projects / mesa.git / blob
? search:


15d6759ced16f3fe768f7bc9dd5207e8c28de345
[mesa.git] / src / amd / common / ac_nir_to_llvm.c
1 /*
2 * Copyright © 2016 Bas Nieuwenhuizen
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * 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 NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include "ac_nir_to_llvm.h"
25 #include "ac_llvm_build.h"
26 #include "ac_llvm_util.h"
27 #include "ac_binary.h"
28 #include "sid.h"
29 #include "nir/nir.h"
30 #include "../vulkan/radv_descriptor_set.h"
31 #include "util/bitscan.h"
32 #include <llvm-c/Transforms/Scalar.h>
33 #include "ac_shader_abi.h"
34 #include "ac_shader_info.h"
35 #include "ac_shader_util.h"
36 #include "ac_exp_param.h"
37
38 enum radeon_llvm_calling_convention {
39 RADEON_LLVM_AMDGPU_VS = 87,
40 RADEON_LLVM_AMDGPU_GS = 88,
41 RADEON_LLVM_AMDGPU_PS = 89,
42 RADEON_LLVM_AMDGPU_CS = 90,
43 RADEON_LLVM_AMDGPU_HS = 93,
44 };
45
46 #define RADEON_LLVM_MAX_INPUTS (VARYING_SLOT_VAR31 + 1)
47 #define RADEON_LLVM_MAX_OUTPUTS (VARYING_SLOT_VAR31 + 1)
48
49 struct nir_to_llvm_context;
50
51 struct ac_nir_context {
52 struct ac_llvm_context ac;
53 struct ac_shader_abi *abi;
54
55 gl_shader_stage stage;
56
57 struct hash_table *defs;
58 struct hash_table *phis;
59 struct hash_table *vars;
60
61 LLVMValueRef main_function;
62 LLVMBasicBlockRef continue_block;
63 LLVMBasicBlockRef break_block;
64
65 LLVMValueRef outputs[RADEON_LLVM_MAX_OUTPUTS * 4];
66
67 int num_locals;
68 LLVMValueRef *locals;
69
70 struct nir_to_llvm_context *nctx; /* TODO get rid of this */
71 };
72
73 struct nir_to_llvm_context {
74 struct ac_llvm_context ac;
75 const struct ac_nir_compiler_options *options;
76 struct ac_shader_variant_info *shader_info;
77 struct ac_shader_abi abi;
78 struct ac_nir_context *nir;
79
80 unsigned max_workgroup_size;
81 LLVMContextRef context;
82 LLVMModuleRef module;
83 LLVMBuilderRef builder;
84 LLVMValueRef main_function;
85
86 struct hash_table *defs;
87 struct hash_table *phis;
88
89 LLVMValueRef descriptor_sets[AC_UD_MAX_SETS];
90 LLVMValueRef ring_offsets;
91 LLVMValueRef push_constants;
92 LLVMValueRef view_index;
93 LLVMValueRef tg_size;
94
95 LLVMValueRef vertex_buffers;
96 LLVMValueRef rel_auto_id;
97 LLVMValueRef vs_prim_id;
98 LLVMValueRef ls_out_layout;
99 LLVMValueRef es2gs_offset;
100
101 LLVMValueRef tcs_offchip_layout;
102 LLVMValueRef tcs_out_offsets;
103 LLVMValueRef tcs_out_layout;
104 LLVMValueRef tcs_in_layout;
105 LLVMValueRef oc_lds;
106 LLVMValueRef merged_wave_info;
107 LLVMValueRef tess_factor_offset;
108 LLVMValueRef tes_rel_patch_id;
109 LLVMValueRef tes_u;
110 LLVMValueRef tes_v;
111
112 LLVMValueRef gsvs_ring_stride;
113 LLVMValueRef gsvs_num_entries;
114 LLVMValueRef gs2vs_offset;
115 LLVMValueRef gs_wave_id;
116 LLVMValueRef gs_vtx_offset[6];
117
118 LLVMValueRef esgs_ring;
119 LLVMValueRef gsvs_ring;
120 LLVMValueRef hs_ring_tess_offchip;
121 LLVMValueRef hs_ring_tess_factor;
122
123 LLVMValueRef sample_pos_offset;
124 LLVMValueRef persp_sample, persp_center, persp_centroid;
125 LLVMValueRef linear_sample, linear_center, linear_centroid;
126
127 gl_shader_stage stage;
128
129 LLVMValueRef inputs[RADEON_LLVM_MAX_INPUTS * 4];
130
131 uint64_t input_mask;
132 uint64_t output_mask;
133 uint8_t num_output_clips;
134 uint8_t num_output_culls;
135
136 bool is_gs_copy_shader;
137 LLVMValueRef gs_next_vertex;
138 unsigned gs_max_out_vertices;
139
140 unsigned tes_primitive_mode;
141 uint64_t tess_outputs_written;
142 uint64_t tess_patch_outputs_written;
143
144 uint32_t tcs_patch_outputs_read;
145 uint64_t tcs_outputs_read;
146 };
147
148 static inline struct nir_to_llvm_context *
149 nir_to_llvm_context_from_abi(struct ac_shader_abi *abi)
150 {
151 struct nir_to_llvm_context *ctx = NULL;
152 return container_of(abi, ctx, abi);
153 }
154
155 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
156 const nir_deref_var *deref,
157 enum ac_descriptor_type desc_type,
158 const nir_tex_instr *instr,
159 bool image, bool write);
160
161 static unsigned radeon_llvm_reg_index_soa(unsigned index, unsigned chan)
162 {
163 return (index * 4) + chan;
164 }
165
166 static unsigned shader_io_get_unique_index(gl_varying_slot slot)
167 {
168 /* handle patch indices separate */
169 if (slot == VARYING_SLOT_TESS_LEVEL_OUTER)
170 return 0;
171 if (slot == VARYING_SLOT_TESS_LEVEL_INNER)
172 return 1;
173 if (slot >= VARYING_SLOT_PATCH0 && slot <= VARYING_SLOT_TESS_MAX)
174 return 2 + (slot - VARYING_SLOT_PATCH0);
175
176 if (slot == VARYING_SLOT_POS)
177 return 0;
178 if (slot == VARYING_SLOT_PSIZ)
179 return 1;
180 if (slot == VARYING_SLOT_CLIP_DIST0)
181 return 2;
182 /* 3 is reserved for clip dist as well */
183 if (slot >= VARYING_SLOT_VAR0 && slot <= VARYING_SLOT_VAR31)
184 return 4 + (slot - VARYING_SLOT_VAR0);
185 unreachable("illegal slot in get unique index\n");
186 }
187
188 static void set_llvm_calling_convention(LLVMValueRef func,
189 gl_shader_stage stage)
190 {
191 enum radeon_llvm_calling_convention calling_conv;
192
193 switch (stage) {
194 case MESA_SHADER_VERTEX:
195 case MESA_SHADER_TESS_EVAL:
196 calling_conv = RADEON_LLVM_AMDGPU_VS;
197 break;
198 case MESA_SHADER_GEOMETRY:
199 calling_conv = RADEON_LLVM_AMDGPU_GS;
200 break;
201 case MESA_SHADER_TESS_CTRL:
202 calling_conv = HAVE_LLVM >= 0x0500 ? RADEON_LLVM_AMDGPU_HS : RADEON_LLVM_AMDGPU_VS;
203 break;
204 case MESA_SHADER_FRAGMENT:
205 calling_conv = RADEON_LLVM_AMDGPU_PS;
206 break;
207 case MESA_SHADER_COMPUTE:
208 calling_conv = RADEON_LLVM_AMDGPU_CS;
209 break;
210 default:
211 unreachable("Unhandle shader type");
212 }
213
214 LLVMSetFunctionCallConv(func, calling_conv);
215 }
216
217 #define MAX_ARGS 23
218 struct arg_info {
219 LLVMTypeRef types[MAX_ARGS];
220 LLVMValueRef *assign[MAX_ARGS];
221 unsigned array_params_mask;
222 uint8_t count;
223 uint8_t sgpr_count;
224 uint8_t num_sgprs_used;
225 uint8_t num_vgprs_used;
226 };
227
228 enum ac_arg_regfile {
229 ARG_SGPR,
230 ARG_VGPR,
231 };
232
233 static void
234 add_arg(struct arg_info *info, enum ac_arg_regfile regfile, LLVMTypeRef type,
235 LLVMValueRef *param_ptr)
236 {
237 assert(info->count < MAX_ARGS);
238
239 info->assign[info->count] = param_ptr;
240 info->types[info->count] = type;
241 info->count++;
242
243 if (regfile == ARG_SGPR) {
244 info->num_sgprs_used += ac_get_type_size(type) / 4;
245 info->sgpr_count++;
246 } else {
247 assert(regfile == ARG_VGPR);
248 info->num_vgprs_used += ac_get_type_size(type) / 4;
249 }
250 }
251
252 static inline void
253 add_array_arg(struct arg_info *info, LLVMTypeRef type, LLVMValueRef *param_ptr)
254 {
255 info->array_params_mask |= (1 << info->count);
256 add_arg(info, ARG_SGPR, type, param_ptr);
257 }
258
259 static void assign_arguments(LLVMValueRef main_function,
260 struct arg_info *info)
261 {
262 unsigned i;
263 for (i = 0; i < info->count; i++) {
264 if (info->assign[i])
265 *info->assign[i] = LLVMGetParam(main_function, i);
266 }
267 }
268
269 static LLVMValueRef
270 create_llvm_function(LLVMContextRef ctx, LLVMModuleRef module,
271 LLVMBuilderRef builder, LLVMTypeRef *return_types,
272 unsigned num_return_elems,
273 struct arg_info *args,
274 unsigned max_workgroup_size,
275 bool unsafe_math)
276 {
277 LLVMTypeRef main_function_type, ret_type;
278 LLVMBasicBlockRef main_function_body;
279
280 if (num_return_elems)
281 ret_type = LLVMStructTypeInContext(ctx, return_types,
282 num_return_elems, true);
283 else
284 ret_type = LLVMVoidTypeInContext(ctx);
285
286 /* Setup the function */
287 main_function_type =
288 LLVMFunctionType(ret_type, args->types, args->count, 0);
289 LLVMValueRef main_function =
290 LLVMAddFunction(module, "main", main_function_type);
291 main_function_body =
292 LLVMAppendBasicBlockInContext(ctx, main_function, "main_body");
293 LLVMPositionBuilderAtEnd(builder, main_function_body);
294
295 LLVMSetFunctionCallConv(main_function, RADEON_LLVM_AMDGPU_CS);
296 for (unsigned i = 0; i < args->sgpr_count; ++i) {
297 ac_add_function_attr(ctx, main_function, i + 1, AC_FUNC_ATTR_INREG);
298
299 if (args->array_params_mask & (1 << i)) {
300 LLVMValueRef P = LLVMGetParam(main_function, i);
301 ac_add_function_attr(ctx, main_function, i + 1, AC_FUNC_ATTR_NOALIAS);
302 ac_add_attr_dereferenceable(P, UINT64_MAX);
303 }
304 }
305
306 if (max_workgroup_size) {
307 ac_llvm_add_target_dep_function_attr(main_function,
308 "amdgpu-max-work-group-size",
309 max_workgroup_size);
310 }
311 if (unsafe_math) {
312 /* These were copied from some LLVM test. */
313 LLVMAddTargetDependentFunctionAttr(main_function,
314 "less-precise-fpmad",
315 "true");
316 LLVMAddTargetDependentFunctionAttr(main_function,
317 "no-infs-fp-math",
318 "true");
319 LLVMAddTargetDependentFunctionAttr(main_function,
320 "no-nans-fp-math",
321 "true");
322 LLVMAddTargetDependentFunctionAttr(main_function,
323 "unsafe-fp-math",
324 "true");
325 LLVMAddTargetDependentFunctionAttr(main_function,
326 "no-signed-zeros-fp-math",
327 "true");
328 }
329 return main_function;
330 }
331
332 static int get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
333 {
334 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
335 type = LLVMGetElementType(type);
336
337 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
338 return LLVMGetIntTypeWidth(type);
339
340 if (type == ctx->f16)
341 return 16;
342 if (type == ctx->f32)
343 return 32;
344 if (type == ctx->f64)
345 return 64;
346
347 unreachable("Unhandled type kind in get_elem_bits");
348 }
349
350 static LLVMValueRef unpack_param(struct ac_llvm_context *ctx,
351 LLVMValueRef param, unsigned rshift,
352 unsigned bitwidth)
353 {
354 LLVMValueRef value = param;
355 if (rshift)
356 value = LLVMBuildLShr(ctx->builder, value,
357 LLVMConstInt(ctx->i32, rshift, false), "");
358
359 if (rshift + bitwidth < 32) {
360 unsigned mask = (1 << bitwidth) - 1;
361 value = LLVMBuildAnd(ctx->builder, value,
362 LLVMConstInt(ctx->i32, mask, false), "");
363 }
364 return value;
365 }
366
367 static LLVMValueRef get_rel_patch_id(struct nir_to_llvm_context *ctx)
368 {
369 switch (ctx->stage) {
370 case MESA_SHADER_TESS_CTRL:
371 return unpack_param(&ctx->ac, ctx->abi.tcs_rel_ids, 0, 8);
372 case MESA_SHADER_TESS_EVAL:
373 return ctx->tes_rel_patch_id;
374 break;
375 default:
376 unreachable("Illegal stage");
377 }
378 }
379
380 /* Tessellation shaders pass outputs to the next shader using LDS.
381 *
382 * LS outputs = TCS inputs
383 * TCS outputs = TES inputs
384 *
385 * The LDS layout is:
386 * - TCS inputs for patch 0
387 * - TCS inputs for patch 1
388 * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
389 * - ...
390 * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
391 * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
392 * - TCS outputs for patch 1
393 * - Per-patch TCS outputs for patch 1
394 * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
395 * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
396 * - ...
397 *
398 * All three shaders VS(LS), TCS, TES share the same LDS space.
399 */
400 static LLVMValueRef
401 get_tcs_in_patch_stride(struct nir_to_llvm_context *ctx)
402 {
403 if (ctx->stage == MESA_SHADER_VERTEX)
404 return unpack_param(&ctx->ac, ctx->ls_out_layout, 0, 13);
405 else if (ctx->stage == MESA_SHADER_TESS_CTRL)
406 return unpack_param(&ctx->ac, ctx->tcs_in_layout, 0, 13);
407 else {
408 assert(0);
409 return NULL;
410 }
411 }
412
413 static LLVMValueRef
414 get_tcs_out_patch_stride(struct nir_to_llvm_context *ctx)
415 {
416 return unpack_param(&ctx->ac, ctx->tcs_out_layout, 0, 13);
417 }
418
419 static LLVMValueRef
420 get_tcs_out_patch0_offset(struct nir_to_llvm_context *ctx)
421 {
422 return LLVMBuildMul(ctx->builder,
423 unpack_param(&ctx->ac, ctx->tcs_out_offsets, 0, 16),
424 LLVMConstInt(ctx->ac.i32, 4, false), "");
425 }
426
427 static LLVMValueRef
428 get_tcs_out_patch0_patch_data_offset(struct nir_to_llvm_context *ctx)
429 {
430 return LLVMBuildMul(ctx->builder,
431 unpack_param(&ctx->ac, ctx->tcs_out_offsets, 16, 16),
432 LLVMConstInt(ctx->ac.i32, 4, false), "");
433 }
434
435 static LLVMValueRef
436 get_tcs_in_current_patch_offset(struct nir_to_llvm_context *ctx)
437 {
438 LLVMValueRef patch_stride = get_tcs_in_patch_stride(ctx);
439 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
440
441 return LLVMBuildMul(ctx->builder, patch_stride, rel_patch_id, "");
442 }
443
444 static LLVMValueRef
445 get_tcs_out_current_patch_offset(struct nir_to_llvm_context *ctx)
446 {
447 LLVMValueRef patch0_offset = get_tcs_out_patch0_offset(ctx);
448 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
449 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
450
451 return LLVMBuildAdd(ctx->builder, patch0_offset,
452 LLVMBuildMul(ctx->builder, patch_stride,
453 rel_patch_id, ""),
454 "");
455 }
456
457 static LLVMValueRef
458 get_tcs_out_current_patch_data_offset(struct nir_to_llvm_context *ctx)
459 {
460 LLVMValueRef patch0_patch_data_offset =
461 get_tcs_out_patch0_patch_data_offset(ctx);
462 LLVMValueRef patch_stride = get_tcs_out_patch_stride(ctx);
463 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
464
465 return LLVMBuildAdd(ctx->builder, patch0_patch_data_offset,
466 LLVMBuildMul(ctx->builder, patch_stride,
467 rel_patch_id, ""),
468 "");
469 }
470
471 static void
472 set_loc(struct ac_userdata_info *ud_info, uint8_t *sgpr_idx, uint8_t num_sgprs,
473 uint32_t indirect_offset)
474 {
475 ud_info->sgpr_idx = *sgpr_idx;
476 ud_info->num_sgprs = num_sgprs;
477 ud_info->indirect = indirect_offset > 0;
478 ud_info->indirect_offset = indirect_offset;
479 *sgpr_idx += num_sgprs;
480 }
481
482 static void
483 set_loc_shader(struct nir_to_llvm_context *ctx, int idx, uint8_t *sgpr_idx,
484 uint8_t num_sgprs)
485 {
486 struct ac_userdata_info *ud_info =
487 &ctx->shader_info->user_sgprs_locs.shader_data[idx];
488 assert(ud_info);
489
490 set_loc(ud_info, sgpr_idx, num_sgprs, 0);
491 }
492
493 static void
494 set_loc_desc(struct nir_to_llvm_context *ctx, int idx, uint8_t *sgpr_idx,
495 uint32_t indirect_offset)
496 {
497 struct ac_userdata_info *ud_info =
498 &ctx->shader_info->user_sgprs_locs.descriptor_sets[idx];
499 assert(ud_info);
500
501 set_loc(ud_info, sgpr_idx, 2, indirect_offset);
502 }
503
504 struct user_sgpr_info {
505 bool need_ring_offsets;
506 uint8_t sgpr_count;
507 bool indirect_all_descriptor_sets;
508 };
509
510 static bool needs_view_index_sgpr(struct nir_to_llvm_context *ctx,
511 gl_shader_stage stage)
512 {
513 switch (stage) {
514 case MESA_SHADER_VERTEX:
515 if (ctx->shader_info->info.needs_multiview_view_index ||
516 (!ctx->options->key.vs.as_es && !ctx->options->key.vs.as_ls && ctx->options->key.has_multiview_view_index))
517 return true;
518 break;
519 case MESA_SHADER_TESS_EVAL:
520 if (ctx->shader_info->info.needs_multiview_view_index || (!ctx->options->key.tes.as_es && ctx->options->key.has_multiview_view_index))
521 return true;
522 break;
523 case MESA_SHADER_GEOMETRY:
524 case MESA_SHADER_TESS_CTRL:
525 if (ctx->shader_info->info.needs_multiview_view_index)
526 return true;
527 break;
528 default:
529 break;
530 }
531 return false;
532 }
533
534 static void allocate_user_sgprs(struct nir_to_llvm_context *ctx,
535 gl_shader_stage stage,
536 bool needs_view_index,
537 struct user_sgpr_info *user_sgpr_info)
538 {
539 memset(user_sgpr_info, 0, sizeof(struct user_sgpr_info));
540
541 /* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
542 if (stage == MESA_SHADER_GEOMETRY ||
543 stage == MESA_SHADER_VERTEX ||
544 stage == MESA_SHADER_TESS_CTRL ||
545 stage == MESA_SHADER_TESS_EVAL ||
546 ctx->is_gs_copy_shader)
547 user_sgpr_info->need_ring_offsets = true;
548
549 if (stage == MESA_SHADER_FRAGMENT &&
550 ctx->shader_info->info.ps.needs_sample_positions)
551 user_sgpr_info->need_ring_offsets = true;
552
553 /* 2 user sgprs will nearly always be allocated for scratch/rings */
554 if (ctx->options->supports_spill || user_sgpr_info->need_ring_offsets) {
555 user_sgpr_info->sgpr_count += 2;
556 }
557
558 /* FIXME: fix the number of user sgprs for merged shaders on GFX9 */
559 switch (stage) {
560 case MESA_SHADER_COMPUTE:
561 if (ctx->shader_info->info.cs.uses_grid_size)
562 user_sgpr_info->sgpr_count += 3;
563 break;
564 case MESA_SHADER_FRAGMENT:
565 user_sgpr_info->sgpr_count += ctx->shader_info->info.ps.needs_sample_positions;
566 break;
567 case MESA_SHADER_VERTEX:
568 if (!ctx->is_gs_copy_shader) {
569 user_sgpr_info->sgpr_count += ctx->shader_info->info.vs.has_vertex_buffers ? 2 : 0;
570 if (ctx->shader_info->info.vs.needs_draw_id) {
571 user_sgpr_info->sgpr_count += 3;
572 } else {
573 user_sgpr_info->sgpr_count += 2;
574 }
575 }
576 if (ctx->options->key.vs.as_ls)
577 user_sgpr_info->sgpr_count++;
578 break;
579 case MESA_SHADER_TESS_CTRL:
580 user_sgpr_info->sgpr_count += 4;
581 break;
582 case MESA_SHADER_TESS_EVAL:
583 user_sgpr_info->sgpr_count += 1;
584 break;
585 case MESA_SHADER_GEOMETRY:
586 user_sgpr_info->sgpr_count += 2;
587 break;
588 default:
589 break;
590 }
591
592 if (needs_view_index)
593 user_sgpr_info->sgpr_count++;
594
595 if (ctx->shader_info->info.loads_push_constants)
596 user_sgpr_info->sgpr_count += 2;
597
598 uint32_t available_sgprs = ctx->options->chip_class >= GFX9 ? 32 : 16;
599 uint32_t remaining_sgprs = available_sgprs - user_sgpr_info->sgpr_count;
600
601 if (remaining_sgprs / 2 < util_bitcount(ctx->shader_info->info.desc_set_used_mask)) {
602 user_sgpr_info->sgpr_count += 2;
603 user_sgpr_info->indirect_all_descriptor_sets = true;
604 } else {
605 user_sgpr_info->sgpr_count += util_bitcount(ctx->shader_info->info.desc_set_used_mask) * 2;
606 }
607 }
608
609 static void
610 declare_global_input_sgprs(struct nir_to_llvm_context *ctx,
611 gl_shader_stage stage,
612 bool has_previous_stage,
613 gl_shader_stage previous_stage,
614 const struct user_sgpr_info *user_sgpr_info,
615 struct arg_info *args,
616 LLVMValueRef *desc_sets)
617 {
618 LLVMTypeRef type = ac_array_in_const_addr_space(ctx->ac.i8);
619 unsigned num_sets = ctx->options->layout ?
