1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
4 * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 **************************************************************************/
31 * TGSI to LLVM IR translation -- SoA.
33 * @author Jose Fonseca <jfonseca@vmware.com>
35 * Based on tgsi_sse2.c code written by Michal Krol, Keith Whitwell,
36 * Brian Paul, and others.
39 #include "pipe/p_config.h"
40 #include "pipe/p_shader_tokens.h"
41 #include "util/u_debug.h"
42 #include "util/u_math.h"
43 #include "util/u_memory.h"
44 #include "tgsi/tgsi_dump.h"
45 #include "tgsi/tgsi_info.h"
46 #include "tgsi/tgsi_parse.h"
47 #include "tgsi/tgsi_util.h"
48 #include "tgsi/tgsi_exec.h"
49 #include "lp_bld_type.h"
50 #include "lp_bld_const.h"
51 #include "lp_bld_arit.h"
52 #include "lp_bld_logic.h"
53 #include "lp_bld_swizzle.h"
54 #include "lp_bld_flow.h"
55 #include "lp_bld_tgsi.h"
56 #include "lp_bld_debug.h"
59 #define LP_MAX_TEMPS 256
60 #define LP_MAX_IMMEDIATES 256
63 #define FOR_EACH_CHANNEL( CHAN )\
64 for (CHAN = 0; CHAN < NUM_CHANNELS; CHAN++)
66 #define IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
67 ((INST)->Dst[0].Register.WriteMask & (1 << (CHAN)))
69 #define IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
70 if (IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
72 #define FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
73 FOR_EACH_CHANNEL( CHAN )\
74 IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
81 #define QUAD_TOP_LEFT 0
82 #define QUAD_TOP_RIGHT 1
83 #define QUAD_BOTTOM_LEFT 2
84 #define QUAD_BOTTOM_RIGHT 3
86 #define LP_TGSI_MAX_NESTING 16
89 struct lp_build_context
*bld
;
93 LLVMTypeRef int_vec_type
;
95 LLVMValueRef cond_stack
[LP_TGSI_MAX_NESTING
];
97 LLVMValueRef cond_mask
;
99 LLVMValueRef break_stack
[LP_TGSI_MAX_NESTING
];
100 int break_stack_size
;
101 LLVMValueRef break_mask
;
103 LLVMValueRef cont_stack
[LP_TGSI_MAX_NESTING
];
105 LLVMValueRef cont_mask
;
107 LLVMBasicBlockRef loop_stack
[LP_TGSI_MAX_NESTING
];
109 LLVMBasicBlockRef loop_block
;
112 LLVMValueRef exec_mask
;
115 struct lp_build_tgsi_soa_context
117 struct lp_build_context base
;
119 LLVMValueRef consts_ptr
;
120 const LLVMValueRef
*pos
;
121 const LLVMValueRef (*inputs
)[NUM_CHANNELS
];
122 LLVMValueRef (*outputs
)[NUM_CHANNELS
];
124 struct lp_build_sampler_soa
*sampler
;
126 LLVMValueRef immediates
[LP_MAX_IMMEDIATES
][NUM_CHANNELS
];
127 LLVMValueRef temps
[LP_MAX_TEMPS
][NUM_CHANNELS
];
129 struct lp_build_mask_context
*mask
;
130 struct lp_exec_mask exec_mask
;
133 static const unsigned char
135 QUAD_TOP_LEFT
, QUAD_TOP_LEFT
,
136 QUAD_BOTTOM_LEFT
, QUAD_BOTTOM_LEFT
139 static const unsigned char
141 QUAD_TOP_RIGHT
, QUAD_TOP_RIGHT
,
142 QUAD_BOTTOM_RIGHT
, QUAD_BOTTOM_RIGHT
145 static const unsigned char
147 QUAD_TOP_LEFT
, QUAD_TOP_RIGHT
,
148 QUAD_TOP_LEFT
, QUAD_TOP_RIGHT
151 static const unsigned char
152 swizzle_bottom
[4] = {
153 QUAD_BOTTOM_LEFT
, QUAD_BOTTOM_RIGHT
,
154 QUAD_BOTTOM_LEFT
, QUAD_BOTTOM_RIGHT
157 static void lp_exec_mask_init(struct lp_exec_mask
*mask
, struct lp_build_context
*bld
)
160 mask
->has_mask
= FALSE
;
161 mask
->cond_stack_size
= 0;
162 mask
->loop_stack_size
= 0;
163 mask
->break_stack_size
= 0;
164 mask
->cont_stack_size
= 0;
166 mask
->int_vec_type
= lp_build_int_vec_type(mask
->bld
->type
);
169 static void lp_exec_mask_update(struct lp_exec_mask
*mask
)
171 if (mask
->loop_stack_size
) {
172 /*for loops we need to update the entire mask at
175 tmp
= LLVMBuildAnd(mask
->bld
->builder
,
179 mask
->exec_mask
= LLVMBuildAnd(mask
->bld
->builder
,
184 mask
->exec_mask
= mask
->cond_mask
;
187 mask
->has_mask
= (mask
->cond_stack_size
> 0 ||
188 mask
->loop_stack_size
> 0);
191 static void lp_exec_mask_cond_push(struct lp_exec_mask
*mask
,
194 mask
->cond_stack
[mask
->cond_stack_size
++] = mask
->cond_mask
;
195 mask
->cond_mask
= LLVMBuildBitCast(mask
->bld
->builder
, val
,
196 mask
->int_vec_type
, "");
198 lp_exec_mask_update(mask
);
201 static void lp_exec_mask_cond_invert(struct lp_exec_mask
*mask
)
203 LLVMValueRef prev_mask
= mask
->cond_stack
[mask
->cond_stack_size
- 1];
204 LLVMValueRef inv_mask
= LLVMBuildNot(mask
->bld
->builder
,
205 mask
->cond_mask
, "");
207 /* means that we didn't have any mask before and that
208 * we were fully enabled */
209 if (mask
->cond_stack_size
<= 1) {
210 prev_mask
= LLVMConstAllOnes(mask
->int_vec_type
);
213 mask
->cond_mask
= LLVMBuildAnd(mask
->bld
->builder
,
216 lp_exec_mask_update(mask
);
219 static void lp_exec_mask_cond_pop(struct lp_exec_mask
*mask
)
221 mask
->cond_mask
= mask
->cond_stack
[--mask
->cond_stack_size
];
222 lp_exec_mask_update(mask
);
225 static void lp_exec_bgnloop(struct lp_exec_mask
*mask
)
228 if (mask
->cont_stack_size
== 0)
229 mask
->cont_mask
