projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallium/draw: remove double semicolon
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
1 /**************************************************************************
2 *
3 * Copyright 2010 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 #include "draw_llvm.h"
29
30 #include "draw_context.h"
31 #include "draw_vs.h"
32 #include "draw_gs.h"
33
34 #include "gallivm/lp_bld_arit.h"
35 #include "gallivm/lp_bld_arit_overflow.h"
36 #include "gallivm/lp_bld_logic.h"
37 #include "gallivm/lp_bld_const.h"
38 #include "gallivm/lp_bld_swizzle.h"
39 #include "gallivm/lp_bld_struct.h"
40 #include "gallivm/lp_bld_type.h"
41 #include "gallivm/lp_bld_flow.h"
42 #include "gallivm/lp_bld_debug.h"
43 #include "gallivm/lp_bld_tgsi.h"
44 #include "gallivm/lp_bld_printf.h"
45 #include "gallivm/lp_bld_intr.h"
46 #include "gallivm/lp_bld_init.h"
47 #include "gallivm/lp_bld_type.h"
48 #include "gallivm/lp_bld_pack.h"
49 #include "gallivm/lp_bld_format.h"
50
51 #include "tgsi/tgsi_exec.h"
52 #include "tgsi/tgsi_dump.h"
53
54 #include "util/u_math.h"
55 #include "util/u_pointer.h"
56 #include "util/u_string.h"
57 #include "util/u_simple_list.h"
58
59
60 #define DEBUG_STORE 0
61
62
63 static void
64 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
65 boolean elts);
66
67
68 struct draw_gs_llvm_iface {
69 struct lp_build_tgsi_gs_iface base;
70
71 struct draw_gs_llvm_variant *variant;
72 LLVMValueRef input;
73 };
74
75 static INLINE const struct draw_gs_llvm_iface *
76 draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
77 {
78 return (const struct draw_gs_llvm_iface *)iface;
79 }
80
81 /**
82 * Create LLVM type for draw_vertex_buffer.
83 */
84 static LLVMTypeRef
85 create_jit_dvbuffer_type(struct gallivm_state *gallivm,
86 const char *struct_name)
87 {
88 LLVMTargetDataRef target = gallivm->target;
89 LLVMTypeRef dvbuffer_type;
90 LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
91 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
92
93 elem_types[DRAW_JIT_DVBUFFER_MAP] =
94 LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
95 elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
96
97 dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
98 Elements(elem_types), 0);
99
100 #if HAVE_LLVM < 0x0300
101 LLVMAddTypeName(gallivm->module, struct_name, dvbuffer_type);
102
103 /* Make sure the target's struct layout cache doesn't return
104 * stale/invalid data.
105 */
106 LLVMInvalidateStructLayout(gallivm->target, dvbuffer_type);
107 #endif
108
109 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
110 target, dvbuffer_type,
111 DRAW_JIT_DVBUFFER_MAP);
112 LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
113 target, dvbuffer_type,
114 DRAW_JIT_DVBUFFER_SIZE);
115
116 return dvbuffer_type;
117 }
118
119 /**
120 * Create LLVM type for struct draw_jit_texture
121 */
122 static LLVMTypeRef
123 create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
124 {
125 LLVMTargetDataRef target = gallivm->target;
126 LLVMTypeRef texture_type;
127 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
128 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
129
130 elem_types[DRAW_JIT_TEXTURE_WIDTH] =
131 elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
132 elem_types[DRAW_JIT_TEXTURE_DEPTH] =
133 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
134 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
135 elem_types[DRAW_JIT_TEXTURE_BASE] =
136 LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
137 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
138 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
139 elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
140 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
141
142 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
143 Elements(elem_types), 0);
144
145 #if HAVE_LLVM < 0x0300
146 LLVMAddTypeName(gallivm->module, struct_name, texture_type);
147
148 /* Make sure the target's struct layout cache doesn't return
149 * stale/invalid data.
150 */
151 LLVMInvalidateStructLayout(gallivm->target, texture_type);
152 #endif
153
154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
155 target, texture_type,
156 DRAW_JIT_TEXTURE_WIDTH);
157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
158 target, texture_type,
159 DRAW_JIT_TEXTURE_HEIGHT);
160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
161 target, texture_type,
162 DRAW_JIT_TEXTURE_DEPTH);
163 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
164 target, texture_type,
165 DRAW_JIT_TEXTURE_FIRST_LEVEL);
166 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
167 target, texture_type,
168 DRAW_JIT_TEXTURE_LAST_LEVEL);
169 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
170 target, texture_type,
171 DRAW_JIT_TEXTURE_BASE);
172 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
173 target, texture_type,
174 DRAW_JIT_TEXTURE_ROW_STRIDE);
175 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
176 target, texture_type,
177 DRAW_JIT_TEXTURE_IMG_STRIDE);
178 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
179 target, texture_type,
180 DRAW_JIT_TEXTURE_MIP_OFFSETS);
181
182 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
183
184 return texture_type;
185 }
186
187
188 /**
189 * Create LLVM type for struct draw_jit_sampler
190 */
191 static LLVMTypeRef
192 create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
193 {
194 LLVMTargetDataRef target = gallivm->target;
195 LLVMTypeRef sampler_type;
196 LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
197
198 elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
199 elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
200 elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
201 elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
202 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
203
204 sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
205 Elements(elem_types), 0);
206
207 #if HAVE_LLVM < 0x0300
208 LLVMAddTypeName(gallivm->module, struct_name, sampler_type);
209
210 /* Make sure the target's struct layout cache doesn't return
211 * stale/invalid data.
212 */
213 LLVMInvalidateStructLayout(gallivm->target, sampler_type);
214 #endif
215
216 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
217 target, sampler_type,
218 DRAW_JIT_SAMPLER_MIN_LOD);
219 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
220 target, sampler_type,
221 DRAW_JIT_SAMPLER_MAX_LOD);
222 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
223 target, sampler_type,
224 DRAW_JIT_SAMPLER_LOD_BIAS);
225 LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
226 target, sampler_type,
227 DRAW_JIT_SAMPLER_BORDER_COLOR);
228
229 LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
230
231 return sampler_type;
232 }
233
234
235 /**
236 * Create LLVM type for struct draw_jit_context
237 */
238 static LLVMTypeRef
239 create_jit_context_type(struct gallivm_state *gallivm,
240 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
241 const char *struct_name)
242 {
243 LLVMTargetDataRef target = gallivm->target;
244 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
245 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
246 LLVMTypeRef context_type;
247
248 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
249 LP_MAX_TGSI_CONST_BUFFERS);
250 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
251 DRAW_TOTAL_CLIP_PLANES), 0);
252 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
253 elem_types[3] = LLVMArrayType(texture_type,
254 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
255 elem_types[4] = LLVMArrayType(sampler_type,
256 PIPE_MAX_SAMPLERS); /* samplers */
257 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
258 Elements(elem_types), 0);
259 #if HAVE_LLVM < 0x0300
260 LLVMAddTypeName(gallivm->module, struct_name, context_type);
261
262 LLVMInvalidateStructLayout(gallivm->target, context_type);
263 #endif
264
265 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
266 target, context_type, DRAW_JIT_CTX_CONSTANTS);
267 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
268 target, context_type, DRAW_JIT_CTX_PLANES);
269 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewport,
270 target, context_type, DRAW_JIT_CTX_VIEWPORT);
271 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
272 target, context_type,
273 DRAW_JIT_CTX_TEXTURES);
274 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
