projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: (trivial) fix NULL pointer dereference
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
1 /**************************************************************************
2 *
3 * Copyright 2010 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include "draw_llvm.h"
29
30 #include "draw_context.h"
31 #include "draw_vs.h"
32 #include "draw_gs.h"
33
34 #include "gallivm/lp_bld_arit.h"
35 #include "gallivm/lp_bld_arit_overflow.h"
36 #include "gallivm/lp_bld_logic.h"
37 #include "gallivm/lp_bld_const.h"
38 #include "gallivm/lp_bld_swizzle.h"
39 #include "gallivm/lp_bld_struct.h"
40 #include "gallivm/lp_bld_type.h"
41 #include "gallivm/lp_bld_flow.h"
42 #include "gallivm/lp_bld_debug.h"
43 #include "gallivm/lp_bld_tgsi.h"
44 #include "gallivm/lp_bld_printf.h"
45 #include "gallivm/lp_bld_intr.h"
46 #include "gallivm/lp_bld_init.h"
47 #include "gallivm/lp_bld_type.h"
48 #include "gallivm/lp_bld_pack.h"
49 #include "gallivm/lp_bld_format.h"
50
51 #include "tgsi/tgsi_exec.h"
52 #include "tgsi/tgsi_dump.h"
53
54 #include "util/u_math.h"
55 #include "util/u_pointer.h"
56 #include "util/u_string.h"
57 #include "util/simple_list.h"
58
59
60 #define DEBUG_STORE 0
61
62
63 static void
64 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
65 boolean elts);
66
67
68 struct draw_gs_llvm_iface {
69 struct lp_build_tgsi_gs_iface base;
70
71 struct draw_gs_llvm_variant *variant;
72 LLVMValueRef input;
73 };
74
75 static INLINE const struct draw_gs_llvm_iface *
76 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
77 {
78 return (const struct draw_gs_llvm_iface *)iface;
79 }
80
81 /**
82 * Create LLVM type for draw_vertex_buffer.
83 */
84 static LLVMTypeRef
85 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
86 const char *struct_name)
87 {
88 LLVMTargetDataRef target = gallivm->target;
89 LLVMTypeRef dvbuffer_type;
90 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
91 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
92
93 elem_types[DRAW_JIT_DVBUFFER_MAP] =
94 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
95 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
96
97 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
98 Elements(elem_types), 0);
99
100 (void) target; /* silence unused var warning for non-debug build */
101 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
102 target, dvbuffer_type,
103 DRAW_JIT_DVBUFFER_MAP);
104 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
105 target, dvbuffer_type,
106 DRAW_JIT_DVBUFFER_SIZE);
107
108 return dvbuffer_type;
109 }
110
111 /**
112 * Create LLVM type for struct draw_jit_texture
113 */
114 static LLVMTypeRef
115 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
116 {
117 LLVMTargetDataRef target = gallivm->target;
118 LLVMTypeRef texture_type;
119 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
120 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
121
122 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
123 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
124 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
125 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
126 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
127 elem_types[DRAW_JIT_TEXTURE_BASE] =
128 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
129 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
130 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
131 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
132 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
133
134 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
135 Elements(elem_types), 0);
136
137 (void) target; /* silence unused var warning for non-debug build */
138 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
139 target, texture_type,
140 DRAW_JIT_TEXTURE_WIDTH);
141 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
142 target, texture_type,
143 DRAW_JIT_TEXTURE_HEIGHT);
144 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
145 target, texture_type,
146 DRAW_JIT_TEXTURE_DEPTH);
147 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
148 target, texture_type,
149 DRAW_JIT_TEXTURE_FIRST_LEVEL);
150 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
151 target, texture_type,
152 DRAW_JIT_TEXTURE_LAST_LEVEL);
153 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
154 target, texture_type,
155 DRAW_JIT_TEXTURE_BASE);
156 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
157 target, texture_type,
158 DRAW_JIT_TEXTURE_ROW_STRIDE);
159 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
160 target, texture_type,
161 DRAW_JIT_TEXTURE_IMG_STRIDE);
162 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
163 target, texture_type,
164 DRAW_JIT_TEXTURE_MIP_OFFSETS);
165
166 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
167
168 return texture_type;
169 }
170
171
172 /**
173 * Create LLVM type for struct draw_jit_sampler
174 */
175 static LLVMTypeRef
176 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
177 {
178 LLVMTargetDataRef target = gallivm->target;
179 LLVMTypeRef sampler_type;
180 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
181
182 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
183 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
184 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
185 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
186 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
187
188 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
189 Elements(elem_types), 0);
190
191 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
192 target, sampler_type,
193 DRAW_JIT_SAMPLER_MIN_LOD);
194 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
195 target, sampler_type,
196 DRAW_JIT_SAMPLER_MAX_LOD);
197 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
198 target, sampler_type,
199 DRAW_JIT_SAMPLER_LOD_BIAS);
200 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
201 target, sampler_type,
202 DRAW_JIT_SAMPLER_BORDER_COLOR);
203
204 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
205
206 return sampler_type;
207 }
208
209
210 /**
211 * Create LLVM type for struct draw_jit_context
212 */
213 static LLVMTypeRef
214 create_jit_context_type(struct gallivm_state *gallivm,
215 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
216 const char *struct_name)
217 {
218 LLVMTargetDataRef target = gallivm->target;
219 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
220 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
221 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
222 LLVMTypeRef context_type;
223
224 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
225 LP_MAX_TGSI_CONST_BUFFERS);
226 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
227 LP_MAX_TGSI_CONST_BUFFERS);
228 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
229 DRAW_TOTAL_CLIP_PLANES), 0);
230 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
231 elem_types[4] = LLVMArrayType(texture_type,
232 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
233 elem_types[5] = LLVMArrayType(sampler_type,
234 PIPE_MAX_SAMPLERS); /* samplers */
235 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
236 Elements(elem_types), 0);
237 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
238 target, context_type, DRAW_JIT_CTX_CONSTANTS);
239 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
240 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
241 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
242 target, context_type, DRAW_JIT_CTX_PLANES);
243 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
244 target, context_type, DRAW_JIT_CTX_VIEWPORT);
245 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
246 target, context_type,
247 DRAW_JIT_CTX_TEXTURES);
248 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
249 target, context_type,
250 DRAW_JIT_CTX_SAMPLERS);
251 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
252 target, context_type);
253
254 return context_type;
255 }
256
257
258 /**
259 * Create LLVM type for struct draw_gs_jit_context
260 */
261 static LLVMTypeRef
262 create_gs_jit_context_type(struct gallivm_state *gallivm,
263 unsigned vector_length,
264 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
265 const char *struct_name)
266 {
267 LLVMTargetDataRef target = gallivm->target;
268 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
269 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
270 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
271 LLVMTypeRef context_type;
272
273 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
274 LP_MAX_TGSI_CONST_BUFFERS);
275 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
276 LP_MAX_TGSI_CONST_BUFFERS);
277 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
278 DRAW_TOTAL_CLIP_PLANES), 0);
279 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
280
281 elem_types[4] = LLVMArrayType(texture_type,
282 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
283 elem_types[5] = LLVMArrayType(sampler_type,
284 PIPE_MAX_SAMPLERS); /* samplers */
285
286 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
287 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
288 vector_length), 0);
289 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
290 vector_length), 0);
291
292 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
293 Elements(elem_types), 0);
294
295 (void) target; /* silence unused var warning for non-debug build */
296 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
297 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
298 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
299 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
300 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
301 target, context_type, DRAW_GS_JIT_CTX_PLANES);
302 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
303 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
304 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
305 target, context_type,
306 DRAW_GS_JIT_CTX_TEXTURES);
307 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
308 target, context_type,
309 DRAW_GS_JIT_CTX_SAMPLERS);
310 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
311 target, context_type,
312 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
313 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
314 target, context_type,
315 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
316 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
317 target, context_type,
318 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
319 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
320 target, context_type);
321
322 return context_type;
323 }
324
325
326 static LLVMTypeRef
327 create_gs_jit_input_type(struct gallivm_state *gallivm)
328 {
329 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
330 LLVMTypeRef input_array;
331
332 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
333 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
334 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
335 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
336
337 return input_array;
338 }
339
340 /**
341 * Create LLVM type for struct pipe_vertex_buffer
342 */
343 static LLVMTypeRef
344 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
345 const char *struct_name)
346 {
347 LLVMTargetDataRef target = gallivm->target;
348 LLVMTypeRef elem_types[4];
349 LLVMTypeRef vb_type;
350
351 elem_types[0] =
352 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
353 elem_types[2] =
354 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
355
356 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
357 Elements(elem_types), 0);
358
359 (void) target; /* silence unused var warning for non-debug build */
360 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
361 target, vb_type, 0);
362 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
363 target, vb_type, 1);
364
365 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
366
367 return vb_type;
368 }
369
370
371 /**
372 * Create LLVM type for struct vertex_header;
373 */
374 static LLVMTypeRef
375 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
376 {
377 LLVMTargetDataRef target = gallivm->target;
378 LLVMTypeRef elem_types[4];
379 LLVMTypeRef vertex_header;
380 char struct_name[24];
381
382 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
383
384 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
385 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
386 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
387 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
388
389 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
390 Elements(elem_types), 0);
391
392 /* these are bit-fields and we can't take address of them
393 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
394 target, vertex_header,
395 DRAW_JIT_VERTEX_CLIPMASK);
396 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
397 target, vertex_header,
398 DRAW_JIT_VERTEX_EDGEFLAG);
399 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
400 target, vertex_header,
401 DRAW_JIT_VERTEX_PAD);
402 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
403 target, vertex_header,
404 DRAW_JIT_VERTEX_VERTEX_ID);
405 */
406 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
407 target, vertex_header,
408 DRAW_JIT_VERTEX_CLIP);
409 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
410 target, vertex_header,
411 DRAW_JIT_VERTEX_PRE_CLIP_POS);
412 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
413 target, vertex_header,
414 DRAW_JIT_VERTEX_DATA);
415
416 assert(LLVMABISizeOfType(target, vertex_header) ==
417 offsetof(struct vertex_header, data[data_elems]));
418
419 return vertex_header;
420 }
421
422
423 /**
424 * Create LLVM types for various structures.
