projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: use ac_build_imad
[mesa.git] / src / gallium / drivers / radeonsi / si_shader_tgsi_setup.c
1 /*
2 * Copyright 2016 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 #include "si_shader_internal.h"
26 #include "si_pipe.h"
27 #include "ac_llvm_util.h"
28 #include "util/u_memory.h"
29
30 enum si_llvm_calling_convention {
31 RADEON_LLVM_AMDGPU_VS = 87,
32 RADEON_LLVM_AMDGPU_GS = 88,
33 RADEON_LLVM_AMDGPU_PS = 89,
34 RADEON_LLVM_AMDGPU_CS = 90,
35 RADEON_LLVM_AMDGPU_HS = 93,
36 };
37
38 struct si_llvm_diagnostics {
39 struct pipe_debug_callback *debug;
40 unsigned retval;
41 };
42
43 static void si_diagnostic_handler(LLVMDiagnosticInfoRef di, void *context)
44 {
45 struct si_llvm_diagnostics *diag = (struct si_llvm_diagnostics *)context;
46 LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
47 char *description = LLVMGetDiagInfoDescription(di);
48 const char *severity_str = NULL;
49
50 switch (severity) {
51 case LLVMDSError:
52 severity_str = "error";
53 break;
54 case LLVMDSWarning:
55 severity_str = "warning";
56 break;
57 case LLVMDSRemark:
58 severity_str = "remark";
59 break;
60 case LLVMDSNote:
61 severity_str = "note";
62 break;
63 default:
64 severity_str = "unknown";
65 }
66
67 pipe_debug_message(diag->debug, SHADER_INFO,
68 "LLVM diagnostic (%s): %s", severity_str, description);
69
70 if (severity == LLVMDSError) {
71 diag->retval = 1;
72 fprintf(stderr,"LLVM triggered Diagnostic Handler: %s\n", description);
73 }
74
75 LLVMDisposeMessage(description);
76 }
77
78 /**
79 * Compile an LLVM module to machine code.
80 *
81 * @returns 0 for success, 1 for failure
82 */
83 unsigned si_llvm_compile(LLVMModuleRef M, struct ac_shader_binary *binary,
84 struct ac_llvm_compiler *compiler,
85 struct pipe_debug_callback *debug,
86 bool less_optimized)
87 {
88 struct ac_compiler_passes *passes =
89 less_optimized && compiler->low_opt_passes ?
90 compiler->low_opt_passes : compiler->passes;
91 struct si_llvm_diagnostics diag;
92 LLVMContextRef llvm_ctx;
93
94 diag.debug = debug;
95 diag.retval = 0;
96
97 /* Setup Diagnostic Handler*/
98 llvm_ctx = LLVMGetModuleContext(M);
99
100 LLVMContextSetDiagnosticHandler(llvm_ctx, si_diagnostic_handler, &diag);
101
102 /* Compile IR. */
103 if (!ac_compile_module_to_binary(passes, M, binary))
104 diag.retval = 1;
105
106 if (diag.retval != 0)
107 pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed");
108 return diag.retval;
109 }
110
111 LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
112 enum tgsi_opcode_type type)
113 {
114 struct si_shader_context *ctx = si_shader_context(bld_base);
115
116 switch (type) {
117 case TGSI_TYPE_UNSIGNED:
118 case TGSI_TYPE_SIGNED:
119 return ctx->ac.i32;
120 case TGSI_TYPE_UNSIGNED64:
121 case TGSI_TYPE_SIGNED64:
122 return ctx->ac.i64;
123 case TGSI_TYPE_DOUBLE:
124 return ctx->ac.f64;
125 case TGSI_TYPE_UNTYPED:
126 case TGSI_TYPE_FLOAT:
127 return ctx->ac.f32;
128 default: break;
129 }
130 return 0;
131 }
132
133 LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
134 enum tgsi_opcode_type type, LLVMValueRef value)
135 {
136 struct si_shader_context *ctx = si_shader_context(bld_base);
137 LLVMTypeRef dst_type = tgsi2llvmtype(bld_base, type);
138
139 if (dst_type)
140 return LLVMBuildBitCast(ctx->ac.builder, value, dst_type, "");
141 else
142 return value;
143 }
144
145 /**
146 * Return a value that is equal to the given i32 \p index if it lies in [0,num)
147 * or an undefined value in the same interval otherwise.
148 */
149 LLVMValueRef si_llvm_bound_index(struct si_shader_context *ctx,
150 LLVMValueRef index,
151 unsigned num)
152 {
153 LLVMBuilderRef builder = ctx->ac.builder;
154 LLVMValueRef c_max = LLVMConstInt(ctx->i32, num - 1, 0);
155 LLVMValueRef cc;
156
157 if (util_is_power_of_two_or_zero(num)) {
158 index = LLVMBuildAnd(builder, index, c_max, "");
159 } else {
160 /* In theory, this MAX pattern should result in code that is
161 * as good as the bit-wise AND above.
162 *
163 * In practice, LLVM generates worse code (at the time of
164 * writing), because its value tracking is not strong enough.
165 */
166 cc = LLVMBuildICmp(builder, LLVMIntULE, index, c_max, "");
167 index = LLVMBuildSelect(builder, cc, index, c_max, "");
168 }
169
170 return index;
171 }
172
173 static LLVMValueRef emit_swizzle(struct lp_build_tgsi_context *bld_base,
174 LLVMValueRef value,
175 unsigned swizzle_x,
176 unsigned swizzle_y,
177 unsigned swizzle_z,
178 unsigned swizzle_w)
179 {
180 struct si_shader_context *ctx = si_shader_context(bld_base);
181 LLVMValueRef swizzles[4];
182
183 swizzles[0] = LLVMConstInt(ctx->i32, swizzle_x, 0);
184 swizzles[1] = LLVMConstInt(ctx->i32, swizzle_y, 0);
185 swizzles[2] = LLVMConstInt(ctx->i32, swizzle_z, 0);
186 swizzles[3] = LLVMConstInt(ctx->i32, swizzle_w, 0);
187
188 return LLVMBuildShuffleVector(ctx->ac.builder,
189 value,
190 LLVMGetUndef(LLVMTypeOf(value)),
191 LLVMConstVector(swizzles, 4), "");
192 }
193
194 /**
195 * Return the description of the array covering the given temporary register
196 * index.
