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_bitarit.h"
37#include "gallivm/lp_bld_gather.h"
38#include "gallivm/lp_bld_logic.h"
39#include "gallivm/lp_bld_const.h"
40#include "gallivm/lp_bld_swizzle.h"
41#include "gallivm/lp_bld_struct.h"
42#include "gallivm/lp_bld_type.h"
43#include "gallivm/lp_bld_flow.h"
44#include "gallivm/lp_bld_debug.h"
45#include "gallivm/lp_bld_tgsi.h"
46#include "gallivm/lp_bld_printf.h"
47#include "gallivm/lp_bld_intr.h"
48#include "gallivm/lp_bld_init.h"
49#include "gallivm/lp_bld_type.h"
50#include "gallivm/lp_bld_pack.h"
51#include "gallivm/lp_bld_format.h"
52
53#include "tgsi/tgsi_exec.h"
54#include "tgsi/tgsi_dump.h"
55
56#include "util/u_math.h"
57#include "util/u_pointer.h"
58#include "util/u_string.h"
59#include "util/simple_list.h"
60
61
62#define DEBUG_STORE 0
63
64
65static void
66draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
67
68
69struct draw_gs_llvm_iface {
70   struct lp_build_tgsi_gs_iface base;
71
72   struct draw_gs_llvm_variant *variant;
73   LLVMValueRef input;
74};
75
76static inline const struct draw_gs_llvm_iface *
77draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
78{
79   return (const struct draw_gs_llvm_iface *)iface;
80}
81
82/**
83 * Create LLVM type for draw_vertex_buffer.
84 */
85static LLVMTypeRef
86create_jit_dvbuffer_type(struct gallivm_state *gallivm,
87                         const char *struct_name)
88{
89   LLVMTargetDataRef target = gallivm->target;
90   LLVMTypeRef dvbuffer_type;
91   LLVMTypeRef elem_types[DRAW_JIT_DVBUFFER_NUM_FIELDS];
92   LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
93
94   elem_types[DRAW_JIT_DVBUFFER_MAP] =
95      LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0);
96   elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
97
98   dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
99                                           ARRAY_SIZE(elem_types), 0);
100
101   (void) target; /* silence unused var warning for non-debug build */
102   LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
103                          target, dvbuffer_type,
104                          DRAW_JIT_DVBUFFER_MAP);
105   LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, size,
106                          target, dvbuffer_type,
107                          DRAW_JIT_DVBUFFER_SIZE);
108
109   return dvbuffer_type;
110}
111
112/**
113 * Create LLVM type for struct draw_jit_texture
114 */
115static LLVMTypeRef
116create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name)
117{
118   LLVMTargetDataRef target = gallivm->target;
119   LLVMTypeRef texture_type;
120   LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS];
121   LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
122
123   elem_types[DRAW_JIT_TEXTURE_WIDTH]  =
124   elem_types[DRAW_JIT_TEXTURE_HEIGHT] =
125   elem_types[DRAW_JIT_TEXTURE_DEPTH] =
126   elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] =
127   elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type;
128   elem_types[DRAW_JIT_TEXTURE_BASE] =
129      LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
130   elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
131   elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
132   elem_types[DRAW_JIT_TEXTURE_MIP_OFFSETS] =
133      LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
134
135   texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
136                                          ARRAY_SIZE(elem_types), 0);
137
138   (void) target; /* silence unused var warning for non-debug build */
139   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
140                          target, texture_type,
141                          DRAW_JIT_TEXTURE_WIDTH);
142   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
143                          target, texture_type,
144                          DRAW_JIT_TEXTURE_HEIGHT);
145   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth,
146                          target, texture_type,
147                          DRAW_JIT_TEXTURE_DEPTH);
148   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level,
149                          target, texture_type,
150                          DRAW_JIT_TEXTURE_FIRST_LEVEL);
151   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level,
152                          target, texture_type,
153                          DRAW_JIT_TEXTURE_LAST_LEVEL);
154   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, base,
155                          target, texture_type,
156                          DRAW_JIT_TEXTURE_BASE);
157   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride,
158                          target, texture_type,
159                          DRAW_JIT_TEXTURE_ROW_STRIDE);
160   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride,
161                          target, texture_type,
162                          DRAW_JIT_TEXTURE_IMG_STRIDE);
163   LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, mip_offsets,
164                          target, texture_type,
165                          DRAW_JIT_TEXTURE_MIP_OFFSETS);
166
167   LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type);
168
169   return texture_type;
170}
171
172
173/**
174 * Create LLVM type for struct draw_jit_sampler
175 */
176static LLVMTypeRef
177create_jit_sampler_type(struct gallivm_state *gallivm, const char *struct_name)
178{
179   LLVMTargetDataRef target = gallivm->target;
180   LLVMTypeRef sampler_type;
181   LLVMTypeRef elem_types[DRAW_JIT_SAMPLER_NUM_FIELDS];
182
183   elem_types[DRAW_JIT_SAMPLER_MIN_LOD] =
184   elem_types[DRAW_JIT_SAMPLER_MAX_LOD] =
185   elem_types[DRAW_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context);
186   elem_types[DRAW_JIT_SAMPLER_BORDER_COLOR] =
187      LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
188
189   sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
190                                          ARRAY_SIZE(elem_types), 0);
191
192   (void) target; /* silence unused var warning for non-debug build */
193   LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
194                          target, sampler_type,
195                          DRAW_JIT_SAMPLER_MIN_LOD);
196   LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, max_lod,
197                          target, sampler_type,
198                          DRAW_JIT_SAMPLER_MAX_LOD);
199   LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, lod_bias,
200                          target, sampler_type,
201                          DRAW_JIT_SAMPLER_LOD_BIAS);
202   LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, border_color,
203                          target, sampler_type,
204                          DRAW_JIT_SAMPLER_BORDER_COLOR);
205
206   LP_CHECK_STRUCT_SIZE(struct draw_jit_sampler, target, sampler_type);
207
208   return sampler_type;
209}
210
211
212/**
213 * Create LLVM type for struct draw_jit_context
214 */
215static LLVMTypeRef
216create_jit_context_type(struct gallivm_state *gallivm,
217                        LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
218                        const char *struct_name)
219{
220   LLVMTargetDataRef target = gallivm->target;
221   LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
222   LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
223   LLVMTypeRef elem_types[DRAW_JIT_CTX_NUM_FIELDS];
224   LLVMTypeRef context_type;
225
226   elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* vs_constants */
227                                 LP_MAX_TGSI_CONST_BUFFERS);
228   elem_types[1] = LLVMArrayType(int_type, /* num_vs_constants */
229                                 LP_MAX_TGSI_CONST_BUFFERS);
230   elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
231                                                 DRAW_TOTAL_CLIP_PLANES), 0);
232   elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
233   elem_types[4] = LLVMArrayType(texture_type,
234                                 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
235   elem_types[5] = LLVMArrayType(sampler_type,
236                                 PIPE_MAX_SAMPLERS); /* samplers */
237   context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
238                                          ARRAY_SIZE(elem_types), 0);
239
240   (void) target; /* silence unused var warning for non-debug build */
241   LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
242                          target, context_type, DRAW_JIT_CTX_CONSTANTS);
243   LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
244                          target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
245   LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
246                          target, context_type, DRAW_JIT_CTX_PLANES);
247   LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
248                          target, context_type, DRAW_JIT_CTX_VIEWPORT);
249   LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
250                          target, context_type,
251                          DRAW_JIT_CTX_TEXTURES);
252   LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
253                          target, context_type,
254                          DRAW_JIT_CTX_SAMPLERS);
255   LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
256                        target, context_type);
257
258   return context_type;
259}
260
261
262/**
263 * Create LLVM type for struct draw_gs_jit_context
264 */
265static LLVMTypeRef
266create_gs_jit_context_type(struct gallivm_state *gallivm,
267                           unsigned vector_length,
268                           LLVMTypeRef texture_type, LLVMTypeRef sampler_type,
269                           const char *struct_name)
270{
271   LLVMTargetDataRef target = gallivm->target;
272   LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
273   LLVMTypeRef int_type = LLVMInt32TypeInContext(gallivm->context);
274   LLVMTypeRef elem_types[DRAW_GS_JIT_CTX_NUM_FIELDS];
275   LLVMTypeRef context_type;
276
277   elem_types[0] = LLVMArrayType(LLVMPointerType(float_type, 0), /* constants */
278                                 LP_MAX_TGSI_CONST_BUFFERS);
279   elem_types[1] = LLVMArrayType(int_type, /* num_constants */
280                                 LP_MAX_TGSI_CONST_BUFFERS);
281   elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
282                                                 DRAW_TOTAL_CLIP_PLANES), 0);
283   elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
284
285   elem_types[4] = LLVMArrayType(texture_type,
286                                 PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
287   elem_types[5] = LLVMArrayType(sampler_type,
288                                 PIPE_MAX_SAMPLERS); /* samplers */
289
290   elem_types[6] = LLVMPointerType(LLVMPointerType(int_type, 0), 0);
291   elem_types[7] = LLVMPointerType(LLVMVectorType(int_type,
292                                                  vector_length), 0);
293   elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
294                                                  vector_length), 0);
295
296   context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
297                                          ARRAY_SIZE(elem_types), 0);
298
299   (void) target; /* silence unused var warning for non-debug build */
300   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
301                          target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
302   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
303                          target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
304   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
305                          target, context_type, DRAW_GS_JIT_CTX_PLANES);
306   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
307                          target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
308   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
309                          target, context_type,
310                          DRAW_GS_JIT_CTX_TEXTURES);
311   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, samplers,
312                          target, context_type,
313                          DRAW_GS_JIT_CTX_SAMPLERS);
314   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, prim_lengths,
315                          target, context_type,
316                          DRAW_GS_JIT_CTX_PRIM_LENGTHS);
317   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_vertices,
318                          target, context_type,
319                          DRAW_GS_JIT_CTX_EMITTED_VERTICES);
320   LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
321                          target, context_type,
322                          DRAW_GS_JIT_CTX_EMITTED_PRIMS);
323   LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
324                        target, context_type);
325
326   return context_type;
327}
328
329
330static LLVMTypeRef
331create_gs_jit_input_type(struct gallivm_state *gallivm)
332{
333   LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context);
334   LLVMTypeRef input_array;
335
336   input_array = LLVMVectorType(float_type, TGSI_NUM_CHANNELS); /* num primitives */
337   input_array = LLVMArrayType(input_array, TGSI_NUM_CHANNELS); /* num channels */
338   input_array = LLVMArrayType(input_array, PIPE_MAX_SHADER_INPUTS); /* num attrs per vertex */
339   input_array = LLVMPointerType(input_array, 0); /* num vertices per prim */
340
341   return input_array;
342}
343
344/**
345 * Create LLVM type for struct pipe_vertex_buffer
346 */
347static LLVMTypeRef
348create_jit_vertex_buffer_type(struct gallivm_state *gallivm,
349                              const char *struct_name)
350{
351   LLVMTargetDataRef target = gallivm->target;
352   LLVMTypeRef elem_types[4];
353   LLVMTypeRef vb_type;
354
355   elem_types[0] = LLVMInt16TypeInContext(gallivm->context);
356   elem_types[1] = LLVMInt8TypeInContext(gallivm->context);
357   elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
358   elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
359
360   vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
361                                     ARRAY_SIZE(elem_types), 0);
362
363   (void) target; /* silence unused var warning for non-debug build */
364   LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
365                          target, vb_type, 0);
366   LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
367                          target, vb_type, 1);
368   LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
369                          target, vb_type, 2);
370   LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
371                          target, vb_type, 3);
372
373   LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
374
375   return vb_type;
376}
377
378
379/**
380 * Create LLVM type for struct vertex_header;
381 */
382static LLVMTypeRef
383create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
384{
385   LLVMTargetDataRef target = gallivm->target;
386   LLVMTypeRef elem_types[3];
387   LLVMTypeRef vertex_header;
388   char struct_name[24];
389
390   util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
391
392   elem_types[DRAW_JIT_VERTEX_VERTEX_ID]  = LLVMIntTypeInContext(gallivm->context, 32);
393   elem_types[DRAW_JIT_VERTEX_CLIP_POS]  = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
394   elem_types[DRAW_JIT_VERTEX_DATA]  = LLVMArrayType(elem_types[1], data_elems);
395
396   vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
397                                           ARRAY_SIZE(elem_types), 0);
398
399   /* these are bit-fields and we can't take address of them
400      LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
401      target, vertex_header,
402      DRAW_JIT_VERTEX_CLIPMASK);
403      LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
404      target, vertex_header,
405      DRAW_JIT_VERTEX_EDGEFLAG);
406      LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
407      target, vertex_header,
408      DRAW_JIT_VERTEX_PAD);
409      LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
410      target, vertex_header,
411      DRAW_JIT_VERTEX_VERTEX_ID);
412   */
413   (void) target; /* silence unused var warning for non-debug build */
414   LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
415                          target, vertex_header,
416                          DRAW_JIT_VERTEX_CLIP_POS);
417   LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
418                          target, vertex_header,
419                          DRAW_JIT_VERTEX_DATA);
420
421   assert(LLVMABISizeOfType(target, vertex_header) ==
422          offsetof(struct vertex_header, data[data_elems]));
423
424   return vertex_header;
425}
426
427
428/**
429 * Create LLVM types for various structures.
