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
), "");
145 static bool uses_temp_indirect_addressing(
146 struct lp_build_tgsi_context
*bld_base
)
148 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
149 return (bld
->indirect_files
& (1 << TGSI_FILE_TEMPORARY
));
154 struct lp_build_tgsi_context
*bld_base
,
155 const struct tgsi_full_src_register
*reg
,
156 enum tgsi_opcode_type type
,
159 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
160 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
161 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
162 LLVMValueRef result
, ptr
;
165 LLVMValueRef values
[TGSI_NUM_CHANNELS
];
167 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
168 values
[chan
] = emit_fetch(bld_base
, reg
, type
, chan
);
170 return lp_build_gather_values(bld_base
->base
.gallivm
, values
,
174 if (reg
->Register
.Indirect
) {
175 struct tgsi_declaration_range range
= get_array_range(bld_base
,
176 reg
->Register
.File
, ®
->Indirect
);
177 return LLVMBuildExtractElement(builder
,
178 emit_array_fetch(bld_base
, reg
->Register
.File
, type
, range
, swizzle
),
179 emit_array_index(bld
, ®
->Indirect
, reg
->Register
.Index
- range
.First
),
183 switch(reg
->Register
.File
) {
184 case TGSI_FILE_IMMEDIATE
: {
185 LLVMTypeRef ctype
= tgsi2llvmtype(bld_base
, type
);
186 return LLVMConstBitCast(bld
->immediates
[reg
->Register
.Index
][swizzle
], ctype
);
189 case TGSI_FILE_INPUT
:
190 result
= ctx
->inputs
[radeon_llvm_reg_index_soa(reg
->Register
.Index
, swizzle
)];
193 case TGSI_FILE_TEMPORARY
:
194 if (uses_temp_indirect_addressing(bld_base
)) {
195 ptr
= lp_get_temp_ptr_soa(bld
, reg
->Register
.Index
, swizzle
);
198 ptr
= ctx
->temps
[reg
->Register
.Index
* TGSI_NUM_CHANNELS
+ swizzle
];
199 result
= LLVMBuildLoad(builder
, ptr
, "");
202 case TGSI_FILE_OUTPUT
:
203 ptr
= lp_get_output_ptr(bld
, reg
->Register
.Index
, swizzle
);
204 result
= LLVMBuildLoad(builder
, ptr
, "");
208 return LLVMGetUndef(tgsi2llvmtype(bld_base
, type
));
211 return bitcast(bld_base
, type
, result
);
214 static LLVMValueRef
fetch_system_value(
215 struct lp_build_tgsi_context
* bld_base
,
216 const struct tgsi_full_src_register
*reg
,
217 enum tgsi_opcode_type type
,
220 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
221 LLVMValueRef cval
= ctx
->system_values
[reg
->Register
.Index
];
222 return bitcast(bld_base
, type
, cval
);
225 static void emit_declaration(
226 struct lp_build_tgsi_context
* bld_base
,
227 const struct tgsi_full_declaration
*decl
)
229 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
230 unsigned first
, last
, i
, idx
;
231 switch(decl
->Declaration
.File
) {
232 case TGSI_FILE_ADDRESS
:
235 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
237 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
238 ctx
->soa
.addr
[idx
][chan
] = lp_build_alloca(
240 ctx
->soa
.bld_base
.uint_bld
.elem_type
, "");
246 case TGSI_FILE_TEMPORARY
:
247 if (decl
->Declaration
.Array
&& decl
->Array
.ArrayID
<= RADEON_LLVM_MAX_ARRAYS
)
248 ctx
->arrays
[decl
->Array
.ArrayID
- 1] = decl
->Range
;
249 if (uses_temp_indirect_addressing(bld_base
)) {
250 lp_emit_declaration_soa(bld_base
, decl
);
253 first
= decl
->Range
.First
;
254 last
= decl
->Range
.Last
;
255 if (!ctx
->temps_count
) {
256 ctx
->temps_count
= bld_base
->info
->file_max
[TGSI_FILE_TEMPORARY
] + 1;
257 ctx
->temps
= MALLOC(TGSI_NUM_CHANNELS
* ctx
->temps_count
* sizeof(LLVMValueRef
));
259 for (idx
= first
; idx
<= last
; idx
++) {
260 for (i
= 0; i
< TGSI_NUM_CHANNELS
; i
++) {
261 ctx
->temps
[idx
* TGSI_NUM_CHANNELS
+ i
] =
262 lp_build_alloca(bld_base
->base
.gallivm
, bld_base
->base
.vec_type
,
268 case TGSI_FILE_INPUT
:
271 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
272 ctx
->load_input(ctx
, idx
, decl
);
277 case TGSI_FILE_SYSTEM_VALUE
:
280 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
281 ctx
->load_system_value(ctx
, idx
, decl
);
286 case TGSI_FILE_OUTPUT
:
289 for (idx
= decl
->Range
.First
; idx
<= decl
->Range
.Last
; idx
++) {
291 assert(idx
< RADEON_LLVM_MAX_OUTPUTS
);
292 for (chan
= 0; chan
< TGSI_NUM_CHANNELS
; chan
++) {
293 ctx
->soa
.outputs
[idx
][chan
] = lp_build_alloca(&ctx
->gallivm
,
294 ctx
->soa
.bld_base
.base
.elem_type
, "");
298 ctx
->output_reg_count
= MAX2(ctx
->output_reg_count
,
299 decl
->Range
.Last
+ 1);
310 struct lp_build_tgsi_context
* bld_base
,
311 const struct tgsi_full_instruction
* inst
,
312 const struct tgsi_opcode_info
* info
,
315 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
316 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
317 struct gallivm_state
*gallivm
= bld
->bld_base
.base
.gallivm
;
318 struct lp_build_context base
= bld
->bld_base
.base
;
319 const struct tgsi_full_dst_register
*reg
= &inst
->Dst
[0];
320 LLVMBuilderRef builder
= bld
->bld_base
.base
.gallivm
->builder
;
321 LLVMValueRef temp_ptr
;
322 unsigned chan
, chan_index
;
323 boolean is_vec_store
= FALSE
;
326 LLVMTypeKind k
= LLVMGetTypeKind(LLVMTypeOf(dst
[0]));
327 is_vec_store
= (k
== LLVMVectorTypeKind
);
331 LLVMValueRef values
[4] = {};
332 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst
, chan
) {
333 LLVMValueRef index
= lp_build_const_int32(gallivm
, chan
);
334 values
[chan
] = LLVMBuildExtractElement(gallivm
->builder
,
337 bld_base
->emit_store(bld_base
, inst
, info
, values
);
341 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst
, chan_index
) {
342 LLVMValueRef value
= dst
[chan_index
];
344 if (inst
->Instruction
.Saturate
!= TGSI_SAT_NONE
) {
345 struct lp_build_emit_data clamp_emit_data
;
347 memset(&clamp_emit_data
, 0, sizeof(clamp_emit_data
));
348 clamp_emit_data
.arg_count
= 3;
349 clamp_emit_data
.args
[0] = value
;
350 clamp_emit_data
.args
[2] = base
.one
;
352 switch(inst
->Instruction
.Saturate
) {
353 case TGSI_SAT_ZERO_ONE
:
354 clamp_emit_data
.args
[1] = base
.zero
;
356 case TGSI_SAT_MINUS_PLUS_ONE
:
357 clamp_emit_data
.args
[1] = LLVMConstReal(
358 base
.elem_type
, -1.0f
);
363 value
= lp_build_emit_llvm(bld_base
, TGSI_OPCODE_CLAMP
,
367 if (reg
->Register
.File
== TGSI_FILE_ADDRESS
) {
368 temp_ptr
= bld
->addr
[reg
->Register
.Index
][chan_index
];
369 LLVMBuildStore(builder
, value
, temp_ptr
);
373 value
= bitcast(bld_base