197 */
198 static unsigned
199 get_temp_array_id(struct lp_build_tgsi_context *bld_base,
200 unsigned reg_index,
201 const struct tgsi_ind_register *reg)
202 {
203 struct si_shader_context *ctx = si_shader_context(bld_base);
204 unsigned num_arrays = ctx->bld_base.info->array_max[TGSI_FILE_TEMPORARY];
205 unsigned i;
206
207 if (reg && reg->ArrayID > 0 && reg->ArrayID <= num_arrays)
208 return reg->ArrayID;
209
210 for (i = 0; i < num_arrays; i++) {
211 const struct tgsi_array_info *array = &ctx->temp_arrays[i];
212
213 if (reg_index >= array->range.First && reg_index <= array->range.Last)
214 return i + 1;
215 }
216
217 return 0;
218 }
219
220 static struct tgsi_declaration_range
221 get_array_range(struct lp_build_tgsi_context *bld_base,
222 unsigned File, unsigned reg_index,
223 const struct tgsi_ind_register *reg)
224 {
225 struct si_shader_context *ctx = si_shader_context(bld_base);
226 struct tgsi_declaration_range range;
227
228 if (File == TGSI_FILE_TEMPORARY) {
229 unsigned array_id = get_temp_array_id(bld_base, reg_index, reg);
230 if (array_id)
231 return ctx->temp_arrays[array_id - 1].range;
232 }
233
234 range.First = 0;
235 range.Last = bld_base->info->file_max[File];
236 return range;
237 }
238
239 /**
240 * For indirect registers, construct a pointer directly to the requested
241 * element using getelementptr if possible.
242 *
243 * Returns NULL if the insertelement/extractelement fallback for array access
244 * must be used.
245 */
246 static LLVMValueRef
247 get_pointer_into_array(struct si_shader_context *ctx,
248 unsigned file,
249 unsigned swizzle,
250 unsigned reg_index,
251 const struct tgsi_ind_register *reg_indirect)
252 {
253 unsigned array_id;
254 struct tgsi_array_info *array;
255 LLVMValueRef idxs[2];
256 LLVMValueRef index;
257 LLVMValueRef alloca;
258
259 if (file != TGSI_FILE_TEMPORARY)
260 return NULL;
261
262 array_id = get_temp_array_id(&ctx->bld_base, reg_index, reg_indirect);
263 if (!array_id)
264 return NULL;
265
266 alloca = ctx->temp_array_allocas[array_id - 1];
267 if (!alloca)
268 return NULL;
269
270 array = &ctx->temp_arrays[array_id - 1];
271
272 if (!(array->writemask & (1 << swizzle)))
273 return ctx->undef_alloca;
274
275 index = si_get_indirect_index(ctx, reg_indirect, 1,
276 reg_index - ctx->temp_arrays[array_id - 1].range.First);
277
278 /* Ensure that the index is within a valid range, to guard against
279 * VM faults and overwriting critical data (e.g. spilled resource
280 * descriptors).
281 *
282 * TODO It should be possible to avoid the additional instructions
283 * if LLVM is changed so that it guarantuees:
284 * 1. the scratch space descriptor isolates the current wave (this
285 * could even save the scratch offset SGPR at the cost of an
286 * additional SALU instruction)
287 * 2. the memory for allocas must be allocated at the _end_ of the
288 * scratch space (after spilled registers)
289 */
290 index = si_llvm_bound_index(ctx, index, array->range.Last - array->range.First + 1);
291
292 index = ac_build_imad(&ctx->ac, index,
293 LLVMConstInt(ctx->i32, util_bitcount(array->writemask), 0),
294 LLVMConstInt(ctx->i32,
295 util_bitcount(array->writemask & ((1 << swizzle) - 1)), 0));
296 idxs[0] = ctx->i32_0;
297 idxs[1] = index;
298 return LLVMBuildGEP(ctx->ac.builder, alloca, idxs, 2, "");
299 }
300
301 LLVMValueRef
302 si_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
303 LLVMTypeRef type,
304 LLVMValueRef ptr,
305 LLVMValueRef ptr2)
306 {
307 struct si_shader_context *ctx = si_shader_context(bld_base);
308 LLVMValueRef result;
309
310 result = LLVMGetUndef(LLVMVectorType(ctx->i32, 2));
311
312 result = LLVMBuildInsertElement(ctx->ac.builder,
313 result,
314 ac_to_integer(&ctx->ac, ptr),
315 ctx->i32_0, "");
316 result = LLVMBuildInsertElement(ctx->ac.builder,
317 result,
318 ac_to_integer(&ctx->ac, ptr2),
319 ctx->i32_1, "");
320 return LLVMBuildBitCast(ctx->ac.builder, result, type, "");
321 }
322
323 static LLVMValueRef
324 emit_array_fetch(struct lp_build_tgsi_context *bld_base,
325 unsigned File, enum tgsi_opcode_type type,
326 struct tgsi_declaration_range range,
327 unsigned swizzle)
328 {
329 struct si_shader_context *ctx = si_shader_context(bld_base);
330 unsigned i, size = range.Last - range.First + 1;
331 LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
332 LLVMValueRef result = LLVMGetUndef(vec);
333
334 struct tgsi_full_src_register tmp_reg = {};
335 tmp_reg.Register.File = File;
336
337 for (i = 0; i < size; ++i) {
338 tmp_reg.Register.Index = i + range.First;
339 LLVMValueRef temp = si_llvm_emit_fetch(bld_base, &tmp_reg, type, swizzle);
340 result = LLVMBuildInsertElement(ctx->ac.builder, result, temp,
341 LLVMConstInt(ctx->i32, i, 0), "array_vector");
342 }
343 return result;
344 }
345
346 static LLVMValueRef
347 load_value_from_array(struct lp_build_tgsi_context *bld_base,
348 unsigned file,
349 enum tgsi_opcode_type type,
350 unsigned swizzle,
351 unsigned reg_index,
