radeon/llvm: Don't use lp_build_swizzle_aos() for swizzles
[mesa.git] / src / gallium / drivers / radeon / radeon_setup_tgsi_llvm.c
1 /*
2 * Copyright 2011 Advanced Micro Devices, Inc.
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * SOFTWARE.
22 *
23 * Authors: Tom Stellard <thomas.stellard@amd.com>
24 *
25 */
26 #include "radeon_llvm.h"
27
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"
39
40 #include <llvm-c/Core.h>
41 #include <llvm-c/Transforms/Scalar.h>
42
43 static struct radeon_llvm_loop * get_current_loop(struct radeon_llvm_context * ctx)
44 {
45 return ctx->loop_depth > 0 ? ctx->loop + (ctx->loop_depth - 1) : NULL;
46 }
47
48 static struct radeon_llvm_branch * get_current_branch(
49 struct radeon_llvm_context * ctx)
50 {
51 return ctx->branch_depth > 0 ?
52 ctx->branch + (ctx->branch_depth - 1) : NULL;
53 }
54
55 unsigned radeon_llvm_reg_index_soa(unsigned index, unsigned chan)
56 {
57 return (index * 4) + chan;
58 }
59
60 static LLVMValueRef emit_swizzle(
61 struct lp_build_tgsi_context * bld_base,
62 LLVMValueRef value,
63 unsigned swizzle_x,
64 unsigned swizzle_y,
65 unsigned swizzle_z,
66 unsigned swizzle_w)
67 {
68 LLVMValueRef swizzles[4];
69 LLVMTypeRef i32t =
70 LLVMInt32TypeInContext(bld_base->base.gallivm->context);
71
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);
76
77 return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
78 value,
79 LLVMGetUndef(LLVMTypeOf(value)),
80 LLVMConstVector(swizzles, 4), "");
81 }
82
83 static LLVMValueRef
84 emit_array_index(
85 struct lp_build_tgsi_soa_context *bld,
86 const struct tgsi_full_src_register *reg,
87 unsigned swizzle)
88 {
89 struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
90
91 LLVMValueRef addr = LLVMBuildLoad(gallivm->builder,
92 bld->addr[reg->Indirect.Index][swizzle], "");
93 LLVMValueRef offset = lp_build_const_int32(gallivm, reg->Register.Index);
94 LLVMValueRef hw_index = LLVMBuildAdd(gallivm->builder, addr, offset, "");
95 LLVMValueRef soa_index = LLVMBuildMul(gallivm->builder, hw_index,
96 lp_build_const_int32(gallivm, 4), "");
97 LLVMValueRef array_index = LLVMBuildAdd(gallivm->builder, soa_index,
98 lp_build_const_int32(gallivm, swizzle), "");
99
100 return array_index;
101 }
102
103 static LLVMValueRef
104 emit_fetch_immediate(
105 struct lp_build_tgsi_context *bld_base,
106 const struct tgsi_full_src_register *reg,
107 enum tgsi_opcode_type type,
108 unsigned swizzle)
109 {
110 LLVMTypeRef ctype;
111 LLVMContextRef ctx = bld_base->base.gallivm->context;
112
113 switch (type) {
114 case TGSI_TYPE_UNSIGNED:
115 case TGSI_TYPE_SIGNED:
116 ctype = LLVMInt32TypeInContext(ctx);
117 break;
118 case TGSI_TYPE_UNTYPED:
119 case TGSI_TYPE_FLOAT:
120 ctype = LLVMFloatTypeInContext(ctx);
121 break;
122 default:
123 ctype = 0;
124 break;
125 }
126
127 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
128 return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
129 }
130
131 static LLVMValueRef
132 emit_fetch_input(
133 struct lp_build_tgsi_context *bld_base,
134 const struct tgsi_full_src_register *reg,
135 enum tgsi_opcode_type type,
136 unsigned swizzle)
137 {
138 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
139 if (swizzle == ~0) {
140 LLVMValueRef values[TGSI_NUM_CHANNELS] = {};
141 unsigned chan;
142 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
143 values[chan] = ctx->inputs[radeon_llvm_reg_index_soa(
144 reg->Register.Index, chan)];
145 }
146 return lp_build_gather_values(bld_base->base.gallivm, values,
147 TGSI_NUM_CHANNELS);
148 } else {
149 return bitcast(bld_base, type, ctx->inputs[radeon_llvm_reg_index_soa(reg->Register.Index, swizzle)]);
150 }
151 }
152
153 static LLVMValueRef
154 emit_fetch_temporary(
155 struct lp_build_tgsi_context *bld_base,
156 const struct tgsi_full_src_register *reg,
157 enum tgsi_opcode_type type,
158 unsigned swizzle)
159 {
160 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
161 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
162 if (reg->Register.Indirect) {
163 LLVMValueRef array_index = emit_array_index(bld, reg, swizzle);
164 LLVMValueRef ptr = LLVMBuildGEP(builder, bld->temps_array, &array_index,
165 1, "");
166 return LLVMBuildLoad(builder, ptr, "");
167 } else {
168 LLVMValueRef temp_ptr;
169 temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
170 return bitcast(bld_base,type,LLVMBuildLoad(builder, temp_ptr, ""));
171 }
172 }
173
174 static LLVMValueRef
175 emit_fetch_output(
176 struct lp_build_tgsi_context *bld_base,
177 const struct tgsi_full_src_register *reg,
178 enum tgsi_opcode_type type,
179 unsigned swizzle)
180 {
181 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
182 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
183 if (reg->Register.Indirect) {
184 LLVMValueRef array_index = emit_array_index(bld, reg, swizzle);
185 LLVMValueRef ptr = LLVMBuildGEP(builder, bld->outputs_array, &array_index,
186 1, "");
187 return LLVMBuildLoad(builder, ptr, "");
188 } else {
189 LLVMValueRef temp_ptr;
190 temp_ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
191 return LLVMBuildLoad(builder, temp_ptr, "");
192 }
193 }
194
195 static void emit_declaration(
196 struct lp_build_tgsi_context * bld_base,
197 const struct tgsi_full_declaration *decl)
198 {
199 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
200 switch(decl->Declaration.File) {
201 case TGSI_FILE_ADDRESS:
202 {
203 unsigned idx;
204 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
205 unsigned chan;