275 target, context_type,
276 DRAW_JIT_CTX_SAMPLERS);
277 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
278 target, context_type);
279
280 return context_type;
281 }
282
283
284 /**
285 * Create LLVM type for struct draw_gs_jit_context
286 */
287 static LLVMTypeRef
288 create_gs_jit_context_type(struct gallivm_state *gallivm,
289 unsigned vector_length,
290 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
291 const char *struct_name)
292 {
293 LLVMTargetDataRef target = gallivm->target;
294 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
295 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
296 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
297 LLVMTypeRef context_type;
298
299 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
300 LP_MAX_TGSI_CONST_BUFFERS);
301 elem_types[1] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
302 DRAW_TOTAL_CLIP_PLANES), 0);
303 elem_types[2] = LLVMPointerType(float_type, 0); /* viewport */
304
305 elem_types[3] = LLVMArrayType(texture_type,
306 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
307 elem_types[4] = LLVMArrayType(sampler_type,
308 PIPE_MAX_SAMPLERS); /* samplers */
309
310 elem_types[5] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
311 elem_types[6] = LLVMPointerType(LLVMVectorType(int_type,
312 vector_length), 0);
313 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
314 vector_length), 0);
315
316 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
317 Elements(elem_types), 0);
318 #if HAVE_LLVM < 0x0300
319 LLVMAddTypeName(gallivm->module, struct_name, context_type);
320
321 LLVMInvalidateStructLayout(gallivm->target, context_type);
322 #endif
323
324 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
325 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
326 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
327 target, context_type, DRAW_GS_JIT_CTX_PLANES);
328 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewport,
329 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
330 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
331 target, context_type,
332 DRAW_GS_JIT_CTX_TEXTURES);
333 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
334 target, context_type,
335 DRAW_GS_JIT_CTX_SAMPLERS);
336 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
337 target, context_type,
338 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
339 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
340 target, context_type,
341 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
342 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
343 target, context_type,
344 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
345 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
346 target, context_type);
347
348 return context_type;
349 }
350
351
352 static LLVMTypeRef
353 create_gs_jit_input_type(struct gallivm_state *gallivm)
354 {
355 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
356 LLVMTypeRef input_array;
357
358 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
359 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
360 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
361 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
362
363 return input_array;
364 }
365
366 /**
367 * Create LLVM type for struct pipe_vertex_buffer
368 */
369 static LLVMTypeRef
370 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
371 const char *struct_name)
372 {
373 LLVMTargetDataRef target = gallivm->target;
374 LLVMTypeRef elem_types[4];
375 LLVMTypeRef vb_type;
376
377 elem_types[0] =
378 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
379 elem_types[2] =
380 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
381
382 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
383 Elements(elem_types), 0);
384 #if HAVE_LLVM < 0x0300
385 LLVMAddTypeName(gallivm->module, struct_name, vb_type);
386
387 LLVMInvalidateStructLayout(gallivm->target, vb_type);
388 #endif
389
390 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
391 target, vb_type, 0);
392 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
393 target, vb_type, 1);
394
395 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
396
397 return vb_type;
398 }
399
400
401 /**
402 * Create LLVM type for struct vertex_header;
403 */
404 static LLVMTypeRef
405 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
406 {
407 LLVMTargetDataRef target = gallivm->target;
408 LLVMTypeRef elem_types[4];
409 LLVMTypeRef vertex_header;
410 char struct_name[24];
411
412 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
413
414 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
415 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
416 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
417 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
418
419 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
420 Elements(elem_types), 0);
421 #if HAVE_LLVM < 0x0300
422 LLVMAddTypeName(gallivm->module, struct_name, vertex_header);
423
424 LLVMInvalidateStructLayout(gallivm->target, vertex_header);
425 #endif
426
427 /* these are bit-fields and we can't take address of them
428 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
429 target, vertex_header,
430 DRAW_JIT_VERTEX_CLIPMASK);
431 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
432 target, vertex_header,
433 DRAW_JIT_VERTEX_EDGEFLAG);
434 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
435 target, vertex_header,
436 DRAW_JIT_VERTEX_PAD);
437 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
438 target, vertex_header,
439 DRAW_JIT_VERTEX_VERTEX_ID);
440 */
441 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
442 target, vertex_header,
443 DRAW_JIT_VERTEX_CLIP);
444 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
445 target, vertex_header,
446 DRAW_JIT_VERTEX_PRE_CLIP_POS);
447 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
448 target, vertex_header,
449 DRAW_JIT_VERTEX_DATA);
450
451 assert(LLVMABISizeOfType(target, vertex_header) ==
452 offsetof(struct vertex_header, data[data_elems]));
453
454 return vertex_header;
455 }
456
457
458 /**
459 * Create LLVM types for various structures.
460 */
461 static void
462 create_jit_types(struct draw_llvm_variant *variant)
463 {
464 struct gallivm_state *gallivm = variant->gallivm;
465 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
466 vb_type;
467
468 texture_type = create_jit_texture_type(gallivm, "texture");
469 sampler_type = create_jit_sampler_type(gallivm, "sampler");
470
471 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
472 "draw_jit_context");
473 variant->context_ptr_type = LLVMPointerType(context_type, 0);
474
475 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
476 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
477
478 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
479 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
480 }
481
482
483 static LLVMTypeRef
484 get_context_ptr_type(struct draw_llvm_variant *variant)
485 {
486 if (!variant->context_ptr_type)
487 create_jit_types(variant);
488 return variant->context_ptr_type;
489 }
490
491
492 static LLVMTypeRef
493 get_buffer_ptr_type(struct draw_llvm_variant *variant)
494 {
495 if (!variant->buffer_ptr_type)
496 create_jit_types(variant);
497 return variant->buffer_ptr_type;
498 }
499
500
501 static LLVMTypeRef
502 get_vb_ptr_type(struct draw_llvm_variant *variant)
503 {
504 if (!variant->vb_ptr_type)
505 create_jit_types(variant);
506 return variant->vb_ptr_type;
507 }
508
509 static LLVMTypeRef
510 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
511 {
512 if (!variant->vertex_header_ptr_type)
513 create_jit_types(variant);
514 return variant->vertex_header_ptr_type;
515 }
516
517
518 /**
519 * Create per-context LLVM info.
520 */
521 struct draw_llvm *
522 draw_llvm_create(struct draw_context *draw)
523 {
524 struct draw_llvm *llvm;
525
526 llvm = CALLOC_STRUCT( draw_llvm );
527 if (!llvm)
528 return NULL;
529
530 lp_build_init();
531
532 llvm->draw = draw;
533
534 llvm->nr_variants = 0;
535 make_empty_list(&llvm->vs_variants_list);
536
537 llvm->nr_gs_variants = 0;
538 make_empty_list(&llvm->gs_variants_list);
539
540 return llvm;
541 }
542
543
544 /**
545 * Free per-context LLVM info.
546 */
547 void
548 draw_llvm_destroy(struct draw_llvm *llvm)
549 {
550 /* XXX free other draw_llvm data? */
551 FREE(llvm);
552 }
553
554
555 /**
556 * Create LLVM-generated code for a vertex shader.