352 const struct tgsi_ind_register *reg_indirect)
353 {
354 struct si_shader_context *ctx = si_shader_context(bld_base);
355 LLVMBuilderRef builder = ctx->ac.builder;
356 LLVMValueRef ptr;
357
358 ptr = get_pointer_into_array(ctx, file, swizzle, reg_index, reg_indirect);
359 if (ptr) {
360 LLVMValueRef val = LLVMBuildLoad(builder, ptr, "");
361 if (tgsi_type_is_64bit(type)) {
362 LLVMValueRef ptr_hi, val_hi;
363 ptr_hi = LLVMBuildGEP(builder, ptr, &ctx->i32_1, 1, "");
364 val_hi = LLVMBuildLoad(builder, ptr_hi, "");
365 val = si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
366 val, val_hi);
367 }
368
369 return val;
370 } else {
371 struct tgsi_declaration_range range =
372 get_array_range(bld_base, file, reg_index, reg_indirect);
373 LLVMValueRef index =
374 si_get_indirect_index(ctx, reg_indirect, 1, reg_index - range.First);
375 LLVMValueRef array =
376 emit_array_fetch(bld_base, file, type, range, swizzle);
377 return LLVMBuildExtractElement(builder, array, index, "");
378 }
379 }
380
381 static void
382 store_value_to_array(struct lp_build_tgsi_context *bld_base,
383 LLVMValueRef value,
384 unsigned file,
385 unsigned chan_index,
386 unsigned reg_index,
387 const struct tgsi_ind_register *reg_indirect)
388 {
389 struct si_shader_context *ctx = si_shader_context(bld_base);
390 LLVMBuilderRef builder = ctx->ac.builder;
391 LLVMValueRef ptr;
392
393 ptr = get_pointer_into_array(ctx, file, chan_index, reg_index, reg_indirect);
394 if (ptr) {
395 LLVMBuildStore(builder, value, ptr);
396 } else {
397 unsigned i, size;
398 struct tgsi_declaration_range range = get_array_range(bld_base, file, reg_index, reg_indirect);
399 LLVMValueRef index = si_get_indirect_index(ctx, reg_indirect, 1, reg_index - range.First);
400 LLVMValueRef array =
401 emit_array_fetch(bld_base, file, TGSI_TYPE_FLOAT, range, chan_index);
402 LLVMValueRef temp_ptr;
403
404 array = LLVMBuildInsertElement(builder, array, value, index, "");
405
406 size = range.Last - range.First + 1;
407 for (i = 0; i < size; ++i) {
408 switch(file) {
409 case TGSI_FILE_OUTPUT:
410 temp_ptr = ctx->outputs[i + range.First][chan_index];
411 break;
412
413 case TGSI_FILE_TEMPORARY:
414 if (range.First + i >= ctx->temps_count)
415 continue;
416 temp_ptr = ctx->temps[(i + range.First) * TGSI_NUM_CHANNELS + chan_index];
417 break;
418
419 default:
420 continue;
421 }
422 value = LLVMBuildExtractElement(builder, array,
423 LLVMConstInt(ctx->i32, i, 0), "");
424 LLVMBuildStore(builder, value, temp_ptr);
425 }
426 }
427 }
428
429 /* If this is true, preload FS inputs at the beginning of shaders. Otherwise,
430 * reload them at each use. This must be true if the shader is using
431 * derivatives and KILL, because KILL can leave the WQM and then a lazy
432 * input load isn't in the WQM anymore.
433 */
434 static bool si_preload_fs_inputs(struct si_shader_context *ctx)
435 {
436 struct si_shader_selector *sel = ctx->shader->selector;
437
438 return sel->info.uses_derivatives &&
439 sel->info.uses_kill;
440 }
441
442 static LLVMValueRef
443 get_output_ptr(struct lp_build_tgsi_context *bld_base, unsigned index,
444 unsigned chan)
445 {
446 struct si_shader_context *ctx = si_shader_context(bld_base);
447
448 assert(index <= ctx->bld_base.info->file_max[TGSI_FILE_OUTPUT]);
449 return ctx->outputs[index][chan];
450 }
451
452 LLVMValueRef si_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
453 const struct tgsi_full_src_register *reg,
454 enum tgsi_opcode_type type,
455 unsigned swizzle)
456 {
457 struct si_shader_context *ctx = si_shader_context(bld_base);
458 LLVMBuilderRef builder = ctx->ac.builder;
459 LLVMValueRef result = NULL, ptr, ptr2;
460
461 if (swizzle == ~0) {
462 LLVMValueRef values[TGSI_NUM_CHANNELS];
463 unsigned chan;
464 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
465 values[chan] = si_llvm_emit_fetch(bld_base, reg, type, chan);
466 }
467 return ac_build_gather_values(&ctx->ac, values,
468 TGSI_NUM_CHANNELS);
469 }
470
471 if (reg->Register.Indirect) {
472 LLVMValueRef load = load_value_from_array(bld_base, reg->Register.File, type,
473 swizzle, reg->Register.Index, &reg->Indirect);
474 return bitcast(bld_base, type, load);
475 }
476
477 switch(reg->Register.File) {
478 case TGSI_FILE_IMMEDIATE: {
479 LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
480 if (tgsi_type_is_64bit(type)) {
481 result = LLVMGetUndef(LLVMVectorType(ctx->i32, 2));
482 result = LLVMConstInsertElement(result,
483 ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle],
484 ctx->i32_0);
485 result = LLVMConstInsertElement(result,
486 ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1],
487 ctx->i32_1);
488 return LLVMConstBitCast(result, ctype);
489 } else {
490 return LLVMConstBitCast(ctx->imms[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle], ctype);
491 }
492 }
493
494 case TGSI_FILE_INPUT: {
495 unsigned index = reg->Register.Index;
496 LLVMValueRef input[4];
497
498 /* I don't think doing this for vertex shaders is beneficial.
