projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: rename prefixes from radeon to si
[mesa.git] / src / gallium / drivers / radeonsi / si_shader_tgsi_setup.c
1 /*
2 * Copyright 2016 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 */
23
24 #include "si_shader_internal.h"
25 #include "radeon/radeon_elf_util.h"
26
27 #include "gallivm/lp_bld_const.h"
28 #include "gallivm/lp_bld_gather.h"
29 #include "gallivm/lp_bld_flow.h"
30 #include "gallivm/lp_bld_init.h"
31 #include "gallivm/lp_bld_intr.h"
32 #include "gallivm/lp_bld_misc.h"
33 #include "gallivm/lp_bld_swizzle.h"
34 #include "tgsi/tgsi_info.h"
35 #include "tgsi/tgsi_parse.h"
36 #include "util/u_math.h"
37 #include "util/u_memory.h"
38 #include "util/u_debug.h"
39
40 #include <stdio.h>
41 #include <llvm-c/Transforms/Scalar.h>
42
43 /* Data for if/else/endif and bgnloop/endloop control flow structures.
44 */
45 struct si_llvm_flow {
46 /* Loop exit or next part of if/else/endif. */
47 LLVMBasicBlockRef next_block;
48 LLVMBasicBlockRef loop_entry_block;
49 };
50
51 #define CPU_STRING_LEN 30
52 #define FS_STRING_LEN 30
53 #define TRIPLE_STRING_LEN 7
54
55 /**
56 * Shader types for the LLVM backend.
57 */
58 enum si_llvm_shader_type {
59 RADEON_LLVM_SHADER_PS = 0,
60 RADEON_LLVM_SHADER_VS = 1,
61 RADEON_LLVM_SHADER_GS = 2,
62 RADEON_LLVM_SHADER_CS = 3,
63 };
64
65 enum si_llvm_calling_convention {
66 RADEON_LLVM_AMDGPU_VS = 87,
67 RADEON_LLVM_AMDGPU_GS = 88,
68 RADEON_LLVM_AMDGPU_PS = 89,
69 RADEON_LLVM_AMDGPU_CS = 90,
70 };
71
72 void si_llvm_add_attribute(LLVMValueRef F, const char *name, int value)
73 {
74 char str[16];
75
76 snprintf(str, sizeof(str), "%i", value);
77 LLVMAddTargetDependentFunctionAttr(F, name, str);
78 }
79
80 /**
81 * Set the shader type we want to compile
82 *
83 * @param type shader type to set
84 */
85 void si_llvm_shader_type(LLVMValueRef F, unsigned type)
86 {
87 enum si_llvm_shader_type llvm_type;
88 enum si_llvm_calling_convention calling_conv;
89
90 switch (type) {
91 case PIPE_SHADER_VERTEX:
92 case PIPE_SHADER_TESS_CTRL:
93 case PIPE_SHADER_TESS_EVAL:
94 llvm_type = RADEON_LLVM_SHADER_VS;
95 calling_conv = RADEON_LLVM_AMDGPU_VS;
96 break;
97 case PIPE_SHADER_GEOMETRY:
98 llvm_type = RADEON_LLVM_SHADER_GS;
99 calling_conv = RADEON_LLVM_AMDGPU_GS;
100 break;
101 case PIPE_SHADER_FRAGMENT:
102 llvm_type = RADEON_LLVM_SHADER_PS;
103 calling_conv = RADEON_LLVM_AMDGPU_PS;
104 break;
105 case PIPE_SHADER_COMPUTE:
106 llvm_type = RADEON_LLVM_SHADER_CS;
107 calling_conv = RADEON_LLVM_AMDGPU_CS;
108 break;
109 default:
110 unreachable("Unhandle shader type");
111 }
112
113 if (HAVE_LLVM >= 0x309)
114 LLVMSetFunctionCallConv(F, calling_conv);
115 else
116 si_llvm_add_attribute(F, "ShaderType", llvm_type);
117 }
118
119 static void init_amdgpu_target()
120 {
121 gallivm_init_llvm_targets();
122 #if HAVE_LLVM < 0x0307
123 LLVMInitializeR600TargetInfo();
124 LLVMInitializeR600Target();
125 LLVMInitializeR600TargetMC();
126 LLVMInitializeR600AsmPrinter();
127 #else
128 LLVMInitializeAMDGPUTargetInfo();
129 LLVMInitializeAMDGPUTarget();
130 LLVMInitializeAMDGPUTargetMC();
131 LLVMInitializeAMDGPUAsmPrinter();
132
133 #endif
134 }
135
136 static once_flag init_amdgpu_target_once_flag = ONCE_FLAG_INIT;
137
138 LLVMTargetRef si_llvm_get_amdgpu_target(const char *triple)
139 {
140 LLVMTargetRef target = NULL;
141 char *err_message = NULL;
142
143 call_once(&init_amdgpu_target_once_flag, init_amdgpu_target);
144
145 if (LLVMGetTargetFromTriple(triple, &target, &err_message)) {
146 fprintf(stderr, "Cannot find target for triple %s ", triple);
147 if (err_message) {
148 fprintf(stderr, "%s\n", err_message);
149 }
150 LLVMDisposeMessage(err_message);
151 return NULL;
152 }
153 return target;
154 }
155
156 struct si_llvm_diagnostics {
157 struct pipe_debug_callback *debug;
158 unsigned retval;
159 };
160
161 static void si_diagnostic_handler(LLVMDiagnosticInfoRef di, void *context)
162 {
163 struct si_llvm_diagnostics *diag = (struct si_llvm_diagnostics *)context;
164 LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
165 char *description = LLVMGetDiagInfoDescription(di);
166 const char *severity_str = NULL;
167
168 switch (severity) {
169 case LLVMDSError:
170 severity_str = "error";
171 break;
172 case LLVMDSWarning:
173 severity_str = "warning";
174 break;
175 case LLVMDSRemark:
176 severity_str = "remark";
177 break;
178 case LLVMDSNote:
179 severity_str = "note";
180 break;
181 default:
182 severity_str = "unknown";
183 }
184
185 pipe_debug_message(diag->debug, SHADER_INFO,
186 "LLVM diagnostic (%s): %s", severity_str, description);
187
188 if (severity == LLVMDSError) {
189 diag->retval = 1;
190 fprintf(stderr,"LLVM triggered Diagnostic Handler: %s\n", description);
191 }
192
193 LLVMDisposeMessage(description);
194 }
195
196 /**
197 * Compile an LLVM module to machine code.