206 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
207 ctx->soa.addr[idx][chan] = lp_build_alloca(
208 &ctx->gallivm,
209 ctx->soa.bld_base.uint_bld.elem_type, "");
210 }
211 }
212 break;
213 }
214
215 case TGSI_FILE_TEMPORARY:
216 lp_emit_declaration_soa(bld_base, decl);
217 break;
218
219 case TGSI_FILE_INPUT:
220 {
221 unsigned idx;
222 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
223 ctx->load_input(ctx, idx, decl);
224 }
225 }
226 break;
227
228 case TGSI_FILE_SYSTEM_VALUE:
229 {
230 unsigned idx;
231 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
232 ctx->load_system_value(ctx, idx, decl);
233 }
234 }
235 break;
236
237 case TGSI_FILE_OUTPUT:
238 {
239 unsigned idx;
240 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
241 unsigned chan;
242 assert(idx < RADEON_LLVM_MAX_OUTPUTS);
243 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
244 ctx->soa.outputs[idx][chan] = lp_build_alloca(&ctx->gallivm,
245 ctx->soa.bld_base.base.elem_type, "");
246 }
247 }
248
249 ctx->output_reg_count = MAX2(ctx->output_reg_count,
250 decl->Range.Last + 1);
251 break;
252 }
253
254 default:
255 break;
256 }
257 }
258
259 static void
260 emit_store(
261 struct lp_build_tgsi_context * bld_base,
262 const struct tgsi_full_instruction * inst,
263 const struct tgsi_opcode_info * info,
264 LLVMValueRef dst[4])
265 {
266 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
267 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
268 struct lp_build_context base = bld->bld_base.base;
269 const struct tgsi_full_dst_register *reg = &inst->Dst[0];
270 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
271 LLVMValueRef temp_ptr;
272 unsigned chan, chan_index;
273 boolean is_vec_store = FALSE;
274 if (dst[0]) {
275 LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
276 is_vec_store = (k == LLVMVectorTypeKind);
277 }
278
279 if (is_vec_store) {
280 LLVMValueRef values[4] = {};
281 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan) {
282 LLVMValueRef index = lp_build_const_int32(gallivm, chan);
283 values[chan] = LLVMBuildExtractElement(gallivm->builder,
284 dst[0], index, "");
285 }
286 bld_base->emit_store(bld_base, inst, info, values);
287 return;
288 }
289
290 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
291 LLVMValueRef value = dst[chan_index];
292
293 if (inst->Instruction.Saturate != TGSI_SAT_NONE) {
294 struct lp_build_emit_data clamp_emit_data;
295
296 memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
297 clamp_emit_data.arg_count = 3;
298 clamp_emit_data.args[0] = value;
299 clamp_emit_data.args[2] = base.one;
300
301 switch(inst->Instruction.Saturate) {
302 case TGSI_SAT_ZERO_ONE:
303 clamp_emit_data.args[1] = base.zero;
304 break;
305 case TGSI_SAT_MINUS_PLUS_ONE:
306 clamp_emit_data.args[1] = LLVMConstReal(
307 base.elem_type, -1.0f);
308 break;
309 default:
310 assert(0);
311 }
312 value = lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
313 &clamp_emit_data);
314 }
315
316 switch(reg->Register.File) {
317 case TGSI_FILE_OUTPUT:
318 temp_ptr = bld->outputs[reg->Register.Index][chan_index];
319 break;
320
321 case TGSI_FILE_TEMPORARY:
322 temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, chan_index);
323 break;
324
325 default:
326 return;
327 }
328
329 value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
330
331 LLVMBuildStore(builder, value, temp_ptr);
332 }
333 }
334
335 static void bgnloop_emit(
336 const struct lp_build_tgsi_action * action,
337 struct lp_build_tgsi_context * bld_base,
338 struct lp_build_emit_data * emit_data)
339 {
340 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
341 struct gallivm_state * gallivm = bld_base->base.gallivm;
342 LLVMBasicBlockRef loop_block;
343 LLVMBasicBlockRef endloop_block;
344 endloop_block = LLVMAppendBasicBlockInContext(gallivm->context,
345 ctx->main_fn, "ENDLOOP");
346 loop_block = LLVMInsertBasicBlockInContext(gallivm->context,
347 endloop_block, "LOOP");
348 LLVMBuildBr(gallivm->builder, loop_block);
349 LLVMPositionBuilderAtEnd(gallivm->builder, loop_block);
350 ctx->loop_depth++;
351 ctx->loop[ctx->loop_depth - 1].loop_block = loop_block;
352 ctx->loop[ctx->loop_depth - 1].endloop_block = endloop_block;
353 }
354
355 static void brk_emit(
356 const struct lp_build_tgsi_action * action,
357 struct lp_build_tgsi_context * bld_base,
358 struct lp_build_emit_data * emit_data)
359 {
360 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
361 struct gallivm_state * gallivm = bld_base->base.gallivm;
362 struct radeon_llvm_loop * current_loop = get_current_loop(ctx);
363
364 LLVMBuildBr(gallivm->builder, current_loop->endloop_block);
365 }
366
367 static void cont_emit(
368 const struct lp_build_tgsi_action * action,
369 struct lp_build_tgsi_context * bld_base,
370 struct lp_build_emit_data * emit_data)
371 {
372 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
373 struct gallivm_state * gallivm = bld_base->base.gallivm;
374 struct radeon_llvm_loop * current_loop = get_current_loop(ctx);
375
376 LLVMBuildBr(gallivm->builder, current_loop->loop_block);
377 }
378
379 static void else_emit(
380 const struct lp_build_tgsi_action * action,
381 struct lp_build_tgsi_context * bld_base,
382 struct lp_build_emit_data * emit_data)
383 {
384 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
385 struct gallivm_state * gallivm = bld_base->base.gallivm;
386 struct radeon_llvm_branch * current_branch = get_current_branch(ctx);