430 */
431static void
432create_jit_types(struct draw_llvm_variant *variant)
433{
434   struct gallivm_state *gallivm = variant->gallivm;
435   LLVMTypeRef texture_type, sampler_type, context_type, buffer_type,
436      vb_type;
437
438   texture_type = create_jit_texture_type(gallivm, "texture");
439   sampler_type = create_jit_sampler_type(gallivm, "sampler");
440
441   context_type = create_jit_context_type(gallivm, texture_type, sampler_type,
442                                          "draw_jit_context");
443   variant->context_ptr_type = LLVMPointerType(context_type, 0);
444
445   buffer_type = create_jit_dvbuffer_type(gallivm, "draw_vertex_buffer");
446   variant->buffer_ptr_type = LLVMPointerType(buffer_type, 0);
447
448   vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer");
449   variant->vb_ptr_type = LLVMPointerType(vb_type, 0);
450}
451
452
453static LLVMTypeRef
454get_context_ptr_type(struct draw_llvm_variant *variant)
455{
456   if (!variant->context_ptr_type)
457      create_jit_types(variant);
458   return variant->context_ptr_type;
459}
460
461
462static LLVMTypeRef
463get_buffer_ptr_type(struct draw_llvm_variant *variant)
464{
465   if (!variant->buffer_ptr_type)
466      create_jit_types(variant);
467   return variant->buffer_ptr_type;
468}
469
470
471static LLVMTypeRef
472get_vb_ptr_type(struct draw_llvm_variant *variant)
473{
474   if (!variant->vb_ptr_type)
475      create_jit_types(variant);
476   return variant->vb_ptr_type;
477}
478
479static LLVMTypeRef
480get_vertex_header_ptr_type(struct draw_llvm_variant *variant)
481{
482   if (!variant->vertex_header_ptr_type)
483      create_jit_types(variant);
484   return variant->vertex_header_ptr_type;
485}
486
487
488/**
489 * Create per-context LLVM info.
490 */
491struct draw_llvm *
492draw_llvm_create(struct draw_context *draw, LLVMContextRef context)
493{
494   struct draw_llvm *llvm;
495
496   if (!lp_build_init())
497      return NULL;
498
499   llvm = CALLOC_STRUCT( draw_llvm );
500   if (!llvm)
501      return NULL;
502
503   llvm->draw = draw;
504
505   llvm->context = context;
506   if (!llvm->context) {
507      llvm->context = LLVMContextCreate();
508      llvm->context_owned = true;
509   }
510   if (!llvm->context)
511      goto fail;
512
513   llvm->nr_variants = 0;
514   make_empty_list(&llvm->vs_variants_list);
515
516   llvm->nr_gs_variants = 0;
517   make_empty_list(&llvm->gs_variants_list);
518
519   return llvm;
520
521fail:
522   draw_llvm_destroy(llvm);
523   return NULL;
524}
525
526
527/**
528 * Free per-context LLVM info.
529 */
530void
531draw_llvm_destroy(struct draw_llvm *llvm)
532{
533   if (llvm->context_owned)
534      LLVMContextDispose(llvm->context);
535   llvm->context = NULL;
536
537   /* XXX free other draw_llvm data? */
538   FREE(llvm);
539}
540
541
542/**
543 * Create LLVM-generated code for a vertex shader.
544 */
545struct draw_llvm_variant *
546draw_llvm_create_variant(struct draw_llvm *llvm,
547                         unsigned num_inputs,
548                         const struct draw_llvm_variant_key *key)
549{
550   struct draw_llvm_variant *variant;
551   struct llvm_vertex_shader *shader =
552      llvm_vertex_shader(llvm->draw->vs.vertex_shader);
553   LLVMTypeRef vertex_header;
554   char module_name[64];
555
556   variant = MALLOC(sizeof *variant +
557                    shader->variant_key_size -
558                    sizeof variant->key);
559   if (!variant)
560      return NULL;
561
562   variant->llvm = llvm;
563   variant->shader = shader;
564
565   util_snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
566                 variant->shader->variants_cached);
567
568   variant->gallivm = gallivm_create(module_name, llvm->context);
569
570   create_jit_types(variant);
571
572   memcpy(&variant->key, key, shader->variant_key_size);
573
574   if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
575      tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
576      draw_llvm_dump_variant_key(&variant->key);
577   }
578
579   vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
580
581   variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
582
583   draw_llvm_generate(llvm, variant);
584
585   gallivm_compile_module(variant->gallivm);
586
587   variant->jit_func = (draw_jit_vert_func)
588         gallivm_jit_function(variant->gallivm, variant->function);
589
590   gallivm_free_ir(variant->gallivm);
591
592   variant->list_item_global.base = variant;
593   variant->list_item_local.base = variant;
594   /*variant->no = */shader->variants_created++;
595   variant->list_item_global.base = variant;
596
597   return variant;
598}
599
600
601static void
602generate_vs(struct draw_llvm_variant *variant,
603            LLVMBuilderRef builder,
604            struct lp_type vs_type,
605            LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
606            const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
607            const struct lp_bld_tgsi_system_values *system_values,
608            LLVMValueRef context_ptr,
609            const struct lp_build_sampler_soa *draw_sampler,
610            boolean clamp_vertex_color)
611{
612   struct draw_llvm *llvm = variant->llvm;
613   const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
614   LLVMValueRef consts_ptr =
615      draw_jit_context_vs_constants(variant->gallivm, context_ptr);
616   LLVMValueRef num_consts_ptr =
617      draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
618
619   lp_build_tgsi_soa(variant->gallivm,
620                     tokens,
621                     vs_type,
622                     NULL /*struct lp_build_mask_context *mask*/,
623                     consts_ptr,
624                     num_consts_ptr,
625                     system_values,
626                     inputs,
627                     outputs,
628                     context_ptr,
629                     NULL,
630                     draw_sampler,
631                     &llvm->draw->vs.vertex_shader->info,
632                     NULL);
633
634   {
635      LLVMValueRef out;
636      unsigned chan, attrib;
637      struct lp_build_context bld;
638      struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info;
639      lp_build_context_init(&bld, variant->gallivm, vs_type);
640
641      for (attrib = 0; attrib < info->num_outputs; ++attrib) {
642         for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
643            if (outputs[attrib][chan]) {
644               switch (info->output_semantic_name[attrib]) {
645               case TGSI_SEMANTIC_COLOR:
646               case TGSI_SEMANTIC_BCOLOR:
647                  if (clamp_vertex_color) {
648                     out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
649                     out = lp_build_clamp(&bld, out, bld.zero, bld.one);
650                     LLVMBuildStore(builder, out, outputs[attrib][chan]);
651                  }
652                  break;
653               }
654            }
655         }
656      }
657   }
658}
659
660
661static void
662fetch_instanced(struct gallivm_state *gallivm,
663                const struct util_format_description *format_desc,
664                struct lp_type vs_type,
665                LLVMValueRef vb_stride,
666                LLVMValueRef map_ptr,
667                LLVMValueRef buffer_size_adj,
668                LLVMValueRef *inputs,
669                LLVMValueRef index)
670{
671   LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
672   LLVMTypeRef aosf_t, aosi_t;
673   LLVMValueRef zero = LLVMConstNull(i32_t);
674   LLVMBuilderRef builder = gallivm->builder;
675   LLVMValueRef stride, buffer_overflowed, aos, index_valid;
676   unsigned i;
677
678   aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
679   aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
680
681   /* This mul can overflow. Wraparound is ok. */
682   stride = LLVMBuildMul(builder, vb_stride, index, "");
683
684   buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
685                                     stride, buffer_size_adj,
686                                     "buffer_overflowed");
687
688   if (0) {
689      lp_build_print_value(gallivm, "   instance index = ", index);
690      lp_build_print_value(gallivm, "   buffer overflowed = ", buffer_overflowed);
691   }
692
693   index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
694   index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
695   stride = LLVMBuildAnd(builder, stride, index_valid, "");
696
697   aos = lp_build_fetch_rgba_aos(gallivm,
698                                 format_desc,
699                                 lp_float32_vec4_type(),
700                                 FALSE,
701                                 map_ptr,
702                                 stride, zero, zero,
703                                 NULL);
704
705   index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
706   aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
707   aos = LLVMBuildAnd(builder, aos, index_valid, "");
708   aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
709
710   for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
711      LLVMValueRef index = lp_build_const_int32(gallivm, i);
712      inputs[i] = lp_build_extract_broadcast(gallivm,
713                                             lp_float32_vec4_type(),
714                                             vs_type, aos, index);
715   }
716}
717
718
719static void
720fetch_vector(struct gallivm_state *gallivm,
721             const struct util_format_description *format_desc,
722             struct lp_type vs_type,
723             LLVMValueRef vb_stride,
724             LLVMValueRef map_ptr,
725             LLVMValueRef buffer_size_adj,
726             LLVMValueRef *inputs,
727             LLVMValueRef indices)
728{
729   LLVMBuilderRef builder = gallivm->builder;
730   struct lp_build_context blduivec;
731   struct lp_type fetch_type = vs_type;
732   LLVMValueRef offset, valid_mask;
733   unsigned i;
734
735   lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
736
737   vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
738   buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
739
740   /* This mul can overflow. Wraparound is ok. */
741   offset = lp_build_mul(&blduivec, vb_stride, indices);
742
743   valid_mask = lp_build_compare(gallivm, blduivec.type,
744                                 PIPE_FUNC_LESS, offset, buffer_size_adj);
745
746   /* not valid elements use offset 0 */
747   offset = LLVMBuildAnd(builder, offset, valid_mask, "");
748
749   if (0) {
750      lp_build_print_value(gallivm, "   indices = ", indices);
751      lp_build_print_value(gallivm, "   offsets = ", offset);
752      lp_build_print_value(gallivm, "   valid_mask = ", valid_mask);
753   }
754
755   /*
756    * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
757    * This should always produce better code.
