projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: don't crash on vertex buffer overflow
[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_logic.h"
36 #include "gallivm/lp_bld_const.h"
37 #include "gallivm/lp_bld_swizzle.h"
38 #include "gallivm/lp_bld_struct.h"
39 #include "gallivm/lp_bld_type.h"
40 #include "gallivm/lp_bld_flow.h"
41 #include "gallivm/lp_bld_debug.h"
42 #include "gallivm/lp_bld_tgsi.h"
43 #include "gallivm/lp_bld_printf.h"
44 #include "gallivm/lp_bld_intr.h"
45 #include "gallivm/lp_bld_init.h"
46 #include "gallivm/lp_bld_type.h"
47 #include "gallivm/lp_bld_pack.h"
48 #include "gallivm/lp_bld_format.h"
49
50 #include "tgsi/tgsi_exec.h"
51 #include "tgsi/tgsi_dump.h"
52
53 #include "util/u_math.h"
54 #include "util/u_pointer.h"
55 #include "util/u_string.h"
56 #include "util/u_simple_list.h"
57
58
59 #define DEBUG_STORE 0
60
61
62 static void
63 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
64 boolean elts);
65
66
67 struct draw_gs_llvm_iface {
68 struct lp_build_tgsi_gs_iface base;
69
70 struct draw_gs_llvm_variant *variant;
71 LLVMValueRef input;
72 };
73
74 static INLINE const struct draw_gs_llvm_iface *
75 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
76 {
77 return (const struct draw_gs_llvm_iface *)iface;
78 }
79
80 /**
81 * Create LLVM type for draw_vertex_buffer.
82 */
83 static LLVMTypeRef
84 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
85 const char *struct_name)
86 {
87 LLVMTargetDataRef target = gallivm->target;
88 LLVMTypeRef dvbuffer_type;
89 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
90 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
91
92 elem_types[DRAW_JIT_DVBUFFER_MAP] =
93 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
94 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
95
96 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
97 Elements(elem_types), 0);
98
99 #if HAVE_LLVM < 0x0300
100 LLVMAddTypeName(gallivm->module, struct_name, dvbuffer_type);
101
102 /* Make sure the target's struct layout cache doesn't return
103 * stale/invalid data.
104 */
105 LLVMInvalidateStructLayout(gallivm->target, dvbuffer_type);
106 #endif
107
108 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
109 target, dvbuffer_type,
110 DRAW_JIT_DVBUFFER_MAP);
111 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
112 target, dvbuffer_type,
113 DRAW_JIT_DVBUFFER_SIZE);
114
115 return dvbuffer_type;
116 }
117
118 /**
119 * Create LLVM type for struct draw_jit_texture
120 */
121 static LLVMTypeRef
122 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
123 {
124 LLVMTargetDataRef target = gallivm->target;
125 LLVMTypeRef texture_type;
126 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
127 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
128
129 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
130 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
131 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
132 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
133 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
134 elem_types[DRAW_JIT_TEXTURE_BASE] =
135 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
136 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
137 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
138 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
139 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
140
141 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
142 Elements(elem_types), 0);
143
144 #if HAVE_LLVM < 0x0300
145 LLVMAddTypeName(gallivm->module, struct_name, texture_type);
146
147 /* Make sure the target's struct layout cache doesn't return
148 * stale/invalid data.
149 */
150 LLVMInvalidateStructLayout(gallivm->target, texture_type);
151 #endif
152
153 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
154 target, texture_type,
155 DRAW_JIT_TEXTURE_WIDTH);
156 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
157 target, texture_type,
158 DRAW_JIT_TEXTURE_HEIGHT);
159 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
160 target, texture_type,
161 DRAW_JIT_TEXTURE_DEPTH);
162 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
163 target, texture_type,
164 DRAW_JIT_TEXTURE_FIRST_LEVEL);
165 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
166 target, texture_type,
167 DRAW_JIT_TEXTURE_LAST_LEVEL);
168 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
169 target, texture_type,
170 DRAW_JIT_TEXTURE_BASE);
171 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
172 target, texture_type,
173 DRAW_JIT_TEXTURE_ROW_STRIDE);
174 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
175 target, texture_type,
176 DRAW_JIT_TEXTURE_IMG_STRIDE);
177 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
178 target, texture_type,
179 DRAW_JIT_TEXTURE_MIP_OFFSETS);
180
181 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
182
183 return texture_type;
184 }
185
186
187 /**
188 * Create LLVM type for struct draw_jit_sampler
189 */
190 static LLVMTypeRef
191 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
192 {
193 LLVMTargetDataRef target = gallivm->target;
194 LLVMTypeRef sampler_type;
195 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
196
197 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
198 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
199 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
200 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
201 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
202
203 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
204 Elements(elem_types), 0);
205
206 #if HAVE_LLVM < 0x0300
207 LLVMAddTypeName(gallivm->module, struct_name, sampler_type);
208
209 /* Make sure the target's struct layout cache doesn't return
210 * stale/invalid data.
211 */
212 LLVMInvalidateStructLayout(gallivm->target, sampler_type);
213 #endif
214
215 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
216 target, sampler_type,
217 DRAW_JIT_SAMPLER_MIN_LOD);
218 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
219 target, sampler_type,
220 DRAW_JIT_SAMPLER_MAX_LOD);
221 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
222 target, sampler_type,
223 DRAW_JIT_SAMPLER_LOD_BIAS);
224 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
225 target, sampler_type,
226 DRAW_JIT_SAMPLER_BORDER_COLOR);
227
228 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
229
230 return sampler_type;
231 }
232
233
234 /**
235 * Create LLVM type for struct draw_jit_context
236 */
237 static LLVMTypeRef
238 create_jit_context_type(struct gallivm_state *gallivm,
239 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
240 const char *struct_name)
241 {
242 LLVMTargetDataRef target = gallivm->target;
243 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
244 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
245 LLVMTypeRef context_type;
246
247 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
248 LP_MAX_TGSI_CONST_BUFFERS);
249 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
250 DRAW_TOTAL_CLIP_PLANES), 0);
251 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
252 elem_types[3] = LLVMArrayType(texture_type,
253 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
254 elem_types[4] = LLVMArrayType(sampler_type,
255 PIPE_MAX_SAMPLERS); /* samplers */
256 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
257 Elements(elem_types), 0);
258 #if HAVE_LLVM < 0x0300
259 LLVMAddTypeName(gallivm->module, struct_name, context_type);
260
261 LLVMInvalidateStructLayout(gallivm->target, context_type);
262 #endif
263
264 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
265 target, context_type, DRAW_JIT_CTX_CONSTANTS);
266 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
267 target, context_type, DRAW_JIT_CTX_PLANES);
268 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewport,
269 target, context_type, DRAW_JIT_CTX_VIEWPORT);
270 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
271 target, context_type,
272 DRAW_JIT_CTX_TEXTURES);
273 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
274 target, context_type,
275 DRAW_JIT_CTX_SAMPLERS);
276 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
277 target, context_type);
278
279 return context_type;
280 }
281
