projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
draw/llvm: improve debugging output a bit
[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 int_type = LLVMInt32TypeInContext(gallivm->context);
246 LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
247 LLVMTypeRef context_type;
248
249 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
250 LP_MAX_TGSI_CONST_BUFFERS);
251 elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
252 LP_MAX_TGSI_CONST_BUFFERS);
253 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
254 DRAW_TOTAL_CLIP_PLANES), 0);
255 elem_types[3] = LLVMPointerType(float_type, 0); /* viewport */
256 elem_types[4] = LLVMArrayType(texture_type,
257 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
258 elem_types[5] = LLVMArrayType(sampler_type,
259 PIPE_MAX_SAMPLERS); /* samplers */
260 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
261 Elements(elem_types), 0);
262 #if HAVE_LLVM < 0x0300
263 LLVMAddTypeName(gallivm->module, struct_name, context_type);
264
265 LLVMInvalidateStructLayout(gallivm->target, context_type);
266 #endif
267
268 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
269 target, context_type, DRAW_JIT_CTX_CONSTANTS);
270 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
271 target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
272 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
273 target, context_type, DRAW_JIT_CTX_PLANES);
274 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewport,
275 target, context_type, DRAW_JIT_CTX_VIEWPORT);
276 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
277 target, context_type,
278 DRAW_JIT_CTX_TEXTURES);
279 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
280 target, context_type,
281 DRAW_JIT_CTX_SAMPLERS);
282 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
283 target, context_type);
284
285 return context_type;
286 }
287
288
289 /**
290 * Create LLVM type for struct draw_gs_jit_context
291 */
292 static LLVMTypeRef
293 create_gs_jit_context_type(struct gallivm_state *gallivm,
294 unsigned vector_length,
295 LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
296 const char *struct_name)
297 {
298 LLVMTargetDataRef target = gallivm->target;
299 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
300 LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
301 LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
302 LLVMTypeRef context_type;
303
304 elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
305 LP_MAX_TGSI_CONST_BUFFERS);
306 elem_types[1] = LLVMArrayType(int_type, /* num_constants */
307 LP_MAX_TGSI_CONST_BUFFERS);
308 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
309 DRAW_TOTAL_CLIP_PLANES), 0);
310 elem_types[3] = LLVMPointerType(float_type, 0); /* viewport */
311
312 elem_types[4] = LLVMArrayType(texture_type,
313 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
314 elem_types[5] = LLVMArrayType(sampler_type,
315 PIPE_MAX_SAMPLERS); /* samplers */
316
317 elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
318 elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
319 vector_length), 0);
320 elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
321 vector_length), 0);
322
323 context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
324 Elements(elem_types), 0);
325 #if HAVE_LLVM < 0x0300
326 LLVMAddTypeName(gallivm->module, struct_name, context_type);
327
328 LLVMInvalidateStructLayout(gallivm->target, context_type);
329 #endif
330
331 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
332 target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
333 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
334 target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
335 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
336 target, context_type, DRAW_GS_JIT_CTX_PLANES);
337 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewport,
338 target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
339 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
340 target, context_type,
341 DRAW_GS_JIT_CTX_TEXTURES);
342 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
343 target, context_type,
344 DRAW_GS_JIT_CTX_SAMPLERS);
345 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
346 target, context_type,
347 DRAW_GS_JIT_CTX_PRIM_LENGTHS);
348 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
349 target, context_type,
350 DRAW_GS_JIT_CTX_EMITTED_VERTICES);
351 LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
352 target, context_type,
353 DRAW_GS_JIT_CTX_EMITTED_PRIMS);
354 LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
355 target, context_type);
356
357 return context_type;
358 }
359
360
361 static LLVMTypeRef
362 create_gs_jit_input_type(struct gallivm_state *gallivm)
363 {
364 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
365 LLVMTypeRef input_array;
366
367 input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
368 input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
369 input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
370 input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
371
372 return input_array;
373 }
374
375 /**
376 * Create LLVM type for struct pipe_vertex_buffer
377 */
378 static LLVMTypeRef
379 create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
380 const char *struct_name)
381 {
382 LLVMTargetDataRef target = gallivm->target;
383 LLVMTypeRef elem_types[4];
384 LLVMTypeRef vb_type;
385
386 elem_types[0] =
387 elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
388 elem_types[2] =
389 elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
390
391 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
392 Elements(elem_types), 0);
393 #if HAVE_LLVM < 0x0300
394 LLVMAddTypeName(gallivm->module, struct_name, vb_type);
395
396 LLVMInvalidateStructLayout(gallivm->target, vb_type);
397 #endif
398
399 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
400 target, vb_type, 0);
401 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
402 target, vb_type, 1);
403
404 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
405
406 return vb_type;
407 }
408
409
410 /**
411 * Create LLVM type for struct vertex_header;
412 */
413 static LLVMTypeRef
414 create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
415 {
416 LLVMTargetDataRef target = gallivm->target;
417 LLVMTypeRef elem_types[4];
418 LLVMTypeRef vertex_header;
419 char struct_name[24];
420
421 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
422
423 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
424 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
425 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
426 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
427
428 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
429 Elements(elem_types), 0);
430 #if HAVE_LLVM < 0x0300
431 LLVMAddTypeName(gallivm->module, struct_name, vertex_header);
432
433 LLVMInvalidateStructLayout(gallivm->target, vertex_header);
434 #endif
435
436 /* these are bit-fields and we can't take address of them
437 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
438 target, vertex_header,
439 DRAW_JIT_VERTEX_CLIPMASK);
440 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
441 target, vertex_header,
442 DRAW_JIT_VERTEX_EDGEFLAG);
443 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
444 target, vertex_header,
445 DRAW_JIT_VERTEX_PAD);
446 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
447 target, vertex_header,
448 DRAW_JIT_VERTEX_VERTEX_ID);
449 */
450 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
451 target, vertex_header,
452 DRAW_JIT_VERTEX_CLIP);
453 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
454 target, vertex_header,
455 DRAW_JIT_VERTEX_PRE_CLIP_POS);
456 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
457 target, vertex_header,
458 DRAW_JIT_VERTEX_DATA);
459
460 assert(LLVMABISizeOfType(target, vertex_header) ==
461 offsetof(struct vertex_header, data[data_elems]));
462
463 return vertex_header;
464 }
465
466
467 /**
468 * Create LLVM types for various structures.
469 */
470 static void
471 create_jit_types(struct draw_llvm_variant *variant)
472 {
473 struct gallivm_state *gallivm = variant->gallivm;
474 LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
475 vb_type;
476
477 texture_type = create_jit_texture_type(gallivm, "texture");
478 sampler_type = create_jit_sampler_type(gallivm, "sampler");
479
480 context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
481 "draw_jit_context");
482 variant->context_ptr_type = LLVMPointerType(context_type, 0);
483
484 buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
485 variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
486
487 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
488 variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
489 }
490
491
492 static LLVMTypeRef
493 get_context_ptr_type(struct draw_llvm_variant *variant)
494 {
495 if (!variant->context_ptr_type)
496 create_jit_types(variant);
497 return variant->context_ptr_type;
498 }
499
500
501 static LLVMTypeRef
502 get_buffer_ptr_type(struct draw_llvm_variant *variant)
503 {
504 if (!variant->buffer_ptr_type)
505 create_jit_types(variant);
506 return variant->buffer_ptr_type;
507 }
508
509
510 static LLVMTypeRef
511 get_vb_ptr_type(struct draw_llvm_variant *variant)
512 {
513 if (!variant->vb_ptr_type)
514 create_jit_types(variant);
515 return variant->vb_ptr_type;
516 }
517
518 static LLVMTypeRef
519 get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
520 {
521 if (!variant->vertex_header_ptr_type)
522 create_jit_types(variant);
523 return variant->vertex_header_ptr_type;
524 }
525
526
527 /**
528 * Create per-context LLVM info.
