Merge branch 'gallium-index-bias'
[mesa.git] / src / gallium / auxiliary / draw / draw_llvm.c
1 #include "draw_llvm.h"
2
3 #include "draw_context.h"
4 #include "draw_vs.h"
5
6 #include "gallivm/lp_bld_arit.h"
7 #include "gallivm/lp_bld_struct.h"
8 #include "gallivm/lp_bld_type.h"
9 #include "gallivm/lp_bld_flow.h"
10 #include "gallivm/lp_bld_debug.h"
11 #include "gallivm/lp_bld_tgsi.h"
12 #include "gallivm/lp_bld_printf.h"
13
14 #include "tgsi/tgsi_exec.h"
15
16 #include "util/u_cpu_detect.h"
17 #include "util/u_string.h"
18
19 #include <llvm-c/Transforms/Scalar.h>
20
21 #define DEBUG_STORE 0
22
23
24 /* generates the draw jit function */
25 static void
26 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
27
28 static void
29 init_globals(struct draw_llvm *llvm)
30 {
31 LLVMTypeRef texture_type;
32
33 /* struct draw_jit_texture */
34 {
35 LLVMTypeRef elem_types[4];
36
37 elem_types[DRAW_JIT_TEXTURE_WIDTH] = LLVMInt32Type();
38 elem_types[DRAW_JIT_TEXTURE_HEIGHT] = LLVMInt32Type();
39 elem_types[DRAW_JIT_TEXTURE_STRIDE] = LLVMInt32Type();
40 elem_types[DRAW_JIT_TEXTURE_DATA] = LLVMPointerType(LLVMInt8Type(), 0);
41
42 texture_type = LLVMStructType(elem_types, Elements(elem_types), 0);
43
44 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
45 llvm->target, texture_type,
46 DRAW_JIT_TEXTURE_WIDTH);
47 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
48 llvm->target, texture_type,
49 DRAW_JIT_TEXTURE_HEIGHT);
50 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, stride,
51 llvm->target, texture_type,
52 DRAW_JIT_TEXTURE_STRIDE);
53 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data,
54 llvm->target, texture_type,
55 DRAW_JIT_TEXTURE_DATA);
56 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture,
57 llvm->target, texture_type);
58
59 LLVMAddTypeName(llvm->module, "texture", texture_type);
60 }
61
62
63 /* struct draw_jit_context */
64 {
65 LLVMTypeRef elem_types[3];
66 LLVMTypeRef context_type;
67
68 elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
69 elem_types[1] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
70 elem_types[2] = LLVMArrayType(texture_type, PIPE_MAX_SAMPLERS); /* textures */
71
72 context_type = LLVMStructType(elem_types, Elements(elem_types), 0);
73
74 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
75 llvm->target, context_type, 0);
76 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, gs_constants,
77 llvm->target, context_type, 1);
78 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
79 llvm->target, context_type,
80 DRAW_JIT_CONTEXT_TEXTURES_INDEX);
81 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
82 llvm->target, context_type);
83
84 LLVMAddTypeName(llvm->module, "draw_jit_context", context_type);
85
86 llvm->context_ptr_type = LLVMPointerType(context_type, 0);
87 }
88 {
89 LLVMTypeRef buffer_ptr = LLVMPointerType(LLVMIntType(8), 0);
90 llvm->buffer_ptr_type = LLVMPointerType(buffer_ptr, 0);
91 }
92 /* struct pipe_vertex_buffer */
93 {
94 LLVMTypeRef elem_types[4];
95 LLVMTypeRef vb_type;
96
97 elem_types[0] = LLVMInt32Type();
98 elem_types[1] = LLVMInt32Type();
