projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: use ac helpers for bitcasts
[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 "si_pipe.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/IPO.h>
42 #include <llvm-c/Transforms/Scalar.h>
43
44 /* Data for if/else/endif and bgnloop/endloop control flow structures.
45 */
46 struct si_llvm_flow {
47 /* Loop exit or next part of if/else/endif. */
48 LLVMBasicBlockRef next_block;
49 LLVMBasicBlockRef loop_entry_block;
50 };
51
52 enum si_llvm_calling_convention {
53 RADEON_LLVM_AMDGPU_VS = 87,
54 RADEON_LLVM_AMDGPU_GS = 88,
55 RADEON_LLVM_AMDGPU_PS = 89,
56 RADEON_LLVM_AMDGPU_CS = 90,
57 RADEON_LLVM_AMDGPU_HS = 93,
58 };
59
60 void si_llvm_add_attribute(LLVMValueRef F, const char *name, int value)
61 {
62 char str[16];
63
64 snprintf(str, sizeof(str), "%i", value);
65 LLVMAddTargetDependentFunctionAttr(F, name, str);
66 }
67
68 struct si_llvm_diagnostics {
69 struct pipe_debug_callback *debug;
70 unsigned retval;
71 };
72
73 static void si_diagnostic_handler(LLVMDiagnosticInfoRef di, void *context)
74 {
75 struct si_llvm_diagnostics *diag = (struct si_llvm_diagnostics *)context;
76 LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
77 char *description = LLVMGetDiagInfoDescription(di);
78 const char *severity_str = NULL;
79
80 switch (severity) {
81 case LLVMDSError:
82 severity_str = "error";
83 break;
84 case LLVMDSWarning:
85 severity_str = "warning";
86 break;
87 case LLVMDSRemark:
88 severity_str = "remark";
89 break;
90 case LLVMDSNote:
91 severity_str = "note";
92 break;
93 default:
94 severity_str = "unknown";
95 }
96
97 pipe_debug_message(diag->debug, SHADER_INFO,
98 "LLVM diagnostic (%s): %s", severity_str, description);
99
100 if (severity == LLVMDSError) {
101 diag->retval = 1;
102 fprintf(stderr,"LLVM triggered Diagnostic Handler: %s\n", description);
103 }
104
105 LLVMDisposeMessage(description);
106 }
107
108 /**
109 * Compile an LLVM module to machine code.
110 *
111 * @returns 0 for success, 1 for failure
112 */
113 unsigned si_llvm_compile(LLVMModuleRef M, struct ac_shader_binary *binary,
114 LLVMTargetMachineRef tm,
115 struct pipe_debug_callback *debug)
116 {
117 struct si_llvm_diagnostics diag;
118 char *err;
119 LLVMContextRef llvm_ctx;
120 LLVMMemoryBufferRef out_buffer;
121 unsigned buffer_size;
122 const char *buffer_data;
123 LLVMBool mem_err;
124
125 diag.debug = debug;
126 diag.retval = 0;
127
128 /* Setup Diagnostic Handler*/
129 llvm_ctx = LLVMGetModuleContext(M);
130
131 LLVMContextSetDiagnosticHandler(llvm_ctx, si_diagnostic_handler, &diag);
132
133 /* Compile IR*/
134 mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile, &err,
135 &out_buffer);
136
137 /* Process Errors/Warnings */
138 if (mem_err) {
139 fprintf(stderr, "%s: %s", __FUNCTION__, err);
140 pipe_debug_message(debug, SHADER_INFO,
141 "LLVM emit error: %s", err);
142 FREE(err);
143 diag.retval = 1;
144 goto out;
145 }
146
147 /* Extract Shader Code*/
148 buffer_size = LLVMGetBufferSize(out_buffer);
149 buffer_data = LLVMGetBufferStart(out_buffer);
150
151 if (!ac_elf_read(buffer_data, buffer_size, binary)) {
152 fprintf(stderr, "radeonsi: cannot read an ELF shader binary\n");
153 diag.retval = 1;
154 }
155
156 /* Clean up */
157 LLVMDisposeMemoryBuffer(out_buffer);
158
159 out:
160 if (diag.retval != 0)
161 pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed");
162 return diag.retval;
163 }
164
165 LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
166 enum tgsi_opcode_type type)
167 {
168 LLVMContextRef ctx = bld_base->base.gallivm->context;
169
170 switch (type) {
171 case TGSI_TYPE_UNSIGNED:
172 case TGSI_TYPE_SIGNED:
173 return LLVMInt32TypeInContext(ctx);
174 case TGSI_TYPE_UNSIGNED64:
175 case TGSI_TYPE_SIGNED64:
176 return LLVMInt64TypeInContext(ctx);
177 case TGSI_TYPE_DOUBLE:
178 return LLVMDoubleTypeInContext(ctx);
179 case TGSI_TYPE_UNTYPED:
180 case TGSI_TYPE_FLOAT:
181 return LLVMFloatTypeInContext(ctx);
182 default: break;
183 }
184 return 0;
185 }
186
187 LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
188 enum tgsi_opcode_type type, LLVMValueRef value)
189 {
190 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
191 LLVMTypeRef dst_type = tgsi2llvmtype(bld_base, type);
192
193 if (dst_type)
194 return LLVMBuildBitCast(builder, value, dst_type, "");
195 else
196 return value;
197 }
198
199 /**
200 * Return a value that is equal to the given i32 \p index if it lies in [0,num)
201 * or an undefined value in the same interval otherwise.
202 */
203 LLVMValueRef si_llvm_bound_index(struct si_shader_context *ctx,
204 LLVMValueRef index,
205 unsigned num)
206 {
207 struct gallivm_state *gallivm = &ctx->gallivm;
208 LLVMBuilderRef builder = gallivm->builder;
209 LLVMValueRef c_max = LLVMConstInt(ctx->i32, num - 1, 0);
210 LLVMValueRef cc;
211
212 if (util_is_power_of_two(num)) {
213 index = LLVMBuildAnd(builder, index, c_max, "");
214 } else {
215 /* In theory, this MAX pattern should result in code that is
216 * as good as the bit-wise AND above.
217 *
218 * In practice, LLVM generates worse code (at the time of
219 * writing), because its value tracking is not strong enough.