557 */
558 struct draw_llvm_variant *
559 draw_llvm_create_variant(struct draw_llvm *llvm,
560 unsigned num_inputs,
561 const struct draw_llvm_variant_key *key)
562 {
563 struct draw_llvm_variant *variant;
564 struct llvm_vertex_shader *shader =
565 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
566 LLVMTypeRef vertex_header;
567
568 variant = MALLOC(sizeof *variant +
569 shader->variant_key_size -
570 sizeof variant->key);
571 if (variant == NULL)
572 return NULL;
573
574 variant->llvm = llvm;
575
576 variant->gallivm = gallivm_create();
577
578 create_jit_types(variant);
579
580 memcpy(&variant->key, key, shader->variant_key_size);
581
582 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
583
584 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
585
586 draw_llvm_generate(llvm, variant, FALSE); /* linear */
587 draw_llvm_generate(llvm, variant, TRUE); /* elts */
588
589 gallivm_compile_module(variant->gallivm);
590
591 variant->jit_func = (draw_jit_vert_func)
592 gallivm_jit_function(variant->gallivm, variant->function);
593
594 variant->jit_func_elts = (draw_jit_vert_func_elts)
595 gallivm_jit_function(variant->gallivm, variant->function_elts);
596
597 variant->shader = shader;
598 variant->list_item_global.base = variant;
599 variant->list_item_local.base = variant;
600 /*variant->no = */shader->variants_created++;
601 variant->list_item_global.base = variant;
602
603 return variant;
604 }
605
606
607 static void
608 generate_vs(struct draw_llvm_variant *variant,
609 LLVMBuilderRef builder,
610 struct lp_type vs_type,
611 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
612 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
613 const struct lp_bld_tgsi_system_values *system_values,
614 LLVMValueRef context_ptr,
615 struct lp_build_sampler_soa *draw_sampler,
616 boolean clamp_vertex_color)
617 {
618 struct draw_llvm *llvm = variant->llvm;
619 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
620 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(variant->gallivm, context_ptr);
621 struct lp_build_sampler_soa *sampler = 0;
622
623 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
624 tgsi_dump(tokens, 0);
625 draw_llvm_dump_variant_key(&variant->key);
626 }
627
628 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers)
629 sampler = draw_sampler;
630
631 lp_build_tgsi_soa(variant->gallivm,
632 tokens,
633 vs_type,
634 NULL /*struct lp_build_mask_context *mask*/,
635 consts_ptr,
636 system_values,
637 inputs,
638 outputs,
639 sampler,
640 &llvm->draw->vs.vertex_shader->info,
641 NULL);
642
643 {
644 LLVMValueRef out;
645 unsigned chan, attrib;
646 struct lp_build_context bld;
647 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
648 lp_build_context_init(&bld, variant->gallivm, vs_type);
649
650 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
651 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
652 if (outputs[attrib][chan]) {
653 switch (info->output_semantic_name[attrib]) {
654 case TGSI_SEMANTIC_COLOR:
655 case TGSI_SEMANTIC_BCOLOR:
656 if (clamp_vertex_color) {
657 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
658 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
659 LLVMBuildStore(builder, out, outputs[attrib][chan]);
660 }
661 break;
662 }
663 }
664 }
665 }
666 }
667 }
668
669 static void
670 generate_fetch(struct gallivm_state *gallivm,
671 struct draw_context *draw,
672 LLVMValueRef vbuffers_ptr,
673 LLVMValueRef *res,
674 struct pipe_vertex_element *velem,
675 LLVMValueRef vbuf,
676 LLVMValueRef index,
677 LLVMValueRef instance_id)
678 {
679 const struct util_format_description *format_desc =
680 util_format_description(velem->src_format);
681 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
682 LLVMBuilderRef builder = gallivm->builder;
683 LLVMValueRef indices =
684 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
685 velem->vertex_buffer_index, 0);
686 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
687 &indices, 1, "");
688 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
689 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
690 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
691 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
692 LLVMValueRef stride;
693 LLVMValueRef buffer_overflowed;
694 LLVMValueRef needed_buffer_size;
695 LLVMValueRef temp_ptr =
696 lp_build_alloca(gallivm,
697 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
698 LLVMValueRef ofbit = NULL;
699 struct lp_build_if_state if_ctx;
700
701 if (velem->instance_divisor) {
702 /* Index is equal to the start instance plus the number of current
703 * instance divided by the divisor. In this case we compute it as:
704 * index = start_instance + (instance_id / divisor)
705 */
706 LLVMValueRef current_instance;
707 index = lp_build_const_int32(gallivm, draw->start_instance);
708 current_instance = LLVMBuildUDiv(builder, instance_id,
709 lp_build_const_int32(gallivm, velem->instance_divisor),
710 "instance_divisor");
711 index = lp_build_uadd_overflow(gallivm, index, current_instance, &ofbit);
712 }
713
714 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
715 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
716 stride = lp_build_uadd_overflow(
717 gallivm, stride,
718 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
719 needed_buffer_size = lp_build_uadd_overflow(
720 gallivm, stride,
721 lp_build_const_int32(gallivm,
722 util_format_get_blocksize(velem->src_format)),
723 &ofbit);
724
725 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
726 needed_buffer_size, buffer_size,
727 "buffer_overflowed");
728 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
729 #if 0
730 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
731 index, vb_stride);
732 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
733 vb_buffer_offset,
734 lp_build_const_int32(gallivm, velem->src_offset));
735 lp_build_print_value(gallivm, " blocksize = ",
736 lp_build_const_int32(
737 gallivm,
738 util_format_get_blocksize(velem->src_format)));
739 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
740 lp_build_printf(gallivm, " stride = %u\n", stride);
741 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
742 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
743 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
744 #endif
745
746 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
747 {
748 LLVMValueRef val =
749 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
750 LLVMBuildStore(builder, val, temp_ptr);
751 }
752 lp_build_else(&if_ctx);
753 {
754 LLVMValueRef val;
755 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
756
757 val = lp_build_fetch_rgba_aos(gallivm,
758 format_desc,
759 lp_float32_vec4_type(),
760 map_ptr,
761 zero, zero, zero);
762 LLVMBuildStore(builder, val, temp_ptr);
763 }
764 lp_build_endif(&if_ctx);
765
766 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
767 }
768
769 static void
770 convert_to_soa(struct gallivm_state *gallivm,
771 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
772 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
773 unsigned num_attribs, const struct lp_type soa_type)
774 {
775 unsigned i, j, k;
776 struct lp_type aos_channel_type = soa_type;
777
778 debug_assert(TGSI_NUM_CHANNELS == 4);
779 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
780
781 aos_channel_type.length >>= 1;
782
783 for (i = 0; i < num_attribs; ++i) {
784 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
785 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
786
787 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
788 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
789
790 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
791
792 for (k = 0; k < pixels_per_channel; ++k) {
793 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
794 }
795
796 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
797 }
798
799 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
800 }
801 }
802
803
804 static void
805 store_aos(struct gallivm_state *gallivm,
806 LLVMValueRef io_ptr,
807 LLVMValueRef index,
808 LLVMValueRef value)
809 {
810 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
811 LLVMBuilderRef builder = gallivm->builder;
812 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
813 LLVMValueRef indices[3];
814
815 indices[0] = lp_build_const_int32(gallivm, 0);
816 indices[1] = index;
817 indices[2] = lp_build_const_int32(gallivm, 0);
818
819 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
820 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
821
822 #if DEBUG_STORE
823 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
824 #endif
825
826 /* Unaligned store due to the vertex header */
827 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
828 }
829
830 /**
831 * Adjust the mask to architecture endianess. The mask will the store in struct:
832 *
833 * struct vertex_header {
834 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
835 * unsigned edgeflag:1;
836 * unsigned have_clipdist:1;
837 * unsigned vertex_id:16;
838 * [...]
839 * }
840 *
841 * On little-endian machine nothing needs to done, however on bit-endian machine
842 * the mask's fields need to be adjusted with the algorithm:
843 *
844 * uint32_t reverse (uint32_t x)
845 * {
846 * return (x >> 16) | // vertex_id
847 * ((x & 0x3fff) << 18) | // clipmask
848 * ((x & 0x4000) << 3) | // have_clipdist
849 * ((x & 0x8000) << 1); // edgeflag
850 * }
851 */
852 static LLVMValueRef
853 adjust_mask(struct gallivm_state *gallivm,
854 LLVMValueRef mask)
855 {
856 #ifdef PIPE_ARCH_BIG_ENDIAN
857 LLVMBuilderRef builder = gallivm->builder;
858 LLVMValueRef vertex_id;
859 LLVMValueRef clipmask;
860 LLVMValueRef have_clipdist;
861 LLVMValueRef edgeflag;
862
863 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
864 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
865 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
866 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
867 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
868 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
869 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
870
871 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
872 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
873 mask = LLVMBuildOr(builder, mask, edgeflag, "");
874 #endif
875 return mask;
876 }
877
878 static void
879 store_aos_array(struct gallivm_state *gallivm,
880 struct lp_type soa_type,
881 LLVMValueRef io_ptr,
882 LLVMValueRef *indices,
883 LLVMValueRef* aos,
884 int attrib,
885 int num_outputs,
886 LLVMValueRef clipmask,
887 boolean have_clipdist)
888 {
889 LLVMBuilderRef builder = gallivm->builder;
890 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
891 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
892 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
893 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
894 int vector_length = soa_type.length;
895 int i;
896
897 debug_assert(TGSI_NUM_CHANNELS == 4);
898
899 for (i = 0; i < vector_length; i++) {
900 linear_inds[i] = lp_build_const_int32(gallivm, i);
901 if (indices) {
902 inds[i] = indices[i];
903 } else {
904 inds[i] = linear_inds[i];
905 }
906 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
907 }
908
909 if (attrib == 0) {
910 /* store vertex header for each of the n vertices */
911 LLVMValueRef val, cliptmp;
912 int vertex_id_pad_edgeflag;
913
914 /* If this assertion fails, it means we need to update the bit twidding
915 * code here. See struct vertex_header in draw_private.h.
