nvc0: try to swap immediates to first source too
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
1 /**************************************************************************
2 *
3 * Copyright 2010 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include "draw_llvm.h"
29
30 #include "draw_context.h"
31 #include "draw_vs.h"
32
33 #include "gallivm/lp_bld_arit.h"
34 #include "gallivm/lp_bld_logic.h"
35 #include "gallivm/lp_bld_const.h"
36 #include "gallivm/lp_bld_swizzle.h"
37 #include "gallivm/lp_bld_struct.h"
38 #include "gallivm/lp_bld_type.h"
39 #include "gallivm/lp_bld_flow.h"
40 #include "gallivm/lp_bld_debug.h"
41 #include "gallivm/lp_bld_tgsi.h"
42 #include "gallivm/lp_bld_printf.h"
43 #include "gallivm/lp_bld_intr.h"
44 #include "gallivm/lp_bld_init.h"
45 #include "gallivm/lp_bld_type.h"
46
47 #include "tgsi/tgsi_exec.h"
48 #include "tgsi/tgsi_dump.h"
49
50 #include "util/u_math.h"
51 #include "util/u_pointer.h"
52 #include "util/u_string.h"
53 #include "util/u_simple_list.h"
54
55
56 #define DEBUG_STORE 0
57
58
59 /**
60 * This function is called by the gallivm "garbage collector" when
61 * the LLVM global data structures are freed. We must free all LLVM-related
62 * data. Specifically, all JIT'd shader variants.
63 */
64 static void
65 draw_llvm_garbage_collect_callback(void *cb_data)
66 {
67 struct draw_llvm *llvm = (struct draw_llvm *) cb_data;
68 struct draw_llvm_variant_list_item *li;
69
70 /* free all shader variants */
71 li = first_elem(&llvm->vs_variants_list);
72 while (!at_end(&llvm->vs_variants_list, li)) {
73 struct draw_llvm_variant_list_item *next = next_elem(li);
74 draw_llvm_destroy_variant(li->base);
75 li = next;
76 }
77
78 /* Null-out these pointers so they get remade next time they're needed.
79 * See the accessor functions below.
80 */
81 llvm->context_ptr_type = NULL;
82 llvm->buffer_ptr_type = NULL;
83 llvm->vb_ptr_type = NULL;
84 llvm->vertex_header_ptr_type = NULL;
85 }
86
87
88 static void
89 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
90
91 static void
92 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *var);
93
94
95 /**
96 * Create LLVM type for struct draw_jit_texture
97 */
98 static LLVMTypeRef
99 create_jit_texture_type(struct gallivm_state *gallivm)
100 {
101 LLVMTargetDataRef target = gallivm->target;
102 LLVMTypeRef texture_type;
103 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
104 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
105
106 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
107 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
108 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
109 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
110 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
111 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
112 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
113 elem_types[DRAW_JIT_TEXTURE_DATA] =
114 LLVMArrayType(LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0),
115 PIPE_MAX_TEXTURE_LEVELS);
116 elem_types[DRAW_JIT_TEXTURE_MIN_LOD] =
117 elem_types[DRAW_JIT_TEXTURE_MAX_LOD] =
118 elem_types[DRAW_JIT_TEXTURE_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
119 elem_types[DRAW_JIT_TEXTURE_BORDER_COLOR] =
120 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
121
122 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
123 Elements(elem_types), 0);
124
125 /* Make sure the target's struct layout cache doesn't return
126 * stale/invalid data.
127 */
128 LLVMInvalidateStructLayout(gallivm->target, texture_type);
129
130 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
131 target, texture_type,
132 DRAW_JIT_TEXTURE_WIDTH);
133 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
134 target, texture_type,
135 DRAW_JIT_TEXTURE_HEIGHT);
136 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
137 target, texture_type,
138 DRAW_JIT_TEXTURE_DEPTH);
139 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
140 target, texture_type,
141 DRAW_JIT_TEXTURE_LAST_LEVEL);
142 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
143 target, texture_type,
144 DRAW_JIT_TEXTURE_ROW_STRIDE);
145 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
146 target, texture_type,
147 DRAW_JIT_TEXTURE_IMG_STRIDE);
148 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data,
149 target, texture_type,
150 DRAW_JIT_TEXTURE_DATA);
151 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, min_lod,
152 target, texture_type,
153 DRAW_JIT_TEXTURE_MIN_LOD);
154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, max_lod,
155 target, texture_type,
156 DRAW_JIT_TEXTURE_MAX_LOD);
157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, lod_bias,
158 target, texture_type,
159 DRAW_JIT_TEXTURE_LOD_BIAS);
160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, border_color,
161 target, texture_type,
162 DRAW_JIT_TEXTURE_BORDER_COLOR);
163
164 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
165
166 return texture_type;
167 }
168
169
170 /**
171 * Create LLVM type for struct draw_jit_texture
172 */
173 static LLVMTypeRef
174 create_jit_context_type(struct gallivm_state *gallivm,
175 LLVMTypeRef texture_type)
176 {
177 LLVMTargetDataRef target = gallivm->target;
178 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
179 LLVMTypeRef elem_types[5];
180 LLVMTypeRef context_type;
181
182 elem_types[0] = LLVMPointerType(float_type, 0); /* vs_constants */
183 elem_types[1] = LLVMPointerType(float_type, 0); /* gs_constants */
184 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4), 12), 0); /* planes */
185 elem_types[3] = LLVMPointerType(float_type, 0); /* viewport */
186 elem_types[4] = LLVMArrayType(texture_type,
187 PIPE_MAX_VERTEX_SAMPLERS); /* textures */
188
189 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
190 Elements(elem_types), 0);
191
192 LLVMInvalidateStructLayout(gallivm->target, context_type);
193
194 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
195 target, context_type, 0);
196 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, gs_constants,
197 target, context_type, 1);
198 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
199 target, context_type, 2);
200 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
201 target, context_type,
202 DRAW_JIT_CTX_TEXTURES);
203 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
204 target, context_type);
205
206 return context_type;
207 }
208
209
210 /**
211 * Create LLVM type for struct pipe_vertex_buffer
212 */
213 static LLVMTypeRef
214 create_jit_vertex_buffer_type(struct gallivm_state *gallivm)
215 {
216 LLVMTargetDataRef target = gallivm->target;
217 LLVMTypeRef elem_types[4];
218 LLVMTypeRef vb_type;
219
220 elem_types[0] =
221 elem_types[1] =
222 elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
223 elem_types[3] = LLVMPointerType(LLVMOpaqueTypeInContext(gallivm->context), 0); /* vs_constants */
224
225 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
226 Elements(elem_types), 0);
227
228 LLVMInvalidateStructLayout(gallivm->target, vb_type);
229
230 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
231 target, vb_type, 0);
232 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
233 target, vb_type, 2);
234
235 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
236
237 return vb_type;
238 }
239
240
241 /**
242 * Create LLVM type for struct vertex_header;
243 */
244 static LLVMTypeRef
245 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
246 {
247 LLVMTargetDataRef target = gallivm->target;
248 LLVMTypeRef elem_types[3];
249 LLVMTypeRef vertex_header;
250 char struct_name[24];
251
252 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
253
254 elem_types[0] = LLVMIntTypeInContext(gallivm->context, 32);
255 elem_types[1] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
256 elem_types[2] = LLVMArrayType(elem_types[1], data_elems);
257
258 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
259 Elements(elem_types), 0);
260
261 LLVMInvalidateStructLayout(gallivm->target, vertex_header);
262
263 /* these are bit-fields and we can't take address of them
264 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
265 target, vertex_header,
266 DRAW_JIT_VERTEX_CLIPMASK);
267 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
268 target, vertex_header,
269 DRAW_JIT_VERTEX_EDGEFLAG);
270 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
271 target, vertex_header,
272 DRAW_JIT_VERTEX_PAD);
273 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
274 target, vertex_header,
275 DRAW_JIT_VERTEX_VERTEX_ID);
276 */
277 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
278 target, vertex_header,
279 DRAW_JIT_VERTEX_CLIP);
280 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
281 target, vertex_header,
282 DRAW_JIT_VERTEX_DATA);
283
284 LLVMAddTypeName(gallivm->module, struct_name, vertex_header);
285
286 return vertex_header;
287 }
288
289
290 /**
291 * Create LLVM types for various structures.