529 */
530 struct draw_llvm *
531 draw_llvm_create(struct draw_context *draw)
532 {
533 struct draw_llvm *llvm;
534
535 llvm = CALLOC_STRUCT( draw_llvm );
536 if (!llvm)
537 return NULL;
538
539 lp_build_init();
540
541 llvm->draw = draw;
542
543 llvm->nr_variants = 0;
544 make_empty_list(&llvm->vs_variants_list);
545
546 llvm->nr_gs_variants = 0;
547 make_empty_list(&llvm->gs_variants_list);
548
549 return llvm;
550 }
551
552
553 /**
554 * Free per-context LLVM info.
555 */
556 void
557 draw_llvm_destroy(struct draw_llvm *llvm)
558 {
559 /* XXX free other draw_llvm data? */
560 FREE(llvm);
561 }
562
563
564 /**
565 * Create LLVM-generated code for a vertex shader.
566 */
567 struct draw_llvm_variant *
568 draw_llvm_create_variant(struct draw_llvm *llvm,
569 unsigned num_inputs,
570 const struct draw_llvm_variant_key *key)
571 {
572 struct draw_llvm_variant *variant;
573 struct llvm_vertex_shader *shader =
574 llvm_vertex_shader(llvm->draw->vs.vertex_shader);
575 LLVMTypeRef vertex_header;
576
577 variant = MALLOC(sizeof *variant +
578 shader->variant_key_size -
579 sizeof variant->key);
580 if (variant == NULL)
581 return NULL;
582
583 variant->llvm = llvm;
584
585 variant->gallivm = gallivm_create();
586
587 create_jit_types(variant);
588
589 memcpy(&variant->key, key, shader->variant_key_size);
590
591 vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
592
593 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
594
595 draw_llvm_generate(llvm, variant, FALSE); /* linear */
596 draw_llvm_generate(llvm, variant, TRUE); /* elts */
597
598 gallivm_compile_module(variant->gallivm);
599
600 variant->jit_func = (draw_jit_vert_func)
601 gallivm_jit_function(variant->gallivm, variant->function);
602
603 variant->jit_func_elts = (draw_jit_vert_func_elts)
604 gallivm_jit_function(variant->gallivm, variant->function_elts);
605
606 variant->shader = shader;
607 variant->list_item_global.base = variant;
608 variant->list_item_local.base = variant;
609 /*variant->no = */shader->variants_created++;
610 variant->list_item_global.base = variant;
611
612 return variant;
613 }
614
615
616 static void
617 generate_vs(struct draw_llvm_variant *variant,
618 LLVMBuilderRef builder,
619 struct lp_type vs_type,
620 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
621 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
622 const struct lp_bld_tgsi_system_values *system_values,
623 LLVMValueRef context_ptr,
624 struct lp_build_sampler_soa *draw_sampler,
625 boolean clamp_vertex_color)
626 {
627 struct draw_llvm *llvm = variant->llvm;
628 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
629 LLVMValueRef consts_ptr =
630 draw_jit_context_vs_constants(variant->gallivm, context_ptr);
631 LLVMValueRef num_consts_ptr =
632 draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
633 struct lp_build_sampler_soa *sampler = 0;
634
635 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
636 tgsi_dump(tokens, 0);
637 draw_llvm_dump_variant_key(&variant->key);
638 }
639
640 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers)
641 sampler = draw_sampler;
642
643 lp_build_tgsi_soa(variant->gallivm,
644 tokens,
645 vs_type,
646 NULL /*struct lp_build_mask_context *mask*/,
647 consts_ptr,
648 num_consts_ptr,
649 system_values,
650 inputs,
651 outputs,
652 sampler,
653 &llvm->draw->vs.vertex_shader->info,
654 NULL);
655
656 {
657 LLVMValueRef out;
658 unsigned chan, attrib;
659 struct lp_build_context bld;
660 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
661 lp_build_context_init(&bld, variant->gallivm, vs_type);
662
663 for (attrib = 0; attrib < info->num_outputs; ++attrib) {
664 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
665 if (outputs[attrib][chan]) {
666 switch (info->output_semantic_name[attrib]) {
667 case TGSI_SEMANTIC_COLOR:
668 case TGSI_SEMANTIC_BCOLOR:
669 if (clamp_vertex_color) {
670 out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
671 out = lp_build_clamp(&bld, out, bld.zero, bld.one);
672 LLVMBuildStore(builder, out, outputs[attrib][chan]);
673 }
674 break;
675 }
676 }
677 }
678 }
679 }
680 }
681
682 static void
683 generate_fetch(struct gallivm_state *gallivm,
684 struct draw_context *draw,
685 LLVMValueRef vbuffers_ptr,
686 LLVMValueRef *res,
687 struct pipe_vertex_element *velem,
688 LLVMValueRef vbuf,
689 LLVMValueRef index,
690 LLVMValueRef instance_id)
691 {
692 const struct util_format_description *format_desc =
693 util_format_description(velem->src_format);
694 LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
695 LLVMBuilderRef builder = gallivm->builder;
696 LLVMValueRef indices =
697 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
698 velem->vertex_buffer_index, 0);
699 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
700 &indices, 1, "");
701 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
702 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
703 LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
704 LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
705 LLVMValueRef stride;
706 LLVMValueRef buffer_overflowed;
707 LLVMValueRef needed_buffer_size;
708 LLVMValueRef temp_ptr =
709 lp_build_alloca(gallivm,
710 lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
711 LLVMValueRef ofbit = NULL;
712 struct lp_build_if_state if_ctx;
713
714 if (velem->instance_divisor) {
715 /* Index is equal to the start instance plus the number of current
716 * instance divided by the divisor. In this case we compute it as:
717 * index = start_instance + (instance_id / divisor)
718 */
719 LLVMValueRef current_instance;
720 index = lp_build_const_int32(gallivm, draw->start_instance);
721 current_instance = LLVMBuildUDiv(builder, instance_id,
722 lp_build_const_int32(gallivm, velem->instance_divisor),
723 "instance_divisor");
724 index = lp_build_uadd_overflow(gallivm, index, current_instance, &ofbit);
725 }
726
727 stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
728 stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
729 stride = lp_build_uadd_overflow(
730 gallivm, stride,
731 lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
732 needed_buffer_size = lp_build_uadd_overflow(
733 gallivm, stride,
734 lp_build_const_int32(gallivm,
735 util_format_get_blocksize(velem->src_format)),
736 &ofbit);
737
738 buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
739 needed_buffer_size, buffer_size,
740 "buffer_overflowed");
741 buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
742 #if 0
743 lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
744 index, vb_stride);
745 lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
746 vb_buffer_offset,
747 lp_build_const_int32(gallivm, velem->src_offset));
748 lp_build_print_value(gallivm, " blocksize = ",
749 lp_build_const_int32(
750 gallivm,
751 util_format_get_blocksize(velem->src_format)));
752 lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
753 lp_build_printf(gallivm, " stride = %u\n", stride);
754 lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
755 lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
756 lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
757 #endif
758
759 lp_build_if(&if_ctx, gallivm, buffer_overflowed);
760 {
761 LLVMValueRef val =
762 lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
763 LLVMBuildStore(builder, val, temp_ptr);
764 }
765 lp_build_else(&if_ctx);
766 {
767 LLVMValueRef val;
768 map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
769
770 val = lp_build_fetch_rgba_aos(gallivm,
771 format_desc,
772 lp_float32_vec4_type(),
773 map_ptr,
774 zero, zero, zero);
775 LLVMBuildStore(builder, val, temp_ptr);
776 }
777 lp_build_endif(&if_ctx);
778
779 *res = LLVMBuildLoad(builder, temp_ptr, "aos");
780 }
781
782 static void
783 convert_to_soa(struct gallivm_state *gallivm,
784 LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
785 LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
786 unsigned num_attribs, const struct lp_type soa_type)
787 {
788 unsigned i, j, k;
789 struct lp_type aos_channel_type = soa_type;
790
791 debug_assert(TGSI_NUM_CHANNELS == 4);
792 debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
793
794 aos_channel_type.length >>= 1;
795
796 for (i = 0; i < num_attribs; ++i) {
797 LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
798 unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
799
800 for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
801 LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
802
803 assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
804
805 for (k = 0; k < pixels_per_channel; ++k) {
806 channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
807 }
808
809 aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
810 }
811
812 lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
813 }
814 }
815
816
817 static void
818 store_aos(struct gallivm_state *gallivm,
819 LLVMValueRef io_ptr,
820 LLVMValueRef index,
821 LLVMValueRef value)
822 {
823 LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
824 LLVMBuilderRef builder = gallivm->builder;
825 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
826 LLVMValueRef indices[3];
827
828 indices[0] = lp_build_const_int32(gallivm, 0);
829 indices[1] = index;
830 indices[2] = lp_build_const_int32(gallivm, 0);
831
832 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
833 data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
834
835 #if DEBUG_STORE
836 lp_build_printf(gallivm, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
837 #endif
838
839 /* Unaligned store due to the vertex header */
840 lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
841 }
842
843 /**
844 * Adjust the mask to architecture endianess. The mask will the store in struct:
845 *
846 * struct vertex_header {
847 * unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
848 * unsigned edgeflag:1;
849 * unsigned have_clipdist:1;
850 * unsigned vertex_id:16;
851 * [...]