220 */
221 cc = LLVMBuildICmp(builder, LLVMIntULE, index, c_max, "");
222 index = LLVMBuildSelect(builder, cc, index, c_max, "");
223 }
224
225 return index;
226 }
227
228 static struct si_llvm_flow *
229 get_current_flow(struct si_shader_context *ctx)
230 {
231 if (ctx->flow_depth > 0)
232 return &ctx->flow[ctx->flow_depth - 1];
233 return NULL;
234 }
235
236 static struct si_llvm_flow *
237 get_innermost_loop(struct si_shader_context *ctx)
238 {
239 for (unsigned i = ctx->flow_depth; i > 0; --i) {
240 if (ctx->flow[i - 1].loop_entry_block)
241 return &ctx->flow[i - 1];
242 }
243 return NULL;
244 }
245
246 static struct si_llvm_flow *
247 push_flow(struct si_shader_context *ctx)
248 {
249 struct si_llvm_flow *flow;
250
251 if (ctx->flow_depth >= ctx->flow_depth_max) {
252 unsigned new_max = MAX2(ctx->flow_depth << 1, RADEON_LLVM_INITIAL_CF_DEPTH);
253 ctx->flow = REALLOC(ctx->flow,
254 ctx->flow_depth_max * sizeof(*ctx->flow),
255 new_max * sizeof(*ctx->flow));
256 ctx->flow_depth_max = new_max;
257 }
258
259 flow = &ctx->flow[ctx->flow_depth];
260 ctx->flow_depth++;
261
262 flow->next_block = NULL;
263 flow->loop_entry_block = NULL;
264 return flow;
265 }
266
267 static LLVMValueRef emit_swizzle(struct lp_build_tgsi_context *bld_base,
268 LLVMValueRef value,
269 unsigned swizzle_x,
270 unsigned swizzle_y,
271 unsigned swizzle_z,
272 unsigned swizzle_w)
273 {
274 LLVMValueRef swizzles[4];
275 LLVMTypeRef i32t =
276 LLVMInt32TypeInContext(bld_base->base.gallivm->context);
277
278 swizzles[0] = LLVMConstInt(i32t, swizzle_x, 0);
279 swizzles[1] = LLVMConstInt(i32t, swizzle_y, 0);
280 swizzles[2] = LLVMConstInt(i32t, swizzle_z, 0);
281 swizzles[3] = LLVMConstInt(i32t, swizzle_w, 0);
282
283 return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
284 value,
285 LLVMGetUndef(LLVMTypeOf(value)),
286 LLVMConstVector(swizzles, 4), "");
287 }
288
289 /**
290 * Return the description of the array covering the given temporary register
291 * index.
292 */
293 static unsigned
294 get_temp_array_id(struct lp_build_tgsi_context *bld_base,
295 unsigned reg_index,
296 const struct tgsi_ind_register *reg)
297 {
298 struct si_shader_context *ctx = si_shader_context(bld_base);
299 unsigned num_arrays = ctx->bld_base.info->array_max[TGSI_FILE_TEMPORARY];
300 unsigned i;
301
302 if (reg && reg->ArrayID > 0 && reg->ArrayID <= num_arrays)
303 return reg->ArrayID;
304
305 for (i = 0; i < num_arrays; i++) {
306 const struct tgsi_array_info *array = &ctx->temp_arrays[i];
307
308 if (reg_index >= array->range.First && reg_index <= array->range.Last)
309 return i + 1;
310 }
311
312 return 0;
313 }
314
315 static struct tgsi_declaration_range
316 get_array_range(struct lp_build_tgsi_context *bld_base,
317 unsigned File, unsigned reg_index,
318 const struct tgsi_ind_register *reg)
319 {
320 struct si_shader_context *ctx = si_shader_context(bld_base);
321 struct tgsi_declaration_range range;
322
323 if (File == TGSI_FILE_TEMPORARY) {
324 unsigned array_id = get_temp_array_id(bld_base, reg_index, reg);
325 if (array_id)
326 return ctx->temp_arrays[array_id - 1].range;
327 }
328
329 range.First = 0;
330 range.Last = bld_base->info->file_max[File];
331 return range;
332 }
333
334 /**
335 * For indirect registers, construct a pointer directly to the requested
336 * element using getelementptr if possible.
337 *
338 * Returns NULL if the insertelement/extractelement fallback for array access
339 * must be used.
340 */
341 static LLVMValueRef
342 get_pointer_into_array(struct si_shader_context *ctx,
343 unsigned file,
344 unsigned swizzle,
345 unsigned reg_index,
346 const struct tgsi_ind_register *reg_indirect)
347 {
348 unsigned array_id;
349 struct tgsi_array_info *array;
350 struct gallivm_state *gallivm = &ctx->gallivm;
351 LLVMBuilderRef builder = gallivm->builder;
352 LLVMValueRef idxs[2];
353 LLVMValueRef index;
354 LLVMValueRef alloca;
355
356 if (file != TGSI_FILE_TEMPORARY)
357 return NULL;
358
359 array_id = get_temp_array_id(&ctx->bld_base, reg_index, reg_indirect);
360 if (!array_id)
361 return NULL;
362
363 alloca = ctx->temp_array_allocas[array_id - 1];
364 if (!alloca)
365 return NULL;
366
367 array = &ctx->temp_arrays[array_id - 1];
368
369 if (!(array->writemask & (1 << swizzle)))
370 return ctx->undef_alloca;
371
372 index = si_get_indirect_index(ctx, reg_indirect, 1,
373 reg_index - ctx->temp_arrays[array_id - 1].range.First);
374
375 /* Ensure that the index is within a valid range, to guard against
376 * VM faults and overwriting critical data (e.g. spilled resource
377 * descriptors).