198 *
199 * @returns 0 for success, 1 for failure
200 */
201 unsigned si_llvm_compile(LLVMModuleRef M, struct radeon_shader_binary *binary,
202 LLVMTargetMachineRef tm,
203 struct pipe_debug_callback *debug)
204 {
205 struct si_llvm_diagnostics diag;
206 char *err;
207 LLVMContextRef llvm_ctx;
208 LLVMMemoryBufferRef out_buffer;
209 unsigned buffer_size;
210 const char *buffer_data;
211 LLVMBool mem_err;
212
213 diag.debug = debug;
214 diag.retval = 0;
215
216 /* Setup Diagnostic Handler*/
217 llvm_ctx = LLVMGetModuleContext(M);
218
219 LLVMContextSetDiagnosticHandler(llvm_ctx, si_diagnostic_handler, &diag);
220
221 /* Compile IR*/
222 mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile, &err,
223 &out_buffer);
224
225 /* Process Errors/Warnings */
226 if (mem_err) {
227 fprintf(stderr, "%s: %s", __FUNCTION__, err);
228 pipe_debug_message(debug, SHADER_INFO,
229 "LLVM emit error: %s", err);
230 FREE(err);
231 diag.retval = 1;
232 goto out;
233 }
234
235 /* Extract Shader Code*/
236 buffer_size = LLVMGetBufferSize(out_buffer);
237 buffer_data = LLVMGetBufferStart(out_buffer);
238
239 radeon_elf_read(buffer_data, buffer_size, binary);
240
241 /* Clean up */
242 LLVMDisposeMemoryBuffer(out_buffer);
243
244 out:
245 if (diag.retval != 0)
246 pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed");
247 return diag.retval;
248 }
249
250 LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
251 enum tgsi_opcode_type type)
252 {
253 LLVMContextRef ctx = bld_base->base.gallivm->context;
254
255 switch (type) {
256 case TGSI_TYPE_UNSIGNED:
257 case TGSI_TYPE_SIGNED:
258 return LLVMInt32TypeInContext(ctx);
259 case TGSI_TYPE_UNSIGNED64:
260 case TGSI_TYPE_SIGNED64:
261 return LLVMInt64TypeInContext(ctx);
262 case TGSI_TYPE_DOUBLE:
263 return LLVMDoubleTypeInContext(ctx);
264 case TGSI_TYPE_UNTYPED:
265 case TGSI_TYPE_FLOAT:
266 return LLVMFloatTypeInContext(ctx);
267 default: break;
268 }
269 return 0;
270 }
271
272 LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
273 enum tgsi_opcode_type type, LLVMValueRef value)
274 {
275 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
276 LLVMTypeRef dst_type = tgsi2llvmtype(bld_base, type);
277
278 if (dst_type)
279 return LLVMBuildBitCast(builder, value, dst_type, "");
280 else
281 return value;
282 }
283
284 /**
285 * Return a value that is equal to the given i32 \p index if it lies in [0,num)
286 * or an undefined value in the same interval otherwise.
287 */
288 LLVMValueRef si_llvm_bound_index(struct si_shader_context *ctx,
289 LLVMValueRef index,
290 unsigned num)
291 {
292 struct gallivm_state *gallivm = &ctx->gallivm;
293 LLVMBuilderRef builder = gallivm->builder;
294 LLVMValueRef c_max = lp_build_const_int32(gallivm, num - 1);
295 LLVMValueRef cc;
296
297 if (util_is_power_of_two(num)) {
298 index = LLVMBuildAnd(builder, index, c_max, "");
299 } else {
300 /* In theory, this MAX pattern should result in code that is
301 * as good as the bit-wise AND above.
302 *
303 * In practice, LLVM generates worse code (at the time of
304 * writing), because its value tracking is not strong enough.
305 */
306 cc = LLVMBuildICmp(builder, LLVMIntULE, index, c_max, "");
307 index = LLVMBuildSelect(builder, cc, index, c_max, "");
308 }
309
310 return index;
311 }
312
313 static struct si_llvm_flow *
314 get_current_flow(struct si_shader_context *ctx)
315 {
316 if (ctx->flow_depth > 0)
317 return &ctx->flow[ctx->flow_depth - 1];
318 return NULL;
319 }
320
321 static struct si_llvm_flow *
322 get_innermost_loop(struct si_shader_context *ctx)
323 {
324 for (unsigned i = ctx->flow_depth; i > 0; --i) {
325 if (ctx->flow[i - 1].loop_entry_block)
326 return &ctx->flow[i - 1];
327 }
328 return NULL;
329 }
330
331 static struct si_llvm_flow *
332 push_flow(struct si_shader_context *ctx)
333 {
334 struct si_llvm_flow *flow;
335
336 if (ctx->flow_depth >= ctx->flow_depth_max) {
337 unsigned new_max = MAX2(ctx->flow_depth << 1, RADEON_LLVM_INITIAL_CF_DEPTH);
338 ctx->flow = REALLOC(ctx->flow,
339 ctx->flow_depth_max * sizeof(*ctx->flow),
340 new_max * sizeof(*ctx->flow));
341 ctx->flow_depth_max = new_max;
342 }
343
344 flow = &ctx->flow[ctx->flow_depth];
345 ctx->flow_depth++;
346
347 flow->next_block = NULL;
348 flow->loop_entry_block = NULL;
349 return flow;
350 }
351
352 static LLVMValueRef emit_swizzle(struct lp_build_tgsi_context *bld_base,
353 LLVMValueRef value,
354 unsigned swizzle_x,
355 unsigned swizzle_y,
356 unsigned swizzle_z,
357 unsigned swizzle_w)
358 {
359 LLVMValueRef swizzles[4];
360 LLVMTypeRef i32t =
361 LLVMInt32TypeInContext(bld_base->base.gallivm->context);
362
363 swizzles[0] = LLVMConstInt(i32t, swizzle_x, 0);
364 swizzles[1] = LLVMConstInt(i32t, swizzle_y, 0);
365 swizzles[2] = LLVMConstInt(i32t, swizzle_z, 0);
366 swizzles[3] = LLVMConstInt(i32t, swizzle_w, 0);
367
368 return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
369 value,
370 LLVMGetUndef(LLVMTypeOf(value)),
371 LLVMConstVector(swizzles, 4), "");
372 }
373
374 /**
375 * Return the description of the array covering the given temporary register
376 * index.
