gallivm: Add an alternative to LLVMDumpValue that works with Windows GUI apps.
[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 static void
28 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *var);
29
30 static void
31 init_globals(struct draw_llvm *llvm)
32 {
33 LLVMTypeRef texture_type;
34
35 /* struct draw_jit_texture */
36 {
37 LLVMTypeRef elem_types[4];
38
39 elem_types[DRAW_JIT_TEXTURE_WIDTH] = LLVMInt32Type();
40 elem_types[DRAW_JIT_TEXTURE_HEIGHT] = LLVMInt32Type();
41 elem_types[DRAW_JIT_TEXTURE_STRIDE] = LLVMInt32Type();
42 elem_types[DRAW_JIT_TEXTURE_DATA] = LLVMPointerType(LLVMInt8Type(), 0);
43
44 texture_type = LLVMStructType(elem_types, Elements(elem_types), 0);
45
46 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
47 llvm->target, texture_type,
48 DRAW_JIT_TEXTURE_WIDTH);
49 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
50 llvm->target, texture_type,
51 DRAW_JIT_TEXTURE_HEIGHT);
52 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, stride,
53 llvm->target, texture_type,
54 DRAW_JIT_TEXTURE_STRIDE);
55 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data,
56 llvm->target, texture_type,
57 DRAW_JIT_TEXTURE_DATA);
58 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture,
59 llvm->target, texture_type);
60
61 LLVMAddTypeName(llvm->module, "texture", texture_type);
62 }
63
64
65 /* struct draw_jit_context */
66 {
67 LLVMTypeRef elem_types[3];
68 LLVMTypeRef context_type;
69
70 elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
71 elem_types[1] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
72 elem_types[2] = LLVMArrayType(texture_type, PIPE_MAX_SAMPLERS); /* textures */
73
74 context_type = LLVMStructType(elem_types, Elements(elem_types), 0);
75
76 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
77 llvm->target, context_type, 0);
78 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, gs_constants,
79 llvm->target, context_type, 1);
80 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
81 llvm->target, context_type,
82 DRAW_JIT_CONTEXT_TEXTURES_INDEX);
83 LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
84 llvm->target, context_type);
85
86 LLVMAddTypeName(llvm->module, "draw_jit_context", context_type);
87
88 llvm->context_ptr_type = LLVMPointerType(context_type, 0);
89 }
90 {
91 LLVMTypeRef buffer_ptr = LLVMPointerType(LLVMIntType(8), 0);
92 llvm->buffer_ptr_type = LLVMPointerType(buffer_ptr, 0);
93 }
94 /* struct pipe_vertex_buffer */
95 {
96 LLVMTypeRef elem_types[4];
97 LLVMTypeRef vb_type;
98
99 elem_types[0] = LLVMInt32Type();
100 elem_types[1] = LLVMInt32Type();
101 elem_types[2] = LLVMInt32Type();
102 elem_types[3] = LLVMPointerType(LLVMOpaqueType(), 0); /* vs_constants */
103
104 vb_type = LLVMStructType(elem_types, Elements(elem_types), 0);
105
106 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
107 llvm->target, vb_type, 0);
108 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
109 llvm->target, vb_type, 2);
110 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer,
111 llvm->target, vb_type);
112
113 LLVMAddTypeName(llvm->module, "pipe_vertex_buffer", vb_type);
114
115 llvm->vb_ptr_type = LLVMPointerType(vb_type, 0);
116 }
117 }
118
119 static LLVMTypeRef
120 create_vertex_header(struct draw_llvm *llvm, int data_elems)
121 {
122 /* struct vertex_header */
123 LLVMTypeRef elem_types[3];
124 LLVMTypeRef vertex_header;
125 char struct_name[24];
126
127 util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
128
129 elem_types[0] = LLVMIntType(32);
130 elem_types[1] = LLVMArrayType(LLVMFloatType(), 4);
131 elem_types[2] = LLVMArrayType(elem_types[1], data_elems);
132
133 vertex_header = LLVMStructType(elem_types, Elements(elem_types), 0);
134
135 /* these are bit-fields and we can't take address of them
136 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
137 llvm->target, vertex_header,
138 DRAW_JIT_VERTEX_CLIPMASK);
139 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