, TGSI_TYPE_FLOAT
, value
);
375 if (reg
->Register
.Indirect
) {
376 struct tgsi_declaration_range range
= get_array_range(bld_base
,
377 reg
->Register
.File
, ®
->Indirect
);
379 unsigned i
, size
= range
.Last
- range
.First
+ 1;
380 LLVMValueRef array
= LLVMBuildInsertElement(builder
,
381 emit_array_fetch(bld_base
, reg
->Register
.File
, TGSI_TYPE_FLOAT
, range
, chan_index
),
382 value
, emit_array_index(bld
, ®
->Indirect
, reg
->Register
.Index
- range
.First
), "");
384 for (i
= 0; i
< size
; ++i
) {
385 switch(reg
->Register
.File
) {
386 case TGSI_FILE_OUTPUT
:
387 temp_ptr
= bld
->outputs
[i
+ range
.First
][chan_index
];
390 case TGSI_FILE_TEMPORARY
:
391 if (uses_temp_indirect_addressing(bld_base
))
392 temp_ptr
= lp_get_temp_ptr_soa(bld
, i
+ range
.First
, chan_index
);
394 temp_ptr
= ctx
->temps
[(i
+ range
.First
) * TGSI_NUM_CHANNELS
+ chan_index
];
400 value
= LLVMBuildExtractElement(builder
, array
,
401 lp_build_const_int32(gallivm
, i
), "");
402 LLVMBuildStore(builder
, value
, temp_ptr
);
406 switch(reg
->Register
.File
) {
407 case TGSI_FILE_OUTPUT
:
408 temp_ptr
= bld
->outputs
[reg
->Register
.Index
][chan_index
];
411 case TGSI_FILE_TEMPORARY
:
412 if (uses_temp_indirect_addressing(bld_base
))
414 temp_ptr
= ctx
->temps
[ TGSI_NUM_CHANNELS
* reg
->Register
.Index
+ chan_index
];
420 LLVMBuildStore(builder
, value
, temp_ptr
);
425 static void bgnloop_emit(
426 const struct lp_build_tgsi_action
* action
,
427 struct lp_build_tgsi_context
* bld_base
,
428 struct lp_build_emit_data
* emit_data
)
430 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
431 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
432 LLVMBasicBlockRef loop_block
;
433 LLVMBasicBlockRef endloop_block
;
434 endloop_block
= LLVMAppendBasicBlockInContext(gallivm
->context
,
435 ctx
->main_fn
, "ENDLOOP");
436 loop_block
= LLVMInsertBasicBlockInContext(gallivm
->context
,
437 endloop_block
, "LOOP");
438 LLVMBuildBr(gallivm
->builder
, loop_block
);
439 LLVMPositionBuilderAtEnd(gallivm
->builder
, loop_block
);
441 ctx
->loop
[ctx
->loop_depth
- 1].loop_block
= loop_block
;
442 ctx
->loop
[ctx
->loop_depth
- 1].endloop_block
= endloop_block
;
445 static void brk_emit(
446 const struct lp_build_tgsi_action
* action
,
447 struct lp_build_tgsi_context
* bld_base
,
448 struct lp_build_emit_data
* emit_data
)
450 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
451 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
452 struct radeon_llvm_loop
* current_loop
= get_current_loop(ctx
);
454 LLVMBuildBr(gallivm
->builder
, current_loop
->endloop_block
);
457 static void cont_emit(
458 const struct lp_build_tgsi_action
* action
,
459 struct lp_build_tgsi_context
* bld_base
,
460 struct lp_build_emit_data
* emit_data
)
462 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
463 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
464 struct radeon_llvm_loop
* current_loop
= get_current_loop(ctx
);
466 LLVMBuildBr(gallivm
->builder
, current_loop
->loop_block
);
469 static void else_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 /* We need to add a terminator to the current block if the previous
480 * instruction was an ENDIF.Example:
493 if (current_block
!= current_branch
->if_block
) {
494 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
496 if (!LLVMGetBasicBlockTerminator(current_branch
->if_block
)) {
497 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
499 current_branch
->has_else
= 1;
500 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->else_block
);
503 static void endif_emit(
504 const struct lp_build_tgsi_action
* action
,
505 struct lp_build_tgsi_context
* bld_base
,
506 struct lp_build_emit_data
* emit_data
)
508 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
509 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
510 struct radeon_llvm_branch
* current_branch
= get_current_branch(ctx
);
511 LLVMBasicBlockRef current_block
= LLVMGetInsertBlock(gallivm
->builder
);
513 /* If we have consecutive ENDIF instructions, then the first ENDIF
514 * will not have a terminator, so we need to add one. */
515 if (current_block
!= current_branch
->if_block
516 && current_block
!= current_branch
->else_block
517 && !LLVMGetBasicBlockTerminator(current_block
)) {
519 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
521 if (!LLVMGetBasicBlockTerminator(current_branch
->else_block
)) {
522 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->else_block
);
523 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
526 if (!LLVMGetBasicBlockTerminator(current_branch
->if_block
)) {
527 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->if_block
);
528 LLVMBuildBr(gallivm
->builder
, current_branch
->endif_block
);
531 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_branch
->endif_block
);
535 static void endloop_emit(
536 const struct lp_build_tgsi_action
* action
,
537 struct lp_build_tgsi_context
* bld_base
,
538 struct lp_build_emit_data
* emit_data
)
540 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
541 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
542 struct radeon_llvm_loop
* current_loop
= get_current_loop(ctx
);
544 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(gallivm
->builder
))) {
545 LLVMBuildBr(gallivm
->builder
, current_loop
->loop_block
);
548 LLVMPositionBuilderAtEnd(gallivm
->builder
, current_loop
->endloop_block
);
552 static void if_cond_emit(
553 const struct lp_build_tgsi_action
* action
,
554 struct lp_build_tgsi_context
* bld_base
,
555 struct lp_build_emit_data
* emit_data
,
558 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
559 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
560 LLVMBasicBlockRef if_block
, else_block
, endif_block
;
562 endif_block
= LLVMAppendBasicBlockInContext(gallivm
->context
,
563 ctx
->main_fn
, "ENDIF");
564 if_block
= LLVMInsertBasicBlockInContext(gallivm
->context
,
566 else_block
= LLVMInsertBasicBlockInContext(gallivm
->context
,
567 endif_block
, "ELSE");
568 LLVMBuildCondBr(gallivm
->builder
, cond
, if_block
, else_block
);
569 LLVMPositionBuilderAtEnd(gallivm
->builder
, if_block
);
572 ctx
->branch
[ctx
->branch_depth
- 1].endif_block
= endif_block
;
573 ctx
->branch
[ctx
->branch_depth
- 1].if_block