387 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(gallivm->builder);
388
389 /* We need to add a terminator to the current block if the previous
390 * instruction was an ENDIF.Example:
391 * IF
392 * [code]
393 * IF
394 * [code]
395 * ELSE
396 * [code]
397 * ENDIF <--
398 * ELSE<--
399 * [code]
400 * ENDIF
401 */
402
403 if (current_block != current_branch->if_block) {
404 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
405 }
406 if (!LLVMGetBasicBlockTerminator(current_branch->if_block)) {
407 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
408 }
409 current_branch->has_else = 1;
410 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->else_block);
411 }
412
413 static void endif_emit(
414 const struct lp_build_tgsi_action * action,
415 struct lp_build_tgsi_context * bld_base,
416 struct lp_build_emit_data * emit_data)
417 {
418 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
419 struct gallivm_state * gallivm = bld_base->base.gallivm;
420 struct radeon_llvm_branch * current_branch = get_current_branch(ctx);
421 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(gallivm->builder);
422
423 /* If we have consecutive ENDIF instructions, then the first ENDIF
424 * will not have a terminator, so we need to add one. */
425 if (current_block != current_branch->if_block
426 && current_block != current_branch->else_block
427 && !LLVMGetBasicBlockTerminator(current_block)) {
428
429 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
430 }
431 if (!LLVMGetBasicBlockTerminator(current_branch->else_block)) {
432 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->else_block);
433 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
434 }
435
436 if (!LLVMGetBasicBlockTerminator(current_branch->if_block)) {
437 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->if_block);
438 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
439 }
440
441 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->endif_block);
442 ctx->branch_depth--;
443 }
444
445 static void endloop_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)
449 {
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);
453
454 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(gallivm->builder))) {
455 LLVMBuildBr(gallivm->builder, current_loop->loop_block);
456 }
457
458 LLVMPositionBuilderAtEnd(gallivm->builder, current_loop->endloop_block);
459 ctx->loop_depth--;
460 }
461
462 static void if_emit(
463 const struct lp_build_tgsi_action * action,
464 struct lp_build_tgsi_context * bld_base,
465 struct lp_build_emit_data * emit_data)
466 {
467 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
468 struct gallivm_state * gallivm = bld_base->base.gallivm;
469 LLVMValueRef cond;
470 LLVMBasicBlockRef if_block, else_block, endif_block;
471
472 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
473 bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
474 bld_base->int_bld.zero, "");
475
476 endif_block = LLVMAppendBasicBlockInContext(gallivm->context,
477 ctx->main_fn, "ENDIF");
478 if_block = LLVMInsertBasicBlockInContext(gallivm->context,
479 endif_block, "IF");
480 else_block = LLVMInsertBasicBlockInContext(gallivm->context,
481 endif_block, "ELSE");
482 LLVMBuildCondBr(gallivm->builder, cond, if_block, else_block);
483 LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
484
485 ctx->branch_depth++;
486 ctx->branch[ctx->branch_depth - 1].endif_block = endif_block;
487 ctx->branch[ctx->branch_depth - 1].if_block = if_block;
488 ctx->branch[ctx->branch_depth - 1].else_block = else_block;
489 ctx->branch[ctx->branch_depth - 1].has_else = 0;
490 }
491
492 static void kil_emit(
493 const struct lp_build_tgsi_action * action,
494 struct lp_build_tgsi_context * bld_base,
495 struct lp_build_emit_data * emit_data)
496 {
497 unsigned i;
498 for (i = 0; i < emit_data->arg_count; i++) {
499 emit_data->output[i] = lp_build_intrinsic_unary(
500 bld_base->base.gallivm->builder,
501 action->intr_name,
502 emit_data->dst_type, emit_data->args[i]);
503 }
504 }
505
506
507 static void emit_prepare_cube_coords(
508 struct lp_build_tgsi_context * bld_base,
509 struct lp_build_emit_data * emit_data)
510 {
511 boolean shadowcube = (emit_data->inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE);
512 struct gallivm_state * gallivm = bld_base->base.gallivm;
513 LLVMBuilderRef builder = gallivm->builder;
514 LLVMTypeRef type = bld_base->base.elem_type;
515 LLVMValueRef coords[4];
516 LLVMValueRef mad_args[3];
517 unsigned i, cnt;
518
519 LLVMValueRef v = build_intrinsic(builder, "llvm.AMDGPU.cube",
520 LLVMVectorType(type, 4),
521 &emit_data->args[0],1, LLVMReadNoneAttribute);
522
523 /* save src.w for shadow cube */
524 cnt = shadowcube ? 3 : 4;
525
526 for (i = 0; i < cnt; ++i) {
527 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
528 coords[i] = LLVMBuildExtractElement(builder, v, idx, "");
529 }
530
531 coords[2] = build_intrinsic(builder, "llvm.AMDIL.fabs.",
532 type, &coords[2], 1, LLVMReadNoneAttribute);
533 coords[2] = build_intrinsic(builder, "llvm.AMDGPU.rcp",
534 type, &coords[2], 1, LLVMReadNoneAttribute);
535
536 mad_args[1] = coords[2];
537 mad_args[2] = LLVMConstReal(type, 1.5);
538
539 mad_args[0] = coords[0];
540 coords[0] = build_intrinsic(builder, "llvm.AMDIL.mad.",
541 type, mad_args, 3, LLVMReadNoneAttribute);
542
543 mad_args[0] = coords[1];
544 coords[1] = build_intrinsic(builder, "llvm.AMDIL.mad.",
545 type, mad_args, 3, LLVMReadNoneAttribute);
546
