1 /**************************************************************************
3 * Copyright 2010 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
30 * TGSI to LLVM IR translation -- AoS.
33 * - No control flow support: the existing control flow code should be factored
34 * out into from the SoA code into a common module and shared.
35 * - No derivatives. Derivate logic should be pluggable, just like the samplers.
37 * @author Jose Fonseca <jfonseca@vmware.com>
40 #include "pipe/p_config.h"
41 #include "pipe/p_shader_tokens.h"
42 #include "util/u_debug.h"
43 #include "util/u_math.h"
44 #include "util/u_memory.h"
45 #include "tgsi/tgsi_dump.h"
46 #include "tgsi/tgsi_info.h"
47 #include "tgsi/tgsi_parse.h"
48 #include "tgsi/tgsi_util.h"
49 #include "tgsi/tgsi_scan.h"
50 #include "lp_bld_type.h"
51 #include "lp_bld_const.h"
52 #include "lp_bld_arit.h"
53 #include "lp_bld_logic.h"
54 #include "lp_bld_swizzle.h"
55 #include "lp_bld_flow.h"
56 #include "lp_bld_quad.h"
57 #include "lp_bld_tgsi.h"
58 #include "lp_bld_debug.h"
59 #include "lp_bld_sample.h"
63 * Wrapper around lp_build_swizzle_aos which translates swizzles to another
67 swizzle_aos(struct lp_build_tgsi_context
*bld_base
,
74 unsigned char swizzles
[4];
75 struct lp_build_tgsi_aos_context
*bld
= lp_aos_context(bld_base
);
77 assert(swizzle_x
< 4);
78 assert(swizzle_y
< 4);
79 assert(swizzle_z
< 4);
80 assert(swizzle_w
< 4);
82 swizzles
[bld
->inv_swizzles
[0]] = bld
->swizzles
[swizzle_x
];
83 swizzles
[bld
->inv_swizzles
[1]] = bld
->swizzles
[swizzle_y
];
84 swizzles
[bld
->inv_swizzles
[2]] = bld
->swizzles
[swizzle_z
];
85 swizzles
[bld
->inv_swizzles
[3]] = bld
->swizzles
[swizzle_w
];
87 return lp_build_swizzle_aos(&bld
->bld_base
.base
, a
, swizzles
);
92 swizzle_scalar_aos(struct lp_build_tgsi_aos_context
*bld
,
96 chan
= bld
->swizzles
[chan
];
97 return lp_build_swizzle_scalar_aos(&bld
->bld_base
.base
, a
, chan
, 4);
103 struct lp_build_tgsi_context
* bld_base
,
104 const struct tgsi_full_src_register
* reg
,
105 enum tgsi_opcode_type stype
,
108 struct lp_build_tgsi_aos_context
* bld
= lp_aos_context(bld_base
);
109 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
110 struct lp_type type
= bld_base
->base
.type
;
114 assert(!reg
->Register
.Indirect
);
117 * Get the constants components
120 res
= bld
->bld_base
.base
.undef
;
121 for (chan
= 0; chan
< 4; ++chan
) {
123 LLVMValueRef scalar_ptr
;
125 LLVMValueRef swizzle
;
127 index
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
,
128 reg
->Register
.Index
* 4 + chan
);
130 scalar_ptr
= LLVMBuildGEP(builder
, bld
->consts_ptr
, &index
, 1, "");
132 scalar
= LLVMBuildLoad(builder
, scalar_ptr
, "");
134 lp_build_name(scalar
, "const[%u].%c", reg
->Register
.Index
, "xyzw"[chan
]);
137 * NOTE: constants array is always assumed to be RGBA
140 swizzle
= lp_build_const_int32(bld
->bld_base
.base
.gallivm
,
141 bld
->swizzles
[chan
]);
143 res
= LLVMBuildInsertElement(builder
, res
, scalar
, swizzle
, "");
147 * Broadcast the first quaternion to all others.