= if_block
;
574 ctx
->branch
[ctx
->branch_depth
- 1].else_block
= else_block
;
575 ctx
->branch
[ctx
->branch_depth
- 1].has_else
= 0;
579 const struct lp_build_tgsi_action
* action
,
580 struct lp_build_tgsi_context
* bld_base
,
581 struct lp_build_emit_data
* emit_data
)
583 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
586 cond
= LLVMBuildFCmp(gallivm
->builder
, LLVMRealUNE
,
588 bld_base
->base
.zero
, "");
590 if_cond_emit(action
, bld_base
, emit_data
, cond
);
593 static void uif_emit(
594 const struct lp_build_tgsi_action
* action
,
595 struct lp_build_tgsi_context
* bld_base
,
596 struct lp_build_emit_data
* emit_data
)
598 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
601 cond
= LLVMBuildICmp(gallivm
->builder
, LLVMIntNE
,
602 bitcast(bld_base
, TGSI_TYPE_UNSIGNED
, emit_data
->args
[0]),
603 bld_base
->int_bld
.zero
, "");
605 if_cond_emit(action
, bld_base
, emit_data
, cond
);
608 static void kil_emit(
609 const struct lp_build_tgsi_action
* action
,
610 struct lp_build_tgsi_context
* bld_base
,
611 struct lp_build_emit_data
* emit_data
)
614 for (i
= 0; i
< emit_data
->arg_count
; i
++) {
615 emit_data
->output
[i
] = lp_build_intrinsic_unary(
616 bld_base
->base
.gallivm
->builder
,
618 emit_data
->dst_type
, emit_data
->args
[i
]);
622 void radeon_llvm_emit_prepare_cube_coords(
623 struct lp_build_tgsi_context
* bld_base
,
624 struct lp_build_emit_data
* emit_data
,
625 LLVMValueRef
*coords_arg
)
628 unsigned target
= emit_data
->inst
->Texture
.Texture
;
629 unsigned opcode
= emit_data
->inst
->Instruction
.Opcode
;
630 struct gallivm_state
* gallivm
= bld_base
->base
.gallivm
;
631 LLVMBuilderRef builder
= gallivm
->builder
;
632 LLVMTypeRef type
= bld_base
->base
.elem_type
;
633 LLVMValueRef coords
[4];
634 LLVMValueRef mad_args
[3];
636 struct LLVMOpaqueValue
*cube_vec
;
640 cube_vec
= lp_build_gather_values(bld_base
->base
.gallivm
, coords_arg
, 4);
641 v
= build_intrinsic(builder
, "llvm.AMDGPU.cube", LLVMVectorType(type
, 4),
642 &cube_vec
, 1, LLVMReadNoneAttribute
);
644 for (i
= 0; i
< 4; ++i
) {
645 idx
= lp_build_const_int32(gallivm
, i
);
646 coords
[i
] = LLVMBuildExtractElement(builder
, v
, idx
, "");
649 coords
[2] = build_intrinsic(builder
, "fabs",
650 type
, &coords
[2], 1, LLVMReadNoneAttribute
);
651 coords
[2] = lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_RCP
, coords
[2]);
653 mad_args
[1] = coords
[2];
654 mad_args
[2] = LLVMConstReal(type
, 1.5);
656 mad_args
[0] = coords
[0];
657 coords
[0] = lp_build_emit_llvm_ternary(bld_base
, TGSI_OPCODE_MAD
,
658 mad_args
[0], mad_args
[1], mad_args
[2]);
660 mad_args
[0] = coords
[1];
661 coords
[1] = lp_build_emit_llvm_ternary(bld_base
, TGSI_OPCODE_MAD
,
662 mad_args
[0], mad_args
[1], mad_args
[2]);
664 /* apply xyz = yxw swizzle to cooords */
665 coords
[2] = coords
[3];
666 coords
[3] = coords
[1];
667 coords
[1] = coords
[0];
668 coords
[0] = coords
[3];
670 if (target
== TGSI_TEXTURE_CUBE_ARRAY
||
671 target
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) {
672 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
673 /* coords_arg.w component - array_index for cube arrays */
674 coords
[2] = lp_build_emit_llvm_ternary(bld_base
, TGSI_OPCODE_MAD
,
675 coords_arg
[3], lp_build_const_float(gallivm
, 8.0), coords
[2]);
678 /* Preserve compare/lod/bias. Put it in coords.w. */
679 if (opcode
== TGSI_OPCODE_TEX2
||
680 opcode
== TGSI_OPCODE_TXB2
||
681 opcode
== TGSI_OPCODE_TXL2
) {
682 coords
[3] = coords_arg
[4];
683 } else if (opcode
== TGSI_OPCODE_TXB
||
684 opcode
== TGSI_OPCODE_TXL
||
685 target
== TGSI_TEXTURE_SHADOWCUBE
) {
686 coords
[3] = coords_arg
[3];
689 memcpy(coords_arg
, coords
, sizeof(coords
));
692 static void txd_fetch_args(
693 struct lp_build_tgsi_context
* bld_base
,
694 struct lp_build_emit_data
* emit_data
)
696 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
698 LLVMValueRef coords
[4];
700 for (src
= 0; src
< 3; src
++) {
701 for (chan
= 0; chan
< 4; chan
++)
702 coords
[chan
] = lp_build_emit_fetch(bld_base
, inst
, src
, chan
);
704 emit_data
->args
[src
] = lp_build_gather_values(bld_base
->base
.gallivm
,
707 emit_data
->arg_count
= 3;
708 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
712 static void txp_fetch_args(
713 struct lp_build_tgsi_context
* bld_base
,
714 struct lp_build_emit_data
* emit_data
)
716 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
719 LLVMValueRef coords
[4];
721 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
722 src_w
= lp_build_emit_fetch(bld_base
, emit_data
->inst
, 0, TGSI_CHAN_W
);
724 for (chan
= 0; chan
< 3; chan
++ ) {
725 LLVMValueRef arg
= lp_build_emit_fetch(bld_base
,
726 emit_data
->inst
, 0, chan
);
727 coords
[chan
] = lp_build_emit_llvm_binary(bld_base
,
728 TGSI_OPCODE_DIV
, arg
, src_w
);
730 coords
[3] = bld_base
->base
.one
;
732 if ((inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE
||
733 inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
||
734 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE
||
735 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) &&
736 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ
&&
737 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ_LZ
) {
738 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
);
741 emit_data
->args
[0] = lp_build_gather_values(bld_base
->base
.gallivm
,
743 emit_data
->arg_count
= 1;
746 static void tex_fetch_args(
747 struct lp_build_tgsi_context
* bld_base
,
748 struct lp_build_emit_data
* emit_data
)
750 /* XXX: lp_build_swizzle_aos() was failing with wrong arg types,
751 * when we used CHAN_ALL. We should be able to get this to work,
752 * but for now we will swizzle it ourselves
753 emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
758 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
760 LLVMValueRef coords
[5];
762 for (chan
= 0; chan
< 4; chan
++) {
763 coords
[chan
] = lp_build_emit_fetch(bld_base
, inst
, 0, chan
);
766 if (inst
->Instruction
.Opcode
== TGSI_OPCODE_TEX2
||
767 inst
->Instruction
.Opcode
== TGSI_OPCODE_TXB2
||
768 inst
->Instruction
.Opcode
== TGSI_OPCODE_TXL2
) {
769 /* These instructions have additional operand that should be packed
770 * into the cube coord vector by radeon_llvm_emit_prepare_cube_coords.