378 *
379 * TODO It should be possible to avoid the additional instructions
380 * if LLVM is changed so that it guarantuees:
381 * 1. the scratch space descriptor isolates the current wave (this
382 * could even save the scratch offset SGPR at the cost of an
383 * additional SALU instruction)
384 * 2. the memory for allocas must be allocated at the _end_ of the
385 * scratch space (after spilled registers)
386 */
387 index = si_llvm_bound_index(ctx, index, array->range.Last - array->range.First + 1);
388
389 index = LLVMBuildMul(
390 builder, index,
391 LLVMConstInt(ctx->i32, util_bitcount(array->writemask), 0),
392 "");
393 index = LLVMBuildAdd(
394 builder, index,
395 LLVMConstInt(ctx->i32,
396 util_bitcount(array->writemask & ((1 << swizzle) - 1)), 0),
397 "");
398 idxs[0] = ctx->i32_0;
399 idxs[1] = index;
400 return LLVMBuildGEP(builder, alloca, idxs, 2, "");
401 }
402
403 LLVMValueRef
404 si_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
405 enum tgsi_opcode_type type,
406 LLVMValueRef ptr,
407 LLVMValueRef ptr2)
408 {
409 struct si_shader_context *ctx = si_shader_context(bld_base);
410 LLVMBuilderRef builder = bld_base->base.gallivm->builder;
411 LLVMValueRef result;
412
413 result = LLVMGetUndef(LLVMVectorType(LLVMIntTypeInContext(bld_base->base.gallivm->context, 32), bld_base->base.type.length * 2));
414
415 result = LLVMBuildInsertElement(builder,
416 result,
417 ac_to_integer(&ctx->ac, ptr),
418 bld_base->int_bld.zero, "");
419 result = LLVMBuildInsertElement(builder,
420 result,
421 ac_to_integer(&ctx->ac, ptr2),
422 bld_base->int_bld.one, "");
423 return bitcast(bld_base, type, result);
424 }
425
426 static LLVMValueRef
427 emit_array_fetch(struct lp_build_tgsi_context *bld_base,
428 unsigned File, enum tgsi_opcode_type type,
429 struct tgsi_declaration_range range,
430 unsigned swizzle)
431 {
432 struct si_shader_context *ctx = si_shader_context(bld_base);
433
434 LLVMBuilderRef builder = ctx->gallivm.builder;
435
436 unsigned i, size = range.Last - range.First + 1;
437 LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
438 LLVMValueRef result = LLVMGetUndef(vec);
439
440 struct tgsi_full_src_register tmp_reg = {};
441 tmp_reg.Register.File = File;
442
443 for (i = 0; i < size; ++i) {
444 tmp_reg.Register.Index = i + range.First;
445 LLVMValueRef temp = si_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
446 result = LLVMBuildInsertElement(builder, result, temp,
447 LLVMConstInt(ctx->i32, i, 0), "array_vector");
448 }
449 return result;
450 }
451
452 static LLVMValueRef
453 load_value_from_array(struct lp_build_tgsi_context *bld_base,
454 unsigned file,
455 enum tgsi_opcode_type type,
456 unsigned swizzle,
457 unsigned reg_index,
458 const struct tgsi_ind_register *reg_indirect)
459 {
460 struct si_shader_context *ctx = si_shader_context(bld_base);
461 struct gallivm_state *gallivm = &ctx->gallivm;
462 LLVMBuilderRef builder = gallivm->builder;
463 LLVMValueRef ptr;
464
465 ptr = get_pointer_into_array(ctx, file, swizzle, reg_index, reg_indirect);
466 if (ptr) {
467 LLVMValueRef val = LLVMBuildLoad(builder, ptr, "");
468 if (tgsi_type_is_64bit(type)) {
469 LLVMValueRef ptr_hi, val_hi;
470 ptr_hi = LLVMBuildGEP(builder, ptr, &ctx->i32_1, 1, "");
471 val_hi = LLVMBuildLoad(builder, ptr_hi, "");
472 val = si_llvm_emit_fetch_64bit(bld_base, type, val, val_hi);
473 }
474
475 return val;
476 } else {
477 struct tgsi_declaration_range range =
478 get_array_range(bld_base, file, reg_index, reg_indirect);
479 LLVMValueRef index =
480 si_get_indirect_index(ctx, reg_indirect, 1, reg_index - range.First);
481 LLVMValueRef array =
482 emit_array_fetch(bld_base, file, type, range, swizzle);
483 return LLVMBuildExtractElement(builder, array, index, "");
484 }
485 }
486
487 static void
488 store_value_to_array(struct lp_build_tgsi_context *bld_base,
489 LLVMValueRef value,
490 unsigned file,
491 unsigned chan_index,
492 unsigned reg_index,
493 const struct tgsi_ind_register *reg_indirect)
494 {
495 struct si_shader_context *ctx = si_shader_context(bld_base);
496 struct gallivm_state *gallivm = &ctx->gallivm;
497 LLVMBuilderRef builder = gallivm->builder;
498 LLVMValueRef ptr;
499
500 ptr = get_pointer_into_array(ctx, file, chan_index, reg_index, reg_indirect);
501 if (ptr) {
502 LLVMBuildStore(builder, value, ptr);
503 } else {
504 unsigned i, size;
505 struct tgsi_declaration_range range = get_array_range(bld_base, file, reg_index, reg_indirect);
506 LLVMValueRef index = si_get_indirect_index(ctx, reg_indirect, 1, reg_index - range.First);
507 LLVMValueRef array =
508 emit_array_fetch(bld_base, file, TGSI_TYPE_FLOAT, range, chan_index);
509 LLVMValueRef temp_ptr;
510
511 array = LLVMBuildInsertElement(builder, array, value, index, "");
512
513 size = range.Last - range.First + 1;
514 for (i = 0; i < size; ++i) {
515 switch(file) {
516 case TGSI_FILE_OUTPUT:
517 temp_ptr = ctx->outputs[i + range.First][chan_index];
518 break;
519
520 case TGSI_FILE_TEMPORARY:
521 if (range.First + i >= ctx->temps_count)
522 continue;
523 temp_ptr = ctx->temps[(i + range.First) * TGSI_NUM_CHANNELS + chan_index];
524 break;
525
526 default:
527 continue;
528 }
529 value = LLVMBuildExtractElement(builder, array,
530 LLVMConstInt(ctx->i32, i, 0), "");
531 LLVMBuildStore(builder, value, temp_ptr);
532 }
533 }
534 }
535
536 /* If this is true, preload FS inputs at the beginning of shaders. Otherwise,
537 * reload them at each use. This must be true if the shader is using
538 * derivatives and KILL, because KILL can leave the WQM and then a lazy
539 * input load isn't in the WQM anymore.
540 */
541 static bool si_preload_fs_inputs(struct si_shader_context *ctx)
542 {
543 struct si_shader_selector *sel = ctx->shader->selector;
544
545 return sel->info.uses_derivatives &&
546 sel->info.uses_kill;
547 }
548
549 static LLVMValueRef
550 get_output_ptr(struct lp_build_tgsi_context *bld_base, unsigned index,
551 unsigned chan)
552 {
553 struct si_shader_context *ctx = si_shader_context(bld_base);
554
555 assert(index <= ctx->bld_base.info->file_max[TGSI_FILE_OUTPUT]);
556 return ctx->outputs[index][chan];
557 }
558
559 LLVMValueRef si_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
560 const struct tgsi_full_src_register *reg,
561 enum tgsi_opcode_type type,
562 unsigned swizzle)
563 {
564 struct si_shader_context *ctx = si_shader_context(bld_base);
565 LLVMBuilderRef builder = ctx->gallivm.builder;
566 LLVMValueRef result = NULL, ptr, ptr2;
567
568 if (swizzle == ~0) {
569 LLVMValueRef values[TGSI_NUM_CHANNELS];
570 unsigned chan;
571 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
572 values[chan] = si_llvm_emit_fetch(bld_base, reg, type, chan);
573 }
574 return lp_build_gather_values(&ctx->gallivm, values,
575 TGSI_NUM_CHANNELS);
576 }
577
578 if (reg->Register.Indirect) {
579 LLVMValueRef load = load_value_from_array(bld_base, reg->Register.File, type,
580 swizzle, reg->Register.Index, &reg->Indirect);
581 return bitcast(bld_base, type, load);
582 }
583
584 switch(reg->Register.File) {
585 case TGSI_FILE_IMMEDIATE: {
586 LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
587 if (tgsi_type_is_64bit(type)) {
588 result = LLVMGetUndef(LLVMVectorType(ctx->i32, bld_base->base.type.length * 2));
589 result = LLVMConstInsertElement(result,
590 ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle],
591 ctx->i32_0);
592 result = LLVMConstInsertElement(result,
593 ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1],
594 ctx->i32_1);
595 return LLVMConstBitCast(result, ctype);
596 } else {
597 return LLVMConstBitCast(ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle], ctype);
598 }
599 }
600
601 case TGSI_FILE_INPUT: {
602 unsigned index = reg->Register.Index;
603 LLVMValueRef input[4];
604
605 /* I don't think doing this for vertex shaders is beneficial.