149 * XXX: could be factored into a reusable function.
152 if (type
.length
> 4) {
153 LLVMValueRef shuffles
[LP_MAX_VECTOR_LENGTH
];
156 for (chan
= 0; chan
< 4; ++chan
) {
157 shuffles
[chan
] = lp_build_const_int32(bld
->bld_base
.base
.gallivm
, chan
);
160 for (i
= 4; i
< type
.length
; ++i
) {
161 shuffles
[i
] = shuffles
[i
% 4];
164 res
= LLVMBuildShuffleVector(builder
,
165 res
, bld
->bld_base
.base
.undef
,
166 LLVMConstVector(shuffles
, type
.length
),
173 emit_fetch_immediate(
174 struct lp_build_tgsi_context
* bld_base
,
175 const struct tgsi_full_src_register
* reg
,
176 enum tgsi_opcode_type stype
,
179 struct lp_build_tgsi_aos_context
* bld
= lp_aos_context(bld_base
);
180 LLVMValueRef res
= bld
->immediates
[reg
->Register
.Index
];
187 struct lp_build_tgsi_context
* bld_base
,
188 const struct tgsi_full_src_register
* reg
,
189 enum tgsi_opcode_type stype
,
192 struct lp_build_tgsi_aos_context
* bld
= lp_aos_context(bld_base
);
193 LLVMValueRef res
= bld
->inputs
[reg
->Register
.Index
];
194 assert(!reg
->Register
.Indirect
);
200 emit_fetch_temporary(
201 struct lp_build_tgsi_context
* bld_base
,
202 const struct tgsi_full_src_register
* reg
,
203 enum tgsi_opcode_type stype
,
206 struct lp_build_tgsi_aos_context
* bld
= lp_aos_context(bld_base
);
207 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
208 LLVMValueRef temp_ptr
= bld
->temps
[reg
->Register
.Index
];
209 LLVMValueRef res
= LLVMBuildLoad(builder
, temp_ptr
, "");
210 assert(!reg
->Register
.Indirect
);
212 return bld
->bld_base
.base
.undef
;
222 struct lp_build_tgsi_aos_context
*bld
,
223 const struct tgsi_full_instruction
*inst
,
227 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
228 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[index
];
229 LLVMValueRef mask
= NULL
;
236 switch (inst
->Instruction
.Saturate
) {
240 case TGSI_SAT_ZERO_ONE
:
241 value
= lp_build_max(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.zero
);
242 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
245 case TGSI_SAT_MINUS_PLUS_ONE
:
246 value
= lp_build_max(&bld
->bld_base
.base
, value
, lp_build_const_vec(bld
->bld_base
.base
.gallivm
, bld
->bld_base
.base
.type
, -1.0));
247 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
255 * Translate the register file
258 assert(!reg
->Register
.Indirect
);
260 switch (reg
->Register
.File
) {
261 case TGSI_FILE_OUTPUT
:
262 ptr
= bld
->outputs
[reg
->Register
.Index
];
265 case TGSI_FILE_TEMPORARY
:
266 ptr
= bld
->temps
[reg
->Register
.Index
];
269 case TGSI_FILE_ADDRESS
:
270 ptr
= bld
->addr
[reg
->Indirect
.Index
];
273 case TGSI_FILE_PREDICATE
:
274 ptr
= bld
->preds
[reg
->Register
.Index
];
288 if (inst
->Instruction
.Predicate
) {
291 assert(inst
->Predicate
.Index
< LP_MAX_TGSI_PREDS
);
293 pred
= LLVMBuildLoad(builder
,
294 bld
->preds
[inst
->Predicate
.Index
], "");
297 * Convert the value to an integer mask.