292 */
293 static void
294 create_jit_types(struct draw_llvm *llvm)
295 {
296 struct gallivm_state *gallivm = llvm->gallivm;
297 LLVMTypeRef texture_type, context_type, buffer_type, vb_type;
298
299 texture_type = create_jit_texture_type(gallivm);
300 LLVMAddTypeName(gallivm->module, "texture", texture_type);
301
302 context_type = create_jit_context_type(gallivm, texture_type);
303 LLVMAddTypeName(gallivm->module, "draw_jit_context", context_type);
304 llvm->context_ptr_type = LLVMPointerType(context_type, 0);
305
306 buffer_type = LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
307 LLVMAddTypeName(gallivm->module, "buffer", buffer_type);
308 llvm->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
309
310 vb_type = create_jit_vertex_buffer_type(gallivm);
311 LLVMAddTypeName(gallivm->module, "pipe_vertex_buffer", vb_type);
312 llvm->vb_ptr_type = LLVMPointerType(vb_type, 0);
313 }
314
315
316 static LLVMTypeRef
317 get_context_ptr_type(struct draw_llvm *llvm)
318 {
319 if (!llvm->context_ptr_type)
320 create_jit_types(llvm);
321 return llvm->context_ptr_type;
322 }
323
324
325 static LLVMTypeRef
326 get_buffer_ptr_type(struct draw_llvm *llvm)
327 {
328 if (!llvm->buffer_ptr_type)
329 create_jit_types(llvm);
330 return llvm->buffer_ptr_type;
331 }
332
333
334 static LLVMTypeRef
335 get_vb_ptr_type(struct draw_llvm *llvm)
336 {
337 if (!llvm->vb_ptr_type)
338 create_jit_types(llvm);
339 return llvm->vb_ptr_type;
340 }
341
342 static LLVMTypeRef
343 get_vertex_header_ptr_type(struct draw_llvm *llvm)
344 {
345 if (!llvm->vertex_header_ptr_type)
346 create_jit_types(llvm);
347 return llvm->vertex_header_ptr_type;
348 }
349
350
351 /**
352 * Create per-context LLVM info.
353 */
354 struct draw_llvm *
355 draw_llvm_create(struct draw_context *draw, struct gallivm_state *gallivm)
356 {
357 struct draw_llvm *llvm;
358
359 llvm = CALLOC_STRUCT( draw_llvm );
360 if (!llvm)
361 return NULL;
362
363 lp_build_init();
364
365 llvm->draw = draw;
366 llvm->gallivm = gallivm;
367
368 if (gallivm_debug & GALLIVM_DEBUG_IR) {
369 LLVMDumpModule(llvm->gallivm->module);
370 }
371
372 llvm->nr_variants = 0;
373 make_empty_list(&llvm->vs_variants_list);
374
375 gallivm_register_garbage_collector_callback(
376 draw_llvm_garbage_collect_callback, llvm);
377
378 return llvm;
379 }
380
381
382 /**
383 * Free per-context LLVM info.
384 */
385 void
386 draw_llvm_destroy(struct draw_llvm *llvm)
387 {
388 gallivm_remove_garbage_collector_callback(
389 draw_llvm_garbage_collect_callback, llvm);
390
391 /* XXX free other draw_llvm data? */
392 FREE(llvm);
393 }
394
395
396 /**
397 * Create LLVM-generated code for a vertex shader.
398 */
399 struct draw_llvm_variant *
400 draw_llvm_create_variant(struct draw_llvm *llvm,
401 unsigned num_inputs,
402 const struct draw_llvm_variant_key *key)
403 {
404 struct draw_llvm_variant *variant;
405 struct llvm_vertex_shader *shader =
406 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
407 LLVMTypeRef vertex_header;
408
409 variant = MALLOC(sizeof *variant +
410 shader->variant_key_size -
411 sizeof variant->key);
412 if (variant == NULL)
413 return NULL;
414
415 variant->llvm = llvm;
416
417 memcpy(&variant->key, key, shader->variant_key_size);
418
419 vertex_header = create_jit_vertex_header(llvm->gallivm, num_inputs);
420
421 llvm->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
422
423 draw_llvm_generate(llvm, variant);
424 draw_llvm_generate_elts(llvm, variant);
425
426 variant->shader = shader;
427 variant->list_item_global.base = variant;
428 variant->list_item_local.base = variant;
429 /*variant->no = */shader->variants_created++;
430 variant->list_item_global.base = variant;
431
432 return variant;
433 }
434
435 static void
436 generate_vs(struct draw_llvm *llvm,
437 LLVMBuilderRef builder,
438 LLVMValueRef (*outputs)[NUM_CHANNELS],
439 const LLVMValueRef (*inputs)[NUM_CHANNELS],
440 LLVMValueRef context_ptr,
441 struct lp_build_sampler_soa *draw_sampler)
442 {
443 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
444 struct lp_type vs_type;
445 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(llvm->gallivm, context_ptr);
446 struct lp_build_sampler_soa *sampler = 0;
447
448 memset(&vs_type, 0, sizeof vs_type);
449 vs_type.floating = TRUE; /* floating point values */
450 vs_type.sign = TRUE; /* values are signed */
451 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
452 vs_type.width = 32; /* 32-bit float */
453 vs_type.length = 4; /* 4 elements per vector */
454 #if 0
455 num_vs = 4; /* number of vertices per block */
456 #endif
457
458 if (gallivm_debug & GALLIVM_DEBUG_IR) {
459 tgsi_dump(tokens, 0);
460 }
461
462 if (llvm->draw->num_sampler_views &&
463 llvm->draw->num_samplers)
464 sampler = draw_sampler;
465
466 lp_build_tgsi_soa(llvm->gallivm,
467 tokens,
468 vs_type,
469 NULL /*struct lp_build_mask_context *mask*/,
470 consts_ptr,
471 NULL /*pos*/,
472 inputs,
473 outputs,
474 sampler,
475 &llvm->draw->vs.vertex_shader->info);
476 }
477
478 #if DEBUG_STORE
479 static void print_vectorf(LLVMBuilderRef builder,
480 LLVMValueRef vec)
481 {
482 LLVMValueRef val[4];
483 val[0] = LLVMBuildExtractElement(builder, vec,
484 lp_build_const_int32(gallivm, 0), "");
485 val[1] = LLVMBuildExtractElement(builder, vec,
486 lp_build_const_int32(gallivm, 1), "");
487 val[2] = LLVMBuildExtractElement(builder, vec,
488 lp_build_const_int32(gallivm, 2), "");
489 val[3] = LLVMBuildExtractElement(builder, vec,
490 lp_build_const_int32(gallivm, 3), "");
491 lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n",
492 val[0], val[1], val[2], val[3]);
493 }
494 #endif
495
496 static void
497 generate_fetch(struct gallivm_state *gallivm,
498 LLVMValueRef vbuffers_ptr,
499 LLVMValueRef *res,
500 struct pipe_vertex_element *velem,
501 LLVMValueRef vbuf,
502 LLVMValueRef index,
503 LLVMValueRef instance_id)
504 {
505 LLVMBuilderRef builder = gallivm->builder;
506 LLVMValueRef indices =
507 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
508 velem->vertex_buffer_index, 0);
509 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
510 &indices, 1, "");
511 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
512 LLVMValueRef vb_max_index = draw_jit_vbuffer_max_index(gallivm, vbuf);
513 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
514 LLVMValueRef cond;
515 LLVMValueRef stride;
516
517 if (velem->instance_divisor) {
518 /* array index = instance_id / instance_divisor */
519 index = LLVMBuildUDiv(builder, instance_id,
