Merge branch '7.8'
[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_interp.h"
8 #include "gallivm/lp_bld_struct.h"
9 #include "gallivm/lp_bld_type.h"
10 #include "gallivm/lp_bld_flow.h"
11 #include "gallivm/lp_bld_debug.h"
12 #include "gallivm/lp_bld_tgsi.h"
13 #include "gallivm/lp_bld_printf.h"
14 #include "gallivm/lp_bld_init.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 1
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
226 struct draw_context *draw_create_with_llvm(void)
227 {
228 struct draw_context *draw = CALLOC_STRUCT( draw_context );
229 if (draw == NULL)
230 goto fail;
231
232 assert(lp_build_engine);
233 draw->engine = lp_build_engine;
234
235 if (!draw_init(draw))
236 goto fail;
237
238 return draw;
239
240 fail:
241 draw_destroy( draw );
242 return NULL;
243 }
244
245 static void
246 generate_vs(struct draw_llvm *llvm,
247 LLVMBuilderRef builder,
248 LLVMValueRef (*outputs)[NUM_CHANNELS],
249 const LLVMValueRef (*inputs)[NUM_CHANNELS],
250 LLVMValueRef context_ptr)
251 {
252 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
253 struct lp_type vs_type;
254 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(builder, context_ptr);
255
256 memset(&vs_type, 0, sizeof vs_type);
257 vs_type.floating = TRUE; /* floating point values */
258 vs_type.sign = TRUE; /* values are signed */
259 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
260 vs_type.width = 32; /* 32-bit float */
261 vs_type.length = 4; /* 4 elements per vector */
262 #if 0
263 num_vs = 4; /* number of vertices per block */
264 #endif
265
266 /*tgsi_dump(tokens, 0);*/
267 lp_build_tgsi_soa(builder,
268 tokens,
269 vs_type,
270 NULL /*struct lp_build_mask_context *mask*/,
271 consts_ptr,
272 NULL /*pos*/,
273 inputs,
274 outputs,
275 NULL/*sampler*/);
276 }
277
278 #if DEBUG_STORE
279 static void print_vectorf(LLVMBuilderRef builder,
280 LLVMValueRef vec)
281 {
282 LLVMValueRef val[4];
283 val[0] = LLVMBuildExtractElement(builder, vec,
284 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
285 val[1] = LLVMBuildExtractElement(builder, vec,
286 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
287 val[2] = LLVMBuildExtractElement(builder, vec,
288 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
289 val[3] = LLVMBuildExtractElement(builder, vec,
290 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
291 lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n",
292 val[0], val[1], val[2], val[3]);
293 }
294 #endif
295
296 static void
297 generate_fetch(LLVMBuilderRef builder,
298 LLVMValueRef vbuffers_ptr,
299 LLVMValueRef *res,
300 struct pipe_vertex_element *velem,
301 LLVMValueRef vbuf,
302 LLVMValueRef index)
303 {
304 LLVMValueRef indices = LLVMConstInt(LLVMInt64Type(), velem->vertex_buffer_index, 0);
305 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
306 &indices, 1, "");
307 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(builder, vbuf);
308 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(builder, vbuf);
309 LLVMValueRef stride = LLVMBuildMul(builder,
310 vb_stride,
311 index, "");
312
313 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
314
315 stride = LLVMBuildAdd(builder, stride,
316 vb_buffer_offset,
317 "");
318 stride = LLVMBuildAdd(builder, stride,
319 LLVMConstInt(LLVMInt32Type(), velem->src_offset, 0),
320 "");
321
322 /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/
323 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
324
325 *res = draw_llvm_translate_from(builder, vbuffer_ptr, velem->src_format);
326 }
327
328 static LLVMValueRef
329 aos_to_soa(LLVMBuilderRef builder,
330 LLVMValueRef val0,
331 LLVMValueRef val1,
332 LLVMValueRef val2,
333 LLVMValueRef val3,
334 LLVMValueRef channel)
335 {
336 LLVMValueRef ex, res;
337
338 ex = LLVMBuildExtractElement(builder, val0,
339 channel, "");
340 res = LLVMBuildInsertElement(builder,
341 LLVMConstNull(LLVMTypeOf(val0)),
342 ex,
343 LLVMConstInt(LLVMInt32Type(), 0, 0),
344 "");
345
346 ex = LLVMBuildExtractElement(builder, val1,
347 channel, "");
