projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallivm,draw,llvmpipe: Remove support for versions of LLVM prior to 3.1.
[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 #include "draw_gs.h"
33
34 #include "gallivm/lp_bld_arit.h"
35 #include "gallivm/lp_bld_arit_overflow.h"
36 #include "gallivm/lp_bld_logic.h"
37 #include "gallivm/lp_bld_const.h"
38 #include "gallivm/lp_bld_swizzle.h"
39 #include "gallivm/lp_bld_struct.h"
40 #include "gallivm/lp_bld_type.h"
41 #include "gallivm/lp_bld_flow.h"
42 #include "gallivm/lp_bld_debug.h"
43 #include "gallivm/lp_bld_tgsi.h"
44 #include "gallivm/lp_bld_printf.h"
45 #include "gallivm/lp_bld_intr.h"
46 #include "gallivm/lp_bld_init.h"
47 #include "gallivm/lp_bld_type.h"
48 #include "gallivm/lp_bld_pack.h"
49 #include "gallivm/lp_bld_format.h"
50
51 #include "tgsi/tgsi_exec.h"
52 #include "tgsi/tgsi_dump.h"
53
54 #include "util/u_math.h"
55 #include "util/u_pointer.h"
56 #include "util/u_string.h"
57 #include "util/u_simple_list.h"
58
59
60 #define DEBUG_STORE 0
61
62
63 static void
64 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
65 boolean elts);
66
67
68 struct draw_gs_llvm_iface {
69 struct lp_build_tgsi_gs_iface base;
70
71 struct draw_gs_llvm_variant *variant;
72 LLVMValueRef input;
73 };
74
75 static INLINE const struct draw_gs_llvm_iface *
76 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
77 {
78 return (const struct draw_gs_llvm_iface *)iface;
79 }
80
81 /**
82 * Create LLVM type for draw_vertex_buffer.
83 */
84 static LLVMTypeRef
85 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
86 const char *struct_name)
87 {
88 LLVMTargetDataRef target = gallivm->target;
89 LLVMTypeRef dvbuffer_type;
90 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
91 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
92
93 elem_types[DRAW_JIT_DVBUFFER_MAP] =
94 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
95 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
96
97 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
98 Elements(elem_types), 0);
99
100 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
101 target, dvbuffer_type,
102 DRAW_JIT_DVBUFFER_MAP);
103 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
104 target, dvbuffer_type,
105 DRAW_JIT_DVBUFFER_SIZE);
106
107 return dvbuffer_type;
108 }
109
110 /**
111 * Create LLVM type for struct draw_jit_texture
112 */
113 static LLVMTypeRef
114 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
115 {
116 LLVMTargetDataRef target = gallivm->target;
117 LLVMTypeRef texture_type;
118 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
119 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
120
121 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
122 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
123 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
124 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
125 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
126 elem_types[DRAW_JIT_TEXTURE_BASE] =
127 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
128 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
129 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
130 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
131 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
132
133 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
134 Elements(elem_types), 0);
135
136 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
137 target, texture_type,
138 DRAW_JIT_TEXTURE_WIDTH);
139 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
140 target, texture_type,
141 DRAW_JIT_TEXTURE_HEIGHT);
142 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
143 target, texture_type,
144 DRAW_JIT_TEXTURE_DEPTH);
145 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
146 target, texture_type,
147 DRAW_JIT_TEXTURE_FIRST_LEVEL);
148 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
149 target, texture_type,
150 DRAW_JIT_TEXTURE_LAST_LEVEL);
151 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
152 target, texture_type,
153 DRAW_JIT_TEXTURE_BASE);
154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
155 target, texture_type,
156 DRAW_JIT_TEXTURE_ROW_STRIDE);
157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
158 target, texture_type,
159 DRAW_JIT_TEXTURE_IMG_STRIDE);
160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
161 target, texture_type,
162 DRAW_JIT_TEXTURE_MIP_OFFSETS);
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_sampler
172 */
173 static LLVMTypeRef
174 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
175 {
176 LLVMTargetDataRef target = gallivm->target;
177 LLVMTypeRef sampler_type;
178 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
179
180 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
181 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
182 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
183 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
184 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
185
186 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
187 Elements(elem_types), 0);
188
189 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
190 target, sampler_type,
191 DRAW_JIT_SAMPLER_MIN_LOD);
192 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
193 target, sampler_type,
194 DRAW_JIT_SAMPLER_MAX_LOD);
195 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
196 target, sampler_type,
197 DRAW_JIT_SAMPLER_LOD_BIAS);
198 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
199 target, sampler_type,
200 DRAW_JIT_SAMPLER_BORDER_COLOR);
201
202 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
203
204 return sampler_type;
205 }
206
207
208 /**
209 * Create LLVM type for struct draw_jit_context
210 */
211 static LLVMTypeRef
212 create_jit_context_type(struct gallivm_state *gallivm,
213 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
214 const char *struct_name)
215 {
216 LLVMTargetDataRef target = gallivm->target;
217 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
218 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
219 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
220 LLVMTypeRef context_type;
221
222 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
223 LP_MAX_TGSI_CONST_BUFFERS);
224 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
225 LP_MAX_TGSI_CONST_BUFFERS);
226 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
227 DRAW_TOTAL_CLIP_PLANES), 0);
228 elem_types[3] = LLVMPointerType(float_type, 0); /* viewport */
229 elem_types[4] = LLVMArrayType(texture_type,
230 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
231 elem_types[5] = LLVMArrayType(sampler_type,
232 PIPE_MAX_SAMPLERS); /* samplers */
233 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
234 Elements(elem_types), 0);
235 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
236 target, context_type, DRAW_JIT_CTX_CONSTANTS);
237 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
238 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
239 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
240 target, context_type, DRAW_JIT_CTX_PLANES);
241 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewport,
242 target, context_type, DRAW_JIT_CTX_VIEWPORT);
243 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
244 target, context_type,
245 DRAW_JIT_CTX_TEXTURES);
246 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
247 target, context_type,
248 DRAW_JIT_CTX_SAMPLERS);
249 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
250 target, context_type);
251
252 return context_type;
253 }
254
255
256 /**
257 * Create LLVM type for struct draw_gs_jit_context
258 */
259 static LLVMTypeRef
260 create_gs_jit_context_type(struct gallivm_state *gallivm,
261 unsigned vector_length,
262 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
263 const char *struct_name)
264 {
265 LLVMTargetDataRef target = gallivm->target;
266 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
267 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
268 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
269 LLVMTypeRef context_type;
270
271 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
272 LP_MAX_TGSI_CONST_BUFFERS);
273 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
274 LP_MAX_TGSI_CONST_BUFFERS);
275 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
276 DRAW_TOTAL_CLIP_PLANES), 0);
277 elem_types[3] = LLVMPointerType(float_type, 0); /* viewport */
278
279 elem_types[4] = LLVMArrayType(texture_type,
280 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
281 elem_types[5] = LLVMArrayType(sampler_type,
282 PIPE_MAX_SAMPLERS); /* samplers */
283
284 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
285 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
286 vector_length), 0);
287 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
288 vector_length), 0);
289
290 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
291 Elements(elem_types), 0);
292
293 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
294 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
295 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
296 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
297 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
298 target, context_type, DRAW_GS_JIT_CTX_PLANES);
299 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewport,
300 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
301 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
302 target, context_type,
303 DRAW_GS_JIT_CTX_TEXTURES);
304 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
305 target, context_type,
306 DRAW_GS_JIT_CTX_SAMPLERS);
307 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
308 target, context_type,
309 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
310 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
311 target, context_type,
312 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
313 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
314 target, context_type,
315 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
316 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
317 target, context_type);
318
319 return context_type;
320 }
321
322
323 static LLVMTypeRef
324 create_gs_jit_input_type(struct gallivm_state *gallivm)
325 {
326 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
327 LLVMTypeRef input_array;
328
329 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
330 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
331 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
332 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
333
334 return input_array;
335 }
336
337 /**
338 * Create LLVM type for struct pipe_vertex_buffer
339 */
340 static LLVMTypeRef
341 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
342 const char *struct_name)
343 {
344 LLVMTargetDataRef target = gallivm->target;
345 LLVMTypeRef elem_types[4];
346 LLVMTypeRef vb_type;
347
348 elem_types[0] =
349 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
350 elem_types[2] =
351 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
352
353 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
354 Elements(elem_types), 0);
355
356 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
357 target, vb_type, 0);
358 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
359 target, vb_type, 1);
360
361 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
362
363 return vb_type;
364 }
365
366
367 /**
368 * Create LLVM type for struct vertex_header;
369 */
370 static LLVMTypeRef
371 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
372 {
373 LLVMTargetDataRef target = gallivm->target;
374 LLVMTypeRef elem_types[4];
375 LLVMTypeRef vertex_header;
376 char struct_name[24];
377
378 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
379
380 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
381 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
382 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
383 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
384
385 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
386 Elements(elem_types), 0);
387
388 /* these are bit-fields and we can't take address of them
389 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
390 target, vertex_header,
391 DRAW_JIT_VERTEX_CLIPMASK);
392 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
393 target, vertex_header,
394 DRAW_JIT_VERTEX_EDGEFLAG);
395 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
396 target, vertex_header,
397 DRAW_JIT_VERTEX_PAD);
398 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
399 target, vertex_header,
400 DRAW_JIT_VERTEX_VERTEX_ID);
401 */
402 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
403 target, vertex_header,
404 DRAW_JIT_VERTEX_CLIP);
405 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
406 target, vertex_header,
407 DRAW_JIT_VERTEX_PRE_CLIP_POS);
408 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
409 target, vertex_header,
410 DRAW_JIT_VERTEX_DATA);
411
412 assert(LLVMABISizeOfType(target, vertex_header) ==
413 offsetof(struct vertex_header, data[data_elems]));
414
415 return vertex_header;
416 }
417
418
419 /**
420 * Create LLVM types for various structures.