= LLVMConstAllOnes(mask
->int_vec_type
);
230 if (mask
->break_stack_size
== 0)
231 mask
->break_mask
= LLVMConstAllOnes(mask
->int_vec_type
);
232 if (mask
->cond_stack_size
== 0)
233 mask
->cond_mask
= LLVMConstAllOnes(mask
->int_vec_type
);
234 mask
->loop_stack
[mask
->loop_stack_size
++] = mask
->loop_block
;
235 mask
->loop_block
= lp_build_insert_new_block(mask
->bld
->builder
, "bgnloop");
236 LLVMBuildBr(mask
->bld
->builder
, mask
->loop_block
);
237 LLVMPositionBuilderAtEnd(mask
->bld
->builder
, mask
->loop_block
);
239 lp_exec_mask_update(mask
);
242 static void lp_exec_break(struct lp_exec_mask
*mask
)
244 LLVMValueRef exec_mask
= LLVMBuildNot(mask
->bld
->builder
,
248 mask
->break_stack
[mask
->break_stack_size
++] = mask
->break_mask
;
249 if (mask
->break_stack_size
> 1) {
250 mask
->break_mask
= LLVMBuildAnd(mask
->bld
->builder
,
252 exec_mask
, "break_full");
254 mask
->break_mask
= exec_mask
;
256 lp_exec_mask_update(mask
);
259 static void lp_exec_continue(struct lp_exec_mask
*mask
)
261 LLVMValueRef exec_mask
= LLVMBuildNot(mask
->bld
->builder
,
265 mask
->cont_stack
[mask
->cont_stack_size
++] = mask
->cont_mask
;
266 if (mask
->cont_stack_size
> 1) {
267 mask
->cont_mask
= LLVMBuildAnd(mask
->bld
->builder
,
271 mask
->cont_mask
= exec_mask
;
273 lp_exec_mask_update(mask
);
277 static void lp_exec_endloop(struct lp_exec_mask
*mask
)
279 LLVMBasicBlockRef endloop
;
280 LLVMTypeRef reg_type
= LLVMIntType(mask
->bld
->type
.width
*
281 mask
->bld
->type
.length
);
282 /* i1cond = (mask == 0) */
283 LLVMValueRef i1cond
= LLVMBuildICmp(
286 LLVMBuildBitCast(mask
->bld
->builder
, mask
->break_mask
, reg_type
, ""),
287 LLVMConstNull(reg_type
), "");
289 endloop
= lp_build_insert_new_block(mask
->bld
->builder
, "endloop");
291 LLVMBuildCondBr(mask
->bld
->builder
,
292 i1cond
, mask
->loop_block
, endloop
);
294 LLVMPositionBuilderAtEnd(mask
->bld
->builder
, endloop
);
296 mask
->loop_block
= mask
->loop_stack
[--mask
->loop_stack_size
];
297 /* pop the break mask */
298 if (mask
->cont_stack_size
) {
299 mask
->cont_mask
= mask
->cont_stack
[--mask
->cont_stack_size
];
301 if (mask
->break_stack_size
) {
302 mask
->break_mask
= mask
->cont_stack
[--mask
->break_stack_size
];
305 lp_exec_mask_update(mask
);
308 static void lp_exec_mask_store(struct lp_exec_mask
*mask
,
312 if (mask
->has_mask
) {
313 LLVMValueRef real_val
, dst_val
;
315 dst_val
= LLVMBuildLoad(mask
->bld
->builder
, dst
, "");
316 real_val
= lp_build_select(mask
->bld
,
320 LLVMBuildStore(mask
->bld
->builder
, real_val
, dst
);
322 LLVMBuildStore(mask
->bld
->builder
, val
, dst
);
327 emit_ddx(struct lp_build_tgsi_soa_context
*bld
,
330 LLVMValueRef src_left
= lp_build_swizzle1_aos(&bld
->base
, src
, swizzle_left
);
331 LLVMValueRef src_right
= lp_build_swizzle1_aos(&bld
->base
, src
, swizzle_right
);
332 return lp_build_sub(&bld
->base
, src_right
, src_left
);
337 emit_ddy(struct lp_build_tgsi_soa_context
*bld
,
340 LLVMValueRef src_top
= lp_build_swizzle1_aos(&bld
->base
, src
, swizzle_top
);
341 LLVMValueRef src_bottom
= lp_build_swizzle1_aos(&bld
->base
, src
, swizzle_bottom
);
342 return lp_build_sub(&bld
->base
, src_top
, src_bottom
);
351 struct lp_build_tgsi_soa_context
*bld
,
352 const struct tgsi_full_instruction
*inst
,
354 const unsigned chan_index
)
356 const struct tgsi_full_src_register
*reg
= &inst
->Src
[index
];
357 unsigned swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
366 switch (reg
->Register
.File
) {
367 case TGSI_FILE_CONSTANT
: {
368 LLVMValueRef index
= LLVMConstInt(LLVMInt32Type(), reg
->Register
.Index
*4 + swizzle
, 0);
369 LLVMValueRef scalar_ptr
= LLVMBuildGEP(bld
->base
.builder
, bld
->consts_ptr
, &index
, 1, "");
370 LLVMValueRef scalar
= LLVMBuildLoad(bld
->base
.builder
, scalar_ptr
, "");
371 res
= lp_build_broadcast_scalar(&bld
->base
, scalar
);
375 case TGSI_FILE_IMMEDIATE
:
376 res
= bld
->immediates
[reg
->Register
.Index
][swizzle
];
380 case TGSI_FILE_INPUT
:
381 res
= bld
->inputs
[reg
->Register
.Index
][swizzle
];
385 case TGSI_FILE_TEMPORARY
:
386 res
= LLVMBuildLoad(bld
->base
.builder
, bld
->temps
[reg
->Register
.Index
][swizzle
], "");
388 return bld
->base
.undef
;
393 return bld
->base
.undef
;
399 return bld
->base
.undef
;
402 switch( tgsi_util_get_full_src_register_sign_mode( reg
, chan_index
) ) {
403 case TGSI_UTIL_SIGN_CLEAR
:
404 res
= lp_build_abs( &bld
->base
, res
);
407 case TGSI_UTIL_SIGN_SET
:
408 /* TODO: Use bitwese OR for floating point */
409 res
= lp_build_abs( &bld
->base
, res
);
410 res
= LLVMBuildNeg( bld
->base
.builder
, res
, "" );
413 case TGSI_UTIL_SIGN_TOGGLE
:
414 res
= LLVMBuildNeg( bld
->base
.builder
, res
, "" );
417 case TGSI_UTIL_SIGN_KEEP
:
426 * Register fetch with derivatives.