606 * For those, we want to make sure the VMEM loads are executed
607 * only once. Fragment shaders don't care much, because
608 * v_interp instructions are much cheaper than VMEM loads.
609 */
610 if (!si_preload_fs_inputs(ctx) &&
611 ctx->bld_base.info->processor == PIPE_SHADER_FRAGMENT)
612 ctx->load_input(ctx, index, &ctx->input_decls[index], input);
613 else
614 memcpy(input, &ctx->inputs[index * 4], sizeof(input));
615
616 result = input[swizzle];
617
618 if (tgsi_type_is_64bit(type)) {
619 ptr = result;
620 ptr2 = input[swizzle + 1];
621 return si_llvm_emit_fetch_64bit(bld_base, type, ptr, ptr2);
622 }
623 break;
624 }
625
626 case TGSI_FILE_TEMPORARY:
627 if (reg->Register.Index >= ctx->temps_count)
628 return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
629 ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
630 if (tgsi_type_is_64bit(type)) {
631 ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1];
632 return si_llvm_emit_fetch_64bit(bld_base, type,
633 LLVMBuildLoad(builder, ptr, ""),
634 LLVMBuildLoad(builder, ptr2, ""));
635 }
636 result = LLVMBuildLoad(builder, ptr, "");
637 break;
638
639 case TGSI_FILE_OUTPUT:
640 ptr = get_output_ptr(bld_base, reg->Register.Index, swizzle);
641 if (tgsi_type_is_64bit(type)) {
642 ptr2 = get_output_ptr(bld_base, reg->Register.Index, swizzle + 1);
643 return si_llvm_emit_fetch_64bit(bld_base, type,
644 LLVMBuildLoad(builder, ptr, ""),
645 LLVMBuildLoad(builder, ptr2, ""));
646 }
647 result = LLVMBuildLoad(builder, ptr, "");
648 break;
649
650 default:
651 return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
652 }
653
654 return bitcast(bld_base, type, result);
655 }
656
657 static LLVMValueRef fetch_system_value(struct lp_build_tgsi_context *bld_base,
658 const struct tgsi_full_src_register *reg,
659 enum tgsi_opcode_type type,
660 unsigned swizzle)
661 {
662 struct si_shader_context *ctx = si_shader_context(bld_base);
663 LLVMBuilderRef builder = ctx->gallivm.builder;
664 LLVMValueRef cval = ctx->system_values[reg->Register.Index];
665
666 if (tgsi_type_is_64bit(type)) {
667 LLVMValueRef lo, hi;
668
669 assert(swizzle == 0 || swizzle == 2);
670
671 lo = LLVMBuildExtractElement(
672 builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
673 hi = LLVMBuildExtractElement(
674 builder, cval, LLVMConstInt(ctx->i32, swizzle + 1, 0), "");
675
676 return si_llvm_emit_fetch_64bit(bld_base, type, lo, hi);
677 }
678
679 if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
680 cval = LLVMBuildExtractElement(
681 builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
682 } else {
683 assert(swizzle == 0);
684 }
685
686 return bitcast(bld_base, type, cval);
687 }
688
689 static void emit_declaration(struct lp_build_tgsi_context *bld_base,
690 const struct tgsi_full_declaration *decl)
691 {
692 struct si_shader_context *ctx = si_shader_context(bld_base);
693 LLVMBuilderRef builder = ctx->gallivm.builder;
694 unsigned first, last, i;
695 switch(decl->Declaration.File) {
696 case TGSI_FILE_ADDRESS:
697 {
698 unsigned idx;
699 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
700 unsigned chan;
701 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
702 ctx->addrs[idx][chan] = lp_build_alloca_undef(
703 &ctx->gallivm,
704 ctx->i32, "");
705 }
706 }
707 break;
708 }
709
710 case TGSI_FILE_TEMPORARY:
711 {
712 char name[16] = "";
713 LLVMValueRef array_alloca = NULL;
714 unsigned decl_size;
715 unsigned writemask = decl->Declaration.UsageMask;
716 first = decl->Range.First;
717 last = decl->Range.Last;
718 decl_size = 4 * ((last - first) + 1);
719
720 if (decl->Declaration.Array) {
721 unsigned id = decl->Array.ArrayID - 1;
722 unsigned array_size;
723
724 writemask &= ctx->temp_arrays[id].writemask;
725 ctx->temp_arrays[id].writemask = writemask;
726 array_size = ((last - first) + 1) * util_bitcount(writemask);
727
728 /* If the array has more than 16 elements, store it
729 * in memory using an alloca that spans the entire
730 * array.
731 *
732 * Otherwise, store each array element individually.
733 * We will then generate vectors (per-channel, up to
734 * <16 x float> if the usagemask is a single bit) for
735 * indirect addressing.
736 *
737 * Note that 16 is the number of vector elements that
738 * LLVM will store in a register, so theoretically an
739 * array with up to 4 * 16 = 64 elements could be
740 * handled this way, but whether that's a good idea
741 * depends on VGPR register pressure elsewhere.
742 *
743 * FIXME: We shouldn't need to have the non-alloca
744 * code path for arrays. LLVM should be smart enough to
745 * promote allocas into registers when profitable.
746 */
747 if (array_size > 16 ||
748 !ctx->screen->llvm_has_working_vgpr_indexing) {
749 array_alloca = lp_build_alloca_undef(&ctx->gallivm,
750 LLVMArrayType(ctx->f32,
751 array_size), "array");
752 ctx->temp_array_allocas[id] = array_alloca;
753 }
754 }
755
756 if (!ctx->temps_count) {
757 ctx->temps_count = bld_base->info->file_max[TGSI_FILE_TEMPORARY] + 1;
758 ctx->temps = MALLOC(TGSI_NUM_CHANNELS * ctx->temps_count * sizeof(LLVMValueRef));
759 }
760 if (!array_alloca) {
761 for (i = 0; i < decl_size; ++i) {
762 #ifdef DEBUG
763 snprintf(name, sizeof(name), "TEMP%d.%c",
764 first + i / 4, "xyzw"[i % 4]);
765 #endif
766 ctx->temps[first * TGSI_NUM_CHANNELS + i] =
767 lp_build_alloca_undef(&ctx->gallivm,
768 ctx->f32,
769 name);
770 }
771 } else {
772 LLVMValueRef idxs[2] = {
773 ctx->i32_0,
774 NULL
775 };
776 unsigned j = 0;
777
778 if (writemask != TGSI_WRITEMASK_XYZW &&
779 !ctx->undef_alloca) {
780 /* Create a dummy alloca. We use it so that we
781 * have a pointer that is safe to load from if
782 * a shader ever reads from a channel that
783 * it never writes to.