421 */
422 static void
423 create_jit_types(struct draw_llvm_variant *variant)
424 {
425 struct gallivm_state *gallivm = variant->gallivm;
426 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
427 vb_type;
428
429 texture_type = create_jit_texture_type(gallivm, "texture");
430 sampler_type = create_jit_sampler_type(gallivm, "sampler");
431
432 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
433 "draw_jit_context");
434 variant->context_ptr_type = LLVMPointerType(context_type, 0);
435
436 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
437 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
438
439 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
440 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
441 }
442
443
444 static LLVMTypeRef
445 get_context_ptr_type(struct draw_llvm_variant *variant)
446 {
447 if (!variant->context_ptr_type)
448 create_jit_types(variant);
449 return variant->context_ptr_type;
450 }
451
452
453 static LLVMTypeRef
454 get_buffer_ptr_type(struct draw_llvm_variant *variant)
455 {
456 if (!variant->buffer_ptr_type)
457 create_jit_types(variant);
458 return variant->buffer_ptr_type;
459 }
460
461
462 static LLVMTypeRef
463 get_vb_ptr_type(struct draw_llvm_variant *variant)
464 {
465 if (!variant->vb_ptr_type)
466 create_jit_types(variant);
467 return variant->vb_ptr_type;
468 }
469
470 static LLVMTypeRef
471 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
472 {
473 if (!variant->vertex_header_ptr_type)
474 create_jit_types(variant);
475 return variant->vertex_header_ptr_type;
476 }
477
478
479 /**
480 * Create per-context LLVM info.
481 */
482 struct draw_llvm *
483 draw_llvm_create(struct draw_context *draw)
484 {
485 struct draw_llvm *llvm;
486
487 llvm = CALLOC_STRUCT( draw_llvm );
488 if (!llvm)
489 return NULL;
490
491 lp_build_init();
492
493 llvm->draw = draw;
494
495 llvm->nr_variants = 0;
496 make_empty_list(&llvm->vs_variants_list);
497
498 llvm->nr_gs_variants = 0;
499 make_empty_list(&llvm->gs_variants_list);
500
501 return llvm;
502 }
503
504
505 /**
506 * Free per-context LLVM info.
507 */
508 void
509 draw_llvm_destroy(struct draw_llvm *llvm)
510 {
511 /* XXX free other draw_llvm data? */
512 FREE(llvm);
513 }
514
515
516 /**
517 * Create LLVM-generated code for a vertex shader.
518 */
519 struct draw_llvm_variant *
520 draw_llvm_create_variant(struct draw_llvm *llvm,
521 unsigned num_inputs,
522 const struct draw_llvm_variant_key *key)
523 {
524 struct draw_llvm_variant *variant;
525 struct llvm_vertex_shader *shader =
526 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
527 LLVMTypeRef vertex_header;
528
529 variant = MALLOC(sizeof *variant +
530 shader->variant_key_size -
531 sizeof variant->key);
532 if (variant == NULL)
533 return NULL;
534
535 variant->llvm = llvm;
536
537 variant->gallivm = gallivm_create();
538
539 create_jit_types(variant);
540
541 memcpy(&variant->key, key, shader->variant_key_size);
542
543 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
544
545 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
546
547 draw_llvm_generate(llvm, variant, FALSE); /* linear */
548 draw_llvm_generate(llvm, variant, TRUE); /* elts */
549
550 gallivm_compile_module(variant->gallivm);
551
552 variant->jit_func = (draw_jit_vert_func)
553 gallivm_jit_function(variant->gallivm, variant->function);
554
555 variant->jit_func_elts = (draw_jit_vert_func_elts)
556 gallivm_jit_function(variant->gallivm, variant->function_elts);
557
558 variant->shader = shader;
559 variant->list_item_global.base = variant;
560 variant->list_item_local.base = variant;
561 /*variant->no = */shader->variants_created++;
562 variant->list_item_global.base = variant;
563
564 return variant;
565 }
566
567
568 static void
569 generate_vs(struct draw_llvm_variant *variant,
570 LLVMBuilderRef builder,
571 struct lp_type vs_type,
572 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
573 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
574 const struct lp_bld_tgsi_system_values *system_values,
575 LLVMValueRef context_ptr,
576 struct lp_build_sampler_soa *draw_sampler,
577 boolean clamp_vertex_color)
578 {
579 struct draw_llvm *llvm = variant->llvm;
580 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
581 LLVMValueRef consts_ptr =
582 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
583 LLVMValueRef num_consts_ptr =
584 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
585 struct lp_build_sampler_soa *sampler = 0;
586
587 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
588 tgsi_dump(tokens, 0);
589 draw_llvm_dump_variant_key(&variant->key);
590 }
591
592 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers)
593 sampler = draw_sampler;
594
595 lp_build_tgsi_soa(variant->gallivm,
596 tokens,
597 vs_type,
598 NULL /*struct lp_build_mask_context *mask*/,
599 consts_ptr,
600 num_consts_ptr,
601 system_values,
602 inputs,
603 outputs,
604 sampler,
605 &llvm->draw->vs.vertex_shader->info,
606 NULL);
607
608 {
609 LLVMValueRef out;
610 unsigned chan, attrib;
611 struct lp_build_context bld;
612 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
613 lp_build_context_init(&bld, variant->gallivm, vs_type);
614
615 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
616 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
617 if (outputs[attrib][chan]) {
618 switch (info->output_semantic_name[attrib]) {
619 case TGSI_SEMANTIC_COLOR:
620 case TGSI_SEMANTIC_BCOLOR:
621 if (clamp_vertex_color) {
622 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
623 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
624 LLVMBuildStore(builder, out, outputs[attrib][chan]);
625 }
626 break;
627 }
628 }
629 }
630 }
631 }
632 }
633
634 static void
635 generate_fetch(struct gallivm_state *gallivm,
636 struct draw_context *draw,
637 LLVMValueRef vbuffers_ptr,
638 LLVMValueRef *res,
639 struct pipe_vertex_element *velem,
640 LLVMValueRef vbuf,
641 LLVMValueRef index,
642 LLVMValueRef instance_id)
643 {
644 const struct util_format_description *format_desc =
645 util_format_description(velem->src_format);
646 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
647 LLVMBuilderRef builder = gallivm->builder;
648 LLVMValueRef indices =
649 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
650 velem->vertex_buffer_index, 0);
651 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
652 &indices, 1, "");
653 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
654 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
655 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
656 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
657 LLVMValueRef stride;
658 LLVMValueRef buffer_overflowed;
659 LLVMValueRef needed_buffer_size;
660 LLVMValueRef temp_ptr =
661 lp_build_alloca(gallivm,
662 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
663 LLVMValueRef ofbit = NULL;
664 struct lp_build_if_state if_ctx;
665
666 if (velem->instance_divisor) {
667 /* Index is equal to the start instance plus the number of current