430 struct lp_build_tgsi_soa_context
*bld
,
431 const struct tgsi_full_instruction
*inst
,
433 const unsigned chan_index
,
440 src
= emit_fetch(bld
, inst
, index
, chan_index
);
445 /* TODO: use interpolation coeffs for inputs */
448 *ddx
= emit_ddx(bld
, src
);
451 *ddy
= emit_ddy(bld
, src
);
460 struct lp_build_tgsi_soa_context
*bld
,
461 const struct tgsi_full_instruction
*inst
,
466 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
468 switch( inst
->Instruction
.Saturate
) {
472 case TGSI_SAT_ZERO_ONE
:
473 value
= lp_build_max(&bld
->base
, value
, bld
->base
.zero
);
474 value
= lp_build_min(&bld
->base
, value
, bld
->base
.one
);
477 case TGSI_SAT_MINUS_PLUS_ONE
:
478 value
= lp_build_max(&bld
->base
, value
, lp_build_const_vec(bld
->base
.type
, -1.0));
479 value
= lp_build_min(&bld
->base
, value
, bld
->base
.one
);
486 switch( reg
->Register
.File
) {
487 case TGSI_FILE_OUTPUT
:
488 lp_exec_mask_store(&bld
->exec_mask
, value
,
489 bld
->outputs
[reg
->Register
.Index
][chan_index
]);
492 case TGSI_FILE_TEMPORARY
:
493 lp_exec_mask_store(&bld
->exec_mask
, value
,
494 bld
->temps
[reg
->Register
.Index
][chan_index
]);
497 case TGSI_FILE_ADDRESS
:
502 case TGSI_FILE_PREDICATE
:
514 * High-level instruction translators.
519 emit_tex( struct lp_build_tgsi_soa_context
*bld
,
520 const struct tgsi_full_instruction
*inst
,
521 boolean apply_lodbias
,
525 const uint unit
= inst
->Src
[1].Register
.Index
;
526 LLVMValueRef lodbias
;
527 LLVMValueRef oow
= NULL
;
528 LLVMValueRef coords
[3];
532 switch (inst
->Texture
.Texture
) {
533 case TGSI_TEXTURE_1D
:
536 case TGSI_TEXTURE_2D
:
537 case TGSI_TEXTURE_RECT
:
540 case TGSI_TEXTURE_SHADOW1D
:
541 case TGSI_TEXTURE_SHADOW2D
:
542 case TGSI_TEXTURE_SHADOWRECT
:
543 case TGSI_TEXTURE_3D
:
544 case TGSI_TEXTURE_CUBE
:
553 lodbias
= emit_fetch( bld
, inst
, 0, 3 );
555 lodbias
= bld
->base
.zero
;
558 oow
= emit_fetch( bld
, inst
, 0, 3 );
559 oow
= lp_build_rcp(&bld
->base
, oow
);
562 for (i
= 0; i
< num_coords
; i
++) {
563 coords
[i
] = emit_fetch( bld
, inst
, 0, i
);
565 coords
[i
] = lp_build_mul(&bld
->base
, coords
[i
], oow
);
567 for (i
= num_coords
; i
< 3; i
++) {
568 coords
[i
] = bld
->base
.undef
;
571 bld
->sampler
->emit_fetch_texel(bld
->sampler
,
574 unit
, num_coords
, coords
, lodbias
,
580 * Kill fragment if any of the src register values are negative.
584 struct lp_build_tgsi_soa_context
*bld
,
585 const struct tgsi_full_instruction
*inst
)
587 const struct tgsi_full_src_register
*reg
= &inst
->Src
[0];
588 LLVMValueRef terms
[NUM_CHANNELS
];
592 memset(&terms
, 0, sizeof terms
);
594 FOR_EACH_CHANNEL( chan_index
) {
597 /* Unswizzle channel */
598 swizzle
= tgsi_util_get_full_src_register_swizzle( reg
, chan_index
);
600 /* Check if the component has not been already tested. */
601 assert(swizzle
< NUM_CHANNELS
);
602 if( !terms
[swizzle
] )
603 /* TODO: change the comparison operator instead of setting the sign */
604 terms
[swizzle
] = emit_fetch(bld
, inst
, 0, chan_index
);
608 FOR_EACH_CHANNEL( chan_index
) {
609 if(terms
[chan_index
]) {
610 LLVMValueRef chan_mask
;
613 * If term < 0 then mask = 0 else mask = ~0.
615 chan_mask
= lp_build_cmp(&bld
->base
, PIPE_FUNC_GEQUAL
, terms
[chan_index
], bld
->base
.zero
);
618 mask
= LLVMBuildAnd(bld
->base
.builder
, mask
, chan_mask
, "");
625 lp_build_mask_update(bld
->mask
, mask
);
630 * Predicated fragment kill.
631 * XXX Actually, we do an unconditional kill (as in tgsi_exec.c).
632 * The only predication is the execution mask which will apply if
633 * we're inside a loop or conditional.
