2 * Copyright 2011 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * Authors: Tom Stellard <thomas.stellard@amd.com>
26 #include "radeon_llvm.h"
28 #include "gallivm/lp_bld_const.h"
29 #include "gallivm/lp_bld_gather.h"
30 #include "gallivm/lp_bld_flow.h"
31 #include "gallivm/lp_bld_init.h"
32 #include "gallivm/lp_bld_intr.h"
33 #include "gallivm/lp_bld_swizzle.h"
34 #include "tgsi/tgsi_info.h"
35 #include "tgsi/tgsi_parse.h"
36 #include "util/u_math.h"
37 #include "util/u_memory.h"
38 #include "util/u_debug.h"
40 #include <llvm-c/Core.h>
41 #include <llvm-c/Transforms/Scalar.h>
43 static struct radeon_llvm_loop
* get_current_loop(struct radeon_llvm_context
* ctx
)
45 return ctx
->loop_depth
> 0 ? ctx
->loop
+ (ctx
->loop_depth
- 1) : NULL
;
48 static struct radeon_llvm_branch
* get_current_branch(
49 struct radeon_llvm_context
* ctx
)
51 return ctx
->branch_depth
> 0 ?
52 ctx
->branch
+ (ctx
->branch_depth
- 1) : NULL
;
55 unsigned radeon_llvm_reg_index_soa(unsigned index
, unsigned chan
)
57 return (index
* 4) + chan
;
60 static LLVMValueRef
emit_swizzle(
61 struct lp_build_tgsi_context
* bld_base
,
68 LLVMValueRef swizzles
[4];
70 LLVMInt32TypeInContext(bld_base
->base
.gallivm
->context
);
72 swizzles
[0] = LLVMConstInt(i32t
, swizzle_x
, 0);
73 swizzles
[1] = LLVMConstInt(i32t
, swizzle_y
, 0);
74 swizzles
[2] = LLVMConstInt(i32t
, swizzle_z
, 0);
75 swizzles
[3] = LLVMConstInt(i32t
, swizzle_w
, 0);
77 return LLVMBuildShuffleVector(bld_base
->base
.gallivm
->builder
,
79 LLVMGetUndef(LLVMTypeOf(value
)),
80 LLVMConstVector(swizzles
, 4), "");
83 static struct tgsi_declaration_range
84 get_array_range(struct lp_build_tgsi_context
*bld_base
,
85 unsigned File
, const struct tgsi_ind_register
*reg
)
87 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
88 if (File
!= TGSI_FILE_TEMPORARY
|| reg
->ArrayID
== 0 ||
89 reg
->ArrayID
> RADEON_LLVM_MAX_ARRAYS
) {
90 struct tgsi_declaration_range range
;
92 range
.Last
= bld_base
->info
->file_max
[File
];
96 return ctx
->arrays
[reg
->ArrayID
- 1];
101 struct lp_build_tgsi_soa_context
*bld
,
102 const struct tgsi_ind_register
*reg
,
105 struct gallivm_state
* gallivm
= bld
->bld_base
.base
.gallivm
;
107 LLVMValueRef addr
= LLVMBuildLoad(gallivm
->builder
, bld
->addr
[reg
->Index
][reg
->Swizzle
], "");
108 return LLVMBuildAdd(gallivm
->builder
, addr
, lp_build_const_int32(gallivm
, offset
), "");
113 struct lp_build_tgsi_context
*bld_base
,
114 const struct tgsi_full_src_register
*reg
,
115 enum tgsi_opcode_type type
,
120 struct lp_build_tgsi_context
*bld_base
,
121 unsigned File
, enum tgsi_opcode_type type
,
122 struct tgsi_declaration_range range
,
125 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
126 struct gallivm_state
* gallivm
= bld
->bld_base
.base
.gallivm
;
127 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
129 unsigned i
, size
= range
.Last
- range
.First
+ 1;
130 LLVMTypeRef vec
= LLVMVectorType(tgsi2llvmtype(bld_base
, type
), size
);
131 LLVMValueRef result
= LLVMGetUndef(vec
);
133 struct tgsi_full_src_register tmp_reg
= {};
134 tmp_reg
.Register
.File
= File
;
136 for (i
= 0; i
< size
; ++i
) {
137 tmp_reg
.Register
.Index
= i
+ range
.First
;
138 LLVMValueRef temp
= emit_fetch(bld_base
, &tmp_reg
, type
, swizzle
);
139 result
= LLVMBuildInsertElement(builder
, result
, temp
,
140 lp_build_const_int32(gallivm
, i
), "");
147 struct lp_build_tgsi_context
*bld_base
,
148 const struct tgsi_full_src_register
*reg
,
149 enum tgsi_opcode_type type
,
152 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
153 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
154 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
155 LLVMValueRef result
, ptr
;
158 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
160 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
161 values
[chan
] = emit_fetch(bld_base
, reg
, type
, chan
);
163 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
167 if (reg
->Register
.Indirect
) {
168 struct tgsi_declaration_range range
= get_array_range(bld_base
,
169 reg
->Register
.File
, ®
->Indirect
);
170 return LLVMBuildExtractElement(builder
,
171 emit_array_fetch(bld_base
, reg
->Register
.File
, type
, range
, swizzle
),
172 emit_array_index(bld
, ®
->Indirect
, reg
->Register
.Index
- range
.First
),
176 switch(reg
->Register
.File
) {
177 case TGSI_FILE_IMMEDIATE
: {
178 LLVMTypeRef ctype
= tgsi2llvmtype(bld_base
, type
);
179 return LLVMConstBitCast(bld
->immediates
[reg
->Register
.Index
][swizzle
], ctype
);
182 case TGSI_FILE_INPUT
:
183 result
= ctx
->inputs
[radeon_llvm_reg_index_soa(reg
->Register
.Index
, swizzle
)];
186 case TGSI_FILE_TEMPORARY
:
187 ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
188 result
= LLVMBuildLoad(builder
, ptr
, "");
191 case TGSI_FILE_OUTPUT
:
192 ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
, swizzle
);
193 result
= LLVMBuildLoad(builder
, ptr
, "");
197 return LLVMGetUndef(tgsi2llvmtype(bld_base
, type
));
200 return bitcast(bld_base
, type
, result
);
203 static LLVMValueRef
fetch_system_value(
204 struct lp_build_tgsi_context
* bld_base
,
205 const struct tgsi_full_src_register
*reg
,
206 enum tgsi_opcode_type type
,
209 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
210 LLVMValueRef cval
= ctx
->system_values
[reg
->Register
.Index
];
211 return bitcast(bld_base
, type
, cval
);
214 static void emit_declaration(
215 struct lp_build_tgsi_context
* bld_base
,
216 const struct tgsi_full_declaration
*decl
)
218 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
219 switch(decl
->Declaration
.File
) {
220 case TGSI_FILE_ADDRESS
:
223 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
225 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
226 ctx
->soa
.addr
[idx
][chan
] = lp_build_alloca(
228 ctx
->soa
.bld_base
.uint_bld
.elem_type
, "");
234 case TGSI_FILE_TEMPORARY
:
235 if (decl
->Declaration
.Array
&& decl
->Array
.ArrayID
<= RADEON_LLVM_MAX_ARRAYS
)
236 ctx
->arrays
[decl
->Array
.ArrayID
- 1] = decl
->Range
;
237 lp_emit_declaration_soa(bld_base
, decl
);
240 case TGSI_FILE_INPUT
:
243 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
244 ctx
->load_input(ctx
, idx
, decl
);
249 case TGSI_FILE_SYSTEM_VALUE
:
252 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
253 ctx
->load_system_value(ctx
, idx
, decl
);
258 case TGSI_FILE_OUTPUT
:
261 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
263 assert(idx
< RADEON_LLVM_MAX_OUTPUTS
);
264 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
265 ctx
->soa
.outputs
[idx
][chan
] = lp_build_alloca(&ctx
->gallivm
,
266 ctx
->soa
.bld_base
.base
.elem_type
, "");
270 ctx
->output_reg_count
= MAX2(ctx
->output_reg_count
,
271 decl
->Range
.Last
+ 1);
282 struct lp_build_tgsi_context
* bld_base
,
283 const struct tgsi_full_instruction
* inst
,
284 const struct tgsi_opcode_info
* info
,
287 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
288 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
289 struct lp_build_context base
= bld
->bld_base
.base
;
290 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[0];
291 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