425 */
426 static void
427 create_jit_types(struct draw_llvm_variant *variant)
428 {
429 struct gallivm_state *gallivm = variant->gallivm;
430 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
431 vb_type;
432
433 texture_type = create_jit_texture_type(gallivm, "texture");
434 sampler_type = create_jit_sampler_type(gallivm, "sampler");
435
436 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
437 "draw_jit_context");
438 variant->context_ptr_type = LLVMPointerType(context_type, 0);
439
440 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
441 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
442
443 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
444 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
445 }
446
447
448 static LLVMTypeRef
449 get_context_ptr_type(struct draw_llvm_variant *variant)
450 {
451 if (!variant->context_ptr_type)
452 create_jit_types(variant);
453 return variant->context_ptr_type;
454 }
455
456
457 static LLVMTypeRef
458 get_buffer_ptr_type(struct draw_llvm_variant *variant)
459 {
460 if (!variant->buffer_ptr_type)
461 create_jit_types(variant);
462 return variant->buffer_ptr_type;
463 }
464
465
466 static LLVMTypeRef
467 get_vb_ptr_type(struct draw_llvm_variant *variant)
468 {
469 if (!variant->vb_ptr_type)
470 create_jit_types(variant);
471 return variant->vb_ptr_type;
472 }
473
474 static LLVMTypeRef
475 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
476 {
477 if (!variant->vertex_header_ptr_type)
478 create_jit_types(variant);
479 return variant->vertex_header_ptr_type;
480 }
481
482
483 /**
484 * Create per-context LLVM info.
485 */
486 struct draw_llvm *
487 draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
488 {
489 struct draw_llvm *llvm;
490
491 if (!lp_build_init())
492 return NULL;
493
494 llvm = CALLOC_STRUCT( draw_llvm );
495 if (!llvm)
496 return NULL;
497
498 llvm->draw = draw;
499
500 llvm->context = context;
501 if (!llvm->context) {
502 llvm->context = LLVMContextCreate();
503 llvm->context_owned = true;
504 }
505 if (!llvm->context)
506 goto fail;
507
508 llvm->nr_variants = 0;
509 make_empty_list(&llvm->vs_variants_list);
510
511 llvm->nr_gs_variants = 0;
512 make_empty_list(&llvm->gs_variants_list);
513
514 return llvm;
515
516 fail:
517 draw_llvm_destroy(llvm);
518 return NULL;
519 }
520
521
522 /**
523 * Free per-context LLVM info.
524 */
525 void
526 draw_llvm_destroy(struct draw_llvm *llvm)
527 {
528 if (llvm->context_owned)
529 LLVMContextDispose(llvm->context);
530 llvm->context = NULL;
531
532 /* XXX free other draw_llvm data? */
533 FREE(llvm);
534 }
535
536
537 /**
538 * Create LLVM-generated code for a vertex shader.
539 */
540 struct draw_llvm_variant *
541 draw_llvm_create_variant(struct draw_llvm *llvm,
542 unsigned num_inputs,
543 const struct draw_llvm_variant_key *key)
544 {
545 struct draw_llvm_variant *variant;
546 struct llvm_vertex_shader *shader =
547 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
548 LLVMTypeRef vertex_header;
549 char module_name[64];
550
551 variant = MALLOC(sizeof *variant +
552 shader->variant_key_size -
553 sizeof variant->key);
554 if (variant == NULL)
555 return NULL;
556
557 variant->llvm = llvm;
558 variant->shader = shader;
559
560 util_snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
561 variant->shader->variants_cached);
562
563 variant->gallivm = gallivm_create(module_name, llvm->context);
564
565 create_jit_types(variant);
566
567 memcpy(&variant->key, key, shader->variant_key_size);
568
569 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
570 tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
571 draw_llvm_dump_variant_key(&variant->key);
572 }
573
574 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
575
576 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
577
578 draw_llvm_generate(llvm, variant, FALSE); /* linear */
579 draw_llvm_generate(llvm, variant, TRUE); /* elts */
580
581 gallivm_compile_module(variant->gallivm);
582
583 variant->jit_func = (draw_jit_vert_func)
584 gallivm_jit_function(variant->gallivm, variant->function);
585
586 variant->jit_func_elts = (draw_jit_vert_func_elts)
587 gallivm_jit_function(variant->gallivm, variant->function_elts);
588
589 gallivm_free_ir(variant->gallivm);
590
591 variant->list_item_global.base = variant;
592 variant->list_item_local.base = variant;
593 /*variant->no = */shader->variants_created++;
594 variant->list_item_global.base = variant;
595
596 return variant;
597 }
598
599
600 static void
601 generate_vs(struct draw_llvm_variant *variant,
602 LLVMBuilderRef builder,
603 struct lp_type vs_type,
604 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
605 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
606 const struct lp_bld_tgsi_system_values *system_values,
607 LLVMValueRef context_ptr,
608 struct lp_build_sampler_soa *draw_sampler,
609 boolean clamp_vertex_color)
610 {
611 struct draw_llvm *llvm = variant->llvm;
612 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
613 LLVMValueRef consts_ptr =
614 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
615 LLVMValueRef num_consts_ptr =
616 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
617
618 lp_build_tgsi_soa(variant->gallivm,
619 tokens,
620 vs_type,
621 NULL /*struct lp_build_mask_context *mask*/,
622 consts_ptr,
623 num_consts_ptr,
624 system_values,
625 inputs,
626 outputs,
627 context_ptr,
628 draw_sampler,
629 &llvm->draw->vs.vertex_shader->info,
630 NULL);
631
632 {
633 LLVMValueRef out;
634 unsigned chan, attrib;
635 struct lp_build_context bld;
636 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
637 lp_build_context_init(&bld, variant->gallivm, vs_type);
638
639 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
640 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
641 if (outputs[attrib][chan]) {
642 switch (info->output_semantic_name[attrib]) {
643 case TGSI_SEMANTIC_COLOR:
644 case TGSI_SEMANTIC_BCOLOR:
645 if (clamp_vertex_color) {
646 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
647 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
648 LLVMBuildStore(builder, out, outputs[attrib][chan]);
649 }
650 break;
651 }
652 }
653 }
654 }
655 }
656 }
657
658 static void
659 generate_fetch(struct gallivm_state *gallivm,
660 struct draw_context *draw,
661 LLVMValueRef vbuffers_ptr,
662 LLVMValueRef *res,
663 struct pipe_vertex_element *velem,
664 LLVMValueRef vbuf,
665 LLVMValueRef index,
666 LLVMValueRef instance_id,
667 LLVMValueRef start_instance)
668 {
669 const struct util_format_description *format_desc =
670 util_format_description(velem->src_format);
671 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
672 LLVMBuilderRef builder = gallivm->builder;
673 LLVMValueRef indices =
674 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
675 velem->vertex_buffer_index, 0);
676 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
677 &indices, 1, "");
678 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
679 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
680 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
681 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
682 LLVMValueRef stride;
683 LLVMValueRef buffer_overflowed;
684 LLVMValueRef needed_buffer_size;
