projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw: fix key comparison with uninitialized value
[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 (void) target; /* silence unused var warning for non-debug build */
192 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
193 target, sampler_type,
194 DRAW_JIT_SAMPLER_MIN_LOD);
195 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
196 target, sampler_type,
197 DRAW_JIT_SAMPLER_MAX_LOD);
198 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
199 target, sampler_type,
200 DRAW_JIT_SAMPLER_LOD_BIAS);
201 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
202 target, sampler_type,
203 DRAW_JIT_SAMPLER_BORDER_COLOR);
204
205 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
206
207 return sampler_type;
208 }
209
210
211 /**
212 * Create LLVM type for struct draw_jit_context
213 */
214 static LLVMTypeRef
215 create_jit_context_type(struct gallivm_state *gallivm,
216 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
217 const char *struct_name)
218 {
219 LLVMTargetDataRef target = gallivm->target;
220 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
221 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
222 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
223 LLVMTypeRef context_type;
224
225 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
226 LP_MAX_TGSI_CONST_BUFFERS);
227 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
228 LP_MAX_TGSI_CONST_BUFFERS);
229 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
230 DRAW_TOTAL_CLIP_PLANES), 0);
231 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
232 elem_types[4] = LLVMArrayType(texture_type,
233 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
234 elem_types[5] = LLVMArrayType(sampler_type,
235 PIPE_MAX_SAMPLERS); /* samplers */
236 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
237 Elements(elem_types), 0);
238
239 (void) target; /* silence unused var warning for non-debug build */
240 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
241 target, context_type, DRAW_JIT_CTX_CONSTANTS);
242 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
243 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
244 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
245 target, context_type, DRAW_JIT_CTX_PLANES);
246 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
247 target, context_type, DRAW_JIT_CTX_VIEWPORT);
248 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
249 target, context_type,
250 DRAW_JIT_CTX_TEXTURES);
251 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
252 target, context_type,
253 DRAW_JIT_CTX_SAMPLERS);
254 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
255 target, context_type);
256
257 return context_type;
258 }
259
260
261 /**
262 * Create LLVM type for struct draw_gs_jit_context
263 */
264 static LLVMTypeRef
265 create_gs_jit_context_type(struct gallivm_state *gallivm,
266 unsigned vector_length,
267 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
268 const char *struct_name)
269 {
270 LLVMTargetDataRef target = gallivm->target;
271 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
272 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
273 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
274 LLVMTypeRef context_type;
275
276 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
277 LP_MAX_TGSI_CONST_BUFFERS);
278 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
279 LP_MAX_TGSI_CONST_BUFFERS);
280 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
281 DRAW_TOTAL_CLIP_PLANES), 0);
282 elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
283
284 elem_types[4] = LLVMArrayType(texture_type,
285 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
286 elem_types[5] = LLVMArrayType(sampler_type,
287 PIPE_MAX_SAMPLERS); /* samplers */
288
289 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
290 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
291 vector_length), 0);
292 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
293 vector_length), 0);
294
295 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
296 Elements(elem_types), 0);
297
298 (void) target; /* silence unused var warning for non-debug build */
299 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
300 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
301 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
302 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
303 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
304 target, context_type, DRAW_GS_JIT_CTX_PLANES);
305 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
306 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
307 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
308 target, context_type,
309 DRAW_GS_JIT_CTX_TEXTURES);
310 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
311 target, context_type,
312 DRAW_GS_JIT_CTX_SAMPLERS);
313 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
314 target, context_type,
315 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
316 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
317 target, context_type,
318 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
319 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
320 target, context_type,
321 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
322 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
323 target, context_type);
324
325 return context_type;
326 }
327
328
329 static LLVMTypeRef
330 create_gs_jit_input_type(struct gallivm_state *gallivm)
331 {
332 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
333 LLVMTypeRef input_array;
334
335 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
336 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
337 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
338 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
339
340 return input_array;
341 }
342
343 /**
344 * Create LLVM type for struct pipe_vertex_buffer
345 */
346 static LLVMTypeRef
347 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
348 const char *struct_name)
349 {
350 LLVMTargetDataRef target = gallivm->target;
351 LLVMTypeRef elem_types[4];
352 LLVMTypeRef vb_type;
353
354 elem_types[0] =
355 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
356 elem_types[2] =
357 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
358
359 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
360 Elements(elem_types), 0);
361
362 (void) target; /* silence unused var warning for non-debug build */
363 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
364 target, vb_type, 0);
365 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
366 target, vb_type, 1);
367
368 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
369
370 return vb_type;
371 }
372
373
374 /**
375 * Create LLVM type for struct vertex_header;
376 */
377 static LLVMTypeRef
378 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
379 {
380 LLVMTargetDataRef target = gallivm->target;
381 LLVMTypeRef elem_types[3];
382 LLVMTypeRef vertex_header;
383 char struct_name[24];
384
385 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
386
387 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
388 elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
389 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
390
391 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
392 Elements(elem_types), 0);
393
394 /* these are bit-fields and we can't take address of them
395 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
396 target, vertex_header,
397 DRAW_JIT_VERTEX_CLIPMASK);
398 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
399 target, vertex_header,
400 DRAW_JIT_VERTEX_EDGEFLAG);
401 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
402 target, vertex_header,
403 DRAW_JIT_VERTEX_PAD);
404 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
405 target, vertex_header,
406 DRAW_JIT_VERTEX_VERTEX_ID);
407 */
408 (void) target; /* silence unused var warning for non-debug build */
409 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
410 target, vertex_header,
411 DRAW_JIT_VERTEX_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)
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 NULL,
629 draw_sampler,
630 &llvm->draw->vs.vertex_shader->info,
631 NULL);
632
633 {
634 LLVMValueRef out;
635 unsigned chan, attrib;
636 struct lp_build_context bld;
637 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
638 lp_build_context_init(&bld, variant->gallivm, vs_type);
639
640 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
641 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