140 llvm->target, vertex_header,
141 DRAW_JIT_VERTEX_EDGEFLAG);
142 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
143 llvm->target, vertex_header,
144 DRAW_JIT_VERTEX_PAD);
145 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
146 llvm->target, vertex_header,
147 DRAW_JIT_VERTEX_VERTEX_ID);
148 */
149 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
150 llvm->target, vertex_header,
151 DRAW_JIT_VERTEX_CLIP);
152 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
153 llvm->target, vertex_header,
154 DRAW_JIT_VERTEX_DATA);
155
156 LLVMAddTypeName(llvm->module, struct_name, vertex_header);
157
158 return LLVMPointerType(vertex_header, 0);
159 }
160
161 struct draw_llvm *
162 draw_llvm_create(struct draw_context *draw)
163 {
164 struct draw_llvm *llvm;
165
166 #ifdef PIPE_ARCH_X86
167 util_cpu_detect();
168 /* require SSE2 due to LLVM PR6960. */
169 if (!util_cpu_caps.has_sse2)
170 return NULL;
171 #endif
172
173 llvm = CALLOC_STRUCT( draw_llvm );
174
175 llvm->draw = draw;
176 llvm->engine = draw->engine;
177
178 debug_assert(llvm->engine);
179
180 llvm->module = LLVMModuleCreateWithName("draw_llvm");
181 llvm->provider = LLVMCreateModuleProviderForExistingModule(llvm->module);
182
183 LLVMAddModuleProvider(llvm->engine, llvm->provider);
184
185 llvm->target = LLVMGetExecutionEngineTargetData(llvm->engine);
186
187 llvm->pass = LLVMCreateFunctionPassManager(llvm->provider);
188 LLVMAddTargetData(llvm->target, llvm->pass);
189 /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
190 * but there are more on SVN. */
191 /* TODO: Add more passes */
192 LLVMAddCFGSimplificationPass(llvm->pass);
193 LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
194 LLVMAddConstantPropagationPass(llvm->pass);
195 if(util_cpu_caps.has_sse4_1) {
196 /* FIXME: There is a bug in this pass, whereby the combination of fptosi
197 * and sitofp (necessary for trunc/floor/ceil/round implementation)
198 * somehow becomes invalid code.
199 */
200 LLVMAddInstructionCombiningPass(llvm->pass);
201 }
202 LLVMAddGVNPass(llvm->pass);
203
204 init_globals(llvm);
205
206
207 #if 0
208 LLVMDumpModule(lp_build_module);
209 #endif
210
211 return llvm;
212 }
213
214 void
215 draw_llvm_destroy(struct draw_llvm *llvm)
216 {
217 LLVMDisposePassManager(llvm->pass);
218
219 FREE(llvm);
220 }
221
222 struct draw_llvm_variant *
223 draw_llvm_prepare(struct draw_llvm *llvm, int num_inputs)
224 {
225 struct draw_llvm_variant *variant = MALLOC(sizeof(struct draw_llvm_variant));
226
227 draw_llvm_make_variant_key(llvm, &variant->key);
228
229 llvm->vertex_header_ptr_type = create_vertex_header(llvm, num_inputs);
230
231 draw_llvm_generate(llvm, variant);
232 draw_llvm_generate_elts(llvm, variant);
233
234 return variant;
235 }
236
237 static void
238 generate_vs(struct draw_llvm *llvm,
239 LLVMBuilderRef builder,
240 LLVMValueRef (*outputs)[NUM_CHANNELS],
241 const LLVMValueRef (*inputs)[NUM_CHANNELS],
242 LLVMValueRef context_ptr)
243 {
244 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
245 struct lp_type vs_type;
246 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(builder, context_ptr);
247
248 memset(&vs_type, 0, sizeof vs_type);
249 vs_type.floating = TRUE; /* floating point values */
250 vs_type.sign = TRUE; /* values are signed */
251 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
252 vs_type.width = 32; /* 32-bit float */
253 vs_type.length = 4; /* 4 elements per vector */
254 #if 0
255 num_vs = 4; /* number of vertices per block */
256 #endif
257
258 /*tgsi_dump(tokens, 0);*/
259 lp_build_tgsi_soa(builder,
260 tokens,
261 vs_type,
262 NULL /*struct lp_build_mask_context *mask*/,
263 consts_ptr,
264 NULL /*pos*/,
265 inputs,
266 outputs,
267 NULL/*sampler*/,
268 &llvm->draw->vs.vertex_shader->info);
269 }
270
271 #if DEBUG_STORE
272 static void print_vectorf(LLVMBuilderRef builder,
273 LLVMValueRef vec)
274 {
275 LLVMValueRef val[4];
