radeon/llvm: add support to fetch temps as vectors
[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 (swizzle == ~0) {
163 LLVMValueRef values[TGSI_NUM_CHANNELS] = {};
164 unsigned chan;
165 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
166 values[chan] = emit_fetch_temporary(bld_base, reg, type, chan);
167 }
168 return lp_build_gather_values(bld_base->base.gallivm, values,
169 TGSI_NUM_CHANNELS);
170 }
171
172 if (reg->Register.Indirect) {
173 LLVMValueRef array_index = emit_array_index(bld, reg, swizzle);
174 LLVMValueRef ptr = LLVMBuildGEP(builder, bld->temps_array, &array_index,
175 1, "");
176 return LLVMBuildLoad(builder, ptr, "");
177 } else {
178 LLVMValueRef temp_ptr;
179 temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
180 return bitcast(bld_base,type,LLVMBuildLoad(builder, temp_ptr, ""));
181 }
182 }
183
184 static LLVMValueRef
185 emit_fetch_output(
186 struct lp_build_tgsi_context *bld_base,
187 const struct tgsi_full_src_register *reg,
188 enum tgsi_opcode_type type,
189 unsigned swizzle)
190 {
191 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
192 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
193 if (reg->Register.Indirect) {
194 LLVMValueRef array_index = emit_array_index(bld, reg, swizzle);
195 LLVMValueRef ptr = LLVMBuildGEP(builder, bld->outputs_array, &array_index,
196 1, "");
197 return LLVMBuildLoad(builder, ptr, "");
198 } else {
199 LLVMValueRef temp_ptr;
200 temp_ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
201 return LLVMBuildLoad(builder, temp_ptr, "");
202 }
203 }
204
205 static void emit_declaration(
206 struct lp_build_tgsi_context * bld_base,
207 const struct tgsi_full_declaration *decl)
208 {
209 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
210 switch(decl->Declaration.File) {
211 case TGSI_FILE_ADDRESS:
212 {
213 unsigned idx;
214 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
215 unsigned chan;
216 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
217 ctx->soa.addr[idx][chan] = lp_build_alloca(
218 &ctx->gallivm,
219 ctx->soa.bld_base.uint_bld.elem_type, "");
220 }
221 }
222 break;
223 }
224
225 case TGSI_FILE_TEMPORARY:
226 lp_emit_declaration_soa(bld_base, decl);
227 break;
228
229 case TGSI_FILE_INPUT:
230 {
231 unsigned idx;
232 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
233 ctx->load_input(ctx, idx, decl);
234 }
235 }
236 break;
237
238 case TGSI_FILE_SYSTEM_VALUE:
239 {
240 unsigned idx;
241 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
242 ctx->load_system_value(ctx, idx, decl);
243 }
244 }
245 break;
246
247 case TGSI_FILE_OUTPUT:
248 {
249 unsigned idx;
250 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
251 unsigned chan;
252 assert(idx < RADEON_LLVM_MAX_OUTPUTS);
253 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
254 ctx->soa.outputs[idx][chan] = lp_build_alloca(&ctx->gallivm,
255 ctx->soa.bld_base.base.elem_type, "");
256 }
257 }
258
259 ctx->output_reg_count = MAX2(ctx->output_reg_count,
260 decl->Range.Last + 1);
261 break;
262 }
263
264 default:
265 break;
266 }
267 }
268
269 static void
270 emit_store(
271 struct lp_build_tgsi_context * bld_base,
272 const struct tgsi_full_instruction * inst,
273 const struct tgsi_opcode_info * info,
274 LLVMValueRef dst[4])
275 {
276 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
277 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
278 struct lp_build_context base = bld->bld_base.base;
279 const struct tgsi_full_dst_register *reg = &inst->Dst[0];
280 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
281 LLVMValueRef temp_ptr;
282 unsigned chan, chan_index;
283 boolean is_vec_store = FALSE;
284 if (dst[0]) {
285 LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
286 is_vec_store = (k == LLVMVectorTypeKind);
287 }
288
289 if (is_vec_store) {
290 LLVMValueRef values[4] = {};
291 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan) {
292 LLVMValueRef index = lp_build_const_int32(gallivm, chan);
293 values[chan] = LLVMBuildExtractElement(gallivm->builder,
294 dst[0], index, "");
295 }
296 bld_base->emit_store(bld_base, inst, info, values);
297 return;
298 }
299
300 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
301 LLVMValueRef value = dst[chan_index];
302
303 if (inst->Instruction.Saturate != TGSI_SAT_NONE) {
304 struct lp_build_emit_data clamp_emit_data;
305
306 memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
307 clamp_emit_data.arg_count = 3;
308 clamp_emit_data.args[0] = value;
309 clamp_emit_data.args[2] = base.one;
310
311 switch(inst->Instruction.Saturate) {
312 case TGSI_SAT_ZERO_ONE:
313 clamp_emit_data.args[1] = base.zero;
