs_iface
->emit_vertex(bld
->gs_iface
, &bld
->bld_base
,
3078 total_emitted_vertices_vec
);
3079 increment_vec_ptr_by_mask(bld_base
, bld
->emitted_vertices_vec_ptr
,
3081 increment_vec_ptr_by_mask(bld_base
, bld
->total_emitted_vertices_vec_ptr
,
3084 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3085 " +++ emit vertex masked ones = ",
3087 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3088 " +++ emit vertex emitted = ",
3089 total_emitted_vertices_vec
);
3096 end_primitive_masked(struct lp_build_tgsi_context
* bld_base
,
3099 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3100 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
3102 if (bld
->gs_iface
->end_primitive
) {
3103 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3104 LLVMValueRef emitted_vertices_vec
=
3105 LLVMBuildLoad(builder
, bld
->emitted_vertices_vec_ptr
, "");
3106 LLVMValueRef emitted_prims_vec
=
3107 LLVMBuildLoad(builder
, bld
->emitted_prims_vec_ptr
, "");
3109 LLVMValueRef emitted_mask
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3110 emitted_vertices_vec
,
3112 /* We need to combine the current execution mask with the mask
3113 telling us which, if any, execution slots actually have
3114 unemitted primitives, this way we make sure that end_primitives
3115 executes only on the paths that have unflushed vertices */
3116 mask
= LLVMBuildAnd(builder
, mask
, emitted_mask
, "");
3118 bld
->gs_iface
->end_primitive(bld
->gs_iface
, &bld
->bld_base
,
3119 emitted_vertices_vec
,
3123 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3124 " +++ end prim masked ones = ",
3126 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3127 " +++ end prim emitted verts1 = ",
3128 emitted_vertices_vec
);
3129 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3130 " +++ end prim emitted prims1 = ",
3131 LLVMBuildLoad(builder
,
3132 bld
->emitted_prims_vec_ptr
, ""));
3134 increment_vec_ptr_by_mask(bld_base
, bld
->emitted_prims_vec_ptr
,
3136 clear_uint_vec_ptr_from_mask(bld_base
, bld
->emitted_vertices_vec_ptr
,
3139 lp_build_print_value(bld
->bld_base
.base
.gallivm
,
3140 " +++ end prim emitted verts2 = ",
3141 LLVMBuildLoad(builder
,
3142 bld
->emitted_vertices_vec_ptr
, ""));
3150 const struct lp_build_tgsi_action
* action
,
3151 struct lp_build_tgsi_context
* bld_base
,
3152 struct lp_build_emit_data
* emit_data
)
3154 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3156 if (bld
->gs_iface
->end_primitive
) {
3157 LLVMValueRef mask
= mask_vec(bld_base
);
3158 end_primitive_masked(bld_base
, mask
);
3164 const struct lp_build_tgsi_action
* action
,
3165 struct lp_build_tgsi_context
* bld_base
,
3166 struct lp_build_emit_data
* emit_data
)
3168 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3170 lp_exec_mask_call(&bld
->exec_mask
, emit_data
->inst
->Label
.Label
,
3176 const struct lp_build_tgsi_action
* action
,
3177 struct lp_build_tgsi_context
* bld_base
,
3178 struct lp_build_emit_data
* emit_data
)
3180 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3182 lp_exec_mask_ret(&bld
->exec_mask
, &bld_base
->pc
);
3187 const struct lp_build_tgsi_action
* action
,
3188 struct lp_build_tgsi_context
* bld_base
,
3189 struct lp_build_emit_data
* emit_data
)
3191 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3193 lp_exec_break(&bld
->exec_mask
, bld_base
);
3198 const struct lp_build_tgsi_action
* action
,
3199 struct lp_build_tgsi_context
* bld_base
,
3200 struct lp_build_emit_data
* emit_data
)
3202 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3203 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3204 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3205 LLVMValueRef unsigned_cond