547 /* apply yxwy swizzle to cooords */
548 coords[2] = coords[3];
549 coords[3] = coords[1];
550 coords[1] = coords[0];
551 coords[0] = coords[3];
552
553 emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
554 coords, 4);
555 }
556
557 static void txd_fetch_args(
558 struct lp_build_tgsi_context * bld_base,
559 struct lp_build_emit_data * emit_data)
560 {
561 const struct tgsi_full_instruction * inst = emit_data->inst;
562
563 LLVMValueRef coords[4];
564 unsigned chan, src;
565 for (src = 0; src < 3; src++) {
566 for (chan = 0; chan < 4; chan++)
567 coords[chan] = lp_build_emit_fetch(bld_base, inst, src, chan);
568
569 emit_data->args[src] = lp_build_gather_values(bld_base->base.gallivm,
570 coords, 4);
571 }
572 emit_data->arg_count = 3;
573 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
574 }
575
576
577 static void txp_fetch_args(
578 struct lp_build_tgsi_context * bld_base,
579 struct lp_build_emit_data * emit_data)
580 {
581 const struct tgsi_full_instruction * inst = emit_data->inst;
582 LLVMValueRef src_w;
583 unsigned chan;
584 LLVMValueRef coords[4];
585
586 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
587 src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
588
589 for (chan = 0; chan < 3; chan++ ) {
590 LLVMValueRef arg = lp_build_emit_fetch(bld_base,
591 emit_data->inst, 0, chan);
592 coords[chan] = lp_build_emit_llvm_binary(bld_base,
593 TGSI_OPCODE_DIV, arg, src_w);
594 }
595 coords[3] = bld_base->base.one;
596 emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
597 coords, 4);
598 emit_data->arg_count = 1;
599
600 if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
601 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) &&
602 inst->Instruction.Opcode != TGSI_OPCODE_TXQ) {
603 emit_prepare_cube_coords(bld_base, emit_data);
604 }
605 }
606
607 static void tex_fetch_args(
608 struct lp_build_tgsi_context * bld_base,
609 struct lp_build_emit_data * emit_data)
610 {
611 /* XXX: lp_build_swizzle_aos() was failing with wrong arg types,
612 * when we used CHAN_ALL. We should be able to get this to work,
613 * but for now we will swizzle it ourselves
614 emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
615 0, CHAN_ALL);
616
617 */
618
619 const struct tgsi_full_instruction * inst = emit_data->inst;
620
621 LLVMValueRef coords[4];
622 unsigned chan;
623 for (chan = 0; chan < 4; chan++) {
624 coords[chan] = lp_build_emit_fetch(bld_base, inst, 0, chan);
625 }
626
627 emit_data->arg_count = 1;
628 emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
629 coords, 4);
630 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
631
632 if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
633 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) &&
634 inst->Instruction.Opcode != TGSI_OPCODE_TXQ) {
635 emit_prepare_cube_coords(bld_base, emit_data);
636 }
637 }
638
639 static void txf_fetch_args(
640 struct lp_build_tgsi_context * bld_base,
641 struct lp_build_emit_data * emit_data)
642 {
643 const struct tgsi_full_instruction * inst = emit_data->inst;
644 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
645 const struct tgsi_texture_offset * off = inst->TexOffsets;
646 LLVMTypeRef offset_type = bld_base->int_bld.elem_type;
647
648 /* fetch tex coords */
649 tex_fetch_args(bld_base, emit_data);
650
651 /* fetch tex offsets */
652 if (inst->Texture.NumOffsets) {
653 assert(inst->Texture.NumOffsets == 1);
654
655 emit_data->args[1] = LLVMConstBitCast(
656 bld->immediates[off->Index][off->SwizzleX],
657 offset_type);
658 emit_data->args[2] = LLVMConstBitCast(
659 bld->immediates[off->Index][off->SwizzleY],
660 offset_type);
661 emit_data->args[3] = LLVMConstBitCast(
662 bld->immediates[off->Index][off->SwizzleZ],
663 offset_type);
664 } else {
665 emit_data->args[1] = bld_base->int_bld.zero;
666 emit_data->args[2] = bld_base->int_bld.zero;
667 emit_data->args[3] = bld_base->int_bld.zero;
668 }
669
670 emit_data->arg_count = 4;
671 }
672
673 static void emit_icmp(
674 const struct lp_build_tgsi_action * action,
675 struct lp_build_tgsi_context * bld_base,
676 struct lp_build_emit_data * emit_data)
677 {
678 unsigned pred;
679 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
680 LLVMContextRef context = bld_base->base.gallivm->context;
681
682 switch (emit_data->inst->Instruction.Opcode) {
683 case TGSI_OPCODE_USEQ: pred = LLVMIntEQ; break;
684 case TGSI_OPCODE_USNE: pred = LLVMIntNE; break;
685 case TGSI_OPCODE_USGE: pred = LLVMIntUGE; break;
686 case TGSI_OPCODE_USLT: pred = LLVMIntULT; break;
687 case TGSI_OPCODE_ISGE: pred = LLVMIntSGE; break;
688 case TGSI_OPCODE_ISLT: pred = LLVMIntSLT; break;
689 default:
690 assert(!"unknown instruction");
691 }
692
693 LLVMValueRef v = LLVMBuildICmp(builder, pred,
694 emit_data->args[0], emit_data->args[1],"");
695
696 v = LLVMBuildSExtOrBitCast(builder, v,
697 LLVMInt32TypeInContext(context), "");
698
699 emit_data->output[emit_data->chan] = v;
700 }
701
702 static void emit_cmp(
703 const struct lp_build_tgsi_action *action,
704 struct lp_build_tgsi_context * bld_base,
705 struct lp_build_emit_data * emit_data)
706 {
707 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
708 LLVMRealPredicate pred;
709 LLVMValueRef cond;
710
711 /* XXX I'm not sure whether to do unordered or ordered comparisons,
712 * but llvmpipe uses unordered comparisons, so for consistency we use
713 * unordered. (The authors of llvmpipe aren't sure about using
714 * unordered vs ordered comparisons either.