642 if (outputs[attrib][chan]) {
643 switch (info->output_semantic_name[attrib]) {
644 case TGSI_SEMANTIC_COLOR:
645 case TGSI_SEMANTIC_BCOLOR:
646 if (clamp_vertex_color) {
647 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
648 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
649 LLVMBuildStore(builder, out, outputs[attrib][chan]);
650 }
651 break;
652 }
653 }
654 }
655 }
656 }
657 }
658
659 static void
660 generate_fetch(struct gallivm_state *gallivm,
661 struct draw_context *draw,
662 LLVMValueRef vbuffers_ptr,
663 LLVMValueRef *res,
664 struct pipe_vertex_element *velem,
665 LLVMValueRef vbuf,
666 LLVMValueRef index,
667 LLVMValueRef instance_id,
668 LLVMValueRef start_instance)
669 {
670 const struct util_format_description *format_desc =
671 util_format_description(velem->src_format);
672 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
673 LLVMBuilderRef builder = gallivm->builder;
674 LLVMValueRef indices =
675 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
676 velem->vertex_buffer_index, 0);
677 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
678 &indices, 1, "");
679 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
680 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
681 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
682 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
683 LLVMValueRef stride;
684 LLVMValueRef buffer_overflowed;
685 LLVMValueRef needed_buffer_size;
686 LLVMValueRef temp_ptr =
687 lp_build_alloca(gallivm,
688 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
689 LLVMValueRef ofbit = NULL;
690 struct lp_build_if_state if_ctx;
691
692 if (velem->instance_divisor) {
693 /* Index is equal to the start instance plus the number of current
694 * instance divided by the divisor. In this case we compute it as:
695 * index = start_instance + (instance_id / divisor)
696 */
697 LLVMValueRef current_instance;
698 current_instance = LLVMBuildUDiv(builder, instance_id,
699 lp_build_const_int32(gallivm, velem->instance_divisor),
700 "instance_divisor");
701 index = lp_build_uadd_overflow(gallivm, start_instance,
702 current_instance, &ofbit);
703 }
704
705 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
706 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
707 stride = lp_build_uadd_overflow(
708 gallivm, stride,
709 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
710 needed_buffer_size = lp_build_uadd_overflow(
711 gallivm, stride,
712 lp_build_const_int32(gallivm,
713 util_format_get_blocksize(velem->src_format)),
714 &ofbit);
715
716 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
717 needed_buffer_size, buffer_size,
718 "buffer_overflowed");
719 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
720 #if 0
721 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
722 index, vb_stride);
723 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
724 vb_buffer_offset,
725 lp_build_const_int32(gallivm, velem->src_offset));
726 lp_build_print_value(gallivm, " blocksize = ",
727 lp_build_const_int32(
728 gallivm,
729 util_format_get_blocksize(velem->src_format)));
730 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
731 lp_build_printf(gallivm, " stride = %u\n", stride);
732 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
733 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
734 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
735 #endif
736
737 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
738 {
739 LLVMValueRef val =
740 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
741 LLVMBuildStore(builder, val, temp_ptr);
742 }
743 lp_build_else(&if_ctx);
744 {
745 LLVMValueRef val;
746 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
747
748 val = lp_build_fetch_rgba_aos(gallivm,
749 format_desc,
750 lp_float32_vec4_type(),
751 FALSE,
752 map_ptr,
753 zero, zero, zero,
754 NULL);
755 LLVMBuildStore(builder, val, temp_ptr);
756 }
757 lp_build_endif(&if_ctx);
758
759 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
760 }
761
762 static void
763 convert_to_soa(struct gallivm_state *gallivm,
764 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
765 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
766 unsigned num_attribs, const struct lp_type soa_type)
767 {
768 unsigned i, j, k;
769 struct lp_type aos_channel_type = soa_type;
770
771 debug_assert(TGSI_NUM_CHANNELS == 4);
772 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
773
774 aos_channel_type.length >>= 1;
775
776 for (i = 0; i < num_attribs; ++i) {
777 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
778 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
779
780 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
781 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
782
783 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
784
785 for (k = 0; k < pixels_per_channel; ++k) {
786 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
787 }
788
789 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
790 }
791
792 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
793 }
794 }
795
796
797 static void
798 store_aos(struct gallivm_state *gallivm,
799 LLVMValueRef io_ptr,
800 LLVMValueRef index,
801 LLVMValueRef value)
802 {
803 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
804 LLVMBuilderRef builder = gallivm->builder;
805 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
806 LLVMValueRef indices[3];
807
808 indices[0] = lp_build_const_int32(gallivm, 0);
809 indices[1] = index;
810 indices[2] = lp_build_const_int32(gallivm, 0);
811
812 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
813 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
814
815 #if DEBUG_STORE
816 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
817 #endif
818
819 /* Unaligned store due to the vertex header */
820 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
821 }
822
823 /**
824 * Adjust the mask to architecture endianess. The mask will the store in struct:
825 *
826 * struct vertex_header {
827 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
828 * unsigned edgeflag:1;
829 * unsigned pad:1;
830 * unsigned vertex_id:16;
831 * [...]
832 * }
833 *
834 * On little-endian machine nothing needs to done, however on bit-endian machine
835 * the mask's fields need to be adjusted with the algorithm:
836 *
837 * uint32_t reverse (uint32_t x)
838 * {
839 * return (x >> 16) | // vertex_id
840 * ((x & 0x3fff) << 18) | // clipmask
841 * ((x & 0x4000) << 3) | // pad
842 * ((x & 0x8000) << 1); // edgeflag
843 * }
844 */
845 static LLVMValueRef
846 adjust_mask(struct gallivm_state *gallivm,
847 LLVMValueRef mask)
848 {
849 #ifdef PIPE_ARCH_BIG_ENDIAN
850 LLVMBuilderRef builder = gallivm->builder;
851 LLVMValueRef vertex_id;
852 LLVMValueRef clipmask;
853 LLVMValueRef pad;
854 LLVMValueRef edgeflag;
855
856 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
857 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
858 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
859 if (0) {
860 pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
861 pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
862 }
863 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
864 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
865
866 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
867 if (0) {
868 mask = LLVMBuildOr(builder, mask, pad, "");
869 }
870 mask = LLVMBuildOr(builder, mask, edgeflag, "");
871 #endif
872 return mask;
873 }
874
875 static void
876 store_aos_array(struct gallivm_state *gallivm,
877 struct lp_type soa_type,
878 LLVMValueRef io_ptr,
879 LLVMValueRef *indices,
880 LLVMValueRef* aos,
881 int attrib,
882 int num_outputs,
883 LLVMValueRef clipmask,
884 boolean need_edgeflag)
885 {
886 LLVMBuilderRef builder = gallivm->builder;
887 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
888 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
889 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
890 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
891 int vector_length = soa_type.length;
892 int i;
893
894 debug_assert(TGSI_NUM_CHANNELS == 4);
895
896 for (i = 0; i < vector_length; i++) {
897 linear_inds[i] = lp_build_const_int32(gallivm, i);
898 if (indices) {
899 inds[i] = indices[i];
900 } else {
901 inds[i] = linear_inds[i];
902 }
903 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
904 }
905
906 if (attrib == 0) {
907 /* store vertex header for each of the n vertices */
908 LLVMValueRef val, cliptmp;
909 int vertex_id_pad_edgeflag;
910
911 /* If this assertion fails, it means we need to update the bit twidding
912 * code here. See struct vertex_header in draw_private.h.