=
3206 LLVMBuildBitCast(builder
, emit_data
->args
[0], uint_bld
->vec_type
, "");
3207 LLVMValueRef cond
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3211 lp_exec_break_condition(&bld
->exec_mask
, cond
);
3216 const struct lp_build_tgsi_action
* action
,
3217 struct lp_build_tgsi_context
* bld_base
,
3218 struct lp_build_emit_data
* emit_data
)
3221 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3223 tmp
= lp_build_cmp(&bld_base
->base
, PIPE_FUNC_NOTEQUAL
,
3224 emit_data
->args
[0], bld
->bld_base
.base
.zero
);
3225 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
3230 const struct lp_build_tgsi_action
* action
,
3231 struct lp_build_tgsi_context
* bld_base
,
3232 struct lp_build_emit_data
* emit_data
)
3235 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3236 struct lp_build_context
*uint_bld
= &bld_base
->uint_bld
;
3238 tmp
= lp_build_cmp(uint_bld
, PIPE_FUNC_NOTEQUAL
,
3239 emit_data
->args
[0], uint_bld
->zero
);
3240 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp
);
3245 const struct lp_build_tgsi_action
* action
,
3246 struct lp_build_tgsi_context
* bld_base
,
3247 struct lp_build_emit_data
* emit_data
)
3249 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3251 lp_exec_case(&bld
->exec_mask
, emit_data
->args
[0]);
3256 const struct lp_build_tgsi_action
* action
,
3257 struct lp_build_tgsi_context
* bld_base
,
3258 struct lp_build_emit_data
* emit_data
)
3260 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3262 lp_exec_default(&bld
->exec_mask
, bld_base
);
3267 const struct lp_build_tgsi_action
* action
,
3268 struct lp_build_tgsi_context
* bld_base
,
3269 struct lp_build_emit_data
* emit_data
)
3271 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3273 lp_exec_switch(&bld
->exec_mask
, emit_data
->args
[0]);
3278 const struct lp_build_tgsi_action
* action
,
3279 struct lp_build_tgsi_context
* bld_base
,
3280 struct lp_build_emit_data
* emit_data
)
3282 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3284 lp_exec_endswitch(&bld
->exec_mask
, bld_base
);
3289 const struct lp_build_tgsi_action
* action
,
3290 struct lp_build_tgsi_context
* bld_base
,
3291 struct lp_build_emit_data
* emit_data
)
3293 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3295 lp_exec_bgnloop(&bld
->exec_mask
);
3300 const struct lp_build_tgsi_action
* action
,
3301 struct lp_build_tgsi_context
* bld_base
,
3302 struct lp_build_emit_data
* emit_data
)
3304 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3306 lp_exec_mask_bgnsub(&bld
->exec_mask
);
3311 const struct lp_build_tgsi_action
* action
,
3312 struct lp_build_tgsi_context
* bld_base
,
3313 struct lp_build_emit_data
* emit_data
)
3315 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3317 lp_exec_mask_cond_invert(&bld
->exec_mask
);
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_cond_pop(&bld
->exec_mask
);
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_endloop(bld_base
->base
.gallivm
, &bld
->exec_mask
);
3344 const struct lp_build_tgsi_action
* action
,
3345 struct lp_build_tgsi_context
* bld_base
,
3346 struct lp_build_emit_data
* emit_data
)
3348 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3350 lp_exec_mask_endsub(&bld
->exec_mask
, &bld_base
->pc
);
3355 const struct lp_build_tgsi_action
* action
,
3356 struct lp_build_tgsi_context
* bld_base
,
3357 struct lp_build_emit_data
* emit_data
)
3359 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3361 lp_exec_continue(&bld
->exec_mask
);
3364 /* XXX: Refactor and move it to lp_bld_tgsi_action.c
