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
;
235 if (inst
->Instruction
.Saturate
) {
236 value
= lp_build_max(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.zero
);
237 value
= lp_build_min(&bld
->bld_base
.base
, value
, bld
->bld_base
.base
.one
);
241 * Translate the register file
244 assert(!reg
->Register
.Indirect
);
246 switch (reg
->Register
.File
) {
247 case TGSI_FILE_OUTPUT
:
248 ptr
= bld
->outputs
[reg
->Register
.Index
];
251 case TGSI_FILE_TEMPORARY
:
252 ptr
= bld
->temps
[reg
->Register
.Index
];
255 case TGSI_FILE_ADDRESS
:
256 ptr
= bld
->addr
[reg
->Indirect
.Index
];
259 case TGSI_FILE_PREDICATE
:
260 ptr
= bld
->preds
[reg
->Register
.Index
];
274 if (inst
->Instruction
.Predicate
) {
277 assert(inst
->Predicate
.Index
< LP_MAX_TGSI_PREDS
);
279 pred
= LLVMBuildLoad(builder
,
280 bld
->preds
[inst
->Predicate
.Index
], "");
283 * Convert the value to an integer mask.
285 pred
= lp_build_compare(bld
->bld_base
.base
.gallivm
,
286 bld
->bld_base
.base
.type
,
289 bld
->bld_base
.base
.zero
);
291 if (inst
->Predicate
.Negate
) {
292 pred
= LLVMBuildNot(builder
, pred
, "");
295 pred
= bld
->bld_base
.emit_swizzle(&bld
->bld_base
, pred
,
296 inst
->Predicate
.SwizzleX
,
297 inst
->Predicate
.SwizzleY
,
298 inst
->Predicate
.SwizzleZ
,
299 inst
->Predicate
.SwizzleW
);
302 mask
= LLVMBuildAnd(builder
, mask
, pred
, "");
312 if (reg
->Register
.WriteMask
!= TGSI_WRITEMASK_XYZW
) {
313 LLVMValueRef writemask
;
315 writemask
= lp_build_const_mask_aos_swizzled(bld
->bld_base
.base
.gallivm
,
316 bld
->bld_base
.base
.type
,
317 reg
->Register
.WriteMask
,
322 mask
= LLVMBuildAnd(builder
, mask
, writemask
, "");
329 LLVMValueRef orig_value
;
331 orig_value
= LLVMBuildLoad(builder
, ptr
, "");
332 value
= lp_build_select(&bld
->bld_base
.base
,
333 mask
, value
, orig_value
);
336 LLVMBuildStore(builder
, value
, ptr
);
341 * High-level instruction translators.
345 emit_tex(struct lp_build_tgsi_aos_context
*bld
,
346 const struct tgsi_full_instruction
*inst
,
347 enum lp_build_tex_modifier modifier
)
352 struct lp_derivatives derivs
= { {NULL
}, {NULL
} };
355 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
356 return bld
->bld_base
.base
.undef
;
359 target
= inst
->Texture
.Texture
;
361 coords
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0 , LP_CHAN_ALL
);
363 if (modifier
== LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
) {
364 /* probably not going to work */
365 derivs
.ddx
[0] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 1 , LP_CHAN_ALL
);
366 derivs
.ddy
[0] = lp_build_emit_fetch( &bld
->bld_base
, inst
, 2 , LP_CHAN_ALL
);
367 unit
= inst
->Src
[3].Register
.Index
;
370 unit
= inst
->Src
[1].Register
.Index
;
372 return bld
->sampler
->emit_fetch_texel(bld
->sampler
,
381 emit_sample(struct lp_build_tgsi_aos_context
*bld
,
382 const struct tgsi_full_instruction
*inst
,
383 enum lp_build_tex_modifier modifier
)
388 struct lp_derivatives derivs
= { {NULL
}, {NULL
} };
391 _debug_printf("warning: found texture instruction but no sampler generator supplied\n");
392 return bld
->bld_base
.base
.undef
;
395 coords
= lp_build_emit_fetch( &bld
->bld_base
, inst
, 0 , LP_CHAN_ALL
);
397 /* ignore modifiers, can't handle different sampler / sampler view, etc... */
398 unit
= inst
->Src
[1].Register
.Index
;
399 assert(inst
->Src
[2].Register
.Index
== unit
);
401 target
= bld
->sv
[unit
].Resource
;
403 return bld
->sampler
->emit_fetch_texel(bld
->sampler
,
412 lp_emit_declaration_aos(
413 struct lp_build_tgsi_aos_context
*bld
,
414 const struct tgsi_full_declaration
*decl
)
416 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