282
283 /**
284 * Create LLVM type for struct draw_gs_jit_context
285 */
286 static LLVMTypeRef
287 create_gs_jit_context_type(struct gallivm_state *gallivm,
288 unsigned vector_length,
289 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
290 const char *struct_name)
291 {
292 LLVMTargetDataRef target = gallivm->target;
293 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
294 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
295 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
296 LLVMTypeRef context_type;
297
298 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
299 LP_MAX_TGSI_CONST_BUFFERS);
300 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
301 DRAW_TOTAL_CLIP_PLANES), 0);
302 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
303
304 elem_types[3] = LLVMArrayType(texture_type,
305 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
306 elem_types[4] = LLVMArrayType(sampler_type,
307 PIPE_MAX_SAMPLERS); /* samplers */
308
309 elem_types[5] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
310 elem_types[6] = LLVMPointerType(LLVMVectorType(int_type,
311 vector_length), 0);
312 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
313 vector_length), 0);
314
315 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
316 Elements(elem_types), 0);
317 #if HAVE_LLVM < 0x0300
318 LLVMAddTypeName(gallivm->module, struct_name, context_type);
319
320 LLVMInvalidateStructLayout(gallivm->target, context_type);
321 #endif
322
323 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
324 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
325 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
326 target, context_type, DRAW_GS_JIT_CTX_PLANES);
327 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewport,
328 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
329 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
330 target, context_type,
331 DRAW_GS_JIT_CTX_TEXTURES);
332 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
333 target, context_type,
334 DRAW_GS_JIT_CTX_SAMPLERS);
335 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
336 target, context_type,
337 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
338 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
339 target, context_type,
340 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
341 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
342 target, context_type,
343 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
344 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
345 target, context_type);
346
347 return context_type;
348 }
349
350
351 static LLVMTypeRef
352 create_gs_jit_input_type(struct gallivm_state *gallivm)
353 {
354 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
355 LLVMTypeRef input_array;
356
357 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
358 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
359 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
360 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
361
362 return input_array;
363 }
364
365 /**
366 * Create LLVM type for struct pipe_vertex_buffer
367 */
368 static LLVMTypeRef
369 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
370 const char *struct_name)
371 {
372 LLVMTargetDataRef target = gallivm->target;
373 LLVMTypeRef elem_types[4];
374 LLVMTypeRef vb_type;
375
376 elem_types[0] =
377 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
378 elem_types[2] =
379 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
380
381 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
382 Elements(elem_types), 0);
383 #if HAVE_LLVM < 0x0300
384 LLVMAddTypeName(gallivm->module, struct_name, vb_type);
385
386 LLVMInvalidateStructLayout(gallivm->target, vb_type);
387 #endif
388
389 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
390 target, vb_type, 0);
391 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
392 target, vb_type, 1);
393
394 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
395
396 return vb_type;
397 }
398
399
400 /**
401 * Create LLVM type for struct vertex_header;
402 */
403 static LLVMTypeRef
404 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
405 {
406 LLVMTargetDataRef target = gallivm->target;
407 LLVMTypeRef elem_types[4];
408 LLVMTypeRef vertex_header;
409 char struct_name[24];
410
411 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
412
413 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
414 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
415 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
416 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
417
418 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
419 Elements(elem_types), 0);
420 #if HAVE_LLVM < 0x0300
421 LLVMAddTypeName(gallivm->module, struct_name, vertex_header);
422
423 LLVMInvalidateStructLayout(gallivm->target, vertex_header);
424 #endif
425
426 /* these are bit-fields and we can't take address of them
427 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
428 target, vertex_header,
429 DRAW_JIT_VERTEX_CLIPMASK);
430 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
431 target, vertex_header,
432 DRAW_JIT_VERTEX_EDGEFLAG);
433 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
434 target, vertex_header,
435 DRAW_JIT_VERTEX_PAD);
436 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
437 target, vertex_header,
438 DRAW_JIT_VERTEX_VERTEX_ID);
439 */
440 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
441 target, vertex_header,
442 DRAW_JIT_VERTEX_CLIP);
443 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
444 target, vertex_header,
445 DRAW_JIT_VERTEX_PRE_CLIP_POS);
446 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
447 target, vertex_header,
448 DRAW_JIT_VERTEX_DATA);
449
450 assert(LLVMABISizeOfType(target, vertex_header) ==
451 offsetof(struct vertex_header, data[data_elems]));
452
453 return vertex_header;
454 }
455
456
457 /**
458 * Create LLVM types for various structures.
459 */
460 static void
461 create_jit_types(struct draw_llvm_variant *variant)
462 {
463 struct gallivm_state *gallivm = variant->gallivm;
464 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
465 vb_type;
466
467 texture_type = create_jit_texture_type(gallivm, "texture");
468 sampler_type = create_jit_sampler_type(gallivm, "sampler");
469
470 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
471 "draw_jit_context");
472 variant->context_ptr_type = LLVMPointerType(context_type, 0);
473
474 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
475 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
476
477 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
478 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
479 }
480
481
482 static LLVMTypeRef
483 get_context_ptr_type(struct draw_llvm_variant *variant)
484 {
485 if (!variant->context_ptr_type)
486 create_jit_types(variant);
487 return variant->context_ptr_type;
488 }
489
490
491 static LLVMTypeRef
492 get_buffer_ptr_type(struct draw_llvm_variant *variant)
493 {
494 if (!variant->buffer_ptr_type)
495 create_jit_types(variant);
496 return variant->buffer_ptr_type;
497 }
498
499
500 static LLVMTypeRef
501 get_vb_ptr_type(struct draw_llvm_variant *variant)
502 {
503 if (!variant->vb_ptr_type)
504 create_jit_types(variant);
505 return variant->vb_ptr_type;
506 }
507
508 static LLVMTypeRef
509 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
510 {
511 if (!variant->vertex_header_ptr_type)
512 create_jit_types(variant);
513 return variant->vertex_header_ptr_type;
514 }
515
516
517 /**
518 * Create per-context LLVM info.
519 */
520 struct draw_llvm *
521 draw_llvm_create(struct draw_context *draw)
522 {
523 struct draw_llvm *llvm;
524
525 llvm = CALLOC_STRUCT( draw_llvm );
526 if (!llvm)
527 return NULL;
528
529 lp_build_init();
530
531 llvm->draw = draw;
532
533 llvm->nr_variants = 0;
534 make_empty_list(&llvm->vs_variants_list);
535
536 llvm->nr_gs_variants = 0;
537 make_empty_list(&llvm->gs_variants_list);
538
539 return llvm;
540 }
541
542
543 /**
544 * Free per-context LLVM info.
545 */
546 void
547 draw_llvm_destroy(struct draw_llvm *llvm)
548 {
549 /* XXX free other draw_llvm data? */
550 FREE(llvm);
551 }
552
553
554 /**
555 * Create LLVM-generated code for a vertex shader.