771 * That operand should be passed as a float value in the args array
772 * right after the coord vector. After packing it's not used anymore,
773 * that's why arg_count is not increased */
774 coords
[4] = lp_build_emit_fetch(bld_base
, inst
, 1, 0);
777 if ((inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE
||
778 inst
->Texture
.Texture
== TGSI_TEXTURE_CUBE_ARRAY
||
779 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE
||
780 inst
->Texture
.Texture
== TGSI_TEXTURE_SHADOWCUBE_ARRAY
) &&
781 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ
&&
782 inst
->Instruction
.Opcode
!= TGSI_OPCODE_TXQ_LZ
) {
783 radeon_llvm_emit_prepare_cube_coords(bld_base
, emit_data
, coords
);
786 emit_data
->arg_count
= 1;
787 emit_data
->args
[0] = lp_build_gather_values(bld_base
->base
.gallivm
,
789 emit_data
->dst_type
= LLVMVectorType(bld_base
->base
.elem_type
, 4);
792 static void txf_fetch_args(
793 struct lp_build_tgsi_context
* bld_base
,
794 struct lp_build_emit_data
* emit_data
)
796 const struct tgsi_full_instruction
* inst
= emit_data
->inst
;
797 struct lp_build_tgsi_soa_context
*bld
= lp_soa_context(bld_base
);
798 const struct tgsi_texture_offset
* off
= inst
->TexOffsets
;
799 LLVMTypeRef offset_type
= bld_base
->int_bld
.elem_type
;
801 /* fetch tex coords */
802 tex_fetch_args(bld_base
, emit_data
);
804 /* fetch tex offsets */
805 if (inst
->Texture
.NumOffsets
) {
806 assert(inst
->Texture
.NumOffsets
== 1);
808 emit_data
->args
[1] = LLVMConstBitCast(
809 bld
->immediates
[off
->Index
][off
->SwizzleX
],
811 emit_data
->args
[2] = LLVMConstBitCast(
812 bld
->immediates
[off
->Index
][off
->SwizzleY
],
814 emit_data
->args
[3] = LLVMConstBitCast(
815 bld
->immediates
[off
->Index
][off
->SwizzleZ
],
818 emit_data
->args
[1] = bld_base
->int_bld
.zero
;
819 emit_data
->args
[2] = bld_base
->int_bld
.zero
;
820 emit_data
->args
[3] = bld_base
->int_bld
.zero
;
823 emit_data
->arg_count
= 4;
826 static void emit_icmp(
827 const struct lp_build_tgsi_action
* action
,
828 struct lp_build_tgsi_context
* bld_base
,
829 struct lp_build_emit_data
* emit_data
)
832 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
833 LLVMContextRef context
= bld_base
->base
.gallivm
->context
;
835 switch (emit_data
->inst
->Instruction
.Opcode
) {
836 case TGSI_OPCODE_USEQ
: pred
= LLVMIntEQ
; break;
837 case TGSI_OPCODE_USNE
: pred
= LLVMIntNE
; break;
838 case TGSI_OPCODE_USGE
: pred
= LLVMIntUGE
; break;
839 case TGSI_OPCODE_USLT
: pred
= LLVMIntULT
; break;
840 case TGSI_OPCODE_ISGE
: pred
= LLVMIntSGE
; break;
841 case TGSI_OPCODE_ISLT
: pred
= LLVMIntSLT
; break;
843 assert(!"unknown instruction");
848 LLVMValueRef v
= LLVMBuildICmp(builder
, pred
,
849 emit_data
->args
[0], emit_data
->args
[1],"");
851 v
= LLVMBuildSExtOrBitCast(builder
, v
,
852 LLVMInt32TypeInContext(context
), "");
854 emit_data
->output
[emit_data
->chan
] = v
;
857 static void emit_ucmp(
858 const struct lp_build_tgsi_action
* action
,
859 struct lp_build_tgsi_context
* bld_base
,
860 struct lp_build_emit_data
* emit_data
)
862 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
864 LLVMValueRef arg0
= LLVMBuildBitCast(builder
, emit_data
->args
[0],
865 bld_base
->uint_bld
.elem_type
, "");
867 LLVMValueRef v
= LLVMBuildICmp(builder
, LLVMIntNE
, arg0
,
868 bld_base
->uint_bld
.zero
, "");
870 emit_data
->output
[emit_data
->chan
] =
871 LLVMBuildSelect(builder
, v
, emit_data
->args
[1], emit_data
->args
[2], "");
874 static void emit_cmp(
875 const struct lp_build_tgsi_action
*action
,
876 struct lp_build_tgsi_context
* bld_base
,
877 struct lp_build_emit_data
* emit_data
)
879 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
880 LLVMRealPredicate pred
;
883 /* Use ordered for everything but NE (which is usual for
886 switch (emit_data
->inst
->Instruction
.Opcode
) {
887 case TGSI_OPCODE_SGE
: pred
= LLVMRealOGE
; break;
888 case TGSI_OPCODE_SEQ
: pred
= LLVMRealOEQ
; break;
889 case TGSI_OPCODE_SLE
: pred
= LLVMRealOLE
; break;
890 case TGSI_OPCODE_SLT
: pred
= LLVMRealOLT
; break;
891 case TGSI_OPCODE_SNE
: pred
= LLVMRealUNE
; break;
892 case TGSI_OPCODE_SGT
: pred
= LLVMRealOGT
; break;
893 default: assert(!"unknown instruction"); pred
= 0; break;
896 cond
= LLVMBuildFCmp(builder
,
897 pred
, emit_data
->args
[0], emit_data
->args
[1], "");
899 emit_data
->output
[emit_data
->chan
] = LLVMBuildSelect(builder
,
900 cond
, bld_base
->base
.one
, bld_base
->base
.zero
, "");
903 static void emit_fcmp(
904 const struct lp_build_tgsi_action
*action
,
905 struct lp_build_tgsi_context
* bld_base
,
906 struct lp_build_emit_data
* emit_data
)
908 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
909 LLVMContextRef context
= bld_base
->base
.gallivm
->context
;
910 LLVMRealPredicate pred
;
912 /* Use ordered for everything but NE (which is usual for
915 switch (emit_data
->inst
->Instruction
.Opcode
) {
916 case TGSI_OPCODE_FSEQ
: pred
= LLVMRealOEQ
; break;
917 case TGSI_OPCODE_FSGE
: pred
= LLVMRealOGE
; break;
918 case TGSI_OPCODE_FSLT
: pred
= LLVMRealOLT
; break;
919 case TGSI_OPCODE_FSNE
: pred
= LLVMRealUNE
; break;
