projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: check for integer overflows in instance computation
[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 needed_buffer_size;
699 LLVMValueRef temp_ptr =
700 lp_build_alloca(gallivm,
701 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
702 LLVMValueRef ofbit, oresult;
703 LLVMTypeRef oelems[2] = {
704 LLVMInt32TypeInContext(gallivm->context),
705 LLVMInt1TypeInContext(gallivm->context)
706 };
707 LLVMTypeRef otype = LLVMStructTypeInContext(gallivm->context,
708 oelems, 2, FALSE);
709 struct lp_build_if_state if_ctx;
710
711 if (velem->instance_divisor) {
712 /* array index = instance_id / instance_divisor */
713 index = LLVMBuildUDiv(builder, instance_id,
714 lp_build_const_int32(gallivm, velem->instance_divisor),
715 "instance_divisor");
716 }
717
718 oresult = lp_build_intrinsic_binary(builder,
719 "llvm.umul.with.overflow.i32",
720 otype, vb_stride, index);
721 ofbit = LLVMBuildExtractValue(builder, oresult, 1, "");
722 stride = LLVMBuildExtractValue(builder, oresult, 0, "");
723
724 oresult = lp_build_intrinsic_binary(builder,
725 "llvm.uadd.with.overflow.i32",
726 otype, stride, vb_buffer_offset);
727 ofbit = LLVMBuildOr(
728 builder, ofbit,
729 LLVMBuildExtractValue(builder, oresult, 1, ""),
730 "");
731 stride = LLVMBuildExtractValue(builder, oresult, 0, "");
732
733 oresult = lp_build_intrinsic_binary(
734 builder,
735 "llvm.uadd.with.overflow.i32",
736 otype, stride,
737 lp_build_const_int32(gallivm, velem->src_offset));
738 ofbit = LLVMBuildOr(
739 builder, ofbit,
740 LLVMBuildExtractValue(builder, oresult, 1, ""),
741 "");
742 stride = LLVMBuildExtractValue(builder, oresult, 0, "");
743
744
745 oresult = lp_build_intrinsic_binary(
746 builder,
747 "llvm.uadd.with.overflow.i32",
748 otype, stride,
749 lp_build_const_int32(gallivm,
750 util_format_get_blocksize(velem->src_format)));
751 ofbit = LLVMBuildOr(
752 builder, ofbit,
753 LLVMBuildExtractValue(builder, oresult, 1, ""),
754 "");
755 needed_buffer_size = LLVMBuildExtractValue(builder, oresult, 0, "");
756
757 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
758 needed_buffer_size, buffer_size,
759 "buffer_overflowed");
760 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
761 #if 0
762 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
763 index, vb_stride);
764 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
765 vb_buffer_offset,
766 lp_build_const_int32(gallivm, velem->src_offset));
767 lp_build_print_value(gallivm, " blocksize = ",
768 lp_build_const_int32(
769 gallivm,
770 util_format_get_blocksize(velem->src_format)));
771 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
772 lp_build_printf(gallivm, " stride = %u\n", stride);
773 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
774 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
775 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
776 #endif
777
778 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
779 {
780 LLVMValueRef val =
781 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
782 LLVMBuildStore(builder, val, temp_ptr);
783 }
784 lp_build_else(&if_ctx);
785 {
786 LLVMValueRef val;
787 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
788
789 val = lp_build_fetch_rgba_aos(gallivm,
790 format_desc,
791 lp_float32_vec4_type(),
792 map_ptr,
793 zero, zero, zero);
794 LLVMBuildStore(builder, val, temp_ptr);
795 }
796 lp_build_endif(&if_ctx);
797
798 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
799 }
800
801 static void
802 convert_to_soa(struct gallivm_state *gallivm,
803 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
804 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
805 unsigned num_attribs, const struct lp_type soa_type)
806 {
807 unsigned i, j, k;
808 struct lp_type aos_channel_type = soa_type;
809
810 debug_assert(TGSI_NUM_CHANNELS == 4);
811 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
812
813 aos_channel_type.length >>= 1;
814
815 for (i = 0; i < num_attribs; ++i) {
816 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
817 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
818
819 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
820 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
821
822 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
823
824 for (k = 0; k < pixels_per_channel; ++k) {
825 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
826 }
827
828 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
829 }
830
831 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
832 }
833 }
834
835
836 static void
837 store_aos(struct gallivm_state *gallivm,
838 LLVMValueRef io_ptr,
839 LLVMValueRef index,
840 LLVMValueRef value)
841 {
842 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
843 LLVMBuilderRef builder = gallivm->builder;
844 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
845 LLVMValueRef indices[3];
846
847 indices[0] = lp_build_const_int32(gallivm, 0);
848 indices[1] = index;
849 indices[2] = lp_build_const_int32(gallivm, 0);
850
851 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
852 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
853
854 #if DEBUG_STORE
855 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
856 #endif
857
858 /* Unaligned store due to the vertex header */
859 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
860 }
861
862 /**
863 * Adjust the mask to architecture endianess. The mask will the store in struct:
864 *
865 * struct vertex_header {
866 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
867 * unsigned edgeflag:1;
868 * unsigned have_clipdist:1;
869 * unsigned vertex_id:16;
870 * [...]