556 */
557 struct draw_llvm_variant *
558 draw_llvm_create_variant(struct draw_llvm *llvm,
559 unsigned num_inputs,
560 const struct draw_llvm_variant_key *key)
561 {
562 struct draw_llvm_variant *variant;
563 struct llvm_vertex_shader *shader =
564 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
565 LLVMTypeRef vertex_header;
566
567 variant = MALLOC(sizeof *variant +
568 shader->variant_key_size -
569 sizeof variant->key);
570 if (variant == NULL)
571 return NULL;
572
573 variant->llvm = llvm;
574
575 variant->gallivm = gallivm_create();
576
577 create_jit_types(variant);
578
579 memcpy(&variant->key, key, shader->variant_key_size);
580
581 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
582
583 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
584
585 draw_llvm_generate(llvm, variant, FALSE); /* linear */
586 draw_llvm_generate(llvm, variant, TRUE); /* elts */
587
588 gallivm_compile_module(variant->gallivm);
589
590 variant->jit_func = (draw_jit_vert_func)
591 gallivm_jit_function(variant->gallivm, variant->function);
592
593 variant->jit_func_elts = (draw_jit_vert_func_elts)
594 gallivm_jit_function(variant->gallivm, variant->function_elts);
595
596 variant->shader = shader;
597 variant->list_item_global.base = variant;
598 variant->list_item_local.base = variant;
599 /*variant->no = */shader->variants_created++;
600 variant->list_item_global.base = variant;
601
602 return variant;
603 }
604
605
606 static void
607 generate_vs(struct draw_llvm_variant *variant,
608 LLVMBuilderRef builder,
609 struct lp_type vs_type,
610 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
611 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
612 const struct lp_bld_tgsi_system_values *system_values,
613 LLVMValueRef context_ptr,
614 struct lp_build_sampler_soa *draw_sampler,
615 boolean clamp_vertex_color)
616 {
617 struct draw_llvm *llvm = variant->llvm;
618 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
619 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(variant->gallivm, context_ptr);
620 struct lp_build_sampler_soa *sampler = 0;
621
622 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
623 tgsi_dump(tokens, 0);
624 draw_llvm_dump_variant_key(&variant->key);
625 }
626
627 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers)
628 sampler = draw_sampler;
629
630 lp_build_tgsi_soa(variant->gallivm,
631 tokens,
632 vs_type,
633 NULL /*struct lp_build_mask_context *mask*/,
634 consts_ptr,
635 system_values,
636 inputs,
637 outputs,
638 sampler,
639 &llvm->draw->vs.vertex_shader->info,
640 NULL);
641
642 {
643 LLVMValueRef out;
644 unsigned chan, attrib;
645 struct lp_build_context bld;
646 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
647 lp_build_context_init(&bld, variant->gallivm, vs_type);
648
649 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
650 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
651 if (outputs[attrib][chan]) {
652 switch (info->output_semantic_name[attrib]) {
653 case TGSI_SEMANTIC_COLOR:
654 case TGSI_SEMANTIC_BCOLOR:
655 if (clamp_vertex_color) {
656 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
657 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
658 LLVMBuildStore(builder, out, outputs[attrib][chan]);
659 }
660 break;
661 case TGSI_SEMANTIC_FOG:
662 if (chan == 1 || chan == 2)
663 LLVMBuildStore(builder, bld.zero, outputs[attrib][chan]);
664 else if (chan == 3)
665 LLVMBuildStore(builder, bld.one, outputs[attrib][chan]);
666 break;
667 }
668 }
669 }
670 }
671 }
672 }
673
674 static void
675 generate_fetch(struct gallivm_state *gallivm,
676 LLVMValueRef vbuffers_ptr,
677 LLVMValueRef *res,
678 struct pipe_vertex_element *velem,
679 LLVMValueRef vbuf,
680 LLVMValueRef index,
681 LLVMValueRef instance_id)
682 {
683 const struct util_format_description *format_desc =
684 util_format_description(velem->src_format);
685 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
686 LLVMBuilderRef builder = gallivm->builder;
687 LLVMValueRef indices =
688 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
689 velem->vertex_buffer_index, 0);
690 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
691 &indices, 1, "");
692 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
693 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
694 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
695 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
696 LLVMValueRef stride;
697 LLVMValueRef buffer_overflowed;
698 LLVMValueRef temp_ptr =
699 lp_build_alloca(gallivm,
700 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
701 struct lp_build_if_state if_ctx;
702
703 if (velem->instance_divisor) {
704 /* array index = instance_id / instance_divisor */
705 index = LLVMBuildUDiv(builder, instance_id,
706 lp_build_const_int32(gallivm, velem->instance_divisor),
707 "instance_divisor");
708 }
709
710 stride = LLVMBuildMul(builder, vb_stride, index, "");
711
712 stride = LLVMBuildAdd(builder, stride,
713 vb_buffer_offset,
714 "");
715 stride = LLVMBuildAdd(builder, stride,
716 lp_build_const_int32(gallivm, velem->src_offset),
717 "");
718
719 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
720 stride, buffer_size,
721 "buffer_overflowed");
722 /*
723 lp_build_printf(gallivm, "vbuf index = %d, stride is %d\n", indices, stride);
724 lp_build_print_value(gallivm, " buffer size = ", buffer_size);
725 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
726 */
727
728 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
729 {
730 LLVMValueRef val =
731 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
732 LLVMBuildStore(builder, val, temp_ptr);
733 }
734 lp_build_else(&if_ctx);
735 {
736 LLVMValueRef val;
737 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
738
739 val = lp_build_fetch_rgba_aos(gallivm,
740 format_desc,
741 lp_float32_vec4_type(),
742 map_ptr,
743 zero, zero, zero);
744 LLVMBuildStore(builder, val, temp_ptr);
745 }
746 lp_build_endif(&if_ctx);
747
748 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
749 }
750
751 static void
752 convert_to_soa(struct gallivm_state *gallivm,
753 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
754 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
755 unsigned num_attribs, const struct lp_type soa_type)
756 {
757 unsigned i, j, k;
758 struct lp_type aos_channel_type = soa_type;
759
760 debug_assert(TGSI_NUM_CHANNELS == 4);
761 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
762
763 aos_channel_type.length >>= 1;
764
765 for (i = 0; i < num_attribs; ++i) {
766 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
767 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
768
769 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
770 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
771
772 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
773
774 for (k = 0; k < pixels_per_channel; ++k) {
775 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
776 }
777
778 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
779 }
780
781 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
782 }
783 }
784
785
786 static void
787 store_aos(struct gallivm_state *gallivm,
788 LLVMValueRef io_ptr,
789 LLVMValueRef index,
790 LLVMValueRef value)
791 {
792 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
793 LLVMBuilderRef builder = gallivm->builder;
794 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
795 LLVMValueRef indices[3];
796
797 indices[0] = lp_build_const_int32(gallivm, 0);
798 indices[1] = index;
799 indices[2] = lp_build_const_int32(gallivm, 0);
800
801 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
802 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
803
804 #if DEBUG_STORE
805 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
806 #endif
807
808 /* Unaligned store due to the vertex header */
809 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
810 }
811
812 /**
813 * Adjust the mask to architecture endianess. The mask will the store in struct:
814 *
815 * struct vertex_header {
816 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
817 * unsigned edgeflag:1;
818 * unsigned have_clipdist:1;
819 * unsigned vertex_id:16;
820 * [...]
821 * }
822 *
823 * On little-endian machine nothing needs to done, however on bit-endian machine
824 * the mask's fields need to be adjusted with the algorithm:
825 *
826 * uint32_t reverse (uint32_t x)
827 * {
828 * return (x >> 16) | // vertex_id
829 * ((x & 0x3fff) << 18) | // clipmask
830 * ((x & 0x4000) << 3) | // have_clipdist
831 * ((x & 0x8000) << 1); // edgeflag
832 * }
833 */
834 static LLVMValueRef
835 adjust_mask(struct gallivm_state *gallivm,
836 LLVMValueRef mask)
837 {
838 #ifdef PIPE_ARCH_BIG_ENDIAN
839 LLVMBuilderRef builder = gallivm->builder;
840 LLVMValueRef vertex_id;
841 LLVMValueRef clipmask;
842 LLVMValueRef have_clipdist;
843 LLVMValueRef edgeflag;
844
845 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
846 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
847 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
848 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
849 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
850 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
851 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
852
853 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
854 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
855 mask = LLVMBuildOr(builder, mask, edgeflag, "");
856 #endif
857 return mask;
858 }
859
860 static void
861 store_aos_array(struct gallivm_state *gallivm,
862 struct lp_type soa_type,
863 LLVMValueRef io_ptr,
864 LLVMValueRef *indices,
865 LLVMValueRef* aos,
866 int attrib,
867 int num_outputs,
868 LLVMValueRef clipmask,
869 boolean have_clipdist)
870 {
871 LLVMBuilderRef builder = gallivm->builder;
872 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
873 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
874 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
875 int vector_length = soa_type.length;
876 int i;
877
878 debug_assert(TGSI_NUM_CHANNELS == 4);
879
880 for (i = 0; i < vector_length; i++) {
881 if (indices) {
882 inds[i] = indices[i];
883 } else {
884 inds[i] = lp_build_const_int32(gallivm, i);
885 }
886 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
887 }
888
889 if (attrib == 0) {
890 /* store vertex header for each of the n vertices */
891 LLVMValueRef val, cliptmp;
892 int vertex_id_pad_edgeflag;
893
894 /* If this assertion fails, it means we need to update the bit twidding
895 * code here. See struct vertex_header in draw_private.h.