520 lp_build_const_int32(gallivm, velem->instance_divisor),
521 "instance_divisor");
522 }
523
524 /* limit index to min(index, vb_max_index) */
525 cond = LLVMBuildICmp(builder, LLVMIntULE, index, vb_max_index, "");
526 index = LLVMBuildSelect(builder, cond, index, vb_max_index, "");
527
528 stride = LLVMBuildMul(builder, vb_stride, index, "");
529
530 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
531
532 stride = LLVMBuildAdd(builder, stride,
533 vb_buffer_offset,
534 "");
535 stride = LLVMBuildAdd(builder, stride,
536 lp_build_const_int32(gallivm, velem->src_offset),
537 "");
538
539 /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/
540 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
541
542 *res = draw_llvm_translate_from(gallivm, vbuffer_ptr, velem->src_format);
543 }
544
545 static LLVMValueRef
546 aos_to_soa(struct gallivm_state *gallivm,
547 LLVMValueRef val0,
548 LLVMValueRef val1,
549 LLVMValueRef val2,
550 LLVMValueRef val3,
551 LLVMValueRef channel)
552 {
553 LLVMBuilderRef builder = gallivm->builder;
554 LLVMValueRef ex, res;
555
556 ex = LLVMBuildExtractElement(builder, val0,
557 channel, "");
558 res = LLVMBuildInsertElement(builder,
559 LLVMConstNull(LLVMTypeOf(val0)),
560 ex,
561 lp_build_const_int32(gallivm, 0),
562 "");
563
564 ex = LLVMBuildExtractElement(builder, val1,
565 channel, "");
566 res = LLVMBuildInsertElement(builder,
567 res, ex,
568 lp_build_const_int32(gallivm, 1),
569 "");
570
571 ex = LLVMBuildExtractElement(builder, val2,
572 channel, "");
573 res = LLVMBuildInsertElement(builder,
574 res, ex,
575 lp_build_const_int32(gallivm, 2),
576 "");
577
578 ex = LLVMBuildExtractElement(builder, val3,
579 channel, "");
580 res = LLVMBuildInsertElement(builder,
581 res, ex,
582 lp_build_const_int32(gallivm, 3),
583 "");
584
585 return res;
586 }
587
588 static void
589 soa_to_aos(struct gallivm_state *gallivm,
590 LLVMValueRef soa[NUM_CHANNELS],
591 LLVMValueRef aos[NUM_CHANNELS])
592 {
593 LLVMBuilderRef builder = gallivm->builder;
594 LLVMValueRef comp;
595 int i = 0;
596
597 debug_assert(NUM_CHANNELS == 4);
598
599 aos[0] = LLVMConstNull(LLVMTypeOf(soa[0]));
600 aos[1] = aos[2] = aos[3] = aos[0];
601
602 for (i = 0; i < NUM_CHANNELS; ++i) {
603 LLVMValueRef channel = lp_build_const_int32(gallivm, i);
604
605 comp = LLVMBuildExtractElement(builder, soa[i],
606 lp_build_const_int32(gallivm, 0), "");
607 aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, "");
608
609 comp = LLVMBuildExtractElement(builder, soa[i],
610 lp_build_const_int32(gallivm, 1), "");
611 aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, "");
612
613 comp = LLVMBuildExtractElement(builder, soa[i],
614 lp_build_const_int32(gallivm, 2), "");
615 aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, "");
616
617 comp = LLVMBuildExtractElement(builder, soa[i],
618 lp_build_const_int32(gallivm, 3), "");
619 aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, "");
620
621 }
622 }
623
624 static void
625 convert_to_soa(struct gallivm_state *gallivm,
626 LLVMValueRef (*aos)[NUM_CHANNELS],
627 LLVMValueRef (*soa)[NUM_CHANNELS],
628 int num_attribs)
629 {
630 int i;
631
632 debug_assert(NUM_CHANNELS == 4);
633
634 for (i = 0; i < num_attribs; ++i) {
635 LLVMValueRef val0 = aos[i][0];
636 LLVMValueRef val1 = aos[i][1];
637 LLVMValueRef val2 = aos[i][2];
638 LLVMValueRef val3 = aos[i][3];
639
640 soa[i][0] = aos_to_soa(gallivm, val0, val1, val2, val3,
641 lp_build_const_int32(gallivm, 0));
642 soa[i][1] = aos_to_soa(gallivm, val0, val1, val2, val3,
643 lp_build_const_int32(gallivm, 1));
644 soa[i][2] = aos_to_soa(gallivm, val0, val1, val2, val3,
645 lp_build_const_int32(gallivm, 2));
646 soa[i][3] = aos_to_soa(gallivm, val0, val1, val2, val3,
647 lp_build_const_int32(gallivm, 3));
648 }
649 }
650
651 static void
652 store_aos(struct gallivm_state *gallivm,
653 LLVMValueRef io_ptr,
654 LLVMValueRef index,
655 LLVMValueRef value,
656 LLVMValueRef clipmask)
657 {
658 LLVMBuilderRef builder = gallivm->builder;
659 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptr);
660 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
661 LLVMValueRef indices[3];
662 LLVMValueRef val, shift;
663
664 indices[0] = lp_build_const_int32(gallivm, 0);
665 indices[1] = index;
666 indices[2] = lp_build_const_int32(gallivm, 0);
667
668 /* initialize vertex id:16 = 0xffff, pad:3 = 0, edgeflag:1 = 1 */
669 val = lp_build_const_int32(gallivm, 0xffff1);
670 shift = lp_build_const_int32(gallivm, 12);
671 val = LLVMBuildShl(builder, val, shift, "");
672 /* add clipmask:12 */
673 val = LLVMBuildOr(builder, val, clipmask, "");
674
675 /* store vertex header */
676 LLVMBuildStore(builder, val, id_ptr);
677
678
679 #if DEBUG_STORE
680 lp_build_printf(builder, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
681 #endif
682 #if 0
683 /*lp_build_printf(builder, " ---- %p storing at %d (%p) ", io_ptr, index, data_ptr);
684 print_vectorf(builder, value);*/
685 data_ptr = LLVMBuildBitCast(builder, data_ptr,
686 LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0), 0),
687 "datavec");
688 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, "");
689
690 LLVMBuildStore(builder, value, data_ptr);
691 #else
692 {
693 LLVMValueRef x, y, z, w;
694 LLVMValueRef idx0, idx1, idx2, idx3;
695 LLVMValueRef gep0, gep1, gep2, gep3;
696 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
697
698 idx0 = lp_build_const_int32(gallivm, 0);
699 idx1 = lp_build_const_int32(gallivm, 1);
700 idx2 = lp_build_const_int32(gallivm, 2);
701 idx3 = lp_build_const_int32(gallivm, 3);
702
703 x = LLVMBuildExtractElement(builder, value,
704 idx0, "");
705 y = LLVMBuildExtractElement(builder, value,
706 idx1, "");
707 z = LLVMBuildExtractElement(builder, value,
708 idx2, "");
709 w = LLVMBuildExtractElement(builder, value,
710 idx3, "");
711
712 gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, "");
713 gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, "");
714 gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, "");
715 gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, "");
716
717 /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n",
718 x, gep0, y, gep1, z, gep2, w, gep3);*/
719 LLVMBuildStore(builder, x, gep0);
720 LLVMBuildStore(builder, y, gep1);
721 LLVMBuildStore(builder, z, gep2);
722 LLVMBuildStore(builder, w, gep3);
723 }
724 #endif
725 }
726
727 static void
728 store_aos_array(struct gallivm_state *gallivm,
729 LLVMValueRef io_ptr,
730 LLVMValueRef aos[NUM_CHANNELS],
731 int attrib,
732 int num_outputs,