852 * }
853 *
854 * On little-endian machine nothing needs to done, however on bit-endian machine
855 * the mask's fields need to be adjusted with the algorithm:
856 *
857 * uint32_t reverse (uint32_t x)
858 * {
859 * return (x >> 16) | // vertex_id
860 * ((x & 0x3fff) << 18) | // clipmask
861 * ((x & 0x4000) << 3) | // have_clipdist
862 * ((x & 0x8000) << 1); // edgeflag
863 * }
864 */
865 static LLVMValueRef
866 adjust_mask(struct gallivm_state *gallivm,
867 LLVMValueRef mask)
868 {
869 #ifdef PIPE_ARCH_BIG_ENDIAN
870 LLVMBuilderRef builder = gallivm->builder;
871 LLVMValueRef vertex_id;
872 LLVMValueRef clipmask;
873 LLVMValueRef have_clipdist;
874 LLVMValueRef edgeflag;
875
876 vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
877 clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
878 clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
879 have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
880 have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
881 edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
882 edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
883
884 mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
885 mask = LLVMBuildOr(builder, mask, have_clipdist, "");
886 mask = LLVMBuildOr(builder, mask, edgeflag, "");
887 #endif
888 return mask;
889 }
890
891 static void
892 store_aos_array(struct gallivm_state *gallivm,
893 struct lp_type soa_type,
894 LLVMValueRef io_ptr,
895 LLVMValueRef *indices,
896 LLVMValueRef* aos,
897 int attrib,
898 int num_outputs,
899 LLVMValueRef clipmask,
900 boolean have_clipdist)
901 {
902 LLVMBuilderRef builder = gallivm->builder;
903 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
904 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
905 LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
906 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
907 int vector_length = soa_type.length;
908 int i;
909
910 debug_assert(TGSI_NUM_CHANNELS == 4);
911
912 for (i = 0; i < vector_length; i++) {
913 linear_inds[i] = lp_build_const_int32(gallivm, i);
914 if (indices) {
915 inds[i] = indices[i];
916 } else {
917 inds[i] = linear_inds[i];
918 }
919 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
920 }
921
922 if (attrib == 0) {
923 /* store vertex header for each of the n vertices */
924 LLVMValueRef val, cliptmp;
925 int vertex_id_pad_edgeflag;
926
927 /* If this assertion fails, it means we need to update the bit twidding
928 * code here. See struct vertex_header in draw_private.h.
929 */
930 assert(DRAW_TOTAL_CLIP_PLANES==14);
931 /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
932 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
933 if (have_clipdist)
934 vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
935 val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
936 /* OR with the clipmask */
937 cliptmp = LLVMBuildOr(builder, val, clipmask, "");
938 for (i = 0; i < vector_length; i++) {
939 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
940 val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
941 val = adjust_mask(gallivm, val);
942 #if DEBUG_STORE
943 lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
944 io_ptrs[i], inds[i], val);
945 #endif
946 LLVMBuildStore(builder, val, id_ptr);
947 }
948 }
949
950 /* store for each of the n vertices */
951 for (i = 0; i < vector_length; i++) {
952 store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
953 }
954 }
955
956
957 static void
958 convert_to_aos(struct gallivm_state *gallivm,
959 LLVMValueRef io,
960 LLVMValueRef *indices,
961 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
962 LLVMValueRef clipmask,
963 int num_outputs,
964 struct lp_type soa_type,
965 boolean have_clipdist)
966 {
967 LLVMBuilderRef builder = gallivm->builder;
968 unsigned chan, attrib, i;
969
970 #if DEBUG_STORE
971 lp_build_printf(gallivm, " # storing begin\n");
972 #endif
973 for (attrib = 0; attrib < num_outputs; ++attrib) {
974 LLVMValueRef soa[TGSI_NUM_CHANNELS];
975 LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
976 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
977 if (outputs[attrib][chan]) {
978 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
979 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
980 #if DEBUG_STORE
981 lp_build_printf(gallivm, "output %d : %d ",
982 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
983 attrib, 0),
984 LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
985 chan, 0));
986 lp_build_print_value(gallivm, "val = ", out);
987 {
988 LLVMValueRef iv =
989 LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
990
991 lp_build_print_value(gallivm, " ival = ", iv);
992 }
993 #endif
994 soa[chan] = out;
995 }
996 else {
997 soa[chan] = 0;
998 }
999 }
1000
1001
1002 if (soa_type.length == TGSI_NUM_CHANNELS) {
1003 lp_build_transpose_aos(gallivm, soa_type, soa, aos);
1004 } else {
1005 lp_build_transpose_aos(gallivm, soa_type, soa, soa);
1006
1007 for (i = 0; i < soa_type.length; ++i) {
1008 aos[i] = lp_build_extract_range(gallivm,
1009 soa[i % TGSI_NUM_CHANNELS],
1010 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1011 TGSI_NUM_CHANNELS);
1012 }
1013 }
1014
1015 store_aos_array(gallivm,
1016 soa_type,
1017 io, indices,
1018 aos,
1019 attrib,
1020 num_outputs,
1021 clipmask, have_clipdist);
1022 }
1023 #if DEBUG_STORE
1024 lp_build_printf(gallivm, " # storing end\n");
1025 #endif
1026 }
1027
1028
1029 /**
1030 * Stores original vertex positions in clip coordinates
1031 */
1032 static void
1033 store_clip(struct gallivm_state *gallivm,
1034 const struct lp_type vs_type,
1035 LLVMValueRef io_ptr,
1036 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1037 boolean pre_clip_pos, int idx)
1038 {
1039 LLVMBuilderRef builder = gallivm->builder;
1040 LLVMValueRef soa[4];
1041 LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1042 LLVMValueRef indices[2];
1043 LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1044 LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1045 LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1046 LLVMTypeRef clip_ptr_type =
1047 LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1048 4), 0);
1049 int i, j;
1050
1051 indices[0] =
1052 indices[1] = lp_build_const_int32(gallivm, 0);
1053
1054 for (i = 0; i < vs_type.length; i++) {
1055 inds[i] = lp_build_const_int32(gallivm, i);
1056 io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1057 }
1058
1059 soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1060 soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1061 soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1062 soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1063
1064 if (!pre_clip_pos) {
1065 for (i = 0; i < vs_type.length; i++) {
1066 clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
1067 }
1068 } else {
1069 for (i = 0; i < vs_type.length; i++) {
1070 clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
1071 }
1072 }
1073
1074 lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1075 for (i = 0; i < vs_type.length; ++i) {
1076 aos[i] = lp_build_extract_range(gallivm,
1077 soa[i % TGSI_NUM_CHANNELS],
1078 (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1079 TGSI_NUM_CHANNELS);
1080 }
1081
1082 for (j = 0; j < vs_type.length; j++) {
1083 LLVMValueRef clip_ptr;
1084
1085 clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1086 clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1087
1088 /* Unaligned store */
1089 lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1090 }
1091 }
1092
1093
1094 /**
1095 * Transforms the outputs for viewport mapping
1096 */
1097 static void
1098 generate_viewport(struct draw_llvm_variant *variant,
1099 LLVMBuilderRef builder,
1100 struct lp_type vs_type,
1101 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1102 LLVMValueRef context_ptr)
1103 {
1104 int i;
1105 struct gallivm_state *gallivm = variant->gallivm;
1106 struct lp_type f32_type = vs_type;
1107 const unsigned pos = variant->llvm->draw->vs.position_output;
1108 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1109 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1110 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
1111 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
1112
1113 /* for 1/w convention*/
1114 out3 = LLVMBuildFDiv(builder, const1, out3, "");
1115 LLVMBuildStore(builder, out3, outputs[pos][3]);
1116
1117 /* Viewport Mapping */
1118 for (i=0; i<3; i++) {