3366 * XXX: What do the comments about xmm registers mean? Maybe they are left over
3367 * from old code, but there is no garauntee that LLVM will use those registers
3370 * XXX: There should be no calls to lp_build_emit_fetch in this function. This
3371 * should be handled by the emit_data->fetch_args function. */
3374 const struct lp_build_tgsi_action
* action
,
3375 struct lp_build_tgsi_context
* bld_base
,
3376 struct lp_build_emit_data
* emit_data
)
3378 LLVMValueRef tmp0
, tmp1
;
3379 LLVMValueRef tmp4
= NULL
;
3380 LLVMValueRef tmp5
= NULL
;
3381 LLVMValueRef tmp6
= NULL
;
3382 LLVMValueRef tmp7
= NULL
;
3383 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3385 uint dims
= (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_NRM
) ? 3 : 4;
3387 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) ||
3388 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
) ||
3389 TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
) ||
3390 (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 4)) {
3392 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
3395 /* xmm0 = src.x * src.x */
3396 tmp0
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_X
);
3397 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
3400 tmp0
= lp_build_mul( &bld
->bld_base
.base
, tmp0
, tmp0
);
3403 /* xmm0 = xmm0 + src.y * src.y */
3404 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Y
);
3405 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
3408 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
3409 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
3412 /* xmm0 = xmm0 + src.z * src.z */
3413 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_Z
);
3414 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
3417 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
3418 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
3422 /* xmm0 = xmm0 + src.w * src.w */
3423 tmp1
= lp_build_emit_fetch(&bld
->bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
3424 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
)) {
3427 tmp1
= lp_build_mul( &bld
->bld_base
.base
, tmp1
, tmp1
);
3428 tmp0
= lp_build_add( &bld
->bld_base
.base
, tmp0
, tmp1
);
3430 /* xmm1 = 1 / sqrt(xmm0) */
3431 tmp1
= lp_build_rsqrt( &bld
->bld_base
.base
, tmp0
);
3432 /* dst.x = xmm1 * src.x */
3433 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
)) {
3434 emit_data
->output
[TGSI_CHAN_X
] = lp_build_mul( &bld
->bld_base
.base
, tmp4
, tmp1
);
3436 /* dst.y = xmm1 * src.y */
3437 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Y
)) {
3438 emit_data
->output
[TGSI_CHAN_Y
] = lp_build_mul( &bld
->bld_base
.base
, tmp5
, tmp1
);
3441 /* dst.z = xmm1 * src.z */
3442 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_Z
)) {
3443 emit_data
->output
[TGSI_CHAN_Z
] = lp_build_mul( &bld
->bld_base
.base
, tmp6
, tmp1
);
3445 /* dst.w = xmm1 * src.w */
3446 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_X
) && dims
== 4) {
3447 emit_data
->output
[TGSI_CHAN_W
] = lp_build_mul( &bld
->bld_base
.base
, tmp7
, tmp1
);
3452 if (TGSI_IS_DST0_CHANNEL_ENABLED(emit_data
->inst
, TGSI_CHAN_W
) && dims
== 3) {
3453 emit_data
->output
[TGSI_CHAN_W
] = bld
->bld_base
.base
.one
;
3457 static void emit_prologue(struct lp_build_tgsi_context
* bld_base
)
3459 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3460 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
3462 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
3463 LLVMValueRef array_size
=
3464 lp_build_const_int32(gallivm
,
3465 bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] * 4 + 4);