620 ctx->options->layout->num_sets : 0;
621 unsigned stage_mask = 1 << stage;
622
623 if (has_previous_stage)
624 stage_mask |= 1 << previous_stage;
625
626 /* 1 for each descriptor set */
627 if (!user_sgpr_info->indirect_all_descriptor_sets) {
628 for (unsigned i = 0; i < num_sets; ++i) {
629 if (ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
630 add_array_arg(args, type,
631 &ctx->descriptor_sets[i]);
632 }
633 }
634 } else {
635 add_array_arg(args, ac_array_in_const_addr_space(type), desc_sets);
636 }
637
638 if (ctx->shader_info->info.loads_push_constants) {
639 /* 1 for push constants and dynamic descriptors */
640 add_array_arg(args, type, &ctx->push_constants);
641 }
642 }
643
644 static void
645 declare_vs_specific_input_sgprs(struct nir_to_llvm_context *ctx,
646 gl_shader_stage stage,
647 bool has_previous_stage,
648 gl_shader_stage previous_stage,
649 struct arg_info *args)
650 {
651 if (!ctx->is_gs_copy_shader &&
652 (stage == MESA_SHADER_VERTEX ||
653 (has_previous_stage && previous_stage == MESA_SHADER_VERTEX))) {
654 if (ctx->shader_info->info.vs.has_vertex_buffers) {
655 add_arg(args, ARG_SGPR, ac_array_in_const_addr_space(ctx->ac.v4i32),
656 &ctx->vertex_buffers);
657 }
658 add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->abi.base_vertex);
659 add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->abi.start_instance);
660 if (ctx->shader_info->info.vs.needs_draw_id) {
661 add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->abi.draw_id);
662 }
663 }
664 }
665
666 static void
667 declare_vs_input_vgprs(struct nir_to_llvm_context *ctx, struct arg_info *args)
668 {
669 add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.vertex_id);
670 if (!ctx->is_gs_copy_shader) {
671 if (ctx->options->key.vs.as_ls) {
672 add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->rel_auto_id);
673 add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.instance_id);
674 } else {
675 add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.instance_id);
676 add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->vs_prim_id);
677 }
678 add_arg(args, ARG_VGPR, ctx->ac.i32, NULL); /* unused */
679 }
680 }
681
682 static void
683 declare_tes_input_vgprs(struct nir_to_llvm_context *ctx, struct arg_info *args)
684 {
685 add_arg(args, ARG_VGPR, ctx->ac.f32, &ctx->tes_u);
686 add_arg(args, ARG_VGPR, ctx->ac.f32, &ctx->tes_v);
687 add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->tes_rel_patch_id);
688 add_arg(args, ARG_VGPR, ctx->ac.i32, &ctx->abi.tes_patch_id);
689 }
690
691 static void
692 set_global_input_locs(struct nir_to_llvm_context *ctx, gl_shader_stage stage,
693 bool has_previous_stage, gl_shader_stage previous_stage,
694 const struct user_sgpr_info *user_sgpr_info,
695 LLVMValueRef desc_sets, uint8_t *user_sgpr_idx)
696 {
697 unsigned num_sets = ctx->options->layout ?
698 ctx->options->layout->num_sets : 0;
699 unsigned stage_mask = 1 << stage;
700
701 if (has_previous_stage)
702 stage_mask |= 1 << previous_stage;
703
704 if (!user_sgpr_info->indirect_all_descriptor_sets) {
705 for (unsigned i = 0; i < num_sets; ++i) {
706 if (ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
707 set_loc_desc(ctx, i, user_sgpr_idx, 0);
708 } else
709 ctx->descriptor_sets[i] = NULL;
710 }
711 } else {
712 set_loc_shader(ctx, AC_UD_INDIRECT_DESCRIPTOR_SETS,
713 user_sgpr_idx, 2);
714
715 for (unsigned i = 0; i < num_sets; ++i) {
716 if (ctx->options->layout->set[i].layout->shader_stages & stage_mask) {
717 set_loc_desc(ctx, i, user_sgpr_idx, i * 8);
718 ctx->descriptor_sets[i] =
719 ac_build_load_to_sgpr(&ctx->ac,
720 desc_sets,
721 LLVMConstInt(ctx->ac.i32, i, false));
722
723 } else
724 ctx->descriptor_sets[i] = NULL;
725 }
726 ctx->shader_info->need_indirect_descriptor_sets = true;
727 }
728
729 if (ctx->shader_info->info.loads_push_constants) {
730 set_loc_shader(ctx, AC_UD_PUSH_CONSTANTS, user_sgpr_idx, 2);
731 }
732 }
733
734 static void
735 set_vs_specific_input_locs(struct nir_to_llvm_context *ctx,
736 gl_shader_stage stage, bool has_previous_stage,
737 gl_shader_stage previous_stage,
738 uint8_t *user_sgpr_idx)
739 {
740 if (!ctx->is_gs_copy_shader &&
741 (stage == MESA_SHADER_VERTEX ||
742 (has_previous_stage && previous_stage == MESA_SHADER_VERTEX))) {
743 if (ctx->shader_info->info.vs.has_vertex_buffers) {
744 set_loc_shader(ctx, AC_UD_VS_VERTEX_BUFFERS,
745 user_sgpr_idx, 2);
746 }
747
748 unsigned vs_num = 2;
749 if (ctx->shader_info->info.vs.needs_draw_id)
750 vs_num++;
751
752 set_loc_shader(ctx, AC_UD_VS_BASE_VERTEX_START_INSTANCE,
753 user_sgpr_idx, vs_num);
754 }
755 }
756
757 static void create_function(struct nir_to_llvm_context *ctx,
758 gl_shader_stage stage,
759 bool has_previous_stage,
760 gl_shader_stage previous_stage)
761 {
762 uint8_t user_sgpr_idx;
763 struct user_sgpr_info user_sgpr_info;
764 struct arg_info args = {};
765 LLVMValueRef desc_sets;
766 bool needs_view_index = needs_view_index_sgpr(ctx, stage);
767 allocate_user_sgprs(ctx, stage, needs_view_index, &user_sgpr_info);
768
769 if (user_sgpr_info.need_ring_offsets && !ctx->options->supports_spill) {
770 add_arg(&args, ARG_SGPR, ac_array_in_const_addr_space(ctx->ac.v4i32),
771 &ctx->ring_offsets);
772 }
773
774 switch (stage) {
775 case MESA_SHADER_COMPUTE:
776 declare_global_input_sgprs(ctx, stage, has_previous_stage,
777 previous_stage, &user_sgpr_info,
778 &args, &desc_sets);
779
780 if (ctx->shader_info->info.cs.uses_grid_size) {
781 add_arg(&args, ARG_SGPR, ctx->ac.v3i32,
782 &ctx->abi.num_work_groups);
783 }
784
785 for (int i = 0; i < 3; i++) {
786 ctx->abi.workgroup_ids[i] = NULL;
787 if (ctx->shader_info->info.cs.uses_block_id[i]) {
788 add_arg(&args, ARG_SGPR, ctx->ac.i32,
789 &ctx->abi.workgroup_ids[i]);
790 }
791 }
792
793 if (ctx->shader_info->info.cs.uses_local_invocation_idx)
794 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->tg_size);
795 add_arg(&args, ARG_VGPR, ctx->ac.v3i32,
796 &ctx->abi.local_invocation_ids);
797 break;
798 case MESA_SHADER_VERTEX:
799 declare_global_input_sgprs(ctx, stage, has_previous_stage,
800 previous_stage, &user_sgpr_info,
801 &args, &desc_sets);
802 declare_vs_specific_input_sgprs(ctx, stage, has_previous_stage,
803 previous_stage, &args);
804
805 if (needs_view_index)
806 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->view_index);
807 if (ctx->options->key.vs.as_es)
808 add_arg(&args, ARG_SGPR, ctx->ac.i32,
809 &ctx->es2gs_offset);
810 else if (ctx->options->key.vs.as_ls)
811 add_arg(&args, ARG_SGPR, ctx->ac.i32,
812 &ctx->ls_out_layout);
813
814 declare_vs_input_vgprs(ctx, &args);
815 break;
816 case MESA_SHADER_TESS_CTRL:
817 if (has_previous_stage) {
818 // First 6 system regs
819 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->oc_lds);
820 add_arg(&args, ARG_SGPR, ctx->ac.i32,
821 &ctx->merged_wave_info);
822 add_arg(&args, ARG_SGPR, ctx->ac.i32,
823 &ctx->tess_factor_offset);
824
825 add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // scratch offset
826 add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
827 add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
828
829 declare_global_input_sgprs(ctx, stage,
830 has_previous_stage,
831 previous_stage,
832 &user_sgpr_info, &args,
833 &desc_sets);
834 declare_vs_specific_input_sgprs(ctx, stage,
835 has_previous_stage,
836 previous_stage, &args);
837
838 add_arg(&args, ARG_SGPR, ctx->ac.i32,
839 &ctx->ls_out_layout);
840
841 add_arg(&args, ARG_SGPR, ctx->ac.i32,
842 &ctx->tcs_offchip_layout);
843 add_arg(&args, ARG_SGPR, ctx->ac.i32,
844 &ctx->tcs_out_offsets);
845 add_arg(&args, ARG_SGPR, ctx->ac.i32,
846 &ctx->tcs_out_layout);
847 add_arg(&args, ARG_SGPR, ctx->ac.i32,
848 &ctx->tcs_in_layout);
849 if (needs_view_index)
850 add_arg(&args, ARG_SGPR, ctx->ac.i32,
851 &ctx->view_index);
852
853 add_arg(&args, ARG_VGPR, ctx->ac.i32,
854 &ctx->abi.tcs_patch_id);
855 add_arg(&args, ARG_VGPR, ctx->ac.i32,
856 &ctx->abi.tcs_rel_ids);
857
858 declare_vs_input_vgprs(ctx, &args);
859 } else {
860 declare_global_input_sgprs(ctx, stage,
861 has_previous_stage,
862 previous_stage,
863 &user_sgpr_info, &args,
864 &desc_sets);
865
866 add_arg(&args, ARG_SGPR, ctx->ac.i32,
867 &ctx->tcs_offchip_layout);
868 add_arg(&args, ARG_SGPR, ctx->ac.i32,
869 &ctx->tcs_out_offsets);
870 add_arg(&args, ARG_SGPR, ctx->ac.i32,
871 &ctx->tcs_out_layout);
872 add_arg(&args, ARG_SGPR, ctx->ac.i32,
873 &ctx->tcs_in_layout);
874 if (needs_view_index)
875 add_arg(&args, ARG_SGPR, ctx->ac.i32,
876 &ctx->view_index);
877
878 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->oc_lds);
879 add_arg(&args, ARG_SGPR, ctx->ac.i32,
880 &ctx->tess_factor_offset);
881 add_arg(&args, ARG_VGPR, ctx->ac.i32,
882 &ctx->abi.tcs_patch_id);
883 add_arg(&args, ARG_VGPR, ctx->ac.i32,
884 &ctx->abi.tcs_rel_ids);
885 }
886 break;
887 case MESA_SHADER_TESS_EVAL:
888 declare_global_input_sgprs(ctx, stage, has_previous_stage,
889 previous_stage, &user_sgpr_info,
890 &args, &desc_sets);
891
892 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->tcs_offchip_layout);
893 if (needs_view_index)
894 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->view_index);
895
896 if (ctx->options->key.tes.as_es) {
897 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->oc_lds);
898 add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL);
899 add_arg(&args, ARG_SGPR, ctx->ac.i32,
900 &ctx->es2gs_offset);
901 } else {
902 add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL);
903 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->oc_lds);
904 }
905 declare_tes_input_vgprs(ctx, &args);
906 break;
907 case MESA_SHADER_GEOMETRY:
908 if (has_previous_stage) {
909 // First 6 system regs
910 add_arg(&args, ARG_SGPR, ctx->ac.i32,
911 &ctx->gs2vs_offset);
912 add_arg(&args, ARG_SGPR, ctx->ac.i32,
913 &ctx->merged_wave_info);
914 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->oc_lds);
915
916 add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // scratch offset
917 add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
918 add_arg(&args, ARG_SGPR, ctx->ac.i32, NULL); // unknown
919
920 declare_global_input_sgprs(ctx, stage,
921 has_previous_stage,
922 previous_stage,
923 &user_sgpr_info, &args,
924 &desc_sets);
925
926 if (previous_stage == MESA_SHADER_TESS_EVAL) {
927 add_arg(&args, ARG_SGPR, ctx->ac.i32,
928 &ctx->tcs_offchip_layout);
929 } else {
930 declare_vs_specific_input_sgprs(ctx, stage,
931 has_previous_stage,
932 previous_stage,
933 &args);
934 }
935
936 add_arg(&args, ARG_SGPR, ctx->ac.i32,
937 &ctx->gsvs_ring_stride);
938 add_arg(&args, ARG_SGPR, ctx->ac.i32,
939 &ctx->gsvs_num_entries);
940 if (needs_view_index)
941 add_arg(&args, ARG_SGPR, ctx->ac.i32,
942 &ctx->view_index);
943
944 add_arg(&args, ARG_VGPR, ctx->ac.i32,
945 &ctx->gs_vtx_offset[0]);
946 add_arg(&args, ARG_VGPR, ctx->ac.i32,
947 &ctx->gs_vtx_offset[2]);
948 add_arg(&args, ARG_VGPR, ctx->ac.i32,
949 &ctx->abi.gs_prim_id);
950 add_arg(&args, ARG_VGPR, ctx->ac.i32,
951 &ctx->abi.gs_invocation_id);
952 add_arg(&args, ARG_VGPR, ctx->ac.i32,
953 &ctx->gs_vtx_offset[4]);
954
955 if (previous_stage == MESA_SHADER_VERTEX) {
956 declare_vs_input_vgprs(ctx, &args);
957 } else {
958 declare_tes_input_vgprs(ctx, &args);
959 }
960 } else {
961 declare_global_input_sgprs(ctx, stage,
962 has_previous_stage,
963 previous_stage,
964 &user_sgpr_info, &args,
965 &desc_sets);
966
967 add_arg(&args, ARG_SGPR, ctx->ac.i32,
968 &ctx->gsvs_ring_stride);
969 add_arg(&args, ARG_SGPR, ctx->ac.i32,
970 &ctx->gsvs_num_entries);
971 if (needs_view_index)
972 add_arg(&args, ARG_SGPR, ctx->ac.i32,
973 &ctx->view_index);
974
975 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->gs2vs_offset);
976 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->gs_wave_id);
977 add_arg(&args, ARG_VGPR, ctx->ac.i32,
978 &ctx->gs_vtx_offset[0]);
979 add_arg(&args, ARG_VGPR, ctx->ac.i32,
980 &ctx->gs_vtx_offset[1]);
981 add_arg(&args, ARG_VGPR, ctx->ac.i32,
982 &ctx->abi.gs_prim_id);
983 add_arg(&args, ARG_VGPR, ctx->ac.i32,
984 &ctx->gs_vtx_offset[2]);
985 add_arg(&args, ARG_VGPR, ctx->ac.i32,
986 &ctx->gs_vtx_offset[3]);
987 add_arg(&args, ARG_VGPR, ctx->ac.i32,
988 &ctx->gs_vtx_offset[4]);
989 add_arg(&args, ARG_VGPR, ctx->ac.i32,
990 &ctx->gs_vtx_offset[5]);
991 add_arg(&args, ARG_VGPR, ctx->ac.i32,
992 &ctx->abi.gs_invocation_id);
993 }
994 break;
995 case MESA_SHADER_FRAGMENT:
996 declare_global_input_sgprs(ctx, stage, has_previous_stage,
997 previous_stage, &user_sgpr_info,
998 &args, &desc_sets);
999
1000 if (ctx->shader_info->info.ps.needs_sample_positions)
1001 add_arg(&args, ARG_SGPR, ctx->ac.i32,
1002 &ctx->sample_pos_offset);
1003
1004 add_arg(&args, ARG_SGPR, ctx->ac.i32, &ctx->abi.prim_mask);
1005 add_arg(&args, ARG_VGPR, ctx->ac.v2i32, &ctx->persp_sample);
1006 add_arg(&args, ARG_VGPR, ctx->ac.v2i32, &ctx->persp_center);
1007 add_arg(&args, ARG_VGPR, ctx->ac.v2i32, &ctx->persp_centroid);
1008 add_arg(&args, ARG_VGPR, ctx->ac.v3i32, NULL); /* persp pull model */
1009 add_arg(&args, ARG_VGPR, ctx->ac.v2i32, &ctx->linear_sample);
1010 add_arg(&args, ARG_VGPR, ctx->ac.v2i32, &ctx->linear_center);
1011 add_arg(&args, ARG_VGPR, ctx->ac.v2i32, &ctx->linear_centroid);
1012 add_arg(&args, ARG_VGPR, ctx->ac.f32, NULL); /* line stipple tex */
1013 add_arg(&args, ARG_VGPR, ctx->ac.f32, &ctx->abi.frag_pos[0]);
1014 add_arg(&args, ARG_VGPR, ctx->ac.f32, &ctx->abi.frag_pos[1]);
1015 add_arg(&args, ARG_VGPR, ctx->ac.f32, &ctx->abi.frag_pos[2]);
1016 add_arg(&args, ARG_VGPR, ctx->ac.f32, &ctx->abi.frag_pos[3]);
1017 add_arg(&args, ARG_VGPR, ctx->ac.i32, &ctx->abi.front_face);
1018 add_arg(&args, ARG_VGPR, ctx->ac.i32, &ctx->abi.ancillary);
1019 add_arg(&args, ARG_VGPR, ctx->ac.i32, &ctx->abi.sample_coverage);
1020 add_arg(&args, ARG_VGPR, ctx->ac.i32, NULL); /* fixed pt */
1021 break;
1022 default:
1023 unreachable("Shader stage not implemented");
1024 }
1025
1026 ctx->main_function = create_llvm_function(
1027 ctx->context, ctx->module, ctx->builder, NULL, 0, &args,
1028 ctx->max_workgroup_size,
1029 ctx->options->unsafe_math);
1030 set_llvm_calling_convention(ctx->main_function, stage);
1031
1032
1033 ctx->shader_info->num_input_vgprs = 0;
1034 ctx->shader_info->num_input_sgprs = ctx->options->supports_spill ? 2 : 0;
1035
1036 ctx->shader_info->num_input_sgprs += args.num_sgprs_used;
1037
1038 if (ctx->stage != MESA_SHADER_FRAGMENT)
1039 ctx->shader_info->num_input_vgprs = args.num_vgprs_used;
1040
1041 assign_arguments(ctx->main_function, &args);
1042
1043 user_sgpr_idx = 0;
1044
1045 if (ctx->options->supports_spill || user_sgpr_info.need_ring_offsets) {
1046 set_loc_shader(ctx, AC_UD_SCRATCH_RING_OFFSETS,
1047 &user_sgpr_idx, 2);
1048 if (ctx->options->supports_spill) {
1049 ctx->ring_offsets = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.implicit.buffer.ptr",
1050 LLVMPointerType(ctx->ac.i8, AC_CONST_ADDR_SPACE),
1051 NULL, 0, AC_FUNC_ATTR_READNONE);
1052 ctx->ring_offsets = LLVMBuildBitCast(ctx->builder, ctx->ring_offsets,
1053 ac_array_in_const_addr_space(ctx->ac.v4i32), "");
1054 }
1055 }
1056
1057 /* For merged shaders the user SGPRs start at 8, with 8 system SGPRs in front (including
1058 * the rw_buffers at s0/s1. With user SGPR0 = s8, lets restart the count from 0 */
1059 if (has_previous_stage)
1060 user_sgpr_idx = 0;
1061
1062 set_global_input_locs(ctx, stage, has_previous_stage, previous_stage,
1063 &user_sgpr_info, desc_sets, &user_sgpr_idx);
1064
1065 switch (stage) {
1066 case MESA_SHADER_COMPUTE:
1067 if (ctx->shader_info->info.cs.uses_grid_size) {
1068 set_loc_shader(ctx, AC_UD_CS_GRID_SIZE,
1069 &user_sgpr_idx, 3);
1070 }
1071 break;
1072 case MESA_SHADER_VERTEX:
1073 set_vs_specific_input_locs(ctx, stage, has_previous_stage,
1074 previous_stage, &user_sgpr_idx);
1075 if (ctx->view_index)
1076 set_loc_shader(ctx, AC_UD_VIEW_INDEX, &user_sgpr_idx, 1);
1077 if (ctx->options->key.vs.as_ls) {
1078 set_loc_shader(ctx, AC_UD_VS_LS_TCS_IN_LAYOUT,
1079 &user_sgpr_idx, 1);
1080 }
1081 if (ctx->options->key.vs.as_ls)
1082 ac_declare_lds_as_pointer(&ctx->ac);
1083 break;
1084 case MESA_SHADER_TESS_CTRL:
1085 set_vs_specific_input_locs(ctx, stage, has_previous_stage,
1086 previous_stage, &user_sgpr_idx);
1087 if (has_previous_stage)
1088 set_loc_shader(ctx, AC_UD_VS_LS_TCS_IN_LAYOUT,
1089 &user_sgpr_idx, 1);
1090 set_loc_shader(ctx, AC_UD_TCS_OFFCHIP_LAYOUT, &user_sgpr_idx, 4);
1091 if (ctx->view_index)
1092 set_loc_shader(ctx, AC_UD_VIEW_INDEX, &user_sgpr_idx, 1);
1093 ac_declare_lds_as_pointer(&ctx->ac);
1094 break;
1095 case MESA_SHADER_TESS_EVAL:
1096 set_loc_shader(ctx, AC_UD_TES_OFFCHIP_LAYOUT, &user_sgpr_idx, 1);
1097 if (ctx->view_index)
1098 set_loc_shader(ctx, AC_UD_VIEW_INDEX, &user_sgpr_idx, 1);
1099 break;
1100 case MESA_SHADER_GEOMETRY:
1101 if (has_previous_stage) {
1102 if (previous_stage == MESA_SHADER_VERTEX)
1103 set_vs_specific_input_locs(ctx, stage,
1104 has_previous_stage,
1105 previous_stage,