668 * instance divided by the divisor. In this case we compute it as:
669 * index = start_instance + (instance_id / divisor)
670 */
671 LLVMValueRef current_instance;
672 index = lp_build_const_int32(gallivm, draw->start_instance);
673 current_instance = LLVMBuildUDiv(builder, instance_id,
674 lp_build_const_int32(gallivm, velem->instance_divisor),
675 "instance_divisor");
676 index = lp_build_uadd_overflow(gallivm, index, current_instance, &ofbit);
677 }
678
679 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
680 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
681 stride = lp_build_uadd_overflow(
682 gallivm, stride,
683 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
684 needed_buffer_size = lp_build_uadd_overflow(
685 gallivm, stride,
686 lp_build_const_int32(gallivm,
687 util_format_get_blocksize(velem->src_format)),
688 &ofbit);
689
690 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
691 needed_buffer_size, buffer_size,
692 "buffer_overflowed");
693 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
694 #if 0
695 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
696 index, vb_stride);
697 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
698 vb_buffer_offset,
699 lp_build_const_int32(gallivm, velem->src_offset));
700 lp_build_print_value(gallivm, " blocksize = ",
701 lp_build_const_int32(
702 gallivm,
703 util_format_get_blocksize(velem->src_format)));
704 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
705 lp_build_printf(gallivm, " stride = %u\n", stride);
706 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
707 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
708 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
709 #endif
710
711 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
712 {
713 LLVMValueRef val =
714 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
715 LLVMBuildStore(builder, val, temp_ptr);
716 }
717 lp_build_else(&if_ctx);
718 {
719 LLVMValueRef val;
720 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
721
722 val = lp_build_fetch_rgba_aos(gallivm,
723 format_desc,
724 lp_float32_vec4_type(),
725 map_ptr,
726 zero, zero, zero);
727 LLVMBuildStore(builder, val, temp_ptr);
728 }
729 lp_build_endif(&if_ctx);
730
731 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
732 }
733
734 static void
735 convert_to_soa(struct gallivm_state *gallivm,
736 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
737 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
738 unsigned num_attribs, const struct lp_type soa_type)
739 {
740 unsigned i, j, k;
741 struct lp_type aos_channel_type = soa_type;
742
743 debug_assert(TGSI_NUM_CHANNELS == 4);
744 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
745
746 aos_channel_type.length >>= 1;
747
748 for (i = 0; i < num_attribs; ++i) {
749 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
750 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
751
752 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
753 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
754
755 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
756
757 for (k = 0; k < pixels_per_channel; ++k) {
758 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
759 }
760
761 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
762 }
763
764 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
765 }
766 }
767
768
769 static void
770 store_aos(struct gallivm_state *gallivm,
771 LLVMValueRef io_ptr,
772 LLVMValueRef index,
773 LLVMValueRef value)
774 {
775 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
776 LLVMBuilderRef builder = gallivm->builder;
777 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
778 LLVMValueRef indices[3];
779
780 indices[0] = lp_build_const_int32(gallivm, 0);
781 indices[1] = index;
782 indices[2] = lp_build_const_int32(gallivm, 0);
783
784 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
785 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
786
787 #if DEBUG_STORE
788 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
789 #endif
790
791 /* Unaligned store due to the vertex header */
792 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
793 }
794
795 /**
796 * Adjust the mask to architecture endianess. The mask will the store in struct:
797 *
798 * struct vertex_header {
799 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
800 * unsigned edgeflag:1;
801 * unsigned have_clipdist:1;
802 * unsigned vertex_id:16;
803 * [...]
804 * }
805 *
806 * On little-endian machine nothing needs to done, however on bit-endian machine
807 * the mask's fields need to be adjusted with the algorithm:
808 *
809 * uint32_t reverse (uint32_t x)
810 * {
811 * return (x >> 16) | // vertex_id
812 * ((x & 0x3fff) << 18) | // clipmask
813 * ((x & 0x4000) << 3) | // have_clipdist
814 * ((x & 0x8000) << 1); // edgeflag
815 * }
816 */
817 static LLVMValueRef
818 adjust_mask(struct gallivm_state *gallivm,
819 LLVMValueRef mask)
820 {
821 #ifdef PIPE_ARCH_BIG_ENDIAN
822 LLVMBuilderRef builder = gallivm->builder;
823 LLVMValueRef vertex_id;
824 LLVMValueRef clipmask;
825 LLVMValueRef have_clipdist;
826 LLVMValueRef edgeflag;
827
828 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
829 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
830 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
831 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
832 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
833 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
834 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
835
836 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
837 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
838 mask = LLVMBuildOr(builder, mask, edgeflag, "");
839 #endif
840 return mask;
841 }
842
843 static void
844 store_aos_array(struct gallivm_state *gallivm,
845 struct lp_type soa_type,
846 LLVMValueRef io_ptr,
847 LLVMValueRef *indices,
848 LLVMValueRef* aos,
849 int attrib,
850 int num_outputs,
851 LLVMValueRef clipmask,
852 boolean have_clipdist)
853 {
854 LLVMBuilderRef builder = gallivm->builder;
855 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
856 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
857 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
858 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
859 int vector_length = soa_type.length;
860 int i;
861
862 debug_assert(TGSI_NUM_CHANNELS == 4);
863
864 for (i = 0; i < vector_length; i++) {
865 linear_inds[i] = lp_build_const_int32(gallivm, i);
866 if (indices) {
867 inds[i] = indices[i];
868 } else {
869 inds[i] = linear_inds[i];
870 }
871 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
872 }
873
874 if (attrib == 0) {
875 /* store vertex header for each of the n vertices */
876 LLVMValueRef val, cliptmp;
877 int vertex_id_pad_edgeflag;
878
879 /* If this assertion fails, it means we need to update the bit twidding
880 * code here. See struct vertex_header in draw_private.h.