913 */
914 assert(DRAW_TOTAL_CLIP_PLANES==14);
915 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
916 if (!need_edgeflag) {
917 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
918 }
919 else {
920 vertex_id_pad_edgeflag = (0xffff << 16);
921 }
922 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
923 vertex_id_pad_edgeflag);
924 /* OR with the clipmask */
925 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
926 for (i = 0; i < vector_length; i++) {
927 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
928 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
929 val = adjust_mask(gallivm, val);
930 #if DEBUG_STORE
931 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
932 io_ptrs[i], inds[i], val);
933 #endif
934 LLVMBuildStore(builder, val, id_ptr);
935 }
936 }
937
938 /* store for each of the n vertices */
939 for (i = 0; i < vector_length; i++) {
940 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
941 }
942 }
943
944
945 static void
946 convert_to_aos(struct gallivm_state *gallivm,
947 LLVMValueRef io,
948 LLVMValueRef *indices,
949 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
950 LLVMValueRef clipmask,
951 int num_outputs,
952 struct lp_type soa_type,
953 boolean need_edgeflag)
954 {
955 LLVMBuilderRef builder = gallivm->builder;
956 unsigned chan, attrib, i;
957
958 #if DEBUG_STORE
959 lp_build_printf(gallivm, " # storing begin\n");
960 #endif
961 for (attrib = 0; attrib < num_outputs; ++attrib) {
962 LLVMValueRef soa[TGSI_NUM_CHANNELS];
963 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
964 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
965 if (outputs[attrib][chan]) {
966 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
967 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
968 #if DEBUG_STORE
969 lp_build_printf(gallivm, "output %d : %d ",
970 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
971 attrib, 0),
972 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
973 chan, 0));
974 lp_build_print_value(gallivm, "val = ", out);
975 {
976 LLVMValueRef iv =
977 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
978
979 lp_build_print_value(gallivm, " ival = ", iv);
980 }
981 #endif
982 soa[chan] = out;
983 }
984 else {
985 soa[chan] = 0;
986 }
987 }
988
989
990 if (soa_type.length == TGSI_NUM_CHANNELS) {
991 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
992 } else {
993 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
994
995 for (i = 0; i < soa_type.length; ++i) {
996 aos[i] = lp_build_extract_range(gallivm,
997 soa[i % TGSI_NUM_CHANNELS],
998 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
999 TGSI_NUM_CHANNELS);
1000 }
1001 }
1002
1003 store_aos_array(gallivm,
1004 soa_type,
1005 io, indices,
1006 aos,
1007 attrib,
1008 num_outputs,
1009 clipmask,
1010 need_edgeflag);
1011 }
1012 #if DEBUG_STORE
1013 lp_build_printf(gallivm, " # storing end\n");
1014 #endif
1015 }
1016
1017
1018 /**
1019 * Stores original vertex positions in clip coordinates
1020 */
1021 static void
1022 store_clip(struct gallivm_state *gallivm,
1023 const struct lp_type vs_type,
1024 LLVMValueRef io_ptr,
1025 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1026 int idx)
1027 {
1028 LLVMBuilderRef builder = gallivm->builder;
1029 LLVMValueRef soa[4];
1030 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1031 LLVMValueRef indices[2];
1032 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1033 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1034 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1035 LLVMTypeRef clip_ptr_type =
1036 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1037 4), 0);
1038 int i, j;
1039
1040 indices[0] =
1041 indices[1] = lp_build_const_int32(gallivm, 0);
1042
1043 for (i = 0; i < vs_type.length; i++) {
1044 inds[i] = lp_build_const_int32(gallivm, i);
1045 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1046 }
1047
1048 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1049 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1050 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1051 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1052
1053 for (i = 0; i < vs_type.length; i++) {
1054 clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1055 }
1056
1057 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1058 for (i = 0; i < vs_type.length; ++i) {
1059 aos[i] = lp_build_extract_range(gallivm,
1060 soa[i % TGSI_NUM_CHANNELS],
1061 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1062 TGSI_NUM_CHANNELS);
1063 }
1064
1065 for (j = 0; j < vs_type.length; j++) {
1066 LLVMValueRef clip_ptr;
1067
1068 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1069 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1070
1071 /* Unaligned store */
1072 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1073 }
1074 }
1075
1076
1077 /**
1078 * Transforms the outputs for viewport mapping
1079 */
1080 static void
1081 generate_viewport(struct draw_llvm_variant *variant,
1082 LLVMBuilderRef builder,
1083 struct lp_type vs_type,
1084 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1085 LLVMValueRef context_ptr)
1086 {
1087 int i;
1088 struct gallivm_state *gallivm = variant->gallivm;
1089 struct lp_type f32_type = vs_type;
1090 const unsigned pos = variant->llvm->draw->vs.position_output;
1091 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1092 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1093 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1094 LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1095
1096 /* We treat pipe_viewport_state as a float array */
1097 const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1098 const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1099
1100 /* for 1/w convention*/
1101 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1102 LLVMBuildStore(builder, out3, outputs[pos][3]);
1103
1104 /* Viewport Mapping */
1105 for (i=0; i<3; i++) {
1106 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1107 LLVMValueRef scale;
1108 LLVMValueRef trans;
1109 LLVMValueRef scale_i;
1110 LLVMValueRef trans_i;
1111 LLVMValueRef index;
1112
1113 index = lp_build_const_int32(gallivm, i + scale_index_offset);
1114 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1115
1116 index = lp_build_const_int32(gallivm, i + trans_index_offset);
1117 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1118
1119 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1120 LLVMBuildLoad(builder, scale_i, "scale"));
1121 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1122 LLVMBuildLoad(builder, trans_i, "trans"));
1123
1124 /* divide by w */
1125 out = LLVMBuildFMul(builder, out, out3, "");
1126 /* mult by scale */
1127 out = LLVMBuildFMul(builder, out, scale, "");
1128 /* add translation */
1129 out = LLVMBuildFAdd(builder, out, trans, "");
1130
1131 /* store transformed outputs */
1132 LLVMBuildStore(builder, out, outputs[pos][i]);
1133 }
1134
1135 }
1136
1137
1138 /**
1139 * Returns clipmask as nxi32 bitmask for the n vertices
1140 */
1141 static LLVMValueRef
1142 generate_clipmask(struct draw_llvm *llvm,
1143 struct gallivm_state *gallivm,
1144 struct lp_type vs_type,
1145 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1146 struct draw_llvm_variant_key *key,
1147 LLVMValueRef context_ptr,
1148 boolean *have_clipdist)
1149 {
1150 LLVMBuilderRef builder = gallivm->builder;
1151 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1152 LLVMValueRef test, temp;
1153 LLVMValueRef zero, shift;
1154 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1155 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1156 LLVMValueRef plane1, planes, plane_ptr, sum;
1157 struct lp_type f32_type = vs_type;
1158 struct lp_type i32_type = lp_int_type(vs_type);
1159 const unsigned pos = llvm->draw->vs.position_output;
1160 const unsigned cv = llvm->draw->vs.clipvertex_output;
1161 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1162 boolean have_cd = false;
1163 boolean clip_user = key->clip_user;
1164 unsigned ucp_enable = key->ucp_enable;
1165 unsigned cd[2];
1166
1167 cd[0] = llvm->draw->vs.clipdistance_output[0];
1168 cd[1] = llvm->draw->vs.clipdistance_output[1];
1169
1170 if (cd[0] != pos || cd[1] != pos)
1171 have_cd = true;
1172
1173 if (num_written_clipdistance && !clip_user) {
1174 clip_user = true;
1175 ucp_enable = (1 << num_written_clipdistance) - 1;
1176 }
1177
1178 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1179 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1180 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1181 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1182
1183 /*
1184 * load clipvertex and position from correct locations.
1185 * if they are the same just load them once.
1186 */
1187 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1188 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1189 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1190 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1191
1192 if (clip_user && cv != pos) {
1193 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1194 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1195 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1196 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1197 } else {
1198 cv_x = pos_x;
1199 cv_y = pos_y;
1200 cv_z = pos_z;
1201 cv_w = pos_w;
1202 }
1203
1204 /*
1205 * Be careful with the comparisons and NaNs (using llvm's unordered
1206 * comparisons here).