758    */
759
760   /* The type handling is annoying here... */
761   if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
762       format_desc->channel[0].pure_integer) {
763      if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
764         fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
765      }
766      else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
767         fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
768      }
769   }
770
771   lp_build_fetch_rgba_soa(gallivm, format_desc,
772                           fetch_type, FALSE, map_ptr, offset,
773                           blduivec.zero, blduivec.zero,
774                           NULL, inputs);
775
776   for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
777      inputs[i] = LLVMBuildBitCast(builder, inputs[i],
778                                   lp_build_vec_type(gallivm, vs_type), "");
779   }
780
781   /* out-of-bound fetches return all zeros */
782   for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
783      inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
784      inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
785      inputs[i] = LLVMBuildBitCast(builder, inputs[i],
786                                   lp_build_vec_type(gallivm, vs_type), "");
787   }
788}
789
790
791static void
792store_aos(struct gallivm_state *gallivm,
793          LLVMValueRef io_ptr,
794          LLVMValueRef index,
795          LLVMValueRef value)
796{
797   LLVMTypeRef data_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, lp_float32_vec4_type()), 0);
798   LLVMBuilderRef builder = gallivm->builder;
799   LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr);
800   LLVMValueRef indices[3];
801
802   indices[0] = lp_build_const_int32(gallivm, 0);
803   indices[1] = index;
804   indices[2] = lp_build_const_int32(gallivm, 0);
805
806   data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
807   data_ptr = LLVMBuildPointerCast(builder, data_ptr, data_ptr_type, "");
808
809#if DEBUG_STORE
810   lp_build_printf(gallivm, "    ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
811#endif
812
813   /* Unaligned store due to the vertex header */
814   LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
815}
816
817/**
818 * Adjust the mask to architecture endianess. The mask will the store in struct:
819 *
820 * struct vertex_header {
821 *    unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
822 *    unsigned edgeflag:1;
823 *    unsigned pad:1;
824 *    unsigned vertex_id:16;
825 *    [...]
826 * }
827 *
828 * On little-endian machine nothing needs to done, however on bit-endian machine
829 * the mask's fields need to be adjusted with the algorithm:
830 *
831 * uint32_t reverse (uint32_t x)
832 * {
833 *   return (x >> 16) |              // vertex_id
834 *          ((x & 0x3fff) << 18) |   // clipmask
835 *          ((x & 0x4000) << 3) |    // pad
836 *          ((x & 0x8000) << 1);     // edgeflag
837 * }
838 */
839static LLVMValueRef
840adjust_mask(struct gallivm_state *gallivm,
841            LLVMValueRef mask)
842{
843#ifdef PIPE_ARCH_BIG_ENDIAN
844   LLVMBuilderRef builder = gallivm->builder;
845   LLVMValueRef vertex_id;
846   LLVMValueRef clipmask;
847   LLVMValueRef pad;
848   LLVMValueRef edgeflag;
849
850   vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
851   clipmask  = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
852   clipmask  = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
853   if (0) {
854      pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
855      pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
856   }
857   edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
858   edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
859
860   mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
861   if (0) {
862      mask = LLVMBuildOr(builder, mask, pad, "");
863   }
864   mask = LLVMBuildOr(builder, mask, edgeflag, "");
865#endif
866   return mask;
867}
868
869static void
870store_aos_array(struct gallivm_state *gallivm,
871                struct lp_type soa_type,
872                LLVMValueRef io_ptr,
873                LLVMValueRef *indices,
874                LLVMValueRef* aos,
875                int attrib,
876                int num_outputs,
877                LLVMValueRef clipmask,
878                boolean need_edgeflag)
879{
880   LLVMBuilderRef builder = gallivm->builder;
881   LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
882   LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
883   LLVMValueRef linear_inds[LP_MAX_VECTOR_WIDTH / 32];
884   LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
885   int vector_length = soa_type.length;
886   int i;
887
888   debug_assert(TGSI_NUM_CHANNELS == 4);
889
890   for (i = 0; i < vector_length; i++) {
891      linear_inds[i] = lp_build_const_int32(gallivm, i);
892      if (indices) {
893         inds[i] = indices[i];
894      } else {
895         inds[i] = linear_inds[i];
896      }
897      io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
898   }
899
900   if (attrib == 0) {
901      /* store vertex header for each of the n vertices */
902      LLVMValueRef val, cliptmp;
903      int vertex_id_pad_edgeflag;
904
905      /* If this assertion fails, it means we need to update the bit twidding
906       * code here.  See struct vertex_header in draw_private.h.
907       */
908      assert(DRAW_TOTAL_CLIP_PLANES==14);
909      /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
910      if (!need_edgeflag) {
911         vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
912      }
913      else {
914         vertex_id_pad_edgeflag = (0xffff << 16);
915      }
916      val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
917                                   vertex_id_pad_edgeflag);
918      /* OR with the clipmask */
919      cliptmp = LLVMBuildOr(builder, val, clipmask, "");
920      for (i = 0; i < vector_length; i++) {
921         LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptrs[i]);
922         val = LLVMBuildExtractElement(builder, cliptmp, linear_inds[i], "");
923         val = adjust_mask(gallivm, val);
924#if DEBUG_STORE
925         lp_build_printf(gallivm, "io = %p, index %d, clipmask = %x\n",
926                         io_ptrs[i], inds[i], val);
927#endif
928         LLVMBuildStore(builder, val, id_ptr);
929      }
930   }
931
932   /* store for each of the n vertices */
933   for (i = 0; i < vector_length; i++) {
934      store_aos(gallivm, io_ptrs[i], attr_index, aos[i]);
935   }
936}
937
938
939static void
940convert_to_aos(struct gallivm_state *gallivm,
941               LLVMValueRef io,
942               LLVMValueRef *indices,
943               LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
944               LLVMValueRef clipmask,
945               int num_outputs,
946               struct lp_type soa_type,
947               boolean need_edgeflag)
948{
949   LLVMBuilderRef builder = gallivm->builder;
950   unsigned chan, attrib, i;
951
952#if DEBUG_STORE
953   lp_build_printf(gallivm, "   # storing begin\n");
954#endif
955   for (attrib = 0; attrib < num_outputs; ++attrib) {
956      LLVMValueRef soa[TGSI_NUM_CHANNELS];
957      LLVMValueRef aos[LP_MAX_VECTOR_WIDTH / 32];
958      for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
959         if (outputs[attrib][chan]) {
960            LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
961            lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
962#if DEBUG_STORE
963            lp_build_printf(gallivm, "output %d : %d ",
964                            LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
965                                         attrib, 0),
966                            LLVMConstInt(LLVMInt32TypeInContext(gallivm->context),
967                                         chan, 0));
968            lp_build_print_value(gallivm, "val = ", out);
969            {
970               LLVMValueRef iv =
971                  LLVMBuildBitCast(builder, out, lp_build_int_vec_type(gallivm, soa_type), "");
972
973               lp_build_print_value(gallivm, "  ival = ", iv);
974            }
975#endif
976            soa[chan] = out;
977         }
978         else {
979            soa[chan] = 0;
980         }
981      }
982
983
984      if (soa_type.length == TGSI_NUM_CHANNELS) {
985         lp_build_transpose_aos(gallivm, soa_type, soa, aos);
986      } else {
987         lp_build_transpose_aos(gallivm, soa_type, soa, soa);
988
989         for (i = 0; i < soa_type.length; ++i) {
990            aos[i] = lp_build_extract_range(gallivm,
991                                            soa[i % TGSI_NUM_CHANNELS],
992                                            (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
993                                            TGSI_NUM_CHANNELS);
994         }
995      }
996
997      store_aos_array(gallivm,
998                      soa_type,
999                      io, indices,
1000                      aos,
1001                      attrib,
1002                      num_outputs,
1003                      clipmask,
1004                      need_edgeflag);
1005   }
1006#if DEBUG_STORE
1007   lp_build_printf(gallivm, "   # storing end\n");
1008#endif
1009}
1010
1011
1012/**
1013 * Stores original vertex positions in clip coordinates
1014 */
1015static void
1016store_clip(struct gallivm_state *gallivm,
1017           const struct lp_type vs_type,
1018           LLVMValueRef io_ptr,
1019           LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1020           int idx)
1021{
1022   LLVMBuilderRef builder = gallivm->builder;
1023   LLVMValueRef soa[4];
1024   LLVMValueRef aos[LP_MAX_VECTOR_LENGTH];
1025   LLVMValueRef indices[2];
1026   LLVMValueRef io_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1027   LLVMValueRef inds[LP_MAX_VECTOR_WIDTH / 32];
1028   LLVMValueRef clip_ptrs[LP_MAX_VECTOR_WIDTH / 32];
1029   LLVMTypeRef clip_ptr_type =
1030      LLVMPointerType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context),
1031                                     4), 0);
1032   int i, j;
1033
1034   indices[0] =
1035   indices[1] = lp_build_const_int32(gallivm, 0);
1036
1037   for (i = 0; i < vs_type.length; i++) {
1038      inds[i] = lp_build_const_int32(gallivm, i);
1039      io_ptrs[i] = LLVMBuildGEP(builder, io_ptr, &inds[i], 1, "");
1040   }
1041
1042   soa[0] = LLVMBuildLoad(builder, outputs[idx][0], ""); /*x0 x1 .. xn*/
1043   soa[1] = LLVMBuildLoad(builder, outputs[idx][1], ""); /*y0 y1 .. yn*/
1044   soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
1045   soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
1046
1047   for (i = 0; i < vs_type.length; i++) {
1048      clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
1049   }
1050
1051   lp_build_transpose_aos(gallivm, vs_type, soa, soa);
1052   for (i = 0; i < vs_type.length; ++i) {
1053      aos[i] = lp_build_extract_range(gallivm,
1054                                      soa[i % TGSI_NUM_CHANNELS],
1055                                      (i / TGSI_NUM_CHANNELS) * TGSI_NUM_CHANNELS,
1056                                      TGSI_NUM_CHANNELS);
1057   }
1058