276 val[0] = LLVMBuildExtractElement(builder, vec,
277 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
278 val[1] = LLVMBuildExtractElement(builder, vec,
279 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
280 val[2] = LLVMBuildExtractElement(builder, vec,
281 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
282 val[3] = LLVMBuildExtractElement(builder, vec,
283 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
284 lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n",
285 val[0], val[1], val[2], val[3]);
286 }
287 #endif
288
289 static void
290 generate_fetch(LLVMBuilderRef builder,
291 LLVMValueRef vbuffers_ptr,
292 LLVMValueRef *res,
293 struct pipe_vertex_element *velem,
294 LLVMValueRef vbuf,
295 LLVMValueRef index)
296 {
297 LLVMValueRef indices = LLVMConstInt(LLVMInt64Type(), velem->vertex_buffer_index, 0);
298 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
299 &indices, 1, "");
300 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(builder, vbuf);
301 LLVMValueRef vb_max_index = draw_jit_vbuffer_max_index(builder, vbuf);
302 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(builder, vbuf);
303 LLVMValueRef cond;
304 LLVMValueRef stride;
305
306 cond = LLVMBuildICmp(builder, LLVMIntULE, index, vb_max_index, "");
307
308 index = LLVMBuildSelect(builder, cond, index, vb_max_index, "");
309
310 stride = LLVMBuildMul(builder, vb_stride, index, "");
311
312 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
313
314 stride = LLVMBuildAdd(builder, stride,
315 vb_buffer_offset,
316 "");
317 stride = LLVMBuildAdd(builder, stride,
318 LLVMConstInt(LLVMInt32Type(), velem->src_offset, 0),
319 "");
320
321 /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/
322 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
323
324 *res = draw_llvm_translate_from(builder, vbuffer_ptr, velem->src_format);
325 }
326
327 static LLVMValueRef
328 aos_to_soa(LLVMBuilderRef builder,
329 LLVMValueRef val0,
330 LLVMValueRef val1,
331 LLVMValueRef val2,
332 LLVMValueRef val3,
333 LLVMValueRef channel)
334 {
335 LLVMValueRef ex, res;
336
337 ex = LLVMBuildExtractElement(builder, val0,
338 channel, "");
339 res = LLVMBuildInsertElement(builder,
340 LLVMConstNull(LLVMTypeOf(val0)),
341 ex,
342 LLVMConstInt(LLVMInt32Type(), 0, 0),
343 "");
344
345 ex = LLVMBuildExtractElement(builder, val1,
346 channel, "");
347 res = LLVMBuildInsertElement(builder,
348 res, ex,
349 LLVMConstInt(LLVMInt32Type(), 1, 0),
350 "");
351
352 ex = LLVMBuildExtractElement(builder, val2,
353 channel, "");
354 res = LLVMBuildInsertElement(builder,
355 res, ex,
356 LLVMConstInt(LLVMInt32Type(), 2, 0),
357 "");
358
359 ex = LLVMBuildExtractElement(builder, val3,
360 channel, "");
361 res = LLVMBuildInsertElement(builder,
362 res, ex,
363 LLVMConstInt(LLVMInt32Type(), 3, 0),
364 "");
365
366 return res;
367 }
368
369 static void
370 soa_to_aos(LLVMBuilderRef builder,
371 LLVMValueRef soa[NUM_CHANNELS],
372 LLVMValueRef aos[NUM_CHANNELS])
373 {
374 LLVMValueRef comp;
375 int i = 0;
376
377 debug_assert(NUM_CHANNELS == 4);
378
379 aos[0] = LLVMConstNull(LLVMTypeOf(soa[0]));
380 aos[1] = aos[2] = aos[3] = aos[0];
381
382 for (i = 0; i < NUM_CHANNELS; ++i) {
383 LLVMValueRef channel = LLVMConstInt(LLVMInt32Type(), i, 0);
384
385 comp = LLVMBuildExtractElement(builder, soa[i],
386 LLVMConstInt(LLVMInt32Type(), 0, 0), "");
387 aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, "");
388
389 comp = LLVMBuildExtractElement(builder, soa[i],
390 LLVMConstInt(LLVMInt32Type(), 1, 0), "");
391 aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, "");
392
393 comp = LLVMBuildExtractElement(builder, soa[i],
394 LLVMConstInt(LLVMInt32Type(), 2, 0), "");
395 aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, "");
396
397 comp = LLVMBuildExtractElement(builder, soa[i],
398 LLVMConstInt(LLVMInt32Type(), 3, 0), "");
399 aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, "");
400
401 }
402 }
403