871 * }
872 *
873 * On little-endian machine nothing needs to done, however on bit-endian machine
874 * the mask's fields need to be adjusted with the algorithm:
875 *
876 * uint32_t reverse (uint32_t x)
877 * {
878 * return (x >> 16) | // vertex_id
879 * ((x & 0x3fff) << 18) | // clipmask
880 * ((x & 0x4000) << 3) | // have_clipdist
881 * ((x & 0x8000) << 1); // edgeflag
882 * }
883 */
884 static LLVMValueRef
885 adjust_mask(struct gallivm_state *gallivm,
886 LLVMValueRef mask)
887 {
888 #ifdef PIPE_ARCH_BIG_ENDIAN
889 LLVMBuilderRef builder = gallivm->builder;
890 LLVMValueRef vertex_id;
891 LLVMValueRef clipmask;
892 LLVMValueRef have_clipdist;
893 LLVMValueRef edgeflag;
894
895 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
896 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
897 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
898 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
899 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
900 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
901 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
902
903 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
904 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
905 mask = LLVMBuildOr(builder, mask, edgeflag, "");
906 #endif
907 return mask;
908 }
909
910 static void
911 store_aos_array(struct gallivm_state *gallivm,
912 struct lp_type soa_type,
913 LLVMValueRef io_ptr,
914 LLVMValueRef *indices,
915 LLVMValueRef* aos,
916 int attrib,
917 int num_outputs,
918 LLVMValueRef clipmask,
919 boolean have_clipdist)
920 {
921 LLVMBuilderRef builder = gallivm->builder;
922 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
923 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
924 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
925 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
926 int vector_length = soa_type.length;
927 int i;
928
929 debug_assert(TGSI_NUM_CHANNELS == 4);
930
931 for (i = 0; i < vector_length; i++) {
932 linear_inds[i] = lp_build_const_int32(gallivm, i);
933 if (indices) {
934 inds[i] = indices[i];
935 } else {
936 inds[i] = linear_inds[i];
937 }
938 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
939 }
940
941 if (attrib == 0) {
942 /* store vertex header for each of the n vertices */
943 LLVMValueRef val, cliptmp;
944 int vertex_id_pad_edgeflag;
945
946 /* If this assertion fails, it means we need to update the bit twidding
947 * code here. See struct vertex_header in draw_private.h.
948 */
949 assert(DRAW_TOTAL_CLIP_PLANES==14);
950 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
951 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
952 if (have_clipdist)
953 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
954 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
955 /* OR with the clipmask */
956 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
957 for (i = 0; i < vector_length; i++) {
958 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
959 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
960 val = adjust_mask(gallivm, val);
961 LLVMBuildStore(builder, val, id_ptr);
962 #if DEBUG_STORE
963 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
964 io_ptrs[i], inds[i], val);
965 #endif
966 }
967 }
968
969 /* store for each of the n vertices */
970 for (i = 0; i < vector_length; i++) {
971 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
972 }
973 }
974
975
976 static void
977 convert_to_aos(struct gallivm_state *gallivm,
978 LLVMValueRef io,
979 LLVMValueRef *indices,
980 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
981 LLVMValueRef clipmask,
982 int num_outputs,
983 struct lp_type soa_type,
984 boolean have_clipdist)
985 {
986 LLVMBuilderRef builder = gallivm->builder;
987 unsigned chan, attrib, i;
988
989 #if DEBUG_STORE
990 lp_build_printf(gallivm, " # storing begin\n");
991 #endif
992 for (attrib = 0; attrib < num_outputs; ++attrib) {
993 LLVMValueRef soa[TGSI_NUM_CHANNELS];
994 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
995 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
996 if (outputs[attrib][chan]) {
997 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
998 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
999 #if DEBUG_STORE
1000 lp_build_printf(gallivm, "output %d : %d ",
1001 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1002 attrib, 0),
1003 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
1004 chan, 0));
1005 lp_build_print_value(gallivm, "val = ", out);
1006 #endif
1007 soa[chan] = out;
1008 }
1009 else {
1010 soa[chan] = 0;
1011 }
1012 }
1013
1014
1015 if (soa_type.length == TGSI_NUM_CHANNELS) {
1016 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1017 } else {
1018 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1019
1020 for (i = 0; i < soa_type.length; ++i) {
1021 aos[i] = lp_build_extract_range(gallivm,
1022 soa[i % TGSI_NUM_CHANNELS],
1023 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1024 TGSI_NUM_CHANNELS);
1025 }
1026 }
1027
1028 store_aos_array(gallivm,
1029 soa_type,
1030 io, indices,
1031 aos,
1032 attrib,
1033 num_outputs,
1034 clipmask, have_clipdist);
1035 }
1036 #if DEBUG_STORE
1037 lp_build_printf(gallivm, " # storing end\n");
1038 #endif
1039 }
1040
1041
1042 /**
1043 * Stores original vertex positions in clip coordinates
1044 */
1045 static void
1046 store_clip(struct gallivm_state *gallivm,
1047 const struct lp_type vs_type,
1048 LLVMValueRef io_ptr,
1049 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1050 boolean pre_clip_pos, int idx)
1051 {
1052 LLVMBuilderRef builder = gallivm->builder;
1053 LLVMValueRef soa[4];
1054 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1055 LLVMValueRef indices[2];
1056 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1057 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1058 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1059 int i, j;
1060
1061 indices[0] =
1062 indices[1] = lp_build_const_int32(gallivm, 0);
1063
1064 for (i = 0; i < vs_type.length; i++) {
1065 inds[i] = lp_build_const_int32(gallivm, i);
1066 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1067 }
1068
1069 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1070 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1071 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1072 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1073
1074 if (!pre_clip_pos) {
1075 for (i = 0; i < vs_type.length; i++) {
1076 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1077 }
1078 } else {
1079 for (i = 0; i < vs_type.length; i++) {
1080 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1081 }
1082 }
1083
1084 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1085 for (i = 0; i < vs_type.length; ++i) {
1086 aos[i] = lp_build_extract_range(gallivm,
1087 soa[i % TGSI_NUM_CHANNELS],
1088 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1089 TGSI_NUM_CHANNELS);
1090 }
1091
1092 for (j = 0; j < vs_type.length; j++) {
1093 LLVMTypeRef clip_ptr_type = LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0);
1094 LLVMValueRef clip_ptr;
1095
1096 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1097 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1098
1099 /* Unaligned store */
1100 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1101 }
1102 }
1103
1104
1105 /**
1106 * Transforms the outputs for viewport mapping
1107 */
1108 static void
1109 generate_viewport(struct draw_llvm_variant *variant,
1110 LLVMBuilderRef builder,
1111 struct lp_type vs_type,
1112 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1113 LLVMValueRef context_ptr)
1114 {
1115 int i;
1116 struct gallivm_state *gallivm = variant->gallivm;
1117 struct lp_type f32_type = vs_type;
1118 const unsigned pos = draw_current_shader_position_output(variant->llvm->draw);
1119 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1120 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1121 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1122 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
1123
1124 /* for 1/w convention*/
1125 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1126 LLVMBuildStore(builder, out3, outputs[pos][3]);
1127
1128 /* Viewport Mapping */
1129 for (i=0; i<3; i++) {
1130 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1131 LLVMValueRef scale;
1132 LLVMValueRef trans;
1133 LLVMValueRef scale_i;