1059   for (j = 0; j < vs_type.length; j++) {
1060      LLVMValueRef clip_ptr;
1061
1062      clip_ptr = LLVMBuildGEP(builder, clip_ptrs[j], indices, 2, "clipo");
1063      clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
1064
1065      /* Unaligned store */
1066      LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
1067   }
1068}
1069
1070
1071/**
1072 * Transforms the outputs for viewport mapping
1073 */
1074static void
1075generate_viewport(struct draw_llvm_variant *variant,
1076                  LLVMBuilderRef builder,
1077                  struct lp_type vs_type,
1078                  LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1079                  LLVMValueRef context_ptr)
1080{
1081   int i;
1082   struct gallivm_state *gallivm = variant->gallivm;
1083   struct lp_type f32_type = vs_type;
1084   const unsigned pos = variant->llvm->draw->vs.position_output;
1085   LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1086   LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
1087   LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0);       /*1.0 1.0 1.0 1.0*/
1088   LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
1089
1090   /* We treat pipe_viewport_state as a float array */
1091   const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
1092   const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
1093
1094   /* for 1/w convention*/
1095   out3 = LLVMBuildFDiv(builder, const1, out3, "");
1096   LLVMBuildStore(builder, out3, outputs[pos][3]);
1097
1098   /* Viewport Mapping */
1099   for (i=0; i<3; i++) {
1100      LLVMValueRef out = LLVMBuildLoad(builder, outputs[pos][i], ""); /*x0 x1 .. xn*/
1101      LLVMValueRef scale;
1102      LLVMValueRef trans;
1103      LLVMValueRef scale_i;
1104      LLVMValueRef trans_i;
1105      LLVMValueRef index;
1106
1107      index = lp_build_const_int32(gallivm, i + scale_index_offset);
1108      scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1109
1110      index = lp_build_const_int32(gallivm, i + trans_index_offset);
1111      trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
1112
1113      scale = lp_build_broadcast(gallivm, vs_type_llvm,
1114                                 LLVMBuildLoad(builder, scale_i, "scale"));
1115      trans = lp_build_broadcast(gallivm, vs_type_llvm,
1116                                 LLVMBuildLoad(builder, trans_i, "trans"));
1117
1118      /* divide by w */
1119      out = LLVMBuildFMul(builder, out, out3, "");
1120      /* mult by scale, add translation */
1121      out = lp_build_fmuladd(builder, out, scale, trans);
1122
1123      /* store transformed outputs */
1124      LLVMBuildStore(builder, out, outputs[pos][i]);
1125   }
1126
1127}
1128
1129
1130/**
1131 * Returns clipmask as nxi32 bitmask for the n vertices
1132 */
1133static LLVMValueRef
1134generate_clipmask(struct draw_llvm *llvm,
1135                  struct gallivm_state *gallivm,
1136                  struct lp_type vs_type,
1137                  LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
1138                  struct draw_llvm_variant_key *key,
1139                  LLVMValueRef context_ptr,
1140                  boolean *have_clipdist)
1141{
1142   LLVMBuilderRef builder = gallivm->builder;
1143   LLVMValueRef mask; /* stores the <nxi32> clipmasks */
1144   LLVMValueRef test, temp;
1145   LLVMValueRef zero, shift;
1146   LLVMValueRef pos_x, pos_y, pos_z, pos_w;
1147   LLVMValueRef cv_x, cv_y, cv_z, cv_w;
1148   LLVMValueRef plane1, planes, plane_ptr, sum;
1149   struct lp_type f32_type = vs_type;
1150   struct lp_type i32_type = lp_int_type(vs_type);
1151   const unsigned pos = llvm->draw->vs.position_output;
1152   const unsigned cv = llvm->draw->vs.clipvertex_output;
1153   int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
1154   boolean have_cd = false;
1155   boolean clip_user = key->clip_user;
1156   unsigned ucp_enable = key->ucp_enable;
1157   unsigned cd[2];
1158
1159   cd[0] = llvm->draw->vs.ccdistance_output[0];
1160   cd[1] = llvm->draw->vs.ccdistance_output[1];
1161
1162   if (cd[0] != pos || cd[1] != pos)
1163      have_cd = true;
1164
1165   if (num_written_clipdistance && !clip_user) {
1166      clip_user = true;
1167      ucp_enable = (1 << num_written_clipdistance) - 1;
1168   }
1169
1170   mask = lp_build_const_int_vec(gallivm, i32_type, 0);
1171   temp = lp_build_const_int_vec(gallivm, i32_type, 0);
1172   zero = lp_build_const_vec(gallivm, f32_type, 0);         /* 0.0f 0.0f 0.0f 0.0f */
1173   shift = lp_build_const_int_vec(gallivm, i32_type, 1);    /* 1 1 1 1 */
1174
1175   /*
1176    * load clipvertex and position from correct locations.
1177    * if they are the same just load them once.
1178    */
1179   pos_x = LLVMBuildLoad(builder, outputs[pos][0], ""); /*x0 x1 .. xn */
1180   pos_y = LLVMBuildLoad(builder, outputs[pos][1], ""); /*y0 y1 .. yn */
1181   pos_z = LLVMBuildLoad(builder, outputs[pos][2], ""); /*z0 z1 .. zn */
1182   pos_w = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn */
1183
1184   if (clip_user && cv != pos) {
1185      cv_x = LLVMBuildLoad(builder, outputs[cv][0], ""); /*x0 x1 .. xn */
1186      cv_y = LLVMBuildLoad(builder, outputs[cv][1], ""); /*y0 y1 .. yn */
1187      cv_z = LLVMBuildLoad(builder, outputs[cv][2], ""); /*z0 z1 .. zn */
1188      cv_w = LLVMBuildLoad(builder, outputs[cv][3], ""); /*w0 w1 .. wn */
1189   } else {
1190      cv_x = pos_x;
1191      cv_y = pos_y;
1192      cv_z = pos_z;
1193      cv_w = pos_w;
1194   }
1195
1196   /*
1197    * Be careful with the comparisons and NaNs (using llvm's unordered
1198    * comparisons here).
1199    */
1200   /* Cliptest, for hardwired planes */
1201   /*
1202    * XXX should take guardband into account (currently not in key).
1203    * Otherwise might run the draw pipeline stages for nothing.
1204    */
1205   if (key->clip_xy) {
1206      /* plane 1 */
1207      test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
1208      temp = shift;
1209      test = LLVMBuildAnd(builder, test, temp, "");
1210      mask = test;
1211
1212      /* plane 2 */
1213      test = LLVMBuildFAdd(builder, pos_x, pos_w, "");
1214      test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1215      temp = LLVMBuildShl(builder, temp, shift, "");
1216      test = LLVMBuildAnd(builder, test, temp, "");
1217      mask = LLVMBuildOr(builder, mask, test, "");
1218
1219      /* plane 3 */
1220      test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w);
1221      temp = LLVMBuildShl(builder, temp, shift, "");
1222      test = LLVMBuildAnd(builder, test, temp, "");
1223      mask = LLVMBuildOr(builder, mask, test, "");
1224
1225      /* plane 4 */
1226      test = LLVMBuildFAdd(builder, pos_y, pos_w, "");
1227      test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1228      temp = LLVMBuildShl(builder, temp, shift, "");
1229      test = LLVMBuildAnd(builder, test, temp, "");
1230      mask = LLVMBuildOr(builder, mask, test, "");
1231   }
1232
1233   if (key->clip_z) {
1234      temp = lp_build_const_int_vec(gallivm, i32_type, 16);
1235      if (key->clip_halfz) {
1236         /* plane 5 */
1237         test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
1238         test = LLVMBuildAnd(builder, test, temp, "");
1239         mask = LLVMBuildOr(builder, mask, test, "");
1240      }
1241      else {
1242         /* plane 5 */
1243         test = LLVMBuildFAdd(builder, pos_z, pos_w, "");
1244         test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test);
1245         test = LLVMBuildAnd(builder, test, temp, "");
1246         mask = LLVMBuildOr(builder, mask, test, "");
1247      }
1248      /* plane 6 */
1249      test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w);
1250      temp = LLVMBuildShl(builder, temp, shift, "");
1251      test = LLVMBuildAnd(builder, test, temp, "");
1252      mask = LLVMBuildOr(builder, mask, test, "");
1253   }
1254
1255   if (clip_user) {
1256      LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr);
1257      LLVMValueRef indices[3];
1258      LLVMValueRef is_nan_or_inf;
1259
1260      /* userclip planes */
1261      while (ucp_enable) {
1262         unsigned plane_idx = ffs(ucp_enable)-1;
1263         ucp_enable &= ~(1 << plane_idx);
1264         plane_idx += 6;
1265
1266         if (have_cd && num_written_clipdistance) {
1267            LLVMValueRef clipdist;
1268            int i;
1269            i = plane_idx - 6;
1270
1271            *have_clipdist = TRUE;
1272            if (i < 4) {
1273               clipdist = LLVMBuildLoad(builder, outputs[cd[0]][i], "");
1274            } else {
1275               clipdist = LLVMBuildLoad(builder, outputs[cd[1]][i-4], "");
1276            }
1277            test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
1278            is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
1279            test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
1280            temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1281            test = LLVMBuildAnd(builder, test, temp, "");
1282            mask = LLVMBuildOr(builder, mask, test, "");
1283         } else {
1284            LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
1285            indices[0] = lp_build_const_int32(gallivm, 0);
1286            indices[1] = lp_build_const_int32(gallivm, plane_idx);
1287
1288            indices[2] = lp_build_const_int32(gallivm, 0);
1289            plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1290            plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x");
1291            planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1292            sum = LLVMBuildFMul(builder, planes, cv_x, "");
1293
1294            indices[2] = lp_build_const_int32(gallivm, 1);
1295            plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1296            plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
1297            planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1298            sum = lp_build_fmuladd(builder, planes, cv_y, sum);
1299
1300            indices[2] = lp_build_const_int32(gallivm, 2);
1301            plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1302            plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
1303            planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1304            sum = lp_build_fmuladd(builder, planes, cv_z, sum);
1305
1306            indices[2] = lp_build_const_int32(gallivm, 3);
1307            plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
1308            plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
1309            planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
1310            sum = lp_build_fmuladd(builder, planes, cv_w, sum);
1311
1312            test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
1313            temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
1314            test = LLVMBuildAnd(builder, test, temp, "");
1315            mask = LLVMBuildOr(builder, mask, test, "");
1316         }
1317      }
1318   }
1319   if (key->need_edgeflags) {
1320      /*
1321       * This isn't really part of clipmask but stored the same in vertex
1322       * header later, so do it here.