784 */
785 ctx->undef_alloca = lp_build_alloca_undef(
786 &ctx->gallivm,
787 ctx->f32, "undef");
788 }
789
790 for (i = 0; i < decl_size; ++i) {
791 LLVMValueRef ptr;
792 if (writemask & (1 << (i % 4))) {
793 #ifdef DEBUG
794 snprintf(name, sizeof(name), "TEMP%d.%c",
795 first + i / 4, "xyzw"[i % 4]);
796 #endif
797 idxs[1] = LLVMConstInt(ctx->i32, j, 0);
798 ptr = LLVMBuildGEP(builder, array_alloca, idxs, 2, name);
799 j++;
800 } else {
801 ptr = ctx->undef_alloca;
802 }
803 ctx->temps[first * TGSI_NUM_CHANNELS + i] = ptr;
804 }
805 }
806 break;
807 }
808 case TGSI_FILE_INPUT:
809 {
810 unsigned idx;
811 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
812 if (ctx->load_input &&
813 ctx->input_decls[idx].Declaration.File != TGSI_FILE_INPUT) {
814 ctx->input_decls[idx] = *decl;
815 ctx->input_decls[idx].Range.First = idx;
816 ctx->input_decls[idx].Range.Last = idx;
817 ctx->input_decls[idx].Semantic.Index += idx - decl->Range.First;
818
819 if (si_preload_fs_inputs(ctx) ||
820 bld_base->info->processor != PIPE_SHADER_FRAGMENT)
821 ctx->load_input(ctx, idx, &ctx->input_decls[idx],
822 &ctx->inputs[idx * 4]);
823 }
824 }
825 }
826 break;
827
828 case TGSI_FILE_SYSTEM_VALUE:
829 {
830 unsigned idx;
831 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
832 si_load_system_value(ctx, idx, decl);
833 }
834 }
835 break;
836
837 case TGSI_FILE_OUTPUT:
838 {
839 char name[16] = "";
840 unsigned idx;
841 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
842 unsigned chan;
843 assert(idx < RADEON_LLVM_MAX_OUTPUTS);
844 if (ctx->outputs[idx][0])
845 continue;
846 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
847 #ifdef DEBUG
848 snprintf(name, sizeof(name), "OUT%d.%c",
849 idx, "xyzw"[chan % 4]);
850 #endif
851 ctx->outputs[idx][chan] = lp_build_alloca_undef(
852 &ctx->gallivm,
853 ctx->f32, name);
854 }
855 }
856 break;
857 }
858
859 case TGSI_FILE_MEMORY:
860 si_declare_compute_memory(ctx, decl);
861 break;
862
863 default:
864 break;
865 }
866 }
867
868 void si_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
869 const struct tgsi_full_instruction *inst,
870 const struct tgsi_opcode_info *info,
871 unsigned index,
872 LLVMValueRef dst[4])
873 {
874 struct si_shader_context *ctx = si_shader_context(bld_base);
875 struct gallivm_state *gallivm = &ctx->gallivm;
876 const struct tgsi_full_dst_register *reg = &inst->Dst[index];
877 LLVMBuilderRef builder = ctx->gallivm.builder;
878 LLVMValueRef temp_ptr, temp_ptr2 = NULL;
879 bool is_vec_store = false;
880 enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode, index);
881
882 if (dst[0]) {
883 LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
884 is_vec_store = (k == LLVMVectorTypeKind);
885 }
886
887 if (is_vec_store) {
888 LLVMValueRef values[4] = {};
889 uint32_t writemask = reg->Register.WriteMask;
890 while (writemask) {
891 unsigned chan = u_bit_scan(&writemask);
892 LLVMValueRef index = LLVMConstInt(ctx->i32, chan, 0);
893 values[chan] = LLVMBuildExtractElement(gallivm->builder,
894 dst[0], index, "");
895 }
896 bld_base->emit_store(bld_base, inst, info, index, values);
897 return;
898 }
899
900 uint32_t writemask = reg->Register.WriteMask;
901 while (writemask) {
902 unsigned chan_index = u_bit_scan(&writemask);
903 LLVMValueRef value = dst[chan_index];
904
905 if (tgsi_type_is_64bit(dtype) && (chan_index == 1 || chan_index == 3))
906 continue;
907 if (inst->Instruction.Saturate)
908 value = ac_build_clamp(&ctx->ac, value);
909
910 if (reg->Register.File == TGSI_FILE_ADDRESS) {
911 temp_ptr = ctx->addrs[reg->Register.Index][chan_index];
912 LLVMBuildStore(builder, value, temp_ptr);
913 continue;
914 }
915
916 if (!tgsi_type_is_64bit(dtype))
917 value = ac_to_float(&ctx->ac, value);
918
919 if (reg->Register.Indirect) {
920 unsigned file = reg->Register.File;
921 unsigned reg_index = reg->Register.Index;
922 store_value_to_array(bld_base, value, file, chan_index,
923 reg_index, &reg->Indirect);
924 } else {
925 switch(reg->Register.File) {
926 case TGSI_FILE_OUTPUT:
927 temp_ptr = ctx->outputs[reg->Register.Index][chan_index];
928 if (tgsi_type_is_64bit(dtype))
929 temp_ptr2 = ctx->outputs[reg->Register.Index][chan_index + 1];
930 break;
931
932 case TGSI_FILE_TEMPORARY:
933 {
934 if (reg->Register.Index >= ctx->temps_count)
935 continue;
936
937 temp_ptr = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index];
938 if (tgsi_type_is_64bit(dtype))
939 temp_ptr2 = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index + 1];
940
941 break;
942 }
943 default:
944 return;
945 }
946 if (!tgsi_type_is_64bit(dtype))
947 LLVMBuildStore(builder, value, temp_ptr);
948 else {
949 LLVMValueRef ptr = LLVMBuildBitCast(builder, value,
950 LLVMVectorType(ctx->i32, 2), "");
951 LLVMValueRef val2;
952 value = LLVMBuildExtractElement(builder, ptr,
953 ctx->i32_0, "");
954 val2 = LLVMBuildExtractElement(builder, ptr,
955 ctx->i32_1, "");
956
957 LLVMBuildStore(builder, ac_to_float(&ctx->ac, value), temp_ptr);
958 LLVMBuildStore(builder, ac_to_float(&ctx->ac, val2), temp_ptr2);
959 }
960 }
961 }
962 }
963
964 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base, int pc)
965 {
966 char buf[32];
967 /* Subtract 1 so that the number shown is that of the corresponding
968 * opcode in the TGSI dump, e.g. an if block has the same suffix as
969 * the instruction number of the corresponding TGSI IF.