920 default: assert(!"unknown instruction"); pred
= 0; break;
923 LLVMValueRef v
= LLVMBuildFCmp(builder
, pred
,
924 emit_data
->args
[0], emit_data
->args
[1],"");
926 v
= LLVMBuildSExtOrBitCast(builder
, v
,
927 LLVMInt32TypeInContext(context
), "");
929 emit_data
->output
[emit_data
->chan
] = v
;
932 static void emit_not(
933 const struct lp_build_tgsi_action
* action
,
934 struct lp_build_tgsi_context
* bld_base
,
935 struct lp_build_emit_data
* emit_data
)
937 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
938 LLVMValueRef v
= bitcast(bld_base
, TGSI_TYPE_UNSIGNED
,
940 emit_data
->output
[emit_data
->chan
] = LLVMBuildNot(builder
, v
, "");
943 static void emit_arl(
944 const struct lp_build_tgsi_action
* action
,
945 struct lp_build_tgsi_context
* bld_base
,
946 struct lp_build_emit_data
* emit_data
)
948 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
949 LLVMValueRef floor_index
= lp_build_emit_llvm_unary(bld_base
, TGSI_OPCODE_FLR
, emit_data
->args
[0]);
950 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToSI(builder
,
951 floor_index
, bld_base
->base
.int_elem_type
, "");
954 static void emit_and(
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
] = LLVMBuildAnd(builder
,
961 emit_data
->args
[0], emit_data
->args
[1], "");
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
] = LLVMBuildOr(builder
,
971 emit_data
->args
[0], emit_data
->args
[1], "");
974 static void emit_uadd(
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
] = LLVMBuildAdd(builder
,
981 emit_data
->args
[0], emit_data
->args
[1], "");
984 static void emit_udiv(
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
] = LLVMBuildUDiv(builder
,
991 emit_data
->args
[0], emit_data
->args
[1], "");
994 static void emit_idiv(
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
] = LLVMBuildSDiv(builder
,
1001 emit_data
->args
[0], emit_data
->args
[1], "");
1004 static void emit_mod(
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
] = LLVMBuildSRem(builder
,
1011 emit_data
->args
[0], emit_data
->args
[1], "");
1014 static void emit_umod(
1015 const struct lp_build_tgsi_action
* action
,
1016 struct lp_build_tgsi_context
* bld_base
,
1017 struct lp_build_emit_data
* emit_data
)
1019 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1020 emit_data
->output
[emit_data
->chan
] = LLVMBuildURem(builder
,
1021 emit_data
->args
[0], emit_data
->args
[1], "");
1024 static void emit_shl(
1025 const struct lp_build_tgsi_action
* action
,
1026 struct lp_build_tgsi_context
* bld_base
,
1027 struct lp_build_emit_data
* emit_data
)
1029 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1030 emit_data
->output
[emit_data
->chan
] = LLVMBuildShl(builder
,
1031 emit_data
->args
[0], emit_data
->args
[1], "");
1034 static void emit_ushr(
1035 const struct lp_build_tgsi_action
* action
,
1036 struct lp_build_tgsi_context
* bld_base
,
1037 struct lp_build_emit_data
* emit_data
)
1039 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1040 emit_data
->output
[emit_data
->chan
] = LLVMBuildLShr(builder
,
1041 emit_data
->args
[0], emit_data
->args
[1], "");
1043 static void emit_ishr(
1044 const struct lp_build_tgsi_action
* action
,
1045 struct lp_build_tgsi_context
* bld_base
,
1046 struct lp_build_emit_data
* emit_data
)
1048 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1049 emit_data
->output
[emit_data
->chan
] = LLVMBuildAShr(builder
,
1050 emit_data
->args
[0], emit_data
->args
[1], "");
1053 static void emit_xor(
1054 const struct lp_build_tgsi_action
* action
,
1055 struct lp_build_tgsi_context
* bld_base
,
1056 struct lp_build_emit_data
* emit_data
)
1058 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1059 emit_data
->output
[emit_data
->chan
] = LLVMBuildXor(builder
,
1060 emit_data
->args
[0], emit_data
->args
[1], "");
1063 static void emit_ssg(
1064 const struct lp_build_tgsi_action
* action
,
1065 struct lp_build_tgsi_context
* bld_base
,
1066 struct lp_build_emit_data
* emit_data
)
1068 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1070 LLVMValueRef cmp
, val
;
1072 if (emit_data
->inst
->Instruction
.Opcode
== TGSI_OPCODE_ISSG
) {
1073 cmp
= LLVMBuildICmp(builder
, LLVMIntSGT
, emit_data
->args
[0], bld_base
->int_bld
.zero
, "");
1074 val
= LLVMBuildSelect(builder
, cmp
, bld_base
->int_bld
.one
, emit_data
->args
[0], "");
1075 cmp
= LLVMBuildICmp(builder
, LLVMIntSGE
, val
, bld_base
->int_bld
.zero
, "");
1076 val
= LLVMBuildSelect(builder
, cmp
, val
, LLVMConstInt(bld_base
->int_bld
.elem_type
, -1, true), "");
1077 } else { // float SSG
1078 cmp
= LLVMBuildFCmp(builder
, LLVMRealUGT
, emit_data
->args
[0], bld_base
->base
.zero
, "");
1079 val
= LLVMBuildSelect(builder
, cmp
, bld_base
->base
.one
, emit_data
->args
[0], "");
1080 cmp
= LLVMBuildFCmp(builder
, LLVMRealUGE
, val
, bld_base
->base
.zero
, "");
1081 val
= LLVMBuildSelect(builder
, cmp
, val
, LLVMConstReal(bld_base
->base
.elem_type
, -1), "");
1084 emit_data
->output
[emit_data
->chan
] = val
;
1087 static void emit_ineg(
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
] = LLVMBuildNeg(builder
,