1323       */
1324      unsigned edge_attr = llvm->draw->vs.edgeflag_output;
1325      LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
1326      LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
1327      test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
1328      temp = lp_build_const_int_vec(gallivm, i32_type,
1329                                    1LL << DRAW_TOTAL_CLIP_PLANES);
1330      test = LLVMBuildAnd(builder, test, temp, "");
1331      mask = LLVMBuildOr(builder, mask, test, "");
1332   }
1333   return mask;
1334}
1335
1336
1337/**
1338 * Returns boolean if any clipping has occurred
1339 * Used zero/one i8 value to represent boolean
1340 */
1341static LLVMValueRef
1342clipmask_booli8(struct gallivm_state *gallivm,
1343                const struct lp_type vs_type,
1344                LLVMValueRef clipmask_bool_ptr,
1345                boolean edgeflag_in_clipmask)
1346{
1347   LLVMBuilderRef builder = gallivm->builder;
1348   LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
1349   LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1350   LLVMValueRef ret;
1351   struct lp_build_context bldivec;
1352
1353   lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
1354
1355   /*
1356    * We need to invert the edgeflag bit from the clipmask here
1357    * (because the result is really if we want to run the pipeline or not
1358    * and we (may) need it if edgeflag was 0).
1359    */
1360   if (edgeflag_in_clipmask) {
1361      LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1362                                                 1LL << DRAW_TOTAL_CLIP_PLANES);
1363      clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
1364   }
1365
1366   /*
1367    * XXX: probably should mask off bits from the mask which come from
1368    * vertices which were beyond the count (i.e. indices_valid for
1369    * linear fetches, for elts ones we don't have the correct mask
1370    * right now). Otherwise might run the pipeline for nothing,
1371    * though everything should still work.
1372    */
1373   ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
1374   ret = LLVMBuildZExt(builder, ret, int8_type, "");
1375   return ret;
1376}
1377
1378static LLVMValueRef
1379draw_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
1380                         struct lp_build_tgsi_context * bld_base,
1381                         boolean is_vindex_indirect,
1382                         LLVMValueRef vertex_index,
1383                         boolean is_aindex_indirect,
1384                         LLVMValueRef attrib_index,
1385                         LLVMValueRef swizzle_index)
1386{
1387   const struct draw_gs_llvm_iface *gs = draw_gs_llvm_iface(gs_iface);
1388   struct gallivm_state *gallivm = bld_base->base.gallivm;
1389   LLVMBuilderRef builder = gallivm->builder;
1390   LLVMValueRef indices[3];
1391   LLVMValueRef res;
1392   struct lp_type type = bld_base->base.type;
1393
1394   if (is_vindex_indirect || is_aindex_indirect) {
1395      int i;
1396      res = bld_base->base.zero;
1397      for (i = 0; i < type.length; ++i) {
1398         LLVMValueRef idx = lp_build_const_int32(gallivm, i);
1399         LLVMValueRef vert_chan_index = vertex_index;
1400         LLVMValueRef attr_chan_index = attrib_index;
1401         LLVMValueRef channel_vec, value;
1402
1403         if (is_vindex_indirect) {
1404            vert_chan_index = LLVMBuildExtractElement(builder,
1405                                                      vertex_index, idx, "");
1406         }
1407         if (is_aindex_indirect) {
1408            attr_chan_index = LLVMBuildExtractElement(builder,
1409                                                      attrib_index, idx, "");
1410         }
1411
1412         indices[0] = vert_chan_index;
1413         indices[1] = attr_chan_index;
1414         indices[2] = swizzle_index;
1415
1416         channel_vec = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1417         channel_vec = LLVMBuildLoad(builder, channel_vec, "");
1418         value = LLVMBuildExtractElement(builder, channel_vec, idx, "");
1419
1420         res = LLVMBuildInsertElement(builder, res, value, idx, "");
1421      }
1422   } else {
1423      indices[0] = vertex_index;
1424      indices[1] = attrib_index;
1425      indices[2] = swizzle_index;
1426
1427      res = LLVMBuildGEP(builder, gs->input, indices, 3, "");
1428      res = LLVMBuildLoad(builder, res, "");
1429   }
1430
1431   return res;
1432}
1433
1434static void
1435draw_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
1436                         struct lp_build_tgsi_context * bld_base,
1437                         LLVMValueRef (*outputs)[4],
1438                         LLVMValueRef emitted_vertices_vec)
1439{
1440   const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1441   struct draw_gs_llvm_variant *variant = gs_iface->variant;
1442   struct gallivm_state *gallivm = variant->gallivm;
1443   LLVMBuilderRef builder = gallivm->builder;
1444   struct lp_type gs_type = bld_base->base.type;
1445   LLVMValueRef clipmask = lp_build_const_int_vec(gallivm,
1446                                                  lp_int_type(gs_type), 0);
1447   LLVMValueRef indices[LP_MAX_VECTOR_LENGTH];
1448   LLVMValueRef next_prim_offset =
1449      lp_build_const_int32(gallivm, variant->shader->base.primitive_boundary);
1450   LLVMValueRef io = variant->io_ptr;
1451   unsigned i;
1452   const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
1453
1454   for (i = 0; i < gs_type.length; ++i) {
1455      LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1456      LLVMValueRef currently_emitted =
1457         LLVMBuildExtractElement(builder, emitted_vertices_vec, ind, "");
1458      indices[i] = LLVMBuildMul(builder, ind, next_prim_offset, "");
1459      indices[i] = LLVMBuildAdd(builder, indices[i], currently_emitted, "");
1460   }
1461
1462   convert_to_aos(gallivm, io, indices,
1463                  outputs, clipmask,
1464                  gs_info->num_outputs, gs_type,
1465                  FALSE);
1466}
1467
1468static void
1469draw_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
1470                           struct lp_build_tgsi_context * bld_base,
1471                           LLVMValueRef verts_per_prim_vec,
1472                           LLVMValueRef emitted_prims_vec)
1473{
1474   const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1475   struct draw_gs_llvm_variant *variant = gs_iface->variant;
1476   struct gallivm_state *gallivm = variant->gallivm;
1477   LLVMBuilderRef builder = gallivm->builder;
1478   LLVMValueRef prim_lengts_ptr =
1479      draw_gs_jit_prim_lengths(variant->gallivm, variant->context_ptr);
1480   unsigned i;
1481
1482   for (i = 0; i < bld_base->base.type.length; ++i) {
1483      LLVMValueRef ind = lp_build_const_int32(gallivm, i);
1484      LLVMValueRef prims_emitted =
1485         LLVMBuildExtractElement(builder, emitted_prims_vec, ind, "");
1486      LLVMValueRef store_ptr;
1487      LLVMValueRef num_vertices =
1488         LLVMBuildExtractElement(builder, verts_per_prim_vec, ind, "");
1489
1490      store_ptr = LLVMBuildGEP(builder, prim_lengts_ptr, &prims_emitted, 1, "");
1491      store_ptr = LLVMBuildLoad(builder, store_ptr, "");
1492      store_ptr = LLVMBuildGEP(builder, store_ptr, &ind, 1, "");
1493      LLVMBuildStore(builder, num_vertices, store_ptr);
1494   }
1495}
1496
1497static void
1498draw_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
1499                      struct lp_build_tgsi_context * bld_base,
1500                      LLVMValueRef total_emitted_vertices_vec,
1501                      LLVMValueRef emitted_prims_vec)
1502{
1503   const struct draw_gs_llvm_iface *gs_iface = draw_gs_llvm_iface(gs_base);
1504   struct draw_gs_llvm_variant *variant = gs_iface->variant;
1505   struct gallivm_state *gallivm = variant->gallivm;
1506   LLVMBuilderRef builder = gallivm->builder;
1507   LLVMValueRef emitted_verts_ptr =
1508      draw_gs_jit_emitted_vertices(gallivm, variant->context_ptr);
1509   LLVMValueRef emitted_prims_ptr =
1510      draw_gs_jit_emitted_prims(gallivm, variant->context_ptr);
1511   LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
1512
1513   emitted_verts_ptr = LLVMBuildGEP(builder, emitted_verts_ptr, &zero, 0, "");
1514   emitted_prims_ptr = LLVMBuildGEP(builder, emitted_prims_ptr, &zero, 0, "");
1515
1516   LLVMBuildStore(builder, total_emitted_vertices_vec, emitted_verts_ptr);
1517   LLVMBuildStore(builder, emitted_prims_vec, emitted_prims_ptr);
1518}
1519
1520static void
1521draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
1522{
1523   struct gallivm_state *gallivm = variant->gallivm;
1524   LLVMContextRef context = gallivm->context;
1525   LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
1526   LLVMTypeRef arg_types[11];
1527   unsigned num_arg_types = ARRAY_SIZE(arg_types);
1528   LLVMTypeRef func_type;
1529   LLVMValueRef context_ptr;
1530   LLVMBasicBlockRef block;
1531   LLVMBuilderRef builder;
1532   char func_name[64];
1533   struct lp_type vs_type;
1534   LLVMValueRef count, fetch_elts, start_or_maxelt;
1535   LLVMValueRef vertex_id_offset, start_instance;
1536   LLVMValueRef stride, step, io_itr;
1537   LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
1538   LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
1539   LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
1540   LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
1541   LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
1542   LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
1543   LLVMValueRef fake_buf_ptr, fake_buf;
1544
1545   struct draw_context *draw = llvm->draw;
1546   const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
1547   unsigned i, j;
1548   struct lp_build_context bld, blduivec;
1549   struct lp_build_loop_state lp_loop;
1550   struct lp_build_if_state if_ctx;
1551   const int vector_length = lp_native_vector_width / 32;
1552   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1553   struct lp_build_sampler_soa *sampler = 0;
1554   LLVMValueRef ret, clipmask_bool_ptr;
1555   struct draw_llvm_variant_key *key = &variant->key;
1556   /* If geometry shader is present we need to skip both the viewport
1557    * transformation and clipping otherwise the inputs to the geometry
1558    * shader will be incorrect.
1559    * The code can't handle vp transform when vs writes vp index neither
1560    * (though this would be fixable here, but couldn't just broadcast
1561    * the values).
1562    */
1563   const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
1564                                   vs_info->writes_viewport_index;
1565   const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
1566                                                    key->clip_z ||
1567                                                    key->clip_user ||
1568                                                    key->need_edgeflags);
1569   LLVMValueRef variant_func;
1570   const unsigned pos = draw->vs.position_output;
1571   const unsigned cv = draw->vs.clipvertex_output;
1572   boolean have_clipdist = FALSE;
1573   struct lp_bld_tgsi_system_values system_values;
1574
1575   memset(&system_values, 0, sizeof(system_values));
1576
1577   util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u",
1578                 variant->shader->variants_cached);
1579
1580   i = 0;
1581   arg_types[i++] = get_context_ptr_type(variant);       /* context */
1582   arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
1583   arg_types[i++] = get_buffer_ptr_type(variant);        /* vbuffers */
1584   arg_types[i++] = int32_type;                          /* count */
1585   arg_types[i++] = int32_type;                          /* start/fetch_elt_max */
1586   arg_types[i++] = int32_type;                          /* stride */
1587   arg_types[i++] = get_vb_ptr_type(variant);            /* pipe_vertex_buffer's */
1588   arg_types[i++] = int32_type;                          /* instance_id */
1589   arg_types[i++] = int32_type;                          /* vertex_id_offset */
1590   arg_types[i++] = int32_type;                          /* start_instance */
1591   arg_types[i++] = LLVMPointerType(int32_type, 0);      /* fetch_elts  */
1592
1593   func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
1594                                arg_types, num_arg_types, 0);
1595
1596   variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
1597   variant->function = variant_func;
1598
1599   LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
1600   for (i = 0; i < num_arg_types; ++i)
1601      if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
1602         lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
1603
1604   context_ptr               = LLVMGetParam(variant_func, 0);
1605   io_ptr                    = LLVMGetParam(variant_func, 1);
1606   vbuffers_ptr              = LLVMGetParam(variant_func, 2);
1607   count                     = LLVMGetParam(variant_func, 3);
1608   /*
1609    * XXX: the maxelt part is unused. Not really useful, since we cannot
1610    * get index buffer overflows due to vsplit (which provides its own
1611    * elts buffer, with a different size than what's passed in here).