99 elem_types[2] = LLVMInt32Type();
100 elem_types[3] = LLVMPointerType(LLVMOpaqueType(), 0); /* vs_constants */
101
102 vb_type = LLVMStructType(elem_types, Elements(elem_types), 0);
103
104 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
105 llvm->target, vb_type, 0);
106 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
107 llvm->target, vb_type, 2);
108 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer,
109 llvm->target, vb_type);
110
111 LLVMAddTypeName(llvm->module, "pipe_vertex_buffer", vb_type);
112
113 llvm->vb_ptr_type = LLVMPointerType(vb_type, 0);
114 }
115 }
116
117 static LLVMTypeRef
118 create_vertex_header(struct draw_llvm *llvm, int data_elems)
119 {
120 /* struct vertex_header */
121 LLVMTypeRef elem_types[3];
122 LLVMTypeRef vertex_header;
123 char struct_name[24];
124
125 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
126
127 elem_types[0] = LLVMIntType(32);
128 elem_types[1] = LLVMArrayType(LLVMFloatType(), 4);
129 elem_types[2] = LLVMArrayType(elem_types[1], data_elems);
130
131 vertex_header = LLVMStructType(elem_types, Elements(elem_types), 0);
132
133 /* these are bit-fields and we can't take address of them
134 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
135 llvm->target, vertex_header,
136 DRAW_JIT_VERTEX_CLIPMASK);
137 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
138 llvm->target, vertex_header,
139 DRAW_JIT_VERTEX_EDGEFLAG);
140 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
141 llvm->target, vertex_header,
142 DRAW_JIT_VERTEX_PAD);
143 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
144 llvm->target, vertex_header,
145 DRAW_JIT_VERTEX_VERTEX_ID);
146 */
147 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
148 llvm->target, vertex_header,
149 DRAW_JIT_VERTEX_CLIP);
150 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
151 llvm->target, vertex_header,
152 DRAW_JIT_VERTEX_DATA);
153
154 LLVMAddTypeName(llvm->module, struct_name, vertex_header);
155
156 return LLVMPointerType(vertex_header, 0);
157 }
158
159 struct draw_llvm *
160 draw_llvm_create(struct draw_context *draw)
161 {
162 struct draw_llvm *llvm = CALLOC_STRUCT( draw_llvm );
163
164 util_cpu_detect();
165
166 llvm->draw = draw;
167 llvm->engine = draw->engine;
168
169 debug_assert(llvm->engine);
170
171 llvm->module = LLVMModuleCreateWithName("draw_llvm");
172 llvm->provider = LLVMCreateModuleProviderForExistingModule(llvm->module);
173
174 LLVMAddModuleProvider(llvm->engine, llvm->provider);
175
176 llvm->target = LLVMGetExecutionEngineTargetData(llvm->engine);
177
178 llvm->pass = LLVMCreateFunctionPassManager(llvm->provider);
179 LLVMAddTargetData(llvm->target, llvm->pass);
180 /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
181 * but there are more on SVN. */
182 /* TODO: Add more passes */
183 LLVMAddConstantPropagationPass(llvm->pass);
184 if(util_cpu_caps.has_sse4_1) {
185 /* FIXME: There is a bug in this pass, whereby the combination of fptosi
186 * and sitofp (necessary for trunc/floor/ceil/round implementation)
187 * somehow becomes invalid code.