1094 emit_data
->args
[0], "");
1097 static void emit_f2i(
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
] = LLVMBuildFPToSI(builder
,
1104 emit_data
->args
[0], bld_base
->int_bld
.elem_type
, "");
1107 static void emit_f2u(
1108 const struct lp_build_tgsi_action
* action
,
1109 struct lp_build_tgsi_context
* bld_base
,
1110 struct lp_build_emit_data
* emit_data
)
1112 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1113 emit_data
->output
[emit_data
->chan
] = LLVMBuildFPToUI(builder
,
1114 emit_data
->args
[0], bld_base
->uint_bld
.elem_type
, "");
1117 static void emit_i2f(
1118 const struct lp_build_tgsi_action
* action
,
1119 struct lp_build_tgsi_context
* bld_base
,
1120 struct lp_build_emit_data
* emit_data
)
1122 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1123 emit_data
->output
[emit_data
->chan
] = LLVMBuildSIToFP(builder
,
1124 emit_data
->args
[0], bld_base
->base
.elem_type
, "");
1127 static void emit_u2f(
1128 const struct lp_build_tgsi_action
* action
,
1129 struct lp_build_tgsi_context
* bld_base
,
1130 struct lp_build_emit_data
* emit_data
)
1132 LLVMBuilderRef builder
= bld_base
->base
.gallivm
->builder
;
1133 emit_data
->output
[emit_data
->chan
] = LLVMBuildUIToFP(builder
,
1134 emit_data
->args
[0], bld_base
->base
.elem_type
, "");
1137 static void emit_immediate(struct lp_build_tgsi_context
* bld_base
,
1138 const struct tgsi_full_immediate
*imm
)
1141 struct radeon_llvm_context
* ctx
= radeon_llvm_context(bld_base
);
1143 for (i
= 0; i
< 4; ++i
) {
1144 ctx
->soa
.immediates
[ctx
->soa
.num_immediates
][i
] =
1145 LLVMConstInt(bld_base
->uint_bld
.elem_type
, imm
->u
[i
].Uint
, false );
1148 ctx
->soa
.num_immediates
++;
1152 build_intrinsic(LLVMBuilderRef builder
,
1154 LLVMTypeRef ret_type
,
1159 LLVMModuleRef module
= LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder
)));
1160 LLVMValueRef function
;
1162 function
= LLVMGetNamedFunction(module
, name
);
1164 LLVMTypeRef arg_types
[LP_MAX_FUNC_ARGS
];
1167 assert(num_args
<= LP_MAX_FUNC_ARGS
);
1169 for(i
= 0; i
< num_args
; ++i
) {
1171 arg_types
[i
] = LLVMTypeOf(args
[i
]);
1174 function
= lp_declare_intrinsic(module
, name
, ret_type
, arg_types
, num_args
);
1177 LLVMAddFunctionAttr(function
, attr
);
1180 return LLVMBuildCall(builder
, function
, args
, num_args
, "");
1183 static void build_tgsi_intrinsic(
1184 const struct lp_build_tgsi_action
* action
,
1185 struct lp_build_tgsi_context
* bld_base
,
1186 struct lp_build_emit_data
* emit_data
,
1189 struct lp_build_context
* base
= &bld_base
->base
;
1190 emit_data
->output
[emit_data
->chan
] = build_intrinsic(
1191 base
->gallivm
->builder
, action
->intr_name
,
1192 emit_data
->dst_type
, emit_data
->args
,
1193 emit_data
->arg_count
, attr
);
1196 build_tgsi_intrinsic_nomem(
1197 const struct lp_build_tgsi_action
* action
,
1198 struct lp_build_tgsi_context
* bld_base
,
1199 struct lp_build_emit_data
* emit_data
)
1201 build_tgsi_intrinsic(action
, bld_base
, emit_data
, LLVMReadNoneAttribute
);
1204 static void build_tgsi_intrinsic_readonly(
1205 const struct lp_build_tgsi_action
* action
,
1206 struct lp_build_tgsi_context
* bld_base
,
1207 struct lp_build_emit_data
* emit_data
)
1209 build_tgsi_intrinsic(action
, bld_base
, emit_data
, LLVMReadOnlyAttribute
);
1212 void radeon_llvm_context_init(struct radeon_llvm_context
* ctx
)
1214 struct lp_type type
;
1216 /* Initialize the gallivm object:
1217 * We are only using the module, context, and builder fields of this struct.
1218 * This should be enough for us to be able to pass our gallivm struct to the
1219 * helper functions in the gallivm module.
1221 memset(&ctx
->gallivm
, 0, sizeof (ctx
->gallivm
));
1222 memset(&ctx
->soa
, 0, sizeof(ctx
->soa
));
1223 ctx
->gallivm
.context
= LLVMContextCreate();
1224 ctx
->gallivm
.module
= LLVMModuleCreateWithNameInContext("tgsi",
1225 ctx
->gallivm
.context
);
1226 ctx
->gallivm
.builder
= LLVMCreateBuilderInContext(ctx
->gallivm
.context
);
1228 ctx
->store_output_intr
= "llvm.AMDGPU.store.output.";
1229 ctx
->swizzle_intr
= "llvm.AMDGPU.swizzle";
1230 struct lp_build_tgsi_context
* bld_base
= &ctx
->soa
.bld_base
;
1232 /* XXX: We need to revisit this.I think the correct way to do this is
1233 * to use length = 4 here and use the elem_bld for everything. */
1234 type
.floating
= TRUE
;
1241 lp_build_context_init(&bld_base
->base
, &ctx
->gallivm
, type
);
1242 lp_build_context_init(&ctx
->soa
.bld_base
.uint_bld
, &ctx
->gallivm
, lp_uint_type(type
));
1243 lp_build_context_init(&ctx
->soa
.bld_base
.int_bld
, &ctx
->gallivm
, lp_int_type(type
));
1246 bld_base
->emit_store
= emit_store
;
1247 bld_base
->emit_swizzle
= emit_swizzle
;
1248 bld_base
->emit_declaration
= emit_declaration
;
1249 bld_base
->emit_immediate
= emit_immediate
;
1251 bld_base
->emit_fetch_funcs
[TGSI_FILE_IMMEDIATE
] = emit_fetch
;
1252 bld_base
->emit_fetch_funcs
[TGSI_FILE_INPUT
] = emit_fetch
;
1253 bld_base
->emit_fetch_funcs
[TGSI_FILE_TEMPORARY
] = emit_fetch
;
1254 bld_base
->emit_fetch_funcs
[TGSI_FILE_OUTPUT
] = emit_fetch
;
1255 bld_base
->emit_fetch_funcs