636 emit_kilp(struct lp_build_tgsi_soa_context
*bld
,
637 const struct tgsi_full_instruction
*inst
)
641 /* For those channels which are "alive", disable fragment shader
644 if (bld
->exec_mask
.has_mask
) {
645 mask
= LLVMBuildNot(bld
->base
.builder
, bld
->exec_mask
.exec_mask
, "kilp");
648 mask
= bld
->base
.zero
;
651 lp_build_mask_update(bld
->mask
, mask
);
656 * Check if inst src/dest regs use indirect addressing into temporary
660 indirect_temp_reference(const struct tgsi_full_instruction
*inst
)
663 for (i
= 0; i
< inst
->Instruction
.NumSrcRegs
; i
++) {
664 const struct tgsi_full_src_register
*reg
= &inst
->Src
[i
];
665 if (reg
->Register
.File
== TGSI_FILE_TEMPORARY
&&
666 reg
->Register
.Indirect
)
669 for (i
= 0; i
< inst
->Instruction
.NumDstRegs
; i
++) {
670 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[i
];
671 if (reg
->Register
.File
== TGSI_FILE_TEMPORARY
&&
672 reg
->Register
.Indirect
)
680 struct lp_build_tgsi_soa_context
*bld
,
681 const struct tgsi_full_declaration
*decl
)
683 unsigned first
= decl
->Range
.First
;
684 unsigned last
= decl
->Range
.Last
;
686 LLVMBasicBlockRef current_block
=
687 LLVMGetInsertBlock(bld
->base
.builder
);
688 LLVMBasicBlockRef first_block
=
689 LLVMGetEntryBasicBlock(
690 LLVMGetBasicBlockParent(current_block
));
691 LLVMValueRef first_inst
=
692 LLVMGetFirstInstruction(first_block
);
694 /* we want alloca's to be the first instruction
695 * in the function so we need to rewind the builder
696 * to the very beginning */
697 LLVMPositionBuilderBefore(bld
->base
.builder
,
700 for (idx
= first
; idx
<= last
; ++idx
) {
701 switch (decl
->Declaration
.File
) {
702 case TGSI_FILE_TEMPORARY
:
703 for (i
= 0; i
< NUM_CHANNELS
; i
++)
704 bld
->temps
[idx
][i
] = lp_build_alloca(&bld
->base
);
707 case TGSI_FILE_OUTPUT
:
708 for (i
= 0; i
< NUM_CHANNELS
; i
++)
709 bld
->outputs
[idx
][i
] = lp_build_alloca(&bld
->base
);
713 /* don't need to declare other vars */
718 LLVMPositionBuilderAtEnd(bld
->base
.builder
,
725 * Emit LLVM for one TGSI instruction.
726 * \param return TRUE for success, FALSE otherwise
730 struct lp_build_tgsi_soa_context
*bld
,
731 const struct tgsi_full_instruction
*inst
,
732 const struct tgsi_opcode_info
*info
)
735 LLVMValueRef src0
, src1
, src2
;
736 LLVMValueRef tmp0
, tmp1
, tmp2
;
737 LLVMValueRef tmp3
= NULL
;
738 LLVMValueRef tmp4
= NULL
;
739 LLVMValueRef tmp5
= NULL
;
740 LLVMValueRef tmp6
= NULL
;
741 LLVMValueRef tmp7
= NULL
;
743 LLVMValueRef dst0
[NUM_CHANNELS
];
745 /* we can't handle indirect addressing into temp register file yet */
746 if (indirect_temp_reference(inst
))
750 * Stores and write masks are handled in a general fashion after the long
751 * instruction opcode switch statement.
753 * Although not stricitly necessary, we avoid generating instructions for
754 * channels which won't be stored, in cases where's that easy. For some
755 * complex instructions, like texture sampling, it is more convenient to
756 * assume a full writemask and then let LLVM optimization passes eliminate
760 assert(info
->num_dst
<= 1);
762 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
763 dst0
[chan_index
] = bld
->base
.undef
;
767 switch (inst
->Instruction
.Opcode
) {
769 case TGSI_OPCODE_ARL
:
771 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
772 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
775 dst0
[chan_index
] = tmp0
;
780 case TGSI_OPCODE_MOV
:
781 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
782 dst0
[chan_index
] = emit_fetch( bld
, inst
, 0, chan_index
);
786 case TGSI_OPCODE_LIT
:
787 if( IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
) ) {
788 dst0
[CHAN_X
] = bld
->base
.one
;
790 if( IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
) ) {
791 src0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
792 dst0
[CHAN_Y
] = lp_build_max( &bld
->base
, src0
, bld
->base
.zero
);
794 if( IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
) ) {
795 /* XMM[1] = SrcReg[0].yyyy */
796 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Y
);
797 /* XMM[1] = max(XMM[1], 0) */
798 tmp1
= lp_build_max( &bld
->base
, tmp1
, bld
->base
.zero
);
799 /* XMM[2] = SrcReg[0].wwww */
800 tmp2
= emit_fetch( bld
, inst
, 0, CHAN_W
);
801 tmp1
= lp_build_pow( &bld
->base
, tmp1
, tmp2
);
802 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
803 tmp2
= lp_build_cmp(&bld
->base
, PIPE_FUNC_GREATER
, tmp0
, bld
->base
.zero
);
804 dst0
[CHAN_Z
] = lp_build_select(&bld
->base
, tmp2
, tmp1
, bld
->base
.zero
);
806 if( IS_DST0_CHANNEL_ENABLED( inst
, CHAN_W
) ) {
807 dst0
[CHAN_W
] = bld
->base
.one
;
811 case TGSI_OPCODE_RCP
:
812 /* TGSI_OPCODE_RECIP */
813 src0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
814 res
= lp_build_rcp(&bld
->base
, src0
);
815 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
816 dst0
[chan_index
] = res
;
820 case TGSI_OPCODE_RSQ
:
821 /* TGSI_OPCODE_RECIPSQRT */
822 src0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
823 src0
= lp_build_abs(&bld
->base
, src0
);
824 res
= lp_build_rsqrt(&bld
->base
, src0
);
825 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
826 dst0
[chan_index
] = res
;
830 case TGSI_OPCODE_EXP
:
831 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
) ||
832 IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
) ||
833 IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
)) {
834 LLVMValueRef
*p_exp2_int_part
= NULL
;
835 LLVMValueRef
*p_frac_part
= NULL
;
836 LLVMValueRef
*p_exp2
= NULL
;
838 src0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
840 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
))
841 p_exp2_int_part
= &tmp0
;
842 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
))
844 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
))
847 lp_build_exp2_approx(&bld
->base
, src0
, p_exp2_int_part
, p_frac_part
, p_exp2
);
849 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
))
851 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
))
853 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
))
857 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_W
)) {
858 dst0
[CHAN_W
] = bld
->base
.one
;
862 case TGSI_OPCODE_LOG
:
863 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
) ||
864 IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
) ||
865 IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
)) {
866 LLVMValueRef
*p_floor_log2
= NULL
;
867 LLVMValueRef
*p_exp
= NULL
;
868 LLVMValueRef
*p_log2
= NULL
;
870 src0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
871 src0
= lp_build_abs( &bld
->base
, src0
);
873 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
))
874 p_floor_log2
= &tmp0
;
875 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
))
877 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
))
880 lp_build_log2_approx(&bld
->base
, src0
, p_exp
, p_floor_log2
, p_log2
);
882 /* dst.x = floor(lg2(abs(src.x))) */
883 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
))
885 /* dst.y = abs(src)/ex2(floor(lg2(abs(src.x)))) */
886 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
)) {
887 dst0
[CHAN_Y
] = lp_build_div( &bld
->base
, src0
, tmp1
);
889 /* dst.z = lg2(abs(src.x)) */
890 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
))
894 if (IS_DST0_CHANNEL_ENABLED( inst
, CHAN_W
)) {
895 dst0
[CHAN_W
] = bld
->base
.one
;
899 case TGSI_OPCODE_MUL
:
900 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
901 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
902 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
903 dst0
[chan_index
] = lp_build_mul(&bld
->base
, src0
, src1
);
907 case TGSI_OPCODE_ADD
:
908 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
909 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
910 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
911 dst0
[chan_index
] = lp_build_add(&bld
->base
, src0
, src1
);
915 case TGSI_OPCODE_DP3
:
916 /* TGSI_OPCODE_DOT3 */
917 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
918 tmp1
= emit_fetch( bld
, inst
, 1, CHAN_X
);
919 tmp0
= lp_build_mul( &bld
->base
, tmp0
, tmp1
);
920 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Y
);
921 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_Y
);
922 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
);
923 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
924 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Z
);
925 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_Z
);
926 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
);
927 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
928 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
929 dst0
[chan_index
] = tmp0
;
933 case TGSI_OPCODE_DP4
:
934 /* TGSI_OPCODE_DOT4 */
935 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
936 tmp1
= emit_fetch( bld
, inst
, 1, CHAN_X
);
937 tmp0
= lp_build_mul( &bld
->base
, tmp0
, tmp1
);
938 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Y
);
939 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_Y
);
940 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
);
941 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
942 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Z
);
943 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_Z
);
944 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
);
945 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
946 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_W
);
947 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_W
);
948 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
);
949 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
950 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
951 dst0
[chan_index
] = tmp0
;
955 case TGSI_OPCODE_DST
:
956 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
) {
957 dst0
[CHAN_X
] = bld
->base
.one
;
959 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
) {
960 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_Y
);
961 tmp1
= emit_fetch( bld
, inst
, 1, CHAN_Y
);
962 dst0
[CHAN_Y
] = lp_build_mul( &bld
->base
, tmp0
, tmp1
);
964 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
) {
965 dst0
[CHAN_Z
] = emit_fetch( bld
, inst
, 0, CHAN_Z
);
967 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_W
) {
968 dst0
[CHAN_W
] = emit_fetch( bld
, inst
, 1, CHAN_W
);
972 case TGSI_OPCODE_MIN
:
973 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
974 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
975 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
976 dst0
[chan_index
] = lp_build_min( &bld
->base
, src0
, src1
);
980 case TGSI_OPCODE_MAX
:
981 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
982 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
983 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
984 dst0
[chan_index
] = lp_build_max( &bld
->base
, src0
, src1
);
988 case TGSI_OPCODE_SLT
:
989 /* TGSI_OPCODE_SETLT */
990 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
991 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
992 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
993 tmp0
= lp_build_cmp( &bld
->base
, PIPE_FUNC_LESS
, src0
, src1
);
994 dst0
[chan_index
] = lp_build_select( &bld
->base
, tmp0
, bld
->base
.one
, bld
->base
.zero
);
998 case TGSI_OPCODE_SGE
:
999 /* TGSI_OPCODE_SETGE */
1000 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1001 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1002 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1003 tmp0
= lp_build_cmp( &bld
->base
, PIPE_FUNC_GEQUAL
, src0
, src1
);
1004 dst0
[chan_index
] = lp_build_select( &bld
->base
, tmp0
, bld
->base
.one
, bld
->base
.zero
);
1008 case TGSI_OPCODE_MAD
:
1009 /* TGSI_OPCODE_MADD */
1010 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1011 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1012 tmp1
= emit_fetch( bld
, inst
, 1, chan_index
);
1013 tmp2
= emit_fetch( bld
, inst
, 2, chan_index
);
1014 tmp0
= lp_build_mul( &bld
->base
, tmp0
, tmp1
);
1015 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp2
);
1016 dst0
[chan_index
] = tmp0
;
1020 case TGSI_OPCODE_SUB
:
1021 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1022 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1023 tmp1
= emit_fetch( bld
, inst
, 1, chan_index
);
1024 dst0
[chan_index
] = lp_build_sub( &bld
->base
, tmp0
, tmp1
);
1028 case TGSI_OPCODE_LRP
:
1029 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1030 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1031 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1032 src2
= emit_fetch( bld
, inst
, 2, chan_index
);
1033 tmp0
= lp_build_sub( &bld
->base
, src1
, src2
);
1034 tmp0
= lp_build_mul( &bld
->base
, src0
, tmp0
);
1035 dst0
[chan_index
] = lp_build_add( &bld
->base
, tmp0
, src2
);
1039 case TGSI_OPCODE_CND
:
1040 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1041 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1042 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1043 src2
= emit_fetch( bld
, inst
, 2, chan_index
);
1044 tmp1
= lp_build_const_vec(bld
->base
.type
, 0.5);
1045 tmp0
= lp_build_cmp( &bld
->base
, PIPE_FUNC_GREATER
, src2
, tmp1
);
1046 dst0
[chan_index
] = lp_build_select( &bld
->base
, tmp0
, src0
, src1
);
1050 case TGSI_OPCODE_DP2A
:
1051 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
); /* xmm0 = src[0].x */
1052 tmp1
= emit_fetch( bld
, inst
, 1, CHAN_X
); /* xmm1 = src[1].x */
1053 tmp0
= lp_build_mul( &bld
->base
, tmp0
, tmp1
); /* xmm0 = xmm0 * xmm1 */
1054 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Y
); /* xmm1 = src[0].y */
1055 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_Y
); /* xmm2 = src[1].y */
1056 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
); /* xmm1 = xmm1 * xmm2 */
1057 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
); /* xmm0 = xmm0 + xmm1 */
1058 tmp1
= emit_fetch( bld
, inst
, 2, CHAN_X
); /* xmm1 = src[2].x */
1059 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
); /* xmm0 = xmm0 + xmm1 */
1060 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1061 dst0
[chan_index
] = tmp0
; /* dest[ch] = xmm0 */
1065 case TGSI_OPCODE_FRC
:
1066 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1067 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1068 tmp0
= lp_build_floor(&bld
->base
, src0
);
1069 tmp0
= lp_build_sub(&bld
->base
, src0
, tmp0
);
1070 dst0
[chan_index
] = tmp0
;
1074 case TGSI_OPCODE_CLAMP
:
1075 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1076 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1077 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1078 src2
= emit_fetch( bld
, inst
, 2, chan_index
);
1079 tmp0
= lp_build_max(&bld
->base
, tmp0
, src1
);
1080 tmp0
= lp_build_min(&bld
->base
, tmp0
, src2
);
1081 dst0
[chan_index
] = tmp0
;
1085 case TGSI_OPCODE_FLR
:
1086 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1087 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1088 dst0
[chan_index
] = lp_build_floor(&bld
->base
, tmp0
);
1092 case TGSI_OPCODE_ROUND
:
1093 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1094 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1095 dst0
[chan_index
] = lp_build_round(&bld
->base
, tmp0
);
1099 case TGSI_OPCODE_EX2
: {
1100 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1101 tmp0
= lp_build_exp2( &bld
->base
, tmp0
);
1102 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1103 dst0
[chan_index
] = tmp0
;
1108 case TGSI_OPCODE_LG2
:
1109 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1110 tmp0
= lp_build_log2( &bld
->base
, tmp0
);
1111 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1112 dst0
[chan_index
] = tmp0
;
1116 case TGSI_OPCODE_POW
:
1117 src0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1118 src1
= emit_fetch( bld
, inst
, 1, CHAN_X
);
1119 res
= lp_build_pow( &bld
->base
, src0
, src1
);
1120 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1121 dst0
[chan_index
] = res
;
1125 case TGSI_OPCODE_XPD
:
1126 if( IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
) ||
1127 IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
) ) {
1128 tmp1
= emit_fetch( bld
, inst
, 1, CHAN_Z
);
1129 tmp3
= emit_fetch( bld
, inst
, 0, CHAN_Z
);
1131 if( IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
) ||
1132 IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
) ) {
1133 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_Y
);
1134 tmp4
= emit_fetch( bld
, inst
, 1, CHAN_Y
);
1136 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
) {
1138 tmp2
= lp_build_mul( &bld
->base
, tmp2
, tmp1
);
1140 tmp5
= lp_build_mul( &bld
->base
, tmp5
, tmp4
);
1141 tmp2
= lp_build_sub( &bld
->base
, tmp2
, tmp5
);
1142 dst0
[CHAN_X
] = tmp2
;
1144 if( IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
) ||
1145 IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
) ) {
1146 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_X
);
1147 tmp5
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1149 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
) {
1150 tmp3
= lp_build_mul( &bld
->base
, tmp3
, tmp2
);
1151 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp5
);
1152 tmp3
= lp_build_sub( &bld
->base
, tmp3
, tmp1
);
1153 dst0
[CHAN_Y
] = tmp3
;
1155 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
) {
1156 tmp5
= lp_build_mul( &bld
->base
, tmp5
, tmp4
);
1157 tmp0
= lp_build_mul( &bld
->base
, tmp0
, tmp2
);
1158 tmp5
= lp_build_sub( &bld
->base
, tmp5
, tmp0
);
1159 dst0
[CHAN_Z
] = tmp5
;
1161 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_W
) {
1162 dst0
[CHAN_W
] = bld
->base
.one
;
1166 case TGSI_OPCODE_ABS
:
1167 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1168 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1169 dst0
[chan_index
] = lp_build_abs( &bld
->base
, tmp0
);
1173 case TGSI_OPCODE_RCC
:
1178 case TGSI_OPCODE_DPH
:
1179 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1180 tmp1
= emit_fetch( bld
, inst
, 1, CHAN_X
);
1181 tmp0
= lp_build_mul( &bld
->base
, tmp0
, tmp1
);
1182 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Y
);
1183 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_Y
);
1184 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
);
1185 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
1186 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Z
);
1187 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_Z
);
1188 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
);
1189 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
1190 tmp1
= emit_fetch( bld
, inst