314 break;
315 case TGSI_SAT_MINUS_PLUS_ONE:
316 clamp_emit_data.args[1] = LLVMConstReal(
317 base.elem_type, -1.0f);
318 break;
319 default:
320 assert(0);
321 }
322 value = lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
323 &clamp_emit_data);
324 }
325
326 switch(reg->Register.File) {
327 case TGSI_FILE_OUTPUT:
328 temp_ptr = bld->outputs[reg->Register.Index][chan_index];
329 break;
330
331 case TGSI_FILE_TEMPORARY:
332 temp_ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, chan_index);
333 break;
334
335 default:
336 return;
337 }
338
339 value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
340
341 LLVMBuildStore(builder, value, temp_ptr);
342 }
343 }
344
345 static void bgnloop_emit(
346 const struct lp_build_tgsi_action * action,
347 struct lp_build_tgsi_context * bld_base,
348 struct lp_build_emit_data * emit_data)
349 {
350 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
351 struct gallivm_state * gallivm = bld_base->base.gallivm;
352 LLVMBasicBlockRef loop_block;
353 LLVMBasicBlockRef endloop_block;
354 endloop_block = LLVMAppendBasicBlockInContext(gallivm->context,
355 ctx->main_fn, "ENDLOOP");
356 loop_block = LLVMInsertBasicBlockInContext(gallivm->context,
357 endloop_block, "LOOP");
358 LLVMBuildBr(gallivm->builder, loop_block);
359 LLVMPositionBuilderAtEnd(gallivm->builder, loop_block);
360 ctx->loop_depth++;
361 ctx->loop[ctx->loop_depth - 1].loop_block = loop_block;
362 ctx->loop[ctx->loop_depth - 1].endloop_block = endloop_block;
363 }
364
365 static void brk_emit(
366 const struct lp_build_tgsi_action * action,
367 struct lp_build_tgsi_context * bld_base,
368 struct lp_build_emit_data * emit_data)
369 {
370 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
371 struct gallivm_state * gallivm = bld_base->base.gallivm;
372 struct radeon_llvm_loop * current_loop = get_current_loop(ctx);
373
374 LLVMBuildBr(gallivm->builder, current_loop->endloop_block);
375 }
376
377 static void cont_emit(
378 const struct lp_build_tgsi_action * action,
379 struct lp_build_tgsi_context * bld_base,
380 struct lp_build_emit_data * emit_data)
381 {
382 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
383 struct gallivm_state * gallivm = bld_base->base.gallivm;
384 struct radeon_llvm_loop * current_loop = get_current_loop(ctx);
385
386 LLVMBuildBr(gallivm->builder, current_loop->loop_block);
387 }
388
389 static void else_emit(
390 const struct lp_build_tgsi_action * action,
391 struct lp_build_tgsi_context * bld_base,
392 struct lp_build_emit_data * emit_data)
393 {
394 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
395 struct gallivm_state * gallivm = bld_base->base.gallivm;
396 struct radeon_llvm_branch * current_branch = get_current_branch(ctx);
397 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(gallivm->builder);
398
399 /* We need to add a terminator to the current block if the previous
400 * instruction was an ENDIF.Example:
401 * IF
402 * [code]
403 * IF
404 * [code]
405 * ELSE
406 * [code]
407 * ENDIF <--
408 * ELSE<--
409 * [code]
410 * ENDIF
411 */
412
413 if (current_block != current_branch->if_block) {
414 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
415 }
416 if (!LLVMGetBasicBlockTerminator(current_branch->if_block)) {
417 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
418 }
419 current_branch->has_else = 1;
420 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->else_block);
421 }
422
423 static void endif_emit(
424 const struct lp_build_tgsi_action * action,
425 struct lp_build_tgsi_context * bld_base,
426 struct lp_build_emit_data * emit_data)
427 {
428 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
429 struct gallivm_state * gallivm = bld_base->base.gallivm;
430 struct radeon_llvm_branch * current_branch = get_current_branch(ctx);
431 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(gallivm->builder);
432
433 /* If we have consecutive ENDIF instructions, then the first ENDIF
434 * will not have a terminator, so we need to add one. */
435 if (current_block != current_branch->if_block
436 && current_block != current_branch->else_block
437 && !LLVMGetBasicBlockTerminator(current_block)) {
438
439 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
440 }
441 if (!LLVMGetBasicBlockTerminator(current_branch->else_block)) {
442 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->else_block);
443 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
444 }
445
446 if (!LLVMGetBasicBlockTerminator(current_branch->if_block)) {
447 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->if_block);
448 LLVMBuildBr(gallivm->builder, current_branch->endif_block);
449 }
450