417 LLVMTypeRef vec_type
= lp_build_vec_type(bld
->bld_base
.base
.gallivm
, bld
->bld_base
.base
.type
);
419 unsigned first
= decl
->Range
.First
;
420 unsigned last
= decl
->Range
.Last
;
423 for (idx
= first
; idx
<= last
; ++idx
) {
424 switch (decl
->Declaration
.File
) {
425 case TGSI_FILE_TEMPORARY
:
426 assert(idx
< LP_MAX_INLINED_TEMPS
);
427 if (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
)) {
428 LLVMValueRef array_size
= lp_build_const_int32(gallivm
, last
+ 1);
429 bld
->temps_array
= lp_build_array_alloca(bld
->bld_base
.base
.gallivm
,
430 vec_type
, array_size
, "");
432 bld
->temps
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
436 case TGSI_FILE_OUTPUT
:
437 bld
->outputs
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
440 case TGSI_FILE_ADDRESS
:
441 assert(idx
< LP_MAX_TGSI_ADDRS
);
442 bld
->addr
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
445 case TGSI_FILE_PREDICATE
:
446 assert(idx
< LP_MAX_TGSI_PREDS
);
447 bld
->preds
[idx
] = lp_build_alloca(gallivm
, vec_type
, "");
450 case TGSI_FILE_SAMPLER_VIEW
:
452 * The target stored here MUST match whatever there actually
453 * is in the set sampler views (what about return type?).
455 assert(last
< PIPE_MAX_SHADER_SAMPLER_VIEWS
);
456 for (idx
= first
; idx
<= last
; ++idx
) {
457 bld
->sv
[idx
] = decl
->SamplerView
;
462 /* don't need to declare other vars */
470 * Emit LLVM for one TGSI instruction.
471 * \param return TRUE for success, FALSE otherwise
474 lp_emit_instruction_aos(
475 struct lp_build_tgsi_aos_context
*bld
,
476 const struct tgsi_full_instruction
*inst
,
477 const struct tgsi_opcode_info
*info
,
480 LLVMValueRef src0
, src1
, src2
;
482 LLVMValueRef dst0
= NULL
;
485 * Stores and write masks are handled in a general fashion after the long
486 * instruction opcode switch statement.
488 * Although not stricitly necessary, we avoid generating instructions for
489 * channels which won't be stored, in cases where's that easy. For some
490 * complex instructions, like texture sampling, it is more convenient to
491 * assume a full writemask and then let LLVM optimization passes eliminate
497 assert(info
->num_dst
<= 1);
499 dst0
= bld
->bld_base
.base
.undef
;
502 switch (inst
->Instruction
.Opcode
) {
503 case TGSI_OPCODE_ARL
:
504 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
505 dst0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
508 case TGSI_OPCODE_MOV
:
509 dst0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
512 case TGSI_OPCODE_LIT
:
515 case TGSI_OPCODE_RCP
:
516 /* TGSI_OPCODE_RECIP */
517 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
518 dst0
= lp_build_rcp(&bld
->bld_base
.base
, src0
);
521 case TGSI_OPCODE_RSQ
:
522 /* TGSI_OPCODE_RECIPSQRT */
523 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
524 tmp0
= lp_build_abs(&bld
->bld_base
.base
, src0
);
525 dst0
= lp_build_rsqrt(&bld
->bld_base
.base
, tmp0
);
528 case TGSI_OPCODE_EXP
:
531 case TGSI_OPCODE_LOG
:
534 case TGSI_OPCODE_MUL
:
535 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
536 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
537 dst0
= lp_build_mul(&bld
->bld_base
.base
, src0
, src1
);
540 case TGSI_OPCODE_ADD
:
541 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
542 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
543 dst0
= lp_build_add(&bld
->bld_base
.base
, src0
, src1
);
546 case TGSI_OPCODE_DP3
:
547 /* TGSI_OPCODE_DOT3 */
550 case TGSI_OPCODE_DP4
:
551 /* TGSI_OPCODE_DOT4 */
554 case TGSI_OPCODE_DST
:
557 case TGSI_OPCODE_MIN
:
558 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
559 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
560 dst0