685 LLVMValueRef temp_ptr =
686 lp_build_alloca(gallivm,
687 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
688 LLVMValueRef ofbit = NULL;
689 struct lp_build_if_state if_ctx;
690
691 if (velem->instance_divisor) {
692 /* Index is equal to the start instance plus the number of current
693 * instance divided by the divisor. In this case we compute it as:
694 * index = start_instance + (instance_id / divisor)
695 */
696 LLVMValueRef current_instance;
697 current_instance = LLVMBuildUDiv(builder, instance_id,
698 lp_build_const_int32(gallivm, velem->instance_divisor),
699 "instance_divisor");
700 index = lp_build_uadd_overflow(gallivm, start_instance,
701 current_instance, &ofbit);
702 }
703
704 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
705 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
706 stride = lp_build_uadd_overflow(
707 gallivm, stride,
708 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
709 needed_buffer_size = lp_build_uadd_overflow(
710 gallivm, stride,
711 lp_build_const_int32(gallivm,
712 util_format_get_blocksize(velem->src_format)),
713 &ofbit);
714
715 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
716 needed_buffer_size, buffer_size,
717 "buffer_overflowed");
718 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
719 #if 0
720 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
721 index, vb_stride);
722 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
723 vb_buffer_offset,
724 lp_build_const_int32(gallivm, velem->src_offset));
725 lp_build_print_value(gallivm, " blocksize = ",
726 lp_build_const_int32(
727 gallivm,
728 util_format_get_blocksize(velem->src_format)));
729 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
730 lp_build_printf(gallivm, " stride = %u\n", stride);
731 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
732 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
733 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
734 #endif
735
736 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
737 {
738 LLVMValueRef val =
739 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
740 LLVMBuildStore(builder, val, temp_ptr);
741 }
742 lp_build_else(&if_ctx);
743 {
744 LLVMValueRef val;
745 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
746
747 val = lp_build_fetch_rgba_aos(gallivm,
748 format_desc,
749 lp_float32_vec4_type(),
750 FALSE,
751 map_ptr,
752 zero, zero, zero);
753 LLVMBuildStore(builder, val, temp_ptr);
754 }
755 lp_build_endif(&if_ctx);
756
757 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
758 }
759
760 static void
761 convert_to_soa(struct gallivm_state *gallivm,
762 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
763 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
764 unsigned num_attribs, const struct lp_type soa_type)
765 {
766 unsigned i, j, k;
767 struct lp_type aos_channel_type = soa_type;
768
769 debug_assert(TGSI_NUM_CHANNELS == 4);
770 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
771
772 aos_channel_type.length >>= 1;
773
774 for (i = 0; i < num_attribs; ++i) {
775 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
776 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
777
778 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
779 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
780
781 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
782
783 for (k = 0; k < pixels_per_channel; ++k) {
784 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
785 }
786
787 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
788 }
789
790 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
791 }
792 }
793
794
795 static void
796 store_aos(struct gallivm_state *gallivm,
797 LLVMValueRef io_ptr,
798 LLVMValueRef index,
799 LLVMValueRef value)
800 {
801 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
802 LLVMBuilderRef builder = gallivm->builder;
803 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
804 LLVMValueRef indices[3];
805
806 indices[0] = lp_build_const_int32(gallivm, 0);
807 indices[1] = index;
808 indices[2] = lp_build_const_int32(gallivm, 0);
809
810 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
811 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
812
813 #if DEBUG_STORE
814 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
815 #endif
816
817 /* Unaligned store due to the vertex header */
818 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
819 }
820
821 /**
822 * Adjust the mask to architecture endianess. The mask will the store in struct:
823 *
824 * struct vertex_header {
825 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
826 * unsigned edgeflag:1;
827 * unsigned have_clipdist:1;
828 * unsigned vertex_id:16;
829 * [...]
830 * }
831 *
832 * On little-endian machine nothing needs to done, however on bit-endian machine
833 * the mask's fields need to be adjusted with the algorithm:
834 *
835 * uint32_t reverse (uint32_t x)
836 * {
837 * return (x >> 16) | // vertex_id
838 * ((x & 0x3fff) << 18) | // clipmask
839 * ((x & 0x4000) << 3) | // have_clipdist
840 * ((x & 0x8000) << 1); // edgeflag
841 * }
842 */
843 static LLVMValueRef
844 adjust_mask(struct gallivm_state *gallivm,
845 LLVMValueRef mask)
846 {
847 #ifdef PIPE_ARCH_BIG_ENDIAN
848 LLVMBuilderRef builder = gallivm->builder;
849 LLVMValueRef vertex_id;
850 LLVMValueRef clipmask;
851 LLVMValueRef have_clipdist;
852 LLVMValueRef edgeflag;
853
854 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
855 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
856 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
857 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
858 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
859 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
860 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
861
862 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
863 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
864 mask = LLVMBuildOr(builder, mask, edgeflag, "");
865 #endif
866 return mask;
867 }
868
869 static void
870 store_aos_array(struct gallivm_state *gallivm,
871 struct lp_type soa_type,
872 LLVMValueRef io_ptr,
873 LLVMValueRef *indices,
874 LLVMValueRef* aos,
875 int attrib,
876 int num_outputs,
877 LLVMValueRef clipmask,
878 boolean have_clipdist)
879 {
880 LLVMBuilderRef builder = gallivm->builder;
881 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
882 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
883 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
884 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
885 int vector_length = soa_type.length;
886 int i;
887
888 debug_assert(TGSI_NUM_CHANNELS == 4);
889
890 for (i = 0; i < vector_length; i++) {
891 linear_inds[i] = lp_build_const_int32(gallivm, i);
892 if (indices) {
893 inds[i] = indices[i];
894 } else {
895 inds[i] = linear_inds[i];
896 }
897 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
898 }
899
900 if (attrib == 0) {
901 /* store vertex header for each of the n vertices */
902 LLVMValueRef val, cliptmp;
903 int vertex_id_pad_edgeflag;
904
905 /* If this assertion fails, it means we need to update the bit twidding
906 * code here. See struct vertex_header in draw_private.h.