299 pred
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
300 bld
->bld_base
.base
.type
,
303 bld
->bld_base
.base
.zero
);
305 if (inst
->Predicate
.Negate
) {
306 pred
= LLVMBuildNot(builder
, pred
, "");
309 pred
= bld
->bld_base
.emit_swizzle(&bld
->bld_base
, pred
,
310 inst
->Predicate
.SwizzleX
,
311 inst
->Predicate
.SwizzleY
,
312 inst
->Predicate
.SwizzleZ
,
313 inst
->Predicate
.SwizzleW
);
316 mask
= LLVMBuildAnd(builder
, mask
, pred
, "");
326 if (reg
->Register
.WriteMask
!= TGSI_WRITEMASK_XYZW
) {
327 LLVMValueRef writemask
;
329 writemask
= lp_build_const_mask_aos_swizzled(bld
->bld_base
.base
.gallivm
,
330 bld
->bld_base
.base
.type
,
331 reg
->Register
.WriteMask
,
336 mask
= LLVMBuildAnd(builder
, mask
, writemask
, "");
343 LLVMValueRef orig_value
;
345 orig_value
= LLVMBuildLoad(builder
, ptr
, "");
346 value
= lp_build_select(&bld
->bld_base
.base
,
347 mask
, value
, orig_value
);
350 LLVMBuildStore(builder
, value
, ptr
);
355 * High-level instruction translators.
359 emit_tex(struct lp_build_tgsi_aos_context
*bld
,
360 const struct tgsi_full_instruction
*inst
,
361 enum lp_build_tex_modifier modifier
)
366 struct lp_derivatives derivs
= { {NULL
}, {NULL
} };
369 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
370 return bld
->bld_base
.base
.undef
;
373 target
= inst
->Texture
.Texture
;
375 coords
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0 , LP_CHAN_ALL
);
377 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
378 /* probably not going to work */
379 derivs
.ddx
[0] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 1 , LP_CHAN_ALL
);
380 derivs
.ddy
[0] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 2 , LP_CHAN_ALL
);
381 unit
= inst
->Src
[3].Register
.Index
;
384 unit
= inst
->Src
[1].Register
.Index
;
386 return bld
->sampler
->emit_fetch_texel(bld
->sampler
,
395 emit_sample(struct lp_build_tgsi_aos_context
*bld
,
396 const struct tgsi_full_instruction
*inst
,
397 enum lp_build_tex_modifier modifier
)
402 struct lp_derivatives derivs
= { {NULL
}, {NULL
} };
405 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
406 return bld
->bld_base
.base
.undef
;
409 coords
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0 , LP_CHAN_ALL
);
411 /* ignore modifiers, can't handle different sampler / sampler view, etc... */
412 unit
= inst
->Src
[1].Register
.Index
;
413 assert(inst
->Src
[2].Register
.Index
== unit
);
415 target
= bld
->sv
[unit
].Resource
;
417 return bld
->sampler
->emit_fetch_texel(bld
->sampler
,
426 lp_emit_declaration_aos(
427 struct lp_build_tgsi_aos_context
*bld
,
428 const struct tgsi_full_declaration
*decl
)
430 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
431 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->bld_base
.base
.gallivm
, bld
->bld_base
.base
.type
);
433 unsigned first
= decl
->Range
.First
;
434 unsigned last
= decl
->Range
.Last
;
437 for (idx
= first
; idx
<= last
; ++idx
) {
438 switch (decl
->Declaration
.File
) {
439 case TGSI_FILE_TEMPORARY
:
440 assert(idx
< LP_MAX_INLINED_TEMPS
);
441 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
442 LLVMValueRef array_size
= lp_build_const_int32(gallivm
, last
+ 1);
443 bld
->temps_array
= lp_build_array_alloca(bld
->bld_base
.base
.gallivm
,
444 vec_type
, array_size
, "");
446 bld
->temps
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
450 case TGSI_FILE_OUTPUT
:
451 bld
->outputs
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
454 case TGSI_FILE_ADDRESS
:
455 assert(idx
< LP_MAX_TGSI_ADDRS
);
456 bld
->addr
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
459 case TGSI_FILE_PREDICATE
:
460 assert(idx
< LP_MAX_TGSI_PREDS
);
461 bld
->preds
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
464 case TGSI_FILE_SAMPLER_VIEW
:
466 * The target stored here MUST match whatever there actually
467 * is in the set sampler views (what about return type?).