292 LLVMValueRef temp_ptr
;
293 unsigned chan
, chan_index
;
294 boolean is_vec_store
= FALSE
;
297 LLVMTypeKind k
= LLVMGetTypeKind(LLVMTypeOf(dst
[0]));
298 is_vec_store
= (k
== LLVMVectorTypeKind
);
302 LLVMValueRef values
[4] = {};
303 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst
, chan
) {
304 LLVMValueRef index
= lp_build_const_int32(gallivm
, chan
);
305 values
[chan
] = LLVMBuildExtractElement(gallivm
->builder
,
308 bld_base
->emit_store(bld_base
, inst
, info
, values
);
312 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
313 LLVMValueRef value
= dst
[chan_index
];
315 if (inst
->Instruction
.Saturate
!= TGSI_SAT_NONE
) {
316 struct lp_build_emit_data clamp_emit_data
;
318 memset(&clamp_emit_data
, 0, sizeof(clamp_emit_data
));
319 clamp_emit_data
.arg_count
= 3;
320 clamp_emit_data
.args
[0] = value
;
321 clamp_emit_data
.args
[2] = base
.one
;
323 switch(inst
->Instruction
.Saturate
) {
324 case TGSI_SAT_ZERO_ONE
:
325 clamp_emit_data
.args
[1] = base
.zero
;
327 case TGSI_SAT_MINUS_PLUS_ONE
:
328 clamp_emit_data
.args
[1] = LLVMConstReal(
329 base
.elem_type
, -1.0f
);
334 value
= lp_build_emit_llvm(bld_base
, TGSI_OPCODE_CLAMP
,
338 if (reg
->Register
.File
== TGSI_FILE_ADDRESS
) {
339 temp_ptr
= bld
->addr
[reg
->Register
.Index
][chan_index
];
340 LLVMBuildStore(builder
, value
, temp_ptr
);
344 value
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, value
);
346 if (reg
->Register
.Indirect
) {
347 struct tgsi_declaration_range range
= get_array_range(bld_base
,
348 reg
->Register
.File
, ®
->Indirect
);
350 unsigned i
, size
= range
.Last
- range
.First
+ 1;
351 LLVMValueRef array
= LLVMBuildInsertElement(builder
,
352 emit_array_fetch(bld_base
, reg
->Register
.File
, TGSI_TYPE_FLOAT
, range
, chan_index
),
353 value
, emit_array_index(bld
, ®
->Indirect
, reg
->Register
.Index
- range
.First
), "");
355 for (i
= 0; i
< size
; ++i
) {
356 switch(reg
->Register
.File
) {
357 case TGSI_FILE_OUTPUT
:
358 temp_ptr
= bld
->outputs
[i
+ range
.First
][chan_index
];
361 case TGSI_FILE_TEMPORARY
:
362 temp_ptr
= lp_get_temp_ptr_soa(bld
, i
+ range
.First
, chan_index
);
368 value
= LLVMBuildExtractElement(builder
, array
,
369 lp_build_const_int32(gallivm
, i
), "");
370 LLVMBuildStore(builder
, value
, temp_ptr
);
374 switch(reg
->Register
.File
) {
375 case TGSI_FILE_OUTPUT
:
376 temp_ptr
= bld
->outputs
[reg
->Register
.Index
][chan_index
];
379 case TGSI_FILE_TEMPORARY
:
380 temp_ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, chan_index
);
386 LLVMBuildStore(builder
, value
, temp_ptr
);
391 static void bgnloop_emit(
392 const struct lp_build_tgsi_action
* action
,
393 struct lp_build_tgsi_context
* bld_base
,
394 struct lp_build_emit_data
* emit_data
)
396 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
397 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
398 LLVMBasicBlockRef loop_block
;
399 LLVMBasicBlockRef endloop_block
;
400 endloop_block
= LLVMAppendBasicBlockInContext(gallivm
->context
,
401 ctx
->main_fn
, "ENDLOOP");
402 loop_block
= LLVMInsertBasicBlockInContext(gallivm
->context
,
403 endloop_block
, "LOOP");
404 LLVMBuildBr(gallivm
->builder
, loop_block
);
405 LLVMPositionBuilderAtEnd(gallivm
->builder
, loop_block
);
407 ctx
->loop
[ctx
->loop_depth
- 1].loop_block
= loop_block
;
408 ctx
->loop
[ctx
->loop_depth
- 1].endloop_block
= endloop_block
;
411 static void brk_emit(
412 const struct lp_build_tgsi_action
* action
,
413 struct lp_build_tgsi_context
* bld_base
,
414 struct lp_build_emit_data
* emit_data
)
416 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
417 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
418 struct radeon_llvm_loop
* current_loop
= get_current_loop(ctx
);
420 LLVMBuildBr(gallivm
->builder
, current_loop
->endloop_block
);
423 static void cont_emit(
424 const struct lp_build_tgsi_action
* action
,
425 struct lp_build_tgsi_context
* bld_base
,
426 struct lp_build_emit_data
* emit_data
)
428 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
429 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
430 struct radeon_llvm_loop
* current_loop
= get_current_loop(ctx
);
432 LLVMBuildBr(gallivm
->builder
, current_loop
->loop_block
);
435 static void else_emit(
436 const struct lp_build_tgsi_action
* action
,
437 struct lp_build_tgsi_context
* bld_base
,
438 struct lp_build_emit_data
* emit_data
)
440 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
441 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
442 struct radeon_llvm_branch
* current_branch
= get_current_branch(ctx
);
443 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(gallivm
->builder
);
445 /* We need to add a terminator to the current block if the previous
446 * instruction was an ENDIF.Example:
459 if (current_block
!= current_branch
->if_block
) {
460 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
462 if (!LLVMGetBasicBlockTerminator(current_branch
->if_block
)) {
463 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
465 current_branch
->has_else
= 1;
466 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->else_block
);
469 static void endif_emit(
470 const struct lp_build_tgsi_action
* action
,
471 struct lp_build_tgsi_context
* bld_base
,
472 struct lp_build_emit_data
* emit_data
)
474 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
475 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
476 struct radeon_llvm_branch
* current_branch
= get_current_branch(ctx
);
477 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(gallivm
->builder
);
479 /* If we have consecutive ENDIF instructions, then the first ENDIF
480 * will not have a terminator, so we need to add one. */
481 if (current_block
!= current_branch
->if_block
482 && current_block
!= current_branch
->else_block
483 && !LLVMGetBasicBlockTerminator(current_block
)) {
485 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
487 if (!LLVMGetBasicBlockTerminator(current_branch
->else_block
)) {
488 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->else_block
);
489 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
492 if (!LLVMGetBasicBlockTerminator(current_branch
->if_block
)) {
493 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->if_block
);
494 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
497 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->endif_block
);
501 static void endloop_emit(
502 const struct lp_build_tgsi_action
* action
,
503 struct lp_build_tgsi_context
* bld_base
,
504 struct lp_build_emit_data
* emit_data
)
506 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
507 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
508 struct radeon_llvm_loop
* current_loop
= get_current_loop(ctx
);
510 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(gallivm
->builder
))) {
511 LLVMBuildBr(gallivm
->builder
, current_loop
->loop_block
);
514 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_loop
->endloop_block
);
518 static void if_cond_emit(
519 const struct lp_build_tgsi_action
* action
,
520 struct lp_build_tgsi_context
* bld_base
,
521 struct lp_build_emit_data
* emit_data
,