881 */
882 assert(DRAW_TOTAL_CLIP_PLANES==14);
883 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
884 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
885 if (have_clipdist)
886 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
887 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
888 /* OR with the clipmask */
889 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
890 for (i = 0; i < vector_length; i++) {
891 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
892 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
893 val = adjust_mask(gallivm, val);
894 #if DEBUG_STORE
895 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
896 io_ptrs[i], inds[i], val);
897 #endif
898 LLVMBuildStore(builder, val, id_ptr);
899 }
900 }
901
902 /* store for each of the n vertices */
903 for (i = 0; i < vector_length; i++) {
904 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
905 }
906 }
907
908
909 static void
910 convert_to_aos(struct gallivm_state *gallivm,
911 LLVMValueRef io,
912 LLVMValueRef *indices,
913 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
914 LLVMValueRef clipmask,
915 int num_outputs,
916 struct lp_type soa_type,
917 boolean have_clipdist)
918 {
919 LLVMBuilderRef builder = gallivm->builder;
920 unsigned chan, attrib, i;
921
922 #if DEBUG_STORE
923 lp_build_printf(gallivm, " # storing begin\n");
924 #endif
925 for (attrib = 0; attrib < num_outputs; ++attrib) {
926 LLVMValueRef soa[TGSI_NUM_CHANNELS];
927 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
928 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
929 if (outputs[attrib][chan]) {
930 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
931 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
932 #if DEBUG_STORE
933 lp_build_printf(gallivm, "output %d : %d ",
934 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
935 attrib, 0),
936 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
937 chan, 0));
938 lp_build_print_value(gallivm, "val = ", out);
939 {
940 LLVMValueRef iv =
941 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
942
943 lp_build_print_value(gallivm, " ival = ", iv);
944 }
945 #endif
946 soa[chan] = out;
947 }
948 else {
949 soa[chan] = 0;
950 }
951 }
952
953
954 if (soa_type.length == TGSI_NUM_CHANNELS) {
955 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
956 } else {
957 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
958
959 for (i = 0; i < soa_type.length; ++i) {
960 aos[i] = lp_build_extract_range(gallivm,
961 soa[i % TGSI_NUM_CHANNELS],
962 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
963 TGSI_NUM_CHANNELS);
964 }
965 }
966
967 store_aos_array(gallivm,
968 soa_type,
969 io, indices,
970 aos,
971 attrib,
972 num_outputs,
973 clipmask, have_clipdist);
974 }
975 #if DEBUG_STORE
976 lp_build_printf(gallivm, " # storing end\n");
977 #endif
978 }
979
980
981 /**
982 * Stores original vertex positions in clip coordinates
983 */
984 static void
985 store_clip(struct gallivm_state *gallivm,
986 const struct lp_type vs_type,
987 LLVMValueRef io_ptr,
988 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
989 boolean pre_clip_pos, int idx)
990 {
991 LLVMBuilderRef builder = gallivm->builder;
992 LLVMValueRef soa[4];
993 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
994 LLVMValueRef indices[2];
995 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
996 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
997 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
998 LLVMTypeRef clip_ptr_type =
999 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1000 4), 0);
1001 int i, j;
1002
1003 indices[0] =
1004 indices[1] = lp_build_const_int32(gallivm, 0);
1005
1006 for (i = 0; i < vs_type.length; i++) {
1007 inds[i] = lp_build_const_int32(gallivm, i);
1008 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1009 }
1010
1011 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1012 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1013 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1014 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1015
1016 if (!pre_clip_pos) {
1017 for (i = 0; i < vs_type.length; i++) {
1018 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1019 }
1020 } else {
1021 for (i = 0; i < vs_type.length; i++) {
1022 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1023 }
1024 }
1025
1026 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1027 for (i = 0; i < vs_type.length; ++i) {
1028 aos[i] = lp_build_extract_range(gallivm,
1029 soa[i % TGSI_NUM_CHANNELS],
1030 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1031 TGSI_NUM_CHANNELS);
1032 }
1033
1034 for (j = 0; j < vs_type.length; j++) {
1035 LLVMValueRef clip_ptr;
1036
1037 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1038 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1039
1040 /* Unaligned store */
1041 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1042 }
1043 }
1044
1045
1046 /**
1047 * Transforms the outputs for viewport mapping
1048 */
1049 static void
1050 generate_viewport(struct draw_llvm_variant *variant,
1051 LLVMBuilderRef builder,
1052 struct lp_type vs_type,
1053 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1054 LLVMValueRef context_ptr)
1055 {
1056 int i;
1057 struct gallivm_state *gallivm = variant->gallivm;
1058 struct lp_type f32_type = vs_type;
1059 const unsigned pos = variant->llvm->draw->vs.position_output;
1060 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1061 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1062 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1063 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
1064
1065 /* for 1/w convention*/
1066 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1067 LLVMBuildStore(builder, out3, outputs[pos][3]);
1068
1069 /* Viewport Mapping */
1070 for (i=0; i<3; i++) {
1071 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1072 LLVMValueRef scale;
1073 LLVMValueRef trans;
1074 LLVMValueRef scale_i;
1075 LLVMValueRef trans_i;
1076 LLVMValueRef index;
1077
1078 index = lp_build_const_int32(gallivm, i);
1079 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1080
1081 index = lp_build_const_int32(gallivm, i+4);
1082 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1083
1084 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1085 LLVMBuildLoad(builder, scale_i, "scale"));
1086 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1087 LLVMBuildLoad(builder, trans_i, "trans"));
1088
1089 /* divide by w */
1090 out = LLVMBuildFMul(builder, out, out3, "");
1091 /* mult by scale */
1092 out = LLVMBuildFMul(builder, out, scale, "");
1093 /* add translation */
1094 out = LLVMBuildFAdd(builder, out, trans, "");
1095
1096 /* store transformed outputs */
1097 LLVMBuildStore(builder, out, outputs[pos][i]);
1098 }
1099
1100 }
1101
1102
1103 /**
1104 * Returns clipmask as nxi32 bitmask for the n vertices
1105 */
1106 static LLVMValueRef
1107 generate_clipmask(struct draw_llvm *llvm,
1108 struct gallivm_state *gallivm,
1109 struct lp_type vs_type,
1110 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1111 boolean clip_xy,
1112 boolean clip_z,
1113 boolean clip_user,
1114 boolean clip_halfz,
1115 unsigned ucp_enable,
1116 LLVMValueRef context_ptr,
1117 boolean *have_clipdist)
1118 {
1119 LLVMBuilderRef builder = gallivm->builder;
1120 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1121 LLVMValueRef test, temp;
1122 LLVMValueRef zero, shift;
1123 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1124 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1125 LLVMValueRef plane1, planes, plane_ptr, sum;
1126 struct lp_type f32_type = vs_type;
1127 struct lp_type i32_type = lp_int_type(vs_type);
1128 const unsigned pos = llvm->draw->vs.position_output;
1129 const unsigned cv = llvm->draw->vs.clipvertex_output;
1130 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1131 bool have_cd = false;
1132 unsigned cd[2];
1133
1134 cd[0] = llvm->draw->vs.clipdistance_output[0];
1135 cd[1] = llvm->draw->vs.clipdistance_output[1];
1136
1137 if (cd[0] != pos || cd[1] != pos)
1138 have_cd = true;
1139
1140 if (num_written_clipdistance && !clip_user) {
1141 clip_user = true;
1142 ucp_enable = (1 << num_written_clipdistance) - 1;
1143 }
1144
1145 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1146 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1147 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1148 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1149
1150 /*
1151 * load clipvertex and position from correct locations.
1152 * if they are the same just load them once.