= lp_build_min(&bld
->bld_base
.base
, src0
, src1
);
563 case TGSI_OPCODE_MAX
:
564 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
565 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
566 dst0
= lp_build_max(&bld
->bld_base
.base
, src0
, src1
);
569 case TGSI_OPCODE_SLT
:
570 /* TGSI_OPCODE_SETLT */
571 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
572 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
573 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LESS
, src0
, src1
);
574 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
577 case TGSI_OPCODE_SGE
:
578 /* TGSI_OPCODE_SETGE */
579 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
580 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
581 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GEQUAL
, src0
, src1
);
582 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
585 case TGSI_OPCODE_MAD
:
586 /* TGSI_OPCODE_MADD */
587 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
588 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
589 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
590 tmp0
= lp_build_mul(&bld
->bld_base
.base
, src0
, src1
);
591 dst0
= lp_build_add(&bld
->bld_base
.base
, tmp0
, src2
);
594 case TGSI_OPCODE_LRP
:
595 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
596 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
597 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
598 tmp0
= lp_build_sub(&bld
->bld_base
.base
, src1
, src2
);
599 tmp0
= lp_build_mul(&bld
->bld_base
.base
, src0
, tmp0
);
600 dst0
= lp_build_add(&bld
->bld_base
.base
, tmp0
, src2
);
603 case TGSI_OPCODE_DP2A
:
606 case TGSI_OPCODE_FRC
:
607 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
608 tmp0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
609 dst0
= lp_build_sub(&bld
->bld_base
.base
, src0
, tmp0
);
612 case TGSI_OPCODE_FLR
:
613 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
614 dst0
= lp_build_floor(&bld
->bld_base
.base
, src0
);
617 case TGSI_OPCODE_ROUND
:
618 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
619 dst0
= lp_build_round(&bld
->bld_base
.base
, src0
);
622 case TGSI_OPCODE_EX2
:
623 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
624 tmp0
= lp_build_swizzle_scalar_aos(&bld
->bld_base
.base
, src0
, TGSI_SWIZZLE_X
, TGSI_NUM_CHANNELS
);
625 dst0
= lp_build_exp2(&bld
->bld_base
.base
, tmp0
);
628 case TGSI_OPCODE_LG2
:
629 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
630 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
631 dst0
= lp_build_log2(&bld
->bld_base
.base
, tmp0
);
634 case TGSI_OPCODE_POW
:
635 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
636 src0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
637 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
638 src1
= swizzle_scalar_aos(bld
, src1
, TGSI_SWIZZLE_X
);
639 dst0
= lp_build_pow(&bld
->bld_base
.base
, src0
, src1
);
642 case TGSI_OPCODE_XPD
:
645 case TGSI_OPCODE_DPH
:
648 case TGSI_OPCODE_COS
:
649 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
650 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
651 dst0
= lp_build_cos(&bld
->bld_base
.base
, tmp0
);
654 case TGSI_OPCODE_DDX
:
657 case TGSI_OPCODE_DDY
:
660 case TGSI_OPCODE_KILL
:
663 case TGSI_OPCODE_KILL_IF
:
666 case TGSI_OPCODE_PK2H
:
670 case TGSI_OPCODE_PK2US
:
674 case TGSI_OPCODE_PK4B
:
678 case TGSI_OPCODE_PK4UB
:
681 case TGSI_OPCODE_SEQ
:
682 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
683 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
684 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_EQUAL
, src0
, src1
);
685 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