916 */
917 assert(DRAW_TOTAL_CLIP_PLANES==14);
918 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
919 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
920 if (have_clipdist)
921 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
922 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
923 /* OR with the clipmask */
924 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
925 for (i = 0; i < vector_length; i++) {
926 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
927 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
928 val = adjust_mask(gallivm, val);
929 LLVMBuildStore(builder, val, id_ptr);
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 }
935 }
936
937 /* store for each of the n vertices */
938 for (i = 0; i < vector_length; i++) {
939 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
940 }
941 }
942
943
944 static void
945 convert_to_aos(struct gallivm_state *gallivm,
946 LLVMValueRef io,
947 LLVMValueRef *indices,
948 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
949 LLVMValueRef clipmask,
950 int num_outputs,
951 struct lp_type soa_type,
952 boolean have_clipdist)
953 {
954 LLVMBuilderRef builder = gallivm->builder;
955 unsigned chan, attrib, i;
956
957 #if DEBUG_STORE
958 lp_build_printf(gallivm, " # storing begin\n");
959 #endif
960 for (attrib = 0; attrib < num_outputs; ++attrib) {
961 LLVMValueRef soa[TGSI_NUM_CHANNELS];
962 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
963 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
964 if (outputs[attrib][chan]) {
965 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
966 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
967 #if DEBUG_STORE
968 lp_build_printf(gallivm, "output %d : %d ",
969 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
970 attrib, 0),
971 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
972 chan, 0));
973 lp_build_print_value(gallivm, "val = ", out);
974 {
975 LLVMValueRef iv =
976 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
977
978 lp_build_print_value(gallivm, " ival = ", iv);
979 }
980 #endif
981 soa[chan] = out;
982 }
983 else {
984 soa[chan] = 0;
985 }
986 }
987
988
989 if (soa_type.length == TGSI_NUM_CHANNELS) {
990 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
991 } else {
992 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
993
994 for (i = 0; i < soa_type.length; ++i) {
995 aos[i] = lp_build_extract_range(gallivm,
996 soa[i % TGSI_NUM_CHANNELS],
997 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
998 TGSI_NUM_CHANNELS);
999 }
1000 }
1001
1002 store_aos_array(gallivm,
1003 soa_type,
1004 io, indices,
1005 aos,
1006 attrib,
1007 num_outputs,
1008 clipmask, have_clipdist);
1009 }
1010 #if DEBUG_STORE
1011 lp_build_printf(gallivm, " # storing end\n");
1012 #endif
1013 }
1014
1015
1016 /**
1017 * Stores original vertex positions in clip coordinates
1018 */
1019 static void
1020 store_clip(struct gallivm_state *gallivm,
1021 const struct lp_type vs_type,
1022 LLVMValueRef io_ptr,
1023 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1024 boolean pre_clip_pos, int idx)
1025 {
1026 LLVMBuilderRef builder = gallivm->builder;
1027 LLVMValueRef soa[4];
1028 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1029 LLVMValueRef indices[2];
1030 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1031 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1032 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1033 LLVMTypeRef clip_ptr_type =
1034 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1035 4), 0);
1036 int i, j;
1037
1038 indices[0] =
1039 indices[1] = lp_build_const_int32(gallivm, 0);
1040
1041 for (i = 0; i < vs_type.length; i++) {
1042 inds[i] = lp_build_const_int32(gallivm, i);
1043 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1044 }
1045
1046 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1047 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1048 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1049 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1050
1051 if (!pre_clip_pos) {
1052 for (i = 0; i < vs_type.length; i++) {
1053 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1054 }
1055 } else {
1056 for (i = 0; i < vs_type.length; i++) {
1057 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1058 }
1059 }
1060
1061 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1062 for (i = 0; i < vs_type.length; ++i) {
1063 aos[i] = lp_build_extract_range(gallivm,
1064 soa[i % TGSI_NUM_CHANNELS],
1065 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1066 TGSI_NUM_CHANNELS);
1067 }
1068
1069 for (j = 0; j < vs_type.length; j++) {
1070 LLVMValueRef clip_ptr;
1071
1072 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1073 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1074
1075 /* Unaligned store */
1076 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1077 }
1078 }
1079
1080
1081 /**
1082 * Transforms the outputs for viewport mapping
1083 */
1084 static void
1085 generate_viewport(struct draw_llvm_variant *variant,
1086 LLVMBuilderRef builder,
1087 struct lp_type vs_type,
1088 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1089 LLVMValueRef context_ptr)
1090 {
1091 int i;
1092 struct gallivm_state *gallivm = variant->gallivm;
1093 struct lp_type f32_type = vs_type;
1094 const unsigned pos = draw_current_shader_position_output(variant->llvm->draw);
1095 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1096 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1097 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1098 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
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);
1114 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1115
1116 index = lp_build_const_int32(gallivm, i+4);
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 boolean clip_xy,
1147 boolean clip_z,
1148 boolean clip_user,
1149 boolean clip_halfz,
1150 unsigned ucp_enable,
1151 LLVMValueRef context_ptr,
1152 boolean *have_clipdist)
1153 {
1154 LLVMBuilderRef builder = gallivm->builder;
1155 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1156 LLVMValueRef test, temp;
1157 LLVMValueRef zero, shift;
1158 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1159 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1160 LLVMValueRef plane1, planes, plane_ptr, sum;
1161 struct lp_type f32_type = vs_type;
1162 struct lp_type i32_type = lp_int_type(vs_type);
1163 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1164 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1165 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1166 bool have_cd = false;
1167 unsigned cd[2];
1168
1169 cd[0] = draw_current_shader_clipdistance_output(llvm->draw, 0);
1170 cd[1] = draw_current_shader_clipdistance_output(llvm->draw, 1);
1171
1172 if (cd[0] != pos || cd[1] != pos)
1173 have_cd = true;
1174
1175 if (num_written_clipdistance && !clip_user) {
1176 clip_user = true;
1177 ucp_enable = (1 << num_written_clipdistance) - 1;
1178 }
1179
1180 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1181 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1182 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1183 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1184
1185 /*
1186 * load clipvertex and position from correct locations.
1187 * if they are the same just load them once.