715 */
716 switch (emit_data->inst->Instruction.Opcode) {
717 case TGSI_OPCODE_SGE: pred = LLVMRealUGE; break;
718 case TGSI_OPCODE_SEQ: pred = LLVMRealUEQ; break;
719 case TGSI_OPCODE_SLE: pred = LLVMRealULE; break;
720 case TGSI_OPCODE_SLT: pred = LLVMRealULT; break;
721 case TGSI_OPCODE_SNE: pred = LLVMRealUNE; break;
722 case TGSI_OPCODE_SGT: pred = LLVMRealUGT; break;
723 default: assert(!"unknown instruction");
724 }
725
726 cond = LLVMBuildFCmp(builder,
727 pred, emit_data->args[0], emit_data->args[1], "");
728
729 emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
730 cond, bld_base->base.one, bld_base->base.zero, "");
731 }
732
733 static void emit_not(
734 const struct lp_build_tgsi_action * action,
735 struct lp_build_tgsi_context * bld_base,
736 struct lp_build_emit_data * emit_data)
737 {
738 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
739 LLVMValueRef v = bitcast(bld_base, TGSI_TYPE_UNSIGNED,
740 emit_data->args[0]);
741 emit_data->output[emit_data->chan] = LLVMBuildNot(builder, v, "");
742 }
743
744 static void emit_and(
745 const struct lp_build_tgsi_action * action,
746 struct lp_build_tgsi_context * bld_base,
747 struct lp_build_emit_data * emit_data)
748 {
749 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
750 emit_data->output[emit_data->chan] = LLVMBuildAnd(builder,
751 emit_data->args[0], emit_data->args[1], "");
752 }
753
754 static void emit_or(
755 const struct lp_build_tgsi_action * action,
756 struct lp_build_tgsi_context * bld_base,
757 struct lp_build_emit_data * emit_data)
758 {
759 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
760 emit_data->output[emit_data->chan] = LLVMBuildOr(builder,
761 emit_data->args[0], emit_data->args[1], "");
762 }
763
764 static void emit_uadd(
765 const struct lp_build_tgsi_action * action,
766 struct lp_build_tgsi_context * bld_base,
767 struct lp_build_emit_data * emit_data)
768 {
769 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
770 emit_data->output[emit_data->chan] = LLVMBuildAdd(builder,
771 emit_data->args[0], emit_data->args[1], "");
772 }
773
774 static void emit_udiv(
775 const struct lp_build_tgsi_action * action,
776 struct lp_build_tgsi_context * bld_base,
777 struct lp_build_emit_data * emit_data)
778 {
779 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
780 emit_data->output[emit_data->chan] = LLVMBuildUDiv(builder,
781 emit_data->args[0], emit_data->args[1], "");
782 }
783
784 static void emit_idiv(
785 const struct lp_build_tgsi_action * action,
786 struct lp_build_tgsi_context * bld_base,
787 struct lp_build_emit_data * emit_data)
788 {
789 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
790 emit_data->output[emit_data->chan] = LLVMBuildSDiv(builder,
791 emit_data->args[0], emit_data->args[1], "");
792 }
793
794 static void emit_mod(
795 const struct lp_build_tgsi_action * action,
796 struct lp_build_tgsi_context * bld_base,
797 struct lp_build_emit_data * emit_data)
798 {
799 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
800 emit_data->output[emit_data->chan] = LLVMBuildSRem(builder,
801 emit_data->args[0], emit_data->args[1], "");
802 }
803
804 static void emit_umod(
805 const struct lp_build_tgsi_action * action,
806 struct lp_build_tgsi_context * bld_base,
807 struct lp_build_emit_data * emit_data)
808 {
809 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
810 emit_data->output[emit_data->chan] = LLVMBuildURem(builder,
811 emit_data->args[0], emit_data->args[1], "");
812 }
813
814 static void emit_shl(
815 const struct lp_build_tgsi_action * action,
816 struct lp_build_tgsi_context * bld_base,
817 struct lp_build_emit_data * emit_data)
818 {
819 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
820 emit_data->output[emit_data->chan] = LLVMBuildShl(builder,
821 emit_data->args[0], emit_data->args[1], "");
822 }
823
824 static void emit_ushr(
825 const struct lp_build_tgsi_action * action,
826 struct lp_build_tgsi_context * bld_base,
827 struct lp_build_emit_data * emit_data)
828 {
829 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
830 emit_data->output[emit_data->chan] = LLVMBuildLShr(builder,
831 emit_data->args[0], emit_data->args[1], "");
832 }
833 static void emit_ishr(
834 const struct lp_build_tgsi_action * action,
835 struct lp_build_tgsi_context * bld_base,
836 struct lp_build_emit_data * emit_data)
837 {
838 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
839 emit_data->output[emit_data->chan] = LLVMBuildAShr(builder,
840 emit_data->args[0], emit_data->args[1], "");
841 }
842
843 static void emit_xor(
844 const struct lp_build_tgsi_action * action,
845 struct lp_build_tgsi_context * bld_base,
846 struct lp_build_emit_data * emit_data)
847 {
848 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
849 emit_data->output[emit_data->chan] = LLVMBuildXor(builder,
850 emit_data->args[0], emit_data->args[1], "");
851 }
852
853 static void emit_ssg(
854 const struct lp_build_tgsi_action * action,
855 struct lp_build_tgsi_context * bld_base,
856 struct lp_build_emit_data * emit_data)
857 {
858 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
859
860 LLVMValueRef cmp, val;
861
862 if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
863 cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