188 */
189 LLVMAddInstructionCombiningPass(llvm->pass);
190 }
191 LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
192 LLVMAddGVNPass(llvm->pass);
193 LLVMAddCFGSimplificationPass(llvm->pass);
194
195 init_globals(llvm);
196
197
198 #if 0
199 LLVMDumpModule(llvm->module);
200 #endif
201
202 return llvm;
203 }
204
205 void
206 draw_llvm_destroy(struct draw_llvm *llvm)
207 {
208 free(llvm);
209 }
210
211 struct draw_llvm_variant *
212 draw_llvm_prepare(struct draw_llvm *llvm, int num_inputs)
213 {
214 struct draw_llvm_variant *variant = MALLOC(sizeof(struct draw_llvm_variant));
215
216 draw_llvm_make_variant_key(llvm, &variant->key);
217
218 llvm->vertex_header_ptr_type = create_vertex_header(llvm, num_inputs);
219
220 draw_llvm_generate(llvm, variant);
221
222 return variant;
223 }
224
225 static void
226 generate_vs(struct draw_llvm *llvm,
227 LLVMBuilderRef builder,
228 LLVMValueRef (*outputs)[NUM_CHANNELS],
229 const LLVMValueRef (*inputs)[NUM_CHANNELS],
230 LLVMValueRef context_ptr)
231 {
232 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
233 struct lp_type vs_type;
234 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(builder, context_ptr);
235
236 memset(&vs_type, 0, sizeof vs_type);
237 vs_type.floating = TRUE; /* floating point values */
238 vs_type.sign = TRUE; /* values are signed */
239 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
240 vs_type.width = 32; /* 32-bit float */
241 vs_type.length = 4; /* 4 elements per vector */
242 #if 0
243 num_vs = 4; /* number of vertices per block */
244 #endif
245
246 /*tgsi_dump(tokens, 0);*/
247 lp_build_tgsi_soa(builder,
248 tokens,
249 vs_type,
250 NULL /*struct lp_build_mask_context *mask*/,
251 consts_ptr,
252 NULL /*pos*/,
253 inputs,
254 outputs,
255 NULL/*sampler*/);
256 }
257
258 #if DEBUG_STORE
259 static void print_vectorf(LLVMBuilderRef builder,
260 LLVMValueRef vec)
261 {
262 LLVMValueRef val[4];
263 val[0] = LLVMBuildExtractElement(builder, vec,
264 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
265 val[1] = LLVMBuildExtractElement(builder, vec,
266 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
267 val[2] = LLVMBuildExtractElement(builder, vec,
268 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
269 val[3] = LLVMBuildExtractElement(builder, vec,
270 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
271 lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n",
272 val[0], val[1], val[2], val[3]);
273 }
274 #endif
275
276 static void
277 generate_fetch(LLVMBuilderRef builder,
278 LLVMValueRef vbuffers_ptr,
279 LLVMValueRef *res,
280 struct pipe_vertex_element *velem,
281 LLVMValueRef vbuf,
282 LLVMValueRef index)
283 {
284 LLVMValueRef indices = LLVMConstInt(LLVMInt64Type(), velem->vertex_buffer_index, 0);
285 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
286 &indices, 1, "");
287 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(builder, vbuf);
288 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(builder, vbuf);
289 LLVMValueRef stride = LLVMBuildMul(builder,
290 vb_stride,
291 index, "");
292
293 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
294
295 stride = LLVMBuildAdd(builder, stride,
296 vb_buffer_offset,
297 "");
298 stride = LLVMBuildAdd(builder, stride,
299 LLVMConstInt(LLVMInt32Type(), velem->src_offset, 0),
300 "");
301
302 /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/
303 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
304
305 *res = draw_llvm_translate_from(builder, vbuffer_ptr, velem->src_format);
306 }
307
308 static LLVMValueRef
309 aos_to_soa(LLVMBuilderRef builder,
310 LLVMValueRef val0,
311 LLVMValueRef val1,
312 LLVMValueRef val2,
313 LLVMValueRef val3,
314 LLVMValueRef channel)
315 {
316 LLVMValueRef ex, res;
317
318 ex = LLVMBuildExtractElement(builder, val0,
319 channel, "");
320 res = LLVMBuildInsertElement(builder,
321 LLVMConstNull(LLVMTypeOf(val0)),
322 ex,
323 LLVMConstInt(LLVMInt32Type(), 0, 0),
324 "");
325
326 ex = LLVMBuildExtractElement(builder, val1,
327 channel, "");