1134 LLVMValueRef trans_i;
1135 LLVMValueRef index;
1136
1137 index = lp_build_const_int32(gallivm, i);
1138 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1139
1140 index = lp_build_const_int32(gallivm, i+4);
1141 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1142
1143 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1144 LLVMBuildLoad(builder, scale_i, "scale"));
1145 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1146 LLVMBuildLoad(builder, trans_i, "trans"));
1147
1148 /* divide by w */
1149 out = LLVMBuildFMul(builder, out, out3, "");
1150 /* mult by scale */
1151 out = LLVMBuildFMul(builder, out, scale, "");
1152 /* add translation */
1153 out = LLVMBuildFAdd(builder, out, trans, "");
1154
1155 /* store transformed outputs */
1156 LLVMBuildStore(builder, out, outputs[pos][i]);
1157 }
1158
1159 }
1160
1161
1162 /**
1163 * Returns clipmask as nxi32 bitmask for the n vertices
1164 */
1165 static LLVMValueRef
1166 generate_clipmask(struct draw_llvm *llvm,
1167 struct gallivm_state *gallivm,
1168 struct lp_type vs_type,
1169 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1170 boolean clip_xy,
1171 boolean clip_z,
1172 boolean clip_user,
1173 boolean clip_halfz,
1174 unsigned ucp_enable,
1175 LLVMValueRef context_ptr,
1176 boolean *have_clipdist)
1177 {
1178 LLVMBuilderRef builder = gallivm->builder;
1179 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1180 LLVMValueRef test, temp;
1181 LLVMValueRef zero, shift;
1182 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1183 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1184 LLVMValueRef plane1, planes, plane_ptr, sum;
1185 struct lp_type f32_type = vs_type;
1186 struct lp_type i32_type = lp_int_type(vs_type);
1187 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1188 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1189 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1190 bool have_cd = false;
1191 unsigned cd[2];
1192
1193 cd[0] = draw_current_shader_clipdistance_output(llvm->draw, 0);
1194 cd[1] = draw_current_shader_clipdistance_output(llvm->draw, 1);
1195
1196 if (cd[0] != pos || cd[1] != pos)
1197 have_cd = true;
1198
1199 if (num_written_clipdistance && !clip_user) {
1200 clip_user = true;
1201 ucp_enable = (1 << num_written_clipdistance) - 1;
1202 }
1203
1204 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1205 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1206 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1207 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1208
1209 /*
1210 * load clipvertex and position from correct locations.
1211 * if they are the same just load them once.
1212 */
1213 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1214 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1215 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1216 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1217
1218 if (clip_user && cv != pos) {
1219 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1220 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1221 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1222 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1223 } else {
1224 cv_x = pos_x;
1225 cv_y = pos_y;
1226 cv_z = pos_z;
1227 cv_w = pos_w;
1228 }
1229
1230 /* Cliptest, for hardwired planes */
1231 if (clip_xy) {
1232 /* plane 1 */
1233 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1234 temp = shift;
1235 test = LLVMBuildAnd(builder, test, temp, "");
1236 mask = test;
1237
1238 /* plane 2 */
1239 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1240 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1241 temp = LLVMBuildShl(builder, temp, shift, "");
1242 test = LLVMBuildAnd(builder, test, temp, "");
1243 mask = LLVMBuildOr(builder, mask, test, "");
1244
1245 /* plane 3 */
1246 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1247 temp = LLVMBuildShl(builder, temp, shift, "");
1248 test = LLVMBuildAnd(builder, test, temp, "");
1249 mask = LLVMBuildOr(builder, mask, test, "");
1250
1251 /* plane 4 */
1252 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1253 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1254 temp = LLVMBuildShl(builder, temp, shift, "");
1255 test = LLVMBuildAnd(builder, test, temp, "");
1256 mask = LLVMBuildOr(builder, mask, test, "");
1257 }
1258
1259 if (clip_z) {
1260 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1261 if (clip_halfz) {
1262 /* plane 5 */
1263 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1264 test = LLVMBuildAnd(builder, test, temp, "");
1265 mask = LLVMBuildOr(builder, mask, test, "");
1266 }
1267 else {
1268 /* plane 5 */
1269 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1270 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1271 test = LLVMBuildAnd(builder, test, temp, "");
1272 mask = LLVMBuildOr(builder, mask, test, "");
1273 }
1274 /* plane 6 */
1275 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1276 temp = LLVMBuildShl(builder, temp, shift, "");
1277 test = LLVMBuildAnd(builder, test, temp, "");
1278 mask = LLVMBuildOr(builder, mask, test, "");
1279 }
1280
1281 if (clip_user) {
1282 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1283 LLVMValueRef indices[3];
1284
1285 /* userclip planes */
1286 while (ucp_enable) {
1287 unsigned plane_idx = ffs(ucp_enable)-1;
1288 ucp_enable &= ~(1 << plane_idx);
1289 plane_idx += 6;
1290
1291 if (have_cd && num_written_clipdistance) {
1292 LLVMValueRef clipdist;
1293 int i;
1294 i = plane_idx - 6;
1295
1296 *have_clipdist = TRUE;
1297 if (i < 4) {
1298 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1299 } else {
1300 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1301 }
1302 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1303 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1304 test = LLVMBuildAnd(builder, test, temp, "");
1305 mask = LLVMBuildOr(builder, mask, test, "");
1306 } else {
1307 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1308 indices[0] = lp_build_const_int32(gallivm, 0);
1309 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1310
1311 indices[2] = lp_build_const_int32(gallivm, 0);
1312 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1313 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1314 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1315 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1316
1317 indices[2] = lp_build_const_int32(gallivm, 1);
1318 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1319 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1320 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1321 test = LLVMBuildFMul(builder, planes, cv_y, "");
1322 sum = LLVMBuildFAdd(builder, sum, test, "");
1323
1324 indices[2] = lp_build_const_int32(gallivm, 2);
1325 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1326 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1327 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1328 test = LLVMBuildFMul(builder, planes, cv_z, "");
1329 sum = LLVMBuildFAdd(builder, sum, test, "");
1330
1331 indices[2] = lp_build_const_int32(gallivm, 3);
1332 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1333 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1334 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1335 test = LLVMBuildFMul(builder, planes, cv_w, "");
1336 sum = LLVMBuildFAdd(builder, sum, test, "");
1337
1338 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1339 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1340 test = LLVMBuildAnd(builder, test, temp, "");
1341 mask = LLVMBuildOr(builder, mask, test, "");
1342 }
1343 }
1344 }
1345 return mask;
1346 }
1347
1348
1349 /**
1350 * Returns boolean if any clipping has occurred
1351 * Used zero/non-zero i32 value to represent boolean
1352 */
1353 static LLVMValueRef
1354 clipmask_booli32(struct gallivm_state *gallivm,
1355 const struct lp_type vs_type,
1356 LLVMValueRef clipmask_bool_ptr)
1357 {
1358 LLVMBuilderRef builder = gallivm->builder;
1359 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1360 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1361 LLVMValueRef ret = LLVMConstNull(int32_type);
1362 LLVMValueRef temp;
1363 int i;
1364
1365 /*
1366 * Can do this with log2(vector length) pack instructions and one extract
1367 * (as we don't actually need a or) with sse2 which would be way better.