1612    */
1613   start_or_maxelt           = LLVMGetParam(variant_func, 4);
1614   /*
1615    * XXX: stride is actually unused. The stride we use is strictly calculated
1616    * from the number of outputs (including the draw_extra outputs).
1617    * Should probably fix some day (we need a new vs just because of extra
1618    * outputs which the generated vs won't touch).
1619    */
1620   stride                    = LLVMGetParam(variant_func, 5);
1621   vb_ptr                    = LLVMGetParam(variant_func, 6);
1622   system_values.instance_id = LLVMGetParam(variant_func, 7);
1623   vertex_id_offset          = LLVMGetParam(variant_func, 8);
1624   start_instance            = LLVMGetParam(variant_func, 9);
1625   fetch_elts                = LLVMGetParam(variant_func, 10);
1626
1627   lp_build_name(context_ptr, "context");
1628   lp_build_name(io_ptr, "io");
1629   lp_build_name(vbuffers_ptr, "vbuffers");
1630   lp_build_name(count, "count");
1631   lp_build_name(start_or_maxelt, "start_or_maxelt");
1632   lp_build_name(stride, "stride");
1633   lp_build_name(vb_ptr, "vb");
1634   lp_build_name(system_values.instance_id, "instance_id");
1635   lp_build_name(vertex_id_offset, "vertex_id_offset");
1636   lp_build_name(start_instance, "start_instance");
1637   lp_build_name(fetch_elts, "fetch_elts");
1638
1639   /*
1640    * Function body
1641    */
1642
1643   block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
1644   builder = gallivm->builder;
1645   LLVMPositionBuilderAtEnd(builder, block);
1646
1647   memset(&vs_type, 0, sizeof vs_type);
1648   vs_type.floating = TRUE; /* floating point values */
1649   vs_type.sign = TRUE;     /* values are signed */
1650   vs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
1651   vs_type.width = 32;      /* 32-bit float */
1652   vs_type.length = vector_length;
1653
1654   lp_build_context_init(&bld, gallivm, lp_type_uint(32));
1655   lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
1656
1657   /* hold temporary "bool" clipmask */
1658   clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
1659
1660   fake_buf = lp_build_alloca_undef(gallivm,
1661                 LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
1662   fake_buf = LLVMBuildBitCast(builder, fake_buf,
1663                 LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1664   fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
1665
1666   /* code generated texture sampling */
1667   sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
1668
1669   step = lp_build_const_int32(gallivm, vector_length);
1670
1671   ind_vec = blduivec.undef;
1672   for (i = 0; i < vs_type.length; i++) {
1673      LLVMValueRef index = lp_build_const_int32(gallivm, i);
1674      ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
1675   }
1676
1677   have_elts = LLVMBuildICmp(builder, LLVMIntNE,
1678                             LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
1679
1680   fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
1681   fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
1682   /*
1683    * Only needed for non-indexed path.
1684    */
1685   start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
1686
1687   /*
1688    * Pre-calculate everything which is constant per shader invocation.
1689    */
1690   for (j = 0; j < key->nr_vertex_elements; ++j) {
1691      LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
1692      LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
1693      struct pipe_vertex_element *velem = &key->vertex_element[j];
1694      LLVMValueRef vb_index =
1695         lp_build_const_int32(gallivm, velem->vertex_buffer_index);
1696      LLVMValueRef bsize = lp_build_const_int32(gallivm,
1697                                                util_format_get_blocksize(velem->src_format));
1698      LLVMValueRef src_offset = lp_build_const_int32(gallivm,
1699                                                     velem->src_offset);
1700      struct lp_build_if_state if_ctx;
1701
1702      if (velem->src_format != PIPE_FORMAT_NONE) {
1703         vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
1704         vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
1705         vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
1706         vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
1707                                      LLVMInt32TypeInContext(context), "");
1708         vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
1709         map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
1710         buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
1711
1712         ofbit = NULL;
1713         /*
1714          * We'll set buffer_size_adj to zero if we have of, so it will
1715          * always overflow later automatically without having to keep ofbit.
1716          * Overflows (with normal wraparound) doing the actual offset
1717          * calculation should be ok, just not for the buffer size calc.
1718          * It would also be possible to detect such overflows and return
1719          * zeros if that happens, but this would be more complex.
1720          */
1721         buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
1722         tmp = lp_build_sub(&bld, bsize, bld.one);
1723         buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
1724                                                     &ofbit);
1725         buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
1726                                                     buf_offset, &ofbit);
1727
1728         /*
1729          * We can't easily set fake vertex buffers outside the generated code.
1730          * Hence, set fake vertex buffers here instead basically, so fetch
1731          * code can always fetch using offset 0, eliminating all control flow
1732          * inside the main loop.
1733          * (Alternatively, could have control flow per vector skipping fetch
1734          * if ofbit is true.)
1735          */
1736         if (velem->instance_divisor) {
1737            /*
1738             * Index is equal to the start instance plus the number of current
1739             * instance divided by the divisor. In this case we compute it as:
1740             * index = start_instance + (instance_id  / divisor).
1741             * Note we could actually do the fetch here, outside the loop -
1742             * it's all constant, hopefully llvm recognizes this.
1743             */
1744            LLVMValueRef current_instance;
1745            current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
1746                                             lp_build_const_int32(gallivm,
1747                                                                  velem->instance_divisor),
1748                                             "instance_divisor");
1749            instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
1750                                                       current_instance, &ofbit);
1751         }
1752
1753         buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
1754                                              buffer_size_adj[j], "");
1755
1756         temp_ptr = lp_build_alloca_undef(gallivm,
1757                       LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
1758
1759         lp_build_if(&if_ctx, gallivm, ofbit);
1760         {
1761            LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
1762         }
1763         lp_build_else(&if_ctx);
1764         {
1765            map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
1766            LLVMBuildStore(builder, map_ptr[j], temp_ptr);
1767         }
1768         lp_build_endif(&if_ctx);
1769         map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
1770
1771         if (0) {
1772            lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
1773                            lp_build_const_int32(gallivm, j),
1774                            vb_index, vb_stride[j]);
1775            lp_build_printf(gallivm,
1776                            "   vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
1777                            vb_buffer_offset, src_offset, buf_offset);
1778            lp_build_printf(gallivm, "   buffer size = %u, blocksize = %u\n",
1779                            buffer_size, bsize);
1780            lp_build_printf(gallivm, "   instance_id = %u\n", system_values.instance_id);
1781         }
1782      }
1783   }
1784
1785   lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
1786   {
1787      LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1788      LLVMValueRef io;
1789      LLVMValueRef clipmask;   /* holds the clipmask value */
1790      LLVMValueRef true_index_array, index_store;
1791      const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
1792
1793      io_itr = lp_loop.counter;
1794
1795      io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
1796#if DEBUG_STORE
1797      lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
1798                      io_itr, io, lp_loop.counter);
1799#endif
1800
1801      true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
1802      true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
1803
1804      /*
1805       * Limit indices to fetch_max, otherwise might try to access indices
1806       * beyond index buffer (or rather vsplit elt buffer) size.
1807       * Could probably safely (?) skip this for non-indexed draws and
1808       * simplify things minimally (by removing it could combine the ind_vec
1809       * and start_vec adds). I think the only effect for non-indexed draws will
1810       * be that for the invalid elements they will be all fetched from the
1811       * same location as the last valid one, but noone should really care.
1812       */
1813      true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
1814
1815      index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
1816
1817      lp_build_if(&if_ctx, gallivm, have_elts);
1818      {
1819         /*
1820          * Note: you'd expect some comparison/clamp against fetch_elt_max
1821          * here.
1822          * There used to be one here but it was incorrect: overflow was
1823          * detected if index > fetch_elt_max - but the correct condition
1824          * would be index >= fetch_elt_max (since this is just size of elts
1825          * buffer / element size).
1826          * Using the correct condition however will cause failures - due to
1827          * vsplit/vcache code which rebases indices. So, as an example, if
1828          * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
1829          * replace all invalid indices with 0 - which in case of elt_bias
1830          * not being zero will get a different fetch index than the valid
1831          * index 0. So, just rely on vsplit code preventing out-of-bounds
1832          * fetches. This is also why it's safe to do elts fetch even if there
1833          * was no index buffer bound - the real buffer is never seen here, at
1834          * least not if there are index buffer overflows...
1835          */
1836
1837         /*
1838          * XXX should not have to do this, as scale can be handled
1839          * natively by loads (hits asserts though).
1840          */
1841         tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
1842         fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
1843                                       LLVMPointerType(LLVMInt8TypeInContext(context),
1844                                                       0), "");
1845         tmp = lp_build_gather(gallivm, vs_type.length,
1846                               32, bld.type, TRUE,
1847                               fetch_elts, tmp, FALSE);
1848         LLVMBuildStore(builder, tmp, index_store);
1849      }
1850      lp_build_else(&if_ctx);
1851      {
1852         tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
1853         LLVMBuildStore(builder, tmp, index_store);
1854      }
1855      lp_build_endif(&if_ctx);
1856
1857      true_index_array = LLVMBuildLoad(builder, index_store, "");
1858
1859      for (j = 0; j < key->nr_vertex_elements; ++j) {
1860         struct pipe_vertex_element *velem = &key->vertex_element[j];
1861         const struct util_format_description *format_desc =
1862            util_format_description(velem->src_format);
1863
1864         if (format_desc->format == PIPE_FORMAT_NONE) {
1865            for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
1866               inputs[j][i] = lp_build_zero(gallivm, vs_type);
1867            }
1868         }
1869         else if (velem->instance_divisor) {
1870            fetch_instanced(gallivm, format_desc, vs_type,
1871                            vb_stride[j], map_ptr[j],
1872                            buffer_size_adj[j],
1873                            inputs[j], instance_index[j]);
1874         }
1875         else {
1876            fetch_vector(gallivm, format_desc, vs_type,
1877                         vb_stride[j], map_ptr[j],
1878                         buffer_size_adj[j],
1879                         inputs[j], true_index_array);
1880         }
1881      }
1882
1883      /* In the paths with elts vertex id has to be unaffected by the
1884       * index bias and because indices inside our elements array have
1885       * already had index bias applied we need to subtract it here to
1886       * get back to the original index.