[TGSI_FILE_SYSTEM_VALUE
] = fetch_system_value
;
1257 /* Allocate outputs */
1258 ctx
->soa
.outputs
= ctx
->outputs
;
1260 ctx
->num_arrays
= 0;
1262 /* XXX: Is there a better way to initialize all this ? */
1264 lp_set_default_actions(bld_base
);
1266 bld_base
->op_actions
[TGSI_OPCODE_ABS
].emit
= build_tgsi_intrinsic_readonly
;
1267 bld_base
->op_actions
[TGSI_OPCODE_ABS
].intr_name
= "fabs";
1268 bld_base
->op_actions
[TGSI_OPCODE_ARL
].emit
= emit_arl
;
1269 bld_base
->op_actions
[TGSI_OPCODE_AND
].emit
= emit_and
;
1270 bld_base
->op_actions
[TGSI_OPCODE_BGNLOOP
].emit
= bgnloop_emit
;
1271 bld_base
->op_actions
[TGSI_OPCODE_BRK
].emit
= brk_emit
;
1272 bld_base
->op_actions
[TGSI_OPCODE_CEIL
].emit
= build_tgsi_intrinsic_readonly
;
1273 bld_base
->op_actions
[TGSI_OPCODE_CEIL
].intr_name
= "ceil";
1274 bld_base
->op_actions
[TGSI_OPCODE_CLAMP
].emit
= build_tgsi_intrinsic_nomem
;
1275 bld_base
->op_actions
[TGSI_OPCODE_CLAMP
].intr_name
= "llvm.AMDIL.clamp.";
1276 bld_base
->op_actions
[TGSI_OPCODE_CMP
].emit
= build_tgsi_intrinsic_nomem
;
1277 bld_base
->op_actions
[TGSI_OPCODE_CMP
].intr_name
= "llvm.AMDGPU.cndlt";
1278 bld_base
->op_actions
[TGSI_OPCODE_CONT
].emit
= cont_emit
;
1279 bld_base
->op_actions
[TGSI_OPCODE_COS
].emit
= build_tgsi_intrinsic_readonly
;
1280 bld_base
->op_actions
[TGSI_OPCODE_COS
].intr_name
= "llvm.cos.f32";
1281 bld_base
->op_actions
[TGSI_OPCODE_DDX
].intr_name
= "llvm.AMDGPU.ddx";
1282 bld_base
->op_actions
[TGSI_OPCODE_DDX
].fetch_args
= tex_fetch_args
;
1283 bld_base
->op_actions
[TGSI_OPCODE_DDY
].intr_name
= "llvm.AMDGPU.ddy";
1284 bld_base
->op_actions
[TGSI_OPCODE_DDY
].fetch_args
= tex_fetch_args
;
1285 bld_base
->op_actions
[TGSI_OPCODE_ELSE
].emit
= else_emit
;
1286 bld_base
->op_actions
[TGSI_OPCODE_ENDIF
].emit
= endif_emit
;
1287 bld_base
->op_actions
[TGSI_OPCODE_ENDLOOP
].emit
= endloop_emit
;
1288 bld_base
->op_actions
[TGSI_OPCODE_EX2
].emit
= build_tgsi_intrinsic_nomem
;
1289 bld_base
->op_actions
[TGSI_OPCODE_EX2
].intr_name
= "llvm.AMDIL.exp.";
1290 bld_base
->op_actions
[TGSI_OPCODE_FLR
].emit
= build_tgsi_intrinsic_readonly
;
1291 bld_base
->op_actions
[TGSI_OPCODE_FLR
].intr_name
= "floor";
1292 bld_base
->op_actions
[TGSI_OPCODE_FRC
].emit
= build_tgsi_intrinsic_nomem
;
1293 bld_base
->op_actions
[TGSI_OPCODE_FRC
].intr_name
= "llvm.AMDIL.fraction.";
1294 bld_base
->op_actions
[TGSI_OPCODE_F2I
].emit
= emit_f2i
;
1295 bld_base
->op_actions
[TGSI_OPCODE_F2U
].emit
= emit_f2u
;
1296 bld_base
->op_actions
[TGSI_OPCODE_FSEQ
].emit
= emit_fcmp
;
1297 bld_base
->op_actions
[TGSI_OPCODE_FSGE
].emit
= emit_fcmp
;
1298 bld_base
->op_actions
[TGSI_OPCODE_FSLT
].emit
= emit_fcmp
;
1299 bld_base
->op_actions
[TGSI_OPCODE_FSNE
].emit
= emit_fcmp
;
1300 bld_base
->op_actions
[TGSI_OPCODE_IABS
].emit
= build_tgsi_intrinsic_nomem
;
1301 bld_base
->op_actions
[TGSI_OPCODE_IABS
].intr_name
= "llvm.AMDIL.abs.";
1302 bld_base
->op_actions
[TGSI_OPCODE_IDIV
].emit
= emit_idiv
;
1303 bld_base
->op_actions
[TGSI_OPCODE_IF
].emit
= if_emit
;
1304 bld_base
->op_actions
[TGSI_OPCODE_UIF
].emit
= uif_emit
;
1305 bld_base
->op_actions
[TGSI_OPCODE_IMAX
].emit
= build_tgsi_intrinsic_nomem
;
1306 bld_base
->op_actions
[TGSI_OPCODE_IMAX
].intr_name
= "llvm.AMDGPU.imax";
1307 bld_base
->op_actions
[TGSI_OPCODE_IMIN
].emit
= build_tgsi_intrinsic_nomem
;
1308 bld_base
->op_actions
[TGSI_OPCODE_IMIN
].intr_name
= "llvm.AMDGPU.imin";
1309 bld_base
->op_actions
[TGSI_OPCODE_INEG
].emit
= emit_ineg
;
1310 bld_base
->op_actions
[TGSI_OPCODE_ISHR
].emit
= emit_ishr
;
1311 bld_base
->op_actions
[TGSI_OPCODE_ISGE
].emit
= emit_icmp
;
1312 bld_base
->op_actions
[TGSI_OPCODE_ISLT
].emit
= emit_icmp
;
1313 bld_base
->op_actions
[TGSI_OPCODE_ISSG
].emit
= emit_ssg
;
1314 bld_base
->op_actions
[TGSI_OPCODE_I2F
].emit
= emit_i2f
;
1315 bld_base
->op_actions
[TGSI_OPCODE_KILL_IF
].emit
= kil_emit
;
1316 bld_base
->op_actions
[TGSI_OPCODE_KILL_IF
].intr_name
= "llvm.AMDGPU.kill";
1317 bld_base
->op_actions
[TGSI_OPCODE_KILL
].emit
= lp_build_tgsi_intrinsic
;
1318 bld_base
->op_actions
[TGSI_OPCODE_KILL
].intr_name
= "llvm.AMDGPU.kilp";
1319 bld_base
->op_actions
[TGSI_OPCODE_LG2
].emit
= build_tgsi_intrinsic_readonly
;
1320 bld_base
->op_actions
[TGSI_OPCODE_LG2
].intr_name
= "llvm.log2.f32";
1321 bld_base
->op_actions
[TGSI_OPCODE_LRP
].emit
= build_tgsi_intrinsic_nomem
;
1322 bld_base
->op_actions
[TGSI_OPCODE_LRP
].intr_name
= "llvm.AMDGPU.lrp";
1323 bld_base
->op_actions
[TGSI_OPCODE_MOD
].emit
= emit_mod
;
1324 bld_base
->op_actions
[TGSI_OPCODE_NOT
].emit
= emit_not
;
1325 bld_base
->op_actions
[TGSI_OPCODE_OR
].emit
= emit_or
;
1326 bld_base
->op_actions
[TGSI_OPCODE_POW
].emit
= build_tgsi_intrinsic_readonly
;
1327 bld_base
->op_actions
[TGSI_OPCODE_POW
].intr_name
= "llvm.pow.f32";
1328 bld_base
->op_actions
[TGSI_OPCODE_ROUND
].emit
= build_tgsi_intrinsic_nomem
;
1329 bld_base
->op_actions
[TGSI_OPCODE_ROUND
].intr_name
= "llvm.AMDIL.round.nearest.";
1330 bld_base
->op_actions
[TGSI_OPCODE_SGE
].emit
= emit_cmp
;
1331 bld_base
->op_actions
[TGSI_OPCODE_SEQ
].emit
= emit_cmp
;
1332 bld_base