499 * For those, we want to make sure the VMEM loads are executed
500 * only once. Fragment shaders don't care much, because
501 * v_interp instructions are much cheaper than VMEM loads.
502 */
503 if (!si_preload_fs_inputs(ctx) &&
504 ctx->bld_base.info->processor == PIPE_SHADER_FRAGMENT)
505 ctx->load_input(ctx, index, &ctx->input_decls[index], input);
506 else
507 memcpy(input, &ctx->inputs[index * 4], sizeof(input));
508
509 result = input[swizzle];
510
511 if (tgsi_type_is_64bit(type)) {
512 ptr = result;
513 ptr2 = input[swizzle + 1];
514 return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
515 ptr, ptr2);
516 }
517 break;
518 }
519
520 case TGSI_FILE_TEMPORARY:
521 if (reg->Register.Index >= ctx->temps_count)
522 return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
523 ptr = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle];
524 if (tgsi_type_is_64bit(type)) {
525 ptr2 = ctx->temps[reg->Register.Index * TGSI_NUM_CHANNELS + swizzle + 1];
526 return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
527 LLVMBuildLoad(builder, ptr, ""),
528 LLVMBuildLoad(builder, ptr2, ""));
529 }
530 result = LLVMBuildLoad(builder, ptr, "");
531 break;
532
533 case TGSI_FILE_OUTPUT:
534 ptr = get_output_ptr(bld_base, reg->Register.Index, swizzle);
535 if (tgsi_type_is_64bit(type)) {
536 ptr2 = get_output_ptr(bld_base, reg->Register.Index, swizzle + 1);
537 return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
538 LLVMBuildLoad(builder, ptr, ""),
539 LLVMBuildLoad(builder, ptr2, ""));
540 }
541 result = LLVMBuildLoad(builder, ptr, "");
542 break;
543
544 default:
545 return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
546 }
547
548 return bitcast(bld_base, type, result);
549 }
550
551 static LLVMValueRef fetch_system_value(struct lp_build_tgsi_context *bld_base,
552 const struct tgsi_full_src_register *reg,
553 enum tgsi_opcode_type type,
554 unsigned swizzle)
555 {
556 struct si_shader_context *ctx = si_shader_context(bld_base);
557 LLVMBuilderRef builder = ctx->ac.builder;
558 LLVMValueRef cval = ctx->system_values[reg->Register.Index];
559
560 if (tgsi_type_is_64bit(type)) {
561 LLVMValueRef lo, hi;
562
563 assert(swizzle == 0 || swizzle == 2);
564
565 lo = LLVMBuildExtractElement(
566 builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
567 hi = LLVMBuildExtractElement(
568 builder, cval, LLVMConstInt(ctx->i32, swizzle + 1, 0), "");
569
570 return si_llvm_emit_fetch_64bit(bld_base, tgsi2llvmtype(bld_base, type),
571 lo, hi);
572 }
573
574 if (LLVMGetTypeKind(LLVMTypeOf(cval)) == LLVMVectorTypeKind) {
575 cval = LLVMBuildExtractElement(
576 builder, cval, LLVMConstInt(ctx->i32, swizzle, 0), "");
577 } else {
578 assert(swizzle == 0);
579 }
580
581 return bitcast(bld_base, type, cval);
582 }
583
584 static void emit_declaration(struct lp_build_tgsi_context *bld_base,
585 const struct tgsi_full_declaration *decl)
586 {
587 struct si_shader_context *ctx = si_shader_context(bld_base);
588 LLVMBuilderRef builder = ctx->ac.builder;
589 unsigned first, last, i;
590 switch(decl->Declaration.File) {
591 case TGSI_FILE_ADDRESS:
592 {
593 unsigned idx;
594 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
595 unsigned chan;
596 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
597 ctx->addrs[idx][chan] = ac_build_alloca_undef(
598 &ctx->ac, ctx->i32, "");
599 }
600 }
601 break;
602 }
603
604 case TGSI_FILE_TEMPORARY:
605 {
606 char name[18] = "";
607 LLVMValueRef array_alloca = NULL;
608 unsigned decl_size;
609 unsigned writemask = decl->Declaration.UsageMask;
610 first = decl->Range.First;
611 last = decl->Range.Last;
612 decl_size = 4 * ((last - first) + 1);
613
614 if (decl->Declaration.Array) {
615 unsigned id = decl->Array.ArrayID - 1;
616 unsigned array_size;
617
618 writemask &= ctx->temp_arrays[id].writemask;
619 ctx->temp_arrays[id].writemask = writemask;
620 array_size = ((last - first) + 1) * util_bitcount(writemask);
621
622 /* If the array has more than 16 elements, store it
623 * in memory using an alloca that spans the entire
624 * array.
625 *
626 * Otherwise, store each array element individually.
627 * We will then generate vectors (per-channel, up to
628 * <16 x float> if the usagemask is a single bit) for
629 * indirect addressing.
630 *
631 * Note that 16 is the number of vector elements that
632 * LLVM will store in a register, so theoretically an
633 * array with up to 4 * 16 = 64 elements could be
634 * handled this way, but whether that's a good idea
635 * depends on VGPR register pressure elsewhere.