1153 */
1154 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1155 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1156 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1157 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1158
1159 if (clip_user && cv != pos) {
1160 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1161 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1162 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1163 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1164 } else {
1165 cv_x = pos_x;
1166 cv_y = pos_y;
1167 cv_z = pos_z;
1168 cv_w = pos_w;
1169 }
1170
1171 /* Cliptest, for hardwired planes */
1172 if (clip_xy) {
1173 /* plane 1 */
1174 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1175 temp = shift;
1176 test = LLVMBuildAnd(builder, test, temp, "");
1177 mask = test;
1178
1179 /* plane 2 */
1180 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1181 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1182 temp = LLVMBuildShl(builder, temp, shift, "");
1183 test = LLVMBuildAnd(builder, test, temp, "");
1184 mask = LLVMBuildOr(builder, mask, test, "");
1185
1186 /* plane 3 */
1187 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1188 temp = LLVMBuildShl(builder, temp, shift, "");
1189 test = LLVMBuildAnd(builder, test, temp, "");
1190 mask = LLVMBuildOr(builder, mask, test, "");
1191
1192 /* plane 4 */
1193 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1194 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1195 temp = LLVMBuildShl(builder, temp, shift, "");
1196 test = LLVMBuildAnd(builder, test, temp, "");
1197 mask = LLVMBuildOr(builder, mask, test, "");
1198 }
1199
1200 if (clip_z) {
1201 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1202 if (clip_halfz) {
1203 /* plane 5 */
1204 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1205 test = LLVMBuildAnd(builder, test, temp, "");
1206 mask = LLVMBuildOr(builder, mask, test, "");
1207 }
1208 else {
1209 /* plane 5 */
1210 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1211 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1212 test = LLVMBuildAnd(builder, test, temp, "");
1213 mask = LLVMBuildOr(builder, mask, test, "");
1214 }
1215 /* plane 6 */
1216 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1217 temp = LLVMBuildShl(builder, temp, shift, "");
1218 test = LLVMBuildAnd(builder, test, temp, "");
1219 mask = LLVMBuildOr(builder, mask, test, "");
1220 }
1221
1222 if (clip_user) {
1223 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1224 LLVMValueRef indices[3];
1225 LLVMValueRef is_nan_or_inf;
1226
1227 /* userclip planes */
1228 while (ucp_enable) {
1229 unsigned plane_idx = ffs(ucp_enable)-1;
1230 ucp_enable &= ~(1 << plane_idx);
1231 plane_idx += 6;
1232
1233 if (have_cd && num_written_clipdistance) {
1234 LLVMValueRef clipdist;
1235 int i;
1236 i = plane_idx - 6;
1237
1238 *have_clipdist = TRUE;
1239 if (i < 4) {
1240 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1241 } else {
1242 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1243 }
1244 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1245 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1246 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1247 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1248 test = LLVMBuildAnd(builder, test, temp, "");
1249 mask = LLVMBuildOr(builder, mask, test, "");
1250 } else {
1251 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1252 indices[0] = lp_build_const_int32(gallivm, 0);
1253 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1254
1255 indices[2] = lp_build_const_int32(gallivm, 0);
1256 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1257 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1258 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1259 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1260
1261 indices[2] = lp_build_const_int32(gallivm, 1);
1262 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1263 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1264 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1265 test = LLVMBuildFMul(builder, planes, cv_y, "");
1266 sum = LLVMBuildFAdd(builder, sum, test, "");
1267
1268 indices[2] = lp_build_const_int32(gallivm, 2);
1269 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1270 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1271 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1272 test = LLVMBuildFMul(builder, planes, cv_z, "");
1273 sum = LLVMBuildFAdd(builder, sum, test, "");
1274
1275 indices[2] = lp_build_const_int32(gallivm, 3);
1276 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1277 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1278 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1279 test = LLVMBuildFMul(builder, planes, cv_w, "");
1280 sum = LLVMBuildFAdd(builder, sum, test, "");
1281
1282 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1283 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1284 test = LLVMBuildAnd(builder, test, temp, "");
1285 mask = LLVMBuildOr(builder, mask, test, "");
1286 }
1287 }
1288 }
1289 return mask;
1290 }
1291
1292
1293 /**
1294 * Returns boolean if any clipping has occurred
1295 * Used zero/non-zero i32 value to represent boolean
1296 */
1297 static LLVMValueRef
1298 clipmask_booli32(struct gallivm_state *gallivm,
1299 const struct lp_type vs_type,
1300 LLVMValueRef clipmask_bool_ptr)
1301 {
1302 LLVMBuilderRef builder = gallivm->builder;
1303 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1304 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1305 LLVMValueRef ret = LLVMConstNull(int32_type);
1306 LLVMValueRef temp;
1307 int i;
1308
1309 /*
1310 * Can do this with log2(vector length) pack instructions and one extract
1311 * (as we don't actually need a or) with sse2 which would be way better.
1312 */
1313 for (i=0; i < vs_type.length; i++) {
1314 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1315 lp_build_const_int32(gallivm, i) , "");
1316 ret = LLVMBuildOr(builder, ret, temp, "");
1317 }
1318 return ret;
1319 }
1320
1321 static LLVMValueRef
1322 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1323 struct lp_build_tgsi_context * bld_base,
1324 boolean is_vindex_indirect,
1325 LLVMValueRef vertex_index,
1326 boolean is_aindex_indirect,
1327 LLVMValueRef attrib_index,
1328 LLVMValueRef swizzle_index)
1329 {
1330 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1331 struct gallivm_state *gallivm = bld_base->base.gallivm;
1332 LLVMBuilderRef builder = gallivm->builder;
1333 LLVMValueRef indices[3];
1334 LLVMValueRef res;
1335 struct lp_type type = bld_base->base.type;
1336
1337 if (is_vindex_indirect || is_aindex_indirect) {
1338 int i;
1339 res = bld_base->base.zero;
1340 for (i = 0; i < type.length; ++i) {
1341 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1342 LLVMValueRef vert_chan_index = vertex_index;
1343 LLVMValueRef attr_chan_index = attrib_index;
1344 LLVMValueRef channel_vec, value;
1345
1346 if (is_vindex_indirect) {
1347 vert_chan_index = LLVMBuildExtractElement(builder,
1348 vertex_index, idx, "");
1349 }
1350 if (is_aindex_indirect) {
1351 attr_chan_index = LLVMBuildExtractElement(builder,
1352 attrib_index, idx, "");
1353 }
1354
1355 indices[0] = vert_chan_index;
1356 indices[1] = attr_chan_index;
1357 indices[2] = swizzle_index;
1358
1359 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1360 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1361 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1362
1363 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1364 }
1365 } else {
1366 indices[0] = vertex_index;
1367 indices[1] = attrib_index;
1368 indices[2] = swizzle_index;
1369
1370 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1371 res = LLVMBuildLoad(builder, res, "");
1372 }
1373
1374 return res;
1375 }
1376
1377 static void
1378 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1379 struct lp_build_tgsi_context * bld_base,
1380 LLVMValueRef (*outputs)[4],
1381 LLVMValueRef emitted_vertices_vec)
1382 {
1383 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1384 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1385 struct gallivm_state *gallivm = variant->gallivm;
1386 LLVMBuilderRef builder = gallivm->builder;
1387 struct lp_type gs_type = bld_base->base.type;
1388 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1389 lp_int_type(gs_type), 0);
1390 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1391 LLVMValueRef next_prim_offset =
1392 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1393 LLVMValueRef io = variant->io_ptr;
1394 unsigned i;
1395 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1396
1397 for (i = 0; i < gs_type.length; ++i) {
1398 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1399 LLVMValueRef currently_emitted =
1400 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1401 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1402 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1403 }
1404
1405 convert_to_aos(gallivm, io, indices,
1406 outputs, clipmask,
1407 gs_info->num_outputs, gs_type,
1408 FALSE);
1409 }
1410
1411 static void
1412 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1413 struct lp_build_tgsi_context * bld_base,
1414 LLVMValueRef verts_per_prim_vec,
1415 LLVMValueRef emitted_prims_vec)
1416 {
1417 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1418 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1419 struct gallivm_state *gallivm = variant->gallivm;
1420 LLVMBuilderRef builder = gallivm->builder;
1421 LLVMValueRef prim_lengts_ptr =
1422 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1423 unsigned i;
1424
1425 for (i = 0; i < bld_base->base.type.length; ++i) {
1426 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1427 LLVMValueRef prims_emitted =
1428 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1429 LLVMValueRef store_ptr;
1430 LLVMValueRef num_vertices =
1431 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1432
1433 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1434 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1435 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1436 LLVMBuildStore(builder, num_vertices, store_ptr);
1437 }
1438 }
1439
1440 static void
1441 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1442 struct lp_build_tgsi_context * bld_base,
1443 LLVMValueRef total_emitted_vertices_vec,
1444 LLVMValueRef emitted_prims_vec)
1445 {
1446 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1447 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1448 struct gallivm_state *gallivm = variant->gallivm;
1449 LLVMBuilderRef builder = gallivm->builder;
1450 LLVMValueRef emitted_verts_ptr =
1451 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1452 LLVMValueRef emitted_prims_ptr =
1453 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1454 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1455
1456 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1457 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1458
1459 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1460 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1461 }
1462
1463 static void
1464 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1465 boolean elts)
1466 {
1467 struct gallivm_state *gallivm = variant->gallivm;
1468 LLVMContextRef context = gallivm->context;
1469 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1470 LLVMTypeRef arg_types[10];
1471 unsigned num_arg_types =
1472 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1473 LLVMTypeRef func_type;
1474 LLVMValueRef context_ptr;
1475 LLVMBasicBlockRef block;
1476 LLVMBuilderRef builder;
1477 struct lp_type vs_type;
1478 LLVMValueRef end, start;
1479 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1480 LLVMValueRef vertex_id_offset;
1481 LLVMValueRef stride, step, io_itr;
1482 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1483 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1484 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1485 struct draw_context *draw = llvm->draw;
1486 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1487 unsigned i, j;
1488 struct lp_build_context bld;
1489 struct lp_build_loop_state lp_loop;
1490 const int vector_length = lp_native_vector_width / 32;
1491 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1492 LLVMValueRef fetch_max;
1493 struct lp_build_sampler_soa *sampler = 0;
1494 LLVMValueRef ret, clipmask_bool_ptr;
1495 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1496 struct draw_llvm_variant_key *key = &variant->key;
1497 /* If geometry shader is present we need to skip both the viewport
1498 * transformation and clipping otherwise the inputs to the geometry
1499 * shader will be incorrect.