469 assert(last
< PIPE_MAX_SHADER_SAMPLER_VIEWS
);
470 for (idx
= first
; idx
<= last
; ++idx
) {
471 bld
->sv
[idx
] = decl
->SamplerView
;
476 /* don't need to declare other vars */
484 * Emit LLVM for one TGSI instruction.
485 * \param return TRUE for success, FALSE otherwise
488 lp_emit_instruction_aos(
489 struct lp_build_tgsi_aos_context
*bld
,
490 const struct tgsi_full_instruction
*inst
,
491 const struct tgsi_opcode_info
*info
,
494 LLVMValueRef src0
, src1
, src2
;
495 LLVMValueRef tmp0
, tmp1
;
496 LLVMValueRef dst0
= NULL
;
499 * Stores and write masks are handled in a general fashion after the long
500 * instruction opcode switch statement.
502 * Although not stricitly necessary, we avoid generating instructions for
503 * channels which won't be stored, in cases where's that easy. For some
504 * complex instructions, like texture sampling, it is more convenient to
505 * assume a full writemask and then let LLVM optimization passes eliminate
511 assert(info
->num_dst
<= 1);
513 dst0
= bld
->bld_base
.base
.undef
;
516 switch (inst
->Instruction
.Opcode
) {
517 case TGSI_OPCODE_ARL
:
518 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
519 dst0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
522 case TGSI_OPCODE_MOV
:
523 dst0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
526 case TGSI_OPCODE_LIT
:
529 case TGSI_OPCODE_RCP
:
530 /* TGSI_OPCODE_RECIP */
531 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
532 dst0
= lp_build_rcp(&bld
->bld_base
.base
, src0
);
535 case TGSI_OPCODE_RSQ
:
536 /* TGSI_OPCODE_RECIPSQRT */
537 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
538 tmp0
= lp_build_emit_llvm_unary(&bld
->bld_base
, TGSI_OPCODE_ABS
, src0
);
539 dst0
= lp_build_rsqrt(&bld
->bld_base
.base
, tmp0
);
542 case TGSI_OPCODE_EXP
:
545 case TGSI_OPCODE_LOG
:
548 case TGSI_OPCODE_MUL
:
549 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
550 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
551 dst0
= lp_build_mul(&bld
->bld_base
.base
, src0
, src1
);
554 case TGSI_OPCODE_ADD
:
555 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
556 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
557 dst0
= lp_build_add(&bld
->bld_base
.base
, src0
, src1
);
560 case TGSI_OPCODE_DP3
:
561 /* TGSI_OPCODE_DOT3 */
564 case TGSI_OPCODE_DP4
:
565 /* TGSI_OPCODE_DOT4 */
568 case TGSI_OPCODE_DST
:
571 case TGSI_OPCODE_MIN
:
572 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
573 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
574 dst0
= lp_build_min(&bld
->bld_base
.base
, src0
, src1
);
577 case TGSI_OPCODE_MAX
:
578 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
579 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
580 dst0
= lp_build_max(&bld
->bld_base
.base
, src0
, src1
);
583 case TGSI_OPCODE_SLT
:
584 /* TGSI_OPCODE_SETLT */
585 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
586 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
587 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LESS
, src0
, src1
);
588 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
591 case TGSI_OPCODE_SGE
:
592 /* TGSI_OPCODE_SETGE */
593 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
594 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
595 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, src0
, src1
);
596 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
599 case TGSI_OPCODE_MAD
:
600 /* TGSI_OPCODE_MADD */
601 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
602 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
603 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