377 */
378 static unsigned
379 get_temp_array_id(struct lp_build_tgsi_context *bld_base,
380 unsigned reg_index,
381 const struct tgsi_ind_register *reg)
382 {
383 struct si_shader_context *ctx = si_shader_context(bld_base);
384 unsigned num_arrays = ctx->soa.bld_base.info->array_max[TGSI_FILE_TEMPORARY];
385 unsigned i;
386
387 if (reg && reg->ArrayID > 0 && reg->ArrayID <= num_arrays)
388 return reg->ArrayID;
389
390 for (i = 0; i < num_arrays; i++) {
391 const struct tgsi_array_info *array = &ctx->temp_arrays[i];
392
393 if (reg_index >= array->range.First && reg_index <= array->range.Last)
394 return i + 1;
395 }
396
397 return 0;
398 }
399
400 static struct tgsi_declaration_range
401 get_array_range(struct lp_build_tgsi_context *bld_base,
402 unsigned File, unsigned reg_index,
403 const struct tgsi_ind_register *reg)
404 {
405 struct si_shader_context *ctx = si_shader_context(bld_base);
406 struct tgsi_declaration_range range;
407
408 if (File == TGSI_FILE_TEMPORARY) {
409 unsigned array_id = get_temp_array_id(bld_base, reg_index, reg);
410 if (array_id)
411 return ctx->temp_arrays[array_id - 1].range;
412 }
413
414 range.First = 0;
415 range.Last = bld_base->info->file_max[File];
416 return range;
417 }
418
419 static LLVMValueRef
420 emit_array_index(struct lp_build_tgsi_soa_context *bld,
421 const struct tgsi_ind_register *reg,
422 unsigned offset)
423 {
424 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
425
426 if (!reg) {
427 return lp_build_const_int32(gallivm, offset);
428 }
429 LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, bld->addr[reg->Index][reg->Swizzle], "");
430 return LLVMBuildAdd(gallivm->builder, addr, lp_build_const_int32(gallivm, offset), "");
431 }
432
433 /**
434 * For indirect registers, construct a pointer directly to the requested
435 * element using getelementptr if possible.
436 *
437 * Returns NULL if the insertelement/extractelement fallback for array access
438 * must be used.
439 */
440 static LLVMValueRef
441 get_pointer_into_array(struct si_shader_context *ctx,
442 unsigned file,
443 unsigned swizzle,
444 unsigned reg_index,
445 const struct tgsi_ind_register *reg_indirect)
446 {
447 unsigned array_id;
448 struct tgsi_array_info *array;
449 struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
450 LLVMBuilderRef builder = gallivm->builder;
451 LLVMValueRef idxs[2];
452 LLVMValueRef index;
453 LLVMValueRef alloca;
454
455 if (file != TGSI_FILE_TEMPORARY)
456 return NULL;
457
458 array_id = get_temp_array_id(&ctx->soa.bld_base, reg_index, reg_indirect);
459 if (!array_id)
460 return NULL;
461
462 alloca = ctx->temp_array_allocas[array_id - 1];
463 if (!alloca)
464 return NULL;
465
466 array = &ctx->temp_arrays[array_id - 1];
467
468 if (!(array->writemask & (1 << swizzle)))
469 return ctx->undef_alloca;
470
471 index = emit_array_index(&ctx->soa, reg_indirect,
472 reg_index - ctx->temp_arrays[array_id - 1].range.First);
473
474 /* Ensure that the index is within a valid range, to guard against
475 * VM faults and overwriting critical data (e.g. spilled resource
476 * descriptors).
477 *
478 * TODO It should be possible to avoid the additional instructions
479 * if LLVM is changed so that it guarantuees:
480 * 1. the scratch space descriptor isolates the current wave (this
481 * could even save the scratch offset SGPR at the cost of an
482 * additional SALU instruction)
483 * 2. the memory for allocas must be allocated at the _end_ of the
484 * scratch space (after spilled registers)
485 */
486 index = si_llvm_bound_index(ctx, index, array->range.Last - array->range.First + 1);
487
488 index = LLVMBuildMul(
489 builder, index,
490 lp_build_const_int32(gallivm, util_bitcount(array->writemask)),
491 "");
492 index = LLVMBuildAdd(
493 builder, index,
494 lp_build_const_int32(
495 gallivm,
496 util_bitcount(array->writemask & ((1 << swizzle) - 1))),
497 "");
498 idxs[0] = ctx->soa.bld_base.uint_bld.zero;
499 idxs[1] = index;
500 return LLVMBuildGEP(builder, alloca, idxs, 2, "");
501 }
502
503 LLVMValueRef
504 si_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
505 enum tgsi_opcode_type type,
506 LLVMValueRef ptr,
507 LLVMValueRef ptr2)
508 {
509 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
510 LLVMValueRef result;
511
512 result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
513
514 result = LLVMBuildInsertElement(builder,
515 result,
516 bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr),
517 bld_base->int_bld.zero, "");
518 result = LLVMBuildInsertElement(builder,
519 result,
520 bitcast(bld_base, TGSI_TYPE_UNSIGNED, ptr2),
521 bld_base->int_bld.one, "");
522 return bitcast(bld_base, type, result);
523 }
524
525 static LLVMValueRef
526 emit_array_fetch(struct lp_build_tgsi_context *bld_base,
527 unsigned File, enum tgsi_opcode_type type,
528 struct tgsi_declaration_range range,
529 unsigned swizzle)
530 {
531 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
532 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
533 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
534
535 unsigned i, size = range.Last - range.First + 1;
536 LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
537 LLVMValueRef result = LLVMGetUndef(vec);
538