524 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
525 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
526 LLVMBasicBlockRef if_block
, else_block
, endif_block
;
528 endif_block
= LLVMAppendBasicBlockInContext(gallivm
->context
,
529 ctx
->main_fn
, "ENDIF");
530 if_block
= LLVMInsertBasicBlockInContext(gallivm
->context
,
532 else_block
= LLVMInsertBasicBlockInContext(gallivm
->context
,
533 endif_block
, "ELSE");
534 LLVMBuildCondBr(gallivm
->builder
, cond
, if_block
, else_block
);
535 LLVMPositionBuilderAtEnd(gallivm
->builder
, if_block
);
538 ctx
->branch
[ctx
->branch_depth
- 1].endif_block
= endif_block
;
539 ctx
->branch
[ctx
->branch_depth
- 1].if_block
= if_block
;
540 ctx
->branch
[ctx
->branch_depth
- 1].else_block
= else_block
;
541 ctx
->branch
[ctx
->branch_depth
- 1].has_else
= 0;
545 const struct lp_build_tgsi_action
* action
,
546 struct lp_build_tgsi_context
* bld_base
,
547 struct lp_build_emit_data
* emit_data
)
549 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
552 cond
= LLVMBuildFCmp(gallivm
->builder
, LLVMRealUNE
,
554 bld_base
->base
.zero
, "");
556 if_cond_emit(action
, bld_base
, emit_data
, cond
);
559 static void uif_emit(
560 const struct lp_build_tgsi_action
* action
,
561 struct lp_build_tgsi_context
* bld_base
,
562 struct lp_build_emit_data
* emit_data
)
564 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
567 cond
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
568 bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, emit_data
->args
[0]),
569 bld_base
->int_bld
.zero
, "");
571 if_cond_emit(action
, bld_base
, emit_data
, cond
);
574 static void kil_emit(
575 const struct lp_build_tgsi_action
* action
,
576 struct lp_build_tgsi_context
* bld_base
,
577 struct lp_build_emit_data
* emit_data
)
580 for (i
= 0; i
< emit_data
->arg_count
; i
++) {
581 emit_data
->output
[i
] = lp_build_intrinsic_unary(
582 bld_base
->base
.gallivm
->builder
,
584 emit_data
->dst_type
, emit_data
->args
[i
]);
588 void radeon_llvm_emit_prepare_cube_coords(
589 struct lp_build_tgsi_context
* bld_base
,
590 struct lp_build_emit_data
* emit_data
,
591 LLVMValueRef
*coords_arg
)
594 unsigned target
= emit_data
->inst
->Texture
.Texture
;
595 unsigned opcode
= emit_data
->inst
->Instruction
.Opcode
;
596 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
597 LLVMBuilderRef builder
= gallivm
->builder
;
598 LLVMTypeRef type
= bld_base
->base
.elem_type
;
599 LLVMValueRef coords
[4];
600 LLVMValueRef mad_args
[3];
602 struct LLVMOpaqueValue
*cube_vec
;
606 cube_vec
= lp_build_gather_values(bld_base
->base
.gallivm
, coords_arg
, 4);
607 v
= build_intrinsic(builder
, "llvm.AMDGPU.cube", LLVMVectorType(type
, 4),
608 &cube_vec
, 1, LLVMReadNoneAttribute
);
610 for (i
= 0; i
< 4; ++i
) {
611 idx
= lp_build_const_int32(gallivm
, i
);
612 coords
[i
] = LLVMBuildExtractElement(builder
, v
, idx
, "");
615 coords
[2] = build_intrinsic(builder
, "fabs",
616 type
, &coords
[2], 1, LLVMReadNoneAttribute
);
617 coords
[2] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_RCP
, coords
[2]);
619 mad_args
[1] = coords
[2];
620 mad_args
[2] = LLVMConstReal(type
, 1.5);
622 mad_args
[0] = coords
[0];
623 coords
[0] = lp_build_emit_llvm_ternary(bld_base
, TGSI_OPCODE_MAD
,
624 mad_args
[0], mad_args
[1], mad_args
[2]);
626 mad_args
[0] = coords
[1];
627 coords
[1] = lp_build_emit_llvm_ternary(bld_base
, TGSI_OPCODE_MAD
,
628 mad_args
[0], mad_args
[1], mad_args
[2]);
630 /* apply xyz = yxw swizzle to cooords */
631 coords
[2] = coords
[3];
632 coords
[3] = coords
[1];
633 coords
[1] = coords
[0];
634 coords
[0] = coords
[3];
636 /* all cases except simple cube map sampling require special handling
637 * for coord vector */
638 if (target
!= TGSI_TEXTURE_CUBE
||
639 opcode
!= TGSI_OPCODE_TEX
) {
641 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
642 if (target
== TGSI_TEXTURE_CUBE_ARRAY
||
643 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
645 /* coords_arg.w component - array_index for cube arrays or
646 * compare value for SHADOWCUBE */
647 coords
[2] = lp_build_emit_llvm_ternary(bld_base
, TGSI_OPCODE_MAD
,
648 coords_arg
[3], lp_build_const_float(gallivm
, 8.0), coords
[2]);
651 /* for instructions that need additional src (compare/lod/bias),
652 * put it in coord.w */
653 if (opcode
== TGSI_OPCODE_TEX2
||
654 opcode
== TGSI_OPCODE_TXB2
||
655 opcode
== TGSI_OPCODE_TXL2
) {
656 coords
[3] = coords_arg
[4];
657 } else if (opcode
== TGSI_OPCODE_TXB
||
658 opcode
== TGSI_OPCODE_TXL
) {
659 coords
[3] = coords_arg
[3];
663 memcpy(coords_arg
, coords
, sizeof(coords
));
666 static void txd_fetch_args(
667 struct lp_build_tgsi_context
* bld_base
,
668 struct lp_build_emit_data
* emit_data
)
670 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
672 LLVMValueRef coords
[4];
674 for (src
= 0; src
< 3; src
++) {
675 for (chan
= 0; chan
< 4; chan
++)
676 coords
[chan
] = lp_build_emit_fetch(bld_base
, inst
, src
, chan
);
678 emit_data
->args
[src
] = lp_build_gather_values(bld_base
->base
.gallivm
,
681 emit_data
->arg_count
= 3;
682 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
686 static void txp_fetch_args(
687 struct lp_build_tgsi_context
* bld_base
,
688 struct lp_build_emit_data
* emit_data
)
690 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
693 LLVMValueRef coords
[4];
695 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
696 src_w
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
698 for (chan
= 0; chan
< 3; chan
++ ) {
699 LLVMValueRef arg
= lp_build_emit_fetch(bld_base
,
700 emit_data
->inst
, 0, chan
);
701 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
702 TGSI_OPCODE_DIV
, arg
, src_w
);
704 coords
[3] = bld_base
->base
.one
;
706 if ((inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE
||
707 inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
||
708 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE
||
709 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) &&
710 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ
&&
711 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ_LZ
) {
712 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
);
715 emit_data
->args
[0] = lp_build_gather_values(bld_base
->base
.gallivm
,
717 emit_data
->arg_count
= 1;
720 static void tex_fetch_args(
721 struct lp_build_tgsi_context
* bld_base
,
722 struct lp_build_emit_data
* emit_data
)
724 /* XXX: lp_build_swizzle_aos() was failing with wrong arg types,
725 * when we used CHAN_ALL. We should be able to get this to work,
726 * but for now we will swizzle it ourselves
727 emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
732 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
734 LLVMValueRef coords
[5];
736 for (chan
= 0; chan
< 4; chan
++) {
737 coords
[chan
] = lp_build_emit_fetch(bld_base
, inst
, 0, chan
);
740 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_TEX2
||
741 inst
->Instruction
.Opcode
== TGSI_OPCODE_TXB2
||
742 inst
->Instruction
.Opcode
== TGSI_OPCODE_TXL2
) {
743 /* These instructions have additional operand that should be packed
744 * into the cube coord vector by radeon_llvm_emit_prepare_cube_coords.