1368 */
1369 for (i=0; i < vs_type.length; i++) {
1370 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1371 lp_build_const_int32(gallivm, i) , "");
1372 ret = LLVMBuildOr(builder, ret, temp, "");
1373 }
1374 return ret;
1375 }
1376
1377 static LLVMValueRef
1378 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1379 struct lp_build_tgsi_context * bld_base,
1380 boolean is_indirect,
1381 LLVMValueRef vertex_index,
1382 LLVMValueRef attrib_index,
1383 LLVMValueRef swizzle_index)
1384 {
1385 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1386 struct gallivm_state *gallivm = bld_base->base.gallivm;
1387 LLVMBuilderRef builder = gallivm->builder;
1388 LLVMValueRef indices[3];
1389 LLVMValueRef res;
1390 struct lp_type type = bld_base->base.type;
1391
1392 if (is_indirect) {
1393 int i;
1394 res = bld_base->base.zero;
1395 for (i = 0; i < type.length; ++i) {
1396 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1397 LLVMValueRef vert_chan_index = LLVMBuildExtractElement(builder,
1398 vertex_index, idx, "");
1399 LLVMValueRef channel_vec, value;
1400 indices[0] = vert_chan_index;
1401 indices[1] = attrib_index;
1402 indices[2] = swizzle_index;
1403
1404 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1405 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1406 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1407
1408 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1409 }
1410 } else {
1411 indices[0] = vertex_index;
1412 indices[1] = attrib_index;
1413 indices[2] = swizzle_index;
1414
1415 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1416 res = LLVMBuildLoad(builder, res, "");
1417 }
1418
1419 return res;
1420 }
1421
1422 static void
1423 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1424 struct lp_build_tgsi_context * bld_base,
1425 LLVMValueRef (*outputs)[4],
1426 LLVMValueRef emitted_vertices_vec)
1427 {
1428 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1429 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1430 struct gallivm_state *gallivm = variant->gallivm;
1431 LLVMBuilderRef builder = gallivm->builder;
1432 struct lp_type gs_type = bld_base->base.type;
1433 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1434 lp_int_type(gs_type), 0);
1435 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1436 LLVMValueRef next_prim_offset =
1437 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1438 LLVMValueRef io = variant->io_ptr;
1439 unsigned i;
1440 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1441
1442 for (i = 0; i < gs_type.length; ++i) {
1443 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1444 LLVMValueRef currently_emitted =
1445 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1446 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1447 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1448 }
1449
1450 convert_to_aos(gallivm, io, indices,
1451 outputs, clipmask,
1452 gs_info->num_outputs, gs_type,
1453 FALSE);
1454 }
1455
1456 static void
1457 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1458 struct lp_build_tgsi_context * bld_base,
1459 LLVMValueRef verts_per_prim_vec,
1460 LLVMValueRef emitted_prims_vec)
1461 {
1462 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1463 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1464 struct gallivm_state *gallivm = variant->gallivm;
1465 LLVMBuilderRef builder = gallivm->builder;
1466 LLVMValueRef prim_lengts_ptr =
1467 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1468 unsigned i;
1469
1470 for (i = 0; i < bld_base->base.type.length; ++i) {
1471 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1472 LLVMValueRef prims_emitted =
1473 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1474 LLVMValueRef store_ptr;
1475 LLVMValueRef num_vertices =
1476 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1477
1478 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1479 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1480 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1481 LLVMBuildStore(builder, num_vertices, store_ptr);
1482 }
1483 }
1484
1485 static void
1486 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1487 struct lp_build_tgsi_context * bld_base,
1488 LLVMValueRef total_emitted_vertices_vec,
1489 LLVMValueRef emitted_prims_vec)
1490 {
1491 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1492 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1493 struct gallivm_state *gallivm = variant->gallivm;
1494 LLVMBuilderRef builder = gallivm->builder;
1495 LLVMValueRef emitted_verts_ptr =
1496 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1497 LLVMValueRef emitted_prims_ptr =
1498 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1499 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1500
1501 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1502 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1503
1504 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1505 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1506 }
1507
1508 static void
1509 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1510 boolean elts)
1511 {
1512 struct gallivm_state *gallivm = variant->gallivm;
1513 LLVMContextRef context = gallivm->context;
1514 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1515 LLVMTypeRef arg_types[9];
1516 unsigned num_arg_types =
1517 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1518 LLVMTypeRef func_type;
1519 LLVMValueRef context_ptr;
1520 LLVMBasicBlockRef block;
1521 LLVMBuilderRef builder;
1522 struct lp_type vs_type;
1523 LLVMValueRef end, start;
1524 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1525 LLVMValueRef stride, step, io_itr;
1526 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1527 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1528 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1529 struct draw_context *draw = llvm->draw;
1530 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1531 unsigned i, j;
1532 struct lp_build_context bld;
1533 struct lp_build_loop_state lp_loop;
1534 const int vector_length = lp_native_vector_width / 32;
1535 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1536 LLVMValueRef fetch_max;
1537 struct lp_build_sampler_soa *sampler = 0;
1538 LLVMValueRef ret, clipmask_bool_ptr;
1539 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1540 struct draw_llvm_variant_key *key = &variant->key;
1541 /* If geometry shader is present we need to skip both the viewport
1542 * transformation and clipping otherwise the inputs to the geometry
1543 * shader will be incorrect.