1119 LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1120 LLVMValueRef scale;
1121 LLVMValueRef trans;
1122 LLVMValueRef scale_i;
1123 LLVMValueRef trans_i;
1124 LLVMValueRef index;
1125
1126 index = lp_build_const_int32(gallivm, i);
1127 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1128
1129 index = lp_build_const_int32(gallivm, i+4);
1130 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1131
1132 scale = lp_build_broadcast(gallivm, vs_type_llvm,
1133 LLVMBuildLoad(builder, scale_i, "scale"));
1134 trans = lp_build_broadcast(gallivm, vs_type_llvm,
1135 LLVMBuildLoad(builder, trans_i, "trans"));
1136
1137 /* divide by w */
1138 out = LLVMBuildFMul(builder, out, out3, "");
1139 /* mult by scale */
1140 out = LLVMBuildFMul(builder, out, scale, "");
1141 /* add translation */
1142 out = LLVMBuildFAdd(builder, out, trans, "");
1143
1144 /* store transformed outputs */
1145 LLVMBuildStore(builder, out, outputs[pos][i]);
1146 }
1147
1148 }
1149
1150
1151 /**
1152 * Returns clipmask as nxi32 bitmask for the n vertices
1153 */
1154 static LLVMValueRef
1155 generate_clipmask(struct draw_llvm *llvm,
1156 struct gallivm_state *gallivm,
1157 struct lp_type vs_type,
1158 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1159 boolean clip_xy,
1160 boolean clip_z,
1161 boolean clip_user,
1162 boolean clip_halfz,
1163 unsigned ucp_enable,
1164 LLVMValueRef context_ptr,
1165 boolean *have_clipdist)
1166 {
1167 LLVMBuilderRef builder = gallivm->builder;
1168 LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1169 LLVMValueRef test, temp;
1170 LLVMValueRef zero, shift;
1171 LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1172 LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1173 LLVMValueRef plane1, planes, plane_ptr, sum;
1174 struct lp_type f32_type = vs_type;
1175 struct lp_type i32_type = lp_int_type(vs_type);
1176 const unsigned pos = llvm->draw->vs.position_output;
1177 const unsigned cv = llvm->draw->vs.clipvertex_output;
1178 int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1179 bool have_cd = false;
1180 unsigned cd[2];
1181
1182 cd[0] = llvm->draw->vs.clipdistance_output[0];
1183 cd[1] = llvm->draw->vs.clipdistance_output[1];
1184
1185 if (cd[0] != pos || cd[1] != pos)
1186 have_cd = true;
1187
1188 if (num_written_clipdistance && !clip_user) {
1189 clip_user = true;
1190 ucp_enable = (1 << num_written_clipdistance) - 1;
1191 }
1192
1193 mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1194 temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1195 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */
1196 shift = lp_build_const_int_vec(gallivm, i32_type, 1); /* 1 1 1 1 */
1197
1198 /*
1199 * load clipvertex and position from correct locations.
1200 * if they are the same just load them once.
1201 */
1202 pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1203 pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1204 pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1205 pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1206
1207 if (clip_user && cv != pos) {
1208 cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1209 cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1210 cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1211 cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1212 } else {
1213 cv_x = pos_x;
1214 cv_y = pos_y;
1215 cv_z = pos_z;
1216 cv_w = pos_w;
1217 }
1218
1219 /* Cliptest, for hardwired planes */
1220 if (clip_xy) {
1221 /* plane 1 */
1222 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1223 temp = shift;
1224 test = LLVMBuildAnd(builder, test, temp, "");
1225 mask = test;
1226
1227 /* plane 2 */
1228 test = LLVMBuildFAdd(builder, pos_x, 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 /* plane 3 */
1235 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1236 temp = LLVMBuildShl(builder, temp, shift, "");
1237 test = LLVMBuildAnd(builder, test, temp, "");
1238 mask = LLVMBuildOr(builder, mask, test, "");
1239
1240 /* plane 4 */
1241 test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1242 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1243 temp = LLVMBuildShl(builder, temp, shift, "");
1244 test = LLVMBuildAnd(builder, test, temp, "");
1245 mask = LLVMBuildOr(builder, mask, test, "");
1246 }
1247
1248 if (clip_z) {
1249 temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1250 if (clip_halfz) {
1251 /* plane 5 */
1252 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1253 test = LLVMBuildAnd(builder, test, temp, "");
1254 mask = LLVMBuildOr(builder, mask, test, "");
1255 }
1256 else {
1257 /* plane 5 */
1258 test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1259 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1260 test = LLVMBuildAnd(builder, test, temp, "");
1261 mask = LLVMBuildOr(builder, mask, test, "");
1262 }
1263 /* plane 6 */
1264 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1265 temp = LLVMBuildShl(builder, temp, shift, "");
1266 test = LLVMBuildAnd(builder, test, temp, "");
1267 mask = LLVMBuildOr(builder, mask, test, "");
1268 }
1269
1270 if (clip_user) {
1271 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1272 LLVMValueRef indices[3];
1273 LLVMValueRef is_nan_or_inf;
1274
1275 /* userclip planes */
1276 while (ucp_enable) {
1277 unsigned plane_idx = ffs(ucp_enable)-1;
1278 ucp_enable &= ~(1 << plane_idx);
1279 plane_idx += 6;
1280
1281 if (have_cd && num_written_clipdistance) {
1282 LLVMValueRef clipdist;
1283 int i;
1284 i = plane_idx - 6;
1285
1286 *have_clipdist = TRUE;
1287 if (i < 4) {
1288 clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1289 } else {
1290 clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1291 }
1292 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1293 is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1294 test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1295 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1296 test = LLVMBuildAnd(builder, test, temp, "");
1297 mask = LLVMBuildOr(builder, mask, test, "");
1298 } else {
1299 LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1300 indices[0] = lp_build_const_int32(gallivm, 0);
1301 indices[1] = lp_build_const_int32(gallivm, plane_idx);
1302
1303 indices[2] = lp_build_const_int32(gallivm, 0);
1304 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1305 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1306 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1307 sum = LLVMBuildFMul(builder, planes, cv_x, "");
1308
1309 indices[2] = lp_build_const_int32(gallivm, 1);
1310 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1311 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1312 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1313 test = LLVMBuildFMul(builder, planes, cv_y, "");
1314 sum = LLVMBuildFAdd(builder, sum, test, "");
1315
1316 indices[2] = lp_build_const_int32(gallivm, 2);
1317 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1318 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1319 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1320 test = LLVMBuildFMul(builder, planes, cv_z, "");
1321 sum = LLVMBuildFAdd(builder, sum, test, "");
1322
1323 indices[2] = lp_build_const_int32(gallivm, 3);
1324 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1325 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1326 planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1327 test = LLVMBuildFMul(builder, planes, cv_w, "");
1328 sum = LLVMBuildFAdd(builder, sum, test, "");
1329
1330 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1331 temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
1332 test = LLVMBuildAnd(builder, test, temp, "");
1333 mask = LLVMBuildOr(builder, mask, test, "");
1334 }
1335 }
1336 }
1337 return mask;
1338 }
1339
1340
1341 /**
1342 * Returns boolean if any clipping has occurred
1343 * Used zero/non-zero i32 value to represent boolean
1344 */
1345 static LLVMValueRef
1346 clipmask_booli32(struct gallivm_state *gallivm,
1347 const struct lp_type vs_type,
1348 LLVMValueRef clipmask_bool_ptr)
1349 {
1350 LLVMBuilderRef builder = gallivm->builder;
1351 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
1352 LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1353 LLVMValueRef ret = LLVMConstNull(int32_type);
1354 LLVMValueRef temp;
1355 int i;
1356
1357 /*
1358 * Can do this with log2(vector length) pack instructions and one extract
1359 * (as we don't actually need a or) with sse2 which would be way better.