745 * That operand should be passed as a float value in the args array
746 * right after the coord vector. After packing it's not used anymore,
747 * that's why arg_count is not increased */
748 coords
[4] = lp_build_emit_fetch(bld_base
, inst
, 1, 0);
751 if ((inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE
||
752 inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
||
753 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE
||
754 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) &&
755 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ
&&
756 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ_LZ
) {
757 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
);
760 emit_data
->arg_count
= 1;
761 emit_data
->args
[0] = lp_build_gather_values(bld_base
->base
.gallivm
,
763 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
766 static void txf_fetch_args(
767 struct lp_build_tgsi_context
* bld_base
,
768 struct lp_build_emit_data
* emit_data
)
770 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
771 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
772 const struct tgsi_texture_offset
* off
= inst
->TexOffsets
;
773 LLVMTypeRef offset_type
= bld_base
->int_bld
.elem_type
;
775 /* fetch tex coords */
776 tex_fetch_args(bld_base
, emit_data
);
778 /* fetch tex offsets */
779 if (inst
->Texture
.NumOffsets
) {
780 assert(inst
->Texture
.NumOffsets
== 1);
782 emit_data
->args
[1] = LLVMConstBitCast(
783 bld
->immediates
[off
->Index
][off
->SwizzleX
],
785 emit_data
->args
[2] = LLVMConstBitCast(
786 bld
->immediates
[off
->Index
][off
->SwizzleY
],
788 emit_data
->args
[3] = LLVMConstBitCast(
789 bld
->immediates
[off
->Index
][off
->SwizzleZ
],
792 emit_data
->args
[1] = bld_base
->int_bld
.zero
;
793 emit_data
->args
[2] = bld_base
->int_bld
.zero
;
794 emit_data
->args
[3] = bld_base
->int_bld
.zero
;
797 emit_data
->arg_count
= 4;
800 static void emit_icmp(
801 const struct lp_build_tgsi_action
* action
,
802 struct lp_build_tgsi_context
* bld_base
,
803 struct lp_build_emit_data
* emit_data
)
806 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
807 LLVMContextRef context
= bld_base
->base
.gallivm
->context
;
809 switch (emit_data
->inst
->Instruction
.Opcode
) {
810 case TGSI_OPCODE_USEQ
: pred
= LLVMIntEQ
; break;
811 case TGSI_OPCODE_USNE
: pred
= LLVMIntNE
; break;
812 case TGSI_OPCODE_USGE
: pred
= LLVMIntUGE
; break;
813 case TGSI_OPCODE_USLT
: pred
= LLVMIntULT
; break;
814 case TGSI_OPCODE_ISGE
: pred
= LLVMIntSGE
; break;
815 case TGSI_OPCODE_ISLT
: pred
= LLVMIntSLT
; break;
817 assert(!"unknown instruction");
822 LLVMValueRef v
= LLVMBuildICmp(builder
, pred
,
823 emit_data
->args
[0], emit_data
->args
[1],"");
825 v
= LLVMBuildSExtOrBitCast(builder
, v
,
826 LLVMInt32TypeInContext(context
), "");
828 emit_data
->output
[emit_data
->chan
] = v
;
831 static void emit_ucmp(
832 const struct lp_build_tgsi_action
* action
,
833 struct lp_build_tgsi_context
* bld_base
,
834 struct lp_build_emit_data
* emit_data
)
836 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
838 LLVMValueRef v
= LLVMBuildFCmp(builder
, LLVMRealUGE
,
839 emit_data
->args
[0], lp_build_const_float(bld_base
->base
.gallivm
, 0.), "");
841 emit_data
->output
[emit_data
->chan
] = LLVMBuildSelect(builder
, v
, emit_data
->args
[2], emit_data
->args
[1], "");
844 static void emit_cmp(
845 const struct lp_build_tgsi_action
*action
,
846 struct lp_build_tgsi_context
* bld_base
,
847 struct lp_build_emit_data
* emit_data
)
849 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
850 LLVMRealPredicate pred
;
853 /* Use ordered for everything but NE (which is usual for
856 switch (emit_data
->inst
->Instruction
.Opcode
) {
857 case TGSI_OPCODE_SGE
: pred
= LLVMRealOGE
; break;
858 case TGSI_OPCODE_SEQ
: pred
= LLVMRealOEQ
; break;
859 case TGSI_OPCODE_SLE
: pred
= LLVMRealOLE
; break;
860 case TGSI_OPCODE_SLT
: pred
= LLVMRealOLT
; break;
861 case TGSI_OPCODE_SNE
: pred
= LLVMRealUNE
; break;
862 case TGSI_OPCODE_SGT
: pred
= LLVMRealOGT
; break;
863 default: assert(!"unknown instruction"); pred
= 0; break;
866 cond
= LLVMBuildFCmp(builder
,
867 pred
, emit_data
->args
[0], emit_data
->args
[1], "");
869 emit_data
->output
[emit_data
->chan
] = LLVMBuildSelect(builder
,
870 cond
, bld_base
->base
.one
, bld_base
->base
.zero
, "");
873 static void emit_fcmp(
874 const struct lp_build_tgsi_action
*action
,
875 struct lp_build_tgsi_context
* bld_base
,
876 struct lp_build_emit_data
* emit_data
)
878 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
879 LLVMContextRef context
= bld_base
->base
.gallivm
->context
;
880 LLVMRealPredicate pred
;
882 /* Use ordered for everything but NE (which is usual for
885 switch (emit_data
->inst
->Instruction
.Opcode
) {
886 case TGSI_OPCODE_FSEQ
: pred
= LLVMRealOEQ
; break;
887 case TGSI_OPCODE_FSGE
: pred
= LLVMRealOGE
; break;
888 case TGSI_OPCODE_FSLT
: pred
= LLVMRealOLT
; break;
889 case TGSI_OPCODE_FSNE
: pred
= LLVMRealUNE
; break;
890 default: assert(!"unknown instruction"); pred
= 0; break;
893 LLVMValueRef v
= LLVMBuildFCmp(builder
, pred
,
894 emit_data
->args
[0], emit_data
->args
[1],"");
896 v
= LLVMBuildSExtOrBitCast(builder
, v
,
897 LLVMInt32TypeInContext(context
), "");
899 emit_data
->output
[emit_data
->chan
] = v
;
902 static void emit_not(
903 const struct lp_build_tgsi_action
* action
,
904 struct lp_build_tgsi_context