907 */
908 assert(DRAW_TOTAL_CLIP_PLANES==14);
909 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
910 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
911 if (have_clipdist)
912 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
913 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
914 /* OR with the clipmask */
915 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
916 for (i = 0; i < vector_length; i++) {
917 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
918 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
919 val = adjust_mask(gallivm, val);
920 #if DEBUG_STORE
921 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
922 io_ptrs[i], inds[i], val);
923 #endif
924 LLVMBuildStore(builder, val, id_ptr);
925 }
926 }
927
928 /* store for each of the n vertices */
929 for (i = 0; i < vector_length; i++) {
930 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
931 }
932 }
933
934
935 static void
936 convert_to_aos(struct gallivm_state *gallivm,
937 LLVMValueRef io,
938 LLVMValueRef *indices,
939 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
940 LLVMValueRef clipmask,
941 int num_outputs,
942 struct lp_type soa_type,
943 boolean have_clipdist)
944 {
945 LLVMBuilderRef builder = gallivm->builder;
946 unsigned chan, attrib, i;
947
948 #if DEBUG_STORE
949 lp_build_printf(gallivm, " # storing begin\n");
950 #endif
951 for (attrib = 0; attrib < num_outputs; ++attrib) {
952 LLVMValueRef soa[TGSI_NUM_CHANNELS];
953 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
954 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
955 if (outputs[attrib][chan]) {
956 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
957 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
958 #if DEBUG_STORE
959 lp_build_printf(gallivm, "output %d : %d ",
960 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
961 attrib, 0),
962 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
963 chan, 0));
964 lp_build_print_value(gallivm, "val = ", out);
965 {
966 LLVMValueRef iv =
967 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
968
969 lp_build_print_value(gallivm, " ival = ", iv);
970 }
971 #endif
972 soa[chan] = out;
973 }
974 else {
975 soa[chan] = 0;
976 }
977 }
978
979
980 if (soa_type.length == TGSI_NUM_CHANNELS) {
981 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
982 } else {
983 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
984
985 for (i = 0; i < soa_type.length; ++i) {
986 aos[i] = lp_build_extract_range(gallivm,
987 soa[i % TGSI_NUM_CHANNELS],
988 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
989 TGSI_NUM_CHANNELS);
990 }
991 }
992
993 store_aos_array(gallivm,
994 soa_type,
995 io, indices,
996 aos,
997 attrib,
998 num_outputs,
999 clipmask, have_clipdist);
1000 }
1001 #if DEBUG_STORE
1002 lp_build_printf(gallivm, " # storing end\n");
1003 #endif
1004 }
1005
1006
1007 /**
1008 * Stores original vertex positions in clip coordinates
1009 */
1010 static void
1011 store_clip(struct gallivm_state *gallivm,
1012 const struct lp_type vs_type,
1013 LLVMValueRef io_ptr,
1014 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1015 boolean pre_clip_pos, int idx)
1016 {
1017 LLVMBuilderRef builder = gallivm->builder;
1018 LLVMValueRef soa[4];
1019 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1020 LLVMValueRef indices[2];
1021 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1022 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1023 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1024 LLVMTypeRef clip_ptr_type =
1025 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1026 4), 0);
1027 int i, j;
1028
1029 indices[0] =
1030 indices[1] = lp_build_const_int32(gallivm, 0);
1031
1032 for (i = 0; i < vs_type.length; i++) {
1033 inds[i] = lp_build_const_int32(gallivm, i);
1034 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1035 }
1036
1037 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1038 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1039 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1040 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1041
1042 if (!pre_clip_pos) {
1043 for (i = 0; i < vs_type.length; i++) {
1044 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1045 }
1046 } else {
1047 for (i = 0; i < vs_type.length; i++) {
1048 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1049 }
1050 }
1051
1052 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1053 for (i = 0; i < vs_type.length; ++i) {
1054 aos[i] = lp_build_extract_range(gallivm,
1055 soa[i % TGSI_NUM_CHANNELS],
1056 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1057 TGSI_NUM_CHANNELS);
1058 }
1059
1060 for (j = 0; j < vs_type.length; j++) {
1061 LLVMValueRef clip_ptr;
1062
1063 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1064 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1065
1066 /* Unaligned store */
1067 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1068 }
1069 }
1070
1071
1072 /**
1073 * Transforms the outputs for viewport mapping
1074 */
1075 static void
1076 generate_viewport(struct draw_llvm_variant *variant,
1077 LLVMBuilderRef builder,
1078 struct lp_type vs_type,
1079 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1080 LLVMValueRef context_ptr)
1081 {
1082 int i;
1083 struct gallivm_state *gallivm = variant->gallivm;
1084 struct lp_type f32_type = vs_type;
1085 const unsigned pos = variant->llvm->draw->vs.position_output;
1086 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1087 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1088 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1089 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1090
1091 /* We treat pipe_viewport_state as a float array */
1092 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1093 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1094
1095 /* for 1/w convention*/
1096 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1097 LLVMBuildStore(builder, out3, outputs[pos][3]);
1098
1099 /* Viewport Mapping */
1100 for (i=0; i<3; i++) {
1101 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1102 LLVMValueRef scale;
1103 LLVMValueRef trans;
1104 LLVMValueRef scale_i;
1105 LLVMValueRef trans_i;
1106 LLVMValueRef index;
1107
1108 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1109 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1110
1111 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1112 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1113
1114 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1115 LLVMBuildLoad(builder, scale_i, "scale"));
1116 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1117 LLVMBuildLoad(builder, trans_i, "trans"));
1118
1119 /* divide by w */
1120 out = LLVMBuildFMul(builder, out, out3, "");
1121 /* mult by scale */
1122 out = LLVMBuildFMul(builder, out, scale, "");
1123 /* add translation */
1124 out = LLVMBuildFAdd(builder, out, trans, "");
1125
1126 /* store transformed outputs */
1127 LLVMBuildStore(builder, out, outputs[pos][i]);
1128 }
1129
1130 }
1131
1132
1133 /**
1134 * Returns clipmask as nxi32 bitmask for the n vertices
1135 */
1136 static LLVMValueRef
1137 generate_clipmask(struct draw_llvm *llvm,
1138 struct gallivm_state *gallivm,
1139 struct lp_type vs_type,
1140 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1141 boolean clip_xy,
1142 boolean clip_z,
1143 boolean clip_user,
1144 boolean clip_halfz,
1145 unsigned ucp_enable,
1146 LLVMValueRef context_ptr,
1147 boolean *have_clipdist)
1148 {
1149 LLVMBuilderRef builder = gallivm->builder;
1150 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1151 LLVMValueRef test, temp;
1152 LLVMValueRef zero, shift;
1153 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1154 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1155 LLVMValueRef plane1, planes, plane_ptr, sum;
1156 struct lp_type f32_type = vs_type;
1157 struct lp_type i32_type = lp_int_type(vs_type);
1158 const unsigned pos = llvm->draw->vs.position_output;
1159 const unsigned cv = llvm->draw->vs.clipvertex_output;
1160 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1161 bool have_cd = false;
1162 unsigned cd[2];
1163
1164 cd[0] = llvm->draw->vs.clipdistance_output[0];
1165 cd[1] = llvm->draw->vs.clipdistance_output[1];
1166
1167 if (cd[0] != pos || cd[1] != pos)
1168 have_cd = true;
1169
1170 if (num_written_clipdistance && !clip_user) {
1171 clip_user = true;
1172 ucp_enable = (1 << num_written_clipdistance) - 1;
1173 }
1174
1175 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1176 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1177 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1178 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1179
1180 /*
1181 * load clipvertex and position from correct locations.
1182 * if they are the same just load them once.