1360 */
1361 for (i=0; i < vs_type.length; i++) {
1362 temp = LLVMBuildExtractElement(builder, clipmask_bool,
1363 lp_build_const_int32(gallivm, i) , "");
1364 ret = LLVMBuildOr(builder, ret, temp, "");
1365 }
1366 return ret;
1367 }
1368
1369 static LLVMValueRef
1370 draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1371 struct lp_build_tgsi_context * bld_base,
1372 boolean is_vindex_indirect,
1373 LLVMValueRef vertex_index,
1374 boolean is_aindex_indirect,
1375 LLVMValueRef attrib_index,
1376 LLVMValueRef swizzle_index)
1377 {
1378 const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1379 struct gallivm_state *gallivm = bld_base->base.gallivm;
1380 LLVMBuilderRef builder = gallivm->builder;
1381 LLVMValueRef indices[3];
1382 LLVMValueRef res;
1383 struct lp_type type = bld_base->base.type;
1384
1385 if (is_vindex_indirect || is_aindex_indirect) {
1386 int i;
1387 res = bld_base->base.zero;
1388 for (i = 0; i < type.length; ++i) {
1389 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1390 LLVMValueRef vert_chan_index = vertex_index;
1391 LLVMValueRef attr_chan_index = attrib_index;
1392 LLVMValueRef channel_vec, value;
1393
1394 if (is_vindex_indirect) {
1395 vert_chan_index = LLVMBuildExtractElement(builder,
1396 vertex_index, idx, "");
1397 }
1398 if (is_aindex_indirect) {
1399 attr_chan_index = LLVMBuildExtractElement(builder,
1400 attrib_index, idx, "");
1401 }
1402
1403 indices[0] = vert_chan_index;
1404 indices[1] = attr_chan_index;
1405 indices[2] = swizzle_index;
1406
1407 channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1408 channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1409 value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1410
1411 res = LLVMBuildInsertElement(builder, res, value, idx, "");
1412 }
1413 } else {
1414 indices[0] = vertex_index;
1415 indices[1] = attrib_index;
1416 indices[2] = swizzle_index;
1417
1418 res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1419 res = LLVMBuildLoad(builder, res, "");
1420 }
1421
1422 return res;
1423 }
1424
1425 static void
1426 draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1427 struct lp_build_tgsi_context * bld_base,
1428 LLVMValueRef (*outputs)[4],
1429 LLVMValueRef emitted_vertices_vec)
1430 {
1431 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1432 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1433 struct gallivm_state *gallivm = variant->gallivm;
1434 LLVMBuilderRef builder = gallivm->builder;
1435 struct lp_type gs_type = bld_base->base.type;
1436 LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1437 lp_int_type(gs_type), 0);
1438 LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1439 LLVMValueRef next_prim_offset =
1440 lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1441 LLVMValueRef io = variant->io_ptr;
1442 unsigned i;
1443 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1444
1445 for (i = 0; i < gs_type.length; ++i) {
1446 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1447 LLVMValueRef currently_emitted =
1448 LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1449 indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1450 indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1451 }
1452
1453 convert_to_aos(gallivm, io, indices,
1454 outputs, clipmask,
1455 gs_info->num_outputs, gs_type,
1456 FALSE);
1457 }
1458
1459 static void
1460 draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1461 struct lp_build_tgsi_context * bld_base,
1462 LLVMValueRef verts_per_prim_vec,
1463 LLVMValueRef emitted_prims_vec)
1464 {
1465 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1466 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1467 struct gallivm_state *gallivm = variant->gallivm;
1468 LLVMBuilderRef builder = gallivm->builder;
1469 LLVMValueRef prim_lengts_ptr =
1470 draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1471 unsigned i;
1472
1473 for (i = 0; i < bld_base->base.type.length; ++i) {
1474 LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1475 LLVMValueRef prims_emitted =
1476 LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1477 LLVMValueRef store_ptr;
1478 LLVMValueRef num_vertices =
1479 LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1480
1481 store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1482 store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1483 store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1484 LLVMBuildStore(builder, num_vertices, store_ptr);
1485 }
1486 }
1487
1488 static void
1489 draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1490 struct lp_build_tgsi_context * bld_base,
1491 LLVMValueRef total_emitted_vertices_vec,
1492 LLVMValueRef emitted_prims_vec)
1493 {
1494 const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1495 struct draw_gs_llvm_variant *variant = gs_iface->variant;
1496 struct gallivm_state *gallivm = variant->gallivm;
1497 LLVMBuilderRef builder = gallivm->builder;
1498 LLVMValueRef emitted_verts_ptr =
1499 draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1500 LLVMValueRef emitted_prims_ptr =
1501 draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1502 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1503
1504 emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1505 emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1506
1507 LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1508 LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1509 }
1510
1511 static void
1512 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
1513 boolean elts)
1514 {
1515 struct gallivm_state *gallivm = variant->gallivm;
1516 LLVMContextRef context = gallivm->context;
1517 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1518 LLVMTypeRef arg_types[10];
1519 unsigned num_arg_types =
1520 elts ? Elements(arg_types) : Elements(arg_types) - 1;
1521 LLVMTypeRef func_type;
1522 LLVMValueRef context_ptr;
1523 LLVMBasicBlockRef block;
1524 LLVMBuilderRef builder;
1525 struct lp_type vs_type;
1526 LLVMValueRef end, start;
1527 LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
1528 LLVMValueRef vertex_id_offset;
1529 LLVMValueRef stride, step, io_itr;
1530 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1531 LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1532 LLVMValueRef one = lp_build_const_int32(gallivm, 1);
1533 struct draw_context *draw = llvm->draw;
1534 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1535 unsigned i, j;
1536 struct lp_build_context bld;
1537 struct lp_build_loop_state lp_loop;
1538 const int vector_length = lp_native_vector_width / 32;
1539 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1540 LLVMValueRef fetch_max;
1541 struct lp_build_sampler_soa *sampler = 0;
1542 LLVMValueRef ret, clipmask_bool_ptr;
1543 const struct draw_geometry_shader *gs = draw->gs.geometry_shader;
1544 struct draw_llvm_variant_key *key = &variant->key;
1545 /* If geometry shader is present we need to skip both the viewport
1546 * transformation and clipping otherwise the inputs to the geometry
1547 * shader will be incorrect.