733 LLVMValueRef clipmask)
734 {
735 LLVMBuilderRef builder = gallivm->builder;
736 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
737 LLVMValueRef ind0 = lp_build_const_int32(gallivm, 0);
738 LLVMValueRef ind1 = lp_build_const_int32(gallivm, 1);
739 LLVMValueRef ind2 = lp_build_const_int32(gallivm, 2);
740 LLVMValueRef ind3 = lp_build_const_int32(gallivm, 3);
741 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
742 LLVMValueRef clipmask0, clipmask1, clipmask2, clipmask3;
743
744 debug_assert(NUM_CHANNELS == 4);
745
746 io0_ptr = LLVMBuildGEP(builder, io_ptr,
747 &ind0, 1, "");
748 io1_ptr = LLVMBuildGEP(builder, io_ptr,
749 &ind1, 1, "");
750 io2_ptr = LLVMBuildGEP(builder, io_ptr,
751 &ind2, 1, "");
752 io3_ptr = LLVMBuildGEP(builder, io_ptr,
753 &ind3, 1, "");
754
755 clipmask0 = LLVMBuildExtractElement(builder, clipmask,
756 ind0, "");
757 clipmask1 = LLVMBuildExtractElement(builder, clipmask,
758 ind1, "");
759 clipmask2 = LLVMBuildExtractElement(builder, clipmask,
760 ind2, "");
761 clipmask3 = LLVMBuildExtractElement(builder, clipmask,
762 ind3, "");
763
764 #if DEBUG_STORE
765 lp_build_printf(builder, "io = %p, indexes[%d, %d, %d, %d]\n, clipmask0 = %x, clipmask1 = %x, clipmask2 = %x, clipmask3 = %x\n",
766 io_ptr, ind0, ind1, ind2, ind3, clipmask0, clipmask1, clipmask2, clipmask3);
767 #endif
768 /* store for each of the 4 vertices */
769 store_aos(gallivm, io0_ptr, attr_index, aos[0], clipmask0);
770 store_aos(gallivm, io1_ptr, attr_index, aos[1], clipmask1);
771 store_aos(gallivm, io2_ptr, attr_index, aos[2], clipmask2);
772 store_aos(gallivm, io3_ptr, attr_index, aos[3], clipmask3);
773 }
774
775 static void
776 convert_to_aos(struct gallivm_state *gallivm,
777 LLVMValueRef io,
778 LLVMValueRef (*outputs)[NUM_CHANNELS],
779 LLVMValueRef clipmask,
780 int num_outputs,
781 int max_vertices)
782 {
783 LLVMBuilderRef builder = gallivm->builder;
784 unsigned chan, attrib;
785
786 #if DEBUG_STORE
787 lp_build_printf(builder, " # storing begin\n");
788 #endif
789 for (attrib = 0; attrib < num_outputs; ++attrib) {
790 LLVMValueRef soa[4];
791 LLVMValueRef aos[4];
792 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
793 if(outputs[attrib][chan]) {
794 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
795 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
796 /*lp_build_printf(builder, "output %d : %d ",
797 LLVMConstInt(LLVMInt32Type(), attrib, 0),
798 LLVMConstInt(LLVMInt32Type(), chan, 0));
799 print_vectorf(builder, out);*/
800 soa[chan] = out;
801 } else
802 soa[chan] = 0;
803 }
804 soa_to_aos(gallivm, soa, aos);
805 store_aos_array(gallivm,
806 io,
807 aos,
808 attrib,
809 num_outputs,
810 clipmask);
811 }
812 #if DEBUG_STORE
813 lp_build_printf(builder, " # storing end\n");
814 #endif
815 }
816
817 /*
818 * Stores original vertex positions in clip coordinates
819 * There is probably a more efficient way to do this, 4 floats at once
820 * rather than extracting each element one by one.
821 */
822 static void
823 store_clip(struct gallivm_state *gallivm,
824 LLVMValueRef io_ptr,
825 LLVMValueRef (*outputs)[NUM_CHANNELS])
826 {
827 LLVMBuilderRef builder = gallivm->builder;
828 LLVMValueRef out[4];
829 LLVMValueRef indices[2];
830 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
831 LLVMValueRef clip_ptr0, clip_ptr1, clip_ptr2, clip_ptr3;
832 LLVMValueRef clip0_ptr, clip1_ptr, clip2_ptr, clip3_ptr;
833 LLVMValueRef out0elem, out1elem, out2elem, out3elem;
834 int i;
835
836 LLVMValueRef ind0 = lp_build_const_int32(gallivm, 0);
837 LLVMValueRef ind1 = lp_build_const_int32(gallivm, 1);
838 LLVMValueRef ind2 = lp_build_const_int32(gallivm, 2);
839 LLVMValueRef ind3 = lp_build_const_int32(gallivm, 3);
840
841 indices[0] =
842 indices[1] = lp_build_const_int32(gallivm, 0);
843
844 out[0] = LLVMBuildLoad(builder, outputs[0][0], ""); /*x0 x1 x2 x3*/
845 out[1] = LLVMBuildLoad(builder, outputs[0][1], ""); /*y0 y1 y2 y3*/
846 out[2] = LLVMBuildLoad(builder, outputs[0][2], ""); /*z0 z1 z2 z3*/
847 out[3] = LLVMBuildLoad(builder, outputs[0][3], ""); /*w0 w1 w2 w3*/
848
849 io0_ptr = LLVMBuildGEP(builder, io_ptr, &ind0, 1, "");
850 io1_ptr = LLVMBuildGEP(builder, io_ptr, &ind1, 1, "");
851 io2_ptr = LLVMBuildGEP(builder, io_ptr, &ind2, 1, "");
852 io3_ptr = LLVMBuildGEP(builder, io_ptr, &ind3, 1, "");
853
854 clip_ptr0 = draw_jit_header_clip(gallivm, io0_ptr);
855 clip_ptr1 = draw_jit_header_clip(gallivm, io1_ptr);
856 clip_ptr2 = draw_jit_header_clip(gallivm, io2_ptr);
857 clip_ptr3 = draw_jit_header_clip(gallivm, io3_ptr);
858
859 for (i = 0; i<4; i++){
860 clip0_ptr = LLVMBuildGEP(builder, clip_ptr0, indices, 2, ""); /* x0 */
861 clip1_ptr = LLVMBuildGEP(builder, clip_ptr1, indices, 2, ""); /* x1 */
862 clip2_ptr = LLVMBuildGEP(builder, clip_ptr2, indices, 2, ""); /* x2 */
863 clip3_ptr = LLVMBuildGEP(builder, clip_ptr3, indices, 2, ""); /* x3 */
864
865 out0elem = LLVMBuildExtractElement(builder, out[i], ind0, ""); /* x0 */
866 out1elem = LLVMBuildExtractElement(builder, out[i], ind1, ""); /* x1 */
867 out2elem = LLVMBuildExtractElement(builder, out[i], ind2, ""); /* x2 */
868 out3elem = LLVMBuildExtractElement(builder, out[i], ind3, ""); /* x3 */
869
870 LLVMBuildStore(builder, out0elem, clip0_ptr);
871 LLVMBuildStore(builder, out1elem, clip1_ptr);
872 LLVMBuildStore(builder, out2elem, clip2_ptr);
873 LLVMBuildStore(builder, out3elem, clip3_ptr);
874
875 indices[1]= LLVMBuildAdd(builder, indices[1], ind1, "");
876 }
877
878 }
879
880 /* Equivalent of _mm_set1_ps(a)
881 */
882 static LLVMValueRef
883 vec4f_from_scalar(struct gallivm_state *gallivm,
884 LLVMValueRef a,
885 const char *name)
886 {
887 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
888 LLVMValueRef res = LLVMGetUndef(LLVMVectorType(float_type, 4));
889 int i;
890
891 for(i = 0; i < 4; ++i) {
892 LLVMValueRef index = lp_build_const_int32(gallivm, i);
893 res = LLVMBuildInsertElement(gallivm->builder, res, a,
894 index, i == 3 ? name : "");
895 }
896
897 return res;
898 }
899
900 /*
901 * Transforms the outputs for viewport mapping
902 */
903 static void
904 generate_viewport(struct draw_llvm *llvm,
905 LLVMBuilderRef builder,
906 LLVMValueRef (*outputs)[NUM_CHANNELS],
907 LLVMValueRef context_ptr)
908 {
909 int i;
910 struct gallivm_state *gallivm = llvm->gallivm;
911 struct lp_type f32_type = lp_type_float_vec(32);
912 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[0][3], ""); /*w0 w1 w2 w3*/
913 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
914 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