896 */
897 assert(DRAW_TOTAL_CLIP_PLANES==14);
898 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
899 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
900 if (have_clipdist)
901 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
902 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
903 /* OR with the clipmask */
904 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
905 for (i = 0; i < vector_length; i++) {
906 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
907 val = LLVMBuildExtractElement(builder, cliptmp, inds[i], "");
908 val = adjust_mask(gallivm, val);
909 LLVMBuildStore(builder, val, id_ptr);
910 #if DEBUG_STORE
911 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
912 io_ptrs[i], inds[i], val);
913 #endif
914 }
915 }
916
917 /* store for each of the n vertices */
918 for (i = 0; i < vector_length; i++) {
919 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
920 }
921 }
922
923
924 static void
925 convert_to_aos(struct gallivm_state *gallivm,
926 LLVMValueRef io,
927 LLVMValueRef *indices,
928 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
929 LLVMValueRef clipmask,
930 int num_outputs,
931 struct lp_type soa_type,
932 boolean have_clipdist)
933 {
934 LLVMBuilderRef builder = gallivm->builder;
935 unsigned chan, attrib, i;
936
937 #if DEBUG_STORE
938 lp_build_printf(gallivm, " # storing begin\n");
939 #endif
940 for (attrib = 0; attrib < num_outputs; ++attrib) {
941 LLVMValueRef soa[TGSI_NUM_CHANNELS];
942 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
943 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
944 if (outputs[attrib][chan]) {
945 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
946 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
947 #if DEBUG_STORE
948 lp_build_printf(gallivm, "output %d : %d ",
949 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
950 attrib, 0),
951 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
952 chan, 0));
953 lp_build_print_value(gallivm, "val = ", out);
954 #endif
955 soa[chan] = out;
956 }
957 else {
958 soa[chan] = 0;
959 }
960 }
961
962
963 if (soa_type.length == TGSI_NUM_CHANNELS) {
964 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
965 } else {
966 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
967
968 for (i = 0; i < soa_type.length; ++i) {
969 aos[i] = lp_build_extract_range(gallivm,
970 soa[i % TGSI_NUM_CHANNELS],
971 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
972 TGSI_NUM_CHANNELS);
973 }
974 }
975
976 store_aos_array(gallivm,
977 soa_type,
978 io, indices,
979 aos,
980 attrib,
981 num_outputs,
982 clipmask, have_clipdist);
983 }
984 #if DEBUG_STORE
985 lp_build_printf(gallivm, " # storing end\n");
986 #endif
987 }
988
989
990 /**
991 * Stores original vertex positions in clip coordinates
992 */
993 static void
994 store_clip(struct gallivm_state *gallivm,
995 const struct lp_type vs_type,
996 LLVMValueRef io_ptr,
997 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
998 boolean pre_clip_pos, int idx)
999 {
1000 LLVMBuilderRef builder = gallivm->builder;
1001 LLVMValueRef soa[4];
1002 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1003 LLVMValueRef indices[2];
1004 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1005 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1006 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1007 int i, j;
1008
1009 indices[0] =
1010 indices[1] = lp_build_const_int32(gallivm, 0);
1011
1012 for (i = 0; i < vs_type.length; i++) {
1013 inds[i] = lp_build_const_int32(gallivm, i);
1014 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1015 }
1016
1017 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1018 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1019 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1020 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1021
1022 if (!pre_clip_pos) {
1023 for (i = 0; i < vs_type.length; i++) {
1024 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1025 }
1026 } else {
1027 for (i = 0; i < vs_type.length; i++) {
1028 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1029 }
1030 }
1031
1032 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1033 for (i = 0; i < vs_type.length; ++i) {
1034 aos[i] = lp_build_extract_range(gallivm,
1035 soa[i % TGSI_NUM_CHANNELS],
1036 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1037 TGSI_NUM_CHANNELS);
1038 }
1039
1040 for (j = 0; j < vs_type.length; j++) {
1041 LLVMTypeRef clip_ptr_type = LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0);
1042 LLVMValueRef clip_ptr;
1043
1044 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1045 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1046
1047 /* Unaligned store */
1048 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1049 }
1050 }
1051
1052
1053 /**
1054 * Transforms the outputs for viewport mapping
1055 */
1056 static void
1057 generate_viewport(struct draw_llvm_variant *variant,
1058 LLVMBuilderRef builder,
1059 struct lp_type vs_type,
1060 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1061 LLVMValueRef context_ptr)
1062 {
1063 int i;
1064 struct gallivm_state *gallivm = variant->gallivm;
1065 struct lp_type f32_type = vs_type;
1066 const unsigned pos = draw_current_shader_position_output(variant->llvm->draw);
1067 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1068 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1069 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1070 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
1071
1072 /* for 1/w convention*/
1073 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1074 LLVMBuildStore(builder, out3, outputs[pos][3]);
1075
1076 /* Viewport Mapping */
1077 for (i=0; i<3; i++) {
1078 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1079 LLVMValueRef scale;
1080 LLVMValueRef trans;
1081 LLVMValueRef scale_i;
1082 LLVMValueRef trans_i;
1083 LLVMValueRef index;
1084
1085 index = lp_build_const_int32(gallivm, i);
1086 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1087
1088 index = lp_build_const_int32(gallivm, i+4);
1089 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1090
1091 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1092 LLVMBuildLoad(builder, scale_i, "scale"));
1093 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1094 LLVMBuildLoad(builder, trans_i, "trans"));
1095
1096 /* divide by w */
1097 out = LLVMBuildFMul(builder, out, out3, "");
1098 /* mult by scale */
1099 out = LLVMBuildFMul(builder, out, scale, "");
1100 /* add translation */
1101 out = LLVMBuildFAdd(builder, out, trans, "");
1102
1103 /* store transformed outputs */
1104 LLVMBuildStore(builder, out, outputs[pos][i]);
1105 }
1106
1107 }
1108
1109
1110 /**
1111 * Returns clipmask as nxi32 bitmask for the n vertices
1112 */
1113 static LLVMValueRef
1114 generate_clipmask(struct draw_llvm *llvm,
1115 struct gallivm_state *gallivm,
1116 struct lp_type vs_type,
1117 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1118 boolean clip_xy,
1119 boolean clip_z,
1120 boolean clip_user,
1121 boolean clip_halfz,
1122 unsigned ucp_enable,
1123 LLVMValueRef context_ptr,
1124 boolean *have_clipdist)
1125 {
1126 LLVMBuilderRef builder = gallivm->builder;
1127 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1128 LLVMValueRef test, temp;
1129 LLVMValueRef zero, shift;
1130 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1131 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1132 LLVMValueRef plane1, planes, plane_ptr, sum;
1133 struct lp_type f32_type = vs_type;
1134 struct lp_type i32_type = lp_int_type(vs_type);
1135 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1136 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1137 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1138 bool have_cd = false;
1139 unsigned cd[2];
1140
1141 cd[0] = draw_current_shader_clipdistance_output(llvm->draw, 0);
1142 cd[1] = draw_current_shader_clipdistance_output(llvm->draw, 1);
1143
1144 if (cd[0] != pos || cd[1] != pos)
1145 have_cd = true;
1146
1147 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1148 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1149 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1150 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1151
1152 /*
1153 * load clipvertex and position from correct locations.
1154 * if they are the same just load them once.