1548 */
1549 const boolean bypass_viewport = gs || key->bypass_viewport;
1550 const boolean enable_cliptest = !gs && (key->clip_xy ||
1551 key->clip_z ||
1552 key->clip_user);
1553 LLVMValueRef variant_func;
1554 const unsigned pos = llvm->draw->vs.position_output;
1555 const unsigned cv = llvm->draw->vs.clipvertex_output;
1556 boolean have_clipdist = FALSE;
1557 struct lp_bld_tgsi_system_values system_values;
1558
1559 memset(&system_values, 0, sizeof(system_values));
1560
1561 i = 0;
1562 arg_types[i++] = get_context_ptr_type(variant); /* context */
1563 arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1564 arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
1565 if (elts) {
1566 arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
1567 arg_types[i++] = int32_type; /* fetch_elt_max */
1568 } else
1569 arg_types[i++] = int32_type; /* start */
1570 arg_types[i++] = int32_type; /* fetch_count / count */
1571 arg_types[i++] = int32_type; /* stride */
1572 arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
1573 arg_types[i++] = int32_type; /* instance_id */
1574 arg_types[i++] = int32_type; /* vertex_id_offset */
1575
1576 func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
1577
1578 variant_func = LLVMAddFunction(gallivm->module,
1579 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader",
1580 func_type);
1581
1582 if (elts)
1583 variant->function_elts = variant_func;
1584 else
1585 variant->function = variant_func;
1586
1587 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1588 for (i = 0; i < num_arg_types; ++i)
1589 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1590 LLVMAddAttribute(LLVMGetParam(variant_func, i),
1591 LLVMNoAliasAttribute);
1592
1593 context_ptr = LLVMGetParam(variant_func, 0);
1594 io_ptr = LLVMGetParam(variant_func, 1);
1595 vbuffers_ptr = LLVMGetParam(variant_func, 2);
1596 stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
1597 vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
1598 system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
1599 vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
1600
1601 lp_build_name(context_ptr, "context");
1602 lp_build_name(io_ptr, "io");
1603 lp_build_name(vbuffers_ptr, "vbuffers");
1604 lp_build_name(stride, "stride");
1605 lp_build_name(vb_ptr, "vb");
1606 lp_build_name(system_values.instance_id, "instance_id");
1607 lp_build_name(vertex_id_offset, "vertex_id_offset");
1608
1609 if (elts) {
1610 fetch_elts = LLVMGetParam(variant_func, 3);
1611 fetch_elt_max = LLVMGetParam(variant_func, 4);
1612 fetch_count = LLVMGetParam(variant_func, 5);
1613 lp_build_name(fetch_elts, "fetch_elts");
1614 lp_build_name(fetch_elt_max, "fetch_elt_max");
1615 lp_build_name(fetch_count, "fetch_count");
1616 start = count = NULL;
1617 }
1618 else {
1619 start = LLVMGetParam(variant_func, 3);
1620 count = LLVMGetParam(variant_func, 4);
1621 lp_build_name(start, "start");
1622 lp_build_name(count, "count");
1623 fetch_elts = fetch_count = NULL;
1624 }
1625
1626 /*
1627 * Function body
1628 */
1629
1630 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1631 builder = gallivm->builder;
1632 LLVMPositionBuilderAtEnd(builder, block);
1633
1634 lp_build_context_init(&bld, gallivm, lp_type_int(32));
1635
1636 memset(&vs_type, 0, sizeof vs_type);
1637 vs_type.floating = TRUE; /* floating point values */
1638 vs_type.sign = TRUE; /* values are signed */
1639 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
1640 vs_type.width = 32; /* 32-bit float */
1641 vs_type.length = vector_length;
1642
1643 /* hold temporary "bool" clipmask */
1644 clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
1645 LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
1646
1647 /* code generated texture sampling */
1648 sampler = draw_llvm_sampler_soa_create(
1649 draw_llvm_variant_key_samplers(key),
1650 context_ptr);
1651
1652 if (elts) {
1653 start = zero;
1654 end = fetch_count;
1655 count = fetch_count;
1656 }
1657 else {
1658 end = lp_build_add(&bld, start, count);
1659 }
1660
1661 step = lp_build_const_int32(gallivm, vector_length);
1662
1663 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
1664
1665 lp_build_loop_begin(&lp_loop, gallivm, zero);
1666 {
1667 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1668 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
1669 LLVMValueRef io;
1670 LLVMValueRef clipmask; /* holds the clipmask value */
1671 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1672
1673 io_itr = lp_loop.counter;
1674
1675 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1676 #if DEBUG_STORE
1677 lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1678 io_itr, io, lp_loop.counter);
1679 #endif
1680 system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
1681 for (i = 0; i < vector_length; ++i) {
1682 LLVMValueRef vert_index =
1683 LLVMBuildAdd(builder,
1684 lp_loop.counter,
1685 lp_build_const_int32(gallivm, i), "");
1686 LLVMValueRef true_index =
1687 LLVMBuildAdd(builder, start, vert_index, "");
1688 LLVMValueRef vertex_id;
1689
1690 /* make sure we're not out of bounds which can happen
1691 * if fetch_count % 4 != 0, because on the last iteration
1692 * a few of the 4 vertex fetches will be out of bounds */
1693 true_index = lp_build_min(&bld, true_index, fetch_max);
1694
1695 if (elts) {
1696 LLVMValueRef fetch_ptr;
1697 LLVMValueRef index_overflowed;
1698 LLVMValueRef index_ptr =
1699 lp_build_alloca(
1700 gallivm,
1701 lp_build_vec_type(gallivm, lp_type_int(32)), "");
1702 struct lp_build_if_state if_ctx;
1703 index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
1704 true_index, fetch_elt_max,
1705 "index_overflowed");
1706
1707 lp_build_if(&if_ctx, gallivm, index_overflowed);
1708 {
1709 /* Generate maximum possible index so that
1710 * generate_fetch can treat it just like
1711 * any other overflow and return zeros.
1712 * We don't have to worry about the restart
1713 * primitive index because it has already been
1714 * handled
1715 */
1716 LLVMValueRef val =
1717 lp_build_const_int32(gallivm, 0xffffffff);
1718 LLVMBuildStore(builder, val, index_ptr);
1719 }
1720 lp_build_else(&if_ctx);
1721 {
1722 LLVMValueRef val;
1723 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
1724 &true_index, 1, "");
1725 val = LLVMBuildLoad(builder, fetch_ptr, "");
1726 LLVMBuildStore(builder, val, index_ptr);
1727 }
1728 lp_build_endif(&if_ctx);
1729 true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
1730 }
1731 /* in the paths with elts vertex id has to be unaffected by the
1732 * index bias and because indices inside our elements array have
1733 * already had index bias applied we need to subtract it here to
1734 * get back to the original index.
1735 * in the linear paths vertex id has to be unaffected by the
1736 * original start index and because we abuse the 'start' variable
1737 * to either represent the actual start index or the index at which
1738 * the primitive was split (we split rendering into chunks of at
1739 * most 4095-vertices) we need to back out the original start
1740 * index out of our vertex id here.
1741 */
1742 vertex_id = LLVMBuildSub(builder, true_index, vertex_id_offset, "");
1743
1744 system_values.vertex_id = LLVMBuildInsertElement(
1745 gallivm->builder,
1746 system_values.vertex_id, vertex_id,
1747 lp_build_const_int32(gallivm, i), "");
1748
1749 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
1750 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
1751 LLVMValueRef vb_index =
1752 lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1753 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1754 generate_fetch(gallivm, draw, vbuffers_ptr,
1755 &aos_attribs[j][i], velem, vb, true_index,
1756 system_values.instance_id);
1757 }
1758 }
1759 convert_to_soa(gallivm, aos_attribs, inputs,
1760 draw->pt.nr_vertex_elements, vs_type);
1761
1762 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1763 generate_vs(variant,
1764 builder,
1765 vs_type,
1766 outputs,
1767 ptr_aos,
1768 &system_values,
1769 context_ptr,
1770 sampler,
1771 key->clamp_vertex_color);
1772
1773 if (pos != -1 && cv != -1) {
1774 /* store original positions in clip before further manipulation */
1775 store_clip(gallivm, vs_type, io, outputs, 0, cv);
1776 store_clip(gallivm, vs_type, io, outputs, 1, pos);
1777
1778 /* do cliptest */
1779 if (enable_cliptest) {
1780 LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1781 /* allocate clipmask, assign it integer type */
1782 clipmask = generate_clipmask(llvm,
1783 gallivm,
1784 vs_type,
1785 outputs,
1786 key->clip_xy,
1787 key->clip_z,
1788 key->clip_user,
1789 key->clip_halfz,
1790 key->ucp_enable,
1791 context_ptr, &have_clipdist);
1792 temp = LLVMBuildOr(builder, clipmask, temp, "");
1793 /* store temporary clipping boolean value */
1794 LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1795 }
1796 else {
1797 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1798 }
1799
1800 /* do viewport mapping */
1801 if (!bypass_viewport) {
1802 generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1803 }
1804 }
1805 else {
1806 clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
1807 }
1808
1809 /* store clipmask in vertex header,
1810 * original positions in clip
1811 * and transformed positions in data
1812 */
1813 convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1814 vs_info->num_outputs, vs_type,
1815 have_clipdist);
1816 }
1817 lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1818
1819 sampler->destroy(sampler);
1820
1821 /* return clipping boolean value for function */
1822 ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
1823
1824 LLVMBuildRet(builder, ret);
1825
1826 gallivm_verify_function(gallivm, variant_func);
1827 }
1828
1829
1830 struct draw_llvm_variant_key *
1831 draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1832 {
1833 unsigned i;
1834 struct draw_llvm_variant_key *key;
1835 struct draw_sampler_static_state *draw_sampler;
1836
1837 key = (struct draw_llvm_variant_key *)store;
1838
1839 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1840
1841 /* Presumably all variants of the shader should have the same
1842 * number of vertex elements - ie the number of shader inputs.