539 struct tgsi_full_src_register tmp_reg = {};
540 tmp_reg.Register.File = File;
541
542 for (i = 0; i < size; ++i) {
543 tmp_reg.Register.Index = i + range.First;
544 LLVMValueRef temp = si_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
545 result = LLVMBuildInsertElement(builder, result, temp,
546 lp_build_const_int32(gallivm, i), "array_vector");
547 }
548 return result;
549 }
550
551 static LLVMValueRef
552 load_value_from_array(struct lp_build_tgsi_context *bld_base,
553 unsigned file,
554 enum tgsi_opcode_type type,
555 unsigned swizzle,
556 unsigned reg_index,
557 const struct tgsi_ind_register *reg_indirect)
558 {
559 struct si_shader_context *ctx = si_shader_context(bld_base);
560 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
561 struct gallivm_state *gallivm = bld_base->base.gallivm;
562 LLVMBuilderRef builder = gallivm->builder;
563 LLVMValueRef ptr;
564
565 ptr = get_pointer_into_array(ctx, file, swizzle, reg_index, reg_indirect);
566 if (ptr) {
567 LLVMValueRef val = LLVMBuildLoad(builder, ptr, "");
568 if (tgsi_type_is_64bit(type)) {
569 LLVMValueRef ptr_hi, val_hi;
570 ptr_hi = LLVMBuildGEP(builder, ptr, &bld_base->uint_bld.one, 1, "");
571 val_hi = LLVMBuildLoad(builder, ptr_hi, "");
572 val = si_llvm_emit_fetch_64bit(bld_base, type, val, val_hi);
573 }
574
575 return val;
576 } else {
577 struct tgsi_declaration_range range =
578 get_array_range(bld_base, file, reg_index, reg_indirect);
579 LLVMValueRef index =
580 emit_array_index(bld, reg_indirect, reg_index - range.First);
581 LLVMValueRef array =
582 emit_array_fetch(bld_base, file, type, range, swizzle);
583 return LLVMBuildExtractElement(builder, array, index, "");
584 }
585 }
586
587 static void
588 store_value_to_array(struct lp_build_tgsi_context *bld_base,
589 LLVMValueRef value,
590 unsigned file,
591 unsigned chan_index,
592 unsigned reg_index,
593 const struct tgsi_ind_register *reg_indirect)
594 {
595 struct si_shader_context *ctx = si_shader_context(bld_base);
596 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
597 struct gallivm_state *gallivm = bld_base->base.gallivm;
598 LLVMBuilderRef builder = gallivm->builder;
599 LLVMValueRef ptr;
600
601 ptr = get_pointer_into_array(ctx, file, chan_index, reg_index, reg_indirect);
602 if (ptr) {
603 LLVMBuildStore(builder, value, ptr);
604 } else {
605 unsigned i, size;
606 struct tgsi_declaration_range range = get_array_range(bld_base, file, reg_index, reg_indirect);
607 LLVMValueRef index = emit_array_index(bld, reg_indirect, reg_index - range.First);
608 LLVMValueRef array =
609 emit_array_fetch(bld_base, file, TGSI_TYPE_FLOAT, range, chan_index);
610 LLVMValueRef temp_ptr;
611
612 array = LLVMBuildInsertElement(builder, array, value, index, "");
613
614 size = range.Last - range.First + 1;
615 for (i = 0; i < size; ++i) {
616 switch(file) {
617 case TGSI_FILE_OUTPUT:
618 temp_ptr = bld->outputs[i + range.First][chan_index];
619 break;
620
621 case TGSI_FILE_TEMPORARY:
622 if (range.First + i >= ctx->temps_count)
623 continue;
624 temp_ptr = ctx->temps[(i + range.First) * TGSI_NUM_CHANNELS + chan_index];
625 break;
626
627 default:
628 continue;
629 }
630 value = LLVMBuildExtractElement(builder, array,
631 lp_build_const_int32(gallivm, i), "");
632 LLVMBuildStore(builder, value, temp_ptr);
633 }
634 }
635 }
636
637 LLVMValueRef si_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
638 const struct tgsi_full_src_register *reg,
639 enum tgsi_opcode_type type,
640 unsigned swizzle)
641 {
642 struct si_shader_context *ctx = si_shader_context(bld_base);
643 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
644 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
645 LLVMValueRef result = NULL, ptr, ptr2;
646
647 if (swizzle == ~0) {
648 LLVMValueRef values[TGSI_NUM_CHANNELS];
649 unsigned chan;
650 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
651 values[chan] = si_llvm_emit_fetch(bld_base, reg, type, chan);
652 }
653 return lp_build_gather_values(bld_base->base.gallivm, values,
654 TGSI_NUM_CHANNELS);
655 }
656
657 if (reg->Register.Indirect) {
658 LLVMValueRef load = load_value_from_array(bld_base, reg->Register.File, type,
659 swizzle, reg->Register.Index, &reg->Indirect);
660 return bitcast(bld_base, type, load);
661 }
662
663 switch(reg->Register.File) {
664 case TGSI_FILE_IMMEDIATE: {
665 LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
666 if (tgsi_type_is_64bit(type)) {
667 result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
668 result = LLVMConstInsertElement(result,
669 bld->immediates[reg->Register.Index][swizzle],
670 bld_base->int_bld.zero);
671 result = LLVMConstInsertElement(result,
672 bld->immediates[reg->Register.Index][swizzle + 1],
673 bld_base->int_bld.one);
674 return LLVMConstBitCast(result, ctype);
675 } else {
676 return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
677 }
678 }
679
680 case TGSI_FILE_INPUT: {
681 unsigned index = reg->Register.Index;
682 LLVMValueRef input[4];
683
684 /* I don't think doing this for vertex shaders is beneficial.
685 * For those, we want to make sure the VMEM loads are executed
686 * only once. Fragment shaders don't care much, because
687 * v_interp instructions are much cheaper than VMEM loads.