* bld_base
,
905 struct lp_build_emit_data
* emit_data
)
907 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
908 LLVMValueRef v
= bitcast(bld_base
, TGSI_TYPE_UNSIGNED
,
910 emit_data
->output
[emit_data
->chan
] = LLVMBuildNot(builder
, v
, "");
913 static void emit_arl(
914 const struct lp_build_tgsi_action
* action
,
915 struct lp_build_tgsi_context
* bld_base
,
916 struct lp_build_emit_data
* emit_data
)
918 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
919 LLVMValueRef floor_index
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_FLR
, emit_data
->args
[0]);
920 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToSI(builder
,
921 floor_index
, bld_base
->base
.int_elem_type
, "");
924 static void emit_and(
925 const struct lp_build_tgsi_action
* action
,
926 struct lp_build_tgsi_context
* bld_base
,
927 struct lp_build_emit_data
* emit_data
)
929 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
930 emit_data
->output
[emit_data
->chan
] = LLVMBuildAnd(builder
,
931 emit_data
->args
[0], emit_data
->args
[1], "");
935 const struct lp_build_tgsi_action
* action
,
936 struct lp_build_tgsi_context
* bld_base
,
937 struct lp_build_emit_data
* emit_data
)
939 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
940 emit_data
->output
[emit_data
->chan
] = LLVMBuildOr(builder
,
941 emit_data
->args
[0], emit_data
->args
[1], "");
944 static void emit_uadd(
945 const struct lp_build_tgsi_action
* action
,
946 struct lp_build_tgsi_context
* bld_base
,
947 struct lp_build_emit_data
* emit_data
)
949 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
950 emit_data
->output
[emit_data
->chan
] = LLVMBuildAdd(builder
,
951 emit_data
->args
[0], emit_data
->args
[1], "");
954 static void emit_udiv(
955 const struct lp_build_tgsi_action
* action
,
956 struct lp_build_tgsi_context
* bld_base
,
957 struct lp_build_emit_data
* emit_data
)
959 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
960 emit_data
->output
[emit_data
->chan
] = LLVMBuildUDiv(builder
,
961 emit_data
->args
[0], emit_data
->args
[1], "");
964 static void emit_idiv(
965 const struct lp_build_tgsi_action
* action
,
966 struct lp_build_tgsi_context
* bld_base
,
967 struct lp_build_emit_data
* emit_data
)
969 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
970 emit_data
->output
[emit_data
->chan
] = LLVMBuildSDiv(builder
,
971 emit_data
->args
[0], emit_data
->args
[1], "");
974 static void emit_mod(
975 const struct lp_build_tgsi_action
* action
,
976 struct lp_build_tgsi_context
* bld_base
,
977 struct lp_build_emit_data
* emit_data
)
979 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
980 emit_data
->output
[emit_data
->chan
] = LLVMBuildSRem(builder
,
981 emit_data
->args
[0], emit_data
->args
[1], "");
984 static void emit_umod(
985 const struct lp_build_tgsi_action
* action
,
986 struct lp_build_tgsi_context
* bld_base
,
987 struct lp_build_emit_data
* emit_data
)
989 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
990 emit_data
->output
[emit_data
->chan
] = LLVMBuildURem(builder
,
991 emit_data
->args
[0], emit_data
->args
[1], "");
994 static void emit_shl(
995 const struct lp_build_tgsi_action
* action
,
996 struct lp_build_tgsi_context
* bld_base
,
997 struct lp_build_emit_data
* emit_data
)
999 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1000 emit_data
->output
[emit_data
->chan
] = LLVMBuildShl(builder
,
1001 emit_data
->args
[0], emit_data
->args
[1], "");
1004 static void emit_ushr(
1005 const struct lp_build_tgsi_action
* action
,
1006 struct lp_build_tgsi_context
* bld_base
,
1007 struct lp_build_emit_data
* emit_data
)
1009 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1010 emit_data
->output
[emit_data
->chan
] = LLVMBuildLShr(builder
,
1011 emit_data
->args
[0], emit_data
->args
[1], "");
1013 static void emit_ishr(
1014 const struct lp_build_tgsi_action
* action
,
1015 struct lp_build_tgsi_context
* bld_base
,
1016 struct lp_build_emit_data
* emit_data
)
1018 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1019 emit_data
->output
[emit_data
->chan
] = LLVMBuildAShr(builder
,
1020 emit_data
->args
[0], emit_data
->args
[1], "");
1023 static void emit_xor(
1024 const struct lp_build_tgsi_action
* action
,
1025 struct lp_build_tgsi_context
* bld_base
,
1026 struct lp_build_emit_data
* emit_data
)
1028 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1029 emit_data
->output
[emit_data
->chan
] = LLVMBuildXor(builder
,
1030 emit_data
->args
[0], emit_data
->args
[1], "");
1033 static void emit_ssg(
1034 const struct lp_build_tgsi_action
* action
,
1035 struct lp_build_tgsi_context
* bld_base
,
1036 struct lp_build_emit_data
* emit_data
)
1038 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1040 LLVMValueRef cmp
, val
;
1042 if (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_ISSG
) {
1043 cmp
= LLVMBuildICmp(builder
, LLVMIntSGT
, emit_data
->args
[0], bld_base
->int_bld
.zero
, "");
1044 val
= LLVMBuildSelect(builder
, cmp
, bld_base
->int_bld
.one
, emit_data
->args
[0], "");
1045 cmp
= LLVMBuildICmp(builder
, LLVMIntSGE
, val
, bld_base
->int_bld
.zero
, "");
1046 val
= LLVMBuildSelect(builder
, cmp
, val
, LLVMConstInt(bld_base
->int_bld
.elem_type
, -1, true), "");
1047 } else { // float SSG
1048 cmp
= LLVMBuildFCmp(builder
, LLVMRealUGT
, emit_data
->args
[0], bld_base
->base
.zero
, "");
1049 val
= LLVMBuildSelect(builder