, 1, CHAN_W
);
1191 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
1192 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1193 dst0
[chan_index
] = tmp0
;
1197 case TGSI_OPCODE_COS
:
1198 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1199 tmp0
= lp_build_cos( &bld
->base
, tmp0
);
1200 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1201 dst0
[chan_index
] = tmp0
;
1205 case TGSI_OPCODE_DDX
:
1206 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1207 emit_fetch_deriv( bld
, inst
, 0, chan_index
, NULL
, &dst0
[chan_index
], NULL
);
1211 case TGSI_OPCODE_DDY
:
1212 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1213 emit_fetch_deriv( bld
, inst
, 0, chan_index
, NULL
, NULL
, &dst0
[chan_index
]);
1217 case TGSI_OPCODE_KILP
:
1218 /* predicated kill */
1219 emit_kilp( bld
, inst
);
1222 case TGSI_OPCODE_KIL
:
1223 /* conditional kill */
1224 emit_kil( bld
, inst
);
1227 case TGSI_OPCODE_PK2H
:
1231 case TGSI_OPCODE_PK2US
:
1235 case TGSI_OPCODE_PK4B
:
1239 case TGSI_OPCODE_PK4UB
:
1243 case TGSI_OPCODE_RFL
:
1247 case TGSI_OPCODE_SEQ
:
1248 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1249 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1250 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1251 tmp0
= lp_build_cmp( &bld
->base
, PIPE_FUNC_EQUAL
, src0
, src1
);
1252 dst0
[chan_index
] = lp_build_select( &bld
->base
, tmp0
, bld
->base
.one
, bld
->base
.zero
);
1256 case TGSI_OPCODE_SFL
:
1257 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1258 dst0
[chan_index
] = bld
->base
.zero
;
1262 case TGSI_OPCODE_SGT
:
1263 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1264 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1265 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1266 tmp0
= lp_build_cmp( &bld
->base
, PIPE_FUNC_GREATER
, src0
, src1
);
1267 dst0
[chan_index
] = lp_build_select( &bld
->base
, tmp0
, bld
->base
.one
, bld
->base
.zero
);
1271 case TGSI_OPCODE_SIN
:
1272 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1273 tmp0
= lp_build_sin( &bld
->base
, tmp0
);
1274 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1275 dst0
[chan_index
] = tmp0
;
1279 case TGSI_OPCODE_SLE
:
1280 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1281 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1282 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1283 tmp0
= lp_build_cmp( &bld
->base
, PIPE_FUNC_LEQUAL
, src0
, src1
);
1284 dst0
[chan_index
] = lp_build_select( &bld
->base
, tmp0
, bld
->base
.one
, bld
->base
.zero
);
1288 case TGSI_OPCODE_SNE
:
1289 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1290 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1291 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1292 tmp0
= lp_build_cmp( &bld
->base
, PIPE_FUNC_NOTEQUAL
, src0
, src1
);
1293 dst0
[chan_index
] = lp_build_select( &bld
->base
, tmp0
, bld
->base
.one
, bld
->base
.zero
);
1297 case TGSI_OPCODE_STR
:
1298 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1299 dst0
[chan_index
] = bld
->base
.one
;
1303 case TGSI_OPCODE_TEX
:
1304 emit_tex( bld
, inst
, FALSE
, FALSE
, dst0
);
1307 case TGSI_OPCODE_TXD
:
1312 case TGSI_OPCODE_UP2H
:
1318 case TGSI_OPCODE_UP2US
:
1324 case TGSI_OPCODE_UP4B
:
1330 case TGSI_OPCODE_UP4UB
:
1336 case TGSI_OPCODE_X2D
:
1342 case TGSI_OPCODE_ARA
:
1349 case TGSI_OPCODE_ARR
:
1351 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1352 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1353 emit_rnd( bld
, 0, 0 );
1354 emit_f2it( bld
, 0 );
1355 dst0
[chan_index
] = tmp0
;
1360 case TGSI_OPCODE_BRA
:
1366 case TGSI_OPCODE_CAL
:
1371 case TGSI_OPCODE_RET
:
1376 case TGSI_OPCODE_END
:
1379 case TGSI_OPCODE_SSG
:
1380 /* TGSI_OPCODE_SGN */
1381 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1382 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1383 dst0
[chan_index
] = lp_build_sgn( &bld
->base
, tmp0
);
1387 case TGSI_OPCODE_CMP
:
1388 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1389 src0
= emit_fetch( bld
, inst
, 0, chan_index
);
1390 src1
= emit_fetch( bld
, inst
, 1, chan_index
);
1391 src2
= emit_fetch( bld
, inst
, 2, chan_index
);
1392 tmp0
= lp_build_cmp( &bld
->base
, PIPE_FUNC_LESS
, src0
, bld
->base
.zero
);
1393 dst0
[chan_index
] = lp_build_select( &bld
->base
, tmp0
, src1
, src2
);
1397 case TGSI_OPCODE_SCS
:
1398 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_X
) {
1399 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1400 dst0
[CHAN_X
] = lp_build_cos( &bld
->base
, tmp0
);
1402 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Y
) {
1403 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
);
1404 dst0
[CHAN_Y
] = lp_build_sin( &bld
->base
, tmp0
);
1406 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_Z
) {
1407 dst0
[CHAN_Z
] = bld
->base
.zero
;
1409 IF_IS_DST0_CHANNEL_ENABLED( inst
, CHAN_W
) {
1410 dst0
[CHAN_W
] = bld
->base
.one
;
1414 case TGSI_OPCODE_TXB
:
1415 emit_tex( bld
, inst
, TRUE
, FALSE
, dst0
);
1418 case TGSI_OPCODE_NRM
:
1420 case TGSI_OPCODE_NRM4
:
1421 /* 3 or 4-component normalization */
1423 uint dims
= (inst
->Instruction
.Opcode
== TGSI_OPCODE_NRM
) ? 3 : 4;
1425 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_X
) ||
1426 IS_DST0_CHANNEL_ENABLED(inst
, CHAN_Y
) ||
1427 IS_DST0_CHANNEL_ENABLED(inst
, CHAN_Z
) ||
1428 (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_W
) && dims
== 4)) {
1430 /* NOTE: Cannot use xmm regs 2/3 here (see emit_rsqrt() above). */
1433 /* xmm0 = src.x * src.x */
1434 tmp0
= emit_fetch(bld
, inst
, 0, CHAN_X
);
1435 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_X
)) {
1438 tmp0
= lp_build_mul( &bld
->base
, tmp0
, tmp0
);
1441 /* xmm0 = xmm0 + src.y * src.y */
1442 tmp1
= emit_fetch(bld
, inst
, 0, CHAN_Y
);
1443 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_Y
)) {
1446 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp1
);
1447 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
1450 /* xmm0 = xmm0 + src.z * src.z */
1451 tmp1
= emit_fetch(bld
, inst
, 0, CHAN_Z
);
1452 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_Z
)) {
1455 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp1
);