604 tmp0
= lp_build_mul(&bld
->bld_base
.base
, src0
, src1
);
605 dst0
= lp_build_add(&bld
->bld_base
.base
, tmp0
, src2
);
608 case TGSI_OPCODE_SUB
:
609 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
610 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
611 dst0
= lp_build_sub(&bld
->bld_base
.base
, src0
, src1
);
614 case TGSI_OPCODE_LRP
:
615 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
616 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
617 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
618 tmp0
= lp_build_sub(&bld
->bld_base
.base
, src1
, src2
);
619 tmp0
= lp_build_mul(&bld
->bld_base
.base
, src0
, tmp0
);
620 dst0
= lp_build_add(&bld
->bld_base
.base
, tmp0
, src2
);
623 case TGSI_OPCODE_CND
:
624 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
625 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
626 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
627 tmp1
= lp_build_const_vec(bld
->bld_base
.base
.gallivm
, bld
->bld_base
.base
.type
, 0.5);
628 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GREATER
, src2
, tmp1
);
629 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, src0
, src1
);
632 case TGSI_OPCODE_DP2A
:
635 case TGSI_OPCODE_FRC
:
636 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
637 tmp0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
638 dst0
= lp_build_sub(&bld
->bld_base
.base
, src0
, tmp0
);
641 case TGSI_OPCODE_CLAMP
:
642 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
643 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
644 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
645 tmp0
= lp_build_max(&bld
->bld_base
.base
, src0
, src1
);
646 dst0
= lp_build_min(&bld
->bld_base
.base
, tmp0
, src2
);
649 case TGSI_OPCODE_FLR
:
650 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
651 dst0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
654 case TGSI_OPCODE_ROUND
:
655 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
656 dst0
= lp_build_round(&bld
->bld_base
.base
, src0
);
659 case TGSI_OPCODE_EX2
:
660 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
661 tmp0
= lp_build_swizzle_scalar_aos(&bld
->bld_base
.base
, src0
, TGSI_SWIZZLE_X
, TGSI_NUM_CHANNELS
);
662 dst0
= lp_build_exp2(&bld
->bld_base
.base
, tmp0
);
665 case TGSI_OPCODE_LG2
:
666 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
667 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
668 dst0
= lp_build_log2(&bld
->bld_base
.base
, tmp0
);
671 case TGSI_OPCODE_POW
:
672 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
673 src0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
674 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
675 src1
= swizzle_scalar_aos(bld
, src1
, TGSI_SWIZZLE_X
);
676 dst0
= lp_build_pow(&bld
->bld_base
.base
, src0
, src1
);
679 case TGSI_OPCODE_XPD
:
682 case TGSI_OPCODE_DPH
:
685 case TGSI_OPCODE_COS
:
686 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
687 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
688 dst0
= lp_build_cos(&bld
->bld_base
.base
, tmp0
);
691 case TGSI_OPCODE_DDX
:
694 case TGSI_OPCODE_DDY
:
697 case TGSI_OPCODE_KILL
:
700 case TGSI_OPCODE_KILL_IF
:
703 case TGSI_OPCODE_PK2H
:
707 case TGSI_OPCODE_PK2US
:
711 case TGSI_OPCODE_PK4B
:
715 case TGSI_OPCODE_PK4UB
:
718 case TGSI_OPCODE_SEQ
:
719 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
720 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
721 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_EQUAL
, src0
, src1
);
722 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