404 static void
405 convert_to_soa(LLVMBuilderRef builder,
406 LLVMValueRef (*aos)[NUM_CHANNELS],
407 LLVMValueRef (*soa)[NUM_CHANNELS],
408 int num_attribs)
409 {
410 int i;
411
412 debug_assert(NUM_CHANNELS == 4);
413
414 for (i = 0; i < num_attribs; ++i) {
415 LLVMValueRef val0 = aos[i][0];
416 LLVMValueRef val1 = aos[i][1];
417 LLVMValueRef val2 = aos[i][2];
418 LLVMValueRef val3 = aos[i][3];
419
420 soa[i][0] = aos_to_soa(builder, val0, val1, val2, val3,
421 LLVMConstInt(LLVMInt32Type(), 0, 0));
422 soa[i][1] = aos_to_soa(builder, val0, val1, val2, val3,
423 LLVMConstInt(LLVMInt32Type(), 1, 0));
424 soa[i][2] = aos_to_soa(builder, val0, val1, val2, val3,
425 LLVMConstInt(LLVMInt32Type(), 2, 0));
426 soa[i][3] = aos_to_soa(builder, val0, val1, val2, val3,
427 LLVMConstInt(LLVMInt32Type(), 3, 0));
428 }
429 }
430
431 static void
432 store_aos(LLVMBuilderRef builder,
433 LLVMValueRef io_ptr,
434 LLVMValueRef index,
435 LLVMValueRef value)
436 {
437 LLVMValueRef id_ptr = draw_jit_header_id(builder, io_ptr);
438 LLVMValueRef data_ptr = draw_jit_header_data(builder, io_ptr);
439 LLVMValueRef indices[3];
440
441 indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
442 indices[1] = index;
443 indices[2] = LLVMConstInt(LLVMInt32Type(), 0, 0);
444
445 /* undefined vertex */
446 LLVMBuildStore(builder, LLVMConstInt(LLVMInt32Type(),
447 0xffff, 0), id_ptr);
448
449 #if DEBUG_STORE
450 lp_build_printf(builder, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
451 #endif
452 #if 0
453 /*lp_build_printf(builder, " ---- %p storing at %d (%p) ", io_ptr, index, data_ptr);
454 print_vectorf(builder, value);*/
455 data_ptr = LLVMBuildBitCast(builder, data_ptr,
456 LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatType(), 4), 0), 0),
457 "datavec");
458 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, "");
459
460 LLVMBuildStore(builder, value, data_ptr);
461 #else
462 {
463 LLVMValueRef x, y, z, w;
464 LLVMValueRef idx0, idx1, idx2, idx3;
465 LLVMValueRef gep0, gep1, gep2, gep3;
466 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
467
468 idx0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
469 idx1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
470 idx2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
471 idx3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
472
473 x = LLVMBuildExtractElement(builder, value,
474 idx0, "");
475 y = LLVMBuildExtractElement(builder, value,
476 idx1, "");
477 z = LLVMBuildExtractElement(builder, value,
478 idx2, "");
479 w = LLVMBuildExtractElement(builder, value,
480 idx3, "");
481
482 gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, "");
483 gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, "");
484 gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, "");
485 gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, "");
486
487 /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n",
488 x, gep0, y, gep1, z, gep2, w, gep3);*/
489 LLVMBuildStore(builder, x, gep0);
490 LLVMBuildStore(builder, y, gep1);
491 LLVMBuildStore(builder, z, gep2);
492 LLVMBuildStore(builder, w, gep3);
493 }
494 #endif
495 }
496
497 static void
498 store_aos_array(LLVMBuilderRef builder,
499 LLVMValueRef io_ptr,
500 LLVMValueRef aos[NUM_CHANNELS],
501 int attrib,
502 int num_outputs)
503 {
504 LLVMValueRef attr_index = LLVMConstInt(LLVMInt32Type(), attrib, 0);
505 LLVMValueRef ind0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
506 LLVMValueRef ind1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
507 LLVMValueRef ind2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
508 LLVMValueRef ind3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
509 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
510
511 debug_assert(NUM_CHANNELS == 4);
512
513 io0_ptr = LLVMBuildGEP(builder, io_ptr,
514 &ind0, 1, "");
515 io1_ptr = LLVMBuildGEP(builder, io_ptr,
516 &ind1, 1, "");
517 io2_ptr = LLVMBuildGEP(builder, io_ptr,