1106 &user_sgpr_idx);
1107 else
1108 set_loc_shader(ctx, AC_UD_TES_OFFCHIP_LAYOUT,
1109 &user_sgpr_idx, 1);
1110 }
1111 set_loc_shader(ctx, AC_UD_GS_VS_RING_STRIDE_ENTRIES,
1112 &user_sgpr_idx, 2);
1113 if (ctx->view_index)
1114 set_loc_shader(ctx, AC_UD_VIEW_INDEX, &user_sgpr_idx, 1);
1115 if (has_previous_stage)
1116 ac_declare_lds_as_pointer(&ctx->ac);
1117 break;
1118 case MESA_SHADER_FRAGMENT:
1119 if (ctx->shader_info->info.ps.needs_sample_positions) {
1120 set_loc_shader(ctx, AC_UD_PS_SAMPLE_POS_OFFSET,
1121 &user_sgpr_idx, 1);
1122 }
1123 break;
1124 default:
1125 unreachable("Shader stage not implemented");
1126 }
1127
1128 ctx->shader_info->num_user_sgprs = user_sgpr_idx;
1129 }
1130
1131 static LLVMValueRef trim_vector(struct ac_llvm_context *ctx,
1132 LLVMValueRef value, unsigned count)
1133 {
1134 unsigned num_components = ac_get_llvm_num_components(value);
1135 if (count == num_components)
1136 return value;
1137
1138 LLVMValueRef masks[] = {
1139 LLVMConstInt(ctx->i32, 0, false), LLVMConstInt(ctx->i32, 1, false),
1140 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false)};
1141
1142 if (count == 1)
1143 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
1144 "");
1145
1146 LLVMValueRef swizzle = LLVMConstVector(masks, count);
1147 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
1148 }
1149
1150 static void
1151 build_store_values_extended(struct ac_llvm_context *ac,
1152 LLVMValueRef *values,
1153 unsigned value_count,
1154 unsigned value_stride,
1155 LLVMValueRef vec)
1156 {
1157 LLVMBuilderRef builder = ac->builder;
1158 unsigned i;
1159
1160 for (i = 0; i < value_count; i++) {
1161 LLVMValueRef ptr = values[i * value_stride];
1162 LLVMValueRef index = LLVMConstInt(ac->i32, i, false);
1163 LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
1164 LLVMBuildStore(builder, value, ptr);
1165 }
1166 }
1167
1168 static LLVMTypeRef get_def_type(struct ac_nir_context *ctx,
1169 const nir_ssa_def *def)
1170 {
1171 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, def->bit_size);
1172 if (def->num_components > 1) {
1173 type = LLVMVectorType(type, def->num_components);
1174 }
1175 return type;
1176 }
1177
1178 static LLVMValueRef get_src(struct ac_nir_context *nir, nir_src src)
1179 {
1180 assert(src.is_ssa);
1181 struct hash_entry *entry = _mesa_hash_table_search(nir->defs, src.ssa);
1182 return (LLVMValueRef)entry->data;
1183 }
1184
1185
1186 static LLVMBasicBlockRef get_block(struct ac_nir_context *nir,
1187 const struct nir_block *b)
1188 {
1189 struct hash_entry *entry = _mesa_hash_table_search(nir->defs, b);
1190 return (LLVMBasicBlockRef)entry->data;
1191 }
1192
1193 static LLVMValueRef get_alu_src(struct ac_nir_context *ctx,
1194 nir_alu_src src,
1195 unsigned num_components)
1196 {
1197 LLVMValueRef value = get_src(ctx, src.src);
1198 bool need_swizzle = false;
1199
1200 assert(value);
1201 LLVMTypeRef type = LLVMTypeOf(value);
1202 unsigned src_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
1203 ? LLVMGetVectorSize(type)
1204 : 1;
1205
1206 for (unsigned i = 0; i < num_components; ++i) {
1207 assert(src.swizzle[i] < src_components);
1208 if (src.swizzle[i] != i)
1209 need_swizzle = true;
1210 }
1211
1212 if (need_swizzle || num_components != src_components) {
1213 LLVMValueRef masks[] = {
1214 LLVMConstInt(ctx->ac.i32, src.swizzle[0], false),
1215 LLVMConstInt(ctx->ac.i32, src.swizzle[1], false),
1216 LLVMConstInt(ctx->ac.i32, src.swizzle[2], false),
1217 LLVMConstInt(ctx->ac.i32, src.swizzle[3], false)};
1218
1219 if (src_components > 1 && num_components == 1) {
1220 value = LLVMBuildExtractElement(ctx->ac.builder, value,
1221 masks[0], "");
1222 } else if (src_components == 1 && num_components > 1) {
1223 LLVMValueRef values[] = {value, value, value, value};
1224 value = ac_build_gather_values(&ctx->ac, values, num_components);
1225 } else {
1226 LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
1227 value = LLVMBuildShuffleVector(ctx->ac.builder, value, value,
1228 swizzle, "");
1229 }
1230 }
1231 assert(!src.negate);
1232 assert(!src.abs);
1233 return value;
1234 }
1235
1236 static LLVMValueRef emit_int_cmp(struct ac_llvm_context *ctx,
1237 LLVMIntPredicate pred, LLVMValueRef src0,
1238 LLVMValueRef src1)
1239 {
1240 LLVMValueRef result = LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
1241 return LLVMBuildSelect(ctx->builder, result,
1242 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
1243 ctx->i32_0, "");
1244 }
1245
1246 static LLVMValueRef emit_float_cmp(struct ac_llvm_context *ctx,
1247 LLVMRealPredicate pred, LLVMValueRef src0,
1248 LLVMValueRef src1)
1249 {
1250 LLVMValueRef result;
1251 src0 = ac_to_float(ctx, src0);
1252 src1 = ac_to_float(ctx, src1);
1253 result = LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
1254 return LLVMBuildSelect(ctx->builder, result,
1255 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
1256 ctx->i32_0, "");
1257 }
1258
1259 static LLVMValueRef emit_intrin_1f_param(struct ac_llvm_context *ctx,
1260 const char *intrin,
1261 LLVMTypeRef result_type,
1262 LLVMValueRef src0)
1263 {
1264 char name[64];
1265 LLVMValueRef params[] = {
1266 ac_to_float(ctx, src0),
1267 };
1268
1269 MAYBE_UNUSED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
1270 get_elem_bits(ctx, result_type));
1271 assert(length < sizeof(name));
1272 return ac_build_intrinsic(ctx, name, result_type, params, 1, AC_FUNC_ATTR_READNONE);
1273 }
1274
1275 static LLVMValueRef emit_intrin_2f_param(struct ac_llvm_context *ctx,
1276 const char *intrin,
1277 LLVMTypeRef result_type,
1278 LLVMValueRef src0, LLVMValueRef src1)
1279 {
1280 char name[64];
1281 LLVMValueRef params[] = {
1282 ac_to_float(ctx, src0),
1283 ac_to_float(ctx, src1),
1284 };
1285
1286 MAYBE_UNUSED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
1287 get_elem_bits(ctx, result_type));
1288 assert(length < sizeof(name));
1289 return ac_build_intrinsic(ctx, name, result_type, params, 2, AC_FUNC_ATTR_READNONE);
1290 }
1291
1292 static LLVMValueRef emit_intrin_3f_param(struct ac_llvm_context *ctx,
1293 const char *intrin,
1294 LLVMTypeRef result_type,
1295 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
1296 {
1297 char name[64];
1298 LLVMValueRef params[] = {
1299 ac_to_float(ctx, src0),
1300 ac_to_float(ctx, src1),
1301 ac_to_float(ctx, src2),
1302 };
1303
1304 MAYBE_UNUSED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
1305 get_elem_bits(ctx, result_type));
1306 assert(length < sizeof(name));
1307 return ac_build_intrinsic(ctx, name, result_type, params, 3, AC_FUNC_ATTR_READNONE);
1308 }
1309
1310 static LLVMValueRef emit_bcsel(struct ac_llvm_context *ctx,
1311 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
1312 {
1313 LLVMValueRef v = LLVMBuildICmp(ctx->builder, LLVMIntNE, src0,
1314 ctx->i32_0, "");
1315 return LLVMBuildSelect(ctx->builder, v, src1, src2, "");
1316 }
1317
1318 static LLVMValueRef emit_minmax_int(struct ac_llvm_context *ctx,
1319 LLVMIntPredicate pred,
1320 LLVMValueRef src0, LLVMValueRef src1)
1321 {
1322 return LLVMBuildSelect(ctx->builder,
1323 LLVMBuildICmp(ctx->builder, pred, src0, src1, ""),
1324 src0,
1325 src1, "");
1326
1327 }
1328 static LLVMValueRef emit_iabs(struct ac_llvm_context *ctx,
1329 LLVMValueRef src0)
1330 {
1331 return emit_minmax_int(ctx, LLVMIntSGT, src0,
1332 LLVMBuildNeg(ctx->builder, src0, ""));
1333 }
1334
1335 static LLVMValueRef emit_fsign(struct ac_llvm_context *ctx,
1336 LLVMValueRef src0,
1337 unsigned bitsize)
1338 {
1339 LLVMValueRef cmp, val, zero, one;
1340 LLVMTypeRef type;
1341
1342 if (bitsize == 32) {
1343 type = ctx->f32;
1344 zero = ctx->f32_0;
1345 one = ctx->f32_1;
1346 } else {
1347 type = ctx->f64;
1348 zero = ctx->f64_0;
1349 one = ctx->f64_1;
1350 }
1351
1352 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
1353 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
1354 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
1355 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
1356 return val;
1357 }
1358
1359 static LLVMValueRef emit_isign(struct ac_llvm_context *ctx,
1360 LLVMValueRef src0, unsigned bitsize)
1361 {
1362 LLVMValueRef cmp, val, zero, one;
1363 LLVMTypeRef type;
1364
1365 if (bitsize == 32) {
1366 type = ctx->i32;
1367 zero = ctx->i32_0;
1368 one = ctx->i32_1;
1369 } else {
1370 type = ctx->i64;
1371 zero = ctx->i64_0;
1372 one = ctx->i64_1;
1373 }
1374
1375 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
1376 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
1377 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
1378 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
1379 return val;
1380 }
1381
1382 static LLVMValueRef emit_ffract(struct ac_llvm_context *ctx,
1383 LLVMValueRef src0, unsigned bitsize)
1384 {
1385 LLVMTypeRef type;
1386 char *intr;
1387
1388 if (bitsize == 32) {
1389 intr = "llvm.floor.f32";
1390 type = ctx->f32;
1391 } else {
1392 intr = "llvm.floor.f64";
1393 type = ctx->f64;
1394 }
1395
1396 LLVMValueRef fsrc0 = ac_to_float(ctx, src0);
1397 LLVMValueRef params[] = {
1398 fsrc0,
1399 };
1400 LLVMValueRef floor = ac_build_intrinsic(ctx, intr, type, params, 1,
1401 AC_FUNC_ATTR_READNONE);
1402 return LLVMBuildFSub(ctx->builder, fsrc0, floor, "");
1403 }
1404
1405 static LLVMValueRef emit_uint_carry(struct ac_llvm_context *ctx,
1406 const char *intrin,
1407 LLVMValueRef src0, LLVMValueRef src1)
1408 {
1409 LLVMTypeRef ret_type;
1410 LLVMTypeRef types[] = { ctx->i32, ctx->i1 };
1411 LLVMValueRef res;
1412 LLVMValueRef params[] = { src0, src1 };
1413 ret_type = LLVMStructTypeInContext(ctx->context, types,
1414 2, true);
1415
1416 res = ac_build_intrinsic(ctx, intrin, ret_type,
1417 params, 2, AC_FUNC_ATTR_READNONE);
1418
1419 res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
1420 res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
1421 return res;
1422 }
1423
1424 static LLVMValueRef emit_b2f(struct ac_llvm_context *ctx,
1425 LLVMValueRef src0)
1426 {
1427 return LLVMBuildAnd(ctx->builder, src0, LLVMBuildBitCast(ctx->builder, LLVMConstReal(ctx->f32, 1.0), ctx->i32, ""), "");
1428 }
1429
1430 static LLVMValueRef emit_f2b(struct ac_llvm_context *ctx,
1431 LLVMValueRef src0)
1432 {
1433 src0 = ac_to_float(ctx, src0);
1434 return LLVMBuildSExt(ctx->builder,
1435 LLVMBuildFCmp(ctx->builder, LLVMRealUNE, src0, ctx->f32_0, ""),
1436 ctx->i32, "");
1437 }
1438
1439 static LLVMValueRef emit_b2i(struct ac_llvm_context *ctx,
1440 LLVMValueRef src0,
1441 unsigned bitsize)
1442 {
1443 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0, ctx->i32_1, "");
1444
1445 if (bitsize == 32)
1446 return result;
1447
1448 return LLVMBuildZExt(ctx->builder, result, ctx->i64, "");
1449 }
1450
1451 static LLVMValueRef emit_i2b(struct ac_llvm_context *ctx,
1452 LLVMValueRef src0)
1453 {
1454 return LLVMBuildSExt(ctx->builder,
1455 LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, ctx->i32_0, ""),
1456 ctx->i32, "");
1457 }
1458
1459 static LLVMValueRef emit_f2f16(struct nir_to_llvm_context *ctx,
1460 LLVMValueRef src0)
1461 {
1462 LLVMValueRef result;
1463 LLVMValueRef cond = NULL;
1464
1465 src0 = ac_to_float(&ctx->ac, src0);
1466 result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->ac.f16, "");
1467
1468 if (ctx->options->chip_class >= VI) {
1469 LLVMValueRef args[2];
1470 /* Check if the result is a denormal - and flush to 0 if so. */
1471 args[0] = result;
1472 args[1] = LLVMConstInt(ctx->ac.i32, N_SUBNORMAL | P_SUBNORMAL, false);
1473 cond = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.class.f16", ctx->ac.i1, args, 2, AC_FUNC_ATTR_READNONE);
1474 }
1475
1476 /* need to convert back up to f32 */
1477 result = LLVMBuildFPExt(ctx->builder, result, ctx->ac.f32, "");
1478
1479 if (ctx->options->chip_class >= VI)
1480 result = LLVMBuildSelect(ctx->builder, cond, ctx->ac.f32_0, result, "");
1481 else {
1482 /* for SI/CIK */
1483 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
1484 * so compare the result and flush to 0 if it's smaller.
1485 */
1486 LLVMValueRef temp, cond2;
1487 temp = emit_intrin_1f_param(&ctx->ac, "llvm.fabs",
1488 ctx->ac.f32, result);
1489 cond = LLVMBuildFCmp(ctx->builder, LLVMRealUGT,
1490 LLVMBuildBitCast(ctx->builder, LLVMConstInt(ctx->ac.i32, 0x38800000, false), ctx->ac.f32, ""),
1491 temp, "");
1492 cond2 = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
1493 temp, ctx->ac.f32_0, "");
1494 cond = LLVMBuildAnd(ctx->builder, cond, cond2, "");
1495 result = LLVMBuildSelect(ctx->builder, cond, ctx->ac.f32_0, result, "");
1496 }
1497 return result;
1498 }
1499
1500 static LLVMValueRef emit_umul_high(struct ac_llvm_context *ctx,
1501 LLVMValueRef src0, LLVMValueRef src1)
1502 {
1503 LLVMValueRef dst64, result;
1504 src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
1505 src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");
1506
1507 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
1508 dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
1509 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
1510 return result;
1511 }
1512
1513 static LLVMValueRef emit_imul_high(struct ac_llvm_context *ctx,
1514 LLVMValueRef src0, LLVMValueRef src1)
1515 {
1516 LLVMValueRef dst64, result;
1517 src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
1518 src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");
1519
1520 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
1521 dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
1522 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
1523 return result;
1524 }
1525
1526 static LLVMValueRef emit_bitfield_extract(struct ac_llvm_context *ctx,
1527 bool is_signed,
1528 const LLVMValueRef srcs[3])
1529 {
1530 LLVMValueRef result;
1531 LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, srcs[2], LLVMConstInt(ctx->i32, 32, false), "");
1532
1533 result = ac_build_bfe(ctx, srcs[0], srcs[1], srcs[2], is_signed);
1534 result = LLVMBuildSelect(ctx->builder, icond, srcs[0], result, "");
1535 return result;
1536 }
1537
1538 static LLVMValueRef emit_bitfield_insert(struct ac_llvm_context *ctx,
1539 LLVMValueRef src0, LLVMValueRef src1,
1540 LLVMValueRef src2, LLVMValueRef src3)
1541 {
1542 LLVMValueRef bfi_args[3], result;
1543
1544 bfi_args[0] = LLVMBuildShl(ctx->builder,
1545 LLVMBuildSub(ctx->builder,
1546 LLVMBuildShl(ctx->builder,
1547 ctx->i32_1,
1548 src3, ""),
1549 ctx->i32_1, ""),
1550 src2, "");
1551 bfi_args[1] = LLVMBuildShl(ctx->builder, src1, src2, "");
1552 bfi_args[2] = src0;
1553
1554 LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, src3, LLVMConstInt(ctx->i32, 32, false), "");
1555
1556 /* Calculate:
1557 * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
1558 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
1559 */
1560 result = LLVMBuildXor(ctx->builder, bfi_args[2],
1561 LLVMBuildAnd(ctx->builder, bfi_args[0],
1562 LLVMBuildXor(ctx->builder, bfi_args[1], bfi_args[2], ""), ""), "");
1563
1564 result = LLVMBuildSelect(ctx->builder, icond, src1, result, "");
1565 return result;
1566 }
1567
1568 static LLVMValueRef emit_pack_half_2x16(struct ac_llvm_context *ctx,
1569 LLVMValueRef src0)
1570 {
1571 LLVMValueRef comp[2];
1572
1573 src0 = ac_to_float(ctx, src0);
1574 comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
1575 comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");
1576
1577 return ac_build_cvt_pkrtz_f16(ctx, comp);
1578 }
1579
1580 static LLVMValueRef emit_unpack_half_2x16(struct ac_llvm_context *ctx,
1581 LLVMValueRef src0)
1582 {
1583 LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
1584 LLVMValueRef temps[2], result, val;
1585 int i;
1586
1587 for (i = 0; i < 2; i++) {
1588 val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
1589 val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
1590 val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
1591 temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
1592 }
1593
1594 result = LLVMBuildInsertElement(ctx->builder, LLVMGetUndef(ctx->v2f32), temps[0],
1595 ctx->i32_0, "");
1596 result = LLVMBuildInsertElement(ctx->builder, result, temps[1],
1597 ctx->i32_1, "");
1598 return result;
1599 }
1600
1601 static LLVMValueRef emit_ddxy(struct ac_nir_context *ctx,
1602 nir_op op,
1603 LLVMValueRef src0)
1604 {
1605 unsigned mask;
1606 int idx;
1607 LLVMValueRef result;
1608
1609 if (op == nir_op_fddx_fine || op == nir_op_fddx)
1610 mask = AC_TID_MASK_LEFT;
1611 else if (op == nir_op_fddy_fine || op == nir_op_fddy)
1612 mask = AC_TID_MASK_TOP;
1613 else
1614 mask = AC_TID_MASK_TOP_LEFT;
1615
1616 /* for DDX we want to next X pixel, DDY next Y pixel. */
1617 if (op == nir_op_fddx_fine ||
1618 op == nir_op_fddx_coarse ||
1619 op == nir_op_fddx)
1620 idx = 1;
1621 else
1622 idx = 2;
1623
1624 result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
1625 return result;
1626 }
1627
1628 /*
1629 * this takes an I,J coordinate pair,
1630 * and works out the X and Y derivatives.