1500 */
1501 const boolean bypass_viewport = gs || key->bypass_viewport;
1502 const boolean enable_cliptest = !gs && (key->clip_xy ||
1503 key->clip_z ||
1504 key->clip_user);
1505 LLVMValueRef variant_func;
1506 const unsigned pos = llvm->draw->vs.position_output;
1507 const unsigned cv = llvm->draw->vs.clipvertex_output;
1508 boolean have_clipdist = FALSE;
1509 struct lp_bld_tgsi_system_values system_values;
1510
1511 memset(&system_values, 0, sizeof(system_values));
1512
1513 i = 0;
1514 arg_types[i++] = get_context_ptr_type(variant); /* context */
1515 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1516 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1517 if (elts) {
1518 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1519 arg_types[i++] = int32_type; /* fetch_elt_max */
1520 } else
1521 arg_types[i++] = int32_type; /* start */
1522 arg_types[i++] = int32_type; /* fetch_count / count */
1523 arg_types[i++] = int32_type; /* stride */
1524 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1525 arg_types[i++] = int32_type; /* instance_id */
1526 arg_types[i++] = int32_type; /* vertex_id_offset */
1527
1528 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1529
1530 variant_func = LLVMAddFunction(gallivm->module,
1531 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1532 func_type);
1533
1534 if (elts)
1535 variant->function_elts = variant_func;
1536 else
1537 variant->function = variant_func;
1538
1539 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1540 for (i = 0; i < num_arg_types; ++i)
1541 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1542 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1543 LLVMNoAliasAttribute);
1544
1545 context_ptr = LLVMGetParam(variant_func, 0);
1546 io_ptr = LLVMGetParam(variant_func, 1);
1547 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1548 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1549 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1550 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1551 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1552
1553 lp_build_name(context_ptr, "context");
1554 lp_build_name(io_ptr, "io");
1555 lp_build_name(vbuffers_ptr, "vbuffers");
1556 lp_build_name(stride, "stride");
1557 lp_build_name(vb_ptr, "vb");
1558 lp_build_name(system_values.instance_id, "instance_id");
1559 lp_build_name(vertex_id_offset, "vertex_id_offset");
1560
1561 if (elts) {
1562 fetch_elts = LLVMGetParam(variant_func, 3);
1563 fetch_elt_max = LLVMGetParam(variant_func, 4);
1564 fetch_count = LLVMGetParam(variant_func, 5);
1565 lp_build_name(fetch_elts, "fetch_elts");
1566 lp_build_name(fetch_elt_max, "fetch_elt_max");
1567 lp_build_name(fetch_count, "fetch_count");
1568 start = count = NULL;
1569 }
1570 else {
1571 start = LLVMGetParam(variant_func, 3);
1572 count = LLVMGetParam(variant_func, 4);
1573 lp_build_name(start, "start");
1574 lp_build_name(count, "count");
1575 fetch_elts = fetch_count = NULL;
1576 }
1577
1578 /*
1579 * Function body
1580 */
1581
1582 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1583 builder = gallivm->builder;
1584 LLVMPositionBuilderAtEnd(builder, block);
1585
1586 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1587
1588 memset(&vs_type, 0, sizeof vs_type);
1589 vs_type.floating = TRUE; /* floating point values */
1590 vs_type.sign = TRUE; /* values are signed */
1591 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1592 vs_type.width = 32; /* 32-bit float */
1593 vs_type.length = vector_length;
1594
1595 /* hold temporary "bool" clipmask */
1596 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1597 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1598
1599 /* code generated texture sampling */
1600 sampler = draw_llvm_sampler_soa_create(
1601 draw_llvm_variant_key_samplers(key),
1602 context_ptr);
1603
1604 if (elts) {
1605 start = zero;
1606 end = fetch_count;
1607 count = fetch_count;
1608 }
1609 else {
1610 end = lp_build_add(&bld, start, count);
1611 }
1612
1613 step = lp_build_const_int32(gallivm, vector_length);
1614
1615 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1616
1617 lp_build_loop_begin(&lp_loop, gallivm, zero);
1618 {
1619 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1620 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1621 LLVMValueRef io;
1622 LLVMValueRef clipmask; /* holds the clipmask value */
1623 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1624
1625 io_itr = lp_loop.counter;
1626
1627 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1628 #if DEBUG_STORE
1629 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1630 io_itr, io, lp_loop.counter);
1631 #endif
1632 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1633 for (i = 0; i < vector_length; ++i) {
1634 LLVMValueRef vert_index =
1635 LLVMBuildAdd(builder,
1636 lp_loop.counter,
1637 lp_build_const_int32(gallivm, i), "");
1638 LLVMValueRef true_index =
1639 LLVMBuildAdd(builder, start, vert_index, "");
1640 LLVMValueRef vertex_id;
1641
1642 /* make sure we're not out of bounds which can happen
1643 * if fetch_count % 4 != 0, because on the last iteration
1644 * a few of the 4 vertex fetches will be out of bounds */
1645 true_index = lp_build_min(&bld, true_index, fetch_max);
1646
1647 if (elts) {
1648 LLVMValueRef fetch_ptr;
1649 LLVMValueRef index_overflowed;
1650 LLVMValueRef index_ptr =
1651 lp_build_alloca(
1652 gallivm,
1653 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1654 struct lp_build_if_state if_ctx;
1655 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1656 true_index, fetch_elt_max,
1657 "index_overflowed");
1658
1659 lp_build_if(&if_ctx, gallivm, index_overflowed);
1660 {
1661 /* Generate maximum possible index so that
1662 * generate_fetch can treat it just like
1663 * any other overflow and return zeros.
1664 * We don't have to worry about the restart
1665 * primitive index because it has already been
1666 * handled
1667 */
1668 LLVMValueRef val =
1669 lp_build_const_int32(gallivm, 0xffffffff);
1670 LLVMBuildStore(builder, val, index_ptr);
1671 }
1672 lp_build_else(&if_ctx);
1673 {
1674 LLVMValueRef val;
1675 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1676 &true_index, 1, "");
1677 val = LLVMBuildLoad(builder, fetch_ptr, "");
1678 LLVMBuildStore(builder, val, index_ptr);
1679 }
1680 lp_build_endif(&if_ctx);
1681 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1682 }
1683 /* in the paths with elts vertex id has to be unaffected by the
1684 * index bias and because indices inside our elements array have
1685 * already had index bias applied we need to subtract it here to
1686 * get back to the original index.
1687 * in the linear paths vertex id has to be unaffected by the
1688 * original start index and because we abuse the 'start' variable
1689 * to either represent the actual start index or the index at which
1690 * the primitive was split (we split rendering into chunks of at
1691 * most 4095-vertices) we need to back out the original start
1692 * index out of our vertex id here.