451 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->endif_block);
452 ctx->branch_depth--;
453 }
454
455 static void endloop_emit(
456 const struct lp_build_tgsi_action * action,
457 struct lp_build_tgsi_context * bld_base,
458 struct lp_build_emit_data * emit_data)
459 {
460 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
461 struct gallivm_state * gallivm = bld_base->base.gallivm;
462 struct radeon_llvm_loop * current_loop = get_current_loop(ctx);
463
464 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(gallivm->builder))) {
465 LLVMBuildBr(gallivm->builder, current_loop->loop_block);
466 }
467
468 LLVMPositionBuilderAtEnd(gallivm->builder, current_loop->endloop_block);
469 ctx->loop_depth--;
470 }
471
472 static void if_emit(
473 const struct lp_build_tgsi_action * action,
474 struct lp_build_tgsi_context * bld_base,
475 struct lp_build_emit_data * emit_data)
476 {
477 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
478 struct gallivm_state * gallivm = bld_base->base.gallivm;
479 LLVMValueRef cond;
480 LLVMBasicBlockRef if_block, else_block, endif_block;
481
482 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
483 bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
484 bld_base->int_bld.zero, "");
485
486 endif_block = LLVMAppendBasicBlockInContext(gallivm->context,
487 ctx->main_fn, "ENDIF");
488 if_block = LLVMInsertBasicBlockInContext(gallivm->context,
489 endif_block, "IF");
490 else_block = LLVMInsertBasicBlockInContext(gallivm->context,
491 endif_block, "ELSE");
492 LLVMBuildCondBr(gallivm->builder, cond, if_block, else_block);
493 LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
494
495 ctx->branch_depth++;
496 ctx->branch[ctx->branch_depth - 1].endif_block = endif_block;
497 ctx->branch[ctx->branch_depth - 1].if_block = if_block;
498 ctx->branch[ctx->branch_depth - 1].else_block = else_block;
499 ctx->branch[ctx->branch_depth - 1].has_else = 0;
500 }
501
502 static void kil_emit(
503 const struct lp_build_tgsi_action * action,
504 struct lp_build_tgsi_context * bld_base,
505 struct lp_build_emit_data * emit_data)
506 {
507 unsigned i;
508 for (i = 0; i < emit_data->arg_count; i++) {
509 emit_data->output[i] = lp_build_intrinsic_unary(
510 bld_base->base.gallivm->builder,
511 action->intr_name,
512 emit_data->dst_type, emit_data->args[i]);
513 }
514 }
515
516
517 static void emit_prepare_cube_coords(
518 struct lp_build_tgsi_context * bld_base,
519 struct lp_build_emit_data * emit_data)
520 {
521 boolean shadowcube = (emit_data->inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE);
522 struct gallivm_state * gallivm = bld_base->base.gallivm;
523 LLVMBuilderRef builder = gallivm->builder;
524 LLVMTypeRef type = bld_base->base.elem_type;
525 LLVMValueRef coords[4];
526 LLVMValueRef mad_args[3];
527 unsigned i, cnt;
528
529 LLVMValueRef v = build_intrinsic(builder, "llvm.AMDGPU.cube",
530 LLVMVectorType(type, 4),
531 &emit_data->args[0],1, LLVMReadNoneAttribute);
532
533 /* save src.w for shadow cube */
534 cnt = shadowcube ? 3 : 4;
535
536 for (i = 0; i < cnt; ++i) {
537 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
538 coords[i] = LLVMBuildExtractElement(builder, v, idx, "");
539 }
540
541 coords[2] = build_intrinsic(builder, "llvm.AMDIL.fabs.",
542 type, &coords[2], 1, LLVMReadNoneAttribute);
543 coords[2] = build_intrinsic(builder, "llvm.AMDGPU.rcp",
544 type, &coords[2], 1, LLVMReadNoneAttribute);
545
546 mad_args[1] = coords[2];
547 mad_args[2] = LLVMConstReal(type, 1.5);
548
549 mad_args[0] = coords[0];
550 coords[0] = build_intrinsic(builder, "llvm.AMDIL.mad.",
551 type, mad_args, 3, LLVMReadNoneAttribute);
552
553 mad_args[0] = coords[1];
554 coords[1] = build_intrinsic(builder, "llvm.AMDIL.mad.",
555 type, mad_args, 3, LLVMReadNoneAttribute);
556
557 /* apply yxwy swizzle to cooords */
558 coords[2] = coords[3];
559 coords[3] = coords[1];
560 coords[1] = coords[0];
561 coords[0] = coords[3];
562
563 emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
564 coords, 4);
565 }
566
567 static void txd_fetch_args(
568 struct lp_build_tgsi_context * bld_base,
569 struct lp_build_emit_data * emit_data)
570 {
571 const struct tgsi_full_instruction * inst = emit_data->inst;
572
573 LLVMValueRef coords[4];
574 unsigned chan, src;
575 for (src = 0; src < 3; src++) {
576 for (chan = 0; chan < 4; chan++)
577 coords[chan] = lp_build_emit_fetch(bld_base, inst, src, chan);
578
579 emit_data->args[src] = lp_build_gather_values(bld_base->base.gallivm,
580 coords, 4);
581 }
582 emit_data->arg_count = 3;
583 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
584 }
585
586
587 static void txp_fetch_args(