3466 bld
->temps_array
= lp_build_array_alloca(gallivm
,
3467 bld_base
->base
.vec_type
, array_size
,
3471 if (bld
->indirect_files
& (1 << TGSI_FILE_OUTPUT
)) {
3472 LLVMValueRef array_size
=
3473 lp_build_const_int32(gallivm
,
3474 bld_base
->info
->file_max
[TGSI_FILE_OUTPUT
] * 4 + 4);
3475 bld
->outputs_array
= lp_build_array_alloca(gallivm
,
3476 bld_base
->base
.vec_type
, array_size
,
3480 if (bld
->indirect_files
& (1 << TGSI_FILE_IMMEDIATE
)) {
3481 LLVMValueRef array_size
=
3482 lp_build_const_int32(gallivm
,
3483 bld_base
->info
->file_max
[TGSI_FILE_IMMEDIATE
] * 4 + 4);
3484 bld
->imms_array
= lp_build_array_alloca(gallivm
,
3485 bld_base
->base
.vec_type
, array_size
,
3489 /* If we have indirect addressing in inputs we need to copy them into
3490 * our alloca array to be able to iterate over them */
3491 if (bld
->indirect_files
& (1 << TGSI_FILE_INPUT
) && !bld
->gs_iface
) {
3492 unsigned index
, chan
;
3493 LLVMTypeRef vec_type
= bld_base
->base
.vec_type
;
3494 LLVMValueRef array_size
= lp_build_const_int32(gallivm
,
3495 bld_base
->info
->file_max
[TGSI_FILE_INPUT
]*4 + 4);
3496 bld
->inputs_array
= lp_build_array_alloca(gallivm
,
3497 vec_type
, array_size
,
3500 assert(bld_base
->info
->num_inputs
3501 <= bld_base
->info
->file_max
[TGSI_FILE_INPUT
] + 1);
3503 for (index
= 0; index
< bld_base
->info
->num_inputs
; ++index
) {
3504 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; ++chan
) {
3505 LLVMValueRef lindex
=
3506 lp_build_const_int32(gallivm
, index
* 4 + chan
);
3507 LLVMValueRef input_ptr
=
3508 LLVMBuildGEP(gallivm
->builder
, bld
->inputs_array
,
3510 LLVMValueRef value
= bld
->inputs
[index
][chan
];
3512 LLVMBuildStore(gallivm
->builder
, value
, input_ptr
);
3517 if (bld
->gs_iface
) {
3518 struct lp_build_context
*uint_bld
= &bld
->bld_base
.uint_bld
;
3519 bld
->emitted_prims_vec_ptr
=
3520 lp_build_alloca(gallivm
,
3522 "emitted_prims_ptr");
3523 bld
->emitted_vertices_vec_ptr
=
3524 lp_build_alloca(gallivm
,
3526 "emitted_vertices_ptr");
3527 bld
->total_emitted_vertices_vec_ptr
=
3528 lp_build_alloca(gallivm
,
3530 "total_emitted_vertices_ptr");
3532 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3533 bld
->emitted_prims_vec_ptr
);
3534 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3535 bld
->emitted_vertices_vec_ptr
);
3536 LLVMBuildStore(gallivm
->builder
, uint_bld
->zero
,
3537 bld
->total_emitted_vertices_vec_ptr
);
3540 if (DEBUG_EXECUTION
) {
3541 lp_build_printf(gallivm
, "\n");
3542 emit_dump_file(bld
, TGSI_FILE_CONSTANT
);
3543 emit_dump_file(bld
, TGSI_FILE_INPUT
);
3547 static void emit_epilogue(struct lp_build_tgsi_context
* bld_base
)
3549 struct lp_build_tgsi_soa_context
* bld
= lp_soa_context(bld_base
);
3550 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
3552 if (DEBUG_EXECUTION
) {
3555 emit_dump_file(bld
, TGSI_FILE_TEMPORARY
);
3557 emit_dump_file(bld
, TGSI_FILE_OUTPUT
);
3558 lp_build_printf(bld_base
->base
.gallivm
, "\n");
3561 /* If we have indirect addressing in outputs we need to copy our alloca array
3562 * to the outputs slots specified by the caller */
3563 if (bld
->gs_iface
) {
3564 LLVMValueRef total_emitted_vertices_vec
;
3565 LLVMValueRef emitted_prims_vec
;
3566 /* implicit end_primitives, needed in case there are any unflushed
3567 vertices in the cache. Note must not call end_primitive here
3568 since the exec_mask is not valid at this point. */
3569 end_primitive_masked(bld_base
, lp_build_mask_value(bld
->mask
));
3571 total_emitted_vertices_vec
=
3572 LLVMBuildLoad(builder
, bld
->total_emitted_vertices_vec_ptr