636 *
637 * FIXME: We shouldn't need to have the non-alloca
638 * code path for arrays. LLVM should be smart enough to
639 * promote allocas into registers when profitable.
640 */
641 if (array_size > 16 ||
642 !ctx->screen->llvm_has_working_vgpr_indexing) {
643 array_alloca = ac_build_alloca_undef(&ctx->ac,
644 LLVMArrayType(ctx->f32,
645 array_size), "array");
646 ctx->temp_array_allocas[id] = array_alloca;
647 }
648 }
649
650 if (!ctx->temps_count) {
651 ctx->temps_count = bld_base->info->file_max[TGSI_FILE_TEMPORARY] + 1;
652 ctx->temps = MALLOC(TGSI_NUM_CHANNELS * ctx->temps_count * sizeof(LLVMValueRef));
653 }
654 if (!array_alloca) {
655 for (i = 0; i < decl_size; ++i) {
656 #ifdef DEBUG
657 snprintf(name, sizeof(name), "TEMP%d.%c",
658 first + i / 4, "xyzw"[i % 4]);
659 #endif
660 ctx->temps[first * TGSI_NUM_CHANNELS + i] =
661 ac_build_alloca_undef(&ctx->ac,
662 ctx->f32,
663 name);
664 }
665 } else {
666 LLVMValueRef idxs[2] = {
667 ctx->i32_0,
668 NULL
669 };
670 unsigned j = 0;
671
672 if (writemask != TGSI_WRITEMASK_XYZW &&
673 !ctx->undef_alloca) {
674 /* Create a dummy alloca. We use it so that we
675 * have a pointer that is safe to load from if
676 * a shader ever reads from a channel that
677 * it never writes to.
678 */
679 ctx->undef_alloca = ac_build_alloca_undef(
680 &ctx->ac, ctx->f32, "undef");
681 }
682
683 for (i = 0; i < decl_size; ++i) {
684 LLVMValueRef ptr;
685 if (writemask & (1 << (i % 4))) {
686 #ifdef DEBUG
687 snprintf(name, sizeof(name), "TEMP%d.%c",
688 first + i / 4, "xyzw"[i % 4]);
689 #endif
690 idxs[1] = LLVMConstInt(ctx->i32, j, 0);
691 ptr = LLVMBuildGEP(builder, array_alloca, idxs, 2, name);
692 j++;
693 } else {
694 ptr = ctx->undef_alloca;
695 }
696 ctx->temps[first * TGSI_NUM_CHANNELS + i] = ptr;
697 }
698 }
699 break;
700 }
701 case TGSI_FILE_INPUT:
702 {
703 unsigned idx;
704 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
705 if (ctx->load_input &&
706 ctx->input_decls[idx].Declaration.File != TGSI_FILE_INPUT) {
707 ctx->input_decls[idx] = *decl;
708 ctx->input_decls[idx].Range.First = idx;
709 ctx->input_decls[idx].Range.Last = idx;
710 ctx->input_decls[idx].Semantic.Index += idx - decl->Range.First;
711
712 if (si_preload_fs_inputs(ctx) ||
713 bld_base->info->processor != PIPE_SHADER_FRAGMENT)
714 ctx->load_input(ctx, idx, &ctx->input_decls[idx],
715 &ctx->inputs[idx * 4]);
716 }
717 }
718 }
719 break;
720
721 case TGSI_FILE_SYSTEM_VALUE:
722 {
723 unsigned idx;
724 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
725 si_load_system_value(ctx, idx, decl);
726 }
727 }
728 break;
729
730 case TGSI_FILE_OUTPUT:
731 {
732 char name[16] = "";
733 unsigned idx;
734 for (idx = decl->Range.First; idx <= decl->Range.Last; idx++) {
735 unsigned chan;
736 assert(idx < RADEON_LLVM_MAX_OUTPUTS);
737 if (ctx->outputs[idx][0])
738 continue;
739 for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
740 #ifdef DEBUG
741 snprintf(name, sizeof(name), "OUT%d.%c",
742 idx, "xyzw"[chan % 4]);
743 #endif
744 ctx->outputs[idx][chan] = ac_build_alloca_undef(
745 &ctx->ac, ctx->f32, name);
746 }
747 }
748 break;
749 }
750
751 case TGSI_FILE_MEMORY:
752 si_tgsi_declare_compute_memory(ctx, decl);
753 break;
754
755 default:
756 break;
757 }
758 }
759
760 void si_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
761 const struct tgsi_full_instruction *inst,
762 const struct tgsi_opcode_info *info,
763 unsigned index,
764 LLVMValueRef dst[4])
765 {
766 struct si_shader_context *ctx = si_shader_context(bld_base);
767 const struct tgsi_full_dst_register *reg = &inst->Dst[index];
768 LLVMBuilderRef builder = ctx->ac.builder;
769 LLVMValueRef temp_ptr, temp_ptr2 = NULL;
770 bool is_vec_store = false;
771 enum tgsi_opcode_type dtype = tgsi_opcode_infer_dst_type(inst->Instruction.Opcode, index);
772
773 if (dst[0]) {
774 LLVMTypeKind k = LLVMGetTypeKind(LLVMTypeOf(dst[0]));
775 is_vec_store = (k == LLVMVectorTypeKind);
776 }
777
778 if (is_vec_store) {
779 LLVMValueRef values[4] = {};
780 uint32_t writemask = reg->Register.WriteMask;
781 while (writemask) {
782 unsigned chan = u_bit_scan(&writemask);
783 LLVMValueRef index = LLVMConstInt(ctx->i32, chan, 0);
784 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
785 dst[0], index, "");
786 }
787 bld_base->emit_store(bld_base, inst, info, index, values);
788 return;
789 }
790
791 uint32_t writemask = reg->Register.WriteMask;
792 while (writemask) {
793 unsigned chan_index = u_bit_scan(&writemask);
794 LLVMValueRef value = dst[chan_index];
795
796 if (tgsi_type_is_64bit(dtype) && (chan_index == 1 || chan_index == 3))
797 continue;
798 if (inst->Instruction.Saturate)
799 value = ac_build_clamp(&ctx->ac, value);
800
801 if (reg->Register.File == TGSI_FILE_ADDRESS) {
802 temp_ptr = ctx->addrs[reg->Register.Index][chan_index];