518 &ind2, 1, "");
519 io3_ptr = LLVMBuildGEP(builder, io_ptr,
520 &ind3, 1, "");
521
522 #if DEBUG_STORE
523 lp_build_printf(builder, " io = %p, indexes[%d, %d, %d, %d]\n",
524 io_ptr, ind0, ind1, ind2, ind3);
525 #endif
526
527 store_aos(builder, io0_ptr, attr_index, aos[0]);
528 store_aos(builder, io1_ptr, attr_index, aos[1]);
529 store_aos(builder, io2_ptr, attr_index, aos[2]);
530 store_aos(builder, io3_ptr, attr_index, aos[3]);
531 }
532
533 static void
534 convert_to_aos(LLVMBuilderRef builder,
535 LLVMValueRef io,
536 LLVMValueRef (*outputs)[NUM_CHANNELS],
537 int num_outputs,
538 int max_vertices)
539 {
540 unsigned chan, attrib;
541
542 #if DEBUG_STORE
543 lp_build_printf(builder, " # storing begin\n");
544 #endif
545 for (attrib = 0; attrib < num_outputs; ++attrib) {
546 LLVMValueRef soa[4];
547 LLVMValueRef aos[4];
548 for(chan = 0; chan < NUM_CHANNELS; ++chan) {
549 if(outputs[attrib][chan]) {
550 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
551 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
552 /*lp_build_printf(builder, "output %d : %d ",
553 LLVMConstInt(LLVMInt32Type(), attrib, 0),
554 LLVMConstInt(LLVMInt32Type(), chan, 0));
555 print_vectorf(builder, out);*/
556 soa[chan] = out;
557 } else
558 soa[chan] = 0;
559 }
560 soa_to_aos(builder, soa, aos);
561 store_aos_array(builder,
562 io,
563 aos,
564 attrib,
565 num_outputs);
566 }
567 #if DEBUG_STORE
568 lp_build_printf(builder, " # storing end\n");
569 #endif
570 }
571
572 static void
573 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
574 {
575 LLVMTypeRef arg_types[7];
576 LLVMTypeRef func_type;
577 LLVMValueRef context_ptr;
578 LLVMBasicBlockRef block;
579 LLVMBuilderRef builder;
580 LLVMValueRef start, end, count, stride, step, io_itr;
581 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
582 struct draw_context *draw = llvm->draw;
583 unsigned i, j;
584 struct lp_build_context bld;
585 struct lp_build_loop_state lp_loop;
586 struct lp_type vs_type = lp_type_float_vec(32);
587 const int max_vertices = 4;
588 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
589
590 arg_types[0] = llvm->context_ptr_type; /* context */
591 arg_types[1] = llvm->vertex_header_ptr_type; /* vertex_header */
592 arg_types[2] = llvm->buffer_ptr_type; /* vbuffers */
593 arg_types[3] = LLVMInt32Type(); /* start */
594 arg_types[4] = LLVMInt32Type(); /* count */
595 arg_types[5] = LLVMInt32Type(); /* stride */
596 arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
597
598 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
599
600 variant->function = LLVMAddFunction(llvm->module, "draw_llvm_shader", func_type);
601 LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
602 for(i = 0; i < Elements(arg_types); ++i)
603 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
604 LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
605
606 context_ptr = LLVMGetParam(variant->function, 0);
607 io_ptr = LLVMGetParam(variant->function, 1);
608 vbuffers_ptr = LLVMGetParam(variant->function, 2);
609 start = LLVMGetParam(variant->function, 3);
610 count = LLVMGetParam(variant->function, 4);
611 stride = LLVMGetParam(variant->function, 5);
612 vb_ptr = LLVMGetParam(variant->function, 6);
613
614 lp_build_name(context_ptr, "context");
615 lp_build_name(io_ptr, "io");
616 lp_build_name(vbuffers_ptr, "vbuffers");
617 lp_build_name(start, "start");
618 lp_build_name(count, "count");
619 lp_build_name(stride, "stride");
620 lp_build_name(vb_ptr, "vb");
621
622 /*
623 * Function body
624 */
625
626 block = LLVMAppendBasicBlock(variant->function, "entry");
627 builder = LLVMCreateBuilder();
628 LLVMPositionBuilderAtEnd(builder, block);
629
630 lp_build_context_init(&bld, builder, vs_type);
631
632 end = lp_build_add(&bld, start, count);
633
634 step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
635
636 #if DEBUG_STORE
637 lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
638 start, end, step);