1188 */
1189 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1190 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1191 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1192 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1193
1194 if (clip_user && cv != pos) {
1195 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1196 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1197 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1198 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1199 } else {
1200 cv_x = pos_x;
1201 cv_y = pos_y;
1202 cv_z = pos_z;
1203 cv_w = pos_w;
1204 }
1205
1206 /* Cliptest, for hardwired planes */
1207 if (clip_xy) {
1208 /* plane 1 */
1209 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1210 temp = shift;
1211 test = LLVMBuildAnd(builder, test, temp, "");
1212 mask = test;
1213
1214 /* plane 2 */
1215 test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1216 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1217 temp = LLVMBuildShl(builder, temp, shift, "");
1218 test = LLVMBuildAnd(builder, test, temp, "");
1219 mask = LLVMBuildOr(builder, mask, test, "");
1220
1221 /* plane 3 */
1222 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1223 temp = LLVMBuildShl(builder, temp, shift, "");
1224 test = LLVMBuildAnd(builder, test, temp, "");
1225 mask = LLVMBuildOr(builder, mask, test, "");
1226
1227 /* plane 4 */
1228 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1229 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1230 temp = LLVMBuildShl(builder, temp, shift, "");
1231 test = LLVMBuildAnd(builder, test, temp, "");
1232 mask = LLVMBuildOr(builder, mask, test, "");
1233 }
1234
1235 if (clip_z) {
1236 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1237 if (clip_halfz) {
1238 /* plane 5 */
1239 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1240 test = LLVMBuildAnd(builder, test, temp, "");
1241 mask = LLVMBuildOr(builder, mask, test, "");
1242 }
1243 else {
1244 /* plane 5 */
1245 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1246 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1247 test = LLVMBuildAnd(builder, test, temp, "");
1248 mask = LLVMBuildOr(builder, mask, test, "");
1249 }
1250 /* plane 6 */
1251 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1252 temp = LLVMBuildShl(builder, temp, shift, "");
1253 test = LLVMBuildAnd(builder, test, temp, "");
1254 mask = LLVMBuildOr(builder, mask, test, "");
1255 }
1256
1257 if (clip_user) {
1258 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1259 LLVMValueRef indices[3];
1260 LLVMValueRef is_nan_or_inf;
1261
1262 /* userclip planes */
1263 while (ucp_enable) {
1264 unsigned plane_idx = ffs(ucp_enable)-1;
1265 ucp_enable &= ~(1 << plane_idx);
1266 plane_idx += 6;
1267
1268 if (have_cd && num_written_clipdistance) {
1269 LLVMValueRef clipdist;
1270 int i;
1271 i = plane_idx - 6;
1272
1273 *have_clipdist = TRUE;
1274 if (i < 4) {
1275 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1276 } else {
1277 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1278 }
1279 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1280 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1281 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1282 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1283 test = LLVMBuildAnd(builder, test, temp, "");
1284 mask = LLVMBuildOr(builder, mask, test, "");
1285 } else {
1286 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1287 indices[0] = lp_build_const_int32(gallivm, 0);
1288 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1289
1290 indices[2] = lp_build_const_int32(gallivm, 0);
1291 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1292 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1293 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1294 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1295
1296 indices[2] = lp_build_const_int32(gallivm, 1);
1297 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1298 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1299 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1300 test = LLVMBuildFMul(builder, planes, cv_y, "");
1301 sum = LLVMBuildFAdd(builder, sum, test, "");
1302
1303 indices[2] = lp_build_const_int32(gallivm, 2);
1304 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1305 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1306 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1307 test = LLVMBuildFMul(builder, planes, cv_z, "");
1308 sum = LLVMBuildFAdd(builder, sum, test, "");
1309
1310 indices[2] = lp_build_const_int32(gallivm, 3);
1311 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1312 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1313 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1314 test = LLVMBuildFMul(builder, planes, cv_w, "");
1315 sum = LLVMBuildFAdd(builder, sum, test, "");
1316
1317 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1318 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1319 test = LLVMBuildAnd(builder, test, temp, "");
1320 mask = LLVMBuildOr(builder, mask, test, "");
1321 }
1322 }
1323 }
1324 return mask;
1325 }
1326
1327
1328 /**
1329 * Returns boolean if any clipping has occurred
1330 * Used zero/non-zero i32 value to represent boolean
1331 */
1332 static LLVMValueRef
1333 clipmask_booli32(struct gallivm_state *gallivm,
1334 const struct lp_type vs_type,
1335 LLVMValueRef clipmask_bool_ptr)
1336 {
1337 LLVMBuilderRef builder = gallivm->builder;
1338 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1339 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1340 LLVMValueRef ret = LLVMConstNull(int32_type);
1341 LLVMValueRef temp;
1342 int i;
1343
1344 /*
1345 * Can do this with log2(vector length) pack instructions and one extract
1346 * (as we don't actually need a or) with sse2 which would be way better.
1347 */
1348 for (i=0; i < vs_type.length; i++) {
1349 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1350 lp_build_const_int32(gallivm, i) , "");
1351 ret = LLVMBuildOr(builder, ret, temp, "");
1352 }
1353 return ret;
1354 }
1355
1356 static LLVMValueRef
1357 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1358 struct lp_build_tgsi_context * bld_base,
1359 boolean is_vindex_indirect,
1360 LLVMValueRef vertex_index,
1361 boolean is_aindex_indirect,
1362 LLVMValueRef attrib_index,
1363 LLVMValueRef swizzle_index)
1364 {
1365 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1366 struct gallivm_state *gallivm = bld_base->base.gallivm;
1367 LLVMBuilderRef builder = gallivm->builder;
1368 LLVMValueRef indices[3];
1369 LLVMValueRef res;
1370 struct lp_type type = bld_base->base.type;
1371
1372 if (is_vindex_indirect || is_aindex_indirect) {
1373 int i;
1374 res = bld_base->base.zero;
1375 for (i = 0; i < type.length; ++i) {
1376 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1377 LLVMValueRef vert_chan_index = vertex_index;
1378 LLVMValueRef attr_chan_index = attrib_index;
1379 LLVMValueRef channel_vec, value;
1380
1381 if (is_vindex_indirect) {
1382 vert_chan_index = LLVMBuildExtractElement(builder,
1383 vertex_index, idx, "");
1384 }
1385 if (is_aindex_indirect) {
1386 attr_chan_index = LLVMBuildExtractElement(builder,
1387 attrib_index, idx, "");
1388 }
1389
1390 indices[0] = vert_chan_index;
1391 indices[1] = attr_chan_index;
1392 indices[2] = swizzle_index;
1393
1394 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1395 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1396 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1397
1398 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1399 }
1400 } else {
1401 indices[0] = vertex_index;
1402 indices[1] = attrib_index;
1403 indices[2] = swizzle_index;
1404
1405 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1406 res = LLVMBuildLoad(builder, res, "");
1407 }
1408
1409 return res;
1410 }
1411
1412 static void
1413 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1414 struct lp_build_tgsi_context * bld_base,
1415 LLVMValueRef (*outputs)[4],
1416 LLVMValueRef emitted_vertices_vec)
1417 {
1418 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1419 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1420 struct gallivm_state *gallivm = variant->gallivm;
1421 LLVMBuilderRef builder = gallivm->builder;
1422 struct lp_type gs_type = bld_base->base.type;
1423 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1424 lp_int_type(gs_type), 0);
1425 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1426 LLVMValueRef next_prim_offset =
1427 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1428 LLVMValueRef io = variant->io_ptr;
1429 unsigned i;
1430 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1431
1432 for (i = 0; i < gs_type.length; ++i) {
1433 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1434 LLVMValueRef currently_emitted =
1435 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1436 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1437 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1438 }
1439
1440 convert_to_aos(gallivm, io, indices,
1441 outputs, clipmask,
1442 gs_info->num_outputs, gs_type,
1443 FALSE);
1444 }
1445
1446 static void
1447 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1448 struct lp_build_tgsi_context * bld_base,
1449 LLVMValueRef verts_per_prim_vec,
1450 LLVMValueRef emitted_prims_vec)
1451 {
1452 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1453 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1454 struct gallivm_state *gallivm = variant->gallivm;
1455 LLVMBuilderRef builder = gallivm->builder;
1456 LLVMValueRef prim_lengts_ptr =
1457 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1458 unsigned i;
1459
1460 for (i = 0; i < bld_base->base.type.length; ++i) {
1461 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1462 LLVMValueRef prims_emitted =
1463 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1464 LLVMValueRef store_ptr;
1465 LLVMValueRef num_vertices =
1466 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1467
1468 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1469 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1470 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1471 LLVMBuildStore(builder, num_vertices, store_ptr);
1472 }
1473 }
1474
1475 static void
1476 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1477 struct lp_build_tgsi_context * bld_base,
1478 LLVMValueRef total_emitted_vertices_vec,
1479 LLVMValueRef emitted_prims_vec)
1480 {
1481 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1482 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1483 struct gallivm_state *gallivm = variant->gallivm;
1484 LLVMBuilderRef builder = gallivm->builder;
1485 LLVMValueRef emitted_verts_ptr =
1486 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1487 LLVMValueRef emitted_prims_ptr =
1488 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1489 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1490
1491 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1492 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1493
1494 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1495 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1496 }
1497
1498 static void
1499 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1500 boolean elts)
1501 {
1502 struct gallivm_state *gallivm = variant->gallivm;
1503 LLVMContextRef context = gallivm->context;
1504 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1505 LLVMTypeRef arg_types[10];
1506 unsigned num_arg_types =
1507 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1508 LLVMTypeRef func_type;
1509 LLVMValueRef context_ptr;
1510 LLVMBasicBlockRef block;
1511 LLVMBuilderRef builder;
1512 struct lp_type vs_type;
1513 LLVMValueRef end, start;
1514 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1515 LLVMValueRef vertex_id_offset;
1516 LLVMValueRef stride, step, io_itr;
1517 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1518 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1519 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1520 struct draw_context *draw = llvm->draw;
1521 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1522 unsigned i, j;
1523 struct lp_build_context bld;
1524 struct lp_build_loop_state lp_loop;
1525 const int vector_length = lp_native_vector_width / 32;
1526 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1527 LLVMValueRef fetch_max;
1528 struct lp_build_sampler_soa *sampler = 0;
1529 LLVMValueRef ret, clipmask_bool_ptr;
1530 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1531 struct draw_llvm_variant_key *key = &variant->key;
1532 /* If geometry shader is present we need to skip both the viewport
1533 * transformation and clipping otherwise the inputs to the geometry
1534 * shader will be incorrect.