803 LLVMBuildStore(builder, value, temp_ptr);
804 continue;
805 }
806
807 if (!tgsi_type_is_64bit(dtype))
808 value = ac_to_float(&ctx->ac, value);
809
810 if (reg->Register.Indirect) {
811 unsigned file = reg->Register.File;
812 unsigned reg_index = reg->Register.Index;
813 store_value_to_array(bld_base, value, file, chan_index,
814 reg_index, &reg->Indirect);
815 } else {
816 switch(reg->Register.File) {
817 case TGSI_FILE_OUTPUT:
818 temp_ptr = ctx->outputs[reg->Register.Index][chan_index];
819 if (tgsi_type_is_64bit(dtype))
820 temp_ptr2 = ctx->outputs[reg->Register.Index][chan_index + 1];
821 break;
822
823 case TGSI_FILE_TEMPORARY:
824 {
825 if (reg->Register.Index >= ctx->temps_count)
826 continue;
827
828 temp_ptr = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index];
829 if (tgsi_type_is_64bit(dtype))
830 temp_ptr2 = ctx->temps[ TGSI_NUM_CHANNELS * reg->Register.Index + chan_index + 1];
831
832 break;
833 }
834 default:
835 return;
836 }
837 if (!tgsi_type_is_64bit(dtype))
838 LLVMBuildStore(builder, value, temp_ptr);
839 else {
840 LLVMValueRef ptr = LLVMBuildBitCast(builder, value,
841 LLVMVectorType(ctx->i32, 2), "");
842 LLVMValueRef val2;
843 value = LLVMBuildExtractElement(builder, ptr,
844 ctx->i32_0, "");
845 val2 = LLVMBuildExtractElement(builder, ptr,
846 ctx->i32_1, "");
847
848 LLVMBuildStore(builder, ac_to_float(&ctx->ac, value), temp_ptr);
849 LLVMBuildStore(builder, ac_to_float(&ctx->ac, val2), temp_ptr2);
850 }
851 }
852 }
853 }
854
855 static int get_line(int pc)
856 {
857 /* Subtract 1 so that the number shown is that of the corresponding
858 * opcode in the TGSI dump, e.g. an if block has the same suffix as
859 * the instruction number of the corresponding TGSI IF.
860 */
861 return pc - 1;
862 }
863
864 static void bgnloop_emit(const struct lp_build_tgsi_action *action,
865 struct lp_build_tgsi_context *bld_base,
866 struct lp_build_emit_data *emit_data)
867 {
868 struct si_shader_context *ctx = si_shader_context(bld_base);
869 ac_build_bgnloop(&ctx->ac, get_line(bld_base->pc));
870 }
871
872 static void brk_emit(const struct lp_build_tgsi_action *action,
873 struct lp_build_tgsi_context *bld_base,
874 struct lp_build_emit_data *emit_data)
875 {
876 struct si_shader_context *ctx = si_shader_context(bld_base);
877 ac_build_break(&ctx->ac);
878 }
879
880 static void cont_emit(const struct lp_build_tgsi_action *action,
881 struct lp_build_tgsi_context *bld_base,
882 struct lp_build_emit_data *emit_data)
883 {
884 struct si_shader_context *ctx = si_shader_context(bld_base);
885 ac_build_continue(&ctx->ac);
886 }
887
888 static void else_emit(const struct lp_build_tgsi_action *action,
889 struct lp_build_tgsi_context *bld_base,
890 struct lp_build_emit_data *emit_data)
891 {
892 struct si_shader_context *ctx = si_shader_context(bld_base);
893 ac_build_else(&ctx->ac, get_line(bld_base->pc));
894 }
895
896 static void endif_emit(const struct lp_build_tgsi_action *action,
897 struct lp_build_tgsi_context *bld_base,
898 struct lp_build_emit_data *emit_data)
899 {
900 struct si_shader_context *ctx = si_shader_context(bld_base);
901 ac_build_endif(&ctx->ac, get_line(bld_base->pc));
902 }
903
904 static void endloop_emit(const struct lp_build_tgsi_action *action,
905 struct lp_build_tgsi_context *bld_base,
906 struct lp_build_emit_data *emit_data)
907 {
908 struct si_shader_context *ctx = si_shader_context(bld_base);
909 ac_build_endloop(&ctx->ac, get_line(bld_base->pc));
910 }
911
912 static void if_emit(const struct lp_build_tgsi_action *action,
913 struct lp_build_tgsi_context *bld_base,
914 struct lp_build_emit_data *emit_data)
915 {
916 struct si_shader_context *ctx = si_shader_context(bld_base);
917 ac_build_if(&ctx->ac, emit_data->args[0], get_line(bld_base->pc));
918 }
919
920 static void uif_emit(const struct lp_build_tgsi_action *action,
921 struct lp_build_tgsi_context *bld_base,
922 struct lp_build_emit_data *emit_data)
923 {
924 struct si_shader_context *ctx = si_shader_context(bld_base);
925 ac_build_uif(&ctx->ac, emit_data->args[0], get_line(bld_base->pc));
926 }
927
928 static void emit_immediate(struct lp_build_tgsi_context *bld_base,
929 const struct tgsi_full_immediate *imm)
930 {
931 unsigned i;
932 struct si_shader_context *ctx = si_shader_context(bld_base);
933
934 for (i = 0; i < 4; ++i) {
935 ctx->imms[ctx->imms_num * TGSI_NUM_CHANNELS + i] =
936 LLVMConstInt(ctx->i32, imm->u[i].Uint, false );
937 }
938
939 ctx->imms_num++;
940 }
941
942 void si_llvm_context_init(struct si_shader_context *ctx,
943 struct si_screen *sscreen,
944 struct ac_llvm_compiler *compiler)
945 {
946 struct lp_type type;
947
948 /* Initialize the gallivm object:
949 * We are only using the module, context, and builder fields of this struct.