, "");
3574 LLVMBuildLoad(builder
, bld
->emitted_prims_vec_ptr
, "");
3576 bld
->gs_iface
->gs_epilogue(bld
->gs_iface
,
3578 total_emitted_vertices_vec
,
3581 gather_outputs(bld
);
3586 lp_build_tgsi_soa(struct gallivm_state
*gallivm
,
3587 const struct tgsi_token
*tokens
,
3588 struct lp_type type
,
3589 struct lp_build_mask_context
*mask
,
3590 LLVMValueRef consts_ptr
,
3591 LLVMValueRef const_sizes_ptr
,
3592 const struct lp_bld_tgsi_system_values
*system_values
,
3593 const LLVMValueRef (*inputs
)[TGSI_NUM_CHANNELS
],
3594 LLVMValueRef (*outputs
)[TGSI_NUM_CHANNELS
],
3595 struct lp_build_sampler_soa
*sampler
,
3596 const struct tgsi_shader_info
*info
,
3597 const struct lp_build_tgsi_gs_iface
*gs_iface
)
3599 struct lp_build_tgsi_soa_context bld
;
3601 struct lp_type res_type
;
3603 assert(type
.length
<= LP_MAX_VECTOR_LENGTH
);
3604 memset(&res_type
, 0, sizeof res_type
);
3605 res_type
.width
= type
.width
;
3606 res_type
.length
= type
.length
;
3609 /* Setup build context */
3610 memset(&bld
, 0, sizeof bld
);
3611 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
3612 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
3613 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
3614 lp_build_context_init(&bld
.elem_bld
, gallivm
, lp_elem_type(type
));
3616 bld
.inputs
= inputs
;
3617 bld
.outputs
= outputs
;
3618 bld
.consts_ptr
= consts_ptr
;
3619 bld
.const_sizes_ptr
= const_sizes_ptr
;
3620 bld
.sampler
= sampler
;
3621 bld
.bld_base
.info
= info
;
3622 bld
.indirect_files
= info
->indirect_files
;
3625 * If the number of temporaries is rather large then we just
3626 * allocate them as an array right from the start and treat
3627 * like indirect temporaries.
3629 if (info
->file_max
[TGSI_FILE_TEMPORARY
] >= LP_MAX_INLINED_TEMPS
) {
3630 bld
.indirect_files
|= (1 << TGSI_FILE_TEMPORARY
);
3633 bld
.bld_base
.soa
= TRUE
;
3634 bld
.bld_base
.emit_debug
= emit_debug
;
3635 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
3636 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
3637 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
3638 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
3639 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = emit_fetch_system_value
;
3640 bld
.bld_base
.emit_store
= emit_store
;
3642 bld
.bld_base
.emit_declaration
= lp_emit_declaration_soa
;
3643 bld
.bld_base
.emit_immediate
= lp_emit_immediate_soa
;
3645 bld
.bld_base
.emit_prologue
= emit_prologue
;
3646 bld
.bld_base
.emit_epilogue
= emit_epilogue
;
3648 /* Set opcode actions */
3649 lp_set_default_actions_cpu(&bld
.bld_base
);
3651 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
3652 bld
.bld_base
.op_actions
[TGSI_OPCODE_BGNSUB
].emit
= bgnsub_emit
;
3653 bld
.bld_base
.op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
3654 bld
.bld_base
.op_actions
[TGSI_OPCODE_BREAKC
].emit
= breakc_emit
;
3655 bld
.bld_base
.op_actions
[TGSI_OPCODE_CAL
].emit
= cal_emit
;
3656 bld
.bld_base
.op_actions
[TGSI_OPCODE_CASE
].emit
= case_emit
;
3657 bld
.bld_base
.op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
3658 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDX
].emit
= ddx_emit
;
3659 bld
.bld_base
.op_actions
[TGSI_OPCODE_DDY
].emit
= ddy_emit
;
3660 bld
.bld_base
.op_actions
[TGSI_OPCODE_DEFAULT
].emit
= default_emit
;
3661 bld
.bld_base
.op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
3662 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
3663 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
3664 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSUB
].emit
= endsub_emit