1456 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
1460 /* xmm0 = xmm0 + src.w * src.w */
1461 tmp1
= emit_fetch(bld
, inst
, 0, CHAN_W
);
1462 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_W
)) {
1465 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp1
);
1466 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
);
1469 /* xmm1 = 1 / sqrt(xmm0) */
1470 tmp1
= lp_build_rsqrt( &bld
->base
, tmp0
);
1472 /* dst.x = xmm1 * src.x */
1473 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_X
)) {
1474 dst0
[CHAN_X
] = lp_build_mul( &bld
->base
, tmp4
, tmp1
);
1477 /* dst.y = xmm1 * src.y */
1478 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_Y
)) {
1479 dst0
[CHAN_Y
] = lp_build_mul( &bld
->base
, tmp5
, tmp1
);
1482 /* dst.z = xmm1 * src.z */
1483 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_Z
)) {
1484 dst0
[CHAN_Z
] = lp_build_mul( &bld
->base
, tmp6
, tmp1
);
1487 /* dst.w = xmm1 * src.w */
1488 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_X
) && dims
== 4) {
1489 dst0
[CHAN_W
] = lp_build_mul( &bld
->base
, tmp7
, tmp1
);
1494 if (IS_DST0_CHANNEL_ENABLED(inst
, CHAN_W
) && dims
== 3) {
1495 dst0
[CHAN_W
] = bld
->base
.one
;
1500 case TGSI_OPCODE_DIV
:
1506 case TGSI_OPCODE_DP2
:
1507 tmp0
= emit_fetch( bld
, inst
, 0, CHAN_X
); /* xmm0 = src[0].x */
1508 tmp1
= emit_fetch( bld
, inst
, 1, CHAN_X
); /* xmm1 = src[1].x */
1509 tmp0
= lp_build_mul( &bld
->base
, tmp0
, tmp1
); /* xmm0 = xmm0 * xmm1 */
1510 tmp1
= emit_fetch( bld
, inst
, 0, CHAN_Y
); /* xmm1 = src[0].y */
1511 tmp2
= emit_fetch( bld
, inst
, 1, CHAN_Y
); /* xmm2 = src[1].y */
1512 tmp1
= lp_build_mul( &bld
->base
, tmp1
, tmp2
); /* xmm1 = xmm1 * xmm2 */
1513 tmp0
= lp_build_add( &bld
->base
, tmp0
, tmp1
); /* xmm0 = xmm0 + xmm1 */
1514 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1515 dst0
[chan_index
] = tmp0
; /* dest[ch] = xmm0 */
1519 case TGSI_OPCODE_TXL
:
1520 emit_tex( bld
, inst
, TRUE
, FALSE
, dst0
);
1523 case TGSI_OPCODE_TXP
:
1524 emit_tex( bld
, inst
, FALSE
, TRUE
, dst0
);
1527 case TGSI_OPCODE_BRK
:
1528 lp_exec_break(&bld
->exec_mask
);
1531 case TGSI_OPCODE_IF
:
1532 tmp0
= emit_fetch(bld
, inst
, 0, CHAN_X
);
1533 tmp0
= lp_build_cmp(&bld
->base
, PIPE_FUNC_NOTEQUAL
,
1534 tmp0
, bld
->base
.zero
);
1535 lp_exec_mask_cond_push(&bld
->exec_mask
, tmp0
);
1538 case TGSI_OPCODE_BGNFOR
:
1544 case TGSI_OPCODE_BGNLOOP
:
1545 lp_exec_bgnloop(&bld
->exec_mask
);
1548 case TGSI_OPCODE_REP
:
1554 case TGSI_OPCODE_ELSE
:
1555 lp_exec_mask_cond_invert(&bld
->exec_mask
);
1558 case TGSI_OPCODE_ENDIF
:
1559 lp_exec_mask_cond_pop(&bld
->exec_mask
);
1562 case TGSI_OPCODE_ENDFOR
:
1568 case TGSI_OPCODE_ENDLOOP
:
1569 lp_exec_endloop(&bld
->exec_mask
);
1572 case TGSI_OPCODE_ENDREP
:
1578 case TGSI_OPCODE_PUSHA
:
1584 case TGSI_OPCODE_POPA
:
1590 case TGSI_OPCODE_CEIL
:
1591 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1592 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1593 dst0
[chan_index
] = lp_build_ceil(&bld
->base
, tmp0
);
1597 case TGSI_OPCODE_I2F
:
1603 case TGSI_OPCODE_NOT
:
1609 case TGSI_OPCODE_TRUNC
:
1610 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1611 tmp0
= emit_fetch( bld
, inst
, 0, chan_index
);
1612 dst0
[chan_index
] = lp_build_trunc(&bld
->base
, tmp0
);
1616 case TGSI_OPCODE_SHL
:
1622 case TGSI_OPCODE_ISHR
:
1628 case TGSI_OPCODE_AND
:
1634 case TGSI_OPCODE_OR
:
1640 case TGSI_OPCODE_MOD
:
1646 case TGSI_OPCODE_XOR
:
1652 case TGSI_OPCODE_SAD
:
1658 case TGSI_OPCODE_TXF
:
1664 case TGSI_OPCODE_TXQ
:
1670 case TGSI_OPCODE_CONT
:
1671 lp_exec_continue(&bld
->exec_mask
);
1674 case TGSI_OPCODE_EMIT
:
1678 case TGSI_OPCODE_ENDPRIM
:
1682 case TGSI_OPCODE_NOP
:
1690 FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
1691 emit_store( bld
, inst
, 0, chan_index
, dst0
[chan_index
]);
1700 lp_build_tgsi_soa(LLVMBuilderRef builder
,
1701 const struct tgsi_token
*tokens
,
1702 struct lp_type type
,
1703 struct lp_build_mask_context
*mask
,
1704 LLVMValueRef consts_ptr
,
1705 const LLVMValueRef
*pos
,
1706 const LLVMValueRef (*inputs
)[NUM_CHANNELS
],
1707 LLVMValueRef (*outputs
)[NUM_CHANNELS
],
1708 struct lp_build_sampler_soa
*sampler
)
1710 struct lp_build_tgsi_soa_context bld
;
1711 struct tgsi_parse_context parse
;
1712 uint num_immediates
= 0;
1715 /* Setup build context */
1716 memset(&bld
, 0, sizeof bld
);
1717 lp_build_context_init(&bld
.base
, builder
, type
);
1720 bld
.inputs
= inputs
;
1721 bld
.outputs
= outputs
;
1722 bld
.consts_ptr
= consts_ptr
;
1723 bld
.sampler
= sampler
;
1725 lp_exec_mask_init(&bld
.exec_mask
, &bld
.base
);
1727 tgsi_parse_init( &parse
, tokens
);
1729 while( !tgsi_parse_end_of_tokens( &parse
) ) {
1730 tgsi_parse_token( &parse
);
1732 switch( parse
.FullToken
.Token
.Type
) {
1733 case TGSI_TOKEN_TYPE_DECLARATION
:
1734 /* Inputs already interpolated */
1736 if (!emit_declaration( &bld
, &parse
.FullToken
.FullDeclaration
))
1737 _debug_printf("warning: failed to define LLVM variable\n");
1741 case TGSI_TOKEN_TYPE_INSTRUCTION
:
1743 unsigned opcode
= parse
.FullToken
.FullInstruction
.Instruction
.Opcode
;
1744 const struct tgsi_opcode_info
*info
= tgsi_get_opcode_info(opcode
);
1745 if (!emit_instruction( &bld
, &parse
.FullToken
.FullInstruction
, info
))
1746 _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1747 info
? info
->mnemonic
: "<invalid>");
1752 case TGSI_TOKEN_TYPE_IMMEDIATE
:
1753 /* simply copy the immediate values into the next immediates[] slot */
1755 const uint size
= parse
.FullToken
.FullImmediate
.Immediate
.NrTokens
- 1;
1757 assert(num_immediates
< LP_MAX_IMMEDIATES
);
1758 for( i
= 0; i
< size
; ++i
)
1759 bld
.immediates
[num_immediates
][i
] =
1760 lp_build_const_vec(type
, parse
.FullToken
.FullImmediate
.u
[i
].Float
);
1761 for( i
= size
; i
< 4; ++i
)
1762 bld
.immediates
[num_immediates
][i
] = bld
.base
.undef
;
1767 case TGSI_TOKEN_TYPE_PROPERTY
:
1775 LLVMBasicBlockRef block
= LLVMGetInsertBlock(builder
);
1776 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
1777 debug_printf("11111111111111111111111111111 \n");
1778 tgsi_dump(tokens
, 0);
1779 LLVMDumpValue(function
);
1780 debug_printf("2222222222222222222222222222 \n");
1782 tgsi_parse_free( &parse
);