, cmp
, bld_base
->base
.one
, emit_data
->args
[0], "");
1050 cmp
= LLVMBuildFCmp(builder
, LLVMRealUGE
, val
, bld_base
->base
.zero
, "");
1051 val
= LLVMBuildSelect(builder
, cmp
, val
, LLVMConstReal(bld_base
->base
.elem_type
, -1), "");
1054 emit_data
->output
[emit_data
->chan
] = val
;
1057 static void emit_ineg(
1058 const struct lp_build_tgsi_action
* action
,
1059 struct lp_build_tgsi_context
* bld_base
,
1060 struct lp_build_emit_data
* emit_data
)
1062 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1063 emit_data
->output
[emit_data
->chan
] = LLVMBuildNeg(builder
,
1064 emit_data
->args
[0], "");
1067 static void emit_f2i(
1068 const struct lp_build_tgsi_action
* action
,
1069 struct lp_build_tgsi_context
* bld_base
,
1070 struct lp_build_emit_data
* emit_data
)
1072 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1073 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToSI(builder
,
1074 emit_data
->args
[0], bld_base
->int_bld
.elem_type
, "");
1077 static void emit_f2u(
1078 const struct lp_build_tgsi_action
* action
,
1079 struct lp_build_tgsi_context
* bld_base
,
1080 struct lp_build_emit_data
* emit_data
)
1082 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1083 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToUI(builder
,
1084 emit_data
->args
[0], bld_base
->uint_bld
.elem_type
, "");
1087 static void emit_i2f(
1088 const struct lp_build_tgsi_action
* action
,
1089 struct lp_build_tgsi_context
* bld_base
,
1090 struct lp_build_emit_data
* emit_data
)
1092 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1093 emit_data
->output
[emit_data
->chan
] = LLVMBuildSIToFP(builder
,
1094 emit_data
->args
[0], bld_base
->base
.elem_type
, "");
1097 static void emit_u2f(
1098 const struct lp_build_tgsi_action
* action
,
1099 struct lp_build_tgsi_context
* bld_base
,
1100 struct lp_build_emit_data
* emit_data
)
1102 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1103 emit_data
->output
[emit_data
->chan
] = LLVMBuildUIToFP(builder
,
1104 emit_data
->args
[0], bld_base
->base
.elem_type
, "");
1107 static void emit_immediate(struct lp_build_tgsi_context
* bld_base
,
1108 const struct tgsi_full_immediate
*imm
)
1111 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
1113 for (i
= 0; i
< 4; ++i
) {
1114 ctx
->soa
.immediates
[ctx
->soa
.num_immediates
][i
] =
1115 LLVMConstInt(bld_base
->uint_bld
.elem_type
, imm
->u
[i
].Uint
, false );
1118 ctx
->soa
.num_immediates
++;
1122 build_intrinsic(LLVMBuilderRef builder
,
1124 LLVMTypeRef ret_type
,
1129 LLVMModuleRef module
= LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder
)));
1130 LLVMValueRef function
;
1132 function
= LLVMGetNamedFunction(module
, name
);
1134 LLVMTypeRef arg_types
[LP_MAX_FUNC_ARGS
];
1137 assert(num_args
<= LP_MAX_FUNC_ARGS
);
1139 for(i
= 0; i
< num_args
; ++i
) {
1141 arg_types
[i
] = LLVMTypeOf(args
[i
]);
1144 function
= lp_declare_intrinsic(module
, name
, ret_type
, arg_types
, num_args
);
1147 LLVMAddFunctionAttr(function
, attr
);
1150 return LLVMBuildCall(builder
, function
, args
, num_args
, "");
1153 static void build_tgsi_intrinsic(
1154 const struct lp_build_tgsi_action
* action
,
1155 struct lp_build_tgsi_context
* bld_base
,
1156 struct lp_build_emit_data
* emit_data
,
1159 struct lp_build_context
* base
= &bld_base
->base
;
1160 emit_data
->output
[emit_data
->chan
] = build_intrinsic(
1161 base
->gallivm
->builder
, action
->intr_name
,
1162 emit_data
->dst_type
, emit_data
->args
,
1163 emit_data
->arg_count
, attr
);
1166 build_tgsi_intrinsic_nomem(
1167 const struct lp_build_tgsi_action
* action
,
1168 struct lp_build_tgsi_context
* bld_base
,
1169 struct lp_build_emit_data
* emit_data
)
1171 build_tgsi_intrinsic(action
, bld_base
, emit_data
, LLVMReadNoneAttribute
);
1174 static void build_tgsi_intrinsic_readonly(
1175 const struct lp_build_tgsi_action
* action
,
1176 struct lp_build_tgsi_context
* bld_base
,
1177 struct lp_build_emit_data
* emit_data
)
1179 build_tgsi_intrinsic(action
, bld_base
, emit_data
, LLVMReadOnlyAttribute
);
1182 void radeon_llvm_context_init(struct radeon_llvm_context
* ctx
)
1184 struct lp_type type
;
1186 /* Initialize the gallivm object:
1187 * We are only using the module, context, and builder fields of this struct.
1188 * This should be enough for us to be able to pass our gallivm struct to the
1189 * helper functions in the gallivm module.
1191 memset(&ctx
->gallivm
, 0, sizeof (ctx
->gallivm
));
1192 memset(&ctx
->soa
, 0, sizeof(ctx
->soa
));
1193 ctx
->gallivm
.context
= LLVMContextCreate();
1194 ctx
->gallivm
.module
= LLVMModuleCreateWithNameInContext("tgsi",
1195 ctx
->gallivm
.context
);
1196 ctx
->gallivm
.builder
= LLVMCreateBuilderInContext(ctx
->gallivm
.context
);
1198 ctx
->store_output_intr
= "llvm.AMDGPU.store.output.";
1199 ctx
->swizzle_intr
= "llvm.AMDGPU.swizzle";
1200 struct lp_build_tgsi_context
* bld_base
= &ctx
->soa
.bld_base
;
1202 /* XXX: We need to revisit this.I think the correct way to do this is
1203 * to use length = 4 here and use the elem_bld for everything. */
1204 type
.floating
= TRUE
;
1211 lp_build_context_init(&bld_base
->base
, &ctx
->gallivm
, type
);
1212 lp_build_context_init(&ctx
->soa
.bld_base
.uint_bld
, &ctx
->gallivm
, lp_uint_type(type
));
1213 lp_build_context_init(&ctx
->soa
.bld_base
.int_bld
, &ctx
->gallivm
, lp_int_type(type
));
1216 bld_base