1155 */
1156 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1157 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1158 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1159 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1160
1161 if (clip_user && cv != pos) {
1162 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1163 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1164 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1165 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1166 } else {
1167 cv_x = pos_x;
1168 cv_y = pos_y;
1169 cv_z = pos_z;
1170 cv_w = pos_w;
1171 }
1172
1173 /* Cliptest, for hardwired planes */
1174 if (clip_xy) {
1175 /* plane 1 */
1176 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1177 temp = shift;
1178 test = LLVMBuildAnd(builder, test, temp, "");
1179 mask = test;
1180
1181 /* plane 2 */
1182 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1183 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1184 temp = LLVMBuildShl(builder, temp, shift, "");
1185 test = LLVMBuildAnd(builder, test, temp, "");
1186 mask = LLVMBuildOr(builder, mask, test, "");
1187
1188 /* plane 3 */
1189 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1190 temp = LLVMBuildShl(builder, temp, shift, "");
1191 test = LLVMBuildAnd(builder, test, temp, "");
1192 mask = LLVMBuildOr(builder, mask, test, "");
1193
1194 /* plane 4 */
1195 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1196 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1197 temp = LLVMBuildShl(builder, temp, shift, "");
1198 test = LLVMBuildAnd(builder, test, temp, "");
1199 mask = LLVMBuildOr(builder, mask, test, "");
1200 }
1201
1202 if (clip_z) {
1203 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1204 if (clip_halfz) {
1205 /* plane 5 */
1206 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1207 test = LLVMBuildAnd(builder, test, temp, "");
1208 mask = LLVMBuildOr(builder, mask, test, "");
1209 }
1210 else {
1211 /* plane 5 */
1212 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1213 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1214 test = LLVMBuildAnd(builder, test, temp, "");
1215 mask = LLVMBuildOr(builder, mask, test, "");
1216 }
1217 /* plane 6 */
1218 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1219 temp = LLVMBuildShl(builder, temp, shift, "");
1220 test = LLVMBuildAnd(builder, test, temp, "");
1221 mask = LLVMBuildOr(builder, mask, test, "");
1222 }
1223
1224 if (clip_user) {
1225 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1226 LLVMValueRef indices[3];
1227
1228 /* userclip planes */
1229 while (ucp_enable) {
1230 unsigned plane_idx = ffs(ucp_enable)-1;
1231 ucp_enable &= ~(1 << plane_idx);
1232 plane_idx += 6;
1233
1234 if (have_cd && num_written_clipdistance) {
1235 LLVMValueRef clipdist;
1236 int i;
1237 i = plane_idx - 6;
1238
1239 *have_clipdist = TRUE;
1240 if (i < 4) {
1241 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1242 } else {
1243 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1244 }
1245 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1246 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1247 test = LLVMBuildAnd(builder, test, temp, "");
1248 mask = LLVMBuildOr(builder, mask, test, "");
1249 } else {
1250 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1251 indices[0] = lp_build_const_int32(gallivm, 0);
1252 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1253
1254 indices[2] = lp_build_const_int32(gallivm, 0);
1255 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1256 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1257 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1258 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1259
1260 indices[2] = lp_build_const_int32(gallivm, 1);
1261 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1262 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1263 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1264 test = LLVMBuildFMul(builder, planes, cv_y, "");
1265 sum = LLVMBuildFAdd(builder, sum, test, "");
1266
1267 indices[2] = lp_build_const_int32(gallivm, 2);
1268 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1269 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1270 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1271 test = LLVMBuildFMul(builder, planes, cv_z, "");
1272 sum = LLVMBuildFAdd(builder, sum, test, "");
1273
1274 indices[2] = lp_build_const_int32(gallivm, 3);
1275 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1276 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1277 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1278 test = LLVMBuildFMul(builder, planes, cv_w, "");
1279 sum = LLVMBuildFAdd(builder, sum, test, "");
1280
1281 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1282 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1283 test = LLVMBuildAnd(builder, test, temp, "");
1284 mask = LLVMBuildOr(builder, mask, test, "");
1285 }
1286 }
1287 }
1288 return mask;
1289 }
1290
1291
1292 /**
1293 * Returns boolean if any clipping has occurred
1294 * Used zero/non-zero i32 value to represent boolean
1295 */
1296 static LLVMValueRef
1297 clipmask_booli32(struct gallivm_state *gallivm,
1298 const struct lp_type vs_type,
1299 LLVMValueRef clipmask_bool_ptr)
1300 {
1301 LLVMBuilderRef builder = gallivm->builder;
1302 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1303 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1304 LLVMValueRef ret = LLVMConstNull(int32_type);
1305 LLVMValueRef temp;
1306 int i;
1307
1308 /*
1309 * Can do this with log2(vector length) pack instructions and one extract
1310 * (as we don't actually need a or) with sse2 which would be way better.
1311 */
1312 for (i=0; i < vs_type.length; i++) {
1313 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1314 lp_build_const_int32(gallivm, i) , "");
1315 ret = LLVMBuildOr(builder, ret, temp, "");
1316 }
1317 return ret;
1318 }
1319
1320 static LLVMValueRef
1321 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1322 struct lp_build_tgsi_context * bld_base,
1323 boolean is_indirect,
1324 LLVMValueRef vertex_index,
1325 LLVMValueRef attrib_index,
1326 LLVMValueRef swizzle_index)
1327 {
1328 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1329 struct gallivm_state *gallivm = bld_base->base.gallivm;
1330 LLVMBuilderRef builder = gallivm->builder;
1331 LLVMValueRef indices[3];
1332 LLVMValueRef res;
1333 struct lp_type type = bld_base->base.type;
1334
1335 if (is_indirect) {
1336 int i;
1337 res = bld_base->base.zero;
1338 for (i = 0; i < type.length; ++i) {
1339 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1340 LLVMValueRef vert_chan_index = LLVMBuildExtractElement(builder,
1341 vertex_index, idx, "");
1342 LLVMValueRef channel_vec, value;
1343 indices[0] = vert_chan_index;
1344 indices[1] = attrib_index;
1345 indices[2] = swizzle_index;
1346
1347 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1348 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1349 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1350
1351 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1352 }
1353 } else {
1354 indices[0] = vertex_index;
1355 indices[1] = attrib_index;
1356 indices[2] = swizzle_index;
1357
1358 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1359 res = LLVMBuildLoad(builder, res, "");
1360 }
1361
1362 return res;
1363 }
1364
1365 static void
1366 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1367 struct lp_build_tgsi_context * bld_base,
1368 LLVMValueRef (*outputs)[4],
1369 LLVMValueRef emitted_vertices_vec)
1370 {
1371 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1372 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1373 struct gallivm_state *gallivm = variant->gallivm;
1374 LLVMBuilderRef builder = gallivm->builder;
1375 struct lp_type gs_type = bld_base->base.type;
1376 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1377 lp_int_type(gs_type), 0);
1378 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1379 LLVMValueRef next_prim_offset =
1380 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1381 LLVMValueRef io = variant->io_ptr;
1382 unsigned i;
1383 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1384
1385 for (i = 0; i < gs_type.length; ++i) {
1386 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1387 LLVMValueRef currently_emitted =
1388 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1389 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1390 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1391 }
1392
1393 convert_to_aos(gallivm, io, indices,
1394 outputs, clipmask,
1395 gs_info->num_outputs, gs_type,
1396 FALSE);
1397 }
1398
1399 static void
1400 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1401 struct lp_build_tgsi_context * bld_base,
1402 LLVMValueRef verts_per_prim_vec,
1403 LLVMValueRef emitted_prims_vec)
1404 {
1405 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1406 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1407 struct gallivm_state *gallivm = variant->gallivm;
1408 LLVMBuilderRef builder = gallivm->builder;
1409 LLVMValueRef prim_lengts_ptr =
1410 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1411 unsigned i;
1412
1413 for (i = 0; i < bld_base->base.type.length; ++i) {
1414 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1415 LLVMValueRef prims_emitted =
1416 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1417 LLVMValueRef store_ptr;
1418 LLVMValueRef num_vertices =
1419 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1420
1421 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1422 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1423 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1424 LLVMBuildStore(builder, num_vertices, store_ptr);
1425 }
1426 }
1427
1428 static void
1429 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1430 struct lp_build_tgsi_context * bld_base,
1431 LLVMValueRef total_emitted_vertices_vec,
1432 LLVMValueRef emitted_prims_vec)
1433 {
1434 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1435 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1436 struct gallivm_state *gallivm = variant->gallivm;
1437 LLVMBuilderRef builder = gallivm->builder;
1438 LLVMValueRef emitted_verts_ptr =
1439 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1440 LLVMValueRef emitted_prims_ptr =
1441 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1442 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1443
1444 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1445 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1446
1447 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1448 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1449 }
1450
1451 static void
1452 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1453 boolean elts)
1454 {
1455 struct gallivm_state *gallivm = variant->gallivm;
1456 LLVMContextRef context = gallivm->context;
1457 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1458 LLVMTypeRef arg_types[8];
1459 LLVMTypeRef func_type;
1460 LLVMValueRef context_ptr;
1461 LLVMBasicBlockRef block;
1462 LLVMBuilderRef builder;
1463 struct lp_type vs_type;
1464 LLVMValueRef end, start;
1465 LLVMValueRef count, fetch_elts, fetch_count;
1466 LLVMValueRef stride, step, io_itr;
1467 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1468 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1469 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1470 struct draw_context *draw = llvm->draw;
1471 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1472 unsigned i, j;
1473 struct lp_build_context bld;
1474 struct lp_build_loop_state lp_loop;
1475 const int vector_length = lp_native_vector_width / 32;
1476 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1477 LLVMValueRef fetch_max;
1478 struct lp_build_sampler_soa *sampler = 0;
1479 LLVMValueRef ret, clipmask_bool_ptr;
1480 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1481 struct draw_llvm_variant_key *key = &variant->key;
1482 /* If geometry shader is present we need to skip both the viewport
1483 * transformation and clipping otherwise the inputs to the geometry
1484 * shader will be incorrect.