1544 */
1545 const boolean bypass_viewport = gs || key->bypass_viewport;
1546 const boolean enable_cliptest = !gs && (key->clip_xy ||
1547 key->clip_z ||
1548 key->clip_user);
1549 LLVMValueRef variant_func;
1550 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1551 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1552 boolean have_clipdist = FALSE;
1553 struct lp_bld_tgsi_system_values system_values;
1554
1555 memset(&system_values, 0, sizeof(system_values));
1556
1557 i = 0;
1558 arg_types[i++] = get_context_ptr_type(variant); /* context */
1559 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1560 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1561 if (elts) {
1562 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1563 arg_types[i++] = int32_type; /* fetch_elt_max */
1564 } else
1565 arg_types[i++] = int32_type; /* start */
1566 arg_types[i++] = int32_type; /* fetch_count / count */
1567 arg_types[i++] = int32_type; /* stride */
1568 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1569 arg_types[i++] = int32_type; /* instance_id */
1570
1571 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1572
1573 variant_func = LLVMAddFunction(gallivm->module,
1574 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1575 func_type);
1576
1577 if (elts)
1578 variant->function_elts = variant_func;
1579 else
1580 variant->function = variant_func;
1581
1582 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1583 for (i = 0; i < num_arg_types; ++i)
1584 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1585 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1586 LLVMNoAliasAttribute);
1587
1588 context_ptr = LLVMGetParam(variant_func, 0);
1589 io_ptr = LLVMGetParam(variant_func, 1);
1590 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1591 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1592 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1593 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1594
1595 lp_build_name(context_ptr, "context");
1596 lp_build_name(io_ptr, "io");
1597 lp_build_name(vbuffers_ptr, "vbuffers");
1598 lp_build_name(stride, "stride");
1599 lp_build_name(vb_ptr, "vb");
1600 lp_build_name(system_values.instance_id, "instance_id");
1601
1602 if (elts) {
1603 fetch_elts = LLVMGetParam(variant_func, 3);
1604 fetch_elt_max = LLVMGetParam(variant_func, 4);
1605 fetch_count = LLVMGetParam(variant_func, 5);
1606 lp_build_name(fetch_elts, "fetch_elts");
1607 lp_build_name(fetch_elt_max, "fetch_elt_max");
1608 lp_build_name(fetch_count, "fetch_count");
1609 start = count = NULL;
1610 }
1611 else {
1612 start = LLVMGetParam(variant_func, 3);
1613 count = LLVMGetParam(variant_func, 4);
1614 lp_build_name(start, "start");
1615 lp_build_name(count, "count");
1616 fetch_elts = fetch_count = NULL;
1617 }
1618
1619 /*
1620 * Function body
1621 */
1622
1623 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1624 builder = gallivm->builder;
1625 LLVMPositionBuilderAtEnd(builder, block);
1626
1627 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1628
1629 memset(&vs_type, 0, sizeof vs_type);
1630 vs_type.floating = TRUE; /* floating point values */
1631 vs_type.sign = TRUE; /* values are signed */
1632 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1633 vs_type.width = 32; /* 32-bit float */
1634 vs_type.length = vector_length;
1635
1636 /* hold temporary "bool" clipmask */
1637 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1638 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1639
1640 /* code generated texture sampling */
1641 sampler = draw_llvm_sampler_soa_create(
1642 draw_llvm_variant_key_samplers(key),
1643 context_ptr);
1644
1645 if (elts) {
1646 start = zero;
1647 end = fetch_count;
1648 count = fetch_count;
1649 }
1650 else {
1651 end = lp_build_add(&bld, start, count);
1652 }
1653
1654 step = lp_build_const_int32(gallivm, vector_length);
1655
1656 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1657
1658 lp_build_loop_begin(&lp_loop, gallivm, zero);
1659 {
1660 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1661 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1662 LLVMValueRef io;
1663 LLVMValueRef clipmask; /* holds the clipmask value */
1664 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1665
1666 io_itr = lp_loop.counter;
1667
1668 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1669 #if DEBUG_STORE
1670 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1671 io_itr, io, lp_loop.counter);
1672 #endif
1673 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1674 for (i = 0; i < vector_length; ++i) {
1675 LLVMValueRef true_index =
1676 LLVMBuildAdd(builder,
1677 lp_loop.counter,
1678 lp_build_const_int32(gallivm, i), "");
1679 true_index = LLVMBuildAdd(builder, start, true_index, "");
1680
1681 /* make sure we're not out of bounds which can happen
1682 * if fetch_count % 4 != 0, because on the last iteration
1683 * a few of the 4 vertex fetches will be out of bounds */
1684 true_index = lp_build_min(&bld, true_index, fetch_max);
1685
1686 system_values.vertex_id = LLVMBuildInsertElement(
1687 gallivm->builder,
1688 system_values.vertex_id, true_index,
1689 lp_build_const_int32(gallivm, i), "");
1690
1691 if (elts) {
1692 LLVMValueRef fetch_ptr;
1693 LLVMValueRef index_overflowed;
1694 LLVMValueRef index_ptr =
1695 lp_build_alloca(
1696 gallivm,
1697 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1698 struct lp_build_if_state if_ctx;
1699 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
1700 true_index, fetch_elt_max,
1701 "index_overflowed");
1702
1703 lp_build_if(&if_ctx, gallivm, index_overflowed);
1704 {
1705 /* Generate maximum possible index so that
1706 * generate_fetch can treat it just like
1707 * any other overflow and return zeros.