1207 */
1208 /* Cliptest, for hardwired planes */
1209 /*
1210 * XXX should take guardband into account (currently not in key).
1211 * Otherwise might run the draw pipeline stages for nothing.
1212 */
1213 if (key->clip_xy) {
1214 /* plane 1 */
1215 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1216 temp = shift;
1217 test = LLVMBuildAnd(builder, test, temp, "");
1218 mask = test;
1219
1220 /* plane 2 */
1221 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1222 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1223 temp = LLVMBuildShl(builder, temp, shift, "");
1224 test = LLVMBuildAnd(builder, test, temp, "");
1225 mask = LLVMBuildOr(builder, mask, test, "");
1226
1227 /* plane 3 */
1228 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1229 temp = LLVMBuildShl(builder, temp, shift, "");
1230 test = LLVMBuildAnd(builder, test, temp, "");
1231 mask = LLVMBuildOr(builder, mask, test, "");
1232
1233 /* plane 4 */
1234 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1235 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1236 temp = LLVMBuildShl(builder, temp, shift, "");
1237 test = LLVMBuildAnd(builder, test, temp, "");
1238 mask = LLVMBuildOr(builder, mask, test, "");
1239 }
1240
1241 if (key->clip_z) {
1242 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1243 if (key->clip_halfz) {
1244 /* plane 5 */
1245 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1246 test = LLVMBuildAnd(builder, test, temp, "");
1247 mask = LLVMBuildOr(builder, mask, test, "");
1248 }
1249 else {
1250 /* plane 5 */
1251 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1252 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1253 test = LLVMBuildAnd(builder, test, temp, "");
1254 mask = LLVMBuildOr(builder, mask, test, "");
1255 }
1256 /* plane 6 */
1257 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1258 temp = LLVMBuildShl(builder, temp, shift, "");
1259 test = LLVMBuildAnd(builder, test, temp, "");
1260 mask = LLVMBuildOr(builder, mask, test, "");
1261 }
1262
1263 if (clip_user) {
1264 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1265 LLVMValueRef indices[3];
1266 LLVMValueRef is_nan_or_inf;
1267
1268 /* userclip planes */
1269 while (ucp_enable) {
1270 unsigned plane_idx = ffs(ucp_enable)-1;
1271 ucp_enable &= ~(1 << plane_idx);
1272 plane_idx += 6;
1273
1274 if (have_cd && num_written_clipdistance) {
1275 LLVMValueRef clipdist;
1276 int i;
1277 i = plane_idx - 6;
1278
1279 *have_clipdist = TRUE;
1280 if (i < 4) {
1281 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1282 } else {
1283 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1284 }
1285 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1286 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1287 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1288 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1289 test = LLVMBuildAnd(builder, test, temp, "");
1290 mask = LLVMBuildOr(builder, mask, test, "");
1291 } else {
1292 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1293 indices[0] = lp_build_const_int32(gallivm, 0);
1294 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1295
1296 indices[2] = lp_build_const_int32(gallivm, 0);
1297 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1298 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1299 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1300 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1301
1302 indices[2] = lp_build_const_int32(gallivm, 1);
1303 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1304 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1305 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1306 test = LLVMBuildFMul(builder, planes, cv_y, "");
1307 sum = LLVMBuildFAdd(builder, sum, test, "");
1308
1309 indices[2] = lp_build_const_int32(gallivm, 2);
1310 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1311 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1312 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1313 test = LLVMBuildFMul(builder, planes, cv_z, "");
1314 sum = LLVMBuildFAdd(builder, sum, test, "");
1315
1316 indices[2] = lp_build_const_int32(gallivm, 3);
1317 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1318 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1319 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1320 test = LLVMBuildFMul(builder, planes, cv_w, "");
1321 sum = LLVMBuildFAdd(builder, sum, test, "");
1322
1323 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1324 temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1325 test = LLVMBuildAnd(builder, test, temp, "");
1326 mask = LLVMBuildOr(builder, mask, test, "");
1327 }
1328 }
1329 }
1330 if (key->need_edgeflags) {
1331 /*
1332 * This isn't really part of clipmask but stored the same in vertex
1333 * header later, so do it here.
1334 */
1335 unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1336 LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1337 LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1338 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1339 temp = lp_build_const_int_vec(gallivm, i32_type,
1340 1LL << DRAW_TOTAL_CLIP_PLANES);
1341 test = LLVMBuildAnd(builder, test, temp, "");
1342 mask = LLVMBuildOr(builder, mask, test, "");
1343 }
1344 return mask;
1345 }
1346
1347
1348 /**
1349 * Returns boolean if any clipping has occurred
1350 * Used zero/non-zero i32 value to represent boolean
1351 */
1352 static LLVMValueRef
1353 clipmask_booli32(struct gallivm_state *gallivm,
1354 const struct lp_type vs_type,
1355 LLVMValueRef clipmask_bool_ptr,
1356 boolean edgeflag_in_clipmask)
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 * We need to invert the edgeflag bit from the clipmask here
1367 * (because the result is really if we want to run the pipeline or not
1368 * and we (may) need it if edgeflag was 0).
1369 */
1370 if (edgeflag_in_clipmask) {
1371 struct lp_type i32_type = lp_int_type(vs_type);
1372 LLVMValueRef edge = lp_build_const_int_vec(gallivm, i32_type,
1373 1LL << DRAW_TOTAL_CLIP_PLANES);
1374 clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1375 }
1376 /*
1377 * Could do much better with just cmp/movmskps.