1631 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
1632 */
1633 static LLVMValueRef emit_ddxy_interp(
1634 struct ac_nir_context *ctx,
1635 LLVMValueRef interp_ij)
1636 {
1637 LLVMValueRef result[4], a;
1638 unsigned i;
1639
1640 for (i = 0; i < 2; i++) {
1641 a = LLVMBuildExtractElement(ctx->ac.builder, interp_ij,
1642 LLVMConstInt(ctx->ac.i32, i, false), "");
1643 result[i] = emit_ddxy(ctx, nir_op_fddx, a);
1644 result[2+i] = emit_ddxy(ctx, nir_op_fddy, a);
1645 }
1646 return ac_build_gather_values(&ctx->ac, result, 4);
1647 }
1648
1649 static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
1650 {
1651 LLVMValueRef src[4], result = NULL;
1652 unsigned num_components = instr->dest.dest.ssa.num_components;
1653 unsigned src_components;
1654 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.dest.ssa);
1655
1656 assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
1657 switch (instr->op) {
1658 case nir_op_vec2:
1659 case nir_op_vec3:
1660 case nir_op_vec4:
1661 src_components = 1;
1662 break;
1663 case nir_op_pack_half_2x16:
1664 src_components = 2;
1665 break;
1666 case nir_op_unpack_half_2x16:
1667 src_components = 1;
1668 break;
1669 default:
1670 src_components = num_components;
1671 break;
1672 }
1673 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1674 src[i] = get_alu_src(ctx, instr->src[i], src_components);
1675
1676 switch (instr->op) {
1677 case nir_op_fmov:
1678 case nir_op_imov:
1679 result = src[0];
1680 break;
1681 case nir_op_fneg:
1682 src[0] = ac_to_float(&ctx->ac, src[0]);
1683 result = LLVMBuildFNeg(ctx->ac.builder, src[0], "");
1684 break;
1685 case nir_op_ineg:
1686 result = LLVMBuildNeg(ctx->ac.builder, src[0], "");
1687 break;
1688 case nir_op_inot:
1689 result = LLVMBuildNot(ctx->ac.builder, src[0], "");
1690 break;
1691 case nir_op_iadd:
1692 result = LLVMBuildAdd(ctx->ac.builder, src[0], src[1], "");
1693 break;
1694 case nir_op_fadd:
1695 src[0] = ac_to_float(&ctx->ac, src[0]);
1696 src[1] = ac_to_float(&ctx->ac, src[1]);
1697 result = LLVMBuildFAdd(ctx->ac.builder, src[0], src[1], "");
1698 break;
1699 case nir_op_fsub:
1700 src[0] = ac_to_float(&ctx->ac, src[0]);
1701 src[1] = ac_to_float(&ctx->ac, src[1]);
1702 result = LLVMBuildFSub(ctx->ac.builder, src[0], src[1], "");
1703 break;
1704 case nir_op_isub:
1705 result = LLVMBuildSub(ctx->ac.builder, src[0], src[1], "");
1706 break;
1707 case nir_op_imul:
1708 result = LLVMBuildMul(ctx->ac.builder, src[0], src[1], "");
1709 break;
1710 case nir_op_imod:
1711 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
1712 break;
1713 case nir_op_umod:
1714 result = LLVMBuildURem(ctx->ac.builder, src[0], src[1], "");
1715 break;
1716 case nir_op_fmod:
1717 src[0] = ac_to_float(&ctx->ac, src[0]);
1718 src[1] = ac_to_float(&ctx->ac, src[1]);
1719 result = ac_build_fdiv(&ctx->ac, src[0], src[1]);
1720 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
1721 ac_to_float_type(&ctx->ac, def_type), result);
1722 result = LLVMBuildFMul(ctx->ac.builder, src[1] , result, "");
1723 result = LLVMBuildFSub(ctx->ac.builder, src[0], result, "");
1724 break;
1725 case nir_op_frem:
1726 src[0] = ac_to_float(&ctx->ac, src[0]);
1727 src[1] = ac_to_float(&ctx->ac, src[1]);
1728 result = LLVMBuildFRem(ctx->ac.builder, src[0], src[1], "");
1729 break;
1730 case nir_op_irem:
1731 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
1732 break;
1733 case nir_op_idiv:
1734 result = LLVMBuildSDiv(ctx->ac.builder, src[0], src[1], "");
1735 break;
1736 case nir_op_udiv:
1737 result = LLVMBuildUDiv(ctx->ac.builder, src[0], src[1], "");
1738 break;
1739 case nir_op_fmul:
1740 src[0] = ac_to_float(&ctx->ac, src[0]);
1741 src[1] = ac_to_float(&ctx->ac, src[1]);
1742 result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
1743 break;
1744 case nir_op_frcp:
1745 src[0] = ac_to_float(&ctx->ac, src[0]);
1746 result = ac_build_fdiv(&ctx->ac, instr->dest.dest.ssa.bit_size == 32 ? ctx->ac.f32_1 : ctx->ac.f64_1,
1747 src[0]);
1748 break;
1749 case nir_op_iand:
1750 result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
1751 break;
1752 case nir_op_ior:
1753 result = LLVMBuildOr(ctx->ac.builder, src[0], src[1], "");
1754 break;
1755 case nir_op_ixor:
1756 result = LLVMBuildXor(ctx->ac.builder, src[0], src[1], "");
1757 break;
1758 case nir_op_ishl:
1759 result = LLVMBuildShl(ctx->ac.builder, src[0],
1760 LLVMBuildZExt(ctx->ac.builder, src[1],
1761 LLVMTypeOf(src[0]), ""),
1762 "");
1763 break;
1764 case nir_op_ishr:
1765 result = LLVMBuildAShr(ctx->ac.builder, src[0],
1766 LLVMBuildZExt(ctx->ac.builder, src[1],
1767 LLVMTypeOf(src[0]), ""),
1768 "");
1769 break;
1770 case nir_op_ushr:
1771 result = LLVMBuildLShr(ctx->ac.builder, src[0],
1772 LLVMBuildZExt(ctx->ac.builder, src[1],
1773 LLVMTypeOf(src[0]), ""),
1774 "");
1775 break;
1776 case nir_op_ilt:
1777 result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
1778 break;
1779 case nir_op_ine:
1780 result = emit_int_cmp(&ctx->ac, LLVMIntNE, src[0], src[1]);
1781 break;
1782 case nir_op_ieq:
1783 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, src[0], src[1]);
1784 break;
1785 case nir_op_ige:
1786 result = emit_int_cmp(&ctx->ac, LLVMIntSGE, src[0], src[1]);
1787 break;
1788 case nir_op_ult:
1789 result = emit_int_cmp(&ctx->ac, LLVMIntULT, src[0], src[1]);
1790 break;
1791 case nir_op_uge:
1792 result = emit_int_cmp(&ctx->ac, LLVMIntUGE, src[0], src[1]);
1793 break;
1794 case nir_op_feq:
1795 result = emit_float_cmp(&ctx->ac, LLVMRealUEQ, src[0], src[1]);
1796 break;
1797 case nir_op_fne:
1798 result = emit_float_cmp(&ctx->ac, LLVMRealUNE, src[0], src[1]);
1799 break;
1800 case nir_op_flt:
1801 result = emit_float_cmp(&ctx->ac, LLVMRealULT, src[0], src[1]);
1802 break;
1803 case nir_op_fge:
1804 result = emit_float_cmp(&ctx->ac, LLVMRealUGE, src[0], src[1]);
1805 break;
1806 case nir_op_fabs:
1807 result = emit_intrin_1f_param(&ctx->ac, "llvm.fabs",
1808 ac_to_float_type(&ctx->ac, def_type), src[0]);
1809 break;
1810 case nir_op_iabs:
1811 result = emit_iabs(&ctx->ac, src[0]);
1812 break;
1813 case nir_op_imax:
1814 result = emit_minmax_int(&ctx->ac, LLVMIntSGT, src[0], src[1]);
1815 break;
1816 case nir_op_imin:
1817 result = emit_minmax_int(&ctx->ac, LLVMIntSLT, src[0], src[1]);
1818 break;
1819 case nir_op_umax:
1820 result = emit_minmax_int(&ctx->ac, LLVMIntUGT, src[0], src[1]);
1821 break;
1822 case nir_op_umin:
1823 result = emit_minmax_int(&ctx->ac, LLVMIntULT, src[0], src[1]);
1824 break;
1825 case nir_op_isign:
1826 result = emit_isign(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1827 break;
1828 case nir_op_fsign:
1829 src[0] = ac_to_float(&ctx->ac, src[0]);
1830 result = emit_fsign(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1831 break;
1832 case nir_op_ffloor:
1833 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
1834 ac_to_float_type(&ctx->ac, def_type), src[0]);
1835 break;
1836 case nir_op_ftrunc:
1837 result = emit_intrin_1f_param(&ctx->ac, "llvm.trunc",
1838 ac_to_float_type(&ctx->ac, def_type), src[0]);
1839 break;
1840 case nir_op_fceil:
1841 result = emit_intrin_1f_param(&ctx->ac, "llvm.ceil",
1842 ac_to_float_type(&ctx->ac, def_type), src[0]);
1843 break;
1844 case nir_op_fround_even:
1845 result = emit_intrin_1f_param(&ctx->ac, "llvm.rint",
1846 ac_to_float_type(&ctx->ac, def_type),src[0]);
1847 break;
1848 case nir_op_ffract:
1849 result = emit_ffract(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1850 break;
1851 case nir_op_fsin:
1852 result = emit_intrin_1f_param(&ctx->ac, "llvm.sin",
1853 ac_to_float_type(&ctx->ac, def_type), src[0]);
1854 break;
1855 case nir_op_fcos:
1856 result = emit_intrin_1f_param(&ctx->ac, "llvm.cos",
1857 ac_to_float_type(&ctx->ac, def_type), src[0]);
1858 break;
1859 case nir_op_fsqrt:
1860 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
1861 ac_to_float_type(&ctx->ac, def_type), src[0]);
1862 break;
1863 case nir_op_fexp2:
1864 result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
1865 ac_to_float_type(&ctx->ac, def_type), src[0]);
1866 break;
1867 case nir_op_flog2:
1868 result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
1869 ac_to_float_type(&ctx->ac, def_type), src[0]);
1870 break;
1871 case nir_op_frsq:
1872 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
1873 ac_to_float_type(&ctx->ac, def_type), src[0]);
1874 result = ac_build_fdiv(&ctx->ac, instr->dest.dest.ssa.bit_size == 32 ? ctx->ac.f32_1 : ctx->ac.f64_1,
1875 result);
1876 break;
1877 case nir_op_fpow:
1878 result = emit_intrin_2f_param(&ctx->ac, "llvm.pow",
1879 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1880 break;
1881 case nir_op_fmax:
1882 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1883 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1884 if (ctx->ac.chip_class < GFX9 &&
1885 instr->dest.dest.ssa.bit_size == 32) {
1886 /* Only pre-GFX9 chips do not flush denorms. */
1887 result = emit_intrin_1f_param(&ctx->ac, "llvm.canonicalize",
1888 ac_to_float_type(&ctx->ac, def_type),
1889 result);
1890 }
1891 break;
1892 case nir_op_fmin:
1893 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1894 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1895 if (ctx->ac.chip_class < GFX9 &&
1896 instr->dest.dest.ssa.bit_size == 32) {
1897 /* Only pre-GFX9 chips do not flush denorms. */
1898 result = emit_intrin_1f_param(&ctx->ac, "llvm.canonicalize",
1899 ac_to_float_type(&ctx->ac, def_type),
1900 result);
1901 }
1902 break;
1903 case nir_op_ffma:
1904 result = emit_intrin_3f_param(&ctx->ac, "llvm.fmuladd",
1905 ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]);
1906 break;
1907 case nir_op_ibitfield_extract:
1908 result = emit_bitfield_extract(&ctx->ac, true, src);
1909 break;
1910 case nir_op_ubitfield_extract:
1911 result = emit_bitfield_extract(&ctx->ac, false, src);
1912 break;
1913 case nir_op_bitfield_insert:
1914 result = emit_bitfield_insert(&ctx->ac, src[0], src[1], src[2], src[3]);
1915 break;
1916 case nir_op_bitfield_reverse:
1917 result = ac_build_intrinsic(&ctx->ac, "llvm.bitreverse.i32", ctx->ac.i32, src, 1, AC_FUNC_ATTR_READNONE);
1918 break;
1919 case nir_op_bit_count:
1920 result = ac_build_intrinsic(&ctx->ac, "llvm.ctpop.i32", ctx->ac.i32, src, 1, AC_FUNC_ATTR_READNONE);
1921 break;
1922 case nir_op_vec2:
1923 case nir_op_vec3:
1924 case nir_op_vec4:
1925 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1926 src[i] = ac_to_integer(&ctx->ac, src[i]);
1927 result = ac_build_gather_values(&ctx->ac, src, num_components);
1928 break;
1929 case nir_op_f2i32:
1930 case nir_op_f2i64:
1931 src[0] = ac_to_float(&ctx->ac, src[0]);
1932 result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
1933 break;
1934 case nir_op_f2u32:
1935 case nir_op_f2u64:
1936 src[0] = ac_to_float(&ctx->ac, src[0]);
1937 result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
1938 break;
1939 case nir_op_i2f32:
1940 case nir_op_i2f64:
1941 src[0] = ac_to_integer(&ctx->ac, src[0]);
1942 result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1943 break;
1944 case nir_op_u2f32:
1945 case nir_op_u2f64:
1946 src[0] = ac_to_integer(&ctx->ac, src[0]);
1947 result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1948 break;
1949 case nir_op_f2f64:
1950 src[0] = ac_to_float(&ctx->ac, src[0]);
1951 result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1952 break;
1953 case nir_op_f2f32:
1954 result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1955 break;
1956 case nir_op_u2u32:
1957 case nir_op_u2u64:
1958 src[0] = ac_to_integer(&ctx->ac, src[0]);
1959 if (get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < get_elem_bits(&ctx->ac, def_type))
1960 result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
1961 else
1962 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
1963 break;
1964 case nir_op_i2i32:
1965 case nir_op_i2i64:
1966 src[0] = ac_to_integer(&ctx->ac, src[0]);
1967 if (get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < get_elem_bits(&ctx->ac, def_type))
1968 result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
1969 else
1970 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
1971 break;
1972 case nir_op_bcsel:
1973 result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
1974 break;
1975 case nir_op_find_lsb:
1976 src[0] = ac_to_integer(&ctx->ac, src[0]);
1977 result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
1978 break;
1979 case nir_op_ufind_msb:
1980 src[0] = ac_to_integer(&ctx->ac, src[0]);
1981 result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
1982 break;
1983 case nir_op_ifind_msb:
1984 src[0] = ac_to_integer(&ctx->ac, src[0]);
1985 result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
1986 break;
1987 case nir_op_uadd_carry:
1988 src[0] = ac_to_integer(&ctx->ac, src[0]);
1989 src[1] = ac_to_integer(&ctx->ac, src[1]);
1990 result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
1991 break;
1992 case nir_op_usub_borrow:
1993 src[0] = ac_to_integer(&ctx->ac, src[0]);
1994 src[1] = ac_to_integer(&ctx->ac, src[1]);
1995 result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
1996 break;
1997 case nir_op_b2f:
1998 result = emit_b2f(&ctx->ac, src[0]);
1999 break;
2000 case nir_op_f2b:
2001 result = emit_f2b(&ctx->ac, src[0]);
2002 break;
2003 case nir_op_b2i:
2004 result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
2005 break;
2006 case nir_op_i2b:
2007 src[0] = ac_to_integer(&ctx->ac, src[0]);
2008 result = emit_i2b(&ctx->ac, src[0]);
2009 break;
2010 case nir_op_fquantize2f16:
2011 result = emit_f2f16(ctx->nctx, src[0]);
2012 break;
2013 case nir_op_umul_high:
2014 src[0] = ac_to_integer(&ctx->ac, src[0]);
2015 src[1] = ac_to_integer(&ctx->ac, src[1]);
2016 result = emit_umul_high(&ctx->ac, src[0], src[1]);
2017 break;
2018 case nir_op_imul_high:
2019 src[0] = ac_to_integer(&ctx->ac, src[0]);
2020 src[1] = ac_to_integer(&ctx->ac, src[1]);
2021 result = emit_imul_high(&ctx->ac, src[0], src[1]);
2022 break;
2023 case nir_op_pack_half_2x16:
2024 result = emit_pack_half_2x16(&ctx->ac, src[0]);
2025 break;
2026 case nir_op_unpack_half_2x16:
2027 result = emit_unpack_half_2x16(&ctx->ac, src[0]);
2028 break;
2029 case nir_op_fddx:
2030 case nir_op_fddy:
2031 case nir_op_fddx_fine:
2032 case nir_op_fddy_fine:
2033 case nir_op_fddx_coarse:
2034 case nir_op_fddy_coarse:
2035 result = emit_ddxy(ctx, instr->op, src[0]);
2036 break;
2037
2038 case nir_op_unpack_64_2x32_split_x: {
2039 assert(instr->src[0].src.ssa->num_components == 1);
2040 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
2041 ctx->ac.v2i32,
2042 "");
2043 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
2044 ctx->ac.i32_0, "");
2045 break;
2046 }
2047
2048 case nir_op_unpack_64_2x32_split_y: {
2049 assert(instr->src[0].src.ssa->num_components == 1);
2050 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
2051 ctx->ac.v2i32,
2052 "");
2053 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
2054 ctx->ac.i32_1, "");
2055 break;
2056 }
2057
2058 case nir_op_pack_64_2x32_split: {
2059 LLVMValueRef tmp = LLVMGetUndef(ctx->ac.v2i32);
2060 tmp = LLVMBuildInsertElement(ctx->ac.builder, tmp,
2061 src[0], ctx->ac.i32_0, "");
2062 tmp = LLVMBuildInsertElement(ctx->ac.builder, tmp,
2063 src[1], ctx->ac.i32_1, "");
2064 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
2065 break;
2066 }
2067
2068 default:
2069 fprintf(stderr, "Unknown NIR alu instr: ");
2070 nir_print_instr(&instr->instr, stderr);
2071 fprintf(stderr, "\n");
2072 abort();
2073 }
2074
2075 if (result) {
2076 assert(instr->dest.dest.is_ssa);
2077 result = ac_to_integer(&ctx->ac, result);
2078 _mesa_hash_table_insert(ctx->defs, &instr->dest.dest.ssa,
2079 result);
2080 }
2081 }
2082
2083 static void visit_load_const(struct ac_nir_context *ctx,
2084 const nir_load_const_instr *instr)
2085 {
2086 LLVMValueRef values[4], value = NULL;
2087 LLVMTypeRef element_type =
2088 LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
2089
2090 for (unsigned i = 0; i < instr->def.num_components; ++i) {
2091 switch (instr->def.bit_size) {
2092 case 32:
2093 values[i] = LLVMConstInt(element_type,
2094 instr->value.u32[i], false);
2095 break;
2096 case 64:
2097 values[i] = LLVMConstInt(element_type,
2098 instr->value.u64[i], false);
2099 break;
2100 default:
2101 fprintf(stderr,
2102 "unsupported nir load_const bit_size: %d\n",
2103 instr->def.bit_size);
2104 abort();
2105 }
2106 }
2107 if (instr->def.num_components > 1) {
2108 value = LLVMConstVector(values, instr->def.num_components);
2109 } else
2110 value = values[0];
2111
2112 _mesa_hash_table_insert(ctx->defs, &instr->def, value);
2113 }
2114
2115 static LLVMValueRef cast_ptr(struct nir_to_llvm_context *ctx, LLVMValueRef ptr,
2116 LLVMTypeRef type)
2117 {
2118 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
2119 return LLVMBuildBitCast(ctx->builder, ptr,
2120 LLVMPointerType(type, addr_space), "");
2121 }
2122
2123 static LLVMValueRef
2124 get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_elements)
2125 {
2126 LLVMValueRef size =
2127 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
2128 LLVMConstInt(ctx->ac.i32, 2, false), "");
2129
2130 /* VI only */
2131 if (ctx->ac.chip_class == VI && in_elements) {
2132 /* On VI, the descriptor contains the size in bytes,
2133 * but TXQ must return the size in elements.
2134 * The stride is always non-zero for resources using TXQ.
2135 */
2136 LLVMValueRef stride =
2137 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
2138 ctx->ac.i32_1, "");
2139 stride = LLVMBuildLShr(ctx->ac.builder, stride,
2140 LLVMConstInt(ctx->ac.i32, 16, false), "");
2141 stride = LLVMBuildAnd(ctx->ac.builder, stride,
2142 LLVMConstInt(ctx->ac.i32, 0x3fff, false), "");
2143
2144 size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
2145 }
2146 return size;
2147 }
2148
2149 /**
2150 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
2151 * intrinsic names).