725 case TGSI_OPCODE_SFL
:
726 dst0
= bld
->bld_base
.base
.zero
;
729 case TGSI_OPCODE_SGT
:
730 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
731 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
732 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GREATER
, src0
, src1
);
733 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
736 case TGSI_OPCODE_SIN
:
737 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
738 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
739 dst0
= lp_build_sin(&bld
->bld_base
.base
, tmp0
);
742 case TGSI_OPCODE_SLE
:
743 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
744 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
745 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LEQUAL
, src0
, src1
);
746 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
749 case TGSI_OPCODE_SNE
:
750 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
751 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
752 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_NOTEQUAL
, src0
, src1
);
753 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
756 case TGSI_OPCODE_STR
:
757 dst0
= bld
->bld_base
.base
.one
;
760 case TGSI_OPCODE_TEX
:
761 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_NONE
);
764 case TGSI_OPCODE_TXD
:
765 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
);
768 case TGSI_OPCODE_UP2H
:
774 case TGSI_OPCODE_UP2US
:
780 case TGSI_OPCODE_UP4B
:
786 case TGSI_OPCODE_UP4UB
:
792 case TGSI_OPCODE_ARR
:
793 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
794 dst0
= lp_build_round(&bld
->bld_base
.base
, src0
);
797 case TGSI_OPCODE_BRA
:
803 case TGSI_OPCODE_CAL
:
806 case TGSI_OPCODE_RET
:
807 /* safe to ignore at end */
810 case TGSI_OPCODE_END
:
814 case TGSI_OPCODE_SSG
:
815 /* TGSI_OPCODE_SGN */
816 tmp0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
817 dst0
= lp_build_sgn(&bld
->bld_base
.base
, tmp0
);
820 case TGSI_OPCODE_CMP
:
821 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
822 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
823 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
824 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LESS
, src0
, bld
->bld_base
.base
.zero
);
825 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, src1
, src2
);
828 case TGSI_OPCODE_SCS
:
831 case TGSI_OPCODE_TXB
:
832 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
);
835 case TGSI_OPCODE_DIV
:
840 case TGSI_OPCODE_DP2
:
843 case TGSI_OPCODE_TXL
:
844 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
);
847 case TGSI_OPCODE_TXP
:
848 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_PROJECTED
);
851 case TGSI_OPCODE_BRK
:
855 case TGSI_OPCODE_UIF
:
858 case TGSI_OPCODE_BGNLOOP
:
861 case TGSI_OPCODE_BGNSUB
:
864 case TGSI_OPCODE_ELSE
:
867 case TGSI_OPCODE_ENDIF
:
870 case TGSI_OPCODE_ENDLOOP
:
873 case TGSI_OPCODE_ENDSUB
:
876 case TGSI_OPCODE_PUSHA
:
882 case TGSI_OPCODE_POPA
:
888 case TGSI_OPCODE_CEIL
:
889 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
890 dst0
= lp_build_ceil(&bld
->bld_base
.base
, src0
);
893 case TGSI_OPCODE_I2F
:
898 case TGSI_OPCODE_NOT
:
903 case TGSI_OPCODE_TRUNC
:
904 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
905 dst0
= lp_build_trunc(&bld
->bld_base
.base
, src0
);
908 case TGSI_OPCODE_SHL
:
913 case TGSI_OPCODE_ISHR
:
918 case TGSI_OPCODE_AND
:
928 case TGSI_OPCODE_MOD
:
933 case TGSI_OPCODE_XOR
:
938 case TGSI_OPCODE_SAD
:
943 case TGSI_OPCODE_TXF
:
948 case TGSI_OPCODE_TXQ
:
953 case TGSI_OPCODE_CONT
:
956 case TGSI_OPCODE_EMIT
:
960 case TGSI_OPCODE_ENDPRIM
:
964 case TGSI_OPCODE_NOP