->emit_store
= emit_store
;
1217 bld_base
->emit_swizzle
= emit_swizzle
;
1218 bld_base
->emit_declaration
= emit_declaration
;
1219 bld_base
->emit_immediate
= emit_immediate
;
1221 bld_base
->emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch
;
1222 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch
;
1223 bld_base
->emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch
;
1224 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = emit_fetch
;
1225 bld_base
->emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = fetch_system_value
;
1227 /* Allocate outputs */
1228 ctx
->soa
.outputs
= ctx
->outputs
;
1230 ctx
->num_arrays
= 0;
1232 /* XXX: Is there a better way to initialize all this ? */
1234 lp_set_default_actions(bld_base
);
1236 bld_base
->op_actions
[TGSI_OPCODE_ABS
].emit
= build_tgsi_intrinsic_readonly
;
1237 bld_base
->op_actions
[TGSI_OPCODE_ABS
].intr_name
= "fabs";
1238 bld_base
->op_actions
[TGSI_OPCODE_ARL
].emit
= emit_arl
;
1239 bld_base
->op_actions
[TGSI_OPCODE_AND
].emit
= emit_and
;
1240 bld_base
->op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
1241 bld_base
->op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
1242 bld_base
->op_actions
[TGSI_OPCODE_CEIL
].emit
= build_tgsi_intrinsic_readonly
;
1243 bld_base
->op_actions
[TGSI_OPCODE_CEIL
].intr_name
= "ceil";
1244 bld_base
->op_actions
[TGSI_OPCODE_CLAMP
].emit
= build_tgsi_intrinsic_nomem
;
1245 bld_base
->op_actions
[TGSI_OPCODE_CLAMP
].intr_name
= "llvm.AMDIL.clamp.";
1246 bld_base
->op_actions
[TGSI_OPCODE_CMP
].emit
= build_tgsi_intrinsic_nomem
;
1247 bld_base
->op_actions
[TGSI_OPCODE_CMP
].intr_name
= "llvm.AMDGPU.cndlt";
1248 bld_base
->op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
1249 bld_base
->op_actions
[TGSI_OPCODE_COS
].emit
= build_tgsi_intrinsic_readonly
;
1250 bld_base
->op_actions
[TGSI_OPCODE_COS
].intr_name
= "llvm.cos.f32";
1251 bld_base
->op_actions
[TGSI_OPCODE_DDX
].intr_name
= "llvm.AMDGPU.ddx";
1252 bld_base
->op_actions
[TGSI_OPCODE_DDX
].fetch_args
= tex_fetch_args
;
1253 bld_base
->op_actions
[TGSI_OPCODE_DDY
].intr_name
= "llvm.AMDGPU.ddy";
1254 bld_base
->op_actions
[TGSI_OPCODE_DDY
].fetch_args
= tex_fetch_args
;
1255 bld_base
->op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
1256 bld_base
->op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
1257 bld_base
->op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
1258 bld_base
->op_actions
[TGSI_OPCODE_EX2
].emit
= build_tgsi_intrinsic_nomem
;
1259 bld_base
->op_actions
[TGSI_OPCODE_EX2
].intr_name
= "llvm.AMDIL.exp.";
1260 bld_base
->op_actions
[TGSI_OPCODE_FLR
].emit
= build_tgsi_intrinsic_readonly
;
1261 bld_base
->op_actions
[TGSI_OPCODE_FLR
].intr_name
= "floor";
1262 bld_base
->op_actions
[TGSI_OPCODE_FRC
].emit
= build_tgsi_intrinsic_nomem
;
1263 bld_base
->op_actions
[TGSI_OPCODE_FRC
].intr_name
= "llvm.AMDIL.fraction.";
1264 bld_base
->op_actions
[TGSI_OPCODE_F2I
].emit
= emit_f2i
;
1265 bld_base
->op_actions
[TGSI_OPCODE_F2U
].emit
= emit_f2u
;
1266 bld_base
->op_actions
[TGSI_OPCODE_FSEQ
].emit
= emit_fcmp
;
1267 bld_base
->op_actions
[TGSI_OPCODE_FSGE
].emit
= emit_fcmp
;
1268 bld_base
->op_actions
[TGSI_OPCODE_FSLT
].emit
= emit_fcmp
;
1269 bld_base
->op_actions
[TGSI_OPCODE_FSNE
].emit
= emit_fcmp
;
1270 bld_base
->op_actions
[TGSI_OPCODE_IABS
].emit
= build_tgsi_intrinsic_nomem
;
1271 bld_base
->op_actions
[TGSI_OPCODE_IABS
].intr_name
= "llvm.AMDIL.abs.";
1272 bld_base
->op_actions
[TGSI_OPCODE_IDIV
].emit
= emit_idiv
;
1273 bld_base
->op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
1274 bld_base
->op_actions
[TGSI_OPCODE_UIF
].emit
= uif_emit
;
1275 bld_base
->op_actions
[TGSI_OPCODE_IMAX
].emit
= build_tgsi_intrinsic_nomem
;
1276 bld_base
->op_actions
[TGSI_OPCODE_IMAX
].intr_name
= "llvm.AMDGPU.imax";
1277 bld_base
->op_actions
[TGSI_OPCODE_IMIN
].emit
= build_tgsi_intrinsic_nomem
;
1278 bld_base
->op_actions
[TGSI_OPCODE_IMIN
].intr_name
= "llvm.AMDGPU.imin";
1279 bld_base
->op_actions
[TGSI_OPCODE_INEG
].emit
= emit_ineg
;
1280 bld_base
->op_actions
[TGSI_OPCODE_ISHR
].emit
= emit_ishr
;
1281 bld_base
->op_actions
[TGSI_OPCODE_ISGE
].emit
= emit_icmp
;
1282 bld_base
->op_actions
[TGSI_OPCODE_ISLT
].emit
= emit_icmp
;
1283 bld_base
->op_actions
[TGSI_OPCODE_ISSG
].emit
= emit_ssg
;
1284 bld_base
->op_actions
[TGSI_OPCODE_I2F
].emit
= emit_i2f
;
1285 bld_base
->op_actions
[TGSI_OPCODE_KILL_IF
].emit
= kil_emit
;
1286 bld_base
->op_actions
[TGSI_OPCODE_KILL_IF
].intr_name
= "llvm.AMDGPU.kill";
1287 bld_base
->op_actions
[TGSI_OPCODE_KILL
].emit
= lp_build_tgsi_intrinsic
;
1288 bld_base
->op_actions
[TGSI_OPCODE_KILL
].intr_name
= "llvm.AMDGPU.kilp";
1289 bld_base
->op_actions
[TGSI_OPCODE_LG2
].emit
= build_tgsi_intrinsic_readonly
;
1290 bld_base
->op_actions
[TGSI_OPCODE_LG2
].intr_name
= "llvm.log2.f32";
1291 bld_base
->op_actions
[TGSI_OPCODE_LRP
].emit
= build_tgsi_intrinsic_nomem
;
1292 bld_base
->op_actions
[TGSI_OPCODE_LRP
].intr_name
= "llvm.AMDGPU.lrp";
1293 bld_base
->op_actions
[TGSI_OPCODE_MOD
].emit
= emit_mod
;
1294 bld_base
->op_actions
[TGSI_OPCODE_NOT
].emit
= emit_not
;
1295 bld_base
->op_actions
[TGSI_OPCODE_OR
].emit
= emit_or
;
1296 bld_base
->op_actions
[TGSI_OPCODE_POW
].emit
= build_tgsi_intrinsic_readonly
;
1297 bld_base
->op_actions
[TGSI_OPCODE_POW
].intr_name
= "llvm.pow.f32";
1298 bld_base
->op_actions
[TGSI_OPCODE_ROUND
].emit