915
916 /* for 1/w convention*/
917 out3 = LLVMBuildFDiv(builder, const1, out3, "");
918 LLVMBuildStore(builder, out3, outputs[0][3]);
919
920 /* Viewport Mapping */
921 for (i=0; i<3; i++){
922 LLVMValueRef out = LLVMBuildLoad(builder, outputs[0][i], ""); /*x0 x1 x2 x3*/
923 LLVMValueRef scale;
924 LLVMValueRef trans;
925 LLVMValueRef scale_i;
926 LLVMValueRef trans_i;
927 LLVMValueRef index;
928
929 index = lp_build_const_int32(gallivm, i);
930 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
931
932 index = lp_build_const_int32(gallivm, i+4);
933 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
934
935 scale = vec4f_from_scalar(gallivm, LLVMBuildLoad(builder, scale_i, ""), "scale");
936 trans = vec4f_from_scalar(gallivm, LLVMBuildLoad(builder, trans_i, ""), "trans");
937
938 /* divide by w */
939 out = LLVMBuildFMul(builder, out, out3, "");
940 /* mult by scale */
941 out = LLVMBuildFMul(builder, out, scale, "");
942 /* add translation */
943 out = LLVMBuildFAdd(builder, out, trans, "");
944
945 /* store transformed outputs */
946 LLVMBuildStore(builder, out, outputs[0][i]);
947 }
948
949 }
950
951
952 /*
953 * Returns clipmask as 4xi32 bitmask for the 4 vertices
954 */
955 static LLVMValueRef
956 generate_clipmask(struct gallivm_state *gallivm,
957 LLVMValueRef (*outputs)[NUM_CHANNELS],
958 boolean clip_xy,
959 boolean clip_z,
960 boolean clip_user,
961 boolean clip_halfz,
962 unsigned nr,
963 LLVMValueRef context_ptr)
964 {
965 LLVMBuilderRef builder = gallivm->builder;
966 LLVMValueRef mask; /* stores the <4xi32> clipmasks */
967 LLVMValueRef test, temp;
968 LLVMValueRef zero, shift;
969 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
970 LLVMValueRef plane1, planes, plane_ptr, sum;
971
972 unsigned i;
973
974 struct lp_type f32_type = lp_type_float_vec(32);
975
976 mask = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 0);
977 temp = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 0);
978 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
979 shift = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 1); /* 1 1 1 1 */
980
981 /* Assuming position stored at output[0] */
982 pos_x = LLVMBuildLoad(builder, outputs[0][0], ""); /*x0 x1 x2 x3*/
983 pos_y = LLVMBuildLoad(builder, outputs[0][1], ""); /*y0 y1 y2 y3*/
984 pos_z = LLVMBuildLoad(builder, outputs[0][2], ""); /*z0 z1 z2 z3*/
985 pos_w = LLVMBuildLoad(builder, outputs[0][3], ""); /*w0 w1 w2 w3*/
986
987 /* Cliptest, for hardwired planes */
988 if (clip_xy){
989 /* plane 1 */
990 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
991 temp = shift;
992 test = LLVMBuildAnd(builder, test, temp, "");
993 mask = test;
994
995 /* plane 2 */
996 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
997 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
998 temp = LLVMBuildShl(builder, temp, shift, "");
999 test = LLVMBuildAnd(builder, test, temp, "");
1000 mask = LLVMBuildOr(builder, mask, test, "");
1001
1002 /* plane 3 */
1003 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1004 temp = LLVMBuildShl(builder, temp, shift, "");
1005 test = LLVMBuildAnd(builder, test, temp, "");
1006 mask = LLVMBuildOr(builder, mask, test, "");
1007
1008 /* plane 4 */
1009 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1010 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1011 temp = LLVMBuildShl(builder, temp, shift, "");
1012 test = LLVMBuildAnd(builder, test, temp, "");
1013 mask = LLVMBuildOr(builder, mask, test, "");
1014 }
1015
1016 if (clip_z){
1017 temp = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 16);
1018 if (clip_halfz){
1019 /* plane 5 */
1020 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1021 test = LLVMBuildAnd(builder, test, temp, "");
1022 mask = LLVMBuildOr(builder, mask, test, "");
1023 }
1024 else{
1025 /* plane 5 */
1026 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1027 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1028 test = LLVMBuildAnd(builder, test, temp, "");
1029 mask = LLVMBuildOr(builder, mask, test, "");
1030 }
1031 /* plane 6 */
1032 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1033 temp = LLVMBuildShl(builder, temp, shift, "");
1034 test = LLVMBuildAnd(builder, test, temp, "");
1035 mask = LLVMBuildOr(builder, mask, test, "");
1036 }
1037
1038 if (clip_user){
1039 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1040 LLVMValueRef indices[3];
1041 temp = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 32);
1042
1043 /* userclip planes */
1044 for (i = 6; i < nr; i++) {
1045 indices[0] = lp_build_const_int32(gallivm, 0);
1046 indices[1] = lp_build_const_int32(gallivm, i);
1047
1048 indices[2] = lp_build_const_int32(gallivm, 0);
1049 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1050 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1051 planes = vec4f_from_scalar(gallivm, plane1, "plane4_x");
1052 sum = LLVMBuildFMul(builder, planes, pos_x, "");
1053
1054 indices[2] = lp_build_const_int32(gallivm, 1);
1055 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1056 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1057 planes = vec4f_from_scalar(gallivm, plane1, "plane4_y");
1058 test = LLVMBuildFMul(builder, planes, pos_y, "");
1059 sum = LLVMBuildFAdd(builder, sum, test, "");
1060
1061 indices[2] = lp_build_const_int32(gallivm, 2);
1062 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1063 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1064 planes = vec4f_from_scalar(gallivm, plane1, "plane4_z");
1065 test = LLVMBuildFMul(builder, planes, pos_z, "");
1066 sum = LLVMBuildFAdd(builder, sum, test, "");
1067
1068 indices[2] = lp_build_const_int32(gallivm, 3);
1069 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1070 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1071 planes = vec4f_from_scalar(gallivm, plane1, "plane4_w");
1072 test = LLVMBuildFMul(builder, planes, pos_w, "");
1073 sum = LLVMBuildFAdd(builder, sum, test, "");
1074
1075 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1076 temp = LLVMBuildShl(builder, temp, shift, "");
1077 test = LLVMBuildAnd(builder, test, temp, "");
1078 mask = LLVMBuildOr(builder, mask, test, "");
1079 }
1080 }
1081 return mask;
1082 }
1083
1084 /*
1085 * Returns boolean if any clipping has occurred
1086 * Used zero/non-zero i32 value to represent boolean
1087 */
1088 static void
1089 clipmask_bool(struct gallivm_state *gallivm,
1090 LLVMValueRef clipmask,
1091 LLVMValueRef ret_ptr)
1092 {
1093 LLVMBuilderRef builder = gallivm->builder;