1535 */
1536 const boolean bypass_viewport = gs || key->bypass_viewport;
1537 const boolean enable_cliptest = !gs && (key->clip_xy ||
1538 key->clip_z ||
1539 key->clip_user);
1540 LLVMValueRef variant_func;
1541 const unsigned pos = draw_current_shader_position_output(llvm->draw);
1542 const unsigned cv = draw_current_shader_clipvertex_output(llvm->draw);
1543 boolean have_clipdist = FALSE;
1544 struct lp_bld_tgsi_system_values system_values;
1545
1546 memset(&system_values, 0, sizeof(system_values));
1547
1548 i = 0;
1549 arg_types[i++] = get_context_ptr_type(variant); /* context */
1550 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1551 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1552 if (elts) {
1553 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1554 arg_types[i++] = int32_type; /* fetch_elt_max */
1555 } else
1556 arg_types[i++] = int32_type; /* start */
1557 arg_types[i++] = int32_type; /* fetch_count / count */
1558 arg_types[i++] = int32_type; /* stride */
1559 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1560 arg_types[i++] = int32_type; /* instance_id */
1561 arg_types[i++] = int32_type; /* vertex_id_offset */
1562
1563 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1564
1565 variant_func = LLVMAddFunction(gallivm->module,
1566 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1567 func_type);
1568
1569 if (elts)
1570 variant->function_elts = variant_func;
1571 else
1572 variant->function = variant_func;
1573
1574 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1575 for (i = 0; i < num_arg_types; ++i)
1576 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1577 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1578 LLVMNoAliasAttribute);
1579
1580 context_ptr = LLVMGetParam(variant_func, 0);
1581 io_ptr = LLVMGetParam(variant_func, 1);
1582 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1583 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1584 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1585 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1586 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1587
1588 lp_build_name(context_ptr, "context");
1589 lp_build_name(io_ptr, "io");
1590 lp_build_name(vbuffers_ptr, "vbuffers");
1591 lp_build_name(stride, "stride");
1592 lp_build_name(vb_ptr, "vb");
1593 lp_build_name(system_values.instance_id, "instance_id");
1594 lp_build_name(vertex_id_offset, "vertex_id_offset");
1595
1596 if (elts) {
1597 fetch_elts = LLVMGetParam(variant_func, 3);
1598 fetch_elt_max = LLVMGetParam(variant_func, 4);
1599 fetch_count = LLVMGetParam(variant_func, 5);
1600 lp_build_name(fetch_elts, "fetch_elts");
1601 lp_build_name(fetch_elt_max, "fetch_elt_max");
1602 lp_build_name(fetch_count, "fetch_count");
1603 start = count = NULL;
1604 }
1605 else {
1606 start = LLVMGetParam(variant_func, 3);
1607 count = LLVMGetParam(variant_func, 4);
1608 lp_build_name(start, "start");
1609 lp_build_name(count, "count");
1610 fetch_elts = fetch_count = NULL;
1611 }
1612
1613 /*
1614 * Function body
1615 */
1616
1617 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1618 builder = gallivm->builder;
1619 LLVMPositionBuilderAtEnd(builder, block);
1620
1621 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1622
1623 memset(&vs_type, 0, sizeof vs_type);
1624 vs_type.floating = TRUE; /* floating point values */
1625 vs_type.sign = TRUE; /* values are signed */
1626 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1627 vs_type.width = 32; /* 32-bit float */
1628 vs_type.length = vector_length;
1629
1630 /* hold temporary "bool" clipmask */
1631 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1632 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1633
1634 /* code generated texture sampling */
1635 sampler = draw_llvm_sampler_soa_create(
1636 draw_llvm_variant_key_samplers(key),
1637 context_ptr);
1638
1639 if (elts) {
1640 start = zero;
1641 end = fetch_count;
1642 count = fetch_count;
1643 }
1644 else {
1645 end = lp_build_add(&bld, start, count);
1646 }
1647
1648 step = lp_build_const_int32(gallivm, vector_length);
1649
1650 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1651
1652 lp_build_loop_begin(&lp_loop, gallivm, zero);
1653 {
1654 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1655 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1656 LLVMValueRef io;
1657 LLVMValueRef clipmask; /* holds the clipmask value */
1658 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1659
1660 io_itr = lp_loop.counter;
1661
1662 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1663 #if DEBUG_STORE
1664 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1665 io_itr, io, lp_loop.counter);
1666 #endif
1667 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1668 for (i = 0; i < vector_length; ++i) {
1669 LLVMValueRef vert_index =
1670 LLVMBuildAdd(builder,
1671 lp_loop.counter,
1672 lp_build_const_int32(gallivm, i), "");
1673 LLVMValueRef true_index =
1674 LLVMBuildAdd(builder, start, vert_index, "");
1675 LLVMValueRef vertex_id;
1676
1677 /* make sure we're not out of bounds which can happen
1678 * if fetch_count % 4 != 0, because on the last iteration
1679 * a few of the 4 vertex fetches will be out of bounds */
1680 true_index = lp_build_min(&bld, true_index, fetch_max);
1681
1682 if (elts) {
1683 LLVMValueRef fetch_ptr;
1684 LLVMValueRef index_overflowed;
1685 LLVMValueRef index_ptr =
1686 lp_build_alloca(
1687 gallivm,
1688 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1689 struct lp_build_if_state if_ctx;
1690 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1691 true_index, fetch_elt_max,
1692 "index_overflowed");
1693
1694 lp_build_if(&if_ctx, gallivm, index_overflowed);
1695 {
1696 /* Generate maximum possible index so that
1697 * generate_fetch can treat it just like
1698 * any other overflow and return zeros.
1699 * We don't have to worry about the restart
1700 * primitive index because it has already been
1701 * handled
1702 */
1703 LLVMValueRef val =
1704 lp_build_const_int32(gallivm, 0xffffffff);
1705 LLVMBuildStore(builder, val, index_ptr);
1706 }
1707 lp_build_else(&if_ctx);
1708 {
1709 LLVMValueRef val;
1710 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1711 &true_index, 1, "");
1712 val = LLVMBuildLoad(builder, fetch_ptr, "");
1713 LLVMBuildStore(builder, val, index_ptr);
1714 }
1715 lp_build_endif(&if_ctx);
1716 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1717 }
1718 /* in the paths with elts vertex id has to be unaffected by the
1719 * index bias and because indices inside our elements array have
1720 * already had index bias applied we need to subtract it here to
1721 * get back to the original index.
1722 * in the linear paths vertex id has to be unaffected by the
1723 * original start index and because we abuse the 'start' variable
1724 * to either represent the actual start index or the index at which
1725 * the primitive was split (we split rendering into chunks of at
1726 * most 4095-vertices) we need to back out the original start
1727 * index out of our vertex id here.