688 case TGSI_OPCODE_SGT
:
689 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
690 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
691 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_GREATER
, src0
, src1
);
692 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
695 case TGSI_OPCODE_SIN
:
696 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
697 tmp0
= swizzle_scalar_aos(bld
, src0
, TGSI_SWIZZLE_X
);
698 dst0
= lp_build_sin(&bld
->bld_base
.base
, tmp0
);
701 case TGSI_OPCODE_SLE
:
702 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
703 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
704 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LEQUAL
, src0
, src1
);
705 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
708 case TGSI_OPCODE_SNE
:
709 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
710 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
711 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_NOTEQUAL
, src0
, src1
);
712 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, bld
->bld_base
.base
.one
, bld
->bld_base
.base
.zero
);
715 case TGSI_OPCODE_TEX
:
716 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_NONE
);
719 case TGSI_OPCODE_TXD
:
720 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV
);
723 case TGSI_OPCODE_UP2H
:
729 case TGSI_OPCODE_UP2US
:
735 case TGSI_OPCODE_UP4B
:
741 case TGSI_OPCODE_UP4UB
:
747 case TGSI_OPCODE_ARR
:
748 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
749 dst0
= lp_build_round(&bld
->bld_base
.base
, src0
);
752 case TGSI_OPCODE_CAL
:
755 case TGSI_OPCODE_RET
:
756 /* safe to ignore at end */
759 case TGSI_OPCODE_END
:
763 case TGSI_OPCODE_SSG
:
764 /* TGSI_OPCODE_SGN */
765 tmp0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
766 dst0
= lp_build_sgn(&bld
->bld_base
.base
, tmp0
);
769 case TGSI_OPCODE_CMP
:
770 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
771 src1
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 1, LP_CHAN_ALL
);
772 src2
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 2, LP_CHAN_ALL
);
773 tmp0
= lp_build_cmp(&bld
->bld_base
.base
, PIPE_FUNC_LESS
, src0
, bld
->bld_base
.base
.zero
);
774 dst0
= lp_build_select(&bld
->bld_base
.base
, tmp0
, src1
, src2
);
777 case TGSI_OPCODE_SCS
:
780 case TGSI_OPCODE_TXB
:
781 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_LOD_BIAS
);
784 case TGSI_OPCODE_DIV
:
789 case TGSI_OPCODE_DP2
:
792 case TGSI_OPCODE_TXL
:
793 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD
);
796 case TGSI_OPCODE_TXP
:
797 dst0
= emit_tex(bld
, inst
, LP_BLD_TEX_MODIFIER_PROJECTED
);
800 case TGSI_OPCODE_BRK
:
804 case TGSI_OPCODE_UIF
:
807 case TGSI_OPCODE_BGNLOOP
:
810 case TGSI_OPCODE_BGNSUB
:
813 case TGSI_OPCODE_ELSE
:
816 case TGSI_OPCODE_ENDIF
:
819 case TGSI_OPCODE_ENDLOOP
:
822 case TGSI_OPCODE_ENDSUB
:
825 case TGSI_OPCODE_PUSHA
:
831 case TGSI_OPCODE_POPA
:
837 case TGSI_OPCODE_CEIL
:
838 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
839 dst0
= lp_build_ceil(&bld
->bld_base
.base
, src0
);
842 case TGSI_OPCODE_I2F
:
847 case TGSI_OPCODE_NOT
:
852 case TGSI_OPCODE_TRUNC
:
853 src0
= lp_build_emit_fetch(&bld
->bld_base
, inst
, 0, LP_CHAN_ALL
);
854 dst0
= lp_build_trunc(&bld
->bld_base
.base
, src0
);
857 case TGSI_OPCODE_SHL
:
862 case TGSI_OPCODE_ISHR
:
867 case TGSI_OPCODE_AND
:
877 case TGSI_OPCODE_MOD
:
882 case TGSI_OPCODE_XOR
:
887 case TGSI_OPCODE_SAD
:
892 case TGSI_OPCODE_TXF
:
897 case TGSI_OPCODE_TXQ
:
902 case TGSI_OPCODE_CONT
:
905 case TGSI_OPCODE_EMIT
:
909 case TGSI_OPCODE_ENDPRIM
:
913 case TGSI_OPCODE_NOP
:
916 case TGSI_OPCODE_SAMPLE
:
917 dst0
= emit_sample(bld