1094 LLVMValueRef ret = LLVMBuildLoad(builder, ret_ptr, "");
1095 LLVMValueRef temp;
1096 int i;
1097
1098 for (i=0; i<4; i++){
1099 temp = LLVMBuildExtractElement(builder, clipmask,
1100 lp_build_const_int32(gallivm, i) , "");
1101 ret = LLVMBuildOr(builder, ret, temp, "");
1102 }
1103
1104 LLVMBuildStore(builder, ret, ret_ptr);
1105 }
1106
1107 static void
1108 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1109 {
1110 struct gallivm_state *gallivm = llvm->gallivm;
1111 LLVMContextRef context = gallivm->context;
1112 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1113 LLVMTypeRef arg_types[8];
1114 LLVMTypeRef func_type;
1115 LLVMValueRef context_ptr;
1116 LLVMBasicBlockRef block;
1117 LLVMBuilderRef builder;
1118 LLVMValueRef start, end, count, stride, step, io_itr;
1119 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1120 LLVMValueRef instance_id;
1121 struct draw_context *draw = llvm->draw;
1122 unsigned i, j;
1123 struct lp_build_context bld;
1124 struct lp_build_loop_state lp_loop;
1125 const int max_vertices = 4;
1126 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
1127 void *code;
1128 struct lp_build_sampler_soa *sampler = 0;
1129 LLVMValueRef ret, ret_ptr;
1130 boolean bypass_viewport = variant->key.bypass_viewport;
1131 boolean enable_cliptest = variant->key.clip_xy ||
1132 variant->key.clip_z ||
1133 variant->key.clip_user;
1134
1135 arg_types[0] = get_context_ptr_type(llvm); /* context */
1136 arg_types[1] = get_vertex_header_ptr_type(llvm); /* vertex_header */
1137 arg_types[2] = get_buffer_ptr_type(llvm); /* vbuffers */
1138 arg_types[3] = int32_type; /* start */
1139 arg_types[4] = int32_type; /* count */
1140 arg_types[5] = int32_type; /* stride */
1141 arg_types[6] = get_vb_ptr_type(llvm); /* pipe_vertex_buffer's */
1142 arg_types[7] = int32_type; /* instance_id */
1143
1144 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1145
1146 variant->function = LLVMAddFunction(gallivm->module, "draw_llvm_shader",
1147 func_type);
1148 LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
1149 for(i = 0; i < Elements(arg_types); ++i)
1150 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1151 LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
1152
1153 context_ptr = LLVMGetParam(variant->function, 0);
1154 io_ptr = LLVMGetParam(variant->function, 1);
1155 vbuffers_ptr = LLVMGetParam(variant->function, 2);
1156 start = LLVMGetParam(variant->function, 3);
1157 count = LLVMGetParam(variant->function, 4);
1158 stride = LLVMGetParam(variant->function, 5);
1159 vb_ptr = LLVMGetParam(variant->function, 6);
1160 instance_id = LLVMGetParam(variant->function, 7);
1161
1162 lp_build_name(context_ptr, "context");
1163 lp_build_name(io_ptr, "io");
1164 lp_build_name(vbuffers_ptr, "vbuffers");
1165 lp_build_name(start, "start");
1166 lp_build_name(count, "count");
1167 lp_build_name(stride, "stride");
1168 lp_build_name(vb_ptr, "vb");
1169 lp_build_name(instance_id, "instance_id");
1170
1171 /*
1172 * Function body
1173 */
1174
1175 block = LLVMAppendBasicBlockInContext(gallivm->context, variant->function, "entry");
1176 builder = gallivm->builder;
1177 assert(builder);
1178 LLVMPositionBuilderAtEnd(builder, block);
1179
1180 lp_build_context_init(&bld, llvm->gallivm, lp_type_int(32));
1181
1182 end = lp_build_add(&bld, start, count);
1183
1184 step = lp_build_const_int32(gallivm, max_vertices);
1185
1186 /* function will return non-zero i32 value if any clipped vertices */
1187 ret_ptr = lp_build_alloca(gallivm, int32_type, "");
1188 LLVMBuildStore(builder, lp_build_const_int32(gallivm, 0), ret_ptr);
1189
1190 /* code generated texture sampling */
1191 sampler = draw_llvm_sampler_soa_create(
1192 draw_llvm_variant_key_samplers(&variant->key),
1193 context_ptr);
1194
1195 #if DEBUG_STORE
1196 lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
1197 start, end, step);
1198 #endif
1199 lp_build_loop_begin(&lp_loop, llvm->gallivm, start);
1200 {
1201 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
1202 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
1203 LLVMValueRef io;
1204 LLVMValueRef clipmask; /* holds the clipmask value */
1205 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
1206
1207 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
1208 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1209 #if DEBUG_STORE
1210 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
1211 io_itr, io, lp_loop.counter);
1212 #endif
1213 for (i = 0; i < NUM_CHANNELS; ++i) {
1214 LLVMValueRef true_index = LLVMBuildAdd(
1215 builder,
1216 lp_loop.counter,
1217 lp_build_const_int32(gallivm, i), "");
1218 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1219 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1220 LLVMValueRef vb_index = lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1221 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
1222 &vb_index, 1, "");
1223 generate_fetch(llvm->gallivm, vbuffers_ptr,
1224 &aos_attribs[j][i], velem, vb, true_index,
1225 instance_id);
1226 }
1227 }
1228 convert_to_soa(gallivm, aos_attribs, inputs,
1229 draw->pt.nr_vertex_elements);
1230
1231 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
1232 generate_vs(llvm,
1233 builder,
1234 outputs,
1235 ptr_aos,
1236 context_ptr,
1237 sampler);
1238
1239 /* store original positions in clip before further manipulation */
1240 store_clip(gallivm, io, outputs);
1241
1242 /* do cliptest */
1243 if (enable_cliptest){
1244 /* allocate clipmask, assign it integer type */
1245 clipmask = generate_clipmask(gallivm, outputs,
1246 variant->key.clip_xy,
1247 variant->key.clip_z,
1248 variant->key.clip_user,
1249 variant->key.clip_halfz,
1250 variant->key.nr_planes,
1251 context_ptr);
1252 /* return clipping boolean value for function */
1253 clipmask_bool(gallivm, clipmask, ret_ptr);
1254 }
1255 else{
1256 clipmask = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 0);
1257 }
1258
1259 /* do viewport mapping */
1260 if (!bypass_viewport){
1261 generate_viewport(llvm, builder, outputs, context_ptr);
1262 }
1263
1264 /* store clipmask in vertex header and positions in data */
1265 convert_to_aos(gallivm, io, outputs, clipmask,
1266 draw->vs.vertex_shader->info.num_outputs,
1267 max_vertices);
1268 }
1269
1270 lp_build_loop_end_cond(&lp_loop, end, step, LLVMIntUGE);
1271
1272 sampler->destroy(sampler);
1273
1274 ret = LLVMBuildLoad(builder, ret_ptr,"");
1275 LLVMBuildRet(builder, ret);
1276
1277 /*
1278 * Translate the LLVM IR into machine code.