348 res = LLVMBuildInsertElement(builder,
349 res, ex,
350 LLVMConstInt(LLVMInt32Type(), 1, 0),
351 "");
352
353 ex = LLVMBuildExtractElement(builder, val2,
354 channel, "");
355 res = LLVMBuildInsertElement(builder,
356 res, ex,
357 LLVMConstInt(LLVMInt32Type(), 2, 0),
358 "");
359
360 ex = LLVMBuildExtractElement(builder, val3,
361 channel, "");
362 res = LLVMBuildInsertElement(builder,
363 res, ex,
364 LLVMConstInt(LLVMInt32Type(), 3, 0),
365 "");
366
367 return res;
368 }
369
370 static void
371 soa_to_aos(LLVMBuilderRef builder,
372 LLVMValueRef soa[NUM_CHANNELS],
373 LLVMValueRef aos[NUM_CHANNELS])
374 {
375 LLVMValueRef comp;
376 int i = 0;
377
378 debug_assert(NUM_CHANNELS == 4);
379
380 aos[0] = LLVMConstNull(LLVMTypeOf(soa[0]));
381 aos[1] = aos[2] = aos[3] = aos[0];
382
383 for (i = 0; i < NUM_CHANNELS; ++i) {
384 LLVMValueRef channel = LLVMConstInt(LLVMInt32Type(), i, 0);
385
386 comp = LLVMBuildExtractElement(builder, soa[i],
387 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
388 aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, "");
389
390 comp = LLVMBuildExtractElement(builder, soa[i],
391 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
392 aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, "");
393
394 comp = LLVMBuildExtractElement(builder, soa[i],
395 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
396 aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, "");
397
398 comp = LLVMBuildExtractElement(builder, soa[i],
399 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
400 aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, "");
401
402 }
403 }
404
405 static void
406 convert_to_soa(LLVMBuilderRef builder,
407 LLVMValueRef (*aos)[NUM_CHANNELS],
408 LLVMValueRef (*soa)[NUM_CHANNELS],
409 int num_attribs)
410 {
411 int i;
412
413 debug_assert(NUM_CHANNELS == 4);
414
415 for (i = 0; i < num_attribs; ++i) {
416 LLVMValueRef val0 = aos[i][0];
417 LLVMValueRef val1 = aos[i][1];
418 LLVMValueRef val2 = aos[i][2];
419 LLVMValueRef val3 = aos[i][3];
420
421 soa[i][0] = aos_to_soa(builder, val0, val1, val2, val3,
422 LLVMConstInt(LLVMInt32Type(), 0, 0));
423 soa[i][1] = aos_to_soa(builder, val0, val1, val2, val3,
424 LLVMConstInt(LLVMInt32Type(), 1, 0));
425 soa[i][2] = aos_to_soa(builder, val0, val1, val2, val3,
426 LLVMConstInt(LLVMInt32Type(), 2, 0));
427 soa[i][3] = aos_to_soa(builder, val0, val1, val2, val3,
428 LLVMConstInt(LLVMInt32Type(), 3, 0));
429 }
430 }
431
432 static void
433 store_aos(LLVMBuilderRef builder,
434 LLVMValueRef io_ptr,
435 LLVMValueRef index,
436 LLVMValueRef value)
437 {
438 LLVMValueRef id_ptr = draw_jit_header_id(builder, io_ptr);
439 LLVMValueRef data_ptr = draw_jit_header_data(builder, io_ptr);
440 LLVMValueRef indices[3];
441
442 indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
443 indices[1] = index;
444 indices[2] = LLVMConstInt(LLVMInt32Type(), 0, 0);
445
446 /* undefined vertex */
447 LLVMBuildStore(builder, LLVMConstInt(LLVMInt32Type(),
448 0xffff, 0), id_ptr);
449
450 #if DEBUG_STORE
451 lp_build_printf(builder, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
452 #endif
453 #if 0
454 /*lp_build_printf(builder, " ---- %p storing at %d (%p) ", io_ptr, index, data_ptr);
455 print_vectorf(builder, value);*/
456 data_ptr = LLVMBuildBitCast(builder, data_ptr,
457 LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatType(), 4), 0), 0),
458 "datavec");
459 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, "");
460
461 LLVMBuildStore(builder, value, data_ptr);
462 #else
463 {
464 LLVMValueRef x, y, z, w;
465 LLVMValueRef idx0, idx1, idx2, idx3;
466 LLVMValueRef gep0, gep1, gep2, gep3;
467 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
468
469 idx0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
470 idx1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
471 idx2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
472 idx3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
473
474 x = LLVMBuildExtractElement(builder, value,
475 idx0, "");