1693 */
1694 vertex_id = LLVMBuildSub(builder, true_index, vertex_id_offset, "");
1695
1696 system_values.vertex_id = LLVMBuildInsertElement(
1697 gallivm->builder,
1698 system_values.vertex_id, vertex_id,
1699 lp_build_const_int32(gallivm, i), "");
1700
1701 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1702 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1703 LLVMValueRef vb_index =
1704 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1705 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1706 generate_fetch(gallivm, draw, vbuffers_ptr,
1707 &aos_attribs[j][i], velem, vb, true_index,
1708 system_values.instance_id);
1709 }
1710 }
1711 convert_to_soa(gallivm, aos_attribs, inputs,
1712 draw->pt.nr_vertex_elements, vs_type);
1713
1714 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1715 generate_vs(variant,
1716 builder,
1717 vs_type,
1718 outputs,
1719 ptr_aos,
1720 &system_values,
1721 context_ptr,
1722 sampler,
1723 key->clamp_vertex_color);
1724
1725 if (pos != -1 && cv != -1) {
1726 /* store original positions in clip before further manipulation */
1727 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1728 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1729
1730 /* do cliptest */
1731 if (enable_cliptest) {
1732 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1733 /* allocate clipmask, assign it integer type */
1734 clipmask = generate_clipmask(llvm,
1735 gallivm,
1736 vs_type,
1737 outputs,
1738 key->clip_xy,
1739 key->clip_z,
1740 key->clip_user,
1741 key->clip_halfz,
1742 key->ucp_enable,
1743 context_ptr, &have_clipdist);
1744 temp = LLVMBuildOr(builder, clipmask, temp, "");
1745 /* store temporary clipping boolean value */
1746 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1747 }
1748 else {
1749 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1750 }
1751
1752 /* do viewport mapping */
1753 if (!bypass_viewport) {
1754 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1755 }
1756 }
1757 else {
1758 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1759 }
1760
1761 /* store clipmask in vertex header,
1762 * original positions in clip
1763 * and transformed positions in data
1764 */
1765 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1766 vs_info->num_outputs, vs_type,
1767 have_clipdist);
1768 }
1769 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1770
1771 sampler->destroy(sampler);
1772
1773 /* return clipping boolean value for function */
1774 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1775
1776 LLVMBuildRet(builder, ret);
1777
1778 gallivm_verify_function(gallivm, variant_func);
1779 }
1780
1781
1782 struct draw_llvm_variant_key *
1783 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1784 {
1785 unsigned i;
1786 struct draw_llvm_variant_key *key;
1787 struct draw_sampler_static_state *draw_sampler;
1788
1789 key = (struct draw_llvm_variant_key *)store;
1790
1791 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1792
1793 /* Presumably all variants of the shader should have the same
1794 * number of vertex elements - ie the number of shader inputs.
1795 * NOTE: we NEED to store the needed number of needed inputs
1796 * here, not the number of provided elements to match keysize
1797 * (and the offset of sampler state in the key).
1798 */
1799 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1800 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1801
1802 /* will have to rig this up properly later */
1803 key->clip_xy = llvm->draw->clip_xy;
1804 key->clip_z = llvm->draw->clip_z;
1805 key->clip_user = llvm->draw->clip_user;
1806 key->bypass_viewport = llvm->draw->identity_viewport;
1807 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1808 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1809 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1810 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1811 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1812 key->pad1 = 0;
1813
1814 /* All variants of this shader will have the same value for
1815 * nr_samplers. Not yet trying to compact away holes in the
1816 * sampler array.
1817 */
1818 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1819 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1820 key->nr_sampler_views =
1821 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1822 }
1823 else {
1824 key->nr_sampler_views = key->nr_samplers;
1825 }
1826
1827 draw_sampler = draw_llvm_variant_key_samplers(key);
1828
1829 memcpy(key->vertex_element,
1830 llvm->draw->pt.vertex_element,
1831 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1832
1833 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1834
1835 for (i = 0 ; i < key->nr_samplers; i++) {
1836 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1837 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1838 }
1839 for (i = 0 ; i < key->nr_sampler_views; i++) {
1840 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1841 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1842 }
1843
1844 return key;
1845 }
1846
1847
1848 void
1849 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1850 {
1851 unsigned i;
1852 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1853
1854 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1855 debug_printf("clip_xy = %u\n", key->clip_xy);
1856 debug_printf("clip_z = %u\n", key->clip_z);
1857 debug_printf("clip_user = %u\n", key->clip_user);
1858 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1859 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1860 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1861 debug_printf("has_gs = %u\n", key->has_gs);
1862 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1863
1864 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1865 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1866 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1867 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1868 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1869 }
1870
1871 for (i = 0 ; i < key->nr_sampler_views; i++) {
1872 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1873 }
1874 }
1875
1876
1877 void
1878 draw_llvm_set_mapped_texture(struct draw_context *draw,
1879 unsigned shader_stage,
1880 unsigned sview_idx,
1881 uint32_t width, uint32_t height, uint32_t depth,
1882 uint32_t first_level, uint32_t last_level,
1883 const void *base_ptr,
1884 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1885 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1886 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1887 {
1888 unsigned j;
1889 struct draw_jit_texture *jit_tex;
1890
1891 assert(shader_stage == PIPE_SHADER_VERTEX ||
1892 shader_stage == PIPE_SHADER_GEOMETRY);
1893
1894 if (shader_stage == PIPE_SHADER_VERTEX) {
1895 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1896
1897 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1898 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1899 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1900
1901 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1902 } else {
1903 assert(0);
1904 return;
1905 }
1906
1907 jit_tex->width = width;
1908 jit_tex->height = height;
1909 jit_tex->depth = depth;
1910 jit_tex->first_level = first_level;
1911 jit_tex->last_level = last_level;
1912 jit_tex->base = base_ptr;
1913
1914 for (j = first_level; j <= last_level; j++) {
1915 jit_tex->mip_offsets[j] = mip_offsets[j];
1916 jit_tex->row_stride[j] = row_stride[j];
1917 jit_tex->img_stride[j] = img_stride[j];
1918 }
1919 }
1920
1921
1922 void
1923 draw_llvm_set_sampler_state(struct draw_context *draw,
1924 unsigned shader_type)
1925 {
1926 unsigned i;
1927
1928 if (shader_type == PIPE_SHADER_VERTEX) {
1929 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1930 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1931
1932 if (draw->samplers[i]) {
1933 const struct pipe_sampler_state *s
1934 = draw->samplers[PIPE_SHADER_VERTEX][i];
1935 jit_sam->min_lod = s->min_lod;
1936 jit_sam->max_lod = s->max_lod;
1937 jit_sam->lod_bias = s->lod_bias;
1938 COPY_4V(jit_sam->border_color, s->border_color.f);
1939 }
1940 }
1941 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1942 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1943 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1944
1945 if (draw->samplers[i]) {
1946 const struct pipe_sampler_state *s
1947 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1948 jit_sam->min_lod = s->min_lod;
1949 jit_sam->max_lod = s->max_lod;
1950 jit_sam->lod_bias = s->lod_bias;
1951 COPY_4V(jit_sam->border_color, s->border_color.f);
1952 }
1953 }
1954 }
1955 }
1956
1957
1958 void
1959 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1960 {
1961 struct draw_llvm *llvm = variant->llvm;
1962
1963 if (variant->function_elts) {
1964 gallivm_free_function(variant->gallivm,
1965 variant->function_elts, variant->jit_func_elts);
1966 }
1967
1968 if (variant->function) {
1969 gallivm_free_function(variant->gallivm,
1970 variant->function, variant->jit_func);
1971 }
1972
1973 gallivm_destroy(variant->gallivm);
1974
1975 remove_from_list(&variant->list_item_local);
1976 variant->shader->variants_cached--;
1977 remove_from_list(&variant->list_item_global);
1978 llvm->nr_variants--;
1979 FREE(variant);
1980 }
1981
1982
1983 /**
1984 * Create LLVM types for various structures.