1843 * NOTE: we NEED to store the needed number of needed inputs
1844 * here, not the number of provided elements to match keysize
1845 * (and the offset of sampler state in the key).
1846 */
1847 key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
1848 assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
1849
1850 /* will have to rig this up properly later */
1851 key->clip_xy = llvm->draw->clip_xy;
1852 key->clip_z = llvm->draw->clip_z;
1853 key->clip_user = llvm->draw->clip_user;
1854 key->bypass_viewport = llvm->draw->identity_viewport;
1855 key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1856 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1857 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1858 key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1859 key->num_outputs = draw_total_vs_outputs(llvm->draw);
1860 key->pad1 = 0;
1861
1862 /* All variants of this shader will have the same value for
1863 * nr_samplers. Not yet trying to compact away holes in the
1864 * sampler array.
1865 */
1866 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1867 if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1868 key->nr_sampler_views =
1869 llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1870 }
1871 else {
1872 key->nr_sampler_views = key->nr_samplers;
1873 }
1874
1875 draw_sampler = draw_llvm_variant_key_samplers(key);
1876
1877 memcpy(key->vertex_element,
1878 llvm->draw->pt.vertex_element,
1879 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
1880
1881 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
1882
1883 for (i = 0 ; i < key->nr_samplers; i++) {
1884 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
1885 llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
1886 }
1887 for (i = 0 ; i < key->nr_sampler_views; i++) {
1888 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
1889 llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
1890 }
1891
1892 return key;
1893 }
1894
1895
1896 void
1897 draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
1898 {
1899 unsigned i;
1900 struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
1901
1902 debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
1903 debug_printf("clip_xy = %u\n", key->clip_xy);
1904 debug_printf("clip_z = %u\n", key->clip_z);
1905 debug_printf("clip_user = %u\n", key->clip_user);
1906 debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
1907 debug_printf("clip_halfz = %u\n", key->clip_halfz);
1908 debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
1909 debug_printf("has_gs = %u\n", key->has_gs);
1910 debug_printf("ucp_enable = %u\n", key->ucp_enable);
1911
1912 for (i = 0 ; i < key->nr_vertex_elements; i++) {
1913 debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
1914 debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
1915 debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
1916 debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
1917 }
1918
1919 for (i = 0 ; i < key->nr_sampler_views; i++) {
1920 debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
1921 }
1922 }
1923
1924
1925 void
1926 draw_llvm_set_mapped_texture(struct draw_context *draw,
1927 unsigned shader_stage,
1928 unsigned sview_idx,
1929 uint32_t width, uint32_t height, uint32_t depth,
1930 uint32_t first_level, uint32_t last_level,
1931 const void *base_ptr,
1932 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
1933 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
1934 uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
1935 {
1936 unsigned j;
1937 struct draw_jit_texture *jit_tex;
1938
1939 assert(shader_stage == PIPE_SHADER_VERTEX ||
1940 shader_stage == PIPE_SHADER_GEOMETRY);
1941
1942 if (shader_stage == PIPE_SHADER_VERTEX) {
1943 assert(sview_idx < Elements(draw->llvm->jit_context.textures));
1944
1945 jit_tex = &draw->llvm->jit_context.textures[sview_idx];
1946 } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
1947 assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
1948
1949 jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
1950 } else {
1951 assert(0);
1952 return;
1953 }
1954
1955 jit_tex->width = width;
1956 jit_tex->height = height;
1957 jit_tex->depth = depth;
1958 jit_tex->first_level = first_level;
1959 jit_tex->last_level = last_level;
1960 jit_tex->base = base_ptr;
1961
1962 for (j = first_level; j <= last_level; j++) {
1963 jit_tex->mip_offsets[j] = mip_offsets[j];
1964 jit_tex->row_stride[j] = row_stride[j];
1965 jit_tex->img_stride[j] = img_stride[j];
1966 }
1967 }
1968
1969
1970 void
1971 draw_llvm_set_sampler_state(struct draw_context *draw,
1972 unsigned shader_type)
1973 {
1974 unsigned i;
1975
1976 if (shader_type == PIPE_SHADER_VERTEX) {
1977 for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
1978 struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
1979
1980 if (draw->samplers[i]) {
1981 const struct pipe_sampler_state *s
1982 = draw->samplers[PIPE_SHADER_VERTEX][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 } else if (shader_type == PIPE_SHADER_GEOMETRY) {
1990 for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
1991 struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
1992
1993 if (draw->samplers[i]) {
1994 const struct pipe_sampler_state *s
1995 = draw->samplers[PIPE_SHADER_GEOMETRY][i];
1996 jit_sam->min_lod = s->min_lod;
1997 jit_sam->max_lod = s->max_lod;
1998 jit_sam->lod_bias = s->lod_bias;
1999 COPY_4V(jit_sam->border_color, s->border_color.f);
2000 }
2001 }
2002 }
2003 }
2004
2005
2006 void
2007 draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2008 {
2009 struct draw_llvm *llvm = variant->llvm;
2010
2011 if (variant->function_elts) {
2012 gallivm_free_function(variant->gallivm,
2013 variant->function_elts, variant->jit_func_elts);
2014 }
2015
2016 if (variant->function) {
2017 gallivm_free_function(variant->gallivm,
2018 variant->function, variant->jit_func);
2019 }
2020
2021 gallivm_destroy(variant->gallivm);
2022
2023 remove_from_list(&variant->list_item_local);
2024 variant->shader->variants_cached--;
2025 remove_from_list(&variant->list_item_global);
2026 llvm->nr_variants--;
2027 FREE(variant);
2028 }
2029
2030
2031 /**
2032 * Create LLVM types for various structures.