688 */
689 if (ctx->soa.bld_base.info->processor == PIPE_SHADER_FRAGMENT)
690 ctx->load_input(ctx, index, &ctx->input_decls[index], input);
691 else
692 memcpy(input, &ctx->inputs[index * 4], sizeof(input));
693
694 result = input[swizzle];
695
696 if (tgsi_type_is_64bit(type)) {
697 ptr = result;
698 ptr2 = input[swizzle + 1];
699 return si_llvm_emit_fetch_64bit(bld_base, type, ptr, ptr2);
700 }
701 break;
702 }
703
704 case TGSI_FILE_TEMPORARY:
705 if (reg->Register.Index >= ctx->temps_count)
706 return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
707 ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
708 if (tgsi_type_is_64bit(type)) {
709 ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1];
710 return si_llvm_emit_fetch_64bit(bld_base, type,
711 LLVMBuildLoad(builder, ptr, ""),
712 LLVMBuildLoad(builder, ptr2, ""));
713 }
714 result = LLVMBuildLoad(builder, ptr, "");
715 break;
716
717 case TGSI_FILE_OUTPUT:
718 ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
719 if (tgsi_type_is_64bit(type)) {
720 ptr2 = lp_get_output_ptr(bld, reg->Register.Index, swizzle + 1);
721 return si_llvm_emit_fetch_64bit(bld_base, type,
722 LLVMBuildLoad(builder, ptr, ""),
723 LLVMBuildLoad(builder, ptr2, ""));
724 }
725 result = LLVMBuildLoad(builder, ptr, "");
726 break;
727
728 default:
729 return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
730 }
731
732 return bitcast(bld_base, type, result);
733 }
734
735 static LLVMValueRef fetch_system_value(struct lp_build_tgsi_context *bld_base,
736 const struct tgsi_full_src_register *reg,
737 enum tgsi_opcode_type type,
738 unsigned swizzle)
739 {
740 struct si_shader_context *ctx = si_shader_context(bld_base);
741 struct gallivm_state *gallivm = bld_base->base.gallivm;
742
743 LLVMValueRef cval = ctx->system_values[reg->Register.Index];
744 if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
745 cval = LLVMBuildExtractElement(gallivm->builder, cval,
746 lp_build_const_int32(gallivm, swizzle), "");
747 }
748 return bitcast(bld_base, type, cval);
749 }
750
751 static void emit_declaration(struct lp_build_tgsi_context *bld_base,
752 const struct tgsi_full_declaration *decl)
753 {
754 struct si_shader_context *ctx = si_shader_context(bld_base);
755 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
756 unsigned first, last, i;
757 switch(decl->Declaration.File) {
758 case TGSI_FILE_ADDRESS:
759 {
760 unsigned idx;
761 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
762 unsigned chan;
763 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
764 ctx->soa.addr[idx][chan] = lp_build_alloca_undef(
765 &ctx->gallivm,
766 ctx->soa.bld_base.uint_bld.elem_type, "");
767 }
768 }
769 break;
770 }
771
772 case TGSI_FILE_TEMPORARY:
773 {
774 char name[16] = "";
775 LLVMValueRef array_alloca = NULL;
776 unsigned decl_size;
777 unsigned writemask = decl->Declaration.UsageMask;
778 first = decl->Range.First;
779 last = decl->Range.Last;
780 decl_size = 4 * ((last - first) + 1);
781
782 if (decl->Declaration.Array) {
783 unsigned id = decl->Array.ArrayID - 1;
784 unsigned array_size;
785
786 writemask &= ctx->temp_arrays[id].writemask;
787 ctx->temp_arrays[id].writemask = writemask;
788 array_size = ((last - first) + 1) * util_bitcount(writemask);
789
790 /* If the array has more than 16 elements, store it
791 * in memory using an alloca that spans the entire
792 * array.
793 *
794 * Otherwise, store each array element individually.
795 * We will then generate vectors (per-channel, up to
796 * <16 x float> if the usagemask is a single bit) for
797 * indirect addressing.
798 *
799 * Note that 16 is the number of vector elements that
800 * LLVM will store in a register, so theoretically an
801 * array with up to 4 * 16 = 64 elements could be
802 * handled this way, but whether that's a good idea
803 * depends on VGPR register pressure elsewhere.
804 *
805 * FIXME: We shouldn't need to have the non-alloca
806 * code path for arrays. LLVM should be smart enough to
807 * promote allocas into registers when profitable.
808 *
809 * LLVM 3.8 crashes with this.
810 */
811 if (HAVE_LLVM >= 0x0309 && array_size > 16) {
812 array_alloca = LLVMBuildAlloca(builder,
813 LLVMArrayType(bld_base->base.vec_type,
814 array_size), "array");
815 ctx->temp_array_allocas[id] = array_alloca;
816 }
817 }
818
819 if (!ctx->temps_count) {
820 ctx->temps_count = bld_base->info->file_max[TGSI_FILE_TEMPORARY] + 1;
821 ctx->temps = MALLOC(TGSI_NUM_CHANNELS * ctx->temps_count * sizeof(LLVMValueRef));
822 }
823 if (!array_alloca) {
824 for (i = 0; i < decl_size; ++i) {
825 #ifdef DEBUG
826 snprintf(name, sizeof(name), "TEMP%d.%c",
827 first + i / 4, "xyzw"[i % 4]);
828 #endif
829 ctx->temps[first * TGSI_NUM_CHANNELS + i] =
830 lp_build_alloca_undef(bld_base->base.gallivm,
831 bld_base->base.vec_type,
832 name);
833 }
834 } else {
835 LLVMValueRef idxs[2] = {
836 bld_base->uint_bld.zero,
837 NULL
838 };
839 unsigned j = 0;
840
841 if (writemask != TGSI_WRITEMASK_XYZW &&
842 !ctx->undef_alloca) {
843 /* Create a dummy alloca. We use it so that we
844 * have a pointer that is safe to load from if
845 * a shader ever reads from a channel that
846 * it never writes to.