864 val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
865 cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
866 val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
867 } else { // float SSG
868 cmp = LLVMBuildFCmp(builder, LLVMRealUGT, emit_data->args[0], bld_base->int_bld.zero, "");
869 val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
870 cmp = LLVMBuildFCmp(builder, LLVMRealUGE, val, bld_base->base.zero, "");
871 val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
872 }
873
874 emit_data->output[emit_data->chan] = val;
875 }
876
877 static void emit_ineg(
878 const struct lp_build_tgsi_action * action,
879 struct lp_build_tgsi_context * bld_base,
880 struct lp_build_emit_data * emit_data)
881 {
882 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
883 emit_data->output[emit_data->chan] = LLVMBuildNeg(builder,
884 emit_data->args[0], "");
885 }
886
887 static void emit_f2i(
888 const struct lp_build_tgsi_action * action,
889 struct lp_build_tgsi_context * bld_base,
890 struct lp_build_emit_data * emit_data)
891 {
892 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
893 emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
894 emit_data->args[0], bld_base->int_bld.elem_type, "");
895 }
896
897 static void emit_f2u(
898 const struct lp_build_tgsi_action * action,
899 struct lp_build_tgsi_context * bld_base,
900 struct lp_build_emit_data * emit_data)
901 {
902 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
903 emit_data->output[emit_data->chan] = LLVMBuildFPToUI(builder,
904 emit_data->args[0], bld_base->uint_bld.elem_type, "");
905 }
906
907 static void emit_i2f(
908 const struct lp_build_tgsi_action * action,
909 struct lp_build_tgsi_context * bld_base,
910 struct lp_build_emit_data * emit_data)
911 {
912 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
913 emit_data->output[emit_data->chan] = LLVMBuildSIToFP(builder,
914 emit_data->args[0], bld_base->base.elem_type, "");
915 }
916
917 static void emit_u2f(
918 const struct lp_build_tgsi_action * action,
919 struct lp_build_tgsi_context * bld_base,
920 struct lp_build_emit_data * emit_data)
921 {
922 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
923 emit_data->output[emit_data->chan] = LLVMBuildUIToFP(builder,
924 emit_data->args[0], bld_base->base.elem_type, "");
925 }
926
927 static void emit_immediate(struct lp_build_tgsi_context * bld_base,
928 const struct tgsi_full_immediate *imm)
929 {
930 unsigned i;
931 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
932
933 for (i = 0; i < 4; ++i) {
934 ctx->soa.immediates[ctx->soa.num_immediates][i] =
935 LLVMConstInt(bld_base->uint_bld.elem_type, imm->u[i].Uint, false );
936 }
937
938 ctx->soa.num_immediates++;
939 }
940
941 LLVMValueRef
942 build_intrinsic(LLVMBuilderRef builder,
943 const char *name,
944 LLVMTypeRef ret_type,
945 LLVMValueRef *args,
946 unsigned num_args,
947 LLVMAttribute attr)
948 {
949 LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
950 LLVMValueRef function;
951
952 function = LLVMGetNamedFunction(module, name);
953 if(!function) {
954 LLVMTypeRef arg_types[LP_MAX_FUNC_ARGS];
955 unsigned i;
956
957 assert(num_args <= LP_MAX_FUNC_ARGS);
958
959 for(i = 0; i < num_args; ++i) {
960 assert(args[i]);
961 arg_types[i] = LLVMTypeOf(args[i]);
962 }
963
964 function = lp_declare_intrinsic(module, name, ret_type, arg_types, num_args);
965
966 if (attr)
967 LLVMAddFunctionAttr(function, attr);
968 }
969
970 return LLVMBuildCall(builder, function, args, num_args, "");
971 }
972
973 void
974 build_tgsi_intrinsic_nomem(
975 const struct lp_build_tgsi_action * action,
976 struct lp_build_tgsi_context * bld_base,
977 struct lp_build_emit_data * emit_data)
978 {
979 struct lp_build_context * base = &bld_base->base;
980 emit_data->output[emit_data->chan] = build_intrinsic(
981 base->gallivm->builder, action->intr_name,
982 emit_data->dst_type, emit_data->args,
983 emit_data->arg_count, LLVMReadNoneAttribute);
984 }
985
986 void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
987 {
988 struct lp_type type;
989 LLVMTypeRef main_fn_type;
990 LLVMBasicBlockRef main_fn_body;
991
992 /* Initialize the gallivm object:
993 * We are only using the module, context, and builder fields of this struct.
994 * This should be enough for us to be able to pass our gallivm struct to the
995 * helper functions in the gallivm module.
996 */
997 memset(&ctx->gallivm, 0, sizeof (ctx->gallivm));
998 memset(&ctx->soa, 0, sizeof(ctx->soa));
999 ctx->gallivm.context = LLVMContextCreate();
1000 ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
1001 ctx->gallivm.context);
1002 ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
1003
1004 /* Setup the module */
1005 main_fn_type = LLVMFunctionType(LLVMVoidTypeInContext(ctx->gallivm.context),
1006 NULL, 0, 0);
1007 ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
1008 main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
1009 ctx->main_fn, "main_body");
1010 LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
1011
1012 ctx->store_output_intr = "llvm.AMDGPU.store.output.";
1013 ctx->swizzle_intr = "llvm.AMDGPU.swizzle";
1014 struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
1015
1016 /* XXX: We need to revisit this.I think the correct way to do this is