2152 */
2153 static void build_int_type_name(
2154 LLVMTypeRef type,
2155 char *buf, unsigned bufsize)
2156 {
2157 assert(bufsize >= 6);
2158
2159 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
2160 snprintf(buf, bufsize, "v%ui32",
2161 LLVMGetVectorSize(type));
2162 else
2163 strcpy(buf, "i32");
2164 }
2165
2166 static LLVMValueRef radv_lower_gather4_integer(struct ac_llvm_context *ctx,
2167 struct ac_image_args *args,
2168 const nir_tex_instr *instr)
2169 {
2170 enum glsl_base_type stype = glsl_get_sampler_result_type(instr->texture->var->type);
2171 LLVMValueRef coord = args->addr;
2172 LLVMValueRef half_texel[2];
2173 LLVMValueRef compare_cube_wa = NULL;
2174 LLVMValueRef result;
2175 int c;
2176 unsigned coord_vgpr_index = (unsigned)args->offset + (unsigned)args->compare;
2177
2178 //TODO Rect
2179 {
2180 struct ac_image_args txq_args = { 0 };
2181
2182 txq_args.da = instr->is_array || instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE;
2183 txq_args.opcode = ac_image_get_resinfo;
2184 txq_args.dmask = 0xf;
2185 txq_args.addr = ctx->i32_0;
2186 txq_args.resource = args->resource;
2187 LLVMValueRef size = ac_build_image_opcode(ctx, &txq_args);
2188
2189 for (c = 0; c < 2; c++) {
2190 half_texel[c] = LLVMBuildExtractElement(ctx->builder, size,
2191 LLVMConstInt(ctx->i32, c, false), "");
2192 half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
2193 half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
2194 half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
2195 LLVMConstReal(ctx->f32, -0.5), "");
2196 }
2197 }
2198
2199 LLVMValueRef orig_coords = args->addr;
2200
2201 for (c = 0; c < 2; c++) {
2202 LLVMValueRef tmp;
2203 LLVMValueRef index = LLVMConstInt(ctx->i32, coord_vgpr_index + c, 0);
2204 tmp = LLVMBuildExtractElement(ctx->builder, coord, index, "");
2205 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
2206 tmp = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
2207 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
2208 coord = LLVMBuildInsertElement(ctx->builder, coord, tmp, index, "");
2209 }
2210
2211
2212 /*
2213 * Apparantly cube has issue with integer types that the workaround doesn't solve,
2214 * so this tests if the format is 8_8_8_8 and an integer type do an alternate
2215 * workaround by sampling using a scaled type and converting.
2216 * This is taken from amdgpu-pro shaders.
2217 */
2218 /* NOTE this produces some ugly code compared to amdgpu-pro,
2219 * LLVM ends up dumping SGPRs into VGPRs to deal with the compare/select,
2220 * and then reads them back. -pro generates two selects,
2221 * one s_cmp for the descriptor rewriting
2222 * one v_cmp for the coordinate and result changes.
2223 */
2224 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
2225 LLVMValueRef tmp, tmp2;
2226
2227 /* workaround 8/8/8/8 uint/sint cube gather bug */
2228 /* first detect it then change to a scaled read and f2i */
2229 tmp = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
2230 tmp2 = tmp;
2231
2232 /* extract the DATA_FORMAT */
2233 tmp = ac_build_bfe(ctx, tmp, LLVMConstInt(ctx->i32, 20, false),
2234 LLVMConstInt(ctx->i32, 6, false), false);
2235
2236 /* is the DATA_FORMAT == 8_8_8_8 */
2237 compare_cube_wa = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tmp, LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false), "");
2238
2239 if (stype == GLSL_TYPE_UINT)
2240 /* Create a NUM FORMAT - 0x2 or 0x4 - USCALED or UINT */
2241 tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, LLVMConstInt(ctx->i32, 0x8000000, false),
2242 LLVMConstInt(ctx->i32, 0x10000000, false), "");
2243 else
2244 /* Create a NUM FORMAT - 0x3 or 0x5 - SSCALED or SINT */
2245 tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, LLVMConstInt(ctx->i32, 0xc000000, false),
2246 LLVMConstInt(ctx->i32, 0x14000000, false), "");
2247
2248 /* replace the NUM FORMAT in the descriptor */
2249 tmp2 = LLVMBuildAnd(ctx->builder, tmp2, LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT_GFX6, false), "");
2250 tmp2 = LLVMBuildOr(ctx->builder, tmp2, tmp, "");
2251
2252 args->resource = LLVMBuildInsertElement(ctx->builder, args->resource, tmp2, ctx->i32_1, "");
2253
2254 /* don't modify the coordinates for this case */
2255 coord = LLVMBuildSelect(ctx->builder, compare_cube_wa, orig_coords, coord, "");
2256 }
2257 args->addr = coord;
2258 result = ac_build_image_opcode(ctx, args);
2259
2260 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
2261 LLVMValueRef tmp, tmp2;
2262
2263 /* if the cube workaround is in place, f2i the result. */
2264 for (c = 0; c < 4; c++) {
2265 tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
2266 if (stype == GLSL_TYPE_UINT)
2267 tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
2268 else
2269 tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
2270 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
2271 tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
2272 tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, tmp2, tmp, "");
2273 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
2274 result = LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
2275 }
2276 }
2277 return result;
2278 }
2279
2280 static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
2281 const nir_tex_instr *instr,
2282 bool lod_is_zero,
2283 struct ac_image_args *args)
2284 {
2285 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
2286 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
2287
2288 return ac_build_buffer_load_format(&ctx->ac,
2289 args->resource,
2290 args->addr,
2291 ctx->ac.i32_0,
2292 util_last_bit(mask),
2293 false, true);
2294 }
2295
2296 args->opcode = ac_image_sample;
2297 args->compare = instr->is_shadow;
2298
2299 switch (instr->op) {
2300 case nir_texop_txf:
2301 case nir_texop_txf_ms:
2302 case nir_texop_samples_identical:
2303 args->opcode = lod_is_zero ||
2304 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ?
2305 ac_image_load : ac_image_load_mip;
2306 args->compare = false;
2307 args->offset = false;
2308 break;
2309 case nir_texop_txb:
2310 args->bias = true;
2311 break;
2312 case nir_texop_txl:
2313 if (lod_is_zero)
2314 args->level_zero = true;
2315 else
2316 args->lod = true;
2317 break;
2318 case nir_texop_txs:
2319 case nir_texop_query_levels:
2320 args->opcode = ac_image_get_resinfo;
2321 break;
2322 case nir_texop_tex:
2323 if (ctx->stage != MESA_SHADER_FRAGMENT)
2324 args->level_zero = true;
2325 break;
2326 case nir_texop_txd:
2327 args->deriv = true;
2328 break;
2329 case nir_texop_tg4:
2330 args->opcode = ac_image_gather4;
2331 args->level_zero = true;
2332 break;
2333 case nir_texop_lod:
2334 args->opcode = ac_image_get_lod;
2335 args->compare = false;
2336 args->offset = false;
2337 break;
2338 default:
2339 break;
2340 }
2341
2342 if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= VI) {
2343 enum glsl_base_type stype = glsl_get_sampler_result_type(instr->texture->var->type);
2344 if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
2345 return radv_lower_gather4_integer(&ctx->ac, args, instr);
2346 }
2347 }
2348 return ac_build_image_opcode(&ctx->ac, args);
2349 }
2350
2351 static LLVMValueRef visit_vulkan_resource_index(struct nir_to_llvm_context *ctx,
2352 nir_intrinsic_instr *instr)
2353 {
2354 LLVMValueRef index = get_src(ctx->nir, instr->src[0]);
2355 unsigned desc_set = nir_intrinsic_desc_set(instr);
2356 unsigned binding = nir_intrinsic_binding(instr);
2357 LLVMValueRef desc_ptr = ctx->descriptor_sets[desc_set];
2358 struct radv_pipeline_layout *pipeline_layout = ctx->options->layout;
2359 struct radv_descriptor_set_layout *layout = pipeline_layout->set[desc_set].layout;
2360 unsigned base_offset = layout->binding[binding].offset;
2361 LLVMValueRef offset, stride;
2362
2363 if (layout->binding[binding].type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC ||
2364 layout->binding[binding].type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC) {
2365 unsigned idx = pipeline_layout->set[desc_set].dynamic_offset_start +
2366 layout->binding[binding].dynamic_offset_offset;
2367 desc_ptr = ctx->push_constants;
2368 base_offset = pipeline_layout->push_constant_size + 16 * idx;
2369 stride = LLVMConstInt(ctx->ac.i32, 16, false);
2370 } else
2371 stride = LLVMConstInt(ctx->ac.i32, layout->binding[binding].size, false);
2372
2373 offset = LLVMConstInt(ctx->ac.i32, base_offset, false);
2374 index = LLVMBuildMul(ctx->builder, index, stride, "");
2375 offset = LLVMBuildAdd(ctx->builder, offset, index, "");
2376
2377 desc_ptr = ac_build_gep0(&ctx->ac, desc_ptr, offset);
2378 desc_ptr = cast_ptr(ctx, desc_ptr, ctx->ac.v4i32);
2379 LLVMSetMetadata(desc_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
2380
2381 return desc_ptr;
2382 }
2383
2384 static LLVMValueRef visit_vulkan_resource_reindex(struct nir_to_llvm_context *ctx,
2385 nir_intrinsic_instr *instr)
2386 {
2387 LLVMValueRef ptr = get_src(ctx->nir, instr->src[0]);
2388 LLVMValueRef index = get_src(ctx->nir, instr->src[1]);
2389
2390 LLVMValueRef result = LLVMBuildGEP(ctx->builder, ptr, &index, 1, "");
2391 LLVMSetMetadata(result, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
2392 return result;
2393 }
2394
2395 static LLVMValueRef visit_load_push_constant(struct nir_to_llvm_context *ctx,
2396 nir_intrinsic_instr *instr)
2397 {
2398 LLVMValueRef ptr, addr;
2399
2400 addr = LLVMConstInt(ctx->ac.i32, nir_intrinsic_base(instr), 0);
2401 addr = LLVMBuildAdd(ctx->builder, addr, get_src(ctx->nir, instr->src[0]), "");
2402
2403 ptr = ac_build_gep0(&ctx->ac, ctx->push_constants, addr);
2404 ptr = cast_ptr(ctx, ptr, get_def_type(ctx->nir, &instr->dest.ssa));
2405
2406 return LLVMBuildLoad(ctx->builder, ptr, "");
2407 }
2408
2409 static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
2410 const nir_intrinsic_instr *instr)
2411 {
2412 LLVMValueRef index = get_src(ctx, instr->src[0]);
2413
2414 return get_buffer_size(ctx, ctx->abi->load_ssbo(ctx->abi, index, false), false);
2415 }
2416
2417 static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
2418 {
2419 uint32_t new_mask = 0;
2420 for(unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
2421 if (mask & (1u << i))
2422 new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
2423 return new_mask;
2424 }
2425
2426 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
2427 unsigned start, unsigned count)
2428 {
2429 LLVMTypeRef type = LLVMTypeOf(src);
2430
2431 if (LLVMGetTypeKind(type) != LLVMVectorTypeKind) {
2432 assert(start == 0);
2433 assert(count == 1);
2434 return src;
2435 }
2436
2437 unsigned src_elements = LLVMGetVectorSize(type);
2438 assert(start < src_elements);
2439 assert(start + count <= src_elements);
2440
2441 if (start == 0 && count == src_elements)
2442 return src;
2443
2444 if (count == 1)
2445 return LLVMBuildExtractElement(ctx->builder, src, LLVMConstInt(ctx->i32, start, false), "");
2446
2447 assert(count <= 8);
2448 LLVMValueRef indices[8];
2449 for (unsigned i = 0; i < count; ++i)
2450 indices[i] = LLVMConstInt(ctx->i32, start + i, false);
2451
2452 LLVMValueRef swizzle = LLVMConstVector(indices, count);
2453 return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
2454 }
2455
2456 static void visit_store_ssbo(struct ac_nir_context *ctx,
2457 nir_intrinsic_instr *instr)
2458 {
2459 const char *store_name;
2460 LLVMValueRef src_data = get_src(ctx, instr->src[0]);
2461 LLVMTypeRef data_type = ctx->ac.f32;
2462 int elem_size_mult = get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 32;
2463 int components_32bit = elem_size_mult * instr->num_components;
2464 unsigned writemask = nir_intrinsic_write_mask(instr);
2465 LLVMValueRef base_data, base_offset;
2466 LLVMValueRef params[6];
2467
2468 params[1] = ctx->abi->load_ssbo(ctx->abi,
2469 get_src(ctx, instr->src[1]), true);
2470 params[2] = ctx->ac.i32_0; /* vindex */
2471 params[4] = ctx->ac.i1false; /* glc */
2472 params[5] = ctx->ac.i1false; /* slc */
2473
2474 if (components_32bit > 1)
2475 data_type = LLVMVectorType(ctx->ac.f32, components_32bit);
2476
2477 writemask = widen_mask(writemask, elem_size_mult);
2478
2479 base_data = ac_to_float(&ctx->ac, src_data);
2480 base_data = trim_vector(&ctx->ac, base_data, instr->num_components);
2481 base_data = LLVMBuildBitCast(ctx->ac.builder, base_data,
2482 data_type, "");
2483 base_offset = get_src(ctx, instr->src[2]); /* voffset */
2484 while (writemask) {
2485 int start, count;
2486 LLVMValueRef data;
2487 LLVMValueRef offset;
2488
2489 u_bit_scan_consecutive_range(&writemask, &start, &count);
2490
2491 /* Due to an LLVM limitation, split 3-element writes
2492 * into a 2-element and a 1-element write. */
2493 if (count == 3) {
2494 writemask |= 1 << (start + 2);
2495 count = 2;
2496 }
2497
2498 if (count > 4) {
2499 writemask |= ((1u << (count - 4)) - 1u) << (start + 4);
2500 count = 4;
2501 }
2502
2503 if (count == 4) {
2504 store_name = "llvm.amdgcn.buffer.store.v4f32";
2505 } else if (count == 2) {
2506 store_name = "llvm.amdgcn.buffer.store.v2f32";
2507
2508 } else {
2509 assert(count == 1);
2510 store_name = "llvm.amdgcn.buffer.store.f32";
2511 }
2512 data = extract_vector_range(&ctx->ac, base_data, start, count);
2513
2514 offset = base_offset;
2515 if (start != 0) {
2516 offset = LLVMBuildAdd(ctx->ac.builder, offset, LLVMConstInt(ctx->ac.i32, start * 4, false), "");
2517 }
2518 params[0] = data;
2519 params[3] = offset;
2520 ac_build_intrinsic(&ctx->ac, store_name,
2521 ctx->ac.voidt, params, 6, 0);
2522 }
2523 }
2524
2525 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
2526 const nir_intrinsic_instr *instr)
2527 {
2528 const char *name;
2529 LLVMValueRef params[6];
2530 int arg_count = 0;
2531
2532 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
2533 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
2534 }
2535 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2536 params[arg_count++] = ctx->abi->load_ssbo(ctx->abi,
2537 get_src(ctx, instr->src[0]),
2538 true);
2539 params[arg_count++] = ctx->ac.i32_0; /* vindex */
2540 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
2541 params[arg_count++] = LLVMConstInt(ctx->ac.i1, 0, false); /* slc */
2542
2543 switch (instr->intrinsic) {
2544 case nir_intrinsic_ssbo_atomic_add:
2545 name = "llvm.amdgcn.buffer.atomic.add";
2546 break;
2547 case nir_intrinsic_ssbo_atomic_imin:
2548 name = "llvm.amdgcn.buffer.atomic.smin";
2549 break;
2550 case nir_intrinsic_ssbo_atomic_umin:
2551 name = "llvm.amdgcn.buffer.atomic.umin";
2552 break;
2553 case nir_intrinsic_ssbo_atomic_imax:
2554 name = "llvm.amdgcn.buffer.atomic.smax";
2555 break;
2556 case nir_intrinsic_ssbo_atomic_umax:
2557 name = "llvm.amdgcn.buffer.atomic.umax";
2558 break;
2559 case nir_intrinsic_ssbo_atomic_and:
2560 name = "llvm.amdgcn.buffer.atomic.and";
2561 break;
2562 case nir_intrinsic_ssbo_atomic_or:
2563 name = "llvm.amdgcn.buffer.atomic.or";
2564 break;
2565 case nir_intrinsic_ssbo_atomic_xor:
2566 name = "llvm.amdgcn.buffer.atomic.xor";
2567 break;
2568 case nir_intrinsic_ssbo_atomic_exchange:
2569 name = "llvm.amdgcn.buffer.atomic.swap";
2570 break;
2571 case nir_intrinsic_ssbo_atomic_comp_swap:
2572 name = "llvm.amdgcn.buffer.atomic.cmpswap";
2573 break;
2574 default:
2575 abort();
2576 }
2577
2578 return ac_build_intrinsic(&ctx->ac, name, ctx->ac.i32, params, arg_count, 0);
2579 }
2580
2581 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
2582 const nir_intrinsic_instr *instr)
2583 {
2584 LLVMValueRef results[2];
2585 int load_components;
2586 int num_components = instr->num_components;
2587 if (instr->dest.ssa.bit_size == 64)
2588 num_components *= 2;
2589
2590 for (int i = 0; i < num_components; i += load_components) {
2591 load_components = MIN2(num_components - i, 4);
2592 const char *load_name;
2593 LLVMTypeRef data_type = ctx->ac.f32;
2594 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * 4, false);
2595 offset = LLVMBuildAdd(ctx->ac.builder, get_src(ctx, instr->src[1]), offset, "");
2596
2597 if (load_components == 3)
2598 data_type = LLVMVectorType(ctx->ac.f32, 4);
2599 else if (load_components > 1)
2600 data_type = LLVMVectorType(ctx->ac.f32, load_components);
2601
2602 if (load_components >= 3)
2603 load_name = "llvm.amdgcn.buffer.load.v4f32";
2604 else if (load_components == 2)
2605 load_name = "llvm.amdgcn.buffer.load.v2f32";
2606 else if (load_components == 1)
2607 load_name = "llvm.amdgcn.buffer.load.f32";
2608 else
2609 unreachable("unhandled number of components");
2610
2611 LLVMValueRef params[] = {
2612 ctx->abi->load_ssbo(ctx->abi,
2613 get_src(ctx, instr->src[0]),
2614 false),
2615 ctx->ac.i32_0,
2616 offset,
2617 ctx->ac.i1false,
2618 ctx->ac.i1false,
2619 };
2620
2621 results[i > 0 ? 1 : 0] = ac_build_intrinsic(&ctx->ac, load_name, data_type, params, 5, 0);
2622 }
2623
2624 assume(results[0]);
2625 LLVMValueRef ret = results[0];
2626 if (num_components > 4 || num_components == 3) {
2627 LLVMValueRef masks[] = {
2628 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
2629 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
2630 LLVMConstInt(ctx->ac.i32, 4, false), LLVMConstInt(ctx->ac.i32, 5, false),
2631 LLVMConstInt(ctx->ac.i32, 6, false), LLVMConstInt(ctx->ac.i32, 7, false)
2632 };
2633
2634 LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
2635 ret = LLVMBuildShuffleVector(ctx->ac.builder, results[0],
2636 results[num_components > 4 ? 1 : 0], swizzle, "");
2637 }
2638
2639 return LLVMBuildBitCast(ctx->ac.builder, ret,
2640 get_def_type(ctx, &instr->dest.ssa), "");
2641 }
2642
2643 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
2644 const nir_intrinsic_instr *instr)
2645 {
2646 LLVMValueRef ret;
2647 LLVMValueRef rsrc = get_src(ctx, instr->src[0]);
2648 LLVMValueRef offset = get_src(ctx, instr->src[1]);
2649 int num_components = instr->num_components;
2650
2651 if (ctx->abi->load_ubo)
2652 rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
2653
2654 if (instr->dest.ssa.bit_size == 64)
2655 num_components *= 2;
2656
2657 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
2658 NULL, 0, false, false, true, true);
2659 ret = trim_vector(&ctx->ac, ret, num_components);
2660 return LLVMBuildBitCast(ctx->ac.builder, ret,
2661 get_def_type(ctx, &instr->dest.ssa), "");
2662 }
2663
2664 static void
2665 get_deref_offset(struct ac_nir_context *ctx, nir_deref_var *deref,
2666 bool vs_in, unsigned *vertex_index_out,
2667 LLVMValueRef *vertex_index_ref,
2668 unsigned *const_out, LLVMValueRef *indir_out)
2669 {
2670 unsigned const_offset = 0;
2671 nir_deref *tail = &deref->deref;
2672 LLVMValueRef offset = NULL;
2673
2674 if (vertex_index_out != NULL || vertex_index_ref != NULL) {
2675 tail = tail->child;
2676 nir_deref_array *deref_array = nir_deref_as_array(tail);
2677 if (vertex_index_out)
2678 *vertex_index_out = deref_array->base_offset;
2679
2680 if (vertex_index_ref) {
2681 LLVMValueRef vtx = LLVMConstInt(ctx->ac.i32, deref_array->base_offset, false);
2682 if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
2683 vtx = LLVMBuildAdd(ctx->ac.builder, vtx, get_src(ctx, deref_array->indirect), "");
2684 }
2685 *vertex_index_ref = vtx;
2686 }
2687 }
2688
2689 if (deref->var->data.compact) {
2690 assert(tail->child->deref_type == nir_deref_type_array);
2691 assert(glsl_type_is_scalar(glsl_without_array(deref->var->type)));
2692 nir_deref_array *deref_array = nir_deref_as_array(tail->child);
2693 /* We always lower indirect dereferences for "compact" array vars. */
2694 assert(deref_array->deref_array_type == nir_deref_array_type_direct);
2695
2696 const_offset = deref_array->base_offset;
2697 goto out;
2698 }
2699
2700 while (tail->child != NULL) {
2701 const struct glsl_type *parent_type = tail->type;
2702 tail = tail->child;
2703
2704 if (tail->deref_type == nir_deref_type_array) {
2705 nir_deref_array *deref_array = nir_deref_as_array(tail);
2706 LLVMValueRef index, stride, local_offset;
2707 unsigned size = glsl_count_attribute_slots(tail->type, vs_in);
2708
2709 const_offset += size * deref_array->base_offset;
2710 if (deref_array->deref_array_type == nir_deref_array_type_direct)
2711 continue;
2712
2713 assert(deref_array->deref_array_type == nir_deref_array_type_indirect);
2714 index = get_src(ctx, deref_array->indirect);
2715 stride = LLVMConstInt(ctx->ac.i32, size, 0);
2716 local_offset = LLVMBuildMul(ctx->ac.builder, stride, index, "");
2717
2718 if (offset)
2719 offset = LLVMBuildAdd(ctx->ac.builder, offset, local_offset, "");
2720 else
2721 offset = local_offset;
2722 } else if (tail->deref_type == nir_deref_type_struct) {
2723 nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
2724
2725 for (unsigned i = 0; i < deref_struct->index; i++) {
2726 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
2727 const_offset += glsl_count_attribute_slots(ft, vs_in);
2728 }
2729 } else
2730 unreachable("unsupported deref type");
2731
2732 }
2733 out:
2734 if (const_offset && offset)
2735 offset = LLVMBuildAdd(ctx->ac.builder, offset,
2736 LLVMConstInt(ctx->ac.i32, const_offset, 0),
2737 "");
2738
2739 *const_out = const_offset;
2740 *indir_out = offset;
2741 }
2742
2743
2744 /* The offchip buffer layout for TCS->TES is
2745 *
2746 * - attribute 0 of patch 0 vertex 0
2747 * - attribute 0 of patch 0 vertex 1
2748 * - attribute 0 of patch 0 vertex 2
2749 * ...