1485 */
1486 const boolean bypass_viewport = gs || key->bypass_viewport;
1487 const boolean enable_cliptest = !gs && (key->clip_xy ||
1488 key->clip_z ||
1489 key->clip_user);
1490 LLVMValueRef variant_func;
1491 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1492 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1493 boolean have_clipdist = FALSE;
1494 struct lp_bld_tgsi_system_values system_values;
1495
1496 memset(&system_values, 0, sizeof(system_values));
1497
1498 arg_types[0] = get_context_ptr_type(variant); /* context */
1499 arg_types[1] = get_vertex_header_ptr_type(variant); /* vertex_header */
1500 arg_types[2] = get_buffer_ptr_type(variant); /* vbuffers */
1501 if (elts)
1502 arg_types[3] = LLVMPointerType(int32_type, 0);/* fetch_elts * */
1503 else
1504 arg_types[3] = int32_type; /* start */
1505 arg_types[4] = int32_type; /* fetch_count / count */
1506 arg_types[5] = int32_type; /* stride */
1507 arg_types[6] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1508 arg_types[7] = int32_type; /* instance_id */
1509
1510 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
1511
1512 variant_func = LLVMAddFunction(gallivm->module,
1513 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1514 func_type);
1515
1516 if (elts)
1517 variant->function_elts = variant_func;
1518 else
1519 variant->function = variant_func;
1520
1521 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1522 for (i = 0; i < Elements(arg_types); ++i)
1523 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1524 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1525 LLVMNoAliasAttribute);
1526
1527 context_ptr = LLVMGetParam(variant_func, 0);
1528 io_ptr = LLVMGetParam(variant_func, 1);
1529 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1530 stride = LLVMGetParam(variant_func, 5);
1531 vb_ptr = LLVMGetParam(variant_func, 6);
1532 system_values.instance_id = LLVMGetParam(variant_func, 7);
1533
1534 lp_build_name(context_ptr, "context");
1535 lp_build_name(io_ptr, "io");
1536 lp_build_name(vbuffers_ptr, "vbuffers");
1537 lp_build_name(stride, "stride");
1538 lp_build_name(vb_ptr, "vb");
1539 lp_build_name(system_values.instance_id, "instance_id");
1540
1541 if (elts) {
1542 fetch_elts = LLVMGetParam(variant_func, 3);
1543 fetch_count = LLVMGetParam(variant_func, 4);
1544 lp_build_name(fetch_elts, "fetch_elts");
1545 lp_build_name(fetch_count, "fetch_count");
1546 start = count = NULL;
1547 }
1548 else {
1549 start = LLVMGetParam(variant_func, 3);
1550 count = LLVMGetParam(variant_func, 4);
1551 lp_build_name(start, "start");
1552 lp_build_name(count, "count");
1553 fetch_elts = fetch_count = NULL;
1554 }
1555
1556 /*
1557 * Function body
1558 */
1559
1560 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1561 builder = gallivm->builder;
1562 LLVMPositionBuilderAtEnd(builder, block);
1563
1564 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1565
1566 memset(&vs_type, 0, sizeof vs_type);
1567 vs_type.floating = TRUE; /* floating point values */
1568 vs_type.sign = TRUE; /* values are signed */
1569 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1570 vs_type.width = 32; /* 32-bit float */
1571 vs_type.length = vector_length;
1572
1573 /* hold temporary "bool" clipmask */
1574 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1575 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1576
1577 /* code generated texture sampling */
1578 sampler = draw_llvm_sampler_soa_create(
1579 draw_llvm_variant_key_samplers(key),
1580 context_ptr);
1581
1582 if (elts) {
1583 start = zero;
1584 end = fetch_count;
1585 }
1586 else {
1587 end = lp_build_add(&bld, start, count);
1588 }
1589
1590 step = lp_build_const_int32(gallivm, vector_length);
1591
1592 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1593
1594 lp_build_loop_begin(&lp_loop, gallivm, start);
1595 {
1596 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1597 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1598 LLVMValueRef io;
1599 LLVMValueRef clipmask; /* holds the clipmask value */
1600 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1601
1602 if (elts)
1603 io_itr = lp_loop.counter;
1604 else
1605 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
1606
1607 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1608 #if DEBUG_STORE
1609 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1610 io_itr, io, lp_loop.counter);
1611 #endif
1612 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1613 for (i = 0; i < vector_length; ++i) {
1614 LLVMValueRef true_index =
1615 LLVMBuildAdd(builder,
1616 lp_loop.counter,
1617 lp_build_const_int32(gallivm, i), "");
1618
1619 /* make sure we're not out of bounds which can happen
1620 * if fetch_count % 4 != 0, because on the last iteration
1621 * a few of the 4 vertex fetches will be out of bounds */
1622 true_index = lp_build_min(&bld, true_index, fetch_max);
1623
1624 if (elts) {
1625 LLVMValueRef fetch_ptr;
1626 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1627 &true_index, 1, "");
1628 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
1629 }
1630
1631 system_values.vertex_id = LLVMBuildInsertElement(gallivm->builder,
1632 system_values.vertex_id, true_index,
1633 lp_build_const_int32(gallivm, i), "");
1634 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1635 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1636 LLVMValueRef vb_index =
1637 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1638 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1639 generate_fetch(gallivm, vbuffers_ptr,
1640 &aos_attribs[j][i], velem, vb, true_index,
1641 system_values.instance_id);
1642 }
1643 }
1644 convert_to_soa(gallivm, aos_attribs, inputs,
1645 draw->pt.nr_vertex_elements, vs_type);
1646
1647 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1648 generate_vs(variant,
1649 builder,
1650 vs_type,
1651 outputs,
1652 ptr_aos,
1653 &system_values,
1654 context_ptr,
1655 sampler,
1656 key->clamp_vertex_color);
1657
1658 if (pos != -1 && cv != -1) {
1659 /* store original positions in clip before further manipulation */
1660 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1661 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1662
1663 /* do cliptest */
1664 if (enable_cliptest) {
1665 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1666 /* allocate clipmask, assign it integer type */
1667 clipmask = generate_clipmask(llvm,
1668 gallivm,
1669 vs_type,
1670 outputs,
1671 key->clip_xy,
1672 key->clip_z,
1673 key->clip_user,
1674 key->clip_halfz,
1675 key->ucp_enable,
1676 context_ptr, &have_clipdist);
1677 temp = LLVMBuildOr(builder, clipmask, temp, "");
1678 /* store temporary clipping boolean value */
1679 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1680 }
1681 else {
1682 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1683 }
1684
1685 /* do viewport mapping */
1686 if (!bypass_viewport) {
1687 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1688 }
1689 }
1690 else {
1691 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1692 }
1693
1694 /* store clipmask in vertex header,
1695 * original positions in clip
1696 * and transformed positions in data
1697 */
1698 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1699 vs_info->num_outputs, vs_type,
1700 have_clipdist);
1701 }
1702
1703 lp_build_loop_end_cond(&lp_loop, end, step, LLVMIntUGE);
1704
1705 sampler->destroy(sampler);
1706
1707 /* return clipping boolean value for function */
1708 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1709
1710 LLVMBuildRet(builder, ret);
1711
1712 gallivm_verify_function(gallivm, variant_func);
1713 }
1714
1715
1716 struct draw_llvm_variant_key *
1717 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1718 {
1719 unsigned i;
1720 struct draw_llvm_variant_key *key;
1721 struct draw_sampler_static_state *draw_sampler;
1722
1723 key = (struct draw_llvm_variant_key *)store;
1724
1725 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1726
1727 /* Presumably all variants of the shader should have the same
1728 * number of vertex elements - ie the number of shader inputs.