950 * This should be enough for us to be able to pass our gallivm struct to the
951 * helper functions in the gallivm module.
952 */
953 memset(ctx, 0, sizeof(*ctx));
954 ctx->screen = sscreen;
955 ctx->compiler = compiler;
956
957 ac_llvm_context_init(&ctx->ac, sscreen->info.chip_class, sscreen->info.family);
958 ctx->ac.module = ac_create_module(compiler->tm, ctx->ac.context);
959
960 enum ac_float_mode float_mode =
961 sscreen->debug_flags & DBG(UNSAFE_MATH) ?
962 AC_FLOAT_MODE_UNSAFE_FP_MATH :
963 AC_FLOAT_MODE_NO_SIGNED_ZEROS_FP_MATH;
964 ctx->ac.builder = ac_create_builder(ctx->ac.context, float_mode);
965
966 ctx->gallivm.context = ctx->ac.context;
967 ctx->gallivm.module = ctx->ac.module;
968 ctx->gallivm.builder = ctx->ac.builder;
969
970 struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
971
972 type.floating = true;
973 type.fixed = false;
974 type.sign = true;
975 type.norm = false;
976 type.width = 32;
977 type.length = 1;
978
979 lp_build_context_init(&bld_base->base, &ctx->gallivm, type);
980 lp_build_context_init(&ctx->bld_base.uint_bld, &ctx->gallivm, lp_uint_type(type));
981 lp_build_context_init(&ctx->bld_base.int_bld, &ctx->gallivm, lp_int_type(type));
982 type.width *= 2;
983 lp_build_context_init(&ctx->bld_base.dbl_bld, &ctx->gallivm, type);
984 lp_build_context_init(&ctx->bld_base.uint64_bld, &ctx->gallivm, lp_uint_type(type));
985 lp_build_context_init(&ctx->bld_base.int64_bld, &ctx->gallivm, lp_int_type(type));
986
987 bld_base->soa = 1;
988 bld_base->emit_swizzle = emit_swizzle;
989 bld_base->emit_declaration = emit_declaration;
990 bld_base->emit_immediate = emit_immediate;
991
992 bld_base->op_actions[TGSI_OPCODE_BGNLOOP].emit = bgnloop_emit;
993 bld_base->op_actions[TGSI_OPCODE_BRK].emit = brk_emit;
994 bld_base->op_actions[TGSI_OPCODE_CONT].emit = cont_emit;
995 bld_base->op_actions[TGSI_OPCODE_IF].emit = if_emit;
996 bld_base->op_actions[TGSI_OPCODE_UIF].emit = uif_emit;
997 bld_base->op_actions[TGSI_OPCODE_ELSE].emit = else_emit;
998 bld_base->op_actions[TGSI_OPCODE_ENDIF].emit = endif_emit;
999 bld_base->op_actions[TGSI_OPCODE_ENDLOOP].emit = endloop_emit;
1000
1001 si_shader_context_init_alu(&ctx->bld_base);
1002 si_shader_context_init_mem(ctx);
1003
1004 ctx->voidt = LLVMVoidTypeInContext(ctx->ac.context);
1005 ctx->i1 = LLVMInt1TypeInContext(ctx->ac.context);
1006 ctx->i8 = LLVMInt8TypeInContext(ctx->ac.context);
1007 ctx->i32 = LLVMInt32TypeInContext(ctx->ac.context);
1008 ctx->i64 = LLVMInt64TypeInContext(ctx->ac.context);
1009 ctx->i128 = LLVMIntTypeInContext(ctx->ac.context, 128);
1010 ctx->f32 = LLVMFloatTypeInContext(ctx->ac.context);
1011 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
1012 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
1013 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
1014 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
1015
1016 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, 0);
1017 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, 0);
1018 ctx->i1false = LLVMConstInt(ctx->i1, 0, 0);
1019 ctx->i1true = LLVMConstInt(ctx->i1, 1, 0);
1020 }
1021
1022 /* Set the context to a certain TGSI shader. Can be called repeatedly
1023 * to change the shader. */
1024 void si_llvm_context_set_tgsi(struct si_shader_context *ctx,
1025 struct si_shader *shader)
1026 {
1027 const struct tgsi_shader_info *info = NULL;
1028 const struct tgsi_token *tokens = NULL;
1029
1030 if (shader && shader->selector) {
1031 info = &shader->selector->info;
1032 tokens = shader->selector->tokens;
1033 }
1034
1035 ctx->shader = shader;
1036 ctx->type = info ? info->processor : -1;
1037 ctx->bld_base.info = info;
1038
1039 /* Clean up the old contents. */
1040 FREE(ctx->temp_arrays);
1041 ctx->temp_arrays = NULL;
1042 FREE(ctx->temp_array_allocas);
1043 ctx->temp_array_allocas = NULL;
1044
1045 FREE(ctx->imms);
1046 ctx->imms = NULL;
1047 ctx->imms_num = 0;
1048
1049 FREE(ctx->temps);
1050 ctx->temps = NULL;
1051 ctx->temps_count = 0;
1052
1053 if (!info)
1054 return;
1055
1056 ctx->num_const_buffers = util_last_bit(info->const_buffers_declared);
1057 ctx->num_shader_buffers = util_last_bit(info->shader_buffers_declared);