1183 */
1184 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1185 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1186 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1187 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1188
1189 if (clip_user && cv != pos) {
1190 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1191 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1192 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1193 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1194 } else {
1195 cv_x = pos_x;
1196 cv_y = pos_y;
1197 cv_z = pos_z;
1198 cv_w = pos_w;
1199 }
1200
1201 /* Cliptest, for hardwired planes */
1202 if (clip_xy) {
1203 /* plane 1 */
1204 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1205 temp = shift;
1206 test = LLVMBuildAnd(builder, test, temp, "");
1207 mask = test;
1208
1209 /* plane 2 */
1210 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1211 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1212 temp = LLVMBuildShl(builder, temp, shift, "");
1213 test = LLVMBuildAnd(builder, test, temp, "");
1214 mask = LLVMBuildOr(builder, mask, test, "");
1215
1216 /* plane 3 */
1217 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1218 temp = LLVMBuildShl(builder, temp, shift, "");
1219 test = LLVMBuildAnd(builder, test, temp, "");
1220 mask = LLVMBuildOr(builder, mask, test, "");
1221
1222 /* plane 4 */
1223 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1224 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1225 temp = LLVMBuildShl(builder, temp, shift, "");
1226 test = LLVMBuildAnd(builder, test, temp, "");
1227 mask = LLVMBuildOr(builder, mask, test, "");
1228 }
1229
1230 if (clip_z) {
1231 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1232 if (clip_halfz) {
1233 /* plane 5 */
1234 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1235 test = LLVMBuildAnd(builder, test, temp, "");
1236 mask = LLVMBuildOr(builder, mask, test, "");
1237 }
1238 else {
1239 /* plane 5 */
1240 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1241 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1242 test = LLVMBuildAnd(builder, test, temp, "");
1243 mask = LLVMBuildOr(builder, mask, test, "");
1244 }
1245 /* plane 6 */
1246 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1247 temp = LLVMBuildShl(builder, temp, shift, "");
1248 test = LLVMBuildAnd(builder, test, temp, "");
1249 mask = LLVMBuildOr(builder, mask, test, "");
1250 }
1251
1252 if (clip_user) {
1253 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1254 LLVMValueRef indices[3];
1255 LLVMValueRef is_nan_or_inf;
1256
1257 /* userclip planes */
1258 while (ucp_enable) {
1259 unsigned plane_idx = ffs(ucp_enable)-1;
1260 ucp_enable &= ~(1 << plane_idx);
1261 plane_idx += 6;
1262
1263 if (have_cd && num_written_clipdistance) {
1264 LLVMValueRef clipdist;
1265 int i;
1266 i = plane_idx - 6;
1267
1268 *have_clipdist = TRUE;
1269 if (i < 4) {
1270 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1271 } else {
1272 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1273 }
1274 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1275 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1276 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1277 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1278 test = LLVMBuildAnd(builder, test, temp, "");
1279 mask = LLVMBuildOr(builder, mask, test, "");
1280 } else {
1281 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1282 indices[0] = lp_build_const_int32(gallivm, 0);
1283 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1284
1285 indices[2] = lp_build_const_int32(gallivm, 0);
1286 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1287 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1288 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1289 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1290
1291 indices[2] = lp_build_const_int32(gallivm, 1);
1292 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1293 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1294 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1295 test = LLVMBuildFMul(builder, planes, cv_y, "");
1296 sum = LLVMBuildFAdd(builder, sum, test, "");
1297
1298 indices[2] = lp_build_const_int32(gallivm, 2);
1299 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1300 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1301 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1302 test = LLVMBuildFMul(builder, planes, cv_z, "");
1303 sum = LLVMBuildFAdd(builder, sum, test, "");
1304
1305 indices[2] = lp_build_const_int32(gallivm, 3);
1306 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1307 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1308 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1309 test = LLVMBuildFMul(builder, planes, cv_w, "");
1310 sum = LLVMBuildFAdd(builder, sum, test, "");
1311
1312 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1313 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1314 test = LLVMBuildAnd(builder, test, temp, "");
1315 mask = LLVMBuildOr(builder, mask, test, "");
1316 }
1317 }
1318 }
1319 return mask;
1320 }
1321
1322
1323 /**
1324 * Returns boolean if any clipping has occurred
1325 * Used zero/non-zero i32 value to represent boolean
1326 */
1327 static LLVMValueRef
1328 clipmask_booli32(struct gallivm_state *gallivm,
1329 const struct lp_type vs_type,
1330 LLVMValueRef clipmask_bool_ptr)
1331 {
1332 LLVMBuilderRef builder = gallivm->builder;
1333 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1334 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1335 LLVMValueRef ret = LLVMConstNull(int32_type);
1336 LLVMValueRef temp;
1337 int i;
1338
1339 /*
1340 * Can do this with log2(vector length) pack instructions and one extract
1341 * (as we don't actually need a or) with sse2 which would be way better.
1342 */
1343 for (i=0; i < vs_type.length; i++) {
1344 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1345 lp_build_const_int32(gallivm, i) , "");
1346 ret = LLVMBuildOr(builder, ret, temp, "");
1347 }
1348 return ret;
1349 }
1350
1351 static LLVMValueRef
1352 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1353 struct lp_build_tgsi_context * bld_base,
1354 boolean is_vindex_indirect,
1355 LLVMValueRef vertex_index,
1356 boolean is_aindex_indirect,
1357 LLVMValueRef attrib_index,
1358 LLVMValueRef swizzle_index)
1359 {
1360 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1361 struct gallivm_state *gallivm = bld_base->base.gallivm;
1362 LLVMBuilderRef builder = gallivm->builder;
1363 LLVMValueRef indices[3];
1364 LLVMValueRef res;
1365 struct lp_type type = bld_base->base.type;
1366
1367 if (is_vindex_indirect || is_aindex_indirect) {
1368 int i;
1369 res = bld_base->base.zero;
1370 for (i = 0; i < type.length; ++i) {
1371 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1372 LLVMValueRef vert_chan_index = vertex_index;
1373 LLVMValueRef attr_chan_index = attrib_index;
1374 LLVMValueRef channel_vec, value;
1375
1376 if (is_vindex_indirect) {
1377 vert_chan_index = LLVMBuildExtractElement(builder,
1378 vertex_index, idx, "");
1379 }
1380 if (is_aindex_indirect) {
1381 attr_chan_index = LLVMBuildExtractElement(builder,
1382 attrib_index, idx, "");
1383 }
1384
1385 indices[0] = vert_chan_index;
1386 indices[1] = attr_chan_index;
1387 indices[2] = swizzle_index;
1388
1389 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1390 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1391 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1392
1393 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1394 }
1395 } else {
1396 indices[0] = vertex_index;
1397 indices[1] = attrib_index;
1398 indices[2] = swizzle_index;
1399
1400 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1401 res = LLVMBuildLoad(builder, res, "");
1402 }
1403
1404 return res;
1405 }
1406
1407 static void
1408 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1409 struct lp_build_tgsi_context * bld_base,
1410 LLVMValueRef (*outputs)[4],
1411 LLVMValueRef emitted_vertices_vec)
1412 {
1413 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1414 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1415 struct gallivm_state *gallivm = variant->gallivm;
1416 LLVMBuilderRef builder = gallivm->builder;
1417 struct lp_type gs_type = bld_base->base.type;
1418 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1419 lp_int_type(gs_type), 0);
1420 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1421 LLVMValueRef next_prim_offset =
1422 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1423 LLVMValueRef io = variant->io_ptr;
1424 unsigned i;
1425 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1426
1427 for (i = 0; i < gs_type.length; ++i) {
1428 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1429 LLVMValueRef currently_emitted =
1430 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1431 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1432 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1433 }
1434
1435 convert_to_aos(gallivm, io, indices,
1436 outputs, clipmask,
1437 gs_info->num_outputs, gs_type,
1438 FALSE);
1439 }
1440
1441 static void
1442 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1443 struct lp_build_tgsi_context * bld_base,
1444 LLVMValueRef verts_per_prim_vec,
1445 LLVMValueRef emitted_prims_vec)
1446 {
1447 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1448 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1449 struct gallivm_state *gallivm = variant->gallivm;
1450 LLVMBuilderRef builder = gallivm->builder;
1451 LLVMValueRef prim_lengts_ptr =
1452 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1453 unsigned i;
1454
1455 for (i = 0; i < bld_base->base.type.length; ++i) {
1456 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1457 LLVMValueRef prims_emitted =
1458 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1459 LLVMValueRef store_ptr;
1460 LLVMValueRef num_vertices =
1461 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1462
1463 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1464 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1465 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1466 LLVMBuildStore(builder, num_vertices, store_ptr);
1467 }
1468 }
1469
1470 static void
1471 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1472 struct lp_build_tgsi_context * bld_base,
1473 LLVMValueRef total_emitted_vertices_vec,
1474 LLVMValueRef emitted_prims_vec)
1475 {
1476 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1477 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1478 struct gallivm_state *gallivm = variant->gallivm;
1479 LLVMBuilderRef builder = gallivm->builder;
1480 LLVMValueRef emitted_verts_ptr =
1481 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1482 LLVMValueRef emitted_prims_ptr =
1483 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1484 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1485
1486 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1487 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1488
1489 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1490 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1491 }
1492
1493 static void
1494 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1495 boolean elts)
1496 {
1497 struct gallivm_state *gallivm = variant->gallivm;
1498 LLVMContextRef context = gallivm->context;
1499 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1500 LLVMTypeRef arg_types[11];
1501 unsigned num_arg_types =
1502 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1503 LLVMTypeRef func_type;
1504 LLVMValueRef context_ptr;
1505 LLVMBasicBlockRef block;
1506 LLVMBuilderRef builder;
1507 char func_name[64];
1508 struct lp_type vs_type;
1509 LLVMValueRef end, start;
1510 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1511 LLVMValueRef vertex_id_offset, start_instance;
1512 LLVMValueRef stride, step, io_itr;
1513 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1514 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1515 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1516 struct draw_context *draw = llvm->draw;
1517 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1518 unsigned i, j;
1519 struct lp_build_context bld;
1520 struct lp_build_loop_state lp_loop;
1521 const int vector_length = lp_native_vector_width / 32;
1522 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1523 LLVMValueRef fetch_max;
1524 struct lp_build_sampler_soa *sampler = 0;
1525 LLVMValueRef ret, clipmask_bool_ptr;
1526 struct draw_llvm_variant_key *key = &variant->key;
1527 /* If geometry shader is present we need to skip both the viewport
1528 * transformation and clipping otherwise the inputs to the geometry
1529 * shader will be incorrect.