1017 * to use length = 4 here and use the elem_bld for everything. */
1018 type.floating = TRUE;
1019 type.sign = TRUE;
1020 type.width = 32;
1021 type.length = 1;
1022
1023 lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
1024 lp_build_context_init(&ctx->soa.bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
1025 lp_build_context_init(&ctx->soa.bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
1026
1027 bld_base->soa = 1;
1028 bld_base->emit_store = emit_store;
1029 bld_base->emit_swizzle = emit_swizzle;
1030 bld_base->emit_declaration = emit_declaration;
1031 bld_base->emit_immediate = emit_immediate;
1032
1033 bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
1034 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
1035 bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
1036 bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = emit_fetch_output;
1037
1038 /* Allocate outputs */
1039 ctx->soa.outputs = ctx->outputs;
1040
1041 /* XXX: Is there a better way to initialize all this ? */
1042
1043 lp_set_default_actions(bld_base);
1044
1045 bld_base->op_actions[TGSI_OPCODE_IABS].emit = build_tgsi_intrinsic_nomem;
1046 bld_base->op_actions[TGSI_OPCODE_IABS].intr_name = "llvm.AMDIL.abs.";
1047 bld_base->op_actions[TGSI_OPCODE_NOT].emit = emit_not;
1048 bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
1049 bld_base->op_actions[TGSI_OPCODE_XOR].emit = emit_xor;
1050 bld_base->op_actions[TGSI_OPCODE_OR].emit = emit_or;
1051 bld_base->op_actions[TGSI_OPCODE_UADD].emit = emit_uadd;
1052 bld_base->op_actions[TGSI_OPCODE_UDIV].emit = emit_udiv;
1053 bld_base->op_actions[TGSI_OPCODE_IDIV].emit = emit_idiv;
1054 bld_base->op_actions[TGSI_OPCODE_MOD].emit = emit_mod;
1055 bld_base->op_actions[TGSI_OPCODE_UMOD].emit = emit_umod;
1056 bld_base->op_actions[TGSI_OPCODE_INEG].emit = emit_ineg;
1057 bld_base->op_actions[TGSI_OPCODE_SHL].emit = emit_shl;
1058 bld_base->op_actions[TGSI_OPCODE_ISHR].emit = emit_ishr;
1059 bld_base->op_actions[TGSI_OPCODE_USHR].emit = emit_ushr;
1060 bld_base->op_actions[TGSI_OPCODE_SSG].emit = emit_ssg;
1061 bld_base->op_actions[TGSI_OPCODE_ISSG].emit = emit_ssg;
1062 bld_base->op_actions[TGSI_OPCODE_I2F].emit = emit_i2f;
1063 bld_base->op_actions[TGSI_OPCODE_U2F].emit = emit_u2f;
1064 bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
1065 bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
1066 bld_base->op_actions[TGSI_OPCODE_DDX].intr_name = "llvm.AMDGPU.ddx";
1067 bld_base->op_actions[TGSI_OPCODE_DDX].fetch_args = tex_fetch_args;
1068 bld_base->op_actions[TGSI_OPCODE_DDY].intr_name = "llvm.AMDGPU.ddy";
1069 bld_base->op_actions[TGSI_OPCODE_DDY].fetch_args = tex_fetch_args;
1070 bld_base->op_actions[TGSI_OPCODE_USEQ].emit = emit_icmp;
1071 bld_base->op_actions[TGSI_OPCODE_USGE].emit = emit_icmp;
1072 bld_base->op_actions[TGSI_OPCODE_USLT].emit = emit_icmp;
1073 bld_base->op_actions[TGSI_OPCODE_USNE].emit = emit_icmp;
1074 bld_base->op_actions[TGSI_OPCODE_ISGE].emit = emit_icmp;
1075 bld_base->op_actions[TGSI_OPCODE_ISLT].emit = emit_icmp;
1076 bld_base->op_actions[TGSI_OPCODE_ROUND].emit = build_tgsi_intrinsic_nomem;
1077 bld_base->op_actions[TGSI_OPCODE_ROUND].intr_name = "llvm.AMDIL.round.nearest.";
1078 bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
1079 bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.AMDIL.min.";
1080 bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
1081 bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.AMDIL.max.";
1082 bld_base->op_actions[TGSI_OPCODE_IMIN].emit = build_tgsi_intrinsic_nomem;
1083 bld_base->op_actions[TGSI_OPCODE_IMIN].intr_name = "llvm.AMDGPU.imin";
1084 bld_base->op_actions[TGSI_OPCODE_IMAX].emit = build_tgsi_intrinsic_nomem;
1085 bld_base->op_actions[TGSI_OPCODE_IMAX].intr_name = "llvm.AMDGPU.imax";
1086 bld_base->op_actions[TGSI_OPCODE_UMIN].emit = build_tgsi_intrinsic_nomem;
1087 bld_base->op_actions[TGSI_OPCODE_UMIN].intr_name = "llvm.AMDGPU.umin";
1088 bld_base->op_actions[TGSI_OPCODE_UMAX].emit = build_tgsi_intrinsic_nomem;
1089 bld_base->op_actions[TGSI_OPCODE_UMAX].intr_name = "llvm.AMDGPU.umax";
1090 bld_base->op_actions[TGSI_OPCODE_TXF].fetch_args = txf_fetch_args;
1091 bld_base->op_actions[TGSI_OPCODE_TXF].intr_name = "llvm.AMDGPU.txf";
1092 bld_base->op_actions[TGSI_OPCODE_TXQ].fetch_args = tex_fetch_args;
1093 bld_base->op_actions[TGSI_OPCODE_TXQ].intr_name = "llvm.AMDGPU.txq";
1094 bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_nomem;
1095 bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "llvm.AMDIL.round.posinf.";
1096
1097
1098
1099 bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_nomem;
1100 bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "llvm.AMDIL.fabs.";
1101 bld_base->op_actions[TGSI_OPCODE_ARL].emit = build_tgsi_intrinsic_nomem;
1102 bld_base->op_actions[TGSI_OPCODE_ARL].intr_name = "llvm.AMDGPU.arl";
1103 bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
1104 bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
1105 bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
1106 bld_base->op_actions[TGSI_OPCODE_CLAMP].emit = build_tgsi_intrinsic_nomem;
1107 bld_base->op_actions[TGSI_OPCODE_CLAMP].intr_name = "llvm.AMDIL.clamp.";
1108 bld_base->op_actions[TGSI_OPCODE_CMP].emit = build_tgsi_intrinsic_nomem;