;
3665 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDSWITCH
].emit
= endswitch_emit
;
3666 bld
.bld_base
.op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
3667 bld
.bld_base
.op_actions
[TGSI_OPCODE_UIF
].emit
= uif_emit
;
3668 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILL_IF
].emit
= kill_if_emit
;
3669 bld
.bld_base
.op_actions
[TGSI_OPCODE_KILL
].emit
= kill_emit
;
3670 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM
].emit
= nrm_emit
;
3671 bld
.bld_base
.op_actions
[TGSI_OPCODE_NRM4
].emit
= nrm_emit
;
3672 bld
.bld_base
.op_actions
[TGSI_OPCODE_RET
].emit
= ret_emit
;
3673 bld
.bld_base
.op_actions
[TGSI_OPCODE_SWITCH
].emit
= switch_emit
;
3674 bld
.bld_base
.op_actions
[TGSI_OPCODE_TEX
].emit
= tex_emit
;
3675 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXB
].emit
= txb_emit
;
3676 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXD
].emit
= txd_emit
;
3677 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXL
].emit
= txl_emit
;
3678 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXP
].emit
= txp_emit
;
3679 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXQ
].emit
= txq_emit
;
3680 bld
.bld_base
.op_actions
[TGSI_OPCODE_TXF
].emit
= txf_emit
;
3681 /* DX10 sampling ops */
3682 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE
].emit
= sample_emit
;
3683 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_B
].emit
= sample_b_emit
;
3684 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_C
].emit
= sample_c_emit
;
3685 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_C_LZ
].emit
= sample_c_lz_emit
;
3686 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_D
].emit
= sample_d_emit
;
3687 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_I
].emit
= sample_i_emit
;
3688 bld
.bld_base
.op_actions
[TGSI_OPCODE_SAMPLE_L
].emit
= sample_l_emit
;
3689 bld
.bld_base
.op_actions
[TGSI_OPCODE_SVIEWINFO
].emit
= sviewinfo_emit
;
3692 /* There's no specific value for this because it should always
3693 * be set, but apps using ext_geometry_shader4 quite often
3694 * were forgetting so we're using MAX_VERTEX_VARYING from
3695 * that spec even though we could debug_assert if it's not
3696 * set, but that's a lot uglier. */
3697 uint max_output_vertices
= 32;
3699 /* inputs are always indirect with gs */
3700 bld
.indirect_files
|= (1 << TGSI_FILE_INPUT
);
3701 bld
.gs_iface
= gs_iface
;
3702 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_gs_input
;
3703 bld
.bld_base
.op_actions
[TGSI_OPCODE_EMIT
].emit
= emit_vertex
;
3704 bld
.bld_base
.op_actions
[TGSI_OPCODE_ENDPRIM
].emit
= end_primitive
;
3706 for (i
= 0; i
< info
->num_properties
; ++i
) {
3707 if (info
->properties
[i
].name
==
3708 TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
) {
3709 max_output_vertices
= info
->properties
[i
].data
[0];
3712 bld
.max_output_vertices_vec
=
3713 lp_build_const_int_vec(gallivm
, bld
.bld_base
.int_bld
.type
,
3714 max_output_vertices
);
3717 lp_exec_mask_init(&bld
.exec_mask
, &bld
.bld_base
.int_bld
);
3719 bld
.system_values
= *system_values
;
3721 lp_build_tgsi_llvm(&bld
.bld_base
, tokens
);
3724 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
3725 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
3726 debug_printf("11111111111111111111111111111 \n");
3727 tgsi_dump(tokens
, 0);
3728 lp_debug_dump_value(function
);
3729 debug_printf("2222222222222222222222222222 \n");
3733 LLVMModuleRef module
= LLVMGetGlobalParent(
3734 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
3735 LLVMDumpModule(module
);
3738 lp_exec_mask_fini(&bld
.exec_mask
);