476 y = LLVMBuildExtractElement(builder, value,
477 idx1, "");
478 z = LLVMBuildExtractElement(builder, value,
479 idx2, "");
480 w = LLVMBuildExtractElement(builder, value,
481 idx3, "");
482
483 gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, "");
484 gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, "");
485 gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, "");
486 gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, "");
487
488 /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n",
489 x, gep0, y, gep1, z, gep2, w, gep3);*/
490 LLVMBuildStore(builder, x, gep0);
491 LLVMBuildStore(builder, y, gep1);
492 LLVMBuildStore(builder, z, gep2);
493 LLVMBuildStore(builder, w, gep3);
494 }
495 #endif
496 }
497
498 static void
499 store_aos_array(LLVMBuilderRef builder,
500 LLVMValueRef io_ptr,
501 LLVMValueRef aos[NUM_CHANNELS],
502 int attrib,
503 int num_outputs)
504 {
505 LLVMValueRef attr_index = LLVMConstInt(LLVMInt32Type(), attrib, 0);
506 LLVMValueRef ind0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
507 LLVMValueRef ind1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
508 LLVMValueRef ind2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
509 LLVMValueRef ind3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
510 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
511
512 debug_assert(NUM_CHANNELS == 4);
513
514 io0_ptr = LLVMBuildGEP(builder, io_ptr,
515 &ind0, 1, "");
516 io1_ptr = LLVMBuildGEP(builder, io_ptr,
517 &ind1, 1, "");
518 io2_ptr = LLVMBuildGEP(builder, io_ptr,
519 &ind2, 1, "");
520 io3_ptr = LLVMBuildGEP(builder, io_ptr,
521 &ind3, 1, "");
522
523 #if DEBUG_STORE
524 lp_build_printf(builder, " io = %p, indexes[%d, %d, %d, %d]\n",
525 io_ptr, ind0, ind1, ind2, ind3);
526 #endif
527
528 store_aos(builder, io0_ptr, attr_index, aos[0]);
529 store_aos(builder, io1_ptr, attr_index, aos[1]);
530 store_aos(builder, io2_ptr, attr_index, aos[2]);
531 store_aos(builder, io3_ptr, attr_index, aos[3]);
532 }
533
534 static void
535 convert_to_aos(LLVMBuilderRef builder,
536 LLVMValueRef io,
537 LLVMValueRef (*outputs)[NUM_CHANNELS],
538 int num_outputs,
539 int max_vertices)
540 {
541 unsigned chan, attrib;
542
543 #if DEBUG_STORE
544 lp_build_printf(builder, " # storing begin\n");
545 #endif
546 for (attrib = 0; attrib < num_outputs; ++attrib) {
547 LLVMValueRef soa[4];
548 LLVMValueRef aos[4];
549 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
550 if(outputs[attrib][chan]) {
551 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
552 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
553 /*lp_build_printf(builder, "output %d : %d ",
554 LLVMConstInt(LLVMInt32Type(), attrib, 0),
555 LLVMConstInt(LLVMInt32Type(), chan, 0));
556 print_vectorf(builder, out);*/
557 soa[chan] = out;
558 } else
559 soa[chan] = 0;
560 }
561 soa_to_aos(builder, soa, aos);
562 store_aos_array(builder,
563 io,
564 aos,
565 attrib,
566 num_outputs);
567 }
568 #if DEBUG_STORE
569 lp_build_printf(builder, " # storing end\n");
570 #endif
571 }
572
573 static void
574 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
575 {
576 LLVMTypeRef arg_types[7];
577 LLVMTypeRef func_type;
578 LLVMValueRef context_ptr;
579 LLVMBasicBlockRef block;
580 LLVMBuilderRef builder;
581 LLVMValueRef start, end, count, stride, step, io_itr;
582 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
583 struct draw_context *draw = llvm->draw;
584 unsigned i, j;
585 struct lp_build_context bld;
586 struct lp_build_loop_state lp_loop;
587 struct lp_type vs_type = lp_type_float_vec(32);
588 const int max_vertices = 4;
589 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
590
591 arg_types[0] = llvm->context_ptr_type; /* context */
592 arg_types[1] = llvm->vertex_header_ptr_type; /* vertex_header */
593 arg_types[2] = llvm->buffer_ptr_type; /* vbuffers */
594 arg_types[3] = LLVMInt32Type(); /* start */
595 arg_types[4] = LLVMInt32Type(); /* count */
596 arg_types[5] = LLVMInt32Type(); /* stride */
597 arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