1530 * The code can't handle vp transform when vs writes vp index neither
1531 * (though this would be fixable here, but couldn't just broadcast
1532 * the values).
1533 */
1534 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1535 llvm->draw->vs.vertex_shader->info.writes_viewport_index;
1536 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1537 key->clip_z ||
1538 key->clip_user);
1539 LLVMValueRef variant_func;
1540 const unsigned pos = llvm->draw->vs.position_output;
1541 const unsigned cv = llvm->draw->vs.clipvertex_output;
1542 boolean have_clipdist = FALSE;
1543 struct lp_bld_tgsi_system_values system_values;
1544
1545 memset(&system_values, 0, sizeof(system_values));
1546
1547 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1548 variant->shader->variants_cached, elts ? "elts" : "linear");
1549
1550 i = 0;
1551 arg_types[i++] = get_context_ptr_type(variant); /* context */
1552 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1553 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1554 if (elts) {
1555 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1556 arg_types[i++] = int32_type; /* fetch_elt_max */
1557 } else
1558 arg_types[i++] = int32_type; /* start */
1559 arg_types[i++] = int32_type; /* fetch_count / count */
1560 arg_types[i++] = int32_type; /* stride */
1561 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1562 arg_types[i++] = int32_type; /* instance_id */
1563 arg_types[i++] = int32_type; /* vertex_id_offset */
1564 arg_types[i++] = int32_type; /* start_instance */
1565
1566 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1567
1568 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1569
1570 if (elts)
1571 variant->function_elts = variant_func;
1572 else
1573 variant->function = variant_func;
1574
1575 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1576 for (i = 0; i < num_arg_types; ++i)
1577 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1578 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1579 LLVMNoAliasAttribute);
1580
1581 context_ptr = LLVMGetParam(variant_func, 0);
1582 io_ptr = LLVMGetParam(variant_func, 1);
1583 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1584 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1585 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1586 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1587 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1588 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1589
1590 lp_build_name(context_ptr, "context");
1591 lp_build_name(io_ptr, "io");
1592 lp_build_name(vbuffers_ptr, "vbuffers");
1593 lp_build_name(stride, "stride");
1594 lp_build_name(vb_ptr, "vb");
1595 lp_build_name(system_values.instance_id, "instance_id");
1596 lp_build_name(vertex_id_offset, "vertex_id_offset");
1597 lp_build_name(start_instance, "start_instance");
1598
1599 if (elts) {
1600 fetch_elts = LLVMGetParam(variant_func, 3);
1601 fetch_elt_max = LLVMGetParam(variant_func, 4);
1602 fetch_count = LLVMGetParam(variant_func, 5);
1603 lp_build_name(fetch_elts, "fetch_elts");
1604 lp_build_name(fetch_elt_max, "fetch_elt_max");
1605 lp_build_name(fetch_count, "fetch_count");
1606 start = count = NULL;
1607 }
1608 else {
1609 start = LLVMGetParam(variant_func, 3);
1610 count = LLVMGetParam(variant_func, 4);
1611 lp_build_name(start, "start");
1612 lp_build_name(count, "count");
1613 fetch_elts = fetch_count = NULL;
1614 }
1615
1616 /*
1617 * Function body
1618 */
1619
1620 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1621 builder = gallivm->builder;
1622 LLVMPositionBuilderAtEnd(builder, block);
1623
1624 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1625
1626 memset(&vs_type, 0, sizeof vs_type);
1627 vs_type.floating = TRUE; /* floating point values */
1628 vs_type.sign = TRUE; /* values are signed */
1629 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1630 vs_type.width = 32; /* 32-bit float */
1631 vs_type.length = vector_length;
1632
1633 /* hold temporary "bool" clipmask */
1634 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1635 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1636
1637 /* code generated texture sampling */
1638 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1639
1640 if (elts) {
1641 start = zero;
1642 end = fetch_count;
1643 count = fetch_count;
1644 }
1645 else {
1646 end = lp_build_add(&bld, start, count);
1647 }
1648
1649 step = lp_build_const_int32(gallivm, vector_length);
1650
1651 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1652
1653 lp_build_loop_begin(&lp_loop, gallivm, zero);
1654 {
1655 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1656 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1657 LLVMValueRef io;
1658 LLVMValueRef clipmask; /* holds the clipmask value */
1659 LLVMValueRef true_index_array = lp_build_zero(gallivm,
1660 lp_type_uint_vec(32, 32*vector_length));
1661 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1662
1663 io_itr = lp_loop.counter;
1664
1665 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1666 #if DEBUG_STORE
1667 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1668 io_itr, io, lp_loop.counter);
1669 #endif
1670 for (i = 0; i < vector_length; ++i) {
1671 LLVMValueRef vert_index =
1672 LLVMBuildAdd(builder,
1673 lp_loop.counter,
1674 lp_build_const_int32(gallivm, i), "");
1675 LLVMValueRef true_index =
1676 LLVMBuildAdd(builder, start, vert_index, "");
1677
1678 /* make sure we're not out of bounds which can happen
1679 * if fetch_count % 4 != 0, because on the last iteration
1680 * a few of the 4 vertex fetches will be out of bounds */
1681 true_index = lp_build_min(&bld, true_index, fetch_max);
1682
1683 if (elts) {
1684 LLVMValueRef fetch_ptr;
1685 LLVMValueRef index_overflowed;
1686 LLVMValueRef index_ptr =
1687 lp_build_alloca(
1688 gallivm,
1689 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1690 struct lp_build_if_state if_ctx;
1691 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1692 true_index, fetch_elt_max,
1693 "index_overflowed");
1694
1695 lp_build_if(&if_ctx, gallivm, index_overflowed);
1696 {
1697 /* Generate maximum possible index so that
1698 * generate_fetch can treat it just like
1699 * any other overflow and return zeros.
1700 * We don't have to worry about the restart
1701 * primitive index because it has already been
1702 * handled
1703 */
1704 LLVMValueRef val =
1705 lp_build_const_int32(gallivm, 0xffffffff);
1706 LLVMBuildStore(builder, val, index_ptr);
1707 }
1708 lp_build_else(&if_ctx);
1709 {
1710 LLVMValueRef val;
1711 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1712 &true_index, 1, "");
1713 val = LLVMBuildLoad(builder, fetch_ptr, "");
1714 LLVMBuildStore(builder, val, index_ptr);
1715 }
1716 lp_build_endif(&if_ctx);
1717 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1718 }
1719 true_index_array = LLVMBuildInsertElement(
1720 gallivm->builder, true_index_array, true_index,
1721 lp_build_const_int32(gallivm, i), "");
1722
1723 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1724 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1725 LLVMValueRef vb_index =
1726 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1727 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1728 generate_fetch(gallivm, draw, vbuffers_ptr,
1729 &aos_attribs[j][i], velem, vb, true_index,
1730 system_values.instance_id, start_instance);
1731 }
1732 }
1733 convert_to_soa(gallivm, aos_attribs, inputs,
1734 draw->pt.nr_vertex_elements, vs_type);
1735
1736 /* In the paths with elts vertex id has to be unaffected by the
1737 * index bias and because indices inside our elements array have
1738 * already had index bias applied we need to subtract it here to
1739 * get back to the original index.
1740 * in the linear paths vertex id has to be unaffected by the
1741 * original start index and because we abuse the 'start' variable
1742 * to either represent the actual start index or the index at which
1743 * the primitive was split (we split rendering into chunks of at
1744 * most 4095-vertices) we need to back out the original start
1745 * index out of our vertex id here.