588 struct lp_build_tgsi_context * bld_base,
589 struct lp_build_emit_data * emit_data)
590 {
591 const struct tgsi_full_instruction * inst = emit_data->inst;
592 LLVMValueRef src_w;
593 unsigned chan;
594 LLVMValueRef coords[4];
595
596 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
597 src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
598
599 for (chan = 0; chan < 3; chan++ ) {
600 LLVMValueRef arg = lp_build_emit_fetch(bld_base,
601 emit_data->inst, 0, chan);
602 coords[chan] = lp_build_emit_llvm_binary(bld_base,
603 TGSI_OPCODE_DIV, arg, src_w);
604 }
605 coords[3] = bld_base->base.one;
606 emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
607 coords, 4);
608 emit_data->arg_count = 1;
609
610 if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
611 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) &&
612 inst->Instruction.Opcode != TGSI_OPCODE_TXQ) {
613 emit_prepare_cube_coords(bld_base, emit_data);
614 }
615 }
616
617 static void tex_fetch_args(
618 struct lp_build_tgsi_context * bld_base,
619 struct lp_build_emit_data * emit_data)
620 {
621 /* XXX: lp_build_swizzle_aos() was failing with wrong arg types,
622 * when we used CHAN_ALL. We should be able to get this to work,
623 * but for now we will swizzle it ourselves
624 emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
625 0, CHAN_ALL);
626
627 */
628
629 const struct tgsi_full_instruction * inst = emit_data->inst;
630
631 LLVMValueRef coords[4];
632 unsigned chan;
633 for (chan = 0; chan < 4; chan++) {
634 coords[chan] = lp_build_emit_fetch(bld_base, inst, 0, chan);
635 }
636
637 emit_data->arg_count = 1;
638 emit_data->args[0] = lp_build_gather_values(bld_base->base.gallivm,
639 coords, 4);
640 emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
641
642 if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
643 inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) &&
644 inst->Instruction.Opcode != TGSI_OPCODE_TXQ) {
645 emit_prepare_cube_coords(bld_base, emit_data);
646 }
647 }
648
649 static void txf_fetch_args(
650 struct lp_build_tgsi_context * bld_base,
651 struct lp_build_emit_data * emit_data)
652 {
653 const struct tgsi_full_instruction * inst = emit_data->inst;
654 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
655 const struct tgsi_texture_offset * off = inst->TexOffsets;
656 LLVMTypeRef offset_type = bld_base->int_bld.elem_type;
657
658 /* fetch tex coords */
659 tex_fetch_args(bld_base, emit_data);
660
661 /* fetch tex offsets */
662 if (inst->Texture.NumOffsets) {
663 assert(inst->Texture.NumOffsets == 1);
664
665 emit_data->args[1] = LLVMConstBitCast(
666 bld->immediates[off->Index][off->SwizzleX],
667 offset_type);
668 emit_data->args[2] = LLVMConstBitCast(
669 bld->immediates[off->Index][off->SwizzleY],
670 offset_type);
671 emit_data->args[3] = LLVMConstBitCast(
672 bld->immediates[off->Index][off->SwizzleZ],
673 offset_type);
674 } else {
675 emit_data->args[1] = bld_base->int_bld.zero;
676 emit_data->args[2] = bld_base->int_bld.zero;
677 emit_data->args[3] = bld_base->int_bld.zero;
678 }
679
680 emit_data->arg_count = 4;
681 }
682
683 static void emit_icmp(
684 const struct lp_build_tgsi_action * action,
685 struct lp_build_tgsi_context * bld_base,
686 struct lp_build_emit_data * emit_data)
687 {
688 unsigned pred;
689 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
690 LLVMContextRef context = bld_base->base.gallivm->context;
691
692 switch (emit_data->inst->Instruction.Opcode) {
693 case TGSI_OPCODE_USEQ: pred = LLVMIntEQ; break;
694 case TGSI_OPCODE_USNE: pred = LLVMIntNE; break;
695 case TGSI_OPCODE_USGE: pred = LLVMIntUGE; break;
696 case TGSI_OPCODE_USLT: pred = LLVMIntULT; break;
697 case TGSI_OPCODE_ISGE: pred = LLVMIntSGE; break;
698 case TGSI_OPCODE_ISLT: pred = LLVMIntSLT; break;
699 default:
700 assert(!"unknown instruction");
701 }
702
703 LLVMValueRef v = LLVMBuildICmp(builder, pred,
704 emit_data->args[0], emit_data->args[1],"");
705
706 v = LLVMBuildSExtOrBitCast(builder, v,
707 LLVMInt32TypeInContext(context), "");
708
709 emit_data->output[emit_data->chan] = v;
710 }
711
712 static void emit_cmp(
713 const struct lp_build_tgsi_action *action,
714 struct lp_build_tgsi_context * bld_base,
715 struct lp_build_emit_data * emit_data)
716 {
717 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
718 LLVMRealPredicate pred;
719 LLVMValueRef cond;
720
721 /* XXX I'm not sure whether to do unordered or ordered comparisons,
722 * but llvmpipe uses unordered comparisons, so for consistency we use
723 * unordered. (The authors of llvmpipe aren't sure about using
724 * unordered vs ordered comparisons either.