639 #endif
640 lp_build_loop_begin(builder, start, &lp_loop);
641 {
642 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
643 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
644 LLVMValueRef io;
645 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
646
647 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
648 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
649 #if DEBUG_STORE
650 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
651 io_itr, io, lp_loop.counter);
652 #endif
653 for (i = 0; i < NUM_CHANNELS; ++i) {
654 LLVMValueRef true_index = LLVMBuildAdd(
655 builder,
656 lp_loop.counter,
657 LLVMConstInt(LLVMInt32Type(), i, 0), "");
658 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
659 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
660 LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
661 velem->vertex_buffer_index,
662 0);
663 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
664 &vb_index, 1, "");
665 generate_fetch(builder, vbuffers_ptr,
666 &aos_attribs[j][i], velem, vb, true_index);
667 }
668 }
669 convert_to_soa(builder, aos_attribs, inputs,
670 draw->pt.nr_vertex_elements);
671
672 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
673 generate_vs(llvm,
674 builder,
675 outputs,
676 ptr_aos,
677 context_ptr);
678
679 convert_to_aos(builder, io, outputs,
680 draw->vs.vertex_shader->info.num_outputs,
681 max_vertices);
682 }
683 lp_build_loop_end_cond(builder, end, step, LLVMIntUGE, &lp_loop);
684
685 LLVMBuildRetVoid(builder);
686
687 LLVMDisposeBuilder(builder);
688
689 /*
690 * Translate the LLVM IR into machine code.
691 */
692 #ifdef DEBUG
693 if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
694 lp_debug_dump_value(variant->function);
695 assert(0);
696 }
697 #endif
698
699 LLVMRunFunctionPassManager(llvm->pass, variant->function);
700
701 if (0) {
702 lp_debug_dump_value(variant->function);
703 debug_printf("\n");
704 }
705 variant->jit_func = (draw_jit_vert_func)LLVMGetPointerToGlobal(llvm->draw->engine, variant->function);
706
707 if (0)
708 lp_disassemble(variant->jit_func);
709 }
710
711
712 static void
713 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
714 {
715 LLVMTypeRef arg_types[7];
716 LLVMTypeRef func_type;
717 LLVMValueRef context_ptr;
718 LLVMBasicBlockRef block;
719 LLVMBuilderRef builder;
720 LLVMValueRef fetch_elts, fetch_count, stride, step, io_itr;
721 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
722 struct draw_context *draw = llvm->draw;
723 unsigned i, j;
724 struct lp_build_context bld;
725 struct lp_build_context bld_int;
726 struct lp_build_loop_state lp_loop;
727 struct lp_type vs_type = lp_type_float_vec(32);
728 const int max_vertices = 4;
729 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
730 LLVMValueRef fetch_max;
731
732 arg_types[0] = llvm->context_ptr_type; /* context */
733 arg_types[1] = llvm->vertex_header_ptr_type; /* vertex_header */
734 arg_types[2] = llvm->buffer_ptr_type; /* vbuffers */
735 arg_types[3] = LLVMPointerType(LLVMInt32Type(), 0); /* fetch_elts * */
736 arg_types[4] = LLVMInt32Type(); /* fetch_count */
737 arg_types[5] = LLVMInt32Type(); /* stride */
738 arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
739
740 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
741
742 variant->function_elts = LLVMAddFunction(llvm->module, "draw_llvm_shader_elts", func_type);
743 LLVMSetFunctionCallConv(variant->function_elts, LLVMCCallConv);
744 for(i = 0; i < Elements(arg_types); ++i)
745 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
746 LLVMAddAttribute(LLVMGetParam(variant->function_elts, i), LLVMNoAliasAttribute);
747
748 context_ptr = LLVMGetParam(variant->function_elts, 0);
749 io_ptr = LLVMGetParam(variant->function_elts, 1);
750 vbuffers_ptr = LLVMGetParam(variant->function_elts, 2);
751 fetch_elts = LLVMGetParam(variant->function_elts, 3);
752 fetch_count = LLVMGetParam(variant->function_elts, 4);