2750 * - attribute 0 of patch 1 vertex 0
2751 * - attribute 0 of patch 1 vertex 1
2752 * ...
2753 * - attribute 1 of patch 0 vertex 0
2754 * - attribute 1 of patch 0 vertex 1
2755 * ...
2756 * - per patch attribute 0 of patch 0
2757 * - per patch attribute 0 of patch 1
2758 * ...
2759 *
2760 * Note that every attribute has 4 components.
2761 */
2762 static LLVMValueRef get_tcs_tes_buffer_address(struct nir_to_llvm_context *ctx,
2763 LLVMValueRef vertex_index,
2764 LLVMValueRef param_index)
2765 {
2766 LLVMValueRef base_addr, vertices_per_patch, num_patches, total_vertices;
2767 LLVMValueRef param_stride, constant16;
2768 LLVMValueRef rel_patch_id = get_rel_patch_id(ctx);
2769
2770 vertices_per_patch = unpack_param(&ctx->ac, ctx->tcs_offchip_layout, 9, 6);
2771 num_patches = unpack_param(&ctx->ac, ctx->tcs_offchip_layout, 0, 9);
2772 total_vertices = LLVMBuildMul(ctx->builder, vertices_per_patch,
2773 num_patches, "");
2774
2775 constant16 = LLVMConstInt(ctx->ac.i32, 16, false);
2776 if (vertex_index) {
2777 base_addr = LLVMBuildMul(ctx->builder, rel_patch_id,
2778 vertices_per_patch, "");
2779
2780 base_addr = LLVMBuildAdd(ctx->builder, base_addr,
2781 vertex_index, "");
2782
2783 param_stride = total_vertices;
2784 } else {
2785 base_addr = rel_patch_id;
2786 param_stride = num_patches;
2787 }
2788
2789 base_addr = LLVMBuildAdd(ctx->builder, base_addr,
2790 LLVMBuildMul(ctx->builder, param_index,
2791 param_stride, ""), "");
2792
2793 base_addr = LLVMBuildMul(ctx->builder, base_addr, constant16, "");
2794
2795 if (!vertex_index) {
2796 LLVMValueRef patch_data_offset =
2797 unpack_param(&ctx->ac, ctx->tcs_offchip_layout, 16, 16);
2798
2799 base_addr = LLVMBuildAdd(ctx->builder, base_addr,
2800 patch_data_offset, "");
2801 }
2802 return base_addr;
2803 }
2804
2805 static LLVMValueRef get_tcs_tes_buffer_address_params(struct nir_to_llvm_context *ctx,
2806 unsigned param,
2807 unsigned const_index,
2808 bool is_compact,
2809 LLVMValueRef vertex_index,
2810 LLVMValueRef indir_index)
2811 {
2812 LLVMValueRef param_index;
2813
2814 if (indir_index)
2815 param_index = LLVMBuildAdd(ctx->builder, LLVMConstInt(ctx->ac.i32, param, false),
2816 indir_index, "");
2817 else {
2818 if (const_index && !is_compact)
2819 param += const_index;
2820 param_index = LLVMConstInt(ctx->ac.i32, param, false);
2821 }
2822 return get_tcs_tes_buffer_address(ctx, vertex_index, param_index);
2823 }
2824
2825 static void
2826 mark_tess_output(struct nir_to_llvm_context *ctx,
2827 bool is_patch, uint32_t param)
2828
2829 {
2830 if (is_patch) {
2831 ctx->tess_patch_outputs_written |= (1ull << param);
2832 } else
2833 ctx->tess_outputs_written |= (1ull << param);
2834 }
2835
2836 static LLVMValueRef
2837 get_dw_address(struct nir_to_llvm_context *ctx,
2838 LLVMValueRef dw_addr,
2839 unsigned param,
2840 unsigned const_index,
2841 bool compact_const_index,
2842 LLVMValueRef vertex_index,
2843 LLVMValueRef stride,
2844 LLVMValueRef indir_index)
2845
2846 {
2847
2848 if (vertex_index) {
2849 dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
2850 LLVMBuildMul(ctx->builder,
2851 vertex_index,
2852 stride, ""), "");
2853 }
2854
2855 if (indir_index)
2856 dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
2857 LLVMBuildMul(ctx->builder, indir_index,
2858 LLVMConstInt(ctx->ac.i32, 4, false), ""), "");
2859 else if (const_index && !compact_const_index)
2860 dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
2861 LLVMConstInt(ctx->ac.i32, const_index, false), "");
2862
2863 dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
2864 LLVMConstInt(ctx->ac.i32, param * 4, false), "");
2865
2866 if (const_index && compact_const_index)
2867 dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
2868 LLVMConstInt(ctx->ac.i32, const_index, false), "");
2869 return dw_addr;
2870 }
2871
2872 static LLVMValueRef
2873 load_tcs_varyings(struct ac_shader_abi *abi,
2874 LLVMValueRef vertex_index,
2875 LLVMValueRef indir_index,
2876 unsigned const_index,
2877 unsigned location,
2878 unsigned driver_location,
2879 unsigned component,
2880 unsigned num_components,
2881 bool is_patch,
2882 bool is_compact,
2883 bool load_input)
2884 {
2885 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
2886 LLVMValueRef dw_addr, stride;
2887 LLVMValueRef value[4], result;
2888 unsigned param = shader_io_get_unique_index(location);
2889
2890 if (load_input) {
2891 stride = unpack_param(&ctx->ac, ctx->tcs_in_layout, 13, 8);
2892 dw_addr = get_tcs_in_current_patch_offset(ctx);
2893 } else {
2894 if (!is_patch) {
2895 stride = unpack_param(&ctx->ac, ctx->tcs_out_layout, 13, 8);
2896 dw_addr = get_tcs_out_current_patch_offset(ctx);
2897 } else {
2898 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
2899 stride = NULL;
2900 }
2901 }
2902
2903 dw_addr = get_dw_address(ctx, dw_addr, param, const_index, is_compact, vertex_index, stride,
2904 indir_index);
2905
2906 for (unsigned i = 0; i < num_components + component; i++) {
2907 value[i] = ac_lds_load(&ctx->ac, dw_addr);
2908 dw_addr = LLVMBuildAdd(ctx->builder, dw_addr,
2909 ctx->ac.i32_1, "");
2910 }
2911 result = ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
2912 return result;
2913 }
2914
2915 static void
2916 store_tcs_output(struct ac_shader_abi *abi,
2917 LLVMValueRef vertex_index,
2918 LLVMValueRef param_index,
2919 unsigned const_index,
2920 unsigned location,
2921 unsigned driver_location,
2922 LLVMValueRef src,
2923 unsigned component,
2924 bool is_patch,
2925 bool is_compact,
2926 unsigned writemask)
2927 {
2928 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
2929 LLVMValueRef dw_addr;
2930 LLVMValueRef stride = NULL;
2931 LLVMValueRef buf_addr = NULL;
2932 unsigned param;
2933 bool store_lds = true;
2934
2935 if (is_patch) {
2936 if (!(ctx->tcs_patch_outputs_read & (1U << (location - VARYING_SLOT_PATCH0))))
2937 store_lds = false;
2938 } else {
2939 if (!(ctx->tcs_outputs_read & (1ULL << location)))
2940 store_lds = false;
2941 }
2942
2943 param = shader_io_get_unique_index(location);
2944 if (location == VARYING_SLOT_CLIP_DIST0 &&
2945 is_compact && const_index > 3) {
2946 const_index -= 3;
2947 param++;
2948 }
2949
2950 if (!is_patch) {
2951 stride = unpack_param(&ctx->ac, ctx->tcs_out_layout, 13, 8);
2952 dw_addr = get_tcs_out_current_patch_offset(ctx);
2953 } else {
2954 dw_addr = get_tcs_out_current_patch_data_offset(ctx);
2955 }
2956
2957 mark_tess_output(ctx, is_patch, param);
2958
2959 dw_addr = get_dw_address(ctx, dw_addr, param, const_index, is_compact, vertex_index, stride,
2960 param_index);
2961 buf_addr = get_tcs_tes_buffer_address_params(ctx, param, const_index, is_compact,
2962 vertex_index, param_index);
2963
2964 bool is_tess_factor = false;
2965 if (location == VARYING_SLOT_TESS_LEVEL_INNER ||
2966 location == VARYING_SLOT_TESS_LEVEL_OUTER)
2967 is_tess_factor = true;
2968
2969 unsigned base = is_compact ? const_index : 0;
2970 for (unsigned chan = 0; chan < 8; chan++) {
2971 if (!(writemask & (1 << chan)))
2972 continue;
2973 LLVMValueRef value = ac_llvm_extract_elem(&ctx->ac, src, chan - component);
2974
2975 if (store_lds || is_tess_factor) {
2976 LLVMValueRef dw_addr_chan =
2977 LLVMBuildAdd(ctx->builder, dw_addr,
2978 LLVMConstInt(ctx->ac.i32, chan, false), "");
2979 ac_lds_store(&ctx->ac, dw_addr_chan, value);
2980 }
2981
2982 if (!is_tess_factor && writemask != 0xF)
2983 ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, value, 1,
2984 buf_addr, ctx->oc_lds,
2985 4 * (base + chan), 1, 0, true, false);
2986 }
2987
2988 if (writemask == 0xF) {
2989 ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, src, 4,
2990 buf_addr, ctx->oc_lds,
2991 (base * 4), 1, 0, true, false);
2992 }
2993 }
2994
2995 static LLVMValueRef
2996 load_tes_input(struct ac_shader_abi *abi,
2997 LLVMValueRef vertex_index,
2998 LLVMValueRef param_index,
2999 unsigned const_index,
3000 unsigned location,
3001 unsigned driver_location,
3002 unsigned component,
3003 unsigned num_components,
3004 bool is_patch,
3005 bool is_compact,
3006 bool load_input)
3007 {
3008 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
3009 LLVMValueRef buf_addr;
3010 LLVMValueRef result;
3011 unsigned param = shader_io_get_unique_index(location);
3012
3013 if (location == VARYING_SLOT_CLIP_DIST0 && is_compact && const_index > 3) {
3014 const_index -= 3;
3015 param++;
3016 }
3017
3018 buf_addr = get_tcs_tes_buffer_address_params(ctx, param, const_index,
3019 is_compact, vertex_index, param_index);
3020
3021 LLVMValueRef comp_offset = LLVMConstInt(ctx->ac.i32, component * 4, false);
3022 buf_addr = LLVMBuildAdd(ctx->builder, buf_addr, comp_offset, "");
3023
3024 result = ac_build_buffer_load(&ctx->ac, ctx->hs_ring_tess_offchip, num_components, NULL,
3025 buf_addr, ctx->oc_lds, is_compact ? (4 * const_index) : 0, 1, 0, true, false);
3026 result = trim_vector(&ctx->ac, result, num_components);
3027 return result;
3028 }
3029
3030 static LLVMValueRef
3031 load_gs_input(struct ac_shader_abi *abi,
3032 unsigned location,
3033 unsigned driver_location,
3034 unsigned component,
3035 unsigned num_components,
3036 unsigned vertex_index,
3037 unsigned const_index,
3038 LLVMTypeRef type)
3039 {
3040 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
3041 LLVMValueRef vtx_offset;
3042 unsigned param, vtx_offset_param;
3043 LLVMValueRef value[4], result;
3044
3045 vtx_offset_param = vertex_index;
3046 assert(vtx_offset_param < 6);
3047 vtx_offset = LLVMBuildMul(ctx->builder, ctx->gs_vtx_offset[vtx_offset_param],
3048 LLVMConstInt(ctx->ac.i32, 4, false), "");
3049
3050 param = shader_io_get_unique_index(location);
3051
3052 for (unsigned i = component; i < num_components + component; i++) {
3053 if (ctx->ac.chip_class >= GFX9) {
3054 LLVMValueRef dw_addr = ctx->gs_vtx_offset[vtx_offset_param];
3055 dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr,
3056 LLVMConstInt(ctx->ac.i32, param * 4 + i + const_index, 0), "");
3057 value[i] = ac_lds_load(&ctx->ac, dw_addr);
3058 } else {
3059 LLVMValueRef soffset =
3060 LLVMConstInt(ctx->ac.i32,
3061 (param * 4 + i + const_index) * 256,
3062 false);
3063
3064 value[i] = ac_build_buffer_load(&ctx->ac,
3065 ctx->esgs_ring, 1,
3066 ctx->ac.i32_0,
3067 vtx_offset, soffset,
3068 0, 1, 0, true, false);
3069
3070 value[i] = LLVMBuildBitCast(ctx->builder, value[i],
3071 type, "");
3072 }
3073 }
3074 result = ac_build_varying_gather_values(&ctx->ac, value, num_components, component);
3075 result = ac_to_integer(&ctx->ac, result);
3076 return result;
3077 }
3078
3079 static LLVMValueRef
3080 build_gep_for_deref(struct ac_nir_context *ctx,
3081 nir_deref_var *deref)
3082 {
3083 struct hash_entry *entry = _mesa_hash_table_search(ctx->vars, deref->var);
3084 assert(entry->data);
3085 LLVMValueRef val = entry->data;
3086 nir_deref *tail = deref->deref.child;
3087 while (tail != NULL) {
3088 LLVMValueRef offset;
3089 switch (tail->deref_type) {
3090 case nir_deref_type_array: {
3091 nir_deref_array *array = nir_deref_as_array(tail);
3092 offset = LLVMConstInt(ctx->ac.i32, array->base_offset, 0);
3093 if (array->deref_array_type ==
3094 nir_deref_array_type_indirect) {
3095 offset = LLVMBuildAdd(ctx->ac.builder, offset,
3096 get_src(ctx,
3097 array->indirect),
3098 "");
3099 }
3100 break;
3101 }
3102 case nir_deref_type_struct: {
3103 nir_deref_struct *deref_struct =
3104 nir_deref_as_struct(tail);
3105 offset = LLVMConstInt(ctx->ac.i32,
3106 deref_struct->index, 0);
3107 break;
3108 }
3109 default:
3110 unreachable("bad deref type");
3111 }
3112 val = ac_build_gep0(&ctx->ac, val, offset);
3113 tail = tail->child;
3114 }
3115 return val;
3116 }
3117
3118 static LLVMValueRef load_tess_varyings(struct ac_nir_context *ctx,
3119 nir_intrinsic_instr *instr,
3120 bool load_inputs)
3121 {
3122 LLVMValueRef result;
3123 LLVMValueRef vertex_index = NULL;
3124 LLVMValueRef indir_index = NULL;
3125 unsigned const_index = 0;
3126 unsigned location = instr->variables[0]->var->data.location;
3127 unsigned driver_location = instr->variables[0]->var->data.driver_location;
3128 const bool is_patch = instr->variables[0]->var->data.patch;
3129 const bool is_compact = instr->variables[0]->var->data.compact;
3130
3131 get_deref_offset(ctx, instr->variables[0],
3132 false, NULL, is_patch ? NULL : &vertex_index,
3133 &const_index, &indir_index);
3134
3135 result = ctx->abi->load_tess_varyings(ctx->abi, vertex_index, indir_index,
3136 const_index, location, driver_location,
3137 instr->variables[0]->var->data.location_frac,
3138 instr->num_components,
3139 is_patch, is_compact, load_inputs);
3140 return LLVMBuildBitCast(ctx->ac.builder, result, get_def_type(ctx, &instr->dest.ssa), "");
3141 }
3142
3143 static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
3144 nir_intrinsic_instr *instr)
3145 {
3146 LLVMValueRef values[8];
3147 int idx = instr->variables[0]->var->data.driver_location;
3148 int ve = instr->dest.ssa.num_components;
3149 unsigned comp = instr->variables[0]->var->data.location_frac;
3150 LLVMValueRef indir_index;
3151 LLVMValueRef ret;
3152 unsigned const_index;
3153 unsigned stride = instr->variables[0]->var->data.compact ? 1 : 4;
3154 bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
3155 instr->variables[0]->var->data.mode == nir_var_shader_in;
3156 get_deref_offset(ctx, instr->variables[0], vs_in, NULL, NULL,
3157 &const_index, &indir_index);
3158
3159 if (instr->dest.ssa.bit_size == 64)
3160 ve *= 2;
3161
3162 switch (instr->variables[0]->var->data.mode) {
3163 case nir_var_shader_in:
3164 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
3165 ctx->stage == MESA_SHADER_TESS_EVAL) {
3166 return load_tess_varyings(ctx, instr, true);
3167 }
3168
3169 if (ctx->stage == MESA_SHADER_GEOMETRY) {
3170 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
3171 LLVMValueRef indir_index;
3172 unsigned const_index, vertex_index;
3173 get_deref_offset(ctx, instr->variables[0],
3174 false, &vertex_index, NULL,
3175 &const_index, &indir_index);
3176
3177 return ctx->abi->load_inputs(ctx->abi, instr->variables[0]->var->data.location,
3178 instr->variables[0]->var->data.driver_location,
3179 instr->variables[0]->var->data.location_frac, ve,
3180 vertex_index, const_index, type);
3181 }
3182
3183 for (unsigned chan = comp; chan < ve + comp; chan++) {
3184 if (indir_index) {
3185 unsigned count = glsl_count_attribute_slots(
3186 instr->variables[0]->var->type,
3187 ctx->stage == MESA_SHADER_VERTEX);
3188 count -= chan / 4;
3189 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
3190 &ctx->ac, ctx->abi->inputs + idx + chan, count,
3191 stride, false, true);
3192
3193 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
3194 tmp_vec,
3195 indir_index, "");
3196 } else
3197 values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
3198 }
3199 break;
3200 case nir_var_local:
3201 for (unsigned chan = 0; chan < ve; chan++) {
3202 if (indir_index) {
3203 unsigned count = glsl_count_attribute_slots(
3204 instr->variables[0]->var->type, false);
3205 count -= chan / 4;
3206 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
3207 &ctx->ac, ctx->locals + idx + chan, count,
3208 stride, true, true);
3209
3210 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
3211 tmp_vec,
3212 indir_index, "");
3213 } else {
3214 values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * stride], "");
3215 }
3216 }
3217 break;
3218 case nir_var_shared: {
3219 LLVMValueRef address = build_gep_for_deref(ctx,
3220 instr->variables[0]);
3221 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
3222 return LLVMBuildBitCast(ctx->ac.builder, val,
3223 get_def_type(ctx, &instr->dest.ssa),
3224 "");
3225 }
3226 case nir_var_shader_out:
3227 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3228 return load_tess_varyings(ctx, instr, false);
3229 }
3230
3231 for (unsigned chan = comp; chan < ve + comp; chan++) {
3232 if (indir_index) {
3233 unsigned count = glsl_count_attribute_slots(
3234 instr->variables[0]->var->type, false);
3235 count -= chan / 4;
3236 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
3237 &ctx->ac, ctx->outputs + idx + chan, count,
3238 stride, true, true);
3239
3240 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
3241 tmp_vec,
3242 indir_index, "");
3243 } else {
3244 values[chan] = LLVMBuildLoad(ctx->ac.builder,
3245 ctx->outputs[idx + chan + const_index * stride],
3246 "");
3247 }
3248 }
3249 break;
3250 default:
3251 unreachable("unhandle variable mode");
3252 }
3253 ret = ac_build_varying_gather_values(&ctx->ac, values, ve, comp);
3254 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
3255 }
3256
3257 static void
3258 visit_store_var(struct ac_nir_context *ctx,
3259 nir_intrinsic_instr *instr)
3260 {
3261 LLVMValueRef temp_ptr, value;
3262 int idx = instr->variables[0]->var->data.driver_location;
3263 unsigned comp = instr->variables[0]->var->data.location_frac;
3264 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
3265 int writemask = instr->const_index[0] << comp;
3266 LLVMValueRef indir_index;
3267 unsigned const_index;
3268 get_deref_offset(ctx, instr->variables[0], false,
3269 NULL, NULL, &const_index, &indir_index);
3270
3271 if (get_elem_bits(&ctx->ac, LLVMTypeOf(src)) == 64) {
3272
3273 src = LLVMBuildBitCast(ctx->ac.builder, src,
3274 LLVMVectorType(ctx->ac.f32, ac_get_llvm_num_components(src) * 2),
3275 "");
3276
3277 writemask = widen_mask(writemask, 2);
3278 }
3279
3280 switch (instr->variables[0]->var->data.mode) {
3281 case nir_var_shader_out:
3282
3283 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3284 LLVMValueRef vertex_index = NULL;
3285 LLVMValueRef indir_index = NULL;
3286 unsigned const_index = 0;
3287 const unsigned location = instr->variables[0]->var->data.location;
3288 const unsigned driver_location = instr->variables[0]->var->data.driver_location;
3289 const unsigned comp = instr->variables[0]->var->data.location_frac;
3290 const bool is_patch = instr->variables[0]->var->data.patch;
3291 const bool is_compact = instr->variables[0]->var->data.compact;
3292
3293 get_deref_offset(ctx, instr->variables[0],
3294 false, NULL, is_patch ? NULL : &vertex_index,
3295 &const_index, &indir_index);
3296
3297 ctx->abi->store_tcs_outputs(ctx->abi, vertex_index, indir_index,
3298 const_index, location, driver_location,
3299 src, comp, is_patch, is_compact, writemask);
3300 return;
3301 }
3302
3303 for (unsigned chan = 0; chan < 8; chan++) {
3304 int stride = 4;
3305 if (!(writemask & (1 << chan)))
3306 continue;
3307
3308 value = ac_llvm_extract_elem(&ctx->ac, src, chan - comp);
3309
3310 if (instr->variables[0]->var->data.compact)
3311 stride = 1;
3312 if (indir_index) {
3313 unsigned count = glsl_count_attribute_slots(
3314 instr->variables[0]->var->type, false);
3315 count -= chan / 4;
3316 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
3317 &ctx->ac, ctx->outputs + idx + chan, count,
3318 stride, true, true);
3319
3320 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
3321 value, indir_index, "");
3322 build_store_values_extended(&ctx->ac, ctx->outputs + idx + chan,
3323 count, stride, tmp_vec);
3324
3325 } else {
3326 temp_ptr = ctx->outputs[idx + chan + const_index * stride];
3327
3328 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
3329 }
3330 }
3331 break;
3332 case nir_var_local:
3333 for (unsigned chan = 0; chan < 8; chan++) {
3334 if (!(writemask & (1 << chan)))
3335 continue;
3336
3337 value = ac_llvm_extract_elem(&ctx->ac, src, chan);
3338 if (indir_index) {
3339 unsigned count = glsl_count_attribute_slots(
3340 instr->variables[0]->var->type, false);
3341 count -= chan / 4;
3342 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
3343 &ctx->ac, ctx->locals + idx + chan, count,
3344 4, true, true);
3345
3346 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
3347 value, indir_index, "");
3348 build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
3349 count, 4, tmp_vec);
3350 } else {
3351 temp_ptr = ctx->locals[idx + chan + const_index * 4];
3352
3353 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
3354 }
3355 }
3356 break;
3357 case nir_var_shared: {
3358 int writemask = instr->const_index[0];
3359 LLVMValueRef address = build_gep_for_deref(ctx,
3360 instr->variables[0]);
3361 LLVMValueRef val = get_src(ctx, instr->src[0]);
3362 unsigned components =
3363 glsl_get_vector_elements(
3364 nir_deref_tail(&instr->variables[0]->deref)->type);
3365 if (writemask == (1 << components) - 1) {
3366 val = LLVMBuildBitCast(
3367 ctx->ac.builder, val,
3368 LLVMGetElementType(LLVMTypeOf(address)), "");
3369 LLVMBuildStore(ctx->ac.builder, val, address);
3370 } else {
3371 for (unsigned chan = 0; chan < 4; chan++) {
3372 if (!(writemask & (1 << chan)))
3373 continue;
3374 LLVMValueRef ptr =
3375 LLVMBuildStructGEP(ctx->ac.builder,
3376 address, chan, "");
3377 LLVMValueRef src = ac_llvm_extract_elem(&ctx->ac, val,
3378 chan);
3379 src = LLVMBuildBitCast(
3380 ctx->ac.builder, src,
3381 LLVMGetElementType(LLVMTypeOf(ptr)), "");
3382 LLVMBuildStore(ctx->ac.builder, src, ptr);
3383 }
3384 }
3385 break;
3386 }
3387 default:
3388 break;
3389 }
3390 }
3391
3392 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
3393 {
3394 switch (dim) {
3395 case GLSL_SAMPLER_DIM_BUF:
3396 return 1;
3397 case GLSL_SAMPLER_DIM_1D:
3398 return array ? 2 : 1;
3399 case GLSL_SAMPLER_DIM_2D:
3400 return array ? 3 : 2;
3401 case GLSL_SAMPLER_DIM_MS:
3402 return array ? 4 : 3;
3403 case GLSL_SAMPLER_DIM_3D:
3404 case GLSL_SAMPLER_DIM_CUBE:
3405 return 3;
3406 case GLSL_SAMPLER_DIM_RECT:
3407 case GLSL_SAMPLER_DIM_SUBPASS:
3408 return 2;
3409 case GLSL_SAMPLER_DIM_SUBPASS_MS:
3410 return 3;
3411 default:
3412 break;
3413 }
3414 return 0;
3415 }
3416
3417
3418
3419 /* Adjust the sample index according to FMASK.