598
599 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
600
601 variant->function = LLVMAddFunction(llvm->module, "draw_llvm_shader", func_type);
602 LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
603 for(i = 0; i < Elements(arg_types); ++i)
604 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
605 LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
606
607 context_ptr = LLVMGetParam(variant->function, 0);
608 io_ptr = LLVMGetParam(variant->function, 1);
609 vbuffers_ptr = LLVMGetParam(variant->function, 2);
610 start = LLVMGetParam(variant->function, 3);
611 count = LLVMGetParam(variant->function, 4);
612 stride = LLVMGetParam(variant->function, 5);
613 vb_ptr = LLVMGetParam(variant->function, 6);
614
615 lp_build_name(context_ptr, "context");
616 lp_build_name(io_ptr, "io");
617 lp_build_name(vbuffers_ptr, "vbuffers");
618 lp_build_name(start, "start");
619 lp_build_name(count, "count");
620 lp_build_name(stride, "stride");
621 lp_build_name(vb_ptr, "vb");
622
623 /*
624 * Function body
625 */
626
627 block = LLVMAppendBasicBlock(variant->function, "entry");
628 builder = LLVMCreateBuilder();
629 LLVMPositionBuilderAtEnd(builder, block);
630
631 lp_build_context_init(&bld, builder, vs_type);
632
633 end = lp_build_add(&bld, start, count);
634
635 step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
636
637 #if DEBUG_STORE
638 lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
639 start, end, step);
640 #endif
641 lp_build_loop_begin(builder, start, &lp_loop);
642 {
643 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
644 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
645 LLVMValueRef io;
646 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
647
648 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
649 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
650 #if DEBUG_STORE
651 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
652 io_itr, io, lp_loop.counter);
653 #endif
654 for (i = 0; i < NUM_CHANNELS; ++i) {
655 LLVMValueRef true_index = LLVMBuildAdd(
656 builder,
657 lp_loop.counter,
658 LLVMConstInt(LLVMInt32Type(), i, 0), "");
659 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
660 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
661 LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
662 velem->vertex_buffer_index,
663 0);
664 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
665 &vb_index, 1, "");
666 generate_fetch(builder, vbuffers_ptr,
667 &aos_attribs[j][i], velem, vb, true_index);
668 }
669 }
670 convert_to_soa(builder, aos_attribs, inputs,
671 draw->pt.nr_vertex_elements);
672
673 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
674 generate_vs(llvm,
675 builder,
676 outputs,
677 ptr_aos,
678 context_ptr);
679
680 convert_to_aos(builder, io, outputs,
681 draw->vs.vertex_shader->info.num_outputs,
682 max_vertices);
683 }
684 lp_build_loop_end_cond(builder, end, step, LLVMIntUGE, &lp_loop);
685
686 LLVMBuildRetVoid(builder);
687
688 LLVMDisposeBuilder(builder);
689
690 /*
691 * Translate the LLVM IR into machine code.
692 */
693 #ifdef DEBUG
694 if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
695 LLVMDumpValue(variant->function);
696 assert(0);
697 }
698 #endif
699
700 LLVMRunFunctionPassManager(llvm->pass, variant->function);
701
702 if (0) {
703 LLVMDumpValue(variant->function);
704 debug_printf("\n");
705 }
706 variant->jit_func = (draw_jit_vert_func)LLVMGetPointerToGlobal(llvm->draw->engine, variant->function);
707
708 if (0)
709 lp_disassemble(variant->jit_func);
710 }
711
712 void
713 draw_llvm_make_variant_key(struct draw_llvm *llvm,
714 struct draw_llvm_variant_key *key)
715 {
716 memset(key, 0, sizeof(struct draw_llvm_variant_key));
717
718 key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
719
720 memcpy(key->vertex_element,
721 llvm->draw->pt.vertex_element,
722 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
723
724 memcpy(&key->vs,
725 &llvm->draw->vs.vertex_shader->state,
726 sizeof(struct pipe_shader_state));
727 }