1746 */
1747 system_values.basevertex = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm,
1748 lp_type_uint_vec(32, 32*vector_length)),
1749 vertex_id_offset);
1750 system_values.vertex_id = true_index_array;
1751 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1752 system_values.basevertex, "");
1753
1754 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1755 generate_vs(variant,
1756 builder,
1757 vs_type,
1758 outputs,
1759 ptr_aos,
1760 &system_values,
1761 context_ptr,
1762 sampler,
1763 key->clamp_vertex_color);
1764
1765 if (pos != -1 && cv != -1) {
1766 /* store original positions in clip before further manipulation */
1767 store_clip(gallivm, vs_type, io, outputs, FALSE, key->clip_user ? cv : pos);
1768 store_clip(gallivm, vs_type, io, outputs, TRUE, pos);
1769
1770 /* do cliptest */
1771 if (enable_cliptest) {
1772 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1773 /* allocate clipmask, assign it integer type */
1774 clipmask = generate_clipmask(llvm,
1775 gallivm,
1776 vs_type,
1777 outputs,
1778 key->clip_xy,
1779 key->clip_z,
1780 key->clip_user,
1781 key->clip_halfz,
1782 key->ucp_enable,
1783 context_ptr, &have_clipdist);
1784 temp = LLVMBuildOr(builder, clipmask, temp, "");
1785 /* store temporary clipping boolean value */
1786 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1787 }
1788 else {
1789 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1790 }
1791
1792 /* do viewport mapping */
1793 if (!bypass_viewport) {
1794 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1795 }
1796 }
1797 else {
1798 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1799 }
1800
1801 /* store clipmask in vertex header,
1802 * original positions in clip
1803 * and transformed positions in data
1804 */
1805 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1806 vs_info->num_outputs, vs_type,
1807 have_clipdist);
1808 }
1809 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1810
1811 sampler->destroy(sampler);
1812
1813 /* return clipping boolean value for function */
1814 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1815
1816 LLVMBuildRet(builder, ret);
1817
1818 gallivm_verify_function(gallivm, variant_func);
1819 }
1820
1821
1822 struct draw_llvm_variant_key *
1823 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1824 {
1825 unsigned i;
1826 struct draw_llvm_variant_key *key;
1827 struct draw_sampler_static_state *draw_sampler;
1828
1829 key = (struct draw_llvm_variant_key *)store;
1830
1831 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1832
1833 /* Presumably all variants of the shader should have the same
1834 * number of vertex elements - ie the number of shader inputs.
1835 * NOTE: we NEED to store the needed number of needed inputs
1836 * here, not the number of provided elements to match keysize
1837 * (and the offset of sampler state in the key).
1838 */
1839 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1840 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1841
1842 /* will have to rig this up properly later */
1843 key->clip_xy = llvm->draw->clip_xy;
1844 key->clip_z = llvm->draw->clip_z;
1845 key->clip_user = llvm->draw->clip_user;
1846 key->bypass_viewport = llvm->draw->bypass_viewport;
1847 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1848 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1849 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1850 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1851 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1852 key->pad1 = 0;
1853
1854 /* All variants of this shader will have the same value for
1855 * nr_samplers. Not yet trying to compact away holes in the
1856 * sampler array.
1857 */
1858 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1859 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1860 key->nr_sampler_views =
1861 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1862 }
1863 else {
1864 key->nr_sampler_views = key->nr_samplers;
1865 }
1866
1867 draw_sampler = draw_llvm_variant_key_samplers(key);
1868
1869 memcpy(key->vertex_element,
1870 llvm->draw->pt.vertex_element,
1871 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1872
1873 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1874
1875 for (i = 0 ; i < key->nr_samplers; i++) {
1876 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1877 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1878 }
1879 for (i = 0 ; i < key->nr_sampler_views; i++) {
1880 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1881 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1882 }
1883
1884 return key;
1885 }
1886
1887
1888 void
1889 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1890 {
1891 unsigned i;
1892 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1893
1894 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1895 debug_printf("clip_xy = %u\n", key->clip_xy);
1896 debug_printf("clip_z = %u\n", key->clip_z);
1897 debug_printf("clip_user = %u\n", key->clip_user);
1898 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1899 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1900 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1901 debug_printf("has_gs = %u\n", key->has_gs);
1902 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1903
1904 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1905 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1906 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1907 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1908 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1909 }
1910
1911 for (i = 0 ; i < key->nr_sampler_views; i++) {
1912 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1913 }
1914 }
1915
1916
1917 void
1918 draw_llvm_set_mapped_texture(struct draw_context *draw,
1919 unsigned shader_stage,
1920 unsigned sview_idx,
1921 uint32_t width, uint32_t height, uint32_t depth,
1922 uint32_t first_level, uint32_t last_level,
1923 const void *base_ptr,
1924 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1925 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1926 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1927 {
1928 unsigned j;
1929 struct draw_jit_texture *jit_tex;
1930
1931 assert(shader_stage == PIPE_SHADER_VERTEX ||
1932 shader_stage == PIPE_SHADER_GEOMETRY);
1933
1934 if (shader_stage == PIPE_SHADER_VERTEX) {
1935 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1936
1937 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1938 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1939 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1940
1941 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1942 } else {
1943 assert(0);
1944 return;
1945 }
1946
1947 jit_tex->width = width;
1948 jit_tex->height = height;
1949 jit_tex->depth = depth;
1950 jit_tex->first_level = first_level;
1951 jit_tex->last_level = last_level;
1952 jit_tex->base = base_ptr;
1953
1954 for (j = first_level; j <= last_level; j++) {
1955 jit_tex->mip_offsets[j] = mip_offsets[j];
1956 jit_tex->row_stride[j] = row_stride[j];
1957 jit_tex->img_stride[j] = img_stride[j];
1958 }
1959 }
1960
1961
1962 void
1963 draw_llvm_set_sampler_state(struct draw_context *draw,
1964 unsigned shader_type)
1965 {
1966 unsigned i;
1967
1968 if (shader_type == PIPE_SHADER_VERTEX) {
1969 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1970 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1971
1972 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
1973 const struct pipe_sampler_state *s
1974 = draw->samplers[PIPE_SHADER_VERTEX][i];
1975 jit_sam->min_lod = s->min_lod;
1976 jit_sam->max_lod = s->max_lod;
1977 jit_sam->lod_bias = s->lod_bias;
1978 COPY_4V(jit_sam->border_color, s->border_color.f);
1979 }
1980 }
1981 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1982 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1983 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1984
1985 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
1986 const struct pipe_sampler_state *s
1987 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1988 jit_sam->min_lod = s->min_lod;
1989 jit_sam->max_lod = s->max_lod;
1990 jit_sam->lod_bias = s->lod_bias;
1991 COPY_4V(jit_sam->border_color, s->border_color.f);
1992 }
1993 }
1994 }
1995 }
1996
1997
1998 void
1999 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2000 {
2001 struct draw_llvm *llvm = variant->llvm;
2002
2003 gallivm_destroy(variant->gallivm);
2004
2005 remove_from_list(&variant->list_item_local);
2006 variant->shader->variants_cached--;
2007 remove_from_list(&variant->list_item_global);
2008 llvm->nr_variants--;
2009 FREE(variant);
2010 }
2011
2012
2013 /**
2014 * Create LLVM types for various structures.