970 */
971 snprintf(buf, sizeof(buf), "%s%d", base, pc - 1);
972 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
973 }
974
975 /* Append a basic block at the level of the parent flow.
976 */
977 static LLVMBasicBlockRef append_basic_block(struct si_shader_context *ctx,
978 const char *name)
979 {
980 struct gallivm_state *gallivm = &ctx->gallivm;
981
982 assert(ctx->flow_depth >= 1);
983
984 if (ctx->flow_depth >= 2) {
985 struct si_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
986
987 return LLVMInsertBasicBlockInContext(gallivm->context,
988 flow->next_block, name);
989 }
990
991 return LLVMAppendBasicBlockInContext(gallivm->context, ctx->main_fn, name);
992 }
993
994 /* Emit a branch to the given default target for the current block if
995 * applicable -- that is, if the current block does not already contain a
996 * branch from a break or continue.
997 */
998 static void emit_default_branch(LLVMBuilderRef builder, LLVMBasicBlockRef target)
999 {
1000 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
1001 LLVMBuildBr(builder, target);
1002 }
1003
1004 static void bgnloop_emit(const struct lp_build_tgsi_action *action,
1005 struct lp_build_tgsi_context *bld_base,
1006 struct lp_build_emit_data *emit_data)
1007 {
1008 struct si_shader_context *ctx = si_shader_context(bld_base);
1009 struct gallivm_state *gallivm = &ctx->gallivm;
1010 struct si_llvm_flow *flow = push_flow(ctx);
1011 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
1012 flow->next_block = append_basic_block(ctx, "ENDLOOP");
1013 set_basicblock_name(flow->loop_entry_block, "loop", bld_base->pc);
1014 LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
1015 LLVMPositionBuilderAtEnd(gallivm->builder, flow->loop_entry_block);
1016 }
1017
1018 static void brk_emit(const struct lp_build_tgsi_action *action,
1019 struct lp_build_tgsi_context *bld_base,
1020 struct lp_build_emit_data *emit_data)
1021 {
1022 struct si_shader_context *ctx = si_shader_context(bld_base);
1023 struct gallivm_state *gallivm = &ctx->gallivm;
1024 struct si_llvm_flow *flow = get_innermost_loop(ctx);
1025
1026 LLVMBuildBr(gallivm->builder, flow->next_block);
1027 }
1028
1029 static void cont_emit(const struct lp_build_tgsi_action *action,
1030 struct lp_build_tgsi_context *bld_base,
1031 struct lp_build_emit_data *emit_data)
1032 {
1033 struct si_shader_context *ctx = si_shader_context(bld_base);
1034 struct gallivm_state *gallivm = &ctx->gallivm;
1035 struct si_llvm_flow *flow = get_innermost_loop(ctx);
1036
1037 LLVMBuildBr(gallivm->builder, flow->loop_entry_block);
1038 }
1039
1040 static void else_emit(const struct lp_build_tgsi_action *action,
1041 struct lp_build_tgsi_context *bld_base,
1042 struct lp_build_emit_data *emit_data)
1043 {
1044 struct si_shader_context *ctx = si_shader_context(bld_base);
1045 struct gallivm_state *gallivm = &ctx->gallivm;
1046 struct si_llvm_flow *current_branch = get_current_flow(ctx);
1047 LLVMBasicBlockRef endif_block;
1048
1049 assert(!current_branch->loop_entry_block);
1050
1051 endif_block = append_basic_block(ctx, "ENDIF");
1052 emit_default_branch(gallivm->builder, endif_block);
1053
1054 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
1055 set_basicblock_name(current_branch->next_block, "else", bld_base->pc);
1056
1057 current_branch->next_block = endif_block;
1058 }
1059
1060 static void endif_emit(const struct lp_build_tgsi_action *action,
1061 struct lp_build_tgsi_context *bld_base,
1062 struct lp_build_emit_data *emit_data)
1063 {
1064 struct si_shader_context *ctx = si_shader_context(bld_base);
1065 struct gallivm_state *gallivm = &ctx->gallivm;
1066 struct si_llvm_flow *current_branch = get_current_flow(ctx);
1067
1068 assert(!current_branch->loop_entry_block);
1069
1070 emit_default_branch(gallivm->builder, current_branch->next_block);
1071 LLVMPositionBuilderAtEnd(gallivm->builder, current_branch->next_block);
1072 set_basicblock_name(current_branch->next_block, "endif", bld_base->pc);
1073
1074 ctx->flow_depth--;
1075 }
1076
1077 static void endloop_emit(const struct lp_build_tgsi_action *action,
1078 struct lp_build_tgsi_context *bld_base,
1079 struct lp_build_emit_data *emit_data)
1080 {
1081 struct si_shader_context *ctx = si_shader_context(bld_base);
1082 struct gallivm_state *gallivm = &ctx->gallivm;
1083 struct si_llvm_flow *current_loop = get_current_flow(ctx);
1084
1085 assert(current_loop->loop_entry_block);
1086
1087 emit_default_branch(gallivm->builder, current_loop->loop_entry_block);
1088
1089 LLVMPositionBuilderAtEnd(gallivm->builder, current_loop->next_block);
1090 set_basicblock_name(current_loop->next_block, "endloop", bld_base->pc);
1091 ctx->flow_depth--;
1092 }
1093
1094 static void if_cond_emit(const struct lp_build_tgsi_action *action,
1095 struct lp_build_tgsi_context *bld_base,
1096 struct lp_build_emit_data *emit_data,
1097 LLVMValueRef cond)
1098 {
1099 struct si_shader_context *ctx = si_shader_context(bld_base);
1100 struct gallivm_state *gallivm = &ctx->gallivm;
1101 struct si_llvm_flow *flow = push_flow(ctx);
1102 LLVMBasicBlockRef if_block;
1103
1104 if_block = append_basic_block(ctx, "IF");
1105 flow->next_block = append_basic_block(ctx, "ELSE");
1106 set_basicblock_name(if_block, "if", bld_base->pc);
1107 LLVMBuildCondBr(gallivm->builder, cond, if_block, flow->next_block);
1108 LLVMPositionBuilderAtEnd(gallivm->builder, if_block);
1109 }
1110
1111 static void if_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 gallivm_state *gallivm = bld_base->base.gallivm;
1116 LLVMValueRef cond;
1117
1118 cond = LLVMBuildFCmp(gallivm->builder, LLVMRealUNE,
1119 emit_data->args[0],
1120 bld_base->base.zero, "");
1121
1122 if_cond_emit(action, bld_base, emit_data, cond);
1123 }
1124
1125 static void uif_emit(const struct lp_build_tgsi_action *action,
1126 struct lp_build_tgsi_context *bld_base,
1127 struct lp_build_emit_data *emit_data)
1128 {
1129 struct si_shader_context *ctx = si_shader_context(bld_base);
1130 struct gallivm_state *gallivm = bld_base->base.gallivm;
1131 LLVMValueRef cond;
1132
1133 cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE,
1134 ac_to_integer(&ctx->ac, emit_data->args[0]),
1135 bld_base->int_bld.zero, "");
1136
1137 if_cond_emit(action, bld_base, emit_data, cond);
1138 }
1139
1140 static void emit_immediate(struct lp_build_tgsi_context *bld_base,
1141 const struct tgsi_full_immediate *imm)
1142 {
1143 unsigned i;
1144 struct si_shader_context *ctx = si_shader_context(bld_base);
1145
1146 for (i = 0; i < 4; ++i) {
1147 ctx->imms[ctx->imms_num * TGSI_NUM_CHANNELS + i] =
1148 LLVMConstInt(ctx->i32, imm->u[i].Uint, false );
1149 }
1150
1151 ctx->imms_num++;
1152 }
1153
1154 void si_llvm_context_init(struct si_shader_context *ctx,
1155 struct si_screen *sscreen,
1156 LLVMTargetMachineRef tm)
1157 {
1158 struct lp_type type;
1159
1160 /* Initialize the gallivm object:
1161 * We are only using the module, context, and builder fields of this struct.