847 */
848 ctx->undef_alloca = lp_build_alloca_undef(
849 bld_base->base.gallivm,
850 bld_base->base.vec_type, "undef");
851 }
852
853 for (i = 0; i < decl_size; ++i) {
854 LLVMValueRef ptr;
855 if (writemask & (1 << (i % 4))) {
856 #ifdef DEBUG
857 snprintf(name, sizeof(name), "TEMP%d.%c",
858 first + i / 4, "xyzw"[i % 4]);
859 #endif
860 idxs[1] = lp_build_const_int32(bld_base->base.gallivm, j);
861 ptr = LLVMBuildGEP(builder, array_alloca, idxs, 2, name);
862 j++;
863 } else {
864 ptr = ctx->undef_alloca;
865 }
866 ctx->temps[first * TGSI_NUM_CHANNELS + i] = ptr;
867 }
868 }
869 break;
870 }
871 case TGSI_FILE_INPUT:
872 {
873 unsigned idx;
874 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
875 if (ctx->load_input &&
876 ctx->input_decls[idx].Declaration.File != TGSI_FILE_INPUT) {
877 ctx->input_decls[idx] = *decl;
878
879 if (bld_base->info->processor != PIPE_SHADER_FRAGMENT)
880 ctx->load_input(ctx, idx, decl,
881 &ctx->inputs[idx * 4]);
882 }
883 }
884 }
885 break;
886
887 case TGSI_FILE_SYSTEM_VALUE:
888 {
889 unsigned idx;
890 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
891 ctx->load_system_value(ctx, idx, decl);
892 }
893 }
894 break;
895
896 case TGSI_FILE_OUTPUT:
897 {
898 char name[16] = "";
899 unsigned idx;
900 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
901 unsigned chan;
902 assert(idx < RADEON_LLVM_MAX_OUTPUTS);
903 if (ctx->soa.outputs[idx][0])
904 continue;
905 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
906 #ifdef DEBUG
907 snprintf(name, sizeof(name), "OUT%d.%c",
908 idx, "xyzw"[chan % 4]);
909 #endif
910 ctx->soa.outputs[idx][chan] = lp_build_alloca_undef(
911 &ctx->gallivm,
912 ctx->soa.bld_base.base.elem_type, name);
913 }
914 }
915 break;
916 }
917
918 case TGSI_FILE_MEMORY:
919 ctx->declare_memory_region(ctx, decl);
920 break;
921
922 default:
923 break;
924 }
925 }
926
927 LLVMValueRef si_llvm_saturate(struct lp_build_tgsi_context *bld_base,
928 LLVMValueRef value)
929 {
930 struct lp_build_emit_data clamp_emit_data;
931
932 memset(&clamp_emit_data, 0, sizeof(clamp_emit_data));
933 clamp_emit_data.arg_count = 3;
934 clamp_emit_data.args[0] = value;
935 clamp_emit_data.args[2] = bld_base->base.one;
936 clamp_emit_data.args[1] = bld_base->base.zero;
937
938 return lp_build_emit_llvm(bld_base, TGSI_OPCODE_CLAMP,
939 &clamp_emit_data);
940 }
941
942 void si_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
943 const struct tgsi_full_instruction *inst,
944 const struct tgsi_opcode_info *info,
945 LLVMValueRef dst[4])
946 {
947 struct si_shader_context *ctx = si_shader_context(bld_base);
948 struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
949 struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
950 const struct tgsi_full_dst_register *reg = &inst->Dst[0];
951 LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
952 LLVMValueRef temp_ptr, temp_ptr2 = NULL;
953 unsigned chan, chan_index;
954 bool is_vec_store = false;
955 enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode);
956
957 if (dst[0]) {
958 LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
959 is_vec_store = (k == LLVMVectorTypeKind);
960 }
961
962 if (is_vec_store) {
963 LLVMValueRef values[4] = {};
964 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan) {
965 LLVMValueRef index = lp_build_const_int32(gallivm, chan);
966 values[chan] = LLVMBuildExtractElement(gallivm->builder,
967 dst[0], index, "");
968 }
969 bld_base->emit_store(bld_base, inst, info, values);
970 return;
971 }
972
973 TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( inst, chan_index ) {
974 LLVMValueRef value = dst[chan_index];
975
976 if (tgsi_type_is_64bit(dtype) && (chan_index == 1 || chan_index == 3))
977 continue;
978 if (inst->Instruction.Saturate)
979 value = si_llvm_saturate(bld_base, value);
980
981 if (reg->Register.File == TGSI_FILE_ADDRESS) {
982 temp_ptr = bld->addr[reg->Register.Index][chan_index];
983 LLVMBuildStore(builder, value, temp_ptr);
984 continue;
985 }
986
987 if (!tgsi_type_is_64bit(dtype))
988 value = bitcast(bld_base, TGSI_TYPE_FLOAT, value);
989
990 if (reg->Register.Indirect) {
991 unsigned file = reg->Register.File;
992 unsigned reg_index = reg->Register.Index;
993 store_value_to_array(bld_base, value, file, chan_index,
994 reg_index, &reg->Indirect);
995 } else {
996 switch(reg->Register.File) {
997 case TGSI_FILE_OUTPUT:
998 temp_ptr = bld->outputs[reg->Register.Index][chan_index];
999 if (tgsi_type_is_64bit(dtype))
1000 temp_ptr2 = bld->outputs[reg->Register.Index][chan_index + 1];
1001 break;
1002
1003 case TGSI_FILE_TEMPORARY:
1004 {
1005 if (reg->Register.Index >= ctx->temps_count)
1006 continue;
1007
1008 temp_ptr = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index];
1009 if (tgsi_type_is_64bit(dtype))
1010 temp_ptr2 = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index + 1];
1011
1012 break;
1013 }
1014 default:
1015 return;
1016 }
1017 if (!tgsi_type_is_64bit(dtype))
1018 LLVMBuildStore(builder, value, temp_ptr);
1019 else {
1020 LLVMValueRef ptr = LLVMBuildBitCast(builder, value,
1021 LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), 2), "");
1022 LLVMValueRef val2;
1023 value = LLVMBuildExtractElement(builder, ptr,
1024 bld_base->uint_bld.zero, "");
1025 val2 = LLVMBuildExtractElement(builder, ptr,
1026 bld_base->uint_bld.one, "");
1027
1028 LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, value), temp_ptr);
1029 LLVMBuildStore(builder, bitcast(bld_base, TGSI_TYPE_FLOAT, val2), temp_ptr2);
1030 }
1031 }
1032 }
1033 }
1034
1035 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base, int pc)
1036 {
1037 char buf[32];
1038 /* Subtract 1 so that the number shown is that of the corresponding
1039 * opcode in the TGSI dump, e.g. an if block has the same suffix as
1040 * the instruction number of the corresponding TGSI IF.
1041 */
1042 snprintf(buf, sizeof(buf), "%s%d", base, pc - 1);
1043 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
1044 }
1045
1046 /* Append a basic block at the level of the parent flow.
1047 */
1048 static LLVMBasicBlockRef append_basic_block(struct si_shader_context *ctx,
1049 const char *name)
1050 {
1051 struct gallivm_state *gallivm = &ctx->gallivm;
1052
1053 assert(ctx->flow_depth >= 1);
1054
1055 if (ctx->flow_depth >= 2) {
1056 struct si_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
1057
1058 return LLVMInsertBasicBlockInContext(gallivm->context,
1059 flow->next_block, name);
1060 }
1061
1062 return LLVMAppendBasicBlockInContext(gallivm->context, ctx->main_fn, name);
1063 }
1064
1065 /* Emit a branch to the given default target for the current block if
1066 * applicable -- that is, if the current block does not already contain a
1067 * branch from a break or continue.