, inst
, LP_BLD_TEX_MODIFIER_NONE
);
925 lp_emit_store_aos(bld
, inst
, 0, dst0
);
933 lp_build_tgsi_aos(struct gallivm_state
*gallivm
,
934 const struct tgsi_token
*tokens
,
936 const unsigned char swizzles
[4],
937 LLVMValueRef consts_ptr
,
938 const LLVMValueRef
*inputs
,
939 LLVMValueRef
*outputs
,
940 struct lp_build_sampler_aos
*sampler
,
941 const struct tgsi_shader_info
*info
)
943 struct lp_build_tgsi_aos_context bld
;
944 struct tgsi_parse_context parse
;
945 uint num_immediates
= 0;
949 /* Setup build context */
950 memset(&bld
, 0, sizeof bld
);
951 lp_build_context_init(&bld
.bld_base
.base
, gallivm
, type
);
952 lp_build_context_init(&bld
.bld_base
.uint_bld
, gallivm
, lp_uint_type(type
));
953 lp_build_context_init(&bld
.bld_base
.int_bld
, gallivm
, lp_int_type(type
));
954 lp_build_context_init(&bld
.int_bld
, gallivm
, lp_int_type(type
));
956 for (chan
= 0; chan
< 4; ++chan
) {
957 bld
.swizzles
[chan
] = swizzles
[chan
];
958 bld
.inv_swizzles
[swizzles
[chan
]] = chan
;
962 bld
.outputs
= outputs
;
963 bld
.consts_ptr
= consts_ptr
;
964 bld
.sampler
= sampler
;
965 bld
.indirect_files
= info
->indirect_files
;
966 bld
.bld_base
.emit_swizzle
= swizzle_aos
;
967 bld
.bld_base
.info
= info
;
969 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_CONSTANT
] = emit_fetch_constant
;
970 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch_immediate
;
971 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch_input
;
972 bld
.bld_base
.emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch_temporary
;
974 /* Set opcode actions */
975 lp_set_default_actions_cpu(&bld
.bld_base
);
977 if (!lp_bld_tgsi_list_init(&bld
.bld_base
)) {
981 tgsi_parse_init(&parse
, tokens
);
983 while (!tgsi_parse_end_of_tokens(&parse
)) {
984 tgsi_parse_token(&parse
);
986 switch(parse
.FullToken
.Token
.Type
) {
987 case TGSI_TOKEN_TYPE_DECLARATION
:
988 /* Inputs already interpolated */
989 lp_emit_declaration_aos(&bld
, &parse
.FullToken
.FullDeclaration
);
992 case TGSI_TOKEN_TYPE_INSTRUCTION
:
993 /* save expanded instruction */
994 lp_bld_tgsi_add_instruction(&bld
.bld_base
,
995 &parse
.FullToken
.FullInstruction
);
998 case TGSI_TOKEN_TYPE_IMMEDIATE
:
999 /* simply copy the immediate values into the next immediates[] slot */
1001 const uint size
= parse
.FullToken
.FullImmediate
.Immediate
.NrTokens
- 1;
1004 assert(num_immediates
< LP_MAX_INLINED_IMMEDIATES
);
1005 for (chan
= 0; chan
< 4; ++chan
) {
1008 for (chan
= 0; chan
< size
; ++chan
) {
1009 unsigned swizzle
= bld
.swizzles
[chan
];
1010 imm
[swizzle
] = parse
.FullToken
.FullImmediate
.u
[chan
].Float
;
1012 bld
.immediates
[num_immediates
] =
1013 lp_build_const_aos(gallivm
, type
,
1014 imm
[0], imm
[1], imm
[2], imm
[3],
1020 case TGSI_TOKEN_TYPE_PROPERTY
:
1029 struct tgsi_full_instruction
*instr
= bld
.bld_base
.instructions
+ pc
;
1030 const struct tgsi_opcode_info
*opcode_info
=
1031 tgsi_get_opcode_info(instr
->Instruction
.Opcode
);
1032 if (!lp_emit_instruction_aos(&bld
, instr
, opcode_info
, &pc
))
1033 _debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
1034 opcode_info
->mnemonic
);
1038 LLVMBasicBlockRef block
= LLVMGetInsertBlock(gallivm
->builder
);
1039 LLVMValueRef function
= LLVMGetBasicBlockParent(block
);
1040 debug_printf("11111111111111111111111111111 \n");
1041 tgsi_dump(tokens
, 0);
1042 lp_debug_dump_value(function
);
1043 debug_printf("2222222222222222222222222222 \n");
1045 tgsi_parse_free(&parse
);
1046 FREE(bld
.bld_base
.instructions
);
1049 LLVMModuleRef module
= LLVMGetGlobalParent(
1050 LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm
->builder
)));
1051 LLVMDumpModule(module
);