1378 */
1379 for (i=0; i < vs_type.length; i++) {
1380 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1381 lp_build_const_int32(gallivm, i) , "");
1382 ret = LLVMBuildOr(builder, ret, temp, "");
1383 }
1384 return ret;
1385 }
1386
1387 static LLVMValueRef
1388 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1389 struct lp_build_tgsi_context * bld_base,
1390 boolean is_vindex_indirect,
1391 LLVMValueRef vertex_index,
1392 boolean is_aindex_indirect,
1393 LLVMValueRef attrib_index,
1394 LLVMValueRef swizzle_index)
1395 {
1396 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1397 struct gallivm_state *gallivm = bld_base->base.gallivm;
1398 LLVMBuilderRef builder = gallivm->builder;
1399 LLVMValueRef indices[3];
1400 LLVMValueRef res;
1401 struct lp_type type = bld_base->base.type;
1402
1403 if (is_vindex_indirect || is_aindex_indirect) {
1404 int i;
1405 res = bld_base->base.zero;
1406 for (i = 0; i < type.length; ++i) {
1407 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1408 LLVMValueRef vert_chan_index = vertex_index;
1409 LLVMValueRef attr_chan_index = attrib_index;
1410 LLVMValueRef channel_vec, value;
1411
1412 if (is_vindex_indirect) {
1413 vert_chan_index = LLVMBuildExtractElement(builder,
1414 vertex_index, idx, "");
1415 }
1416 if (is_aindex_indirect) {
1417 attr_chan_index = LLVMBuildExtractElement(builder,
1418 attrib_index, idx, "");
1419 }
1420
1421 indices[0] = vert_chan_index;
1422 indices[1] = attr_chan_index;
1423 indices[2] = swizzle_index;
1424
1425 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1426 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1427 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1428
1429 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1430 }
1431 } else {
1432 indices[0] = vertex_index;
1433 indices[1] = attrib_index;
1434 indices[2] = swizzle_index;
1435
1436 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1437 res = LLVMBuildLoad(builder, res, "");
1438 }
1439
1440 return res;
1441 }
1442
1443 static void
1444 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1445 struct lp_build_tgsi_context * bld_base,
1446 LLVMValueRef (*outputs)[4],
1447 LLVMValueRef emitted_vertices_vec)
1448 {
1449 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1450 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1451 struct gallivm_state *gallivm = variant->gallivm;
1452 LLVMBuilderRef builder = gallivm->builder;
1453 struct lp_type gs_type = bld_base->base.type;
1454 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1455 lp_int_type(gs_type), 0);
1456 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1457 LLVMValueRef next_prim_offset =
1458 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1459 LLVMValueRef io = variant->io_ptr;
1460 unsigned i;
1461 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1462
1463 for (i = 0; i < gs_type.length; ++i) {
1464 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1465 LLVMValueRef currently_emitted =
1466 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1467 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1468 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1469 }
1470
1471 convert_to_aos(gallivm, io, indices,
1472 outputs, clipmask,
1473 gs_info->num_outputs, gs_type,
1474 FALSE);
1475 }
1476
1477 static void
1478 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1479 struct lp_build_tgsi_context * bld_base,
1480 LLVMValueRef verts_per_prim_vec,
1481 LLVMValueRef emitted_prims_vec)
1482 {
1483 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1484 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1485 struct gallivm_state *gallivm = variant->gallivm;
1486 LLVMBuilderRef builder = gallivm->builder;
1487 LLVMValueRef prim_lengts_ptr =
1488 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1489 unsigned i;
1490
1491 for (i = 0; i < bld_base->base.type.length; ++i) {
1492 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1493 LLVMValueRef prims_emitted =
1494 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1495 LLVMValueRef store_ptr;
1496 LLVMValueRef num_vertices =
1497 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1498
1499 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1500 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1501 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1502 LLVMBuildStore(builder, num_vertices, store_ptr);
1503 }
1504 }
1505
1506 static void
1507 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1508 struct lp_build_tgsi_context * bld_base,
1509 LLVMValueRef total_emitted_vertices_vec,
1510 LLVMValueRef emitted_prims_vec)
1511 {
1512 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1513 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1514 struct gallivm_state *gallivm = variant->gallivm;
1515 LLVMBuilderRef builder = gallivm->builder;
1516 LLVMValueRef emitted_verts_ptr =
1517 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1518 LLVMValueRef emitted_prims_ptr =
1519 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1520 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1521
1522 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1523 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1524
1525 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1526 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1527 }
1528
1529 static void
1530 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1531 boolean elts)
1532 {
1533 struct gallivm_state *gallivm = variant->gallivm;
1534 LLVMContextRef context = gallivm->context;
1535 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1536 LLVMTypeRef arg_types[11];
1537 unsigned num_arg_types =
1538 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1539 LLVMTypeRef func_type;
1540 LLVMValueRef context_ptr;
1541 LLVMBasicBlockRef block;
1542 LLVMBuilderRef builder;
1543 char func_name[64];
1544 struct lp_type vs_type;
1545 LLVMValueRef end, start;
1546 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1547 LLVMValueRef vertex_id_offset, start_instance;
1548 LLVMValueRef stride, step, io_itr;
1549 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1550 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1551 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1552 struct draw_context *draw = llvm->draw;
1553 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1554 unsigned i, j;
1555 struct lp_build_context bld;
1556 struct lp_build_loop_state lp_loop;
1557 const int vector_length = lp_native_vector_width / 32;
1558 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1559 LLVMValueRef fetch_max;
1560 struct lp_build_sampler_soa *sampler = 0;
1561 LLVMValueRef ret, clipmask_bool_ptr;
1562 struct draw_llvm_variant_key *key = &variant->key;
1563 /* If geometry shader is present we need to skip both the viewport
1564 * transformation and clipping otherwise the inputs to the geometry
1565 * shader will be incorrect.
1566 * The code can't handle vp transform when vs writes vp index neither
1567 * (though this would be fixable here, but couldn't just broadcast
1568 * the values).
1569 */
1570 const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1571 llvm->draw->vs.vertex_shader->info.writes_viewport_index;
1572 const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1573 key->clip_z ||
1574 key->clip_user ||
1575 key->need_edgeflags);
1576 LLVMValueRef variant_func;
1577 const unsigned pos = llvm->draw->vs.position_output;
1578 const unsigned cv = llvm->draw->vs.clipvertex_output;
1579 boolean have_clipdist = FALSE;
1580 struct lp_bld_tgsi_system_values system_values;
1581
1582 memset(&system_values, 0, sizeof(system_values));
1583
1584 util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
1585 variant->shader->variants_cached, elts ? "elts" : "linear");
1586
1587 i = 0;
1588 arg_types[i++] = get_context_ptr_type(variant); /* context */
1589 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1590 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1591 if (elts) {
1592 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1593 arg_types[i++] = int32_type; /* fetch_elt_max */
1594 } else
1595 arg_types[i++] = int32_type; /* start */
1596 arg_types[i++] = int32_type; /* fetch_count / count */
1597 arg_types[i++] = int32_type; /* stride */
1598 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1599 arg_types[i++] = int32_type; /* instance_id */
1600 arg_types[i++] = int32_type; /* vertex_id_offset */
1601 arg_types[i++] = int32_type; /* start_instance */
1602
1603 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1604
1605 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1606
1607 if (elts)
1608 variant->function_elts = variant_func;
1609 else
1610 variant->function = variant_func;
1611
1612 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1613 for (i = 0; i < num_arg_types; ++i)