1985 */
1986 static void
1987 create_gs_jit_types(struct draw_gs_llvm_variant *var)
1988 {
1989 struct gallivm_state *gallivm = var->gallivm;
1990 LLVMTypeRef texture_type, sampler_type, context_type;
1991
1992 texture_type = create_jit_texture_type(gallivm, "texture");
1993 sampler_type = create_jit_sampler_type(gallivm, "sampler");
1994
1995 context_type = create_gs_jit_context_type(gallivm,
1996 var->shader->base.vector_length,
1997 texture_type, sampler_type,
1998 "draw_gs_jit_context");
1999 var->context_ptr_type = LLVMPointerType(context_type, 0);
2000
2001 var->input_array_type = create_gs_jit_input_type(gallivm);
2002 }
2003
2004 static LLVMTypeRef
2005 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2006 {
2007 if (!variant->context_ptr_type)
2008 create_gs_jit_types(variant);
2009 return variant->context_ptr_type;
2010 }
2011
2012 static LLVMValueRef
2013 generate_mask_value(struct draw_gs_llvm_variant *variant,
2014 struct lp_type gs_type)
2015 {
2016 struct gallivm_state *gallivm = variant->gallivm;
2017 LLVMBuilderRef builder = gallivm->builder;
2018 struct lp_type mask_type = lp_int_type(gs_type);
2019 LLVMValueRef num_prims;
2020 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2021 unsigned i;
2022
2023 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2024 variant->num_prims);
2025 for (i = 0; i <= gs_type.length; i++) {
2026 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2027 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2028 }
2029 mask_val = lp_build_compare(gallivm, mask_type,
2030 PIPE_FUNC_GREATER, num_prims, mask_val);
2031
2032 return mask_val;
2033 }
2034
2035 static void
2036 draw_gs_llvm_generate(struct draw_llvm *llvm,
2037 struct draw_gs_llvm_variant *variant)
2038 {
2039 struct gallivm_state *gallivm = variant->gallivm;
2040 LLVMContextRef context = gallivm->context;
2041 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2042 LLVMTypeRef arg_types[6];
2043 LLVMTypeRef func_type;
2044 LLVMValueRef variant_func;
2045 LLVMValueRef context_ptr;
2046 LLVMValueRef prim_id_ptr;
2047 LLVMBasicBlockRef block;
2048 LLVMBuilderRef builder;
2049 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2050 struct lp_build_sampler_soa *sampler = 0;
2051 struct lp_build_context bld;
2052 struct lp_bld_tgsi_system_values system_values;
2053 struct lp_type gs_type;
2054 unsigned i;
2055 struct draw_gs_llvm_iface gs_iface;
2056 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2057 LLVMValueRef consts_ptr, num_consts_ptr;
2058 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2059 struct lp_build_mask_context mask;
2060 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2061 unsigned vector_length = variant->shader->base.vector_length;
2062
2063 memset(&system_values, 0, sizeof(system_values));
2064
2065 assert(variant->vertex_header_ptr_type);
2066
2067 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2068 arg_types[1] = variant->input_array_type; /* input */
2069 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2070 arg_types[3] = int32_type; /* num_prims */
2071 arg_types[4] = int32_type; /* instance_id */
2072 arg_types[5] = LLVMPointerType(
2073 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2074
2075 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2076
2077 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
2078 func_type);
2079 variant->function = variant_func;
2080
2081 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2082
2083 for (i = 0; i < Elements(arg_types); ++i)
2084 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2085 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2086 LLVMNoAliasAttribute);
2087
2088 context_ptr = LLVMGetParam(variant_func, 0);
2089 input_array = LLVMGetParam(variant_func, 1);
2090 io_ptr = LLVMGetParam(variant_func, 2);
2091 num_prims = LLVMGetParam(variant_func, 3);
2092 system_values.instance_id = LLVMGetParam(variant_func, 4);
2093 prim_id_ptr = LLVMGetParam(variant_func, 5);
2094
2095 lp_build_name(context_ptr, "context");
2096 lp_build_name(input_array, "input");
2097 lp_build_name(io_ptr, "io");
2098 lp_build_name(num_prims, "num_prims");
2099 lp_build_name(system_values.instance_id, "instance_id");
2100 lp_build_name(prim_id_ptr, "prim_id_ptr");
2101
2102 variant->context_ptr = context_ptr;
2103 variant->io_ptr = io_ptr;
2104 variant->num_prims = num_prims;
2105
2106 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2107 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2108 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2109 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2110 gs_iface.input = input_array;
2111 gs_iface.variant = variant;
2112
2113 /*
2114 * Function body
2115 */
2116
2117 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2118 builder = gallivm->builder;
2119 LLVMPositionBuilderAtEnd(builder, block);
2120
2121 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2122
2123 memset(&gs_type, 0, sizeof gs_type);
2124 gs_type.floating = TRUE; /* floating point values */
2125 gs_type.sign = TRUE; /* values are signed */
2126 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2127 gs_type.width = 32; /* 32-bit float */
2128 gs_type.length = vector_length;
2129
2130 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2131 num_consts_ptr =
2132 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2133
2134 /* code generated texture sampling */
2135 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2136 context_ptr);
2137
2138 mask_val = generate_mask_value(variant, gs_type);
2139 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2140
2141 if (gs_info->uses_primid) {
2142 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2143 }
2144
2145 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2146 tgsi_dump(tokens, 0);
2147 draw_gs_llvm_dump_variant_key(&variant->key);
2148 }
2149
2150 lp_build_tgsi_soa(variant->gallivm,
2151 tokens,
2152 gs_type,
2153 &mask,
2154 consts_ptr,
2155 num_consts_ptr,
2156 &system_values,
2157 NULL,
2158 outputs,
2159 sampler,
2160 &llvm->draw->gs.geometry_shader->info,
2161 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2162
2163 sampler->destroy(sampler);
2164
2165 lp_build_mask_end(&mask);
2166
2167 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2168
2169 gallivm_verify_function(gallivm, variant_func);
2170 }
2171
2172
2173 struct draw_gs_llvm_variant *
2174 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2175 unsigned num_outputs,
2176 const struct draw_gs_llvm_variant_key *key)
2177 {
2178 struct draw_gs_llvm_variant *variant;
2179 struct llvm_geometry_shader *shader =
2180 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2181 LLVMTypeRef vertex_header;
2182
2183 variant = MALLOC(sizeof *variant +
2184 shader->variant_key_size -
2185 sizeof variant->key);
2186 if (variant == NULL)
2187 return NULL;
2188
2189 variant->llvm = llvm;
2190 variant->shader = shader;
2191
2192 variant->gallivm = gallivm_create();
2193
2194 create_gs_jit_types(variant);
2195
2196 memcpy(&variant->key, key, shader->variant_key_size);
2197
2198 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2199
2200 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2201
2202 draw_gs_llvm_generate(llvm, variant);
2203
2204 gallivm_compile_module(variant->gallivm);
2205
2206 variant->jit_func = (draw_gs_jit_func)
2207 gallivm_jit_function(variant->gallivm, variant->function);
2208
2209 variant->list_item_global.base = variant;
2210 variant->list_item_local.base = variant;
2211 /*variant->no = */shader->variants_created++;
2212 variant->list_item_global.base = variant;
2213
2214 return variant;
2215 }
2216
2217 void
2218 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2219 {
2220 struct draw_llvm *llvm = variant->llvm;
2221
2222 if (variant->function) {
2223 gallivm_free_function(variant->gallivm,
2224 variant->function, variant->jit_func);
2225 }
2226
2227 gallivm_destroy(variant->gallivm);
2228
2229 remove_from_list(&variant->list_item_local);
2230 variant->shader->variants_cached--;
2231 remove_from_list(&variant->list_item_global);
2232 llvm->nr_gs_variants--;
2233 FREE(variant);
2234 }
2235
2236 struct draw_gs_llvm_variant_key *
2237 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2238 {
2239 unsigned i;
2240 struct draw_gs_llvm_variant_key *key;
2241 struct draw_sampler_static_state *draw_sampler;
2242
2243 key = (struct draw_gs_llvm_variant_key *)store;
2244
2245 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2246
2247 /* All variants of this shader will have the same value for
2248 * nr_samplers. Not yet trying to compact away holes in the
2249 * sampler array.
2250 */
2251 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2252 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2253 key->nr_sampler_views =
2254 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2255 }
2256 else {
2257 key->nr_sampler_views = key->nr_samplers;
2258 }
2259
2260 draw_sampler = key->samplers;
2261
2262 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2263
2264 for (i = 0 ; i < key->nr_samplers; i++) {
2265 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2266 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2267 }
2268 for (i = 0 ; i < key->nr_sampler_views; i++) {
2269 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2270 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2271 }
2272
2273 return key;
2274 }
2275
2276 void
2277 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2278 {
2279 unsigned i;
2280 struct draw_sampler_static_state *sampler = key->samplers;
2281
2282 for (i = 0 ; i < key->nr_sampler_views; i++) {
2283 debug_printf("sampler[%i].src_format = %s\n", i,
2284 util_format_name(sampler[i].texture_state.format));
2285 }
2286 }