= build_tgsi_intrinsic_nomem
;
1299 bld_base
->op_actions
[TGSI_OPCODE_ROUND
].intr_name
= "llvm.AMDIL.round.nearest.";
1300 bld_base
->op_actions
[TGSI_OPCODE_SGE
].emit
= emit_cmp
;
1301 bld_base
->op_actions
[TGSI_OPCODE_SEQ
].emit
= emit_cmp
;
1302 bld_base
->op_actions
[TGSI_OPCODE_SHL
].emit
= emit_shl
;
1303 bld_base
->op_actions
[TGSI_OPCODE_SLE
].emit
= emit_cmp
;
1304 bld_base
->op_actions
[TGSI_OPCODE_SLT
].emit
= emit_cmp
;
1305 bld_base
->op_actions
[TGSI_OPCODE_SNE
].emit
= emit_cmp
;
1306 bld_base
->op_actions
[TGSI_OPCODE_SGT
].emit
= emit_cmp
;
1307 bld_base
->op_actions
[TGSI_OPCODE_SIN
].emit
= build_tgsi_intrinsic_readonly
;
1308 bld_base
->op_actions
[TGSI_OPCODE_SIN
].intr_name
= "llvm.sin.f32";
1309 bld_base
->op_actions
[TGSI_OPCODE_SSG
].emit
= emit_ssg
;
1310 bld_base
->op_actions
[TGSI_OPCODE_TEX
].fetch_args
= tex_fetch_args
;
1311 bld_base
->op_actions
[TGSI_OPCODE_TEX
].intr_name
= "llvm.AMDGPU.tex";
1312 bld_base
->op_actions
[TGSI_OPCODE_TEX2
].fetch_args
= tex_fetch_args
;
1313 bld_base
->op_actions
[TGSI_OPCODE_TEX2
].intr_name
= "llvm.AMDGPU.tex";
1314 bld_base
->op_actions
[TGSI_OPCODE_TXB
].fetch_args
= tex_fetch_args
;
1315 bld_base
->op_actions
[TGSI_OPCODE_TXB
].intr_name
= "llvm.AMDGPU.txb";
1316 bld_base
->op_actions
[TGSI_OPCODE_TXB2
].fetch_args
= tex_fetch_args
;
1317 bld_base
->op_actions
[TGSI_OPCODE_TXB2
].intr_name
= "llvm.AMDGPU.txb";
1318 bld_base
->op_actions
[TGSI_OPCODE_TXD
].fetch_args
= txd_fetch_args
;
1319 bld_base
->op_actions
[TGSI_OPCODE_TXD
].intr_name
= "llvm.AMDGPU.txd";
1320 bld_base
->op_actions
[TGSI_OPCODE_TXF
].fetch_args
= txf_fetch_args
;
1321 bld_base
->op_actions
[TGSI_OPCODE_TXF
].intr_name
= "llvm.AMDGPU.txf";
1322 bld_base
->op_actions
[TGSI_OPCODE_TXL
].fetch_args
= tex_fetch_args
;
1323 bld_base
->op_actions
[TGSI_OPCODE_TXL
].intr_name
= "llvm.AMDGPU.txl";
1324 bld_base
->op_actions
[TGSI_OPCODE_TXL2
].fetch_args
= tex_fetch_args
;
1325 bld_base
->op_actions
[TGSI_OPCODE_TXL2
].intr_name
= "llvm.AMDGPU.txl";
1326 bld_base
->op_actions
[TGSI_OPCODE_TXP
].fetch_args
= txp_fetch_args
;
1327 bld_base
->op_actions
[TGSI_OPCODE_TXP
].intr_name
= "llvm.AMDGPU.tex";
1328 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].fetch_args
= tex_fetch_args
;
1329 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].intr_name
= "llvm.AMDGPU.txq";
1330 bld_base
->op_actions
[TGSI_OPCODE_TRUNC
].emit
= build_tgsi_intrinsic_nomem
;
1331 bld_base
->op_actions
[TGSI_OPCODE_TRUNC
].intr_name
= "llvm.AMDGPU.trunc";
1332 bld_base
->op_actions
[TGSI_OPCODE_UADD
].emit
= emit_uadd
;
1333 bld_base
->op_actions
[TGSI_OPCODE_UDIV
].emit
= emit_udiv
;
1334 bld_base
->op_actions
[TGSI_OPCODE_UMAX
].emit
= build_tgsi_intrinsic_nomem
;
1335 bld_base
->op_actions
[TGSI_OPCODE_UMAX
].intr_name
= "llvm.AMDGPU.umax";
1336 bld_base
->op_actions
[TGSI_OPCODE_UMIN
].emit
= build_tgsi_intrinsic_nomem
;
1337 bld_base
->op_actions
[TGSI_OPCODE_UMIN
].intr_name
= "llvm.AMDGPU.umin";
1338 bld_base
->op_actions
[TGSI_OPCODE_UMOD
].emit
= emit_umod
;
1339 bld_base
->op_actions
[TGSI_OPCODE_USEQ
].emit
= emit_icmp
;
1340 bld_base
->op_actions
[TGSI_OPCODE_USGE
].emit
= emit_icmp
;
1341 bld_base
->op_actions
[TGSI_OPCODE_USHR
].emit
= emit_ushr
;
1342 bld_base
->op_actions
[TGSI_OPCODE_USLT
].emit
= emit_icmp
;
1343 bld_base
->op_actions
[TGSI_OPCODE_USNE
].emit
= emit_icmp
;
1344 bld_base
->op_actions
[TGSI_OPCODE_U2F
].emit
= emit_u2f
;
1345 bld_base
->op_actions
[TGSI_OPCODE_XOR
].emit
= emit_xor
;
1346 bld_base
->op_actions
[TGSI_OPCODE_UCMP
].emit
= emit_ucmp
;
1348 bld_base
->rsq_action
.emit
= build_tgsi_intrinsic_nomem
;
1349 bld_base
->rsq_action
.intr_name
= "llvm.AMDGPU.rsq";
1352 void radeon_llvm_create_func(struct radeon_llvm_context
* ctx
,
1353 LLVMTypeRef
*ParamTypes
, unsigned ParamCount
)
1355 LLVMTypeRef main_fn_type
;
1356 LLVMBasicBlockRef main_fn_body
;
1358 /* Setup the function */
1359 main_fn_type
= LLVMFunctionType(LLVMVoidTypeInContext(ctx
->gallivm
.context
),
1360 ParamTypes
, ParamCount
, 0);
1361 ctx
->main_fn
= LLVMAddFunction(ctx
->gallivm
.module
, "main", main_fn_type
);
1362 main_fn_body
= LLVMAppendBasicBlockInContext(ctx
->gallivm
.context
,
1363 ctx
->main_fn
, "main_body");
1364 LLVMPositionBuilderAtEnd(ctx
->gallivm
.builder
, main_fn_body
);
1367 void radeon_llvm_finalize_module(struct radeon_llvm_context
* ctx
)
1369 struct gallivm_state
* gallivm
= ctx
->soa
.bld_base
.base
.gallivm
;
1370 /* End the main function with Return*/
1371 LLVMBuildRetVoid(gallivm
->builder
);
1373 /* Create the pass manager */
1374 ctx
->gallivm
.passmgr
= LLVMCreateFunctionPassManagerForModule(
1377 /* This pass should eliminate all the load and store instructions */
1378 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
1380 /* Add some optimization passes */
1381 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
1382 LLVMAddLICMPass(gallivm
->passmgr
);
1383 LLVMAddAggressiveDCEPass(gallivm
->passmgr
);
1384 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
1387 LLVMRunFunctionPassManager(gallivm
->passmgr
, ctx
->main_fn
);
1389 LLVMDisposeBuilder(gallivm
->builder
);
1390 LLVMDisposePassManager(gallivm
->passmgr
);
1394 void radeon_llvm_dispose(struct radeon_llvm_context
* ctx
)
1396 LLVMDisposeModule(ctx
->soa
.bld_base
.base
.gallivm
->module
);
1397 LLVMContextDispose(ctx
->soa
.bld_base
.base
.gallivm
->context
);