1729 * NOTE: we NEED to store the needed number of needed inputs
1730 * here, not the number of provided elements to match keysize
1731 * (and the offset of sampler state in the key).
1732 */
1733 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1734 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1735
1736 /* will have to rig this up properly later */
1737 key->clip_xy = llvm->draw->clip_xy;
1738 key->clip_z = llvm->draw->clip_z;
1739 key->clip_user = llvm->draw->clip_user;
1740 key->bypass_viewport = llvm->draw->identity_viewport;
1741 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1742 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1743 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1744 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1745 key->pad1 = 0;
1746
1747 /* All variants of this shader will have the same value for
1748 * nr_samplers. Not yet trying to compact away holes in the
1749 * sampler array.
1750 */
1751 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1752 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1753 key->nr_sampler_views =
1754 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1755 }
1756 else {
1757 key->nr_sampler_views = key->nr_samplers;
1758 }
1759
1760 draw_sampler = draw_llvm_variant_key_samplers(key);
1761
1762 memcpy(key->vertex_element,
1763 llvm->draw->pt.vertex_element,
1764 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1765
1766 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1767
1768 for (i = 0 ; i < key->nr_samplers; i++) {
1769 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1770 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1771 }
1772 for (i = 0 ; i < key->nr_sampler_views; i++) {
1773 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1774 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1775 }
1776
1777 return key;
1778 }
1779
1780
1781 void
1782 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1783 {
1784 unsigned i;
1785 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1786
1787 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1788 debug_printf("clip_xy = %u\n", key->clip_xy);
1789 debug_printf("clip_z = %u\n", key->clip_z);
1790 debug_printf("clip_user = %u\n", key->clip_user);
1791 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1792 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1793 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1794 debug_printf("has_gs = %u\n", key->has_gs);
1795 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1796
1797 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1798 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1799 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1800 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1801 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1802 }
1803
1804 for (i = 0 ; i < key->nr_sampler_views; i++) {
1805 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1806 }
1807 }
1808
1809
1810 void
1811 draw_llvm_set_mapped_texture(struct draw_context *draw,
1812 unsigned shader_stage,
1813 unsigned sview_idx,
1814 uint32_t width, uint32_t height, uint32_t depth,
1815 uint32_t first_level, uint32_t last_level,
1816 const void *base_ptr,
1817 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1818 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1819 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1820 {
1821 unsigned j;
1822 struct draw_jit_texture *jit_tex;
1823
1824 assert(shader_stage == PIPE_SHADER_VERTEX ||
1825 shader_stage == PIPE_SHADER_GEOMETRY);
1826
1827 if (shader_stage == PIPE_SHADER_VERTEX) {
1828 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1829
1830 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1831 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1832 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1833
1834 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1835 } else {
1836 assert(0);
1837 return;
1838 }
1839
1840 jit_tex->width = width;
1841 jit_tex->height = height;
1842 jit_tex->depth = depth;
1843 jit_tex->first_level = first_level;
1844 jit_tex->last_level = last_level;
1845 jit_tex->base = base_ptr;
1846
1847 for (j = first_level; j <= last_level; j++) {
1848 jit_tex->mip_offsets[j] = mip_offsets[j];
1849 jit_tex->row_stride[j] = row_stride[j];
1850 jit_tex->img_stride[j] = img_stride[j];
1851 }
1852 }
1853
1854
1855 void
1856 draw_llvm_set_sampler_state(struct draw_context *draw,
1857 unsigned shader_type)
1858 {
1859 unsigned i;
1860
1861 if (shader_type == PIPE_SHADER_VERTEX) {
1862 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1863 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1864
1865 if (draw->samplers[i]) {
1866 const struct pipe_sampler_state *s
1867 = draw->samplers[PIPE_SHADER_VERTEX][i];
1868 jit_sam->min_lod = s->min_lod;
1869 jit_sam->max_lod = s->max_lod;
1870 jit_sam->lod_bias = s->lod_bias;
1871 COPY_4V(jit_sam->border_color, s->border_color.f);
1872 }
1873 }
1874 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1875 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1876 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1877
1878 if (draw->samplers[i]) {
1879 const struct pipe_sampler_state *s
1880 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1881 jit_sam->min_lod = s->min_lod;
1882 jit_sam->max_lod = s->max_lod;
1883 jit_sam->lod_bias = s->lod_bias;
1884 COPY_4V(jit_sam->border_color, s->border_color.f);
1885 }
1886 }
1887 }
1888 }
1889
1890
1891 void
1892 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1893 {
1894 struct draw_llvm *llvm = variant->llvm;
1895
1896 if (variant->function_elts) {
1897 gallivm_free_function(variant->gallivm,
1898 variant->function_elts, variant->jit_func_elts);
1899 }
1900
1901 if (variant->function) {
1902 gallivm_free_function(variant->gallivm,
1903 variant->function, variant->jit_func);
1904 }
1905
1906 gallivm_destroy(variant->gallivm);
1907
1908 remove_from_list(&variant->list_item_local);
1909 variant->shader->variants_cached--;
1910 remove_from_list(&variant->list_item_global);
1911 llvm->nr_variants--;
1912 FREE(variant);
1913 }
1914
1915
1916 /**
1917 * Create LLVM types for various structures.