725 */
726 switch (emit_data->inst->Instruction.Opcode) {
727 case TGSI_OPCODE_SGE: pred = LLVMRealUGE; break;
728 case TGSI_OPCODE_SEQ: pred = LLVMRealUEQ; break;
729 case TGSI_OPCODE_SLE: pred = LLVMRealULE; break;
730 case TGSI_OPCODE_SLT: pred = LLVMRealULT; break;
731 case TGSI_OPCODE_SNE: pred = LLVMRealUNE; break;
732 case TGSI_OPCODE_SGT: pred = LLVMRealUGT; break;
733 default: assert(!"unknown instruction");
734 }
735
736 cond = LLVMBuildFCmp(builder,
737 pred, emit_data->args[0], emit_data->args[1], "");
738
739 emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
740 cond, bld_base->base.one, bld_base->base.zero, "");
741 }
742
743 static void emit_not(
744 const struct lp_build_tgsi_action * action,
745 struct lp_build_tgsi_context * bld_base,
746 struct lp_build_emit_data * emit_data)
747 {
748 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
749 LLVMValueRef v = bitcast(bld_base, TGSI_TYPE_UNSIGNED,
750 emit_data->args[0]);
751 emit_data->output[emit_data->chan] = LLVMBuildNot(builder, v, "");
752 }
753
754 static void emit_and(
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] = LLVMBuildAnd(builder,
761 emit_data->args[0], emit_data->args[1], "");
762 }
763
764 static void emit_or(
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] = LLVMBuildOr(builder,
771 emit_data->args[0], emit_data->args[1], "");
772 }
773
774 static void emit_uadd(
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] = LLVMBuildAdd(builder,
781 emit_data->args[0], emit_data->args[1], "");
782 }
783
784 static void emit_udiv(
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] = LLVMBuildUDiv(builder,
791 emit_data->args[0], emit_data->args[1], "");
792 }
793
794 static void emit_idiv(
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] = LLVMBuildSDiv(builder,
801 emit_data->args[0], emit_data->args[1], "");
802 }
803
804 static void emit_mod(
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] = LLVMBuildSRem(builder,
811 emit_data->args[0], emit_data->args[1], "");
812 }
813
814 static void emit_umod(
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] = LLVMBuildURem(builder,
821 emit_data->args[0], emit_data->args[1], "");
822 }
823
824 static void emit_shl(
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] = LLVMBuildShl(builder,
831 emit_data->args[0], emit_data->args[1], "");
832 }
833
834 static void emit_ushr(
835 const struct lp_build_tgsi_action * action,
836 struct lp_build_tgsi_context * bld_base,
837 struct lp_build_emit_data * emit_data)
838 {
839 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
840 emit_data->output[emit_data->chan] = LLVMBuildLShr(builder,
841 emit_data->args[0], emit_data->args[1], "");
842 }
843 static void emit_ishr(
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] = LLVMBuildAShr(builder,
850 emit_data->args[0], emit_data->args[1], "");
851 }
852
853 static void emit_xor(
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 emit_data->output[emit_data->chan] = LLVMBuildXor(builder,
860 emit_data->args[0], emit_data->args[1], "");
861 }
862
863 static void emit_ssg(
864 const struct lp_build_tgsi_action * action,
865 struct lp_build_tgsi_context * bld_base,
866 struct lp_build_emit_data * emit_data)
867 {
868 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
869
870 LLVMValueRef cmp, val;
871
872 if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
873 cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
874 val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
875 cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
876 val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
877 } else { // float SSG
878 cmp = LLVMBuildFCmp(builder, LLVMRealUGT, emit_data->args[0], bld_base->int_bld.zero, "");
879 val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
880 cmp = LLVMBuildFCmp(builder, LLVMRealUGE, val, bld_base->base.zero, "");
881 val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
882 }
883
884 emit_data->output[emit_data->chan] = val;
885 }
886
887 static void emit_ineg(
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] = LLVMBuildNeg(builder,
894 emit_data->args[0], "");
895 }
896
897 static void emit_f2i(
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] = LLVMBuildFPToSI(builder,
904 emit_data->args[0], bld_base->int_bld.elem_type, "");
905 }
906
907 static void emit_f2u(
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] = LLVMBuildFPToUI(builder,
914 emit_data->args[0], bld_base->uint_bld.elem_type, "");
915 }
916
917 static void emit_i2f(
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] = LLVMBuildSIToFP(builder,
924 emit_data->args[0], bld_base->base.elem_type, "");
925 }
926
927 static void emit_u2f(
928 const struct lp_build_tgsi_action * action,
929 struct lp_build_tgsi_context * bld_base,
930 struct lp_build_emit_data * emit_data)
931 {
932 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
933 emit_data->output[emit_data->chan] = LLVMBuildUIToFP(builder,
934 emit_data->args[0], bld_base->base.elem_type, "");
935 }
936
937 static void emit_immediate(struct lp_build_tgsi_context * bld_base,
938 const struct tgsi_full_immediate *imm)
939 {
940 unsigned i;
941 struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
942
943 for (i = 0; i < 4; ++i) {
944 ctx->soa.immediates[ctx->soa.num_immediates][i] =
945 LLVMConstInt(bld_base->uint_bld.elem_type, imm->u[i].Uint, false );
946 }
947
948 ctx->soa.num_immediates++;
949 }
950
951 LLVMValueRef
952 build_intrinsic(LLVMBuilderRef builder,
953 const char *name,
954 LLVMTypeRef ret_type,
955 LLVMValueRef *args,
956 unsigned num_args,
957 LLVMAttribute attr)
958 {
959 LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
960 LLVMValueRef function;
961
962 function = LLVMGetNamedFunction(module, name);
963 if(!function) {
964 LLVMTypeRef arg_types[LP_MAX_FUNC_ARGS];
965 unsigned i;
966
967 assert(num_args <= LP_MAX_FUNC_ARGS);
968
969 for(i = 0; i < num_args; ++i) {
970 assert(args[i]);
971 arg_types[i] = LLVMTypeOf(args[i]);
972 }
973
974 function = lp_declare_intrinsic(module, name, ret_type, arg_types, num_args);
975
976 if (attr)
977 LLVMAddFunctionAttr(function, attr);
978 }
979
980 return LLVMBuildCall(builder, function, args, num_args, "");
981 }
982
983 void
984 build_tgsi_intrinsic_nomem(
985 const struct lp_build_tgsi_action * action,
986 struct lp_build_tgsi_context * bld_base,
987 struct lp_build_emit_data * emit_data)
988 {
989 struct lp_build_context * base = &bld_base->base;
990 emit_data->output[emit_data->chan] = build_intrinsic(
991 base->gallivm->builder, action->intr_name,
992 emit_data->dst_type, emit_data->args,
993 emit_data->arg_count, LLVMReadNoneAttribute);
994 }
995
996 void radeon_llvm_context_init(struct radeon_llvm_context * ctx)
997 {
998 struct lp_type type;
999 LLVMTypeRef main_fn_type;
1000 LLVMBasicBlockRef main_fn_body;
1001
1002 /* Initialize the gallivm object:
1003 * We are only using the module, context, and builder fields of this struct.