1162 * This should be enough for us to be able to pass our gallivm struct to the
1163 * helper functions in the gallivm module.
1164 */
1165 memset(ctx, 0, sizeof(*ctx));
1166 ctx->screen = sscreen;
1167 ctx->tm = tm;
1168
1169 ctx->gallivm.context = LLVMContextCreate();
1170 ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
1171 ctx->gallivm.context);
1172 LLVMSetTarget(ctx->gallivm.module, "amdgcn--");
1173
1174 LLVMTargetDataRef data_layout = LLVMCreateTargetDataLayout(tm);
1175 char *data_layout_str = LLVMCopyStringRepOfTargetData(data_layout);
1176 LLVMSetDataLayout(ctx->gallivm.module, data_layout_str);
1177 LLVMDisposeTargetData(data_layout);
1178 LLVMDisposeMessage(data_layout_str);
1179
1180 bool unsafe_fpmath = (sscreen->b.debug_flags & DBG_UNSAFE_MATH) != 0;
1181 enum lp_float_mode float_mode =
1182 unsafe_fpmath ? LP_FLOAT_MODE_UNSAFE_FP_MATH :
1183 LP_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH;
1184
1185 ctx->gallivm.builder = lp_create_builder(ctx->gallivm.context,
1186 float_mode);
1187
1188 ac_llvm_context_init(&ctx->ac, ctx->gallivm.context, sscreen->b.chip_class);
1189 ctx->ac.module = ctx->gallivm.module;
1190 ctx->ac.builder = ctx->gallivm.builder;
1191
1192 struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
1193
1194 type.floating = true;
1195 type.fixed = false;
1196 type.sign = true;
1197 type.norm = false;
1198 type.width = 32;
1199 type.length = 1;
1200
1201 lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
1202 lp_build_context_init(&ctx->bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
1203 lp_build_context_init(&ctx->bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
1204 type.width *= 2;
1205 lp_build_context_init(&ctx->bld_base.dbl_bld, &ctx->gallivm, type);
1206 lp_build_context_init(&ctx->bld_base.uint64_bld, &ctx->gallivm, lp_uint_type(type));
1207 lp_build_context_init(&ctx->bld_base.int64_bld, &ctx->gallivm, lp_int_type(type));
1208
1209 bld_base->soa = 1;
1210 bld_base->emit_swizzle = emit_swizzle;
1211 bld_base->emit_declaration = emit_declaration;
1212 bld_base->emit_immediate = emit_immediate;
1213
1214 /* metadata allowing 2.5 ULP */
1215 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->gallivm.context,
1216 "fpmath", 6);
1217 LLVMValueRef arg = lp_build_const_float(&ctx->gallivm, 2.5);
1218 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->gallivm.context,
1219 &arg, 1);
1220
1221 bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
1222 bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
1223 bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
1224 bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
1225 bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
1226 bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
1227 bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
1228 bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1229
1230 si_shader_context_init_alu(&ctx->bld_base);
1231 si_shader_context_init_mem(ctx);
1232
1233 ctx->voidt = LLVMVoidTypeInContext(ctx->gallivm.context);
1234 ctx->i1 = LLVMInt1TypeInContext(ctx->gallivm.context);
1235 ctx->i8 = LLVMInt8TypeInContext(ctx->gallivm.context);
1236 ctx->i32 = LLVMInt32TypeInContext(ctx->gallivm.context);
1237 ctx->i64 = LLVMInt64TypeInContext(ctx->gallivm.context);
1238 ctx->i128 = LLVMIntTypeInContext(ctx->gallivm.context, 128);
1239 ctx->f32 = LLVMFloatTypeInContext(ctx->gallivm.context);
1240 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
1241 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
1242 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
1243 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
1244
1245 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, 0);
1246 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, 0);
1247 }
1248
1249 /* Set the context to a certain TGSI shader. Can be called repeatedly
1250 * to change the shader. */
1251 void si_llvm_context_set_tgsi(struct si_shader_context *ctx,
1252 struct si_shader *shader)
1253 {
1254 const struct tgsi_shader_info *info = NULL;
1255 const struct tgsi_token *tokens = NULL;
1256
1257 if (shader && shader->selector) {
1258 info = &shader->selector->info;
1259 tokens = shader->selector->tokens;
1260 }
1261
1262 ctx->shader = shader;
1263 ctx->type = info ? info->processor : -1;
1264 ctx->bld_base.info = info;
1265
1266 /* Clean up the old contents. */
1267 FREE(ctx->temp_arrays);
1268 ctx->temp_arrays = NULL;
1269 FREE(ctx->temp_array_allocas);
1270 ctx->temp_array_allocas = NULL;
1271
1272 FREE(ctx->imms);
1273 ctx->imms = NULL;
1274 ctx->imms_num = 0;
1275
1276 FREE(ctx->temps);
1277 ctx->temps = NULL;
1278 ctx->temps_count = 0;
1279
1280 if (!info || !tokens)
1281 return;
1282
1283 if (info->array_max[TGSI_FILE_TEMPORARY] > 0) {
1284 int size = info->array_max[TGSI_FILE_TEMPORARY];
1285
1286 ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
1287 ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
1288
1289 tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
1290 ctx->temp_arrays);
1291 }
1292 if (info->file_max[TGSI_FILE_IMMEDIATE] >= 0) {
1293 int size = info->file_max[TGSI_FILE_IMMEDIATE] + 1;