1614 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1615 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1616 LLVMNoAliasAttribute);
1617
1618 context_ptr = LLVMGetParam(variant_func, 0);
1619 io_ptr = LLVMGetParam(variant_func, 1);
1620 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1621 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1622 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1623 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1624 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1625 start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
1626
1627 lp_build_name(context_ptr, "context");
1628 lp_build_name(io_ptr, "io");
1629 lp_build_name(vbuffers_ptr, "vbuffers");
1630 lp_build_name(stride, "stride");
1631 lp_build_name(vb_ptr, "vb");
1632 lp_build_name(system_values.instance_id, "instance_id");
1633 lp_build_name(vertex_id_offset, "vertex_id_offset");
1634 lp_build_name(start_instance, "start_instance");
1635
1636 if (elts) {
1637 fetch_elts = LLVMGetParam(variant_func, 3);
1638 fetch_elt_max = LLVMGetParam(variant_func, 4);
1639 fetch_count = LLVMGetParam(variant_func, 5);
1640 lp_build_name(fetch_elts, "fetch_elts");
1641 lp_build_name(fetch_elt_max, "fetch_elt_max");
1642 lp_build_name(fetch_count, "fetch_count");
1643 start = count = NULL;
1644 }
1645 else {
1646 start = LLVMGetParam(variant_func, 3);
1647 count = LLVMGetParam(variant_func, 4);
1648 lp_build_name(start, "start");
1649 lp_build_name(count, "count");
1650 fetch_elts = fetch_count = NULL;
1651 }
1652
1653 /*
1654 * Function body
1655 */
1656
1657 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1658 builder = gallivm->builder;
1659 LLVMPositionBuilderAtEnd(builder, block);
1660
1661 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1662
1663 memset(&vs_type, 0, sizeof vs_type);
1664 vs_type.floating = TRUE; /* floating point values */
1665 vs_type.sign = TRUE; /* values are signed */
1666 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1667 vs_type.width = 32; /* 32-bit float */
1668 vs_type.length = vector_length;
1669
1670 /* hold temporary "bool" clipmask */
1671 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1672 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1673
1674 /* code generated texture sampling */
1675 sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1676
1677 if (elts) {
1678 start = zero;
1679 end = fetch_count;
1680 count = fetch_count;
1681 }
1682 else {
1683 end = lp_build_add(&bld, start, count);
1684 }
1685
1686 step = lp_build_const_int32(gallivm, vector_length);
1687
1688 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1689
1690 lp_build_loop_begin(&lp_loop, gallivm, zero);
1691 {
1692 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1693 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1694 LLVMValueRef io;
1695 LLVMValueRef clipmask; /* holds the clipmask value */
1696 LLVMValueRef true_index_array = lp_build_zero(gallivm,
1697 lp_type_uint_vec(32, 32*vector_length));
1698 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1699
1700 io_itr = lp_loop.counter;
1701
1702 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1703 #if DEBUG_STORE
1704 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1705 io_itr, io, lp_loop.counter);
1706 #endif
1707 for (i = 0; i < vector_length; ++i) {
1708 LLVMValueRef vert_index =
1709 LLVMBuildAdd(builder,
1710 lp_loop.counter,
1711 lp_build_const_int32(gallivm, i), "");
1712 LLVMValueRef true_index =
1713 LLVMBuildAdd(builder, start, vert_index, "");
1714
1715 /* make sure we're not out of bounds which can happen
1716 * if fetch_count % 4 != 0, because on the last iteration
1717 * a few of the 4 vertex fetches will be out of bounds */
1718 true_index = lp_build_min(&bld, true_index, fetch_max);
1719
1720 if (elts) {
1721 LLVMValueRef fetch_ptr;
1722 LLVMValueRef index_overflowed;
1723 LLVMValueRef index_ptr =
1724 lp_build_alloca(
1725 gallivm,
1726 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1727 struct lp_build_if_state if_ctx;
1728 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1729 true_index, fetch_elt_max,
1730 "index_overflowed");
1731
1732 lp_build_if(&if_ctx, gallivm, index_overflowed);
1733 {
1734 /* Generate maximum possible index so that
1735 * generate_fetch can treat it just like
1736 * any other overflow and return zeros.
1737 * We don't have to worry about the restart
1738 * primitive index because it has already been
1739 * handled
1740 */
1741 LLVMValueRef val =
1742 lp_build_const_int32(gallivm, 0xffffffff);
1743 LLVMBuildStore(builder, val, index_ptr);
1744 }
1745 lp_build_else(&if_ctx);
1746 {
1747 LLVMValueRef val;
1748 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1749 &true_index, 1, "");
1750 val = LLVMBuildLoad(builder, fetch_ptr, "");
1751 LLVMBuildStore(builder, val, index_ptr);
1752 }
1753 lp_build_endif(&if_ctx);
1754 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1755 }
1756 true_index_array = LLVMBuildInsertElement(
1757 gallivm->builder, true_index_array, true_index,
1758 lp_build_const_int32(gallivm, i), "");
1759
1760 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1761 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1762 LLVMValueRef vb_index =
1763 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1764 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1765 generate_fetch(gallivm, draw, vbuffers_ptr,
1766 &aos_attribs[j][i], velem, vb, true_index,
1767 system_values.instance_id, start_instance);
1768 }
1769 }
1770 convert_to_soa(gallivm, aos_attribs, inputs,
1771 draw->pt.nr_vertex_elements, vs_type);
1772
1773 /* In the paths with elts vertex id has to be unaffected by the
1774 * index bias and because indices inside our elements array have
1775 * already had index bias applied we need to subtract it here to
1776 * get back to the original index.
1777 * in the linear paths vertex id has to be unaffected by the
1778 * original start index and because we abuse the 'start' variable
1779 * to either represent the actual start index or the index at which
1780 * the primitive was split (we split rendering into chunks of at
1781 * most 4095-vertices) we need to back out the original start
1782 * index out of our vertex id here.
1783 */
1784 system_values.basevertex = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm,
1785 lp_type_uint_vec(32, 32*vector_length)),
1786 vertex_id_offset);
1787 system_values.vertex_id = true_index_array;
1788 system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1789 system_values.basevertex, "");
1790
1791 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1792 generate_vs(variant,
1793 builder,
1794 vs_type,
1795 outputs,
1796 ptr_aos,
1797 &system_values,
1798 context_ptr,
1799 sampler,
1800 key->clamp_vertex_color);
1801
1802 if (pos != -1 && cv != -1) {
1803 /* store original positions in clip before further manipulation */
1804 store_clip(gallivm, vs_type, io, outputs, pos);
1805
1806 /* do cliptest */
1807 if (enable_cliptest) {
1808 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1809 /* allocate clipmask, assign it integer type */
1810 clipmask = generate_clipmask(llvm,
1811 gallivm,
1812 vs_type,
1813 outputs,
1814 key,
1815 context_ptr, &have_clipdist);
1816 temp = LLVMBuildOr(builder, clipmask, temp, "");
1817 /* store temporary clipping boolean value */
1818 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1819 }
1820 else {
1821 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1822 }
1823
1824 /* do viewport mapping */
1825 if (!bypass_viewport) {
1826 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1827 }
1828 }
1829 else {
1830 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1831 }
1832
1833 /* store clipmask in vertex header,
1834 * original positions in clip
1835 * and transformed positions in data
1836 */
1837 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1838 vs_info->num_outputs, vs_type,
1839 enable_cliptest && key->need_edgeflags);
1840 }
1841 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1842
1843 sampler->destroy(sampler);
1844
1845 /* return clipping boolean value for function */
1846 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr,
1847 enable_cliptest && key->need_edgeflags);
1848
1849 LLVMBuildRet(builder, ret);
1850
1851 gallivm_verify_function(gallivm, variant_func);
1852 }
1853
1854
1855 struct draw_llvm_variant_key *
1856 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1857 {
1858 unsigned i;
1859 struct draw_llvm_variant_key *key;
1860 struct draw_sampler_static_state *draw_sampler;
1861
1862 key = (struct draw_llvm_variant_key *)store;
1863
1864 memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
1865
1866 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1867
1868 /* Presumably all variants of the shader should have the same
1869 * number of vertex elements - ie the number of shader inputs.