328 res = LLVMBuildInsertElement(builder,
329 res, ex,
330 LLVMConstInt(LLVMInt32Type(), 1, 0),
331 "");
332
333 ex = LLVMBuildExtractElement(builder, val2,
334 channel, "");
335 res = LLVMBuildInsertElement(builder,
336 res, ex,
337 LLVMConstInt(LLVMInt32Type(), 2, 0),
338 "");
339
340 ex = LLVMBuildExtractElement(builder, val3,
341 channel, "");
342 res = LLVMBuildInsertElement(builder,
343 res, ex,
344 LLVMConstInt(LLVMInt32Type(), 3, 0),
345 "");
346
347 return res;
348 }
349
350 static void
351 soa_to_aos(LLVMBuilderRef builder,
352 LLVMValueRef soa[NUM_CHANNELS],
353 LLVMValueRef aos[NUM_CHANNELS])
354 {
355 LLVMValueRef comp;
356 int i = 0;
357
358 debug_assert(NUM_CHANNELS == 4);
359
360 aos[0] = LLVMConstNull(LLVMTypeOf(soa[0]));
361 aos[1] = aos[2] = aos[3] = aos[0];
362
363 for (i = 0; i < NUM_CHANNELS; ++i) {
364 LLVMValueRef channel = LLVMConstInt(LLVMInt32Type(), i, 0);
365
366 comp = LLVMBuildExtractElement(builder, soa[i],
367 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
368 aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, "");
369
370 comp = LLVMBuildExtractElement(builder, soa[i],
371 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
372 aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, "");
373
374 comp = LLVMBuildExtractElement(builder, soa[i],
375 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
376 aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, "");
377
378 comp = LLVMBuildExtractElement(builder, soa[i],
379 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
380 aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, "");
381
382 }
383 }
384
385 static void
386 convert_to_soa(LLVMBuilderRef builder,
387 LLVMValueRef (*aos)[NUM_CHANNELS],
388 LLVMValueRef (*soa)[NUM_CHANNELS],
389 int num_attribs)
390 {
391 int i;
392
393 debug_assert(NUM_CHANNELS == 4);
394
395 for (i = 0; i < num_attribs; ++i) {
396 LLVMValueRef val0 = aos[i][0];
397 LLVMValueRef val1 = aos[i][1];
398 LLVMValueRef val2 = aos[i][2];
399 LLVMValueRef val3 = aos[i][3];
400
401 soa[i][0] = aos_to_soa(builder, val0, val1, val2, val3,
402 LLVMConstInt(LLVMInt32Type(), 0, 0));
403 soa[i][1] = aos_to_soa(builder, val0, val1, val2, val3,
404 LLVMConstInt(LLVMInt32Type(), 1, 0));
405 soa[i][2] = aos_to_soa(builder, val0, val1, val2, val3,
406 LLVMConstInt(LLVMInt32Type(), 2, 0));
407 soa[i][3] = aos_to_soa(builder, val0, val1, val2, val3,
408 LLVMConstInt(LLVMInt32Type(), 3, 0));
409 }
410 }
411
412 static void
413 store_aos(LLVMBuilderRef builder,
414 LLVMValueRef io_ptr,
415 LLVMValueRef index,
416 LLVMValueRef value)
417 {
418 LLVMValueRef id_ptr = draw_jit_header_id(builder, io_ptr);
419 LLVMValueRef data_ptr = draw_jit_header_data(builder, io_ptr);
420 LLVMValueRef indices[3];
421
422 indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
423 indices[1] = index;
424 indices[2] = LLVMConstInt(LLVMInt32Type(), 0, 0);
425
426 /* undefined vertex */
427 LLVMBuildStore(builder, LLVMConstInt(LLVMInt32Type(),
428 0xffff, 0), id_ptr);
429
430 #if DEBUG_STORE
431 lp_build_printf(builder, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
432 #endif
433 #if 0
434 /*lp_build_printf(builder, " ---- %p storing at %d (%p) ", io_ptr, index, data_ptr);
435 print_vectorf(builder, value);*/
436 data_ptr = LLVMBuildBitCast(builder, data_ptr,
437 LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatType(), 4), 0), 0),
438 "datavec");
439 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, "");
440
441 LLVMBuildStore(builder, value, data_ptr);
442 #else
443 {
444 LLVMValueRef x, y, z, w;
445 LLVMValueRef idx0, idx1, idx2, idx3;
446 LLVMValueRef gep0, gep1, gep2, gep3;
447 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
448
449 idx0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
450 idx1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
451 idx2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
452 idx3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
453
454 x = LLVMBuildExtractElement(builder, value,
455 idx0, "");