:
967 case TGSI_OPCODE_SAMPLE
:
968 dst0
= emit_sample(bld
, inst
, LP_BLD_TEX_MODIFIER_NONE
);
976 lp_emit_store_aos(bld
, inst
, 0, dst0
);
984 lp_build_tgsi_aos(struct gallivm_state
*gallivm
,
985 const struct tgsi_token
*tokens
,
987 const unsigned char swizzles
[4],
988 LLVMValueRef consts_ptr
,
989 const LLVMValueRef
*inputs
,
990 LLVMValueRef
*outputs
,
991 struct lp_build_sampler_aos
*sampler
,
992 const struct tgsi_shader_info
*info
)
994 struct lp_build_tgsi_aos_context bld
;
995 struct tgsi_parse_context parse
;
996 uint num_immediates
= 0;
1000 /* Setup build context */
1001 memset(&bld
, 0, sizeof bld
);
1002 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
1003 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
1004 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
1005 lp_build_context_init(&bld
.int_bld
, gallivm
, lp_int_type(type
));
1007 for (chan
= 0; chan
< 4; ++chan
) {
1008 bld
.swizzles
[chan
] = swizzles
[chan
];
1009 bld
.inv_swizzles
[swizzles
[chan
]] = chan
;
1012 bld
.inputs
= inputs
;
1013 bld
.outputs
= outputs
;
1014 bld
.consts_ptr
= consts_ptr
;
1015 bld
.sampler
= sampler
;
1016 bld
.indirect_files
= info
->indirect_files
;
1017 bld
.bld_base
.emit_swizzle
= swizzle_aos
;
1018 bld
.bld_base
.info
= info
;
1020 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
1021 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
1022 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
1023 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
1025 /* Set opcode actions */
1026 lp_set_default_actions_cpu(&bld
.bld_base
);
1028 if (!lp_bld_tgsi_list_init(&bld
.bld_base
)) {
1032 tgsi_parse_init(&parse
, tokens
);
1034 while (!tgsi_parse_end_of_tokens(&parse
)) {
1035 tgsi_parse_token(&parse
);
1037 switch(parse
.FullToken
.Token
.Type
) {
1038 case TGSI_TOKEN_TYPE_DECLARATION
:
1039 /* Inputs already interpolated */
1040 lp_emit_declaration_aos(&bld
, &parse
.FullToken
.FullDeclaration
);
1043 case TGSI_TOKEN_TYPE_INSTRUCTION
:
1044 /* save expanded instruction */
1045 lp_bld_tgsi_add_instruction(&bld
.bld_base
,
1046 &parse
.FullToken
.FullInstruction
);
1049 case TGSI_TOKEN_TYPE_IMMEDIATE
:
1050 /* simply copy the immediate values into the next immediates[] slot */
1052 const uint size
= parse
.FullToken
.FullImmediate
.Immediate
.NrTokens
- 1;
1055 assert(num_immediates
< LP_MAX_INLINED_IMMEDIATES
);
1056 for (chan
= 0; chan
< 4; ++chan
) {
1059 for (chan
= 0; chan
< size
; ++chan
) {
1060 unsigned swizzle
= bld
.swizzles
[chan
];
1061 imm
[swizzle
] = parse
.FullToken
.FullImmediate
.u
[chan
].Float
;
1063 bld
.immediates
[num_immediates
] =
1064 lp_build_const_aos(gallivm
, type
,
1065 imm
[0], imm
[1], imm
[2], imm
[3],
1071 case TGSI_TOKEN_TYPE_PROPERTY
:
1080 struct tgsi_full_instruction
*instr
= bld
.bld_base
.instructions
+ pc
;
1081 const struct tgsi_opcode_info
*opcode_info
=
1082 tgsi_get_opcode_info(instr
->Instruction
.Opcode
);
1083 if (!lp_emit_instruction_aos(&bld
, instr
, opcode_info
, &pc
))
1084 _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1085 opcode_info
->mnemonic
);
1089 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
1090 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
1091 debug_printf("11111111111111111111111111111 \n");
1092 tgsi_dump(tokens
, 0);
1093 lp_debug_dump_value(function
);
1094 debug_printf("2222222222222222222222222222 \n");
1096 tgsi_parse_free(&parse
);
1097 FREE(bld
.bld_base
.instructions
);
1100 LLVMModuleRef module
= LLVMGetGlobalParent(
1101 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
1102 LLVMDumpModule(module
);