1058
1059 ctx->num_samplers = util_last_bit(info->samplers_declared);
1060 ctx->num_images = util_last_bit(info->images_declared);
1061
1062 if (!tokens)
1063 return;
1064
1065 if (info->array_max[TGSI_FILE_TEMPORARY] > 0) {
1066 int size = info->array_max[TGSI_FILE_TEMPORARY];
1067
1068 ctx->temp_arrays = CALLOC(size, sizeof(ctx->temp_arrays[0]));
1069 ctx->temp_array_allocas = CALLOC(size, sizeof(ctx->temp_array_allocas[0]));
1070
1071 tgsi_scan_arrays(tokens, TGSI_FILE_TEMPORARY, size,
1072 ctx->temp_arrays);
1073 }
1074 if (info->file_max[TGSI_FILE_IMMEDIATE] >= 0) {
1075 int size = info->file_max[TGSI_FILE_IMMEDIATE] + 1;
1076 ctx->imms = MALLOC(size * TGSI_NUM_CHANNELS * sizeof(LLVMValueRef));
1077 }
1078
1079 /* Re-set these to start with a clean slate. */
1080 ctx->bld_base.num_instructions = 0;
1081 ctx->bld_base.pc = 0;
1082 memset(ctx->outputs, 0, sizeof(ctx->outputs));
1083
1084 ctx->bld_base.emit_store = si_llvm_emit_store;
1085 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = si_llvm_emit_fetch;
1086 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = si_llvm_emit_fetch;
1087 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = si_llvm_emit_fetch;
1088 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_OUTPUT] = si_llvm_emit_fetch;
1089 ctx->bld_base.emit_fetch_funcs[TGSI_FILE_SYSTEM_VALUE] = fetch_system_value;
1090 }
1091
1092 void si_llvm_create_func(struct si_shader_context *ctx,
1093 const char *name,
1094 LLVMTypeRef *return_types, unsigned num_return_elems,
1095 LLVMTypeRef *ParamTypes, unsigned ParamCount)
1096 {
1097 LLVMTypeRef main_fn_type, ret_type;
1098 LLVMBasicBlockRef main_fn_body;
1099 enum si_llvm_calling_convention call_conv;
1100 unsigned real_shader_type;
1101
1102 if (num_return_elems)
1103 ret_type = LLVMStructTypeInContext(ctx->ac.context,
1104 return_types,
1105 num_return_elems, true);
1106 else
1107 ret_type = ctx->voidt;
1108
1109 /* Setup the function */
1110 ctx->return_type = ret_type;
1111 main_fn_type = LLVMFunctionType(ret_type, ParamTypes, ParamCount, 0);
1112 ctx->main_fn = LLVMAddFunction(ctx->gallivm.module, name, main_fn_type);
1113 main_fn_body = LLVMAppendBasicBlockInContext(ctx->ac.context,
1114 ctx->main_fn, "main_body");
1115 LLVMPositionBuilderAtEnd(ctx->ac.builder, main_fn_body);
1116
1117 real_shader_type = ctx->type;
1118
1119 /* LS is merged into HS (TCS), and ES is merged into GS. */
1120 if (ctx->screen->info.chip_class >= GFX9) {
1121 if (ctx->shader->key.as_ls)
1122 real_shader_type = PIPE_SHADER_TESS_CTRL;
1123 else if (ctx->shader->key.as_es)
1124 real_shader_type = PIPE_SHADER_GEOMETRY;
1125 }
1126
1127 switch (real_shader_type) {
1128 case PIPE_SHADER_VERTEX:
1129 case PIPE_SHADER_TESS_EVAL:
1130 call_conv = RADEON_LLVM_AMDGPU_VS;
1131 break;
1132 case PIPE_SHADER_TESS_CTRL:
1133 call_conv = RADEON_LLVM_AMDGPU_HS;
1134 break;
1135 case PIPE_SHADER_GEOMETRY:
1136 call_conv = RADEON_LLVM_AMDGPU_GS;
1137 break;
1138 case PIPE_SHADER_FRAGMENT:
1139 call_conv = RADEON_LLVM_AMDGPU_PS;
1140 break;
1141 case PIPE_SHADER_COMPUTE:
1142 call_conv = RADEON_LLVM_AMDGPU_CS;
1143 break;
1144 default:
1145 unreachable("Unhandle shader type");
1146 }
1147
1148 LLVMSetFunctionCallConv(ctx->main_fn, call_conv);
1149 }
1150
1151 void si_llvm_optimize_module(struct si_shader_context *ctx)
1152 {
1153 /* Dump LLVM IR before any optimization passes */
1154 if (ctx->screen->debug_flags & DBG(PREOPT_IR) &&
1155 si_can_dump_shader(ctx->screen, ctx->type))
1156 LLVMDumpModule(ctx->gallivm.module);
1157
1158 /* Run the pass */
1159 LLVMRunPassManager(ctx->compiler->passmgr, ctx->gallivm.module);
1160 LLVMDisposeBuilder(ctx->ac.builder);
1161 }
1162
1163 void si_llvm_dispose(struct si_shader_context *ctx)
1164 {
1165 LLVMDisposeModule(ctx->gallivm.module);
1166 LLVMContextDispose(ctx->gallivm.context);
1167 FREE(ctx->temp_arrays);
1168 ctx->temp_arrays = NULL;
1169 FREE(ctx->temp_array_allocas);
1170 ctx->temp_array_allocas = NULL;
1171 FREE(ctx->temps);
1172 ctx->temps = NULL;
1173 ctx->temps_count = 0;
1174 FREE(ctx->imms);
1175 ctx->imms = NULL;
1176 ctx->imms_num = 0;
1177 ac_llvm_context_dispose(&ctx->ac);
1178 }