1708 * We don't have to worry about the restart
1709 * primitive index because it has already been
1710 * handled
1711 */
1712 LLVMValueRef val =
1713 lp_build_const_int32(gallivm, 0xffffffff);
1714 LLVMBuildStore(builder, val, index_ptr);
1715 }
1716 lp_build_else(&if_ctx);
1717 {
1718 LLVMValueRef val;
1719 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1720 &true_index, 1, "");
1721 val = LLVMBuildLoad(builder, fetch_ptr, "");
1722 LLVMBuildStore(builder, val, index_ptr);
1723 }
1724 lp_build_endif(&if_ctx);
1725 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1726 }
1727
1728 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1729 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1730 LLVMValueRef vb_index =
1731 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1732 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1733 generate_fetch(gallivm, vbuffers_ptr,
1734 &aos_attribs[j][i], velem, vb, true_index,
1735 system_values.instance_id);
1736 }
1737 }
1738 convert_to_soa(gallivm, aos_attribs, inputs,
1739 draw->pt.nr_vertex_elements, vs_type);
1740
1741 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1742 generate_vs(variant,
1743 builder,
1744 vs_type,
1745 outputs,
1746 ptr_aos,
1747 &system_values,
1748 context_ptr,
1749 sampler,
1750 key->clamp_vertex_color);
1751
1752 if (pos != -1 && cv != -1) {
1753 /* store original positions in clip before further manipulation */
1754 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1755 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1756
1757 /* do cliptest */
1758 if (enable_cliptest) {
1759 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1760 /* allocate clipmask, assign it integer type */
1761 clipmask = generate_clipmask(llvm,
1762 gallivm,
1763 vs_type,
1764 outputs,
1765 key->clip_xy,
1766 key->clip_z,
1767 key->clip_user,
1768 key->clip_halfz,
1769 key->ucp_enable,
1770 context_ptr, &have_clipdist);
1771 temp = LLVMBuildOr(builder, clipmask, temp, "");
1772 /* store temporary clipping boolean value */
1773 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1774 }
1775 else {
1776 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1777 }
1778
1779 /* do viewport mapping */
1780 if (!bypass_viewport) {
1781 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1782 }
1783 }
1784 else {
1785 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1786 }
1787
1788 /* store clipmask in vertex header,
1789 * original positions in clip
1790 * and transformed positions in data
1791 */
1792 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1793 vs_info->num_outputs, vs_type,
1794 have_clipdist);
1795 }
1796 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1797
1798 sampler->destroy(sampler);
1799
1800 /* return clipping boolean value for function */
1801 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1802
1803 LLVMBuildRet(builder, ret);
1804
1805 gallivm_verify_function(gallivm, variant_func);
1806 }
1807
1808
1809 struct draw_llvm_variant_key *
1810 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1811 {
1812 unsigned i;
1813 struct draw_llvm_variant_key *key;
1814 struct draw_sampler_static_state *draw_sampler;
1815
1816 key = (struct draw_llvm_variant_key *)store;
1817
1818 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1819
1820 /* Presumably all variants of the shader should have the same
1821 * number of vertex elements - ie the number of shader inputs.
1822 * NOTE: we NEED to store the needed number of needed inputs
1823 * here, not the number of provided elements to match keysize
1824 * (and the offset of sampler state in the key).
1825 */
1826 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1827 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1828
1829 /* will have to rig this up properly later */
1830 key->clip_xy = llvm->draw->clip_xy;
1831 key->clip_z = llvm->draw->clip_z;
1832 key->clip_user = llvm->draw->clip_user;
1833 key->bypass_viewport = llvm->draw->identity_viewport;
1834 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1835 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1836 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1837 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1838 key->pad1 = 0;
1839
1840 /* All variants of this shader will have the same value for
1841 * nr_samplers. Not yet trying to compact away holes in the
1842 * sampler array.
1843 */
1844 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1845 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1846 key->nr_sampler_views =
1847 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1848 }
1849 else {
1850 key->nr_sampler_views = key->nr_samplers;
1851 }
1852
1853 draw_sampler = draw_llvm_variant_key_samplers(key);
1854
1855 memcpy(key->vertex_element,
1856 llvm->draw->pt.vertex_element,
1857 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1858
1859 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1860
1861 for (i = 0 ; i < key->nr_samplers; i++) {
1862 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1863 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1864 }
1865 for (i = 0 ; i < key->nr_sampler_views; i++) {
1866 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1867 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1868 }
1869
1870 return key;
1871 }
1872
1873
1874 void
1875 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1876 {
1877 unsigned i;
1878 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1879
1880 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1881 debug_printf("clip_xy = %u\n", key->clip_xy);
1882 debug_printf("clip_z = %u\n", key->clip_z);
1883 debug_printf("clip_user = %u\n", key->clip_user);
1884 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1885 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1886 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1887 debug_printf("has_gs = %u\n", key->has_gs);
1888 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1889
1890 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1891 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1892 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1893 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1894 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1895 }
1896
1897 for (i = 0 ; i < key->nr_sampler_views; i++) {
1898 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1899 }
1900 }
1901
1902
1903 void
1904 draw_llvm_set_mapped_texture(struct draw_context *draw,