1068 */
1069 static void emit_default_branch(LLVMBuilderRef builder, LLVMBasicBlockRef target)
1070 {
1071 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
1072 LLVMBuildBr(builder, target);
1073 }
1074
1075 static void bgnloop_emit(const struct lp_build_tgsi_action *action,
1076 struct lp_build_tgsi_context *bld_base,
1077 struct lp_build_emit_data *emit_data)
1078 {
1079 struct si_shader_context *ctx = si_shader_context(bld_base);
1080 struct gallivm_state *gallivm = bld_base->base.gallivm;
1081 struct si_llvm_flow *flow = push_flow(ctx);
1082 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
1083 flow->next_block = append_basic_block(ctx, "ENDLOOP");
1084 set_basicblock_name(flow->loop_entry_block, "loop", bld_base->pc);
1085 LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
1086 LLVMPositionBuilderAtEnd(gallivm->builder, flow->loop_entry_block);
1087 }
1088
1089 static void brk_emit(const struct lp_build_tgsi_action *action,
1090 struct lp_build_tgsi_context *bld_base,
1091 struct lp_build_emit_data *emit_data)
1092 {
1093 struct si_shader_context *ctx = si_shader_context(bld_base);
1094 struct gallivm_state *gallivm = bld_base->base.gallivm;
1095 struct si_llvm_flow *flow = get_innermost_loop(ctx);
1096
1097 LLVMBuildBr(gallivm->builder, flow->next_block);
1098 }
1099
1100 static void cont_emit(const struct lp_build_tgsi_action *action,
1101 struct lp_build_tgsi_context *bld_base,
1102 struct lp_build_emit_data *emit_data)
1103 {
1104 struct si_shader_context *ctx = si_shader_context(bld_base);
1105 struct gallivm_state *gallivm = bld_base->base.gallivm;
1106 struct si_llvm_flow *flow = get_innermost_loop(ctx);
1107
1108 LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
1109 }
1110
1111 static void else_emit(const struct lp_build_tgsi_action *action,
1112 struct lp_build_tgsi_context *bld_base,
1113 struct lp_build_emit_data *emit_data)
1114 {
1115 struct si_shader_context *ctx = si_shader_context(bld_base);
1116 struct gallivm_state *gallivm = bld_base->base.gallivm;
1117 struct si_llvm_flow *current_branch = get_current_flow(ctx);
1118 LLVMBasicBlockRef endif_block;
1119
1120 assert(!current_branch->loop_entry_block);
1121
1122 endif_block = append_basic_block(ctx, "ENDIF");
1123 emit_default_branch(gallivm->builder, endif_block);
1124
1125 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
1126 set_basicblock_name(current_branch->next_block, "else", bld_base->pc);
1127
1128 current_branch->next_block = endif_block;
1129 }
1130
1131 static void endif_emit(const struct lp_build_tgsi_action *action,
1132 struct lp_build_tgsi_context *bld_base,
1133 struct lp_build_emit_data *emit_data)
1134 {
1135 struct si_shader_context *ctx = si_shader_context(bld_base);
1136 struct gallivm_state *gallivm = bld_base->base.gallivm;
1137 struct si_llvm_flow *current_branch = get_current_flow(ctx);
1138
1139 assert(!current_branch->loop_entry_block);
1140
1141 emit_default_branch(gallivm->builder, current_branch->next_block);
1142 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
1143 set_basicblock_name(current_branch->next_block, "endif", bld_base->pc);
1144
1145 ctx->flow_depth--;
1146 }
1147
1148 static void endloop_emit(const struct lp_build_tgsi_action *action,
1149 struct lp_build_tgsi_context *bld_base,
1150 struct lp_build_emit_data *emit_data)
1151 {
1152 struct si_shader_context *ctx = si_shader_context(bld_base);
1153 struct gallivm_state *gallivm = bld_base->base.gallivm;
1154 struct si_llvm_flow *current_loop = get_current_flow(ctx);
1155
1156 assert(current_loop->loop_entry_block);
1157
1158 emit_default_branch(gallivm->builder, current_loop->loop_entry_block);
1159
1160 LLVMPositionBuilderAtEnd(gallivm->builder, current_loop->next_block);
1161 set_basicblock_name(current_loop->next_block, "endloop", bld_base->pc);
1162 ctx->flow_depth--;
1163 }
1164
1165 static void if_cond_emit(const struct lp_build_tgsi_action *action,
1166 struct lp_build_tgsi_context *bld_base,
1167 struct lp_build_emit_data *emit_data,
1168 LLVMValueRef cond)
1169 {
1170 struct si_shader_context *ctx = si_shader_context(bld_base);
1171 struct gallivm_state *gallivm = bld_base->base.gallivm;
1172 struct si_llvm_flow *flow = push_flow(ctx);
1173 LLVMBasicBlockRef if_block;
1174
1175 if_block = append_basic_block(ctx, "IF");
1176 flow->next_block = append_basic_block(ctx, "ELSE");
1177 set_basicblock_name(if_block, "if", bld_base->pc);
1178 LLVMBuildCondBr(gallivm->builder, cond, if_block, flow->next_block);
1179 LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
1180 }
1181
1182 static void if_emit(const struct lp_build_tgsi_action *action,
1183 struct lp_build_tgsi_context *bld_base,
1184 struct lp_build_emit_data *emit_data)
1185 {
1186 struct gallivm_state *gallivm = bld_base->base.gallivm;
1187 LLVMValueRef cond;
1188
1189 cond = LLVMBuildFCmp(gallivm->builder, LLVMRealUNE,
1190 emit_data->args[0],
1191 bld_base->base.zero, "");
1192
1193 if_cond_emit(action, bld_base, emit_data, cond);
1194 }
1195
1196 static void uif_emit(const struct lp_build_tgsi_action *action,
1197 struct lp_build_tgsi_context *bld_base,
1198 struct lp_build_emit_data *emit_data)
1199 {
1200 struct gallivm_state *gallivm = bld_base->base.gallivm;
1201 LLVMValueRef cond;
1202
1203 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
1204 bitcast(bld_base, TGSI_TYPE_UNSIGNED, emit_data->args[0]),
1205 bld_base->int_bld.zero, "");
1206
1207 if_cond_emit(action, bld_base, emit_data, cond);
1208 }
1209
1210 static void emit_immediate(struct lp_build_tgsi_context *bld_base,
1211 const struct tgsi_full_immediate *imm)
1212 {
1213 unsigned i;
1214 struct si_shader_context *ctx = si_shader_context(bld_base);
1215
1216 for (i = 0; i < 4; ++i) {
1217 ctx->soa.immediates[ctx->soa.num_immediates][i] =
1218 LLVMConstInt(bld_base->uint_bld.elem_type, imm->u[i].Uint, false );
1219 }
1220
1221 ctx->soa.num_immediates++;
1222 }
1223
1224 void si_llvm_context_init(struct si_shader_context *ctx, const char *triple,
1225 const struct tgsi_shader_info *info,
1226 const struct tgsi_token *tokens)
1227 {
1228 struct lp_type type;
1229
1230 /* Initialize the gallivm object:
1231 * We are only using the module, context, and builder fields of this struct.