1870 * NOTE: we NEED to store the needed number of needed inputs
1871 * here, not the number of provided elements to match keysize
1872 * (and the offset of sampler state in the key).
1873 */
1874 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1875 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1876
1877 /* will have to rig this up properly later */
1878 key->clip_xy = llvm->draw->clip_xy;
1879 key->clip_z = llvm->draw->clip_z;
1880 key->clip_user = llvm->draw->clip_user;
1881 key->bypass_viewport = llvm->draw->bypass_viewport;
1882 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1883 /* XXX assumes edgeflag output not at 0 */
1884 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1885 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1886 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1887 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1888
1889 /* All variants of this shader will have the same value for
1890 * nr_samplers. Not yet trying to compact away holes in the
1891 * sampler array.
1892 */
1893 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1894 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1895 key->nr_sampler_views =
1896 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1897 }
1898 else {
1899 key->nr_sampler_views = key->nr_samplers;
1900 }
1901
1902 draw_sampler = draw_llvm_variant_key_samplers(key);
1903
1904 memcpy(key->vertex_element,
1905 llvm->draw->pt.vertex_element,
1906 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1907
1908 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1909
1910 for (i = 0 ; i < key->nr_samplers; i++) {
1911 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1912 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1913 }
1914 for (i = 0 ; i < key->nr_sampler_views; i++) {
1915 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1916 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1917 }
1918
1919 return key;
1920 }
1921
1922
1923 void
1924 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1925 {
1926 unsigned i;
1927 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1928
1929 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1930 debug_printf("clip_xy = %u\n", key->clip_xy);
1931 debug_printf("clip_z = %u\n", key->clip_z);
1932 debug_printf("clip_user = %u\n", key->clip_user);
1933 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1934 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1935 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1936 debug_printf("has_gs = %u\n", key->has_gs);
1937 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1938
1939 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1940 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1941 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1942 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1943 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1944 }
1945
1946 for (i = 0 ; i < key->nr_sampler_views; i++) {
1947 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1948 }
1949 }
1950
1951
1952 void
1953 draw_llvm_set_mapped_texture(struct draw_context *draw,
1954 unsigned shader_stage,
1955 unsigned sview_idx,
1956 uint32_t width, uint32_t height, uint32_t depth,
1957 uint32_t first_level, uint32_t last_level,
1958 const void *base_ptr,
1959 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1960 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1961 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1962 {
1963 unsigned j;
1964 struct draw_jit_texture *jit_tex;
1965
1966 assert(shader_stage == PIPE_SHADER_VERTEX ||
1967 shader_stage == PIPE_SHADER_GEOMETRY);
1968
1969 if (shader_stage == PIPE_SHADER_VERTEX) {
1970 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1971
1972 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1973 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1974 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1975
1976 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1977 } else {
1978 assert(0);
1979 return;
1980 }
1981
1982 jit_tex->width = width;
1983 jit_tex->height = height;
1984 jit_tex->depth = depth;
1985 jit_tex->first_level = first_level;
1986 jit_tex->last_level = last_level;
1987 jit_tex->base = base_ptr;
1988
1989 for (j = first_level; j <= last_level; j++) {
1990 jit_tex->mip_offsets[j] = mip_offsets[j];
1991 jit_tex->row_stride[j] = row_stride[j];
1992 jit_tex->img_stride[j] = img_stride[j];
1993 }
1994 }
1995
1996
1997 void
1998 draw_llvm_set_sampler_state(struct draw_context *draw,
1999 unsigned shader_type)
2000 {
2001 unsigned i;
2002
2003 if (shader_type == PIPE_SHADER_VERTEX) {
2004 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2005 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2006
2007 if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2008 const struct pipe_sampler_state *s
2009 = draw->samplers[PIPE_SHADER_VERTEX][i];
2010 jit_sam->min_lod = s->min_lod;
2011 jit_sam->max_lod = s->max_lod;
2012 jit_sam->lod_bias = s->lod_bias;
2013 COPY_4V(jit_sam->border_color, s->border_color.f);
2014 }
2015 }
2016 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2017 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2018 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2019
2020 if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2021 const struct pipe_sampler_state *s
2022 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2023 jit_sam->min_lod = s->min_lod;
2024 jit_sam->max_lod = s->max_lod;
2025 jit_sam->lod_bias = s->lod_bias;
2026 COPY_4V(jit_sam->border_color, s->border_color.f);
2027 }
2028 }
2029 }
2030 }
2031
2032
2033 void
2034 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2035 {
2036 struct draw_llvm *llvm = variant->llvm;
2037
2038 gallivm_destroy(variant->gallivm);
2039
2040 remove_from_list(&variant->list_item_local);
2041 variant->shader->variants_cached--;
2042 remove_from_list(&variant->list_item_global);
2043 llvm->nr_variants--;
2044 FREE(variant);
2045 }
2046
2047
2048 /**
2049 * Create LLVM types for various structures.
2050 */
2051 static void
2052 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2053 {
2054 struct gallivm_state *gallivm = var->gallivm;
2055 LLVMTypeRef texture_type, sampler_type, context_type;
2056
2057 texture_type = create_jit_texture_type(gallivm, "texture");
2058 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2059
2060 context_type = create_gs_jit_context_type(gallivm,
2061 var->shader->base.vector_length,
2062 texture_type, sampler_type,
2063 "draw_gs_jit_context");
2064 var->context_ptr_type = LLVMPointerType(context_type, 0);
2065
2066 var->input_array_type = create_gs_jit_input_type(gallivm);
2067 }
2068
2069 static LLVMTypeRef
2070 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2071 {
2072 if (!variant->context_ptr_type)
2073 create_gs_jit_types(variant);
2074 return variant->context_ptr_type;
2075 }
2076
2077 static LLVMValueRef
2078 generate_mask_value(struct draw_gs_llvm_variant *variant,
2079 struct lp_type gs_type)
2080 {
2081 struct gallivm_state *gallivm = variant->gallivm;
2082 LLVMBuilderRef builder = gallivm->builder;
2083 struct lp_type mask_type = lp_int_type(gs_type);
2084 LLVMValueRef num_prims;
2085 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2086 unsigned i;
2087
2088 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2089 variant->num_prims);
2090 for (i = 0; i < gs_type.length; i++) {
2091 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2092 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2093 }
2094 mask_val = lp_build_compare(gallivm, mask_type,
2095 PIPE_FUNC_GREATER, num_prims, mask_val);
2096
2097 return mask_val;
2098 }
2099
2100 static void
2101 draw_gs_llvm_generate(struct draw_llvm *llvm,
2102 struct draw_gs_llvm_variant *variant)
2103 {
2104 struct gallivm_state *gallivm = variant->gallivm;
2105 LLVMContextRef context = gallivm->context;
2106 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2107 LLVMTypeRef arg_types[7];
2108 LLVMTypeRef func_type;
2109 LLVMValueRef variant_func;
2110 LLVMValueRef context_ptr;
2111 LLVMValueRef prim_id_ptr;