1905 unsigned shader_stage,
1906 unsigned sview_idx,
1907 uint32_t width, uint32_t height, uint32_t depth,
1908 uint32_t first_level, uint32_t last_level,
1909 const void *base_ptr,
1910 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1911 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1912 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1913 {
1914 unsigned j;
1915 struct draw_jit_texture *jit_tex;
1916
1917 assert(shader_stage == PIPE_SHADER_VERTEX ||
1918 shader_stage == PIPE_SHADER_GEOMETRY);
1919
1920 if (shader_stage == PIPE_SHADER_VERTEX) {
1921 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1922
1923 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1924 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1925 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1926
1927 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1928 } else {
1929 assert(0);
1930 return;
1931 }
1932
1933 jit_tex->width = width;
1934 jit_tex->height = height;
1935 jit_tex->depth = depth;
1936 jit_tex->first_level = first_level;
1937 jit_tex->last_level = last_level;
1938 jit_tex->base = base_ptr;
1939
1940 for (j = first_level; j <= last_level; j++) {
1941 jit_tex->mip_offsets[j] = mip_offsets[j];
1942 jit_tex->row_stride[j] = row_stride[j];
1943 jit_tex->img_stride[j] = img_stride[j];
1944 }
1945 }
1946
1947
1948 void
1949 draw_llvm_set_sampler_state(struct draw_context *draw,
1950 unsigned shader_type)
1951 {
1952 unsigned i;
1953
1954 if (shader_type == PIPE_SHADER_VERTEX) {
1955 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1956 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1957
1958 if (draw->samplers[i]) {
1959 const struct pipe_sampler_state *s
1960 = draw->samplers[PIPE_SHADER_VERTEX][i];
1961 jit_sam->min_lod = s->min_lod;
1962 jit_sam->max_lod = s->max_lod;
1963 jit_sam->lod_bias = s->lod_bias;
1964 COPY_4V(jit_sam->border_color, s->border_color.f);
1965 }
1966 }
1967 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1968 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1969 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1970
1971 if (draw->samplers[i]) {
1972 const struct pipe_sampler_state *s
1973 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1974 jit_sam->min_lod = s->min_lod;
1975 jit_sam->max_lod = s->max_lod;
1976 jit_sam->lod_bias = s->lod_bias;
1977 COPY_4V(jit_sam->border_color, s->border_color.f);
1978 }
1979 }
1980 }
1981 }
1982
1983
1984 void
1985 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1986 {
1987 struct draw_llvm *llvm = variant->llvm;
1988
1989 if (variant->function_elts) {
1990 gallivm_free_function(variant->gallivm,
1991 variant->function_elts, variant->jit_func_elts);
1992 }
1993
1994 if (variant->function) {
1995 gallivm_free_function(variant->gallivm,
1996 variant->function, variant->jit_func);
1997 }
1998
1999 gallivm_destroy(variant->gallivm);
2000
2001 remove_from_list(&variant->list_item_local);
2002 variant->shader->variants_cached--;
2003 remove_from_list(&variant->list_item_global);
2004 llvm->nr_variants--;
2005 FREE(variant);
2006 }
2007
2008
2009 /**
2010 * Create LLVM types for various structures.
2011 */
2012 static void
2013 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2014 {
2015 struct gallivm_state *gallivm = var->gallivm;
2016 LLVMTypeRef texture_type, sampler_type, context_type;
2017
2018 texture_type = create_jit_texture_type(gallivm, "texture");
2019 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2020
2021 context_type = create_gs_jit_context_type(gallivm,
2022 var->shader->base.vector_length,
2023 texture_type, sampler_type,
2024 "draw_gs_jit_context");
2025 var->context_ptr_type = LLVMPointerType(context_type, 0);
2026
2027 var->input_array_type = create_gs_jit_input_type(gallivm);
2028 }
2029
2030 static LLVMTypeRef
2031 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2032 {
2033 if (!variant->context_ptr_type)
2034 create_gs_jit_types(variant);
2035 return variant->context_ptr_type;
2036 }
2037
2038 static LLVMValueRef
2039 generate_mask_value(struct draw_gs_llvm_variant *variant,
2040 struct lp_type gs_type)
2041 {
2042 struct gallivm_state *gallivm = variant->gallivm;
2043 LLVMBuilderRef builder = gallivm->builder;
2044 LLVMValueRef bits[16];
2045 struct lp_type mask_type = lp_int_type(gs_type);
2046 struct lp_type mask_elem_type = lp_elem_type(mask_type);
2047 LLVMValueRef mask_val = lp_build_const_vec(gallivm,
2048 mask_type,
2049 0);
2050 unsigned i;
2051
2052 assert(gs_type.length <= Elements(bits));
2053
2054 for (i = gs_type.length; i >= 1; --i) {
2055 int idx = i - 1;
2056 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
2057 bits[idx] = lp_build_compare(gallivm,
2058 mask_elem_type, PIPE_FUNC_GEQUAL,
2059 variant->num_prims, ind);
2060 }
2061 for (i = 0; i < gs_type.length; ++i) {
2062 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
2063 mask_val = LLVMBuildInsertElement(builder, mask_val, bits[i], ind, "");
2064 }
2065 mask_val = lp_build_compare(gallivm,
2066 mask_type, PIPE_FUNC_NOTEQUAL,
2067 mask_val,
2068 lp_build_const_int_vec(gallivm, mask_type, 0));
2069
2070 return mask_val;
2071 }
2072
2073 static void
2074 draw_gs_llvm_generate(struct draw_llvm *llvm,
2075 struct draw_gs_llvm_variant *variant)
2076 {
2077 struct gallivm_state *gallivm = variant->gallivm;
2078 LLVMContextRef context = gallivm->context;
2079 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2080 LLVMTypeRef arg_types[6];
2081 LLVMTypeRef func_type;
2082 LLVMValueRef variant_func;
2083 LLVMValueRef context_ptr;
2084 LLVMValueRef prim_id_ptr;
2085 LLVMBasicBlockRef block;
2086 LLVMBuilderRef builder;
2087 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2088 struct lp_build_sampler_soa *sampler = 0;
2089 struct lp_build_context bld;
2090 struct lp_bld_tgsi_system_values system_values;
2091 struct lp_type gs_type;
2092 unsigned i;
2093 struct draw_gs_llvm_iface gs_iface;
2094 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2095 LLVMValueRef consts_ptr;
2096 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2097 struct lp_build_mask_context mask;
2098 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2099 unsigned vector_length = variant->shader->base.vector_length;
2100
2101 memset(&system_values, 0, sizeof(system_values));
2102
2103 assert(variant->vertex_header_ptr_type);
2104
2105 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2106 arg_types[1] = variant->input_array_type; /* input */