->op_actions
[TGSI_OPCODE_SHL
].emit
= emit_shl
;
1333 bld_base
->op_actions
[TGSI_OPCODE_SLE
].emit
= emit_cmp
;
1334 bld_base
->op_actions
[TGSI_OPCODE_SLT
].emit
= emit_cmp
;
1335 bld_base
->op_actions
[TGSI_OPCODE_SNE
].emit
= emit_cmp
;
1336 bld_base
->op_actions
[TGSI_OPCODE_SGT
].emit
= emit_cmp
;
1337 bld_base
->op_actions
[TGSI_OPCODE_SIN
].emit
= build_tgsi_intrinsic_readonly
;
1338 bld_base
->op_actions
[TGSI_OPCODE_SIN
].intr_name
= "llvm.sin.f32";
1339 bld_base
->op_actions
[TGSI_OPCODE_SSG
].emit
= emit_ssg
;
1340 bld_base
->op_actions
[TGSI_OPCODE_TEX
].fetch_args
= tex_fetch_args
;
1341 bld_base
->op_actions
[TGSI_OPCODE_TEX
].intr_name
= "llvm.AMDGPU.tex";
1342 bld_base
->op_actions
[TGSI_OPCODE_TEX2
].fetch_args
= tex_fetch_args
;
1343 bld_base
->op_actions
[TGSI_OPCODE_TEX2
].intr_name
= "llvm.AMDGPU.tex";
1344 bld_base
->op_actions
[TGSI_OPCODE_TXB
].fetch_args
= tex_fetch_args
;
1345 bld_base
->op_actions
[TGSI_OPCODE_TXB
].intr_name
= "llvm.AMDGPU.txb";
1346 bld_base
->op_actions
[TGSI_OPCODE_TXB2
].fetch_args
= tex_fetch_args
;
1347 bld_base
->op_actions
[TGSI_OPCODE_TXB2
].intr_name
= "llvm.AMDGPU.txb";
1348 bld_base
->op_actions
[TGSI_OPCODE_TXD
].fetch_args
= txd_fetch_args
;
1349 bld_base
->op_actions
[TGSI_OPCODE_TXD
].intr_name
= "llvm.AMDGPU.txd";
1350 bld_base
->op_actions
[TGSI_OPCODE_TXF
].fetch_args
= txf_fetch_args
;
1351 bld_base
->op_actions
[TGSI_OPCODE_TXF
].intr_name
= "llvm.AMDGPU.txf";
1352 bld_base
->op_actions
[TGSI_OPCODE_TXL
].fetch_args
= tex_fetch_args
;
1353 bld_base
->op_actions
[TGSI_OPCODE_TXL
].intr_name
= "llvm.AMDGPU.txl";
1354 bld_base
->op_actions
[TGSI_OPCODE_TXL2
].fetch_args
= tex_fetch_args
;
1355 bld_base
->op_actions
[TGSI_OPCODE_TXL2
].intr_name
= "llvm.AMDGPU.txl";
1356 bld_base
->op_actions
[TGSI_OPCODE_TXP
].fetch_args
= txp_fetch_args
;
1357 bld_base
->op_actions
[TGSI_OPCODE_TXP
].intr_name
= "llvm.AMDGPU.tex";
1358 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].fetch_args
= tex_fetch_args
;
1359 bld_base
->op_actions
[TGSI_OPCODE_TXQ
].intr_name
= "llvm.AMDGPU.txq";
1360 bld_base
->op_actions
[TGSI_OPCODE_TRUNC
].emit
= build_tgsi_intrinsic_nomem
;
1361 bld_base
->op_actions
[TGSI_OPCODE_TRUNC
].intr_name
= "llvm.AMDGPU.trunc";
1362 bld_base
->op_actions
[TGSI_OPCODE_UADD
].emit
= emit_uadd
;
1363 bld_base
->op_actions
[TGSI_OPCODE_UDIV
].emit
= emit_udiv
;
1364 bld_base
->op_actions
[TGSI_OPCODE_UMAX
].emit
= build_tgsi_intrinsic_nomem
;
1365 bld_base
->op_actions
[TGSI_OPCODE_UMAX
].intr_name
= "llvm.AMDGPU.umax";
1366 bld_base
->op_actions
[TGSI_OPCODE_UMIN
].emit
= build_tgsi_intrinsic_nomem
;
1367 bld_base
->op_actions
[TGSI_OPCODE_UMIN
].intr_name
= "llvm.AMDGPU.umin";
1368 bld_base
->op_actions
[TGSI_OPCODE_UMOD
].emit
= emit_umod
;
1369 bld_base
->op_actions
[TGSI_OPCODE_USEQ
].emit
= emit_icmp
;
1370 bld_base
->op_actions
[TGSI_OPCODE_USGE
].emit
= emit_icmp
;
1371 bld_base
->op_actions
[TGSI_OPCODE_USHR
].emit
= emit_ushr
;
1372 bld_base
->op_actions
[TGSI_OPCODE_USLT
].emit
= emit_icmp
;
1373 bld_base
->op_actions
[TGSI_OPCODE_USNE
].emit
= emit_icmp
;
1374 bld_base
->op_actions
[TGSI_OPCODE_U2F
].emit
= emit_u2f
;
1375 bld_base
->op_actions
[TGSI_OPCODE_XOR
].emit
= emit_xor
;
1376 bld_base
->op_actions
[TGSI_OPCODE_UCMP
].emit
= emit_ucmp
;
1378 bld_base
->rsq_action
.emit
= build_tgsi_intrinsic_nomem
;
1379 bld_base
->rsq_action
.intr_name
= "llvm.AMDGPU.rsq";
1382 void radeon_llvm_create_func(struct radeon_llvm_context
* ctx
,
1383 LLVMTypeRef
*ParamTypes
, unsigned ParamCount
)
1385 LLVMTypeRef main_fn_type
;
1386 LLVMBasicBlockRef main_fn_body
;
1388 /* Setup the function */
1389 main_fn_type
= LLVMFunctionType(LLVMVoidTypeInContext(ctx
->gallivm
.context
),
1390 ParamTypes
, ParamCount
, 0);
1391 ctx
->main_fn
= LLVMAddFunction(ctx
->gallivm
.module
, "main", main_fn_type
);
1392 main_fn_body
= LLVMAppendBasicBlockInContext(ctx
->gallivm
.context
,
1393 ctx
->main_fn
, "main_body");
1394 LLVMPositionBuilderAtEnd(ctx
->gallivm
.builder
, main_fn_body
);
1397 void radeon_llvm_finalize_module(struct radeon_llvm_context
* ctx
)
1399 struct gallivm_state
* gallivm
= ctx
->soa
.bld_base
.base
.gallivm
;
1400 /* End the main function with Return*/
1401 LLVMBuildRetVoid(gallivm
->builder
);
1403 /* Create the pass manager */
1404 ctx
->gallivm
.passmgr
= LLVMCreateFunctionPassManagerForModule(
1407 /* This pass should eliminate all the load and store instructions */
1408 LLVMAddPromoteMemoryToRegisterPass(gallivm
->passmgr
);
1410 /* Add some optimization passes */
1411 LLVMAddScalarReplAggregatesPass(gallivm
->passmgr
);
1412 LLVMAddLICMPass(gallivm
->passmgr
);
1413 LLVMAddAggressiveDCEPass(gallivm
->passmgr
);
1414 LLVMAddCFGSimplificationPass(gallivm
->passmgr
);
1417 LLVMRunFunctionPassManager(gallivm
->passmgr
, ctx
->main_fn
);
1419 LLVMDisposeBuilder(gallivm
->builder
);
1420 LLVMDisposePassManager(gallivm
->passmgr
);
1424 void radeon_llvm_dispose(struct radeon_llvm_context
* ctx
)
1426 LLVMDisposeModule(ctx
->soa
.bld_base
.base
.gallivm
->module
);
1427 LLVMContextDispose(ctx
->soa
.bld_base
.base
.gallivm
->context
);