1004 * This should be enough for us to be able to pass our gallivm struct to the
1005 * helper functions in the gallivm module.
1006 */
1007 memset(&ctx->gallivm, 0, sizeof (ctx->gallivm));
1008 memset(&ctx->soa, 0, sizeof(ctx->soa));
1009 ctx->gallivm.context = LLVMContextCreate();
1010 ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
1011 ctx->gallivm.context);
1012 ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
1013
1014 /* Setup the module */
1015 main_fn_type = LLVMFunctionType(LLVMVoidTypeInContext(ctx->gallivm.context),
1016 NULL, 0, 0);
1017 ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
1018 main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
1019 ctx->main_fn, "main_body");
1020 LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
1021
1022 ctx->store_output_intr = "llvm.AMDGPU.store.output.";
1023 ctx->swizzle_intr = "llvm.AMDGPU.swizzle";
1024 struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
1025
1026 /* XXX: We need to revisit this.I think the correct way to do this is
1027 * to use length = 4 here and use the elem_bld for everything. */
1028 type.floating = TRUE;
1029 type.sign = TRUE;
1030 type.width = 32;
1031 type.length = 1;
1032
1033 lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
1034 lp_build_context_init(&ctx->soa.bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
1035 lp_build_context_init(&ctx->soa.bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
1036
1037 bld_base->soa = 1;
1038 bld_base->emit_store = emit_store;
1039 bld_base->emit_swizzle = emit_swizzle;
1040 bld_base->emit_declaration = emit_declaration;
1041 bld_base->emit_immediate = emit_immediate;
1042
1043 bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
1044 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
1045 bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
1046 bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = emit_fetch_output;
1047
1048 /* Allocate outputs */
1049 ctx->soa.outputs = ctx->outputs;
1050
1051 /* XXX: Is there a better way to initialize all this ? */
1052
1053 lp_set_default_actions(bld_base);
1054
1055 bld_base->op_actions[TGSI_OPCODE_IABS].emit = build_tgsi_intrinsic_nomem;
1056 bld_base->op_actions[TGSI_OPCODE_IABS].intr_name = "llvm.AMDIL.abs.";
1057 bld_base->op_actions[TGSI_OPCODE_NOT].emit = emit_not;
1058 bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
1059 bld_base->op_actions[TGSI_OPCODE_XOR].emit = emit_xor;
1060 bld_base->op_actions[TGSI_OPCODE_OR].emit = emit_or;
1061 bld_base->op_actions[TGSI_OPCODE_UADD].emit = emit_uadd;
1062 bld_base->op_actions[TGSI_OPCODE_UDIV].emit = emit_udiv;
1063 bld_base->op_actions[TGSI_OPCODE_IDIV].emit = emit_idiv;
1064 bld_base->op_actions[TGSI_OPCODE_MOD].emit = emit_mod;
1065 bld_base->op_actions[TGSI_OPCODE_UMOD].emit = emit_umod;
1066 bld_base->op_actions[TGSI_OPCODE_INEG].emit = emit_ineg;
1067 bld_base->op_actions[TGSI_OPCODE_SHL].emit = emit_shl;
1068 bld_base->op_actions[TGSI_OPCODE_ISHR].emit = emit_ishr;
1069 bld_base->op_actions[TGSI_OPCODE_USHR].emit = emit_ushr;
1070 bld_base->op_actions[TGSI_OPCODE_SSG].emit = emit_ssg;
1071 bld_base->op_actions[TGSI_OPCODE_ISSG].emit = emit_ssg;
1072 bld_base->op_actions[TGSI_OPCODE_I2F].emit = emit_i2f;
1073 bld_base->op_actions[TGSI_OPCODE_U2F].emit = emit_u2f;
1074 bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
1075 bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
1076 bld_base->op_actions[TGSI_OPCODE_DDX].intr_name = "llvm.AMDGPU.ddx";
1077 bld_base->op_actions[TGSI_OPCODE_DDX].fetch_args = tex_fetch_args;
1078 bld_base->op_actions[TGSI_OPCODE_DDY].intr_name = "llvm.AMDGPU.ddy";
1079 bld_base->op_actions[TGSI_OPCODE_DDY].fetch_args = tex_fetch_args;
1080 bld_base->op_actions[TGSI_OPCODE_USEQ].emit = emit_icmp;
1081 bld_base->op_actions[TGSI_OPCODE_USGE].emit = emit_icmp;