2015 */
2016 static void
2017 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2018 {
2019 struct gallivm_state *gallivm = var->gallivm;
2020 LLVMTypeRef texture_type, sampler_type, context_type;
2021
2022 texture_type = create_jit_texture_type(gallivm, "texture");
2023 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2024
2025 context_type = create_gs_jit_context_type(gallivm,
2026 var->shader->base.vector_length,
2027 texture_type, sampler_type,
2028 "draw_gs_jit_context");
2029 var->context_ptr_type = LLVMPointerType(context_type, 0);
2030
2031 var->input_array_type = create_gs_jit_input_type(gallivm);
2032 }
2033
2034 static LLVMTypeRef
2035 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2036 {
2037 if (!variant->context_ptr_type)
2038 create_gs_jit_types(variant);
2039 return variant->context_ptr_type;
2040 }
2041
2042 static LLVMValueRef
2043 generate_mask_value(struct draw_gs_llvm_variant *variant,
2044 struct lp_type gs_type)
2045 {
2046 struct gallivm_state *gallivm = variant->gallivm;
2047 LLVMBuilderRef builder = gallivm->builder;
2048 struct lp_type mask_type = lp_int_type(gs_type);
2049 LLVMValueRef num_prims;
2050 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2051 unsigned i;
2052
2053 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2054 variant->num_prims);
2055 for (i = 0; i < gs_type.length; i++) {
2056 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2057 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2058 }
2059 mask_val = lp_build_compare(gallivm, mask_type,
2060 PIPE_FUNC_GREATER, num_prims, mask_val);
2061
2062 return mask_val;
2063 }
2064
2065 static void
2066 draw_gs_llvm_generate(struct draw_llvm *llvm,
2067 struct draw_gs_llvm_variant *variant)
2068 {
2069 struct gallivm_state *gallivm = variant->gallivm;
2070 LLVMContextRef context = gallivm->context;
2071 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2072 LLVMTypeRef arg_types[6];
2073 LLVMTypeRef func_type;
2074 LLVMValueRef variant_func;
2075 LLVMValueRef context_ptr;
2076 LLVMValueRef prim_id_ptr;
2077 LLVMBasicBlockRef block;
2078 LLVMBuilderRef builder;
2079 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2080 struct lp_build_sampler_soa *sampler = 0;
2081 struct lp_build_context bld;
2082 struct lp_bld_tgsi_system_values system_values;
2083 char func_name[64];
2084 struct lp_type gs_type;
2085 unsigned i;
2086 struct draw_gs_llvm_iface gs_iface;
2087 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2088 LLVMValueRef consts_ptr, num_consts_ptr;
2089 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2090 struct lp_build_mask_context mask;
2091 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2092 unsigned vector_length = variant->shader->base.vector_length;
2093
2094 memset(&system_values, 0, sizeof(system_values));
2095
2096 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2097 variant->shader->variants_cached);
2098
2099 assert(variant->vertex_header_ptr_type);
2100
2101 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2102 arg_types[1] = variant->input_array_type; /* input */
2103 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2104 arg_types[3] = int32_type; /* num_prims */
2105 arg_types[4] = int32_type; /* instance_id */
2106 arg_types[5] = LLVMPointerType(
2107 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2108
2109 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2110
2111 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2112
2113 variant->function = variant_func;
2114
2115 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2116
2117 for (i = 0; i < Elements(arg_types); ++i)
2118 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2119 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2120 LLVMNoAliasAttribute);
2121
2122 context_ptr = LLVMGetParam(variant_func, 0);
2123 input_array = LLVMGetParam(variant_func, 1);
2124 io_ptr = LLVMGetParam(variant_func, 2);
2125 num_prims = LLVMGetParam(variant_func, 3);
2126 system_values.instance_id = LLVMGetParam(variant_func, 4);
2127 prim_id_ptr = LLVMGetParam(variant_func, 5);
2128
2129 lp_build_name(context_ptr, "context");
2130 lp_build_name(input_array, "input");
2131 lp_build_name(io_ptr, "io");
2132 lp_build_name(num_prims, "num_prims");
2133 lp_build_name(system_values.instance_id, "instance_id");
2134 lp_build_name(prim_id_ptr, "prim_id_ptr");
2135
2136 variant->context_ptr = context_ptr;
2137 variant->io_ptr = io_ptr;
2138 variant->num_prims = num_prims;
2139
2140 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2141 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2142 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2143 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2144 gs_iface.input = input_array;
2145 gs_iface.variant = variant;
2146
2147 /*
2148 * Function body
2149 */
2150
2151 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2152 builder = gallivm->builder;
2153 LLVMPositionBuilderAtEnd(builder, block);
2154
2155 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2156
2157 memset(&gs_type, 0, sizeof gs_type);
2158 gs_type.floating = TRUE; /* floating point values */
2159 gs_type.sign = TRUE; /* values are signed */
2160 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2161 gs_type.width = 32; /* 32-bit float */
2162 gs_type.length = vector_length;
2163
2164 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2165 num_consts_ptr =
2166 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2167
2168 /* code generated texture sampling */
2169 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2170
2171 mask_val = generate_mask_value(variant, gs_type);
2172 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2173
2174 if (gs_info->uses_primid) {
2175 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2176 }
2177
2178 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2179 tgsi_dump(tokens, 0);
2180 draw_gs_llvm_dump_variant_key(&variant->key);
2181 }
2182
2183 lp_build_tgsi_soa(variant->gallivm,
2184 tokens,
2185 gs_type,
2186 &mask,
2187 consts_ptr,
2188 num_consts_ptr,
2189 &system_values,
2190 NULL,
2191 outputs,
2192 context_ptr,
2193 sampler,
2194 &llvm->draw->gs.geometry_shader->info,
2195 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2196
2197 sampler->destroy(sampler);
2198
2199 lp_build_mask_end(&mask);
2200
2201 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2202
2203 gallivm_verify_function(gallivm, variant_func);
2204 }
2205
2206
2207 struct draw_gs_llvm_variant *
2208 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2209 unsigned num_outputs,
2210 const struct draw_gs_llvm_variant_key *key)
2211 {
2212 struct draw_gs_llvm_variant *variant;
2213 struct llvm_geometry_shader *shader =
2214 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2215 LLVMTypeRef vertex_header;
2216 char module_name[64];
2217
2218 variant = MALLOC(sizeof *variant +
2219 shader->variant_key_size -
2220 sizeof variant->key);
2221 if (variant == NULL)
2222 return NULL;
2223
2224 variant->llvm = llvm;
2225 variant->shader = shader;
2226
2227 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2228 variant->shader->variants_cached);
2229
2230 variant->gallivm = gallivm_create(module_name, llvm->context);
2231
2232 create_gs_jit_types(variant);
2233
2234 memcpy(&variant->key, key, shader->variant_key_size);
2235
2236 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2237
2238 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2239
2240 draw_gs_llvm_generate(llvm, variant);
2241
2242 gallivm_compile_module(variant->gallivm);
2243
2244 variant->jit_func = (draw_gs_jit_func)
2245 gallivm_jit_function(variant->gallivm, variant->function);
2246
2247 gallivm_free_ir(variant->gallivm);
2248
2249 variant->list_item_global.base = variant;
2250 variant->list_item_local.base = variant;
2251 /*variant->no = */shader->variants_created++;
2252 variant->list_item_global.base = variant;
2253
2254 return variant;
2255 }
2256
2257 void
2258 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2259 {
2260 struct draw_llvm *llvm = variant->llvm;
2261
2262 gallivm_destroy(variant->gallivm);
2263
2264 remove_from_list(&variant->list_item_local);
2265 variant->shader->variants_cached--;
2266 remove_from_list(&variant->list_item_global);
2267 llvm->nr_gs_variants--;
2268 FREE(variant);
2269 }
2270
2271 struct draw_gs_llvm_variant_key *
2272 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2273 {
2274 unsigned i;
2275 struct draw_gs_llvm_variant_key *key;
2276 struct draw_sampler_static_state *draw_sampler;
2277
2278 key = (struct draw_gs_llvm_variant_key *)store;
2279
2280 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2281
2282 /* All variants of this shader will have the same value for
2283 * nr_samplers. Not yet trying to compact away holes in the
2284 * sampler array.
2285 */
2286 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2287 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2288 key->nr_sampler_views =
2289 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2290 }
2291 else {
2292 key->nr_sampler_views = key->nr_samplers;
2293 }
2294
2295 draw_sampler = key->samplers;
2296
2297 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2298
2299 for (i = 0 ; i < key->nr_samplers; i++) {
2300 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2301 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2302 }
2303 for (i = 0 ; i < key->nr_sampler_views; i++) {
2304 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2305 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2306 }
2307
2308 return key;
2309 }
2310
2311 void
2312 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2313 {
2314 unsigned i;
2315 struct draw_sampler_static_state *sampler = key->samplers;
2316
2317 for (i = 0 ; i < key->nr_sampler_views; i++) {
2318 debug_printf("sampler[%i].src_format = %s\n", i,
2319 util_format_name(sampler[i].texture_state.format));
2320 }
2321 }