456 y = LLVMBuildExtractElement(builder, value,
457 idx1, "");
458 z = LLVMBuildExtractElement(builder, value,
459 idx2, "");
460 w = LLVMBuildExtractElement(builder, value,
461 idx3, "");
462
463 gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, "");
464 gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, "");
465 gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, "");
466 gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, "");
467
468 /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n",
469 x, gep0, y, gep1, z, gep2, w, gep3);*/
470 LLVMBuildStore(builder, x, gep0);
471 LLVMBuildStore(builder, y, gep1);
472 LLVMBuildStore(builder, z, gep2);
473 LLVMBuildStore(builder, w, gep3);
474 }
475 #endif
476 }
477
478 static void
479 store_aos_array(LLVMBuilderRef builder,
480 LLVMValueRef io_ptr,
481 LLVMValueRef aos[NUM_CHANNELS],
482 int attrib,
483 int num_outputs)
484 {
485 LLVMValueRef attr_index = LLVMConstInt(LLVMInt32Type(), attrib, 0);
486 LLVMValueRef ind0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
487 LLVMValueRef ind1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
488 LLVMValueRef ind2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
489 LLVMValueRef ind3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
490 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
491
492 debug_assert(NUM_CHANNELS == 4);
493
494 io0_ptr = LLVMBuildGEP(builder, io_ptr,
495 &ind0, 1, "");
496 io1_ptr = LLVMBuildGEP(builder, io_ptr,
497 &ind1, 1, "");
498 io2_ptr = LLVMBuildGEP(builder, io_ptr,
499 &ind2, 1, "");
500 io3_ptr = LLVMBuildGEP(builder, io_ptr,
501 &ind3, 1, "");
502
503 #if DEBUG_STORE
504 lp_build_printf(builder, " io = %p, indexes[%d, %d, %d, %d]\n",
505 io_ptr, ind0, ind1, ind2, ind3);
506 #endif
507
508 store_aos(builder, io0_ptr, attr_index, aos[0]);
509 store_aos(builder, io1_ptr, attr_index, aos[1]);
510 store_aos(builder, io2_ptr, attr_index, aos[2]);
511 store_aos(builder, io3_ptr, attr_index, aos[3]);
512 }
513
514 static void
515 convert_to_aos(LLVMBuilderRef builder,
516 LLVMValueRef io,
517 LLVMValueRef (*outputs)[NUM_CHANNELS],
518 int num_outputs,
519 int max_vertices)
520 {
521 unsigned chan, attrib;
522
523 #if DEBUG_STORE
524 lp_build_printf(builder, " # storing begin\n");
525 #endif
526 for (attrib = 0; attrib < num_outputs; ++attrib) {
527 LLVMValueRef soa[4];
528 LLVMValueRef aos[4];
529 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
530 if(outputs[attrib][chan]) {
531 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
532 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
533 /*lp_build_printf(builder, "output %d : %d ",
534 LLVMConstInt(LLVMInt32Type(), attrib, 0),
535 LLVMConstInt(LLVMInt32Type(), chan, 0));
536 print_vectorf(builder, out);*/
537 soa[chan] = out;
538 } else
539 soa[chan] = 0;
540 }
541 soa_to_aos(builder, soa, aos);
542 store_aos_array(builder,
543 io,
544 aos,
545 attrib,
546 num_outputs);
547 }
548 #if DEBUG_STORE
549 lp_build_printf(builder, " # storing end\n");
550 #endif
551 }
552
553 static void
554 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
555 {
556 LLVMTypeRef arg_types[7];
557 LLVMTypeRef func_type;
558 LLVMValueRef context_ptr;
559 LLVMBasicBlockRef block;
560 LLVMBuilderRef builder;
561 LLVMValueRef start, end, count, stride, step, io_itr;
562 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
563 struct draw_context *draw = llvm->draw;
564 unsigned i, j;
565 struct lp_build_context bld;
566 struct lp_build_loop_state lp_loop;
567 struct lp_type vs_type = lp_type_float_vec(32);
568 const int max_vertices = 4;
569 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
570
571 arg_types[0] = llvm->context_ptr_type; /* context */
572 arg_types[1] = llvm->vertex_header_ptr_type; /* vertex_header */
573 arg_types[2] = llvm->buffer_ptr_type; /* vbuffers */
574 arg_types[3] = LLVMInt32Type(); /* start */
575 arg_types[4] = LLVMInt32Type(); /* count */
576 arg_types[5] = LLVMInt32Type(); /* stride */
577 arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