1082 bld_base->op_actions[TGSI_OPCODE_USLT].emit = emit_icmp;
1083 bld_base->op_actions[TGSI_OPCODE_USNE].emit = emit_icmp;
1084 bld_base->op_actions[TGSI_OPCODE_ISGE].emit = emit_icmp;
1085 bld_base->op_actions[TGSI_OPCODE_ISLT].emit = emit_icmp;
1086 bld_base->op_actions[TGSI_OPCODE_ROUND].emit = build_tgsi_intrinsic_nomem;
1087 bld_base->op_actions[TGSI_OPCODE_ROUND].intr_name = "llvm.AMDIL.round.nearest.";
1088 bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
1089 bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.AMDIL.min.";
1090 bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
1091 bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.AMDIL.max.";
1092 bld_base->op_actions[TGSI_OPCODE_IMIN].emit = build_tgsi_intrinsic_nomem;
1093 bld_base->op_actions[TGSI_OPCODE_IMIN].intr_name = "llvm.AMDGPU.imin";
1094 bld_base->op_actions[TGSI_OPCODE_IMAX].emit = build_tgsi_intrinsic_nomem;
1095 bld_base->op_actions[TGSI_OPCODE_IMAX].intr_name = "llvm.AMDGPU.imax";
1096 bld_base->op_actions[TGSI_OPCODE_UMIN].emit = build_tgsi_intrinsic_nomem;
1097 bld_base->op_actions[TGSI_OPCODE_UMIN].intr_name = "llvm.AMDGPU.umin";
1098 bld_base->op_actions[TGSI_OPCODE_UMAX].emit = build_tgsi_intrinsic_nomem;
1099 bld_base->op_actions[TGSI_OPCODE_UMAX].intr_name = "llvm.AMDGPU.umax";
1100 bld_base->op_actions[TGSI_OPCODE_TXF].fetch_args = txf_fetch_args;
1101 bld_base->op_actions[TGSI_OPCODE_TXF].intr_name = "llvm.AMDGPU.txf";
1102 bld_base->op_actions[TGSI_OPCODE_TXQ].fetch_args = tex_fetch_args;
1103 bld_base->op_actions[TGSI_OPCODE_TXQ].intr_name = "llvm.AMDGPU.txq";
1104 bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_nomem;
1105 bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "llvm.AMDIL.round.posinf.";
1106
1107
1108
1109 bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_nomem;
1110 bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "llvm.AMDIL.fabs.";
1111 bld_base->op_actions[TGSI_OPCODE_ARL].emit = build_tgsi_intrinsic_nomem;
1112 bld_base->op_actions[TGSI_OPCODE_ARL].intr_name = "llvm.AMDGPU.arl";
1113 bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
1114 bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
1115 bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
1116 bld_base->op_actions[TGSI_OPCODE_CLAMP].emit = build_tgsi_intrinsic_nomem;
1117 bld_base->op_actions[TGSI_OPCODE_CLAMP].intr_name = "llvm.AMDIL.clamp.";
1118 bld_base->op_actions[TGSI_OPCODE_CMP].emit = build_tgsi_intrinsic_nomem;
1119 bld_base->op_actions[TGSI_OPCODE_CMP].intr_name = "llvm.AMDGPU.cndlt";
1120 bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_nomem;
1121 bld_base->op_actions[TGSI_OPCODE_COS].intr_name = "llvm.AMDGPU.cos";
1122 bld_base->op_actions[TGSI_OPCODE_DIV].emit = build_tgsi_intrinsic_nomem;
1123 bld_base->op_actions[TGSI_OPCODE_DIV].intr_name = "llvm.AMDGPU.div";
1124 bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
1125 bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
1126 bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1127 bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
1128 bld_base->op_actions[TGSI_OPCODE_EX2].intr_name = "llvm.AMDIL.exp.";
1129 bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;
1130 bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "llvm.AMDGPU.floor";
1131 bld_base->op_actions[TGSI_OPCODE_FRC].emit = build_tgsi_intrinsic_nomem;
1132 bld_base->op_actions[TGSI_OPCODE_FRC].intr_name = "llvm.AMDIL.fraction.";
1133 bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
1134 bld_base->op_actions[TGSI_OPCODE_KIL].emit = kil_emit;
1135 bld_base->op_actions[TGSI_OPCODE_KIL].intr_name = "llvm.AMDGPU.kill";
1136 bld_base->op_actions[TGSI_OPCODE_KILP].emit = lp_build_tgsi_intrinsic;
1137 bld_base->op_actions[TGSI_OPCODE_KILP].intr_name = "llvm.AMDGPU.kilp";
1138 bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_nomem;