1109 bld_base->op_actions[TGSI_OPCODE_CMP].intr_name = "llvm.AMDGPU.cndlt";
1110 bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_nomem;
1111 bld_base->op_actions[TGSI_OPCODE_COS].intr_name = "llvm.AMDGPU.cos";
1112 bld_base->op_actions[TGSI_OPCODE_DIV].emit = build_tgsi_intrinsic_nomem;
1113 bld_base->op_actions[TGSI_OPCODE_DIV].intr_name = "llvm.AMDGPU.div";
1114 bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
1115 bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
1116 bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1117 bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
1118 bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.AMDIL.exp.";
1119 bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;
1120 bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "llvm.AMDGPU.floor";
1121 bld_base->op_actions[TGSI_OPCODE_FRC].emit = build_tgsi_intrinsic_nomem;
1122 bld_base->op_actions[TGSI_OPCODE_FRC].intr_name = "llvm.AMDIL.fraction.";
1123 bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
1124 bld_base->op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
1125 bld_base->op_actions[TGSI_OPCODE_KIL].intr_name = "llvm.AMDGPU.kill";
1126 bld_base->op_actions[TGSI_OPCODE_KILP].emit = lp_build_tgsi_intrinsic;
1127 bld_base->op_actions[TGSI_OPCODE_KILP].intr_name = "llvm.AMDGPU.kilp";
1128 bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_nomem;
1129 bld_base->op_actions[TGSI_OPCODE_LG2].intr_name = "llvm.AMDIL.log.";
1130 bld_base->op_actions[TGSI_OPCODE_LRP].emit = build_tgsi_intrinsic_nomem;
1131 bld_base->op_actions[TGSI_OPCODE_LRP].intr_name = "llvm.AMDGPU.lrp";
1132 bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
1133 bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.AMDIL.min.";
1134 bld_base->op_actions[TGSI_OPCODE_MAD].emit = build_tgsi_intrinsic_nomem;
1135 bld_base->op_actions[TGSI_OPCODE_MAD].intr_name = "llvm.AMDIL.mad.";
1136 bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
1137 bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.AMDIL.max.";
1138 bld_base->op_actions[TGSI_OPCODE_MUL].emit = build_tgsi_intrinsic_nomem;
1139 bld_base->op_actions[TGSI_OPCODE_MUL].intr_name = "llvm.AMDGPU.mul";
1140 bld_base->op_actions[TGSI_OPCODE_POW].emit = build_tgsi_intrinsic_nomem;
1141 bld_base->op_actions[TGSI_OPCODE_POW].intr_name = "llvm.AMDGPU.pow";
1142 bld_base->op_actions[TGSI_OPCODE_RCP].emit = build_tgsi_intrinsic_nomem;
1143 bld_base->op_actions[TGSI_OPCODE_RCP].intr_name = "llvm.AMDGPU.rcp";
1144 bld_base->op_actions[TGSI_OPCODE_SSG].emit = build_tgsi_intrinsic_nomem;
1145 bld_base->op_actions[TGSI_OPCODE_SSG].intr_name = "llvm.AMDGPU.ssg";
1146 bld_base->op_actions[TGSI_OPCODE_SGE].emit = emit_cmp;
1147 bld_base->op_actions[TGSI_OPCODE_SEQ].emit = emit_cmp;
1148 bld_base->op_actions[TGSI_OPCODE_SLE].emit = emit_cmp;
1149 bld_base->op_actions[TGSI_OPCODE_SLT].emit = emit_cmp;
1150 bld_base->op_actions[TGSI_OPCODE_SNE].emit = emit_cmp;
1151 bld_base->op_actions[TGSI_OPCODE_SGT].emit = emit_cmp;
1152 bld_base->op_actions[TGSI_OPCODE_SIN].emit = build_tgsi_intrinsic_nomem;
1153 bld_base->op_actions[TGSI_OPCODE_SIN].intr_name = "llvm.AMDGPU.sin";
1154 bld_base->op_actions[TGSI_OPCODE_TEX].fetch_args = tex_fetch_args;
1155 bld_base->op_actions[TGSI_OPCODE_TEX].intr_name = "llvm.AMDGPU.tex";
1156 bld_base->op_actions[TGSI_OPCODE_TXB].fetch_args = tex_fetch_args;
1157 bld_base->op_actions[TGSI_OPCODE_TXB].intr_name = "llvm.AMDGPU.txb";
1158 bld_base->op_actions[TGSI_OPCODE_TXD].fetch_args = txd_fetch_args;
1159 bld_base->op_actions[TGSI_OPCODE_TXD].intr_name = "llvm.AMDGPU.txd";
1160 bld_base->op_actions[TGSI_OPCODE_TXL].fetch_args = tex_fetch_args;
1161 bld_base->op_actions[TGSI_OPCODE_TXL].intr_name = "llvm.AMDGPU.txl";
1162 bld_base->op_actions[TGSI_OPCODE_TXP].fetch_args = txp_fetch_args;
1163 bld_base->op_actions[TGSI_OPCODE_TXP].intr_name = "llvm.AMDGPU.tex";
1164 bld_base->op_actions[TGSI_OPCODE_TRUNC].emit = build_tgsi_intrinsic_nomem;
1165 bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.AMDGPU.trunc";
1166
1167 bld_base->rsq_action.emit = build_tgsi_intrinsic_nomem;
1168 bld_base->rsq_action.intr_name = "llvm.AMDGPU.rsq";
1169 }
1170
1171 void radeon_llvm_finalize_module(struct radeon_llvm_context * ctx)
1172 {
1173 struct gallivm_state * gallivm = ctx->soa.bld_base.base.gallivm;
1174 /* End the main function with Return*/
1175 LLVMBuildRetVoid(gallivm->builder);
1176
1177 /* Create the pass manager */
1178 ctx->gallivm.passmgr = LLVMCreateFunctionPassManagerForModule(
1179 gallivm->module);
1180
1181 /* This pass should eliminate all the load and store instructions */
1182 LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
1183
1184 /* Add some optimization passes */
1185 LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
1186 LLVMAddCFGSimplificationPass(gallivm->passmgr);
1187
1188 /* Run the passs */
1189 LLVMRunFunctionPassManager(gallivm->passmgr, ctx->main_fn);
1190
1191 LLVMDisposeBuilder(gallivm->builder);
1192 LLVMDisposePassManager(gallivm->passmgr);
1193
1194 }
1195
1196 void radeon_llvm_dispose(struct radeon_llvm_context * ctx)
1197 {
1198 LLVMDisposeModule(ctx->soa.bld_base.base.gallivm->module);
1199 LLVMContextDispose(ctx->soa.bld_base.base.gallivm->context);
1200 }