1918 */
1919 static void
1920 create_gs_jit_types(struct draw_gs_llvm_variant *var)
1921 {
1922 struct gallivm_state *gallivm = var->gallivm;
1923 LLVMTypeRef texture_type, sampler_type, context_type;
1924
1925 texture_type = create_jit_texture_type(gallivm, "texture");
1926 sampler_type = create_jit_sampler_type(gallivm, "sampler");
1927
1928 context_type = create_gs_jit_context_type(gallivm,
1929 var->shader->base.vector_length,
1930 texture_type, sampler_type,
1931 "draw_gs_jit_context");
1932 var->context_ptr_type = LLVMPointerType(context_type, 0);
1933
1934 var->input_array_type = create_gs_jit_input_type(gallivm);
1935 }
1936
1937 static LLVMTypeRef
1938 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
1939 {
1940 if (!variant->context_ptr_type)
1941 create_gs_jit_types(variant);
1942 return variant->context_ptr_type;
1943 }
1944
1945 static LLVMValueRef
1946 generate_mask_value(struct draw_gs_llvm_variant *variant,
1947 struct lp_type gs_type)
1948 {
1949 struct gallivm_state *gallivm = variant->gallivm;
1950 LLVMBuilderRef builder = gallivm->builder;
1951 LLVMValueRef bits[16];
1952 struct lp_type mask_type = lp_int_type(gs_type);
1953 struct lp_type mask_elem_type = lp_elem_type(mask_type);
1954 LLVMValueRef mask_val = lp_build_const_vec(gallivm,
1955 mask_type,
1956 0);
1957 unsigned i;
1958
1959 assert(gs_type.length <= Elements(bits));
1960
1961 for (i = gs_type.length; i >= 1; --i) {
1962 int idx = i - 1;
1963 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1964 bits[idx] = lp_build_compare(gallivm,
1965 mask_elem_type, PIPE_FUNC_GEQUAL,
1966 variant->num_prims, ind);
1967 }
1968 for (i = 0; i < gs_type.length; ++i) {
1969 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1970 mask_val = LLVMBuildInsertElement(builder, mask_val, bits[i], ind, "");
1971 }
1972 mask_val = lp_build_compare(gallivm,
1973 mask_type, PIPE_FUNC_NOTEQUAL,
1974 mask_val,
1975 lp_build_const_int_vec(gallivm, mask_type, 0));
1976
1977 return mask_val;
1978 }
1979
1980 static void
1981 draw_gs_llvm_generate(struct draw_llvm *llvm,
1982 struct draw_gs_llvm_variant *variant)
1983 {
1984 struct gallivm_state *gallivm = variant->gallivm;
1985 LLVMContextRef context = gallivm->context;
1986 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1987 LLVMTypeRef arg_types[6];
1988 LLVMTypeRef func_type;
1989 LLVMValueRef variant_func;
1990 LLVMValueRef context_ptr;
1991 LLVMValueRef prim_id_ptr;
1992 LLVMBasicBlockRef block;
1993 LLVMBuilderRef builder;
1994 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
1995 struct lp_build_sampler_soa *sampler = 0;
1996 struct lp_build_context bld;
1997 struct lp_bld_tgsi_system_values system_values;
1998 struct lp_type gs_type;
1999 unsigned i;
2000 struct draw_gs_llvm_iface gs_iface;
2001 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2002 LLVMValueRef consts_ptr;
2003 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2004 struct lp_build_mask_context mask;
2005 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2006 unsigned vector_length = variant->shader->base.vector_length;
2007
2008 memset(&system_values, 0, sizeof(system_values));
2009
2010 assert(variant->vertex_header_ptr_type);
2011
2012 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2013 arg_types[1] = variant->input_array_type; /* input */
2014 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2015 arg_types[3] = int32_type; /* num_prims */
2016 arg_types[4] = int32_type; /* instance_id */
2017 arg_types[5] = LLVMPointerType(
2018 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2019
2020 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2021
2022 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
2023 func_type);
2024 variant->function = variant_func;
2025
2026 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2027
2028 for (i = 0; i < Elements(arg_types); ++i)
2029 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2030 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2031 LLVMNoAliasAttribute);
2032
2033 context_ptr = LLVMGetParam(variant_func, 0);
2034 input_array = LLVMGetParam(variant_func, 1);
2035 io_ptr = LLVMGetParam(variant_func, 2);
2036 num_prims = LLVMGetParam(variant_func, 3);
2037 system_values.instance_id = LLVMGetParam(variant_func, 4);
2038 prim_id_ptr = LLVMGetParam(variant_func, 5);
2039
2040 lp_build_name(context_ptr, "context");
2041 lp_build_name(input_array, "input");
2042 lp_build_name(io_ptr, "io");
2043 lp_build_name(num_prims, "num_prims");
2044 lp_build_name(system_values.instance_id, "instance_id");
2045 lp_build_name(prim_id_ptr, "prim_id_ptr");
2046
2047 variant->context_ptr = context_ptr;
2048 variant->io_ptr = io_ptr;
2049 variant->num_prims = num_prims;
2050
2051 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2052 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2053 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2054 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2055 gs_iface.input = input_array;
2056 gs_iface.variant = variant;
2057
2058 /*
2059 * Function body
2060 */
2061
2062 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2063 builder = gallivm->builder;
2064 LLVMPositionBuilderAtEnd(builder, block);
2065
2066 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2067
2068 memset(&gs_type, 0, sizeof gs_type);
2069 gs_type.floating = TRUE; /* floating point values */
2070 gs_type.sign = TRUE; /* values are signed */
2071 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2072 gs_type.width = 32; /* 32-bit float */
2073 gs_type.length = vector_length;
2074
2075 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2076
2077 /* code generated texture sampling */
2078 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2079 context_ptr);
2080
2081 mask_val = generate_mask_value(variant, gs_type);
2082 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2083
2084 if (gs_info->uses_primid) {
2085 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");;
2086 }
2087
2088 lp_build_tgsi_soa(variant->gallivm,
2089 tokens,
2090 gs_type,
2091 &mask,
2092 consts_ptr,
2093 &system_values,
2094 NULL,
2095 outputs,
2096 sampler,
2097 &llvm->draw->gs.geometry_shader->info,
2098 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2099
2100 sampler->destroy(sampler);
2101
2102 lp_build_mask_end(&mask);
2103
2104 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2105
2106 gallivm_verify_function(gallivm, variant_func);
2107 }
2108
2109
2110 struct draw_gs_llvm_variant *
2111 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2112 unsigned num_outputs,
2113 const struct draw_gs_llvm_variant_key *key)
2114 {
2115 struct draw_gs_llvm_variant *variant;
2116 struct llvm_geometry_shader *shader =
2117 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2118 LLVMTypeRef vertex_header;
2119
2120 variant = MALLOC(sizeof *variant +
2121 shader->variant_key_size -
2122 sizeof variant->key);
2123 if (variant == NULL)
2124 return NULL;
2125
2126 variant->llvm = llvm;
2127 variant->shader = shader;
2128
2129 variant->gallivm = gallivm_create();
2130
2131 create_gs_jit_types(variant);
2132
2133 memcpy(&variant->key, key, shader->variant_key_size);
2134
2135 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2136
2137 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2138
2139 draw_gs_llvm_generate(llvm, variant);
2140
2141 gallivm_compile_module(variant->gallivm);
2142
2143 variant->jit_func = (draw_gs_jit_func)
2144 gallivm_jit_function(variant->gallivm, variant->function);
2145
2146 variant->list_item_global.base = variant;
2147 variant->list_item_local.base = variant;
2148 /*variant->no = */shader->variants_created++;
2149 variant->list_item_global.base = variant;
2150
2151 return variant;
2152 }
2153
2154 void
2155 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2156 {
2157 struct draw_llvm *llvm = variant->llvm;
2158
2159 if (variant->function) {
2160 gallivm_free_function(variant->gallivm,
2161 variant->function, variant->jit_func);
2162 }
2163
2164 gallivm_destroy(variant->gallivm);
2165
2166 remove_from_list(&variant->list_item_local);
2167 variant->shader->variants_cached--;
2168 remove_from_list(&variant->list_item_global);
2169 llvm->nr_gs_variants--;
2170 FREE(variant);
2171 }
2172
2173 struct draw_gs_llvm_variant_key *
2174 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2175 {
2176 unsigned i;
2177 struct draw_gs_llvm_variant_key *key;
2178 struct draw_sampler_static_state *draw_sampler;
2179
2180 key = (struct draw_gs_llvm_variant_key *)store;
2181
2182 /* All variants of this shader will have the same value for
2183 * nr_samplers. Not yet trying to compact away holes in the
2184 * sampler array.
2185 */
2186 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2187 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2188 key->nr_sampler_views =
2189 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2190 }
2191 else {
2192 key->nr_sampler_views = key->nr_samplers;
2193 }
2194
2195 draw_sampler = key->samplers;
2196
2197 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2198
2199 for (i = 0 ; i < key->nr_samplers; i++) {
2200 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2201 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2202 }
2203 for (i = 0 ; i < key->nr_sampler_views; i++) {
2204 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2205 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2206 }
2207
2208 return key;
2209 }
2210
2211 void
2212 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2213 {
2214 unsigned i;
2215 struct draw_sampler_static_state *sampler = key->samplers;
2216
2217 for (i = 0 ; i < key->nr_sampler_views; i++) {
2218 debug_printf("sampler[%i].src_format = %s\n", i,
2219 util_format_name(sampler[i].texture_state.format));
2220 }
2221 }