1139 bld_base->op_actions[TGSI_OPCODE_LG2].intr_name = "llvm.AMDIL.log.";
1140 bld_base->op_actions[TGSI_OPCODE_LRP].emit = build_tgsi_intrinsic_nomem;
1141 bld_base->op_actions[TGSI_OPCODE_LRP].intr_name = "llvm.AMDGPU.lrp";
1142 bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
1143 bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.AMDIL.min.";
1144 bld_base->op_actions[TGSI_OPCODE_MAD].emit = build_tgsi_intrinsic_nomem;
1145 bld_base->op_actions[TGSI_OPCODE_MAD].intr_name = "llvm.AMDIL.mad.";
1146 bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
1147 bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.AMDIL.max.";
1148 bld_base->op_actions[TGSI_OPCODE_MUL].emit = build_tgsi_intrinsic_nomem;
1149 bld_base->op_actions[TGSI_OPCODE_MUL].intr_name = "llvm.AMDGPU.mul";
1150 bld_base->op_actions[TGSI_OPCODE_POW].emit = build_tgsi_intrinsic_nomem;
1151 bld_base->op_actions[TGSI_OPCODE_POW].intr_name = "llvm.AMDGPU.pow";
1152 bld_base->op_actions[TGSI_OPCODE_RCP].emit = build_tgsi_intrinsic_nomem;
1153 bld_base->op_actions[TGSI_OPCODE_RCP].intr_name = "llvm.AMDGPU.rcp";
1154 bld_base->op_actions[TGSI_OPCODE_SSG].emit = build_tgsi_intrinsic_nomem;
1155 bld_base->op_actions[TGSI_OPCODE_SSG].intr_name = "llvm.AMDGPU.ssg";
1156 bld_base->op_actions[TGSI_OPCODE_SGE].emit = emit_cmp;
1157 bld_base->op_actions[TGSI_OPCODE_SEQ].emit = emit_cmp;
1158 bld_base->op_actions[TGSI_OPCODE_SLE].emit = emit_cmp;
1159 bld_base->op_actions[TGSI_OPCODE_SLT].emit = emit_cmp;
1160 bld_base->op_actions[TGSI_OPCODE_SNE].emit = emit_cmp;
1161 bld_base->op_actions[TGSI_OPCODE_SGT].emit = emit_cmp;
1162 bld_base->op_actions[TGSI_OPCODE_SIN].emit = build_tgsi_intrinsic_nomem;
1163 bld_base->op_actions[TGSI_OPCODE_SIN].intr_name = "llvm.AMDGPU.sin";
1164 bld_base->op_actions[TGSI_OPCODE_TEX].fetch_args = tex_fetch_args;
1165 bld_base->op_actions[TGSI_OPCODE_TEX].intr_name = "llvm.AMDGPU.tex";
1166 bld_base->op_actions[TGSI_OPCODE_TXB].fetch_args = tex_fetch_args;
1167 bld_base->op_actions[TGSI_OPCODE_TXB].intr_name = "llvm.AMDGPU.txb";
1168 bld_base->op_actions[TGSI_OPCODE_TXD].fetch_args = txd_fetch_args;
1169 bld_base->op_actions[TGSI_OPCODE_TXD].intr_name = "llvm.AMDGPU.txd";
1170 bld_base->op_actions[TGSI_OPCODE_TXL].fetch_args = tex_fetch_args;
1171 bld_base->op_actions[TGSI_OPCODE_TXL].intr_name = "llvm.AMDGPU.txl";
1172 bld_base->op_actions[TGSI_OPCODE_TXP].fetch_args = txp_fetch_args;
1173 bld_base->op_actions[TGSI_OPCODE_TXP].intr_name = "llvm.AMDGPU.tex";
1174 bld_base->op_actions[TGSI_OPCODE_TRUNC].emit = build_tgsi_intrinsic_nomem;
1175 bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.AMDGPU.trunc";
1176
1177 bld_base->rsq_action.emit = build_tgsi_intrinsic_nomem;
1178 bld_base->rsq_action.intr_name = "llvm.AMDGPU.rsq";
1179 }
1180
1181 void radeon_llvm_finalize_module(struct radeon_llvm_context * ctx)
1182 {
1183 struct gallivm_state * gallivm = ctx->soa.bld_base.base.gallivm;
1184 /* End the main function with Return*/
1185 LLVMBuildRetVoid(gallivm->builder);
1186
1187 /* Create the pass manager */
1188 ctx->gallivm.passmgr = LLVMCreateFunctionPassManagerForModule(
1189 gallivm->module);
1190
1191 /* This pass should eliminate all the load and store instructions */
1192 LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
1193
1194 /* Add some optimization passes */
1195 LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
1196 LLVMAddCFGSimplificationPass(gallivm->passmgr);
1197
1198 /* Run the passs */
1199 LLVMRunFunctionPassManager(gallivm->passmgr, ctx->main_fn);
1200
1201 LLVMDisposeBuilder(gallivm->builder);
1202 LLVMDisposePassManager(gallivm->passmgr);
1203
1204 }
1205
1206 void radeon_llvm_dispose(struct radeon_llvm_context * ctx)
1207 {
1208 LLVMDisposeModule(ctx->soa.bld_base.base.gallivm->module);
1209 LLVMContextDispose(ctx->soa.bld_base.base.gallivm->context);
1210 }