578
579 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
580
581 variant->function = LLVMAddFunction(llvm->module, "draw_llvm_shader", func_type);
582 LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
583 for(i = 0; i < Elements(arg_types); ++i)
584 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
585 LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
586
587 context_ptr = LLVMGetParam(variant->function, 0);
588 io_ptr = LLVMGetParam(variant->function, 1);
589 vbuffers_ptr = LLVMGetParam(variant->function, 2);
590 start = LLVMGetParam(variant->function, 3);
591 count = LLVMGetParam(variant->function, 4);
592 stride = LLVMGetParam(variant->function, 5);
593 vb_ptr = LLVMGetParam(variant->function, 6);
594
595 lp_build_name(context_ptr, "context");
596 lp_build_name(io_ptr, "io");
597 lp_build_name(vbuffers_ptr, "vbuffers");
598 lp_build_name(start, "start");
599 lp_build_name(count, "count");
600 lp_build_name(stride, "stride");
601 lp_build_name(vb_ptr, "vb");
602
603 /*
604 * Function body
605 */
606
607 block = LLVMAppendBasicBlock(variant->function, "entry");
608 builder = LLVMCreateBuilder();
609 LLVMPositionBuilderAtEnd(builder, block);
610
611 lp_build_context_init(&bld, builder, vs_type);
612
613 end = lp_build_add(&bld, start, count);
614
615 step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
616
617 #if DEBUG_STORE
618 lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
619 start, end, step);
620 #endif
621 lp_build_loop_begin(builder, start, &lp_loop);
622 {
623 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
624 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
625 LLVMValueRef io;
626 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
627
628 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
629 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
630 #if DEBUG_STORE
631 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
632 io_itr, io, lp_loop.counter);
633 #endif
634 for (i = 0; i < NUM_CHANNELS; ++i) {
635 LLVMValueRef true_index = LLVMBuildAdd(
636 builder,
637 lp_loop.counter,
638 LLVMConstInt(LLVMInt32Type(), i, 0), "");
639 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
640 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
641 LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
642 velem->vertex_buffer_index,
643 0);
644 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
645 &vb_index, 1, "");
646 generate_fetch(builder, vbuffers_ptr,
647 &aos_attribs[j][i], velem, vb, true_index);
648 }
649 }
650 convert_to_soa(builder, aos_attribs, inputs,
651 draw->pt.nr_vertex_elements);
652
653 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
654 generate_vs(llvm,
655 builder,
656 outputs,
657 ptr_aos,
658 context_ptr);
659
660 convert_to_aos(builder, io, outputs,
661 draw->vs.vertex_shader->info.num_outputs,
662 max_vertices);
663 }
664 lp_build_loop_end_cond(builder, end, step, LLVMIntUGE, &lp_loop);
665
666 LLVMBuildRetVoid(builder);
667
668 LLVMDisposeBuilder(builder);
669
670 /*
671 * Translate the LLVM IR into machine code.
672 */
673 #ifdef DEBUG
674 if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
675 LLVMDumpValue(variant->function);
676 assert(0);
677 }
678 #endif
679
680 LLVMRunFunctionPassManager(llvm->pass, variant->function);
681
682 if (0) {
683 LLVMDumpValue(variant->function);
684 debug_printf("\n");
685 }
686 variant->jit_func = (draw_jit_vert_func)LLVMGetPointerToGlobal(llvm->draw->engine, variant->function);
687
688 if (0)
689 lp_disassemble(variant->jit_func);
690 }
691
692 void
693 draw_llvm_make_variant_key(struct draw_llvm *llvm,
694 struct draw_llvm_variant_key *key)
695 {
696 memset(key, 0, sizeof(struct draw_llvm_variant_key));
697
698 key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
699
700 memcpy(key->vertex_element,
701 llvm->draw->pt.vertex_element,
702 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
703
704 memcpy(&key->vs,
705 &llvm->draw->vs.vertex_shader->state,
706 sizeof(struct pipe_shader_state));
707 }