3420 *
3421 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3422 * which is the identity mapping. Each nibble says which physical sample
3423 * should be fetched to get that sample.
3424 *
3425 * For example, 0x11111100 means there are only 2 samples stored and
3426 * the second sample covers 3/4 of the pixel. When reading samples 0
3427 * and 1, return physical sample 0 (determined by the first two 0s
3428 * in FMASK), otherwise return physical sample 1.
3429 *
3430 * The sample index should be adjusted as follows:
3431 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
3432 */
3433 static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
3434 LLVMValueRef coord_x, LLVMValueRef coord_y,
3435 LLVMValueRef coord_z,
3436 LLVMValueRef sample_index,
3437 LLVMValueRef fmask_desc_ptr)
3438 {
3439 LLVMValueRef fmask_load_address[4];
3440 LLVMValueRef res;
3441
3442 fmask_load_address[0] = coord_x;
3443 fmask_load_address[1] = coord_y;
3444 if (coord_z) {
3445 fmask_load_address[2] = coord_z;
3446 fmask_load_address[3] = LLVMGetUndef(ctx->i32);
3447 }
3448
3449 struct ac_image_args args = {0};
3450
3451 args.opcode = ac_image_load;
3452 args.da = coord_z ? true : false;
3453 args.resource = fmask_desc_ptr;
3454 args.dmask = 0xf;
3455 args.addr = ac_build_gather_values(ctx, fmask_load_address, coord_z ? 4 : 2);
3456
3457 res = ac_build_image_opcode(ctx, &args);
3458
3459 res = ac_to_integer(ctx, res);
3460 LLVMValueRef four = LLVMConstInt(ctx->i32, 4, false);
3461 LLVMValueRef F = LLVMConstInt(ctx->i32, 0xf, false);
3462
3463 LLVMValueRef fmask = LLVMBuildExtractElement(ctx->builder,
3464 res,
3465 ctx->i32_0, "");
3466
3467 LLVMValueRef sample_index4 =
3468 LLVMBuildMul(ctx->builder, sample_index, four, "");
3469 LLVMValueRef shifted_fmask =
3470 LLVMBuildLShr(ctx->builder, fmask, sample_index4, "");
3471 LLVMValueRef final_sample =
3472 LLVMBuildAnd(ctx->builder, shifted_fmask, F, "");
3473
3474 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3475 * resource descriptor is 0 (invalid),
3476 */
3477 LLVMValueRef fmask_desc =
3478 LLVMBuildBitCast(ctx->builder, fmask_desc_ptr,
3479 ctx->v8i32, "");
3480
3481 LLVMValueRef fmask_word1 =
3482 LLVMBuildExtractElement(ctx->builder, fmask_desc,
3483 ctx->i32_1, "");
3484
3485 LLVMValueRef word1_is_nonzero =
3486 LLVMBuildICmp(ctx->builder, LLVMIntNE,
3487 fmask_word1, ctx->i32_0, "");
3488
3489 /* Replace the MSAA sample index. */
3490 sample_index =
3491 LLVMBuildSelect(ctx->builder, word1_is_nonzero,
3492 final_sample, sample_index, "");
3493 return sample_index;
3494 }
3495
3496 static LLVMValueRef get_image_coords(struct ac_nir_context *ctx,
3497 const nir_intrinsic_instr *instr)
3498 {
3499 const struct glsl_type *type = glsl_without_array(instr->variables[0]->var->type);
3500
3501 LLVMValueRef src0 = get_src(ctx, instr->src[0]);
3502 LLVMValueRef coords[4];
3503 LLVMValueRef masks[] = {
3504 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
3505 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
3506 };
3507 LLVMValueRef res;
3508 LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[1]), 0);
3509
3510 int count;
3511 enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
3512 bool is_array = glsl_sampler_type_is_array(type);
3513 bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
3514 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
3515 bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
3516 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
3517 bool gfx9_1d = ctx->ac.chip_class >= GFX9 && dim == GLSL_SAMPLER_DIM_1D;
3518 count = image_type_to_components_count(dim, is_array);
3519
3520 if (is_ms) {
3521 LLVMValueRef fmask_load_address[3];
3522 int chan;
3523
3524 fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
3525 fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
3526 if (is_array)
3527 fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
3528 else
3529 fmask_load_address[2] = NULL;
3530 if (add_frag_pos) {
3531 for (chan = 0; chan < 2; ++chan)
3532 fmask_load_address[chan] =
3533 LLVMBuildAdd(ctx->ac.builder, fmask_load_address[chan],
3534 LLVMBuildFPToUI(ctx->ac.builder, ctx->abi->frag_pos[chan],
3535 ctx->ac.i32, ""), "");
3536 fmask_load_address[2] = ac_to_integer(&ctx->ac, ctx->abi->inputs[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
3537 }
3538 sample_index = adjust_sample_index_using_fmask(&ctx->ac,
3539 fmask_load_address[0],
3540 fmask_load_address[1],
3541 fmask_load_address[2],
3542 sample_index,
3543 get_sampler_desc(ctx, instr->variables[0], AC_DESC_FMASK, NULL, true, false));
3544 }
3545 if (count == 1 && !gfx9_1d) {
3546 if (instr->src[0].ssa->num_components)
3547 res = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
3548 else
3549 res = src0;
3550 } else {
3551 int chan;
3552 if (is_ms)
3553 count--;
3554 for (chan = 0; chan < count; ++chan) {
3555 coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
3556 }
3557 if (add_frag_pos) {
3558 for (chan = 0; chan < 2; ++chan)
3559 coords[chan] = LLVMBuildAdd(ctx->ac.builder, coords[chan], LLVMBuildFPToUI(ctx->ac.builder, ctx->abi->frag_pos[chan],
3560 ctx->ac.i32, ""), "");
3561 coords[2] = ac_to_integer(&ctx->ac, ctx->abi->inputs[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
3562 count++;
3563 }
3564
3565 if (gfx9_1d) {
3566 if (is_array) {
3567 coords[2] = coords[1];
3568 coords[1] = ctx->ac.i32_0;
3569 } else
3570 coords[1] = ctx->ac.i32_0;
3571 count++;
3572 }
3573
3574 if (is_ms) {
3575 coords[count] = sample_index;
3576 count++;
3577 }
3578
3579 if (count == 3) {
3580 coords[3] = LLVMGetUndef(ctx->ac.i32);
3581 count = 4;
3582 }
3583 res = ac_build_gather_values(&ctx->ac, coords, count);
3584 }
3585 return res;
3586 }
3587
3588 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
3589 const nir_intrinsic_instr *instr)
3590 {
3591 LLVMValueRef params[7];
3592 LLVMValueRef res;
3593 char intrinsic_name[64];
3594 const nir_variable *var = instr->variables[0]->var;
3595 const struct glsl_type *type = var->type;
3596
3597 if(instr->variables[0]->deref.child)
3598 type = instr->variables[0]->deref.child->type;
3599
3600 type = glsl_without_array(type);
3601
3602 const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
3603 if (dim == GLSL_SAMPLER_DIM_BUF) {
3604 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
3605 unsigned num_channels = util_last_bit(mask);
3606 LLVMValueRef rsrc, vindex;
3607
3608 rsrc = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER, NULL, true, false);
3609 vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[0]),
3610 ctx->ac.i32_0, "");
3611
3612 /* TODO: set "glc" and "can_speculate" when OpenGL needs it. */
3613 res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
3614 ctx->ac.i32_0, num_channels,
3615 false, false);
3616 res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
3617
3618 res = trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
3619 res = ac_to_integer(&ctx->ac, res);
3620 } else {
3621 bool is_da = glsl_sampler_type_is_array(type) ||
3622 dim == GLSL_SAMPLER_DIM_CUBE ||
3623 dim == GLSL_SAMPLER_DIM_3D ||
3624 dim == GLSL_SAMPLER_DIM_SUBPASS ||
3625 dim == GLSL_SAMPLER_DIM_SUBPASS_MS;
3626 LLVMValueRef da = is_da ? ctx->ac.i1true : ctx->ac.i1false;
3627 LLVMValueRef glc = ctx->ac.i1false;
3628 LLVMValueRef slc = ctx->ac.i1false;
3629
3630 params[0] = get_image_coords(ctx, instr);
3631 params[1] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, NULL, true, false);
3632 params[2] = LLVMConstInt(ctx->ac.i32, 15, false); /* dmask */
3633 params[3] = glc;
3634 params[4] = slc;
3635 params[5] = ctx->ac.i1false;
3636 params[6] = da;
3637
3638 ac_get_image_intr_name("llvm.amdgcn.image.load",
3639 ctx->ac.v4f32, /* vdata */
3640 LLVMTypeOf(params[0]), /* coords */
3641 LLVMTypeOf(params[1]), /* rsrc */
3642 intrinsic_name, sizeof(intrinsic_name));
3643
3644 res = ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.v4f32,
3645 params, 7, AC_FUNC_ATTR_READONLY);
3646 }
3647 return ac_to_integer(&ctx->ac, res);
3648 }
3649
3650 static void visit_image_store(struct ac_nir_context *ctx,
3651 nir_intrinsic_instr *instr)
3652 {
3653 LLVMValueRef params[8];
3654 char intrinsic_name[64];
3655 const nir_variable *var = instr->variables[0]->var;
3656 const struct glsl_type *type = glsl_without_array(var->type);
3657 const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
3658 LLVMValueRef glc = ctx->ac.i1false;
3659 bool force_glc = ctx->ac.chip_class == SI;
3660 if (force_glc)
3661 glc = ctx->ac.i1true;
3662
3663 if (dim == GLSL_SAMPLER_DIM_BUF) {
3664 params[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[2])); /* data */
3665 params[1] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER, NULL, true, true);
3666 params[2] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[0]),
3667 ctx->ac.i32_0, ""); /* vindex */
3668 params[3] = ctx->ac.i32_0; /* voffset */
3669 params[4] = glc; /* glc */
3670 params[5] = ctx->ac.i1false; /* slc */
3671 ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.buffer.store.format.v4f32", ctx->ac.voidt,
3672 params, 6, 0);
3673 } else {
3674 bool is_da = glsl_sampler_type_is_array(type) ||
3675 dim == GLSL_SAMPLER_DIM_CUBE ||
3676 dim == GLSL_SAMPLER_DIM_3D;
3677 LLVMValueRef da = is_da ? ctx->ac.i1true : ctx->ac.i1false;
3678 LLVMValueRef slc = ctx->ac.i1false;
3679
3680 params[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[2]));
3681 params[1] = get_image_coords(ctx, instr); /* coords */
3682 params[2] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, NULL, true, true);
3683 params[3] = LLVMConstInt(ctx->ac.i32, 15, false); /* dmask */
3684 params[4] = glc;
3685 params[5] = slc;
3686 params[6] = ctx->ac.i1false;
3687 params[7] = da;
3688
3689 ac_get_image_intr_name("llvm.amdgcn.image.store",
3690 LLVMTypeOf(params[0]), /* vdata */
3691 LLVMTypeOf(params[1]), /* coords */
3692 LLVMTypeOf(params[2]), /* rsrc */
3693 intrinsic_name, sizeof(intrinsic_name));
3694
3695 ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.voidt,
3696 params, 8, 0);
3697 }
3698
3699 }
3700
3701 static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx,
3702 const nir_intrinsic_instr *instr)
3703 {
3704 LLVMValueRef params[7];
3705 int param_count = 0;
3706 const nir_variable *var = instr->variables[0]->var;
3707
3708 const char *atomic_name;
3709 char intrinsic_name[41];
3710 const struct glsl_type *type = glsl_without_array(var->type);
3711 MAYBE_UNUSED int length;
3712
3713 bool is_unsigned = glsl_get_sampler_result_type(type) == GLSL_TYPE_UINT;
3714
3715 switch (instr->intrinsic) {
3716 case nir_intrinsic_image_atomic_add:
3717 atomic_name = "add";
3718 break;
3719 case nir_intrinsic_image_atomic_min:
3720 atomic_name = is_unsigned ? "umin" : "smin";
3721 break;
3722 case nir_intrinsic_image_atomic_max:
3723 atomic_name = is_unsigned ? "umax" : "smax";
3724 break;
3725 case nir_intrinsic_image_atomic_and:
3726 atomic_name = "and";
3727 break;
3728 case nir_intrinsic_image_atomic_or:
3729 atomic_name = "or";
3730 break;
3731 case nir_intrinsic_image_atomic_xor:
3732 atomic_name = "xor";
3733 break;
3734 case nir_intrinsic_image_atomic_exchange:
3735 atomic_name = "swap";
3736 break;
3737 case nir_intrinsic_image_atomic_comp_swap:
3738 atomic_name = "cmpswap";
3739 break;
3740 default:
3741 abort();
3742 }
3743
3744 if (instr->intrinsic == nir_intrinsic_image_atomic_comp_swap)
3745 params[param_count++] = get_src(ctx, instr->src[3]);
3746 params[param_count++] = get_src(ctx, instr->src[2]);
3747
3748 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
3749 params[param_count++] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_BUFFER,
3750 NULL, true, true);
3751 params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[0]),
3752 ctx->ac.i32_0, ""); /* vindex */
3753 params[param_count++] = ctx->ac.i32_0; /* voffset */
3754 params[param_count++] = ctx->ac.i1false; /* slc */
3755
3756 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
3757 "llvm.amdgcn.buffer.atomic.%s", atomic_name);
3758 } else {
3759 char coords_type[8];
3760
3761 bool da = glsl_sampler_type_is_array(type) ||
3762 glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE;
3763
3764 LLVMValueRef coords = params[param_count++] = get_image_coords(ctx, instr);
3765 params[param_count++] = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE,
3766 NULL, true, true);
3767 params[param_count++] = ctx->ac.i1false; /* r128 */
3768 params[param_count++] = da ? ctx->ac.i1true : ctx->ac.i1false; /* da */
3769 params[param_count++] = ctx->ac.i1false; /* slc */
3770
3771 build_int_type_name(LLVMTypeOf(coords),
3772 coords_type, sizeof(coords_type));
3773
3774 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
3775 "llvm.amdgcn.image.atomic.%s.%s", atomic_name, coords_type);
3776 }
3777
3778 assert(length < sizeof(intrinsic_name));
3779 return ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32, params, param_count, 0);
3780 }
3781
3782 static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
3783 const nir_intrinsic_instr *instr)
3784 {
3785 LLVMValueRef res;
3786 const nir_variable *var = instr->variables[0]->var;
3787 const struct glsl_type *type = instr->variables[0]->var->type;
3788 bool da = glsl_sampler_type_is_array(var->type) ||
3789 glsl_get_sampler_dim(var->type) == GLSL_SAMPLER_DIM_CUBE ||
3790 glsl_get_sampler_dim(var->type) == GLSL_SAMPLER_DIM_3D;
3791 if(instr->variables[0]->deref.child)
3792 type = instr->variables[0]->deref.child->type;
3793
3794 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF)
3795 return get_buffer_size(ctx,
3796 get_sampler_desc(ctx, instr->variables[0],
3797 AC_DESC_BUFFER, NULL, true, false), true);
3798
3799 struct ac_image_args args = { 0 };
3800
3801 args.da = da;
3802 args.dmask = 0xf;
3803 args.resource = get_sampler_desc(ctx, instr->variables[0], AC_DESC_IMAGE, NULL, true, false);
3804 args.opcode = ac_image_get_resinfo;
3805 args.addr = ctx->ac.i32_0;
3806
3807 res = ac_build_image_opcode(&ctx->ac, &args);
3808
3809 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
3810
3811 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE &&
3812 glsl_sampler_type_is_array(type)) {
3813 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
3814 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
3815 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
3816 res = LLVMBuildInsertElement(ctx->ac.builder, res, z, two, "");
3817 }
3818 if (ctx->ac.chip_class >= GFX9 &&
3819 glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_1D &&
3820 glsl_sampler_type_is_array(type)) {
3821 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
3822 res = LLVMBuildInsertElement(ctx->ac.builder, res, layers,
3823 ctx->ac.i32_1, "");
3824
3825 }
3826 return res;
3827 }
3828
3829 #define NOOP_WAITCNT 0xf7f
3830 #define LGKM_CNT 0x07f
3831 #define VM_CNT 0xf70
3832
3833 static void emit_membar(struct nir_to_llvm_context *ctx,
3834 const nir_intrinsic_instr *instr)
3835 {
3836 unsigned waitcnt = NOOP_WAITCNT;
3837
3838 switch (instr->intrinsic) {
3839 case nir_intrinsic_memory_barrier:
3840 case nir_intrinsic_group_memory_barrier:
3841 waitcnt &= VM_CNT & LGKM_CNT;
3842 break;
3843 case nir_intrinsic_memory_barrier_atomic_counter:
3844 case nir_intrinsic_memory_barrier_buffer:
3845 case nir_intrinsic_memory_barrier_image:
3846 waitcnt &= VM_CNT;
3847 break;
3848 case nir_intrinsic_memory_barrier_shared:
3849 waitcnt &= LGKM_CNT;
3850 break;
3851 default:
3852 break;
3853 }
3854 if (waitcnt != NOOP_WAITCNT)
3855 ac_build_waitcnt(&ctx->ac, waitcnt);
3856 }
3857
3858 static void emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
3859 {
3860 /* SI only (thanks to a hw bug workaround):
3861 * The real barrier instruction isn’t needed, because an entire patch
3862 * always fits into a single wave.