1279 */
1280 #ifdef DEBUG
1281 if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
1282 lp_debug_dump_value(variant->function);
1283 assert(0);
1284 }
1285 #endif
1286
1287 LLVMRunFunctionPassManager(gallivm->passmgr, variant->function);
1288
1289 if (gallivm_debug & GALLIVM_DEBUG_IR) {
1290 lp_debug_dump_value(variant->function);
1291 debug_printf("\n");
1292 }
1293
1294 code = LLVMGetPointerToGlobal(gallivm->engine, variant->function);
1295 variant->jit_func = (draw_jit_vert_func)pointer_to_func(code);
1296
1297 if (gallivm_debug & GALLIVM_DEBUG_ASM) {
1298 lp_disassemble(code);
1299 }
1300 lp_func_delete_body(variant->function);
1301 }
1302
1303
1304 static void
1305 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1306 {
1307 struct gallivm_state *gallivm = llvm->gallivm;
1308 LLVMContextRef context = gallivm->context;
1309 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1310 LLVMTypeRef arg_types[8];
1311 LLVMTypeRef func_type;
1312 LLVMValueRef context_ptr;
1313 LLVMBasicBlockRef block;
1314 LLVMBuilderRef builder;
1315 LLVMValueRef fetch_elts, fetch_count, stride, step, io_itr;
1316 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1317 LLVMValueRef instance_id;
1318 struct draw_context *draw = llvm->draw;
1319 unsigned i, j;
1320 struct lp_build_context bld;
1321 struct lp_build_loop_state lp_loop;
1322 const int max_vertices = 4;
1323 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
1324 LLVMValueRef fetch_max;
1325 void *code;
1326 struct lp_build_sampler_soa *sampler = 0;
1327 LLVMValueRef ret, ret_ptr;
1328 boolean bypass_viewport = variant->key.bypass_viewport;
1329 boolean enable_cliptest = variant->key.clip_xy ||
1330 variant->key.clip_z ||
1331 variant->key.clip_user;
1332
1333 arg_types[0] = get_context_ptr_type(llvm); /* context */
1334 arg_types[1] = get_vertex_header_ptr_type(llvm); /* vertex_header */
1335 arg_types[2] = get_buffer_ptr_type(llvm); /* vbuffers */
1336 arg_types[3] = LLVMPointerType(int32_type, 0); /* fetch_elts * */
1337 arg_types[4] = int32_type; /* fetch_count */
1338 arg_types[5] = int32_type; /* stride */
1339 arg_types[6] = get_vb_ptr_type(llvm); /* pipe_vertex_buffer's */
1340 arg_types[7] = int32_type; /* instance_id */
1341
1342 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1343
1344 variant->function_elts = LLVMAddFunction(gallivm->module, "draw_llvm_shader_elts", func_type);
1345 LLVMSetFunctionCallConv(variant->function_elts, LLVMCCallConv);
1346 for(i = 0; i < Elements(arg_types); ++i)
1347 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1348 LLVMAddAttribute(LLVMGetParam(variant->function_elts, i),
1349 LLVMNoAliasAttribute);
1350
1351 context_ptr = LLVMGetParam(variant->function_elts, 0);
1352 io_ptr = LLVMGetParam(variant->function_elts, 1);
1353 vbuffers_ptr = LLVMGetParam(variant->function_elts, 2);
1354 fetch_elts = LLVMGetParam(variant->function_elts, 3);
1355 fetch_count = LLVMGetParam(variant->function_elts, 4);
1356 stride = LLVMGetParam(variant->function_elts, 5);
1357 vb_ptr = LLVMGetParam(variant->function_elts, 6);
1358 instance_id = LLVMGetParam(variant->function_elts, 7);
1359
1360 lp_build_name(context_ptr, "context");
1361 lp_build_name(io_ptr, "io");
1362 lp_build_name(vbuffers_ptr, "vbuffers");
1363 lp_build_name(fetch_elts, "fetch_elts");
1364 lp_build_name(fetch_count, "fetch_count");
1365 lp_build_name(stride, "stride");
1366 lp_build_name(vb_ptr, "vb");
1367 lp_build_name(instance_id, "instance_id");
1368
1369 /*
1370 * Function body
1371 */
1372
1373 block = LLVMAppendBasicBlockInContext(gallivm->context, variant->function_elts, "entry");
1374 builder = gallivm->builder;
1375 LLVMPositionBuilderAtEnd(builder, block);
1376
1377 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1378
1379 step = lp_build_const_int32(gallivm, max_vertices);
1380
1381 /* code generated texture sampling */
1382 sampler = draw_llvm_sampler_soa_create(
1383 draw_llvm_variant_key_samplers(&variant->key),
1384 context_ptr);
1385
1386 fetch_max = LLVMBuildSub(builder, fetch_count,
1387 lp_build_const_int32(gallivm, 1),
1388 "fetch_max");
1389
1390 /* function returns non-zero i32 value if any clipped vertices */
1391 ret_ptr = lp_build_alloca(gallivm, int32_type, "");
1392 LLVMBuildStore(builder, lp_build_const_int32(gallivm, 0), ret_ptr);
1393
1394 lp_build_loop_begin(&lp_loop, gallivm, lp_build_const_int32(gallivm, 0));
1395 {
1396 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
1397 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
1398 LLVMValueRef io;
1399 LLVMValueRef clipmask; /* holds the clipmask value */
1400 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
1401
1402 io_itr = lp_loop.counter;
1403 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1404 #if DEBUG_STORE
1405 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
1406 io_itr, io, lp_loop.counter);
1407 #endif
1408 for (i = 0; i < NUM_CHANNELS; ++i) {
1409 LLVMValueRef true_index = LLVMBuildAdd(
1410 builder,
1411 lp_loop.counter,
1412 lp_build_const_int32(gallivm, i), "");
1413 LLVMValueRef fetch_ptr;
1414
1415 /* make sure we're not out of bounds which can happen
1416 * if fetch_count % 4 != 0, because on the last iteration
1417 * a few of the 4 vertex fetches will be out of bounds */
1418 true_index = lp_build_min(&bld, true_index, fetch_max);
1419
1420 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1421 &true_index, 1, "");
1422 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
1423 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1424 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1425 LLVMValueRef vb_index = lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1426 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
1427 &vb_index, 1, "");
1428 generate_fetch(gallivm, vbuffers_ptr,
1429 &aos_attribs[j][i], velem, vb, true_index,
1430 instance_id);
1431 }
1432 }