1728 */
1729 vertex_id = LLVMBuildSub(builder, true_index, vertex_id_offset, "");
1730
1731 system_values.vertex_id = LLVMBuildInsertElement(
1732 gallivm->builder,
1733 system_values.vertex_id, vertex_id,
1734 lp_build_const_int32(gallivm, i), "");
1735
1736 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1737 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1738 LLVMValueRef vb_index =
1739 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1740 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1741 generate_fetch(gallivm, draw, vbuffers_ptr,
1742 &aos_attribs[j][i], velem, vb, true_index,
1743 system_values.instance_id);
1744 }
1745 }
1746 convert_to_soa(gallivm, aos_attribs, inputs,
1747 draw->pt.nr_vertex_elements, vs_type);
1748
1749 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1750 generate_vs(variant,
1751 builder,
1752 vs_type,
1753 outputs,
1754 ptr_aos,
1755 &system_values,
1756 context_ptr,
1757 sampler,
1758 key->clamp_vertex_color);
1759
1760 if (pos != -1 && cv != -1) {
1761 /* store original positions in clip before further manipulation */
1762 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1763 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1764
1765 /* do cliptest */
1766 if (enable_cliptest) {
1767 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1768 /* allocate clipmask, assign it integer type */
1769 clipmask = generate_clipmask(llvm,
1770 gallivm,
1771 vs_type,
1772 outputs,
1773 key->clip_xy,
1774 key->clip_z,
1775 key->clip_user,
1776 key->clip_halfz,
1777 key->ucp_enable,
1778 context_ptr, &have_clipdist);
1779 temp = LLVMBuildOr(builder, clipmask, temp, "");
1780 /* store temporary clipping boolean value */
1781 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1782 }
1783 else {
1784 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1785 }
1786
1787 /* do viewport mapping */
1788 if (!bypass_viewport) {
1789 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1790 }
1791 }
1792 else {
1793 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1794 }
1795
1796 /* store clipmask in vertex header,
1797 * original positions in clip
1798 * and transformed positions in data
1799 */
1800 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1801 vs_info->num_outputs, vs_type,
1802 have_clipdist);
1803 }
1804 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1805
1806 sampler->destroy(sampler);
1807
1808 /* return clipping boolean value for function */
1809 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1810
1811 LLVMBuildRet(builder, ret);
1812
1813 gallivm_verify_function(gallivm, variant_func);
1814 }
1815
1816
1817 struct draw_llvm_variant_key *
1818 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1819 {
1820 unsigned i;
1821 struct draw_llvm_variant_key *key;
1822 struct draw_sampler_static_state *draw_sampler;
1823
1824 key = (struct draw_llvm_variant_key *)store;
1825
1826 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1827
1828 /* Presumably all variants of the shader should have the same
1829 * number of vertex elements - ie the number of shader inputs.
1830 * NOTE: we NEED to store the needed number of needed inputs
1831 * here, not the number of provided elements to match keysize
1832 * (and the offset of sampler state in the key).
1833 */
1834 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1835 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1836
1837 /* will have to rig this up properly later */
1838 key->clip_xy = llvm->draw->clip_xy;
1839 key->clip_z = llvm->draw->clip_z;
1840 key->clip_user = llvm->draw->clip_user;
1841 key->bypass_viewport = llvm->draw->identity_viewport;
1842 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1843 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1844 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1845 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1846 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1847 key->pad1 = 0;
1848
1849 /* All variants of this shader will have the same value for
1850 * nr_samplers. Not yet trying to compact away holes in the
1851 * sampler array.
1852 */
1853 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1854 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1855 key->nr_sampler_views =
1856 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1857 }
1858 else {
1859 key->nr_sampler_views = key->nr_samplers;
1860 }
1861
1862 draw_sampler = draw_llvm_variant_key_samplers(key);
1863
1864 memcpy(key->vertex_element,
1865 llvm->draw->pt.vertex_element,
1866 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1867
1868 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1869
1870 for (i = 0 ; i < key->nr_samplers; i++) {
1871 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1872 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1873 }
1874 for (i = 0 ; i < key->nr_sampler_views; i++) {
1875 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1876 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1877 }
1878
1879 return key;
1880 }
1881
1882
1883 void
1884 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1885 {
1886 unsigned i;
1887 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1888
1889 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1890 debug_printf("clip_xy = %u\n", key->clip_xy);
1891 debug_printf("clip_z = %u\n", key->clip_z);
1892 debug_printf("clip_user = %u\n", key->clip_user);
1893 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1894 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1895 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1896 debug_printf("has_gs = %u\n", key->has_gs);
1897 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1898
1899 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1900 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1901 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1902 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1903 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1904 }
1905
1906 for (i = 0 ; i < key->nr_sampler_views; i++) {
1907 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1908 }
1909 }
1910
1911
1912 void
1913 draw_llvm_set_mapped_texture(struct draw_context *draw,
1914 unsigned shader_stage,
1915 unsigned sview_idx,
1916 uint32_t width, uint32_t height, uint32_t depth,
1917 uint32_t first_level, uint32_t last_level,
1918 const void *base_ptr,
1919 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1920 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1921 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1922 {
1923 unsigned j;
1924 struct draw_jit_texture *jit_tex;
1925
1926 assert(shader_stage == PIPE_SHADER_VERTEX ||
1927 shader_stage == PIPE_SHADER_GEOMETRY);
1928
1929 if (shader_stage == PIPE_SHADER_VERTEX) {
1930 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1931
1932 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1933 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1934 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1935
1936 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1937 } else {
1938 assert(0);
1939 return;
1940 }
1941
1942 jit_tex->width = width;
1943 jit_tex->height = height;
1944 jit_tex->depth = depth;
1945 jit_tex->first_level = first_level;
1946 jit_tex->last_level = last_level;
1947 jit_tex->base = base_ptr;
1948
1949 for (j = first_level; j <= last_level; j++) {
1950 jit_tex->mip_offsets[j] = mip_offsets[j];
1951 jit_tex->row_stride[j] = row_stride[j];
1952 jit_tex->img_stride[j] = img_stride[j];
1953 }
1954 }
1955
1956
1957 void
1958 draw_llvm_set_sampler_state(struct draw_context *draw,
1959 unsigned shader_type)
1960 {
1961 unsigned i;
1962
1963 if (shader_type == PIPE_SHADER_VERTEX) {
1964 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1965 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1966
1967 if (draw->samplers[i]) {
1968 const struct pipe_sampler_state *s
1969 = draw->samplers[PIPE_SHADER_VERTEX][i];
1970 jit_sam->min_lod = s->min_lod;
1971 jit_sam->max_lod = s->max_lod;
1972 jit_sam->lod_bias = s->lod_bias;
1973 COPY_4V(jit_sam->border_color, s->border_color.f);
1974 }
1975 }
1976 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1977 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1978 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1979
1980 if (draw->samplers[i]) {
1981 const struct pipe_sampler_state *s
1982 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1983 jit_sam->min_lod = s->min_lod;
1984 jit_sam->max_lod = s->max_lod;
1985 jit_sam->lod_bias = s->lod_bias;
1986 COPY_4V(jit_sam->border_color, s->border_color.f);
1987 }
1988 }
1989 }
1990 }
1991
1992
1993 void
1994 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
1995 {
1996 struct draw_llvm *llvm = variant->llvm;
1997
1998 if (variant->function_elts) {
1999 gallivm_free_function(variant->gallivm,
2000 variant->function_elts, variant->jit_func_elts);
2001 }
2002
2003 if (variant->function) {
2004 gallivm_free_function(variant->gallivm,
2005 variant->function, variant->jit_func);
2006 }
2007
2008 gallivm_destroy(variant->gallivm);
2009
2010 remove_from_list(&variant->list_item_local);
2011 variant->shader->variants_cached--;
2012 remove_from_list(&variant->list_item_global);
2013 llvm->nr_variants--;
2014 FREE(variant);
2015 }
2016
2017
2018 /**
2019 * Create LLVM types for various structures.