1887       * in the linear paths vertex id has to be unaffected by the
1888       * original start index and because we abuse the 'start' variable
1889       * to either represent the actual start index or the index at which
1890       * the primitive was split (we split rendering into chunks of at
1891       * most 4095-vertices) we need to back out the original start
1892       * index out of our vertex id here.
1893       */
1894      system_values.basevertex = lp_build_broadcast_scalar(&blduivec,
1895                                                           vertex_id_offset);
1896      system_values.vertex_id = true_index_array;
1897      system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
1898                                                      system_values.basevertex, "");
1899
1900      ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
1901      generate_vs(variant,
1902                  builder,
1903                  vs_type,
1904                  outputs,
1905                  ptr_aos,
1906                  &system_values,
1907                  context_ptr,
1908                  sampler,
1909                  key->clamp_vertex_color);
1910
1911      if (pos != -1 && cv != -1) {
1912         /* store original positions in clip before further manipulation */
1913         store_clip(gallivm, vs_type, io, outputs, pos);
1914
1915         /* do cliptest */
1916         if (enable_cliptest) {
1917            LLVMValueRef temp = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
1918            /* allocate clipmask, assign it integer type */
1919            clipmask = generate_clipmask(llvm,
1920                                         gallivm,
1921                                         vs_type,
1922                                         outputs,
1923                                         key,
1924                                         context_ptr, &have_clipdist);
1925            temp = LLVMBuildOr(builder, clipmask, temp, "");
1926            /* store temporary clipping boolean value */
1927            LLVMBuildStore(builder, temp, clipmask_bool_ptr);
1928         }
1929         else {
1930            clipmask = blduivec.zero;
1931         }
1932
1933         /* do viewport mapping */
1934         if (!bypass_viewport) {
1935            generate_viewport(variant, builder, vs_type, outputs, context_ptr);
1936         }
1937      }
1938      else {
1939         clipmask = blduivec.zero;
1940      }
1941
1942      /* store clipmask in vertex header,
1943       * original positions in clip
1944       * and transformed positions in data
1945       */
1946      convert_to_aos(gallivm, io, NULL, outputs, clipmask,
1947                     vs_info->num_outputs, vs_type,
1948                     enable_cliptest && key->need_edgeflags);
1949   }
1950   lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
1951
1952   sampler->destroy(sampler);
1953
1954   /* return clipping boolean value for function */
1955   ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
1956                         enable_cliptest && key->need_edgeflags);
1957
1958   LLVMBuildRet(builder, ret);
1959
1960   gallivm_verify_function(gallivm, variant_func);
1961}
1962
1963
1964struct draw_llvm_variant_key *
1965draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
1966{
1967   unsigned i;
1968   struct draw_llvm_variant_key *key;
1969   struct draw_sampler_static_state *draw_sampler;
1970
1971   key = (struct draw_llvm_variant_key *)store;
1972
1973   memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
1974
1975   key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
1976
1977   /* will have to rig this up properly later */
1978   key->clip_xy = llvm->draw->clip_xy;
1979   key->clip_z = llvm->draw->clip_z;
1980   key->clip_user = llvm->draw->clip_user;
1981   key->bypass_viewport = llvm->draw->bypass_viewport;
1982   key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
1983   /* XXX assumes edgeflag output not at 0 */
1984   key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
1985   key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
1986   key->has_gs = llvm->draw->gs.geometry_shader != NULL;
1987   key->num_outputs = draw_total_vs_outputs(llvm->draw);
1988
1989   /* All variants of this shader will have the same value for
1990    * nr_samplers.  Not yet trying to compact away holes in the
1991    * sampler array.
1992    */
1993   key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1994   if (llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
1995      key->nr_sampler_views =
1996         llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
1997   }
1998   else {
1999      key->nr_sampler_views = key->nr_samplers;
2000   }
2001
2002   /* Presumably all variants of the shader should have the same
2003    * number of vertex elements - ie the number of shader inputs.
2004    * NOTE: we NEED to store the needed number of needed inputs
2005    * here, not the number of provided elements to match keysize
2006    * (and the offset of sampler state in the key).
2007    * If we have excess number of vertex elements, this is valid,
2008    * but the excess ones don't matter.
2009    * If we don't have enough vertex elements (which looks not really
2010    * valid but we'll handle it gracefully) fill out missing ones with
2011    * zero (we'll recognize these later by PIPE_FORMAT_NONE).
2012    */
2013   key->nr_vertex_elements =
2014      llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
2015
2016   if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
2017      debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
2018                   key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
2019      memset(key->vertex_element, 0,
2020             sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
2021   }
2022   memcpy(key->vertex_element,
2023          llvm->draw->pt.vertex_element,
2024          sizeof(struct pipe_vertex_element) *
2025             MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
2026
2027   draw_sampler = draw_llvm_variant_key_samplers(key);
2028   memset(draw_sampler, 0,
2029          MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2030
2031   for (i = 0 ; i < key->nr_samplers; i++) {
2032      lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2033                                      llvm->draw->samplers[PIPE_SHADER_VERTEX][i]);
2034   }
2035   for (i = 0 ; i < key->nr_sampler_views; i++) {
2036      lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2037                                      llvm->draw->sampler_views[PIPE_SHADER_VERTEX][i]);
2038   }
2039
2040   return key;
2041}
2042
2043
2044void
2045draw_llvm_dump_variant_key(struct draw_llvm_variant_key *key)
2046{
2047   unsigned i;
2048   struct draw_sampler_static_state *sampler = draw_llvm_variant_key_samplers(key);
2049
2050   debug_printf("clamp_vertex_color = %u\n", key->clamp_vertex_color);
2051   debug_printf("clip_xy = %u\n", key->clip_xy);
2052   debug_printf("clip_z = %u\n", key->clip_z);
2053   debug_printf("clip_user = %u\n", key->clip_user);
2054   debug_printf("bypass_viewport = %u\n", key->bypass_viewport);
2055   debug_printf("clip_halfz = %u\n", key->clip_halfz);
2056   debug_printf("need_edgeflags = %u\n", key->need_edgeflags);
2057   debug_printf("has_gs = %u\n", key->has_gs);
2058   debug_printf("ucp_enable = %u\n", key->ucp_enable);
2059
2060   for (i = 0 ; i < key->nr_vertex_elements; i++) {
2061      debug_printf("vertex_element[%i].src_offset = %u\n", i, key->vertex_element[i].src_offset);
2062      debug_printf("vertex_element[%i].instance_divisor = %u\n", i, key->vertex_element[i].instance_divisor);
2063      debug_printf("vertex_element[%i].vertex_buffer_index = %u\n", i, key->vertex_element[i].vertex_buffer_index);
2064      debug_printf("vertex_element[%i].src_format = %s\n", i, util_format_name(key->vertex_element[i].src_format));
2065   }
2066
2067   for (i = 0 ; i < key->nr_sampler_views; i++) {
2068      debug_printf("sampler[%i].src_format = %s\n", i, util_format_name(sampler[i].texture_state.format));
2069   }
2070}
2071
2072
2073void
2074draw_llvm_set_mapped_texture(struct draw_context *draw,
2075                             enum pipe_shader_type shader_stage,
2076                             unsigned sview_idx,
2077                             uint32_t width, uint32_t height, uint32_t depth,
2078                             uint32_t first_level, uint32_t last_level,
2079                             const void *base_ptr,
2080                             uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
2081                             uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
2082                             uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS])
2083{
2084   unsigned j;
2085   struct draw_jit_texture *jit_tex;
2086
2087   assert(shader_stage == PIPE_SHADER_VERTEX ||
2088          shader_stage == PIPE_SHADER_GEOMETRY);
2089
2090   if (shader_stage == PIPE_SHADER_VERTEX) {
2091      assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
2092
2093      jit_tex = &draw->llvm->jit_context.textures[sview_idx];
2094   } else if (shader_stage == PIPE_SHADER_GEOMETRY) {
2095      assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
2096
2097      jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
2098   } else {
2099      assert(0);
2100      return;
2101   }
2102
2103   jit_tex->width = width;
2104   jit_tex->height = height;
2105   jit_tex->depth = depth;
2106   jit_tex->first_level = first_level;
2107   jit_tex->last_level = last_level;
2108   jit_tex->base = base_ptr;
2109
2110   for (j = first_level; j <= last_level; j++) {
2111      jit_tex->mip_offsets[j] = mip_offsets[j];
2112      jit_tex->row_stride[j] = row_stride[j];
2113      jit_tex->img_stride[j] = img_stride[j];
2114   }
2115}
2116
2117
2118void
2119draw_llvm_set_sampler_state(struct draw_context *draw,
2120                            enum pipe_shader_type shader_type)
2121{
2122   unsigned i;
2123
2124   if (shader_type == PIPE_SHADER_VERTEX) {
2125      for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
2126         struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
2127
2128         if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
2129            const struct pipe_sampler_state *s
2130               = draw->samplers[PIPE_SHADER_VERTEX][i];
2131            jit_sam->min_lod = s->min_lod;
2132            jit_sam->max_lod = s->max_lod;
2133            jit_sam->lod_bias = s->lod_bias;
2134            COPY_4V(jit_sam->border_color, s->border_color.f);
2135         }
2136      }
2137   } else if (shader_type == PIPE_SHADER_GEOMETRY) {
2138      for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
2139         struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
2140
2141         if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
2142            const struct pipe_sampler_state *s
2143               = draw->samplers[PIPE_SHADER_GEOMETRY][i];
2144            jit_sam->min_lod = s->min_lod;
2145            jit_sam->max_lod = s->max_lod;
2146            jit_sam->lod_bias = s->lod_bias;
2147            COPY_4V(jit_sam->border_color, s->border_color.f);
2148         }
2149      }
2150   }
2151}
2152
2153
2154void
2155draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
2156{
2157   struct draw_llvm *llvm = variant->llvm;
2158
2159   if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2160      debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
2161                    variant->shader->variants_cached, llvm->nr_variants);
2162   }
2163
2164   gallivm_destroy(variant->gallivm);
2165
2166   remove_from_list(&variant->list_item_local);
2167   variant->shader->variants_cached--;
2168   remove_from_list(&variant->list_item_global);
2169   llvm->nr_variants--;
2170   FREE(variant);
2171}
2172
2173
2174/**
2175 * Create LLVM types for various structures.