753 stride = LLVMGetParam(variant->function_elts, 5);
754 vb_ptr = LLVMGetParam(variant->function_elts, 6);
755
756 lp_build_name(context_ptr, "context");
757 lp_build_name(io_ptr, "io");
758 lp_build_name(vbuffers_ptr, "vbuffers");
759 lp_build_name(fetch_elts, "fetch_elts");
760 lp_build_name(fetch_count, "fetch_count");
761 lp_build_name(stride, "stride");
762 lp_build_name(vb_ptr, "vb");
763
764 /*
765 * Function body
766 */
767
768 block = LLVMAppendBasicBlock(variant->function_elts, "entry");
769 builder = LLVMCreateBuilder();
770 LLVMPositionBuilderAtEnd(builder, block);
771
772 lp_build_context_init(&bld, builder, vs_type);
773 lp_build_context_init(&bld_int, builder, lp_type_int(32));
774
775 step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
776
777 fetch_max = LLVMBuildSub(builder, fetch_count,
778 LLVMConstInt(LLVMInt32Type(), 1, 0),
779 "fetch_max");
780
781 lp_build_loop_begin(builder, LLVMConstInt(LLVMInt32Type(), 0, 0), &lp_loop);
782 {
783 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
784 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
785 LLVMValueRef io;
786 const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
787
788 io_itr = lp_loop.counter;
789 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
790 #if DEBUG_STORE
791 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
792 io_itr, io, lp_loop.counter);
793 #endif
794 for (i = 0; i < NUM_CHANNELS; ++i) {
795 LLVMValueRef true_index = LLVMBuildAdd(
796 builder,
797 lp_loop.counter,
798 LLVMConstInt(LLVMInt32Type(), i, 0), "");
799 LLVMValueRef fetch_ptr;
800
801 /* make sure we're not out of bounds which can happen
802 * if fetch_count % 4 != 0, because on the last iteration
803 * a few of the 4 vertex fetches will be out of bounds */
804 true_index = lp_build_min(&bld_int, true_index, fetch_max);
805
806 fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
807 &true_index, 1, "");
808 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
809 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
810 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
811 LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
812 velem->vertex_buffer_index,
813 0);
814 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
815 &vb_index, 1, "");
816 generate_fetch(builder, vbuffers_ptr,
817 &aos_attribs[j][i], velem, vb, true_index);
818 }
819 }
820 convert_to_soa(builder, aos_attribs, inputs,
821 draw->pt.nr_vertex_elements);
822
823 ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
824 generate_vs(llvm,
825 builder,
826 outputs,
827 ptr_aos,
828 context_ptr);
829
830 convert_to_aos(builder, io, outputs,
831 draw->vs.vertex_shader->info.num_outputs,
832 max_vertices);
833 }
834 lp_build_loop_end_cond(builder, fetch_count, step, LLVMIntUGE, &lp_loop);
835
836 LLVMBuildRetVoid(builder);
837
838 LLVMDisposeBuilder(builder);
839
840 /*
841 * Translate the LLVM IR into machine code.
842 */
843 #ifdef DEBUG
844 if(LLVMVerifyFunction(variant->function_elts, LLVMPrintMessageAction)) {
845 lp_debug_dump_value(variant->function_elts);
846 assert(0);
847 }
848 #endif
849
850 LLVMRunFunctionPassManager(llvm->pass, variant->function_elts);
851
852 if (0) {
853 lp_debug_dump_value(variant->function_elts);
854 debug_printf("\n");
855 }
856 variant->jit_func_elts = (draw_jit_vert_func_elts)LLVMGetPointerToGlobal(
857 llvm->draw->engine, variant->function_elts);
858
859 if (0)
860 lp_disassemble(variant->jit_func_elts);
861 }
862
863 void
864 draw_llvm_make_variant_key(struct draw_llvm *llvm,
865 struct draw_llvm_variant_key *key)
866 {
867 memset(key, 0, sizeof(struct draw_llvm_variant_key));
868
869 key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
870
871 memcpy(key->vertex_element,
872 llvm->draw->pt.vertex_element,
873 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
874
875 memcpy(&key->vs,
876 &llvm->draw->vs.vertex_shader->state,
877 sizeof(struct pipe_shader_state));
878 }