2020 */
2021 static void
2022 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2023 {
2024 struct gallivm_state *gallivm = var->gallivm;
2025 LLVMTypeRef texture_type, sampler_type, context_type;
2026
2027 texture_type = create_jit_texture_type(gallivm, "texture");
2028 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2029
2030 context_type = create_gs_jit_context_type(gallivm,
2031 var->shader->base.vector_length,
2032 texture_type, sampler_type,
2033 "draw_gs_jit_context");
2034 var->context_ptr_type = LLVMPointerType(context_type, 0);
2035
2036 var->input_array_type = create_gs_jit_input_type(gallivm);
2037 }
2038
2039 static LLVMTypeRef
2040 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2041 {
2042 if (!variant->context_ptr_type)
2043 create_gs_jit_types(variant);
2044 return variant->context_ptr_type;
2045 }
2046
2047 static LLVMValueRef
2048 generate_mask_value(struct draw_gs_llvm_variant *variant,
2049 struct lp_type gs_type)
2050 {
2051 struct gallivm_state *gallivm = variant->gallivm;
2052 LLVMBuilderRef builder = gallivm->builder;
2053 struct lp_type mask_type = lp_int_type(gs_type);
2054 LLVMValueRef num_prims;
2055 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2056 unsigned i;
2057
2058 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2059 variant->num_prims);
2060 for (i = 0; i <= gs_type.length; i++) {
2061 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2062 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2063 }
2064 mask_val = lp_build_compare(gallivm, mask_type,
2065 PIPE_FUNC_GREATER, num_prims, mask_val);
2066
2067 return mask_val;
2068 }
2069
2070 static void
2071 draw_gs_llvm_generate(struct draw_llvm *llvm,
2072 struct draw_gs_llvm_variant *variant)
2073 {
2074 struct gallivm_state *gallivm = variant->gallivm;
2075 LLVMContextRef context = gallivm->context;
2076 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2077 LLVMTypeRef arg_types[6];
2078 LLVMTypeRef func_type;
2079 LLVMValueRef variant_func;
2080 LLVMValueRef context_ptr;
2081 LLVMValueRef prim_id_ptr;
2082 LLVMBasicBlockRef block;
2083 LLVMBuilderRef builder;
2084 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2085 struct lp_build_sampler_soa *sampler = 0;
2086 struct lp_build_context bld;
2087 struct lp_bld_tgsi_system_values system_values;
2088 struct lp_type gs_type;
2089 unsigned i;
2090 struct draw_gs_llvm_iface gs_iface;
2091 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2092 LLVMValueRef consts_ptr;
2093 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2094 struct lp_build_mask_context mask;
2095 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2096 unsigned vector_length = variant->shader->base.vector_length;
2097
2098 memset(&system_values, 0, sizeof(system_values));
2099
2100 assert(variant->vertex_header_ptr_type);
2101
2102 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2103 arg_types[1] = variant->input_array_type; /* input */
2104 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2105 arg_types[3] = int32_type; /* num_prims */
2106 arg_types[4] = int32_type; /* instance_id */
2107 arg_types[5] = LLVMPointerType(
2108 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2109
2110 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2111
2112 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
2113 func_type);
2114 variant->function = variant_func;
2115
2116 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2117
2118 for (i = 0; i < Elements(arg_types); ++i)
2119 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2120 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2121 LLVMNoAliasAttribute);
2122
2123 context_ptr = LLVMGetParam(variant_func, 0);
2124 input_array = LLVMGetParam(variant_func, 1);
2125 io_ptr = LLVMGetParam(variant_func, 2);
2126 num_prims = LLVMGetParam(variant_func, 3);
2127 system_values.instance_id = LLVMGetParam(variant_func, 4);
2128 prim_id_ptr = LLVMGetParam(variant_func, 5);
2129
2130 lp_build_name(context_ptr, "context");
2131 lp_build_name(input_array, "input");
2132 lp_build_name(io_ptr, "io");
2133 lp_build_name(num_prims, "num_prims");
2134 lp_build_name(system_values.instance_id, "instance_id");
2135 lp_build_name(prim_id_ptr, "prim_id_ptr");
2136
2137 variant->context_ptr = context_ptr;
2138 variant->io_ptr = io_ptr;
2139 variant->num_prims = num_prims;
2140
2141 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2142 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2143 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2144 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2145 gs_iface.input = input_array;
2146 gs_iface.variant = variant;
2147
2148 /*
2149 * Function body
2150 */
2151
2152 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2153 builder = gallivm->builder;
2154 LLVMPositionBuilderAtEnd(builder, block);
2155
2156 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2157
2158 memset(&gs_type, 0, sizeof gs_type);
2159 gs_type.floating = TRUE; /* floating point values */
2160 gs_type.sign = TRUE; /* values are signed */
2161 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2162 gs_type.width = 32; /* 32-bit float */
2163 gs_type.length = vector_length;
2164
2165 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2166
2167 /* code generated texture sampling */
2168 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2169 context_ptr);
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 &system_values,
2189 NULL,
2190 outputs,
2191 sampler,
2192 &llvm->draw->gs.geometry_shader->info,
2193 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2194
2195 sampler->destroy(sampler);
2196
2197 lp_build_mask_end(&mask);
2198
2199 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2200
2201 gallivm_verify_function(gallivm, variant_func);
2202 }
2203
2204
2205 struct draw_gs_llvm_variant *
2206 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2207 unsigned num_outputs,
2208 const struct draw_gs_llvm_variant_key *key)
2209 {
2210 struct draw_gs_llvm_variant *variant;
2211 struct llvm_geometry_shader *shader =
2212 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2213 LLVMTypeRef vertex_header;
2214
2215 variant = MALLOC(sizeof *variant +
2216 shader->variant_key_size -
2217 sizeof variant->key);
2218 if (variant == NULL)
2219 return NULL;
2220
2221 variant->llvm = llvm;
2222 variant->shader = shader;
2223
2224 variant->gallivm = gallivm_create();
2225
2226 create_gs_jit_types(variant);
2227
2228 memcpy(&variant->key, key, shader->variant_key_size);
2229
2230 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2231
2232 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2233
2234 draw_gs_llvm_generate(llvm, variant);
2235
2236 gallivm_compile_module(variant->gallivm);
2237
2238 variant->jit_func = (draw_gs_jit_func)
2239 gallivm_jit_function(variant->gallivm, variant->function);
2240
2241 variant->list_item_global.base = variant;
2242 variant->list_item_local.base = variant;
2243 /*variant->no = */shader->variants_created++;
2244 variant->list_item_global.base = variant;
2245
2246 return variant;
2247 }
2248
2249 void
2250 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2251 {
2252 struct draw_llvm *llvm = variant->llvm;
2253
2254 if (variant->function) {
2255 gallivm_free_function(variant->gallivm,
2256 variant->function, variant->jit_func);
2257 }
2258
2259 gallivm_destroy(variant->gallivm);
2260
2261 remove_from_list(&variant->list_item_local);
2262 variant->shader->variants_cached--;
2263 remove_from_list(&variant->list_item_global);
2264 llvm->nr_gs_variants--;
2265 FREE(variant);
2266 }
2267
2268 struct draw_gs_llvm_variant_key *
2269 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2270 {
2271 unsigned i;
2272 struct draw_gs_llvm_variant_key *key;
2273 struct draw_sampler_static_state *draw_sampler;
2274
2275 key = (struct draw_gs_llvm_variant_key *)store;
2276
2277 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2278
2279 /* All variants of this shader will have the same value for
2280 * nr_samplers. Not yet trying to compact away holes in the
2281 * sampler array.
2282 */
2283 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2284 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2285 key->nr_sampler_views =
2286 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2287 }
2288 else {
2289 key->nr_sampler_views = key->nr_samplers;
2290 }
2291
2292 draw_sampler = key->samplers;
2293
2294 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2295
2296 for (i = 0 ; i < key->nr_samplers; i++) {
2297 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2298 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2299 }
2300 for (i = 0 ; i < key->nr_sampler_views; i++) {
2301 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2302 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2303 }
2304
2305 return key;
2306 }
2307
2308 void
2309 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2310 {
2311 unsigned i;
2312 struct draw_sampler_static_state *sampler = key->samplers;
2313
2314 for (i = 0 ; i < key->nr_sampler_views; i++) {
2315 debug_printf("sampler[%i].src_format = %s\n", i,
2316 util_format_name(sampler[i].texture_state.format));
2317 }
2318 }