2033 */
2034 static void
2035 create_gs_jit_types(struct draw_gs_llvm_variant *var)
2036 {
2037 struct gallivm_state *gallivm = var->gallivm;
2038 LLVMTypeRef texture_type, sampler_type, context_type;
2039
2040 texture_type = create_jit_texture_type(gallivm, "texture");
2041 sampler_type = create_jit_sampler_type(gallivm, "sampler");
2042
2043 context_type = create_gs_jit_context_type(gallivm,
2044 var->shader->base.vector_length,
2045 texture_type, sampler_type,
2046 "draw_gs_jit_context");
2047 var->context_ptr_type = LLVMPointerType(context_type, 0);
2048
2049 var->input_array_type = create_gs_jit_input_type(gallivm);
2050 }
2051
2052 static LLVMTypeRef
2053 get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2054 {
2055 if (!variant->context_ptr_type)
2056 create_gs_jit_types(variant);
2057 return variant->context_ptr_type;
2058 }
2059
2060 static LLVMValueRef
2061 generate_mask_value(struct draw_gs_llvm_variant *variant,
2062 struct lp_type gs_type)
2063 {
2064 struct gallivm_state *gallivm = variant->gallivm;
2065 LLVMBuilderRef builder = gallivm->builder;
2066 struct lp_type mask_type = lp_int_type(gs_type);
2067 LLVMValueRef num_prims;
2068 LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2069 unsigned i;
2070
2071 num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2072 variant->num_prims);
2073 for (i = 0; i <= gs_type.length; i++) {
2074 LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2075 mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2076 }
2077 mask_val = lp_build_compare(gallivm, mask_type,
2078 PIPE_FUNC_GREATER, num_prims, mask_val);
2079
2080 return mask_val;
2081 }
2082
2083 static void
2084 draw_gs_llvm_generate(struct draw_llvm *llvm,
2085 struct draw_gs_llvm_variant *variant)
2086 {
2087 struct gallivm_state *gallivm = variant->gallivm;
2088 LLVMContextRef context = gallivm->context;
2089 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2090 LLVMTypeRef arg_types[6];
2091 LLVMTypeRef func_type;
2092 LLVMValueRef variant_func;
2093 LLVMValueRef context_ptr;
2094 LLVMValueRef prim_id_ptr;
2095 LLVMBasicBlockRef block;
2096 LLVMBuilderRef builder;
2097 LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2098 struct lp_build_sampler_soa *sampler = 0;
2099 struct lp_build_context bld;
2100 struct lp_bld_tgsi_system_values system_values;
2101 struct lp_type gs_type;
2102 unsigned i;
2103 struct draw_gs_llvm_iface gs_iface;
2104 const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2105 LLVMValueRef consts_ptr, num_consts_ptr;
2106 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2107 struct lp_build_mask_context mask;
2108 const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2109 unsigned vector_length = variant->shader->base.vector_length;
2110
2111 memset(&system_values, 0, sizeof(system_values));
2112
2113 assert(variant->vertex_header_ptr_type);
2114
2115 arg_types[0] = get_gs_context_ptr_type(variant); /* context */
2116 arg_types[1] = variant->input_array_type; /* input */
2117 arg_types[2] = variant->vertex_header_ptr_type; /* vertex_header */
2118 arg_types[3] = int32_type; /* num_prims */
2119 arg_types[4] = int32_type; /* instance_id */
2120 arg_types[5] = LLVMPointerType(
2121 LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
2122
2123 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
2124
2125 variant_func = LLVMAddFunction(gallivm->module, "draw_geometry_shader",
2126 func_type);
2127 variant->function = variant_func;
2128
2129 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2130
2131 for (i = 0; i < Elements(arg_types); ++i)
2132 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2133 LLVMAddAttribute(LLVMGetParam(variant_func, i),
2134 LLVMNoAliasAttribute);
2135
2136 context_ptr = LLVMGetParam(variant_func, 0);
2137 input_array = LLVMGetParam(variant_func, 1);
2138 io_ptr = LLVMGetParam(variant_func, 2);
2139 num_prims = LLVMGetParam(variant_func, 3);
2140 system_values.instance_id = LLVMGetParam(variant_func, 4);
2141 prim_id_ptr = LLVMGetParam(variant_func, 5);
2142
2143 lp_build_name(context_ptr, "context");
2144 lp_build_name(input_array, "input");
2145 lp_build_name(io_ptr, "io");
2146 lp_build_name(num_prims, "num_prims");
2147 lp_build_name(system_values.instance_id, "instance_id");
2148 lp_build_name(prim_id_ptr, "prim_id_ptr");
2149
2150 variant->context_ptr = context_ptr;
2151 variant->io_ptr = io_ptr;
2152 variant->num_prims = num_prims;
2153
2154 gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2155 gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2156 gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2157 gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2158 gs_iface.input = input_array;
2159 gs_iface.variant = variant;
2160
2161 /*
2162 * Function body
2163 */
2164
2165 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2166 builder = gallivm->builder;
2167 LLVMPositionBuilderAtEnd(builder, block);
2168
2169 lp_build_context_init(&bld, gallivm, lp_type_int(32));
2170
2171 memset(&gs_type, 0, sizeof gs_type);
2172 gs_type.floating = TRUE; /* floating point values */
2173 gs_type.sign = TRUE; /* values are signed */
2174 gs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
2175 gs_type.width = 32; /* 32-bit float */
2176 gs_type.length = vector_length;
2177
2178 consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2179 num_consts_ptr =
2180 draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2181
2182 /* code generated texture sampling */
2183 sampler = draw_llvm_sampler_soa_create(variant->key.samplers,
2184 context_ptr);
2185
2186 mask_val = generate_mask_value(variant, gs_type);
2187 lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2188
2189 if (gs_info->uses_primid) {
2190 system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2191 }
2192
2193 if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2194 tgsi_dump(tokens, 0);
2195 draw_gs_llvm_dump_variant_key(&variant->key);
2196 }
2197
2198 lp_build_tgsi_soa(variant->gallivm,
2199 tokens,
2200 gs_type,
2201 &mask,
2202 consts_ptr,
2203 num_consts_ptr,
2204 &system_values,
2205 NULL,
2206 outputs,
2207 sampler,
2208 &llvm->draw->gs.geometry_shader->info,
2209 (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2210
2211 sampler->destroy(sampler);
2212
2213 lp_build_mask_end(&mask);
2214
2215 LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2216
2217 gallivm_verify_function(gallivm, variant_func);
2218 }
2219
2220
2221 struct draw_gs_llvm_variant *
2222 draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2223 unsigned num_outputs,
2224 const struct draw_gs_llvm_variant_key *key)
2225 {
2226 struct draw_gs_llvm_variant *variant;
2227 struct llvm_geometry_shader *shader =
2228 llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2229 LLVMTypeRef vertex_header;
2230
2231 variant = MALLOC(sizeof *variant +
2232 shader->variant_key_size -
2233 sizeof variant->key);
2234 if (variant == NULL)
2235 return NULL;
2236
2237 variant->llvm = llvm;
2238 variant->shader = shader;
2239
2240 variant->gallivm = gallivm_create();
2241
2242 create_gs_jit_types(variant);
2243
2244 memcpy(&variant->key, key, shader->variant_key_size);
2245
2246 vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2247
2248 variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2249
2250 draw_gs_llvm_generate(llvm, variant);
2251
2252 gallivm_compile_module(variant->gallivm);
2253
2254 variant->jit_func = (draw_gs_jit_func)
2255 gallivm_jit_function(variant->gallivm, variant->function);
2256
2257 variant->list_item_global.base = variant;
2258 variant->list_item_local.base = variant;
2259 /*variant->no = */shader->variants_created++;
2260 variant->list_item_global.base = variant;
2261
2262 return variant;
2263 }
2264
2265 void
2266 draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2267 {
2268 struct draw_llvm *llvm = variant->llvm;
2269
2270 if (variant->function) {
2271 gallivm_free_function(variant->gallivm,
2272 variant->function, variant->jit_func);
2273 }
2274
2275 gallivm_destroy(variant->gallivm);
2276
2277 remove_from_list(&variant->list_item_local);
2278 variant->shader->variants_cached--;
2279 remove_from_list(&variant->list_item_global);
2280 llvm->nr_gs_variants--;
2281 FREE(variant);
2282 }
2283
2284 struct draw_gs_llvm_variant_key *
2285 draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2286 {
2287 unsigned i;
2288 struct draw_gs_llvm_variant_key *key;
2289 struct draw_sampler_static_state *draw_sampler;
2290
2291 key = (struct draw_gs_llvm_variant_key *)store;
2292
2293 key->num_outputs = draw_total_gs_outputs(llvm->draw);
2294
2295 /* All variants of this shader will have the same value for
2296 * nr_samplers. Not yet trying to compact away holes in the
2297 * sampler array.
2298 */
2299 key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2300 if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2301 key->nr_sampler_views =
2302 llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2303 }
2304 else {
2305 key->nr_sampler_views = key->nr_samplers;
2306 }
2307
2308 draw_sampler = key->samplers;
2309
2310 memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2311
2312 for (i = 0 ; i < key->nr_samplers; i++) {
2313 lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2314 llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2315 }
2316 for (i = 0 ; i < key->nr_sampler_views; i++) {
2317 lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2318 llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2319 }
2320
2321 return key;
2322 }
2323
2324 void
2325 draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2326 {
2327 unsigned i;
2328 struct draw_sampler_static_state *sampler = key->samplers;
2329
2330 for (i = 0 ; i < key->nr_sampler_views; i++) {
2331 debug_printf("sampler[%i].src_format = %s\n", i,
2332 util_format_name(sampler[i].texture_state.format));
2333 }
2334 }