2107 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2108 arg_types[3] = int32_type; /* num_prims */
2109 arg_types[4] = int32_type; /* instance_id */
2110 arg_types[5] = LLVMPointerType(
2111 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2112
2113 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2114
2115 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
2116 func_type);
2117 variant->function = variant_func;
2118
2119 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2120
2121 for (i = 0; i < Elements(arg_types); ++i)
2122 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2123 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2124 LLVMNoAliasAttribute);
2125
2126 context_ptr = LLVMGetParam(variant_func, 0);
2127 input_array = LLVMGetParam(variant_func, 1);
2128 io_ptr = LLVMGetParam(variant_func, 2);
2129 num_prims = LLVMGetParam(variant_func, 3);
2130 system_values.instance_id = LLVMGetParam(variant_func, 4);
2131 prim_id_ptr = LLVMGetParam(variant_func, 5);
2132
2133 lp_build_name(context_ptr, "context");
2134 lp_build_name(input_array, "input");
2135 lp_build_name(io_ptr, "io");
2136 lp_build_name(num_prims, "num_prims");
2137 lp_build_name(system_values.instance_id, "instance_id");
2138 lp_build_name(prim_id_ptr, "prim_id_ptr");
2139
2140 variant->context_ptr = context_ptr;
2141 variant->io_ptr = io_ptr;
2142 variant->num_prims = num_prims;
2143
2144 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2145 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2146 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2147 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2148 gs_iface.input = input_array;
2149 gs_iface.variant = variant;
2150
2151 /*
2152 * Function body
2153 */
2154
2155 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2156 builder = gallivm->builder;
2157 LLVMPositionBuilderAtEnd(builder, block);
2158
2159 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2160
2161 memset(&gs_type, 0, sizeof gs_type);
2162 gs_type.floating = TRUE; /* floating point values */
2163 gs_type.sign = TRUE; /* values are signed */
2164 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2165 gs_type.width = 32; /* 32-bit float */
2166 gs_type.length = vector_length;
2167
2168 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2169
2170 /* code generated texture sampling */
2171 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2172 context_ptr);
2173
2174 mask_val = generate_mask_value(variant, gs_type);
2175 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2176
2177 if (gs_info->uses_primid) {
2178 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");;
2179 }
2180
2181 lp_build_tgsi_soa(variant->gallivm,
2182 tokens,
2183 gs_type,
2184 &mask,
2185 consts_ptr,
2186 &system_values,
2187 NULL,
2188 outputs,
2189 sampler,
2190 &llvm->draw->gs.geometry_shader->info,
2191 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2192
2193 sampler->destroy(sampler);
2194
2195 lp_build_mask_end(&mask);
2196
2197 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2198
2199 gallivm_verify_function(gallivm, variant_func);
2200 }
2201
2202
2203 struct draw_gs_llvm_variant *
2204 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2205 unsigned num_outputs,
2206 const struct draw_gs_llvm_variant_key *key)
2207 {
2208 struct draw_gs_llvm_variant *variant;
2209 struct llvm_geometry_shader *shader =
2210 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2211 LLVMTypeRef vertex_header;
2212
2213 variant = MALLOC(sizeof *variant +
2214 shader->variant_key_size -
2215 sizeof variant->key);
2216 if (variant == NULL)
2217 return NULL;
2218
2219 variant->llvm = llvm;
2220 variant->shader = shader;
2221
2222 variant->gallivm = gallivm_create();
2223
2224 create_gs_jit_types(variant);
2225
2226 memcpy(&variant->key, key, shader->variant_key_size);
2227
2228 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2229
2230 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2231
2232 draw_gs_llvm_generate(llvm, variant);
2233
2234 gallivm_compile_module(variant->gallivm);
2235
2236 variant->jit_func = (draw_gs_jit_func)
2237 gallivm_jit_function(variant->gallivm, variant->function);
2238
2239 variant->list_item_global.base = variant;
2240 variant->list_item_local.base = variant;
2241 /*variant->no = */shader->variants_created++;
2242 variant->list_item_global.base = variant;
2243
2244 return variant;
2245 }
2246
2247 void
2248 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2249 {
2250 struct draw_llvm *llvm = variant->llvm;
2251
2252 if (variant->function) {
2253 gallivm_free_function(variant->gallivm,
2254 variant->function, variant->jit_func);
2255 }
2256
2257 gallivm_destroy(variant->gallivm);
2258
2259 remove_from_list(&variant->list_item_local);
2260 variant->shader->variants_cached--;
2261 remove_from_list(&variant->list_item_global);
2262 llvm->nr_gs_variants--;
2263 FREE(variant);
2264 }
2265
2266 struct draw_gs_llvm_variant_key *
2267 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2268 {
2269 unsigned i;
2270 struct draw_gs_llvm_variant_key *key;
2271 struct draw_sampler_static_state *draw_sampler;
2272
2273 key = (struct draw_gs_llvm_variant_key *)store;
2274
2275 /* All variants of this shader will have the same value for
2276 * nr_samplers. Not yet trying to compact away holes in the
2277 * sampler array.
2278 */
2279 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2280 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2281 key->nr_sampler_views =
2282 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2283 }
2284 else {
2285 key->nr_sampler_views = key->nr_samplers;
2286 }
2287
2288 draw_sampler = key->samplers;
2289
2290 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2291
2292 for (i = 0 ; i < key->nr_samplers; i++) {
2293 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2294 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2295 }
2296 for (i = 0 ; i < key->nr_sampler_views; i++) {
2297 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2298 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2299 }
2300
2301 return key;
2302 }
2303
2304 void
2305 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2306 {
2307 unsigned i;
2308 struct draw_sampler_static_state *sampler = key->samplers;
2309
2310 for (i = 0 ; i < key->nr_sampler_views; i++) {
2311 debug_printf("sampler[%i].src_format = %s\n", i,
2312 util_format_name(sampler[i].texture_state.format));
2313 }
2314 }