1232 * This should be enough for us to be able to pass our gallivm struct to the
1233 * helper functions in the gallivm module.
1234 */
1235 memset(&ctx->gallivm, 0, sizeof (ctx->gallivm));
1236 memset(&ctx->soa, 0, sizeof(ctx->soa));
1237 ctx->gallivm.context = LLVMContextCreate();
1238 ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
1239 ctx->gallivm.context);
1240 LLVMSetTarget(ctx->gallivm.module, triple);
1241 ctx->gallivm.builder = LLVMCreateBuilderInContext(ctx->gallivm.context);
1242
1243 struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
1244
1245 bld_base->info = info;
1246
1247 if (info && info->array_max[TGSI_FILE_TEMPORARY] > 0) {
1248 int size = info->array_max[TGSI_FILE_TEMPORARY];
1249
1250 ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
1251 ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
1252
1253 if (tokens)
1254 tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
1255 ctx->temp_arrays);
1256 }
1257
1258 type.floating = true;
1259 type.fixed = false;
1260 type.sign = true;
1261 type.norm = false;
1262 type.width = 32;
1263 type.length = 1;
1264
1265 lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
1266 lp_build_context_init(&ctx->soa.bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
1267 lp_build_context_init(&ctx->soa.bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
1268 type.width *= 2;
1269 lp_build_context_init(&ctx->soa.bld_base.dbl_bld, &ctx->gallivm, type);
1270 lp_build_context_init(&ctx->soa.bld_base.uint64_bld, &ctx->gallivm, lp_uint_type(type));
1271 lp_build_context_init(&ctx->soa.bld_base.int64_bld, &ctx->gallivm, lp_int_type(type));
1272
1273 bld_base->soa = 1;
1274 bld_base->emit_store = si_llvm_emit_store;
1275 bld_base->emit_swizzle = emit_swizzle;
1276 bld_base->emit_declaration = emit_declaration;
1277 bld_base->emit_immediate = emit_immediate;
1278
1279 bld_base->emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = si_llvm_emit_fetch;
1280 bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = si_llvm_emit_fetch;
1281 bld_base->emit_fetch_funcs[TGSI_FILE_TEMPORARY] = si_llvm_emit_fetch;
1282 bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = si_llvm_emit_fetch;
1283 bld_base->emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
1284
1285 /* metadata allowing 2.5 ULP */
1286 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->gallivm.context,
1287 "fpmath", 6);
1288 LLVMValueRef arg = lp_build_const_float(&ctx->gallivm, 2.5);
1289 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->gallivm.context,
1290 &arg, 1);
1291
1292 /* Allocate outputs */
1293 ctx->soa.outputs = ctx->outputs;
1294
1295 bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
1296 bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
1297 bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
1298 bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
1299 bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
1300 bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
1301 bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
1302 bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1303 }
1304
1305 void si_llvm_create_func(struct si_shader_context *ctx,
1306 LLVMTypeRef *return_types, unsigned num_return_elems,
1307 LLVMTypeRef *ParamTypes, unsigned ParamCount)
1308 {
1309 LLVMTypeRef main_fn_type, ret_type;
1310 LLVMBasicBlockRef main_fn_body;
1311
1312 if (num_return_elems)
1313 ret_type = LLVMStructTypeInContext(ctx->gallivm.context,
1314 return_types,
1315 num_return_elems, true);
1316 else
1317 ret_type = LLVMVoidTypeInContext(ctx->gallivm.context);
1318
1319 /* Setup the function */
1320 ctx->return_type = ret_type;
1321 main_fn_type = LLVMFunctionType(ret_type, ParamTypes, ParamCount, 0);
1322 ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, "main", main_fn_type);
1323 main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
1324 ctx->main_fn, "main_body");
1325 LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
1326 }
1327
1328 void si_llvm_finalize_module(struct si_shader_context *ctx,
1329 bool run_verifier)
1330 {
1331 struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
1332 const char *triple = LLVMGetTarget(gallivm->module);
1333 LLVMTargetLibraryInfoRef target_library_info;
1334
1335 /* Create the pass manager */
1336 gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(
1337 gallivm->module);
1338
1339 target_library_info = gallivm_create_target_library_info(triple);
1340 LLVMAddTargetLibraryInfo(target_library_info, gallivm->passmgr);
1341
1342 if (run_verifier)
1343 LLVMAddVerifierPass(gallivm->passmgr);
1344
1345 /* This pass should eliminate all the load and store instructions */
1346 LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
1347
1348 /* Add some optimization passes */
1349 LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
1350 LLVMAddLICMPass(gallivm->passmgr);
1351 LLVMAddAggressiveDCEPass(gallivm->passmgr);
1352 LLVMAddCFGSimplificationPass(gallivm->passmgr);
1353 LLVMAddInstructionCombiningPass(gallivm->passmgr);
1354
1355 /* Run the pass */
1356 LLVMInitializeFunctionPassManager(gallivm->passmgr);
1357 LLVMRunFunctionPassManager(gallivm->passmgr, ctx->main_fn);
1358 LLVMFinalizeFunctionPassManager(gallivm->passmgr);
1359
1360 LLVMDisposeBuilder(gallivm->builder);
1361 LLVMDisposePassManager(gallivm->passmgr);
1362 gallivm_dispose_target_library_info(target_library_info);
1363 }
1364
1365 void si_llvm_dispose(struct si_shader_context *ctx)
1366 {
1367 LLVMDisposeModule(ctx->soa.bld_base.base.gallivm->module);
1368 LLVMContextDispose(ctx->soa.bld_base.base.gallivm->context);
1369 FREE(ctx->temp_arrays);
1370 ctx->temp_arrays = NULL;
1371 FREE(ctx->temp_array_allocas);
1372 ctx->temp_array_allocas = NULL;
1373 FREE(ctx->temps);
1374 ctx->temps = NULL;
1375 ctx->temps_count = 0;
1376 FREE(ctx->flow);
1377 ctx->flow = NULL;
1378 ctx->flow_depth_max = 0;
1379 }