1294 ctx->imms = MALLOC(size * TGSI_NUM_CHANNELS * sizeof(LLVMValueRef));
1295 }
1296
1297 /* Re-set these to start with a clean slate. */
1298 ctx->bld_base.num_instructions = 0;
1299 ctx->bld_base.pc = 0;
1300 memset(ctx->outputs, 0, sizeof(ctx->outputs));
1301
1302 ctx->bld_base.emit_store = si_llvm_emit_store;
1303 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = si_llvm_emit_fetch;
1304 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = si_llvm_emit_fetch;
1305 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = si_llvm_emit_fetch;
1306 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_OUTPUT] = si_llvm_emit_fetch;
1307 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
1308
1309 ctx->num_const_buffers = util_last_bit(info->const_buffers_declared);
1310 ctx->num_shader_buffers = util_last_bit(info->shader_buffers_declared);
1311 ctx->num_samplers = util_last_bit(info->samplers_declared);
1312 ctx->num_images = util_last_bit(info->images_declared);
1313 }
1314
1315 void si_llvm_create_func(struct si_shader_context *ctx,
1316 const char *name,
1317 LLVMTypeRef *return_types, unsigned num_return_elems,
1318 LLVMTypeRef *ParamTypes, unsigned ParamCount)
1319 {
1320 LLVMTypeRef main_fn_type, ret_type;
1321 LLVMBasicBlockRef main_fn_body;
1322 enum si_llvm_calling_convention call_conv;
1323 unsigned real_shader_type;
1324
1325 if (num_return_elems)
1326 ret_type = LLVMStructTypeInContext(ctx->gallivm.context,
1327 return_types,
1328 num_return_elems, true);
1329 else
1330 ret_type = LLVMVoidTypeInContext(ctx->gallivm.context);
1331
1332 /* Setup the function */
1333 ctx->return_type = ret_type;
1334 main_fn_type = LLVMFunctionType(ret_type, ParamTypes, ParamCount, 0);
1335 ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, name, main_fn_type);
1336 main_fn_body = LLVMAppendBasicBlockInContext(ctx->gallivm.context,
1337 ctx->main_fn, "main_body");
1338 LLVMPositionBuilderAtEnd(ctx->gallivm.builder, main_fn_body);
1339
1340 real_shader_type = ctx->type;
1341
1342 /* LS is merged into HS (TCS), and ES is merged into GS. */
1343 if (ctx->screen->b.chip_class >= GFX9) {
1344 if (ctx->shader->key.as_ls)
1345 real_shader_type = PIPE_SHADER_TESS_CTRL;
1346 else if (ctx->shader->key.as_es)
1347 real_shader_type = PIPE_SHADER_GEOMETRY;
1348 }
1349
1350 switch (real_shader_type) {
1351 case PIPE_SHADER_VERTEX:
1352 case PIPE_SHADER_TESS_EVAL:
1353 call_conv = RADEON_LLVM_AMDGPU_VS;
1354 break;
1355 case PIPE_SHADER_TESS_CTRL:
1356 call_conv = HAVE_LLVM >= 0x0500 ? RADEON_LLVM_AMDGPU_HS :
1357 RADEON_LLVM_AMDGPU_VS;
1358 break;
1359 case PIPE_SHADER_GEOMETRY:
1360 call_conv = RADEON_LLVM_AMDGPU_GS;
1361 break;
1362 case PIPE_SHADER_FRAGMENT:
1363 call_conv = RADEON_LLVM_AMDGPU_PS;
1364 break;
1365 case PIPE_SHADER_COMPUTE:
1366 call_conv = RADEON_LLVM_AMDGPU_CS;
1367 break;
1368 default:
1369 unreachable("Unhandle shader type");
1370 }
1371
1372 LLVMSetFunctionCallConv(ctx->main_fn, call_conv);
1373 }
1374
1375 void si_llvm_optimize_module(struct si_shader_context *ctx)
1376 {
1377 struct gallivm_state *gallivm = &ctx->gallivm;
1378 const char *triple = LLVMGetTarget(gallivm->module);
1379 LLVMTargetLibraryInfoRef target_library_info;
1380
1381 /* Dump LLVM IR before any optimization passes */
1382 if (ctx->screen->b.debug_flags & DBG_PREOPT_IR &&
1383 si_can_dump_shader(&ctx->screen->b, ctx->type))
1384 LLVMDumpModule(ctx->gallivm.module);
1385
1386 /* Create the pass manager */
1387 gallivm->passmgr = LLVMCreatePassManager();
1388
1389 target_library_info = gallivm_create_target_library_info(triple);
1390 LLVMAddTargetLibraryInfo(target_library_info, gallivm->passmgr);
1391
1392 if (si_extra_shader_checks(&ctx->screen->b, ctx->type))
1393 LLVMAddVerifierPass(gallivm->passmgr);
1394
1395 LLVMAddAlwaysInlinerPass(gallivm->passmgr);
1396
1397 /* This pass should eliminate all the load and store instructions */
1398 LLVMAddPromoteMemoryToRegisterPass(gallivm->passmgr);
1399
1400 /* Add some optimization passes */
1401 LLVMAddScalarReplAggregatesPass(gallivm->passmgr);
1402 LLVMAddLICMPass(gallivm->passmgr);
1403 LLVMAddAggressiveDCEPass(gallivm->passmgr);
1404 LLVMAddCFGSimplificationPass(gallivm->passmgr);
1405 #if HAVE_LLVM >= 0x0400
1406 /* This is recommended by the instruction combining pass. */
1407 LLVMAddEarlyCSEMemSSAPass(gallivm->passmgr);
1408 #endif
1409 LLVMAddInstructionCombiningPass(gallivm->passmgr);
1410
1411 /* Run the pass */
1412 LLVMRunPassManager(gallivm->passmgr, ctx->gallivm.module);
1413
1414 LLVMDisposeBuilder(gallivm->builder);
1415 LLVMDisposePassManager(gallivm->passmgr);
1416 gallivm_dispose_target_library_info(target_library_info);
1417 }
1418
1419 void si_llvm_dispose(struct si_shader_context *ctx)
1420 {
1421 LLVMDisposeModule(ctx->gallivm.module);
1422 LLVMContextDispose(ctx->gallivm.context);
1423 FREE(ctx->temp_arrays);
1424 ctx->temp_arrays = NULL;
1425 FREE(ctx->temp_array_allocas);
1426 ctx->temp_array_allocas = NULL;
1427 FREE(ctx->temps);
1428 ctx->temps = NULL;
1429 ctx->temps_count = 0;
1430 FREE(ctx->imms);
1431 ctx->imms = NULL;
1432 ctx->imms_num = 0;
1433 FREE(ctx->flow);
1434 ctx->flow = NULL;
1435 ctx->flow_depth_max = 0;
1436 }