2112 LLVMBasicBlockRef block;
2113 LLVMBuilderRef builder;
2114 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2115 struct lp_build_sampler_soa *sampler = 0;
2116 struct lp_build_context bld;
2117 struct lp_bld_tgsi_system_values system_values;
2118 char func_name[64];
2119 struct lp_type gs_type;
2120 unsigned i;
2121 struct draw_gs_llvm_iface gs_iface;
2122 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2123 LLVMValueRef consts_ptr, num_consts_ptr;
2124 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2125 struct lp_build_mask_context mask;
2126 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2127 unsigned vector_length = variant->shader->base.vector_length;
2128
2129 memset(&system_values, 0, sizeof(system_values));
2130
2131 util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2132 variant->shader->variants_cached);
2133
2134 assert(variant->vertex_header_ptr_type);
2135
2136 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2137 arg_types[1] = variant->input_array_type; /* input */
2138 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2139 arg_types[3] = int32_type; /* num_prims */
2140 arg_types[4] = int32_type; /* instance_id */
2141 arg_types[5] = LLVMPointerType(
2142 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2143 arg_types[6] = int32_type;
2144
2145 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2146
2147 variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2148
2149 variant->function = variant_func;
2150
2151 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2152
2153 for (i = 0; i < Elements(arg_types); ++i)
2154 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2155 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2156 LLVMNoAliasAttribute);
2157
2158 context_ptr = LLVMGetParam(variant_func, 0);
2159 input_array = LLVMGetParam(variant_func, 1);
2160 io_ptr = LLVMGetParam(variant_func, 2);
2161 num_prims = LLVMGetParam(variant_func, 3);
2162 system_values.instance_id = LLVMGetParam(variant_func, 4);
2163 prim_id_ptr = LLVMGetParam(variant_func, 5);
2164 system_values.invocation_id = LLVMGetParam(variant_func, 6);
2165
2166 lp_build_name(context_ptr, "context");
2167 lp_build_name(input_array, "input");
2168 lp_build_name(io_ptr, "io");
2169 lp_build_name(num_prims, "num_prims");
2170 lp_build_name(system_values.instance_id, "instance_id");
2171 lp_build_name(prim_id_ptr, "prim_id_ptr");
2172 lp_build_name(system_values.invocation_id, "invocation_id");
2173
2174 variant->context_ptr = context_ptr;
2175 variant->io_ptr = io_ptr;
2176 variant->num_prims = num_prims;
2177
2178 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2179 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2180 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2181 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2182 gs_iface.input = input_array;
2183 gs_iface.variant = variant;
2184
2185 /*
2186 * Function body
2187 */
2188
2189 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2190 builder = gallivm->builder;
2191 LLVMPositionBuilderAtEnd(builder, block);
2192
2193 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2194
2195 memset(&gs_type, 0, sizeof gs_type);
2196 gs_type.floating = TRUE; /* floating point values */
2197 gs_type.sign = TRUE; /* values are signed */
2198 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2199 gs_type.width = 32; /* 32-bit float */
2200 gs_type.length = vector_length;
2201
2202 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2203 num_consts_ptr =
2204 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2205
2206 /* code generated texture sampling */
2207 sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2208
2209 mask_val = generate_mask_value(variant, gs_type);
2210 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2211
2212 if (gs_info->uses_primid) {
2213 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2214 }
2215
2216 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2217 tgsi_dump(tokens, 0);
2218 draw_gs_llvm_dump_variant_key(&variant->key);
2219 }
2220
2221 lp_build_tgsi_soa(variant->gallivm,
2222 tokens,
2223 gs_type,
2224 &mask,
2225 consts_ptr,
2226 num_consts_ptr,
2227 &system_values,
2228 NULL,
2229 outputs,
2230 context_ptr,
2231 NULL,
2232 sampler,
2233 &llvm->draw->gs.geometry_shader->info,
2234 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2235
2236 sampler->destroy(sampler);
2237
2238 lp_build_mask_end(&mask);
2239
2240 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2241
2242 gallivm_verify_function(gallivm, variant_func);
2243 }
2244
2245
2246 struct draw_gs_llvm_variant *
2247 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2248 unsigned num_outputs,
2249 const struct draw_gs_llvm_variant_key *key)
2250 {
2251 struct draw_gs_llvm_variant *variant;
2252 struct llvm_geometry_shader *shader =
2253 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2254 LLVMTypeRef vertex_header;
2255 char module_name[64];
2256
2257 variant = MALLOC(sizeof *variant +
2258 shader->variant_key_size -
2259 sizeof variant->key);
2260 if (!variant)
2261 return NULL;
2262
2263 variant->llvm = llvm;
2264 variant->shader = shader;
2265
2266 util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2267 variant->shader->variants_cached);
2268
2269 variant->gallivm = gallivm_create(module_name, llvm->context);
2270
2271 create_gs_jit_types(variant);
2272
2273 memcpy(&variant->key, key, shader->variant_key_size);
2274
2275 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2276
2277 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2278
2279 draw_gs_llvm_generate(llvm, variant);
2280
2281 gallivm_compile_module(variant->gallivm);
2282
2283 variant->jit_func = (draw_gs_jit_func)
2284 gallivm_jit_function(variant->gallivm, variant->function);
2285
2286 gallivm_free_ir(variant->gallivm);
2287
2288 variant->list_item_global.base = variant;
2289 variant->list_item_local.base = variant;
2290 /*variant->no = */shader->variants_created++;
2291 variant->list_item_global.base = variant;
2292
2293 return variant;
2294 }
2295
2296 void
2297 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2298 {
2299 struct draw_llvm *llvm = variant->llvm;
2300
2301 gallivm_destroy(variant->gallivm);
2302
2303 remove_from_list(&variant->list_item_local);
2304 variant->shader->variants_cached--;
2305 remove_from_list(&variant->list_item_global);
2306 llvm->nr_gs_variants--;
2307 FREE(variant);
2308 }
2309
2310 struct draw_gs_llvm_variant_key *
2311 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2312 {
2313 unsigned i;
2314 struct draw_gs_llvm_variant_key *key;
2315 struct draw_sampler_static_state *draw_sampler;
2316
2317 key = (struct draw_gs_llvm_variant_key *)store;
2318
2319 memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2320
2321 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2322
2323 /* All variants of this shader will have the same value for
2324 * nr_samplers. Not yet trying to compact away holes in the
2325 * sampler array.
2326 */
2327 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2328 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2329 key->nr_sampler_views =
2330 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2331 }
2332 else {
2333 key->nr_sampler_views = key->nr_samplers;
2334 }
2335
2336 draw_sampler = key->samplers;
2337
2338 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2339
2340 for (i = 0 ; i < key->nr_samplers; i++) {
2341 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2342 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2343 }
2344 for (i = 0 ; i < key->nr_sampler_views; i++) {
2345 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2346 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2347 }
2348
2349 return key;
2350 }
2351
2352 void
2353 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2354 {
2355 unsigned i;
2356 struct draw_sampler_static_state *sampler = key->samplers;
2357
2358 for (i = 0 ; i < key->nr_sampler_views; i++) {
2359 debug_printf("sampler[%i].src_format = %s\n", i,
2360 util_format_name(sampler[i].texture_state.format));
2361 }
2362 }