2176 */
2177static void
2178create_gs_jit_types(struct draw_gs_llvm_variant *var)
2179{
2180   struct gallivm_state *gallivm = var->gallivm;
2181   LLVMTypeRef texture_type, sampler_type, context_type;
2182
2183   texture_type = create_jit_texture_type(gallivm, "texture");
2184   sampler_type = create_jit_sampler_type(gallivm, "sampler");
2185
2186   context_type = create_gs_jit_context_type(gallivm,
2187                                             var->shader->base.vector_length,
2188                                             texture_type, sampler_type,
2189                                             "draw_gs_jit_context");
2190   var->context_ptr_type = LLVMPointerType(context_type, 0);
2191
2192   var->input_array_type = create_gs_jit_input_type(gallivm);
2193}
2194
2195static LLVMTypeRef
2196get_gs_context_ptr_type(struct draw_gs_llvm_variant *variant)
2197{
2198   if (!variant->context_ptr_type)
2199      create_gs_jit_types(variant);
2200   return variant->context_ptr_type;
2201}
2202
2203static LLVMValueRef
2204generate_mask_value(struct draw_gs_llvm_variant *variant,
2205                    struct lp_type gs_type)
2206{
2207   struct gallivm_state *gallivm = variant->gallivm;
2208   LLVMBuilderRef builder = gallivm->builder;
2209   struct lp_type mask_type = lp_int_type(gs_type);
2210   LLVMValueRef num_prims;
2211   LLVMValueRef mask_val = lp_build_const_vec(gallivm, mask_type, 0);
2212   unsigned i;
2213
2214   num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
2215                                  variant->num_prims);
2216   for (i = 0; i < gs_type.length; i++) {
2217      LLVMValueRef idx = lp_build_const_int32(gallivm, i);
2218      mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
2219   }
2220   mask_val = lp_build_compare(gallivm, mask_type,
2221                               PIPE_FUNC_GREATER, num_prims, mask_val);
2222
2223   return mask_val;
2224}
2225
2226static void
2227draw_gs_llvm_generate(struct draw_llvm *llvm,
2228                      struct draw_gs_llvm_variant *variant)
2229{
2230   struct gallivm_state *gallivm = variant->gallivm;
2231   LLVMContextRef context = gallivm->context;
2232   LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
2233   LLVMTypeRef arg_types[7];
2234   LLVMTypeRef func_type;
2235   LLVMValueRef variant_func;
2236   LLVMValueRef context_ptr;
2237   LLVMValueRef prim_id_ptr;
2238   LLVMBasicBlockRef block;
2239   LLVMBuilderRef builder;
2240   LLVMValueRef io_ptr, input_array, num_prims, mask_val;
2241   struct lp_build_sampler_soa *sampler = 0;
2242   struct lp_build_context bld;
2243   struct lp_bld_tgsi_system_values system_values;
2244   char func_name[64];
2245   struct lp_type gs_type;
2246   unsigned i;
2247   struct draw_gs_llvm_iface gs_iface;
2248   const struct tgsi_token *tokens = variant->shader->base.state.tokens;
2249   LLVMValueRef consts_ptr, num_consts_ptr;
2250   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
2251   struct lp_build_mask_context mask;
2252   const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
2253   unsigned vector_length = variant->shader->base.vector_length;
2254
2255   memset(&system_values, 0, sizeof(system_values));
2256
2257   util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
2258                 variant->shader->variants_cached);
2259
2260   assert(variant->vertex_header_ptr_type);
2261
2262   arg_types[0] = get_gs_context_ptr_type(variant);    /* context */
2263   arg_types[1] = variant->input_array_type;           /* input */
2264   arg_types[2] = variant->vertex_header_ptr_type;     /* vertex_header */
2265   arg_types[3] = int32_type;                          /* num_prims */
2266   arg_types[4] = int32_type;                          /* instance_id */
2267   arg_types[5] = LLVMPointerType(
2268      LLVMVectorType(int32_type, vector_length), 0);   /* prim_id_ptr */
2269   arg_types[6] = int32_type;
2270
2271   func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
2272
2273   variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
2274
2275   variant->function = variant_func;
2276
2277   LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
2278
2279   for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
2280      if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
2281         lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
2282
2283   context_ptr               = LLVMGetParam(variant_func, 0);
2284   input_array               = LLVMGetParam(variant_func, 1);
2285   io_ptr                    = LLVMGetParam(variant_func, 2);
2286   num_prims                 = LLVMGetParam(variant_func, 3);
2287   system_values.instance_id = LLVMGetParam(variant_func, 4);
2288   prim_id_ptr               = LLVMGetParam(variant_func, 5);
2289   system_values.invocation_id = LLVMGetParam(variant_func, 6);
2290
2291   lp_build_name(context_ptr, "context");
2292   lp_build_name(input_array, "input");
2293   lp_build_name(io_ptr, "io");
2294   lp_build_name(num_prims, "num_prims");
2295   lp_build_name(system_values.instance_id, "instance_id");
2296   lp_build_name(prim_id_ptr, "prim_id_ptr");
2297   lp_build_name(system_values.invocation_id, "invocation_id");
2298
2299   variant->context_ptr = context_ptr;
2300   variant->io_ptr = io_ptr;
2301   variant->num_prims = num_prims;
2302
2303   gs_iface.base.fetch_input = draw_gs_llvm_fetch_input;
2304   gs_iface.base.emit_vertex = draw_gs_llvm_emit_vertex;
2305   gs_iface.base.end_primitive = draw_gs_llvm_end_primitive;
2306   gs_iface.base.gs_epilogue = draw_gs_llvm_epilogue;
2307   gs_iface.input = input_array;
2308   gs_iface.variant = variant;
2309
2310   /*
2311    * Function body
2312    */
2313
2314   block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry");
2315   builder = gallivm->builder;
2316   LLVMPositionBuilderAtEnd(builder, block);
2317
2318   lp_build_context_init(&bld, gallivm, lp_type_int(32));
2319
2320   memset(&gs_type, 0, sizeof gs_type);
2321   gs_type.floating = TRUE; /* floating point values */
2322   gs_type.sign = TRUE;     /* values are signed */
2323   gs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
2324   gs_type.width = 32;      /* 32-bit float */
2325   gs_type.length = vector_length;
2326
2327   consts_ptr = draw_gs_jit_context_constants(variant->gallivm, context_ptr);
2328   num_consts_ptr =
2329      draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
2330
2331   /* code generated texture sampling */
2332   sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
2333
2334   mask_val = generate_mask_value(variant, gs_type);
2335   lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
2336
2337   if (gs_info->uses_primid) {
2338      system_values.prim_id = LLVMBuildLoad(builder, prim_id_ptr, "prim_id");
2339   }
2340
2341   if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2342      tgsi_dump(tokens, 0);
2343      draw_gs_llvm_dump_variant_key(&variant->key);
2344   }
2345
2346   lp_build_tgsi_soa(variant->gallivm,
2347                     tokens,
2348                     gs_type,
2349                     &mask,
2350                     consts_ptr,
2351                     num_consts_ptr,
2352                     &system_values,
2353                     NULL,
2354                     outputs,
2355                     context_ptr,
2356                     NULL,
2357                     sampler,
2358                     &llvm->draw->gs.geometry_shader->info,
2359                     (const struct lp_build_tgsi_gs_iface *)&gs_iface);
2360
2361   sampler->destroy(sampler);
2362
2363   lp_build_mask_end(&mask);
2364
2365   LLVMBuildRet(builder, lp_build_zero(gallivm, lp_type_uint(32)));
2366
2367   gallivm_verify_function(gallivm, variant_func);
2368}
2369
2370
2371struct draw_gs_llvm_variant *
2372draw_gs_llvm_create_variant(struct draw_llvm *llvm,
2373                            unsigned num_outputs,
2374                            const struct draw_gs_llvm_variant_key *key)
2375{
2376   struct draw_gs_llvm_variant *variant;
2377   struct llvm_geometry_shader *shader =
2378      llvm_geometry_shader(llvm->draw->gs.geometry_shader);
2379   LLVMTypeRef vertex_header;
2380   char module_name[64];
2381
2382   variant = MALLOC(sizeof *variant +
2383                    shader->variant_key_size -
2384                    sizeof variant->key);
2385   if (!variant)
2386      return NULL;
2387
2388   variant->llvm = llvm;
2389   variant->shader = shader;
2390
2391   util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
2392                 variant->shader->variants_cached);
2393
2394   variant->gallivm = gallivm_create(module_name, llvm->context);
2395
2396   create_gs_jit_types(variant);
2397
2398   memcpy(&variant->key, key, shader->variant_key_size);
2399
2400   vertex_header = create_jit_vertex_header(variant->gallivm, num_outputs);
2401
2402   variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
2403
2404   draw_gs_llvm_generate(llvm, variant);
2405
2406   gallivm_compile_module(variant->gallivm);
2407
2408   variant->jit_func = (draw_gs_jit_func)
2409         gallivm_jit_function(variant->gallivm, variant->function);
2410
2411   gallivm_free_ir(variant->gallivm);
2412
2413   variant->list_item_global.base = variant;
2414   variant->list_item_local.base = variant;
2415   /*variant->no = */shader->variants_created++;
2416   variant->list_item_global.base = variant;
2417
2418   return variant;
2419}
2420
2421void
2422draw_gs_llvm_destroy_variant(struct draw_gs_llvm_variant *variant)
2423{
2424   struct draw_llvm *llvm = variant->llvm;
2425
2426   if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
2427      debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
2428                    variant->shader->variants_cached, llvm->nr_gs_variants);
2429   }
2430
2431   gallivm_destroy(variant->gallivm);
2432
2433   remove_from_list(&variant->list_item_local);
2434   variant->shader->variants_cached--;
2435   remove_from_list(&variant->list_item_global);
2436   llvm->nr_gs_variants--;
2437   FREE(variant);
2438}
2439
2440struct draw_gs_llvm_variant_key *
2441draw_gs_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
2442{
2443   unsigned i;
2444   struct draw_gs_llvm_variant_key *key;
2445   struct draw_sampler_static_state *draw_sampler;
2446
2447   key = (struct draw_gs_llvm_variant_key *)store;
2448
2449   memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
2450
2451   key->num_outputs = draw_total_gs_outputs(llvm->draw);
2452
2453   /* All variants of this shader will have the same value for
2454    * nr_samplers.  Not yet trying to compact away holes in the
2455    * sampler array.
2456    */
2457   key->nr_samplers = llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
2458   if (llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
2459      key->nr_sampler_views =
2460         llvm->draw->gs.geometry_shader->info.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
2461   }
2462   else {
2463      key->nr_sampler_views = key->nr_samplers;
2464   }
2465
2466   draw_sampler = key->samplers;
2467
2468   memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
2469
2470   for (i = 0 ; i < key->nr_samplers; i++) {
2471      lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
2472                                      llvm->draw->samplers[PIPE_SHADER_GEOMETRY][i]);
2473   }
2474   for (i = 0 ; i < key->nr_sampler_views; i++) {
2475      lp_sampler_static_texture_state(&draw_sampler[i].texture_state,
2476                                      llvm->draw->sampler_views[PIPE_SHADER_GEOMETRY][i]);
2477   }
2478
2479   return key;
2480}
2481
2482void
2483draw_gs_llvm_dump_variant_key(struct draw_gs_llvm_variant_key *key)
2484{
2485   unsigned i;
2486   struct draw_sampler_static_state *sampler = key->samplers;
2487
2488   for (i = 0 ; i < key->nr_sampler_views; i++) {
2489      debug_printf("sampler[%i].src_format = %s\n", i,
2490                   util_format_name(sampler[i].texture_state.format));
2491   }
2492}
2493