1433 convert_to_soa(gallivm, aos_attribs, inputs,
1434 draw->pt.nr_vertex_elements);
1435
1436 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
1437 generate_vs(llvm,
1438 builder,
1439 outputs,
1440 ptr_aos,
1441 context_ptr,
1442 sampler);
1443
1444 /* store original positions in clip before further manipulation */
1445 store_clip(gallivm, io, outputs);
1446
1447 /* do cliptest */
1448 if (enable_cliptest){
1449 /* allocate clipmask, assign it integer type */
1450 clipmask = generate_clipmask(gallivm, outputs,
1451 variant->key.clip_xy,
1452 variant->key.clip_z,
1453 variant->key.clip_user,
1454 variant->key.clip_halfz,
1455 variant->key.nr_planes,
1456 context_ptr);
1457 /* return clipping boolean value for function */
1458 clipmask_bool(gallivm, clipmask, ret_ptr);
1459 }
1460 else{
1461 clipmask = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 0);
1462 }
1463
1464 /* do viewport mapping */
1465 if (!bypass_viewport){
1466 generate_viewport(llvm, builder, outputs, context_ptr);
1467 }
1468
1469 /* store clipmask in vertex header,
1470 * original positions in clip
1471 * and transformed positions in data
1472 */
1473 convert_to_aos(gallivm, io, outputs, clipmask,
1474 draw->vs.vertex_shader->info.num_outputs,
1475 max_vertices);
1476 }
1477
1478 lp_build_loop_end_cond(&lp_loop, fetch_count, step, LLVMIntUGE);
1479
1480 sampler->destroy(sampler);
1481
1482 ret = LLVMBuildLoad(builder, ret_ptr,"");
1483 LLVMBuildRet(builder, ret);
1484
1485 /*
1486 * Translate the LLVM IR into machine code.
1487 */
1488 #ifdef DEBUG
1489 if(LLVMVerifyFunction(variant->function_elts, LLVMPrintMessageAction)) {
1490 lp_debug_dump_value(variant->function_elts);
1491 assert(0);
1492 }
1493 #endif
1494
1495 LLVMRunFunctionPassManager(gallivm->passmgr, variant->function_elts);
1496
1497 if (gallivm_debug & GALLIVM_DEBUG_IR) {
1498 lp_debug_dump_value(variant->function_elts);
1499 debug_printf("\n");
1500 }
1501
1502 code = LLVMGetPointerToGlobal(gallivm->engine, variant->function_elts);
1503 variant->jit_func_elts = (draw_jit_vert_func_elts)pointer_to_func(code);
1504
1505 if (gallivm_debug & GALLIVM_DEBUG_ASM) {
1506 lp_disassemble(code);
1507 }
1508 lp_func_delete_body(variant->function_elts);
1509 }
1510
1511
1512 struct draw_llvm_variant_key *
1513 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1514 {
1515 unsigned i;
1516 struct draw_llvm_variant_key *key;
1517 struct lp_sampler_static_state *sampler;
1518
1519 key = (struct draw_llvm_variant_key *)store;
1520
1521 /* Presumably all variants of the shader should have the same
1522 * number of vertex elements - ie the number of shader inputs.
1523 */
1524 key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
1525
1526 /* will have to rig this up properly later */
1527 key->clip_xy = llvm->draw->clip_xy;
1528 key->clip_z = llvm->draw->clip_z;
1529 key->clip_user = llvm->draw->clip_user;
1530 key->bypass_viewport = llvm->draw->identity_viewport;
1531 key->clip_halfz = !llvm->draw->rasterizer->gl_rasterization_rules;
1532 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1533 key->nr_planes = llvm->draw->nr_planes;
1534 key->pad = 0;
1535
1536 /* All variants of this shader will have the same value for
1537 * nr_samplers. Not yet trying to compact away holes in the
1538 * sampler array.
1539 */
1540 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1541
1542 sampler = draw_llvm_variant_key_samplers(key);
1543
1544 memcpy(key->vertex_element,
1545 llvm->draw->pt.vertex_element,
1546 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1547
1548 memset(sampler, 0, key->nr_samplers * sizeof *sampler);
1549
1550 for (i = 0 ; i < key->nr_samplers; i++) {
1551 lp_sampler_static_state(&sampler[i],
1552 llvm->draw->sampler_views[i],
1553 llvm->draw->samplers[i]);
1554 }
1555
1556 return key;
1557 }
1558
1559 void
1560 draw_llvm_set_mapped_texture(struct draw_context *draw,
1561 unsigned sampler_idx,
1562 uint32_t width, uint32_t height, uint32_t depth,
1563 uint32_t last_level,
1564 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1565 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1566 const void *data[PIPE_MAX_TEXTURE_LEVELS])
1567 {
1568 unsigned j;
1569 struct draw_jit_texture *jit_tex;
1570
1571 assert(sampler_idx < PIPE_MAX_VERTEX_SAMPLERS);
1572
1573
1574 jit_tex = &draw->llvm->jit_context.textures[sampler_idx];
1575
1576 jit_tex->width = width;
1577 jit_tex->height = height;
1578 jit_tex->depth = depth;
1579 jit_tex->last_level = last_level;
1580
1581 for (j = 0; j <= last_level; j++) {
1582 jit_tex->data[j] = data[j];
1583 jit_tex->row_stride[j] = row_stride[j];
1584 jit_tex->img_stride[j] = img_stride[j];
1585 }
1586 }
1587
1588
1589 void
1590 draw_llvm_set_sampler_state(struct draw_context *draw)
1591 {
1592 unsigned i;
1593
1594 for (i = 0; i < draw->num_samplers; i++) {
1595 struct draw_jit_texture *jit_tex = &draw->llvm->jit_context.textures[i];
1596
1597 if (draw->samplers[i]) {
1598 jit_tex->min_lod = draw->samplers[i]->min_lod;
1599 jit_tex->max_lod = draw->samplers[i]->max_lod;
1600 jit_tex->lod_bias = draw->samplers[i]->lod_bias;
1601 COPY_4V(jit_tex->border_color, draw->samplers[i]->border_color);
1602 }
1603 }
1604 }
1605
1606
1607 void
1608 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1609 {
1610 struct draw_llvm *llvm = variant->llvm;
1611
1612 if (variant->function_elts) {
1613 LLVMFreeMachineCodeForFunction(llvm->gallivm->engine,
1614 variant->function_elts);
1615 LLVMDeleteFunction(variant->function_elts);
1616 }
1617
1618 if (variant->function) {
1619 LLVMFreeMachineCodeForFunction(llvm->gallivm->engine,
1620 variant->function);
1621 LLVMDeleteFunction(variant->function);
1622 }
1623
1624 remove_from_list(&variant->list_item_local);
1625 variant->shader->variants_cached--;
1626 remove_from_list(&variant->list_item_global);
1627 llvm->nr_variants--;
1628 FREE(variant);
1629 }