3863 */
3864 if (ac->chip_class == SI && stage == MESA_SHADER_TESS_CTRL) {
3865 ac_build_waitcnt(ac, LGKM_CNT & VM_CNT);
3866 return;
3867 }
3868 ac_build_intrinsic(ac, "llvm.amdgcn.s.barrier",
3869 ac->voidt, NULL, 0, AC_FUNC_ATTR_CONVERGENT);
3870 }
3871
3872 static void emit_discard(struct ac_nir_context *ctx,
3873 const nir_intrinsic_instr *instr)
3874 {
3875 LLVMValueRef cond;
3876
3877 if (instr->intrinsic == nir_intrinsic_discard_if) {
3878 cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
3879 get_src(ctx, instr->src[0]),
3880 ctx->ac.i32_0, "");
3881 } else {
3882 assert(instr->intrinsic == nir_intrinsic_discard);
3883 cond = LLVMConstInt(ctx->ac.i1, false, 0);
3884 }
3885
3886 ac_build_kill_if_false(&ctx->ac, cond);
3887 }
3888
3889 static LLVMValueRef
3890 visit_load_helper_invocation(struct ac_nir_context *ctx)
3891 {
3892 LLVMValueRef result = ac_build_intrinsic(&ctx->ac,
3893 "llvm.amdgcn.ps.live",
3894 ctx->ac.i1, NULL, 0,
3895 AC_FUNC_ATTR_READNONE);
3896 result = LLVMBuildNot(ctx->ac.builder, result, "");
3897 return LLVMBuildSExt(ctx->ac.builder, result, ctx->ac.i32, "");
3898 }
3899
3900 static LLVMValueRef
3901 visit_load_local_invocation_index(struct nir_to_llvm_context *ctx)
3902 {
3903 LLVMValueRef result;
3904 LLVMValueRef thread_id = ac_get_thread_id(&ctx->ac);
3905 result = LLVMBuildAnd(ctx->builder, ctx->tg_size,
3906 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
3907
3908 return LLVMBuildAdd(ctx->builder, result, thread_id, "");
3909 }
3910
3911 static LLVMValueRef visit_var_atomic(struct nir_to_llvm_context *ctx,
3912 const nir_intrinsic_instr *instr)
3913 {
3914 LLVMValueRef ptr, result;
3915 LLVMValueRef src = get_src(ctx->nir, instr->src[0]);
3916 ptr = build_gep_for_deref(ctx->nir, instr->variables[0]);
3917
3918 if (instr->intrinsic == nir_intrinsic_var_atomic_comp_swap) {
3919 LLVMValueRef src1 = get_src(ctx->nir, instr->src[1]);
3920 result = LLVMBuildAtomicCmpXchg(ctx->builder,
3921 ptr, src, src1,
3922 LLVMAtomicOrderingSequentiallyConsistent,
3923 LLVMAtomicOrderingSequentiallyConsistent,
3924 false);
3925 } else {
3926 LLVMAtomicRMWBinOp op;
3927 switch (instr->intrinsic) {
3928 case nir_intrinsic_var_atomic_add:
3929 op = LLVMAtomicRMWBinOpAdd;
3930 break;
3931 case nir_intrinsic_var_atomic_umin:
3932 op = LLVMAtomicRMWBinOpUMin;
3933 break;
3934 case nir_intrinsic_var_atomic_umax:
3935 op = LLVMAtomicRMWBinOpUMax;
3936 break;
3937 case nir_intrinsic_var_atomic_imin:
3938 op = LLVMAtomicRMWBinOpMin;
3939 break;
3940 case nir_intrinsic_var_atomic_imax:
3941 op = LLVMAtomicRMWBinOpMax;
3942 break;
3943 case nir_intrinsic_var_atomic_and:
3944 op = LLVMAtomicRMWBinOpAnd;
3945 break;
3946 case nir_intrinsic_var_atomic_or:
3947 op = LLVMAtomicRMWBinOpOr;
3948 break;
3949 case nir_intrinsic_var_atomic_xor:
3950 op = LLVMAtomicRMWBinOpXor;
3951 break;
3952 case nir_intrinsic_var_atomic_exchange:
3953 op = LLVMAtomicRMWBinOpXchg;
3954 break;
3955 default:
3956 return NULL;
3957 }
3958
3959 result = LLVMBuildAtomicRMW(ctx->builder, op, ptr, ac_to_integer(&ctx->ac, src),
3960 LLVMAtomicOrderingSequentiallyConsistent,
3961 false);
3962 }
3963 return result;
3964 }
3965
3966 static LLVMValueRef lookup_interp_param(struct ac_shader_abi *abi,
3967 enum glsl_interp_mode interp, unsigned location)
3968 {
3969 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
3970
3971 switch (interp) {
3972 case INTERP_MODE_FLAT:
3973 default:
3974 return NULL;
3975 case INTERP_MODE_SMOOTH:
3976 case INTERP_MODE_NONE:
3977 if (location == INTERP_CENTER)
3978 return ctx->persp_center;
3979 else if (location == INTERP_CENTROID)
3980 return ctx->persp_centroid;
3981 else if (location == INTERP_SAMPLE)
3982 return ctx->persp_sample;
3983 break;
3984 case INTERP_MODE_NOPERSPECTIVE:
3985 if (location == INTERP_CENTER)
3986 return ctx->linear_center;
3987 else if (location == INTERP_CENTROID)
3988 return ctx->linear_centroid;
3989 else if (location == INTERP_SAMPLE)
3990 return ctx->linear_sample;
3991 break;
3992 }
3993 return NULL;
3994 }
3995
3996 static LLVMValueRef load_sample_position(struct ac_shader_abi *abi,
3997 LLVMValueRef sample_id)
3998 {
3999 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
4000
4001 LLVMValueRef result;
4002 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_PS_SAMPLE_POSITIONS, false));
4003
4004 ptr = LLVMBuildBitCast(ctx->builder, ptr,
4005 ac_array_in_const_addr_space(ctx->ac.v2f32), "");
4006
4007 sample_id = LLVMBuildAdd(ctx->builder, sample_id, ctx->sample_pos_offset, "");
4008 result = ac_build_load_invariant(&ctx->ac, ptr, sample_id);
4009
4010 return result;
4011 }
4012
4013 static LLVMValueRef load_sample_pos(struct ac_nir_context *ctx)
4014 {
4015 LLVMValueRef values[2];
4016
4017 values[0] = emit_ffract(&ctx->ac, ctx->abi->frag_pos[0], 32);
4018 values[1] = emit_ffract(&ctx->ac, ctx->abi->frag_pos[1], 32);
4019 return ac_build_gather_values(&ctx->ac, values, 2);
4020 }
4021
4022 static LLVMValueRef load_sample_mask_in(struct ac_nir_context *ctx)
4023 {
4024 uint8_t log2_ps_iter_samples = ctx->nctx->shader_info->info.ps.force_persample ? ctx->nctx->options->key.fs.log2_num_samples : ctx->nctx->options->key.fs.log2_ps_iter_samples;
4025
4026 /* The bit pattern matches that used by fixed function fragment
4027 * processing. */
4028 static const uint16_t ps_iter_masks[] = {
4029 0xffff, /* not used */
4030 0x5555,
4031 0x1111,
4032 0x0101,
4033 0x0001,
4034 };
4035 assert(log2_ps_iter_samples < ARRAY_SIZE(ps_iter_masks));
4036
4037 uint32_t ps_iter_mask = ps_iter_masks[log2_ps_iter_samples];
4038
4039 LLVMValueRef result, sample_id;
4040 sample_id = unpack_param(&ctx->ac, ctx->abi->ancillary, 8, 4);
4041 sample_id = LLVMBuildShl(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, ps_iter_mask, false), sample_id, "");
4042 result = LLVMBuildAnd(ctx->ac.builder, sample_id, ctx->abi->sample_coverage, "");
4043 return result;
4044 }
4045
4046 static LLVMValueRef visit_interp(struct ac_nir_context *ctx,
4047 const nir_intrinsic_instr *instr)
4048 {
4049 LLVMValueRef result[4];
4050 LLVMValueRef interp_param, attr_number;
4051 unsigned location;
4052 unsigned chan;
4053 LLVMValueRef src_c0 = NULL;
4054 LLVMValueRef src_c1 = NULL;
4055 LLVMValueRef src0 = NULL;
4056 int input_index = instr->variables[0]->var->data.location - VARYING_SLOT_VAR0;
4057 switch (instr->intrinsic) {
4058 case nir_intrinsic_interp_var_at_centroid:
4059 location = INTERP_CENTROID;
4060 break;
4061 case nir_intrinsic_interp_var_at_sample:
4062 case nir_intrinsic_interp_var_at_offset:
4063 location = INTERP_CENTER;
4064 src0 = get_src(ctx, instr->src[0]);
4065 break;
4066 default:
4067 break;
4068 }
4069
4070 if (instr->intrinsic == nir_intrinsic_interp_var_at_offset) {
4071 src_c0 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, src0, ctx->ac.i32_0, ""));
4072 src_c1 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, src0, ctx->ac.i32_1, ""));
4073 } else if (instr->intrinsic == nir_intrinsic_interp_var_at_sample) {
4074 LLVMValueRef sample_position;
4075 LLVMValueRef halfval = LLVMConstReal(ctx->ac.f32, 0.5f);
4076
4077 /* fetch sample ID */
4078 sample_position = ctx->abi->load_sample_position(ctx->abi, src0);
4079
4080 src_c0 = LLVMBuildExtractElement(ctx->ac.builder, sample_position, ctx->ac.i32_0, "");
4081 src_c0 = LLVMBuildFSub(ctx->ac.builder, src_c0, halfval, "");
4082 src_c1 = LLVMBuildExtractElement(ctx->ac.builder, sample_position, ctx->ac.i32_1, "");
4083 src_c1 = LLVMBuildFSub(ctx->ac.builder, src_c1, halfval, "");
4084 }
4085 interp_param = ctx->abi->lookup_interp_param(ctx->abi, instr->variables[0]->var->data.interpolation, location);
4086 attr_number = LLVMConstInt(ctx->ac.i32, input_index, false);
4087
4088 if (location == INTERP_CENTER) {
4089 LLVMValueRef ij_out[2];
4090 LLVMValueRef ddxy_out = emit_ddxy_interp(ctx, interp_param);
4091
4092 /*
4093 * take the I then J parameters, and the DDX/Y for it, and
4094 * calculate the IJ inputs for the interpolator.
4095 * temp1 = ddx * offset/sample.x + I;
4096 * interp_param.I = ddy * offset/sample.y + temp1;
4097 * temp1 = ddx * offset/sample.x + J;
4098 * interp_param.J = ddy * offset/sample.y + temp1;
4099 */
4100 for (unsigned i = 0; i < 2; i++) {
4101 LLVMValueRef ix_ll = LLVMConstInt(ctx->ac.i32, i, false);
4102 LLVMValueRef iy_ll = LLVMConstInt(ctx->ac.i32, i + 2, false);
4103 LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder,
4104 ddxy_out, ix_ll, "");
4105 LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder,
4106 ddxy_out, iy_ll, "");
4107 LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder,
4108 interp_param, ix_ll, "");
4109 LLVMValueRef temp1, temp2;
4110
4111 interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el,
4112 ctx->ac.f32, "");
4113
4114 temp1 = LLVMBuildFMul(ctx->ac.builder, ddx_el, src_c0, "");
4115 temp1 = LLVMBuildFAdd(ctx->ac.builder, temp1, interp_el, "");
4116
4117 temp2 = LLVMBuildFMul(ctx->ac.builder, ddy_el, src_c1, "");
4118 temp2 = LLVMBuildFAdd(ctx->ac.builder, temp2, temp1, "");
4119
4120 ij_out[i] = LLVMBuildBitCast(ctx->ac.builder,
4121 temp2, ctx->ac.i32, "");
4122 }
4123 interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
4124
4125 }
4126
4127 for (chan = 0; chan < 4; chan++) {
4128 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
4129
4130 if (interp_param) {
4131 interp_param = LLVMBuildBitCast(ctx->ac.builder,
4132 interp_param, ctx->ac.v2f32, "");
4133 LLVMValueRef i = LLVMBuildExtractElement(
4134 ctx->ac.builder, interp_param, ctx->ac.i32_0, "");
4135 LLVMValueRef j = LLVMBuildExtractElement(
4136 ctx->ac.builder, interp_param, ctx->ac.i32_1, "");
4137
4138 result[chan] = ac_build_fs_interp(&ctx->ac,
4139 llvm_chan, attr_number,
4140 ctx->abi->prim_mask, i, j);
4141 } else {
4142 result[chan] = ac_build_fs_interp_mov(&ctx->ac,
4143 LLVMConstInt(ctx->ac.i32, 2, false),
4144 llvm_chan, attr_number,
4145 ctx->abi->prim_mask);
4146 }
4147 }
4148 return ac_build_varying_gather_values(&ctx->ac, result, instr->num_components,
4149 instr->variables[0]->var->data.location_frac);
4150 }
4151
4152 static void
4153 visit_emit_vertex(struct ac_shader_abi *abi, unsigned stream, LLVMValueRef *addrs)
4154 {
4155 LLVMValueRef gs_next_vertex;
4156 LLVMValueRef can_emit;
4157 int idx;
4158 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
4159
4160 assert(stream == 0);
4161
4162 /* Write vertex attribute values to GSVS ring */
4163 gs_next_vertex = LLVMBuildLoad(ctx->builder,
4164 ctx->gs_next_vertex,
4165 "");
4166
4167 /* If this thread has already emitted the declared maximum number of
4168 * vertices, kill it: excessive vertex emissions are not supposed to
4169 * have any effect, and GS threads have no externally observable
4170 * effects other than emitting vertices.
4171 */
4172 can_emit = LLVMBuildICmp(ctx->builder, LLVMIntULT, gs_next_vertex,
4173 LLVMConstInt(ctx->ac.i32, ctx->gs_max_out_vertices, false), "");
4174 ac_build_kill_if_false(&ctx->ac, can_emit);
4175
4176 /* loop num outputs */
4177 idx = 0;
4178 for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
4179 LLVMValueRef *out_ptr = &addrs[i * 4];
4180 int length = 4;
4181 int slot = idx;
4182 int slot_inc = 1;
4183
4184 if (!(ctx->output_mask & (1ull << i)))
4185 continue;
4186
4187 if (i == VARYING_SLOT_CLIP_DIST0) {
4188 /* pack clip and cull into a single set of slots */
4189 length = ctx->num_output_clips + ctx->num_output_culls;
4190 if (length > 4)
4191 slot_inc = 2;
4192 }
4193 for (unsigned j = 0; j < length; j++) {
4194 LLVMValueRef out_val = LLVMBuildLoad(ctx->builder,
4195 out_ptr[j], "");
4196 LLVMValueRef voffset = LLVMConstInt(ctx->ac.i32, (slot * 4 + j) * ctx->gs_max_out_vertices, false);
4197 voffset = LLVMBuildAdd(ctx->builder, voffset, gs_next_vertex, "");
4198 voffset = LLVMBuildMul(ctx->builder, voffset, LLVMConstInt(ctx->ac.i32, 4, false), "");
4199
4200 out_val = LLVMBuildBitCast(ctx->builder, out_val, ctx->ac.i32, "");
4201
4202 ac_build_buffer_store_dword(&ctx->ac, ctx->gsvs_ring,
4203 out_val, 1,
4204 voffset, ctx->gs2vs_offset, 0,
4205 1, 1, true, true);
4206 }
4207 idx += slot_inc;
4208 }
4209
4210 gs_next_vertex = LLVMBuildAdd(ctx->builder, gs_next_vertex,
4211 ctx->ac.i32_1, "");
4212 LLVMBuildStore(ctx->builder, gs_next_vertex, ctx->gs_next_vertex);
4213
4214 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_EMIT | AC_SENDMSG_GS | (0 << 8), ctx->gs_wave_id);
4215 }
4216
4217 static void
4218 visit_end_primitive(struct ac_shader_abi *abi, unsigned stream)
4219 {
4220 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
4221 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_CUT | AC_SENDMSG_GS | (stream << 8), ctx->gs_wave_id);
4222 }
4223
4224 static LLVMValueRef
4225 load_tess_coord(struct ac_shader_abi *abi, LLVMTypeRef type,
4226 unsigned num_components)
4227 {
4228 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
4229
4230 LLVMValueRef coord[4] = {
4231 ctx->tes_u,
4232 ctx->tes_v,
4233 ctx->ac.f32_0,
4234 ctx->ac.f32_0,
4235 };
4236
4237 if (ctx->tes_primitive_mode == GL_TRIANGLES)
4238 coord[2] = LLVMBuildFSub(ctx->builder, ctx->ac.f32_1,
4239 LLVMBuildFAdd(ctx->builder, coord[0], coord[1], ""), "");
4240
4241 LLVMValueRef result = ac_build_gather_values(&ctx->ac, coord, num_components);
4242 return LLVMBuildBitCast(ctx->builder, result, type, "");
4243 }
4244
4245 static LLVMValueRef
4246 load_patch_vertices_in(struct ac_shader_abi *abi)
4247 {
4248 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
4249 return LLVMConstInt(ctx->ac.i32, ctx->options->key.tcs.input_vertices, false);
4250 }
4251
4252 static void visit_intrinsic(struct ac_nir_context *ctx,
4253 nir_intrinsic_instr *instr)
4254 {
4255 LLVMValueRef result = NULL;
4256
4257 switch (instr->intrinsic) {
4258 case nir_intrinsic_ballot:
4259 result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
4260 break;
4261 case nir_intrinsic_read_invocation:
4262 case nir_intrinsic_read_first_invocation: {
4263 LLVMValueRef args[2];
4264
4265 /* Value */
4266 args[0] = get_src(ctx, instr->src[0]);
4267
4268 unsigned num_args;
4269 const char *intr_name;
4270 if (instr->intrinsic == nir_intrinsic_read_invocation) {
4271 num_args = 2;
4272 intr_name = "llvm.amdgcn.readlane";
4273
4274 /* Invocation */
4275 args[1] = get_src(ctx, instr->src[1]);
4276 } else {
4277 num_args = 1;
4278 intr_name = "llvm.amdgcn.readfirstlane";
4279 }
4280
4281 /* We currently have no other way to prevent LLVM from lifting the icmp
4282 * calls to a dominating basic block.
4283 */
4284 ac_build_optimization_barrier(&ctx->ac, &args[0]);
4285
4286 result = ac_build_intrinsic(&ctx->ac, intr_name,
4287 ctx->ac.i32, args, num_args,
4288 AC_FUNC_ATTR_READNONE |
4289 AC_FUNC_ATTR_CONVERGENT);
4290 break;
4291 }
4292 case nir_intrinsic_load_subgroup_invocation:
4293 result = ac_get_thread_id(&ctx->ac);
4294 break;
4295 case nir_intrinsic_load_work_group_id: {
4296 LLVMValueRef values[3];
4297
4298 for (int i = 0; i < 3; i++) {
4299 values[i] = ctx->abi->workgroup_ids[i] ?
4300 ctx->abi->workgroup_ids[i] : ctx->ac.i32_0;
4301 }
4302
4303 result = ac_build_gather_values(&ctx->ac, values, 3);
4304 break;
4305 }
4306 case nir_intrinsic_load_base_vertex: {
4307 result = ctx->abi->base_vertex;
4308 break;
4309 }
4310 case nir_intrinsic_load_local_group_size:
4311 result = ctx->abi->load_local_group_size(ctx->abi);
4312 break;
4313 case nir_intrinsic_load_vertex_id_zero_base: {
4314 result = ctx->abi->vertex_id;
4315 break;
4316 }
4317 case nir_intrinsic_load_local_invocation_id: {
4318 result = ctx->abi->local_invocation_ids;
4319 break;
4320 }
4321 case nir_intrinsic_load_base_instance:
4322 result = ctx->abi->start_instance;
4323 break;
4324 case nir_intrinsic_load_draw_id:
4325 result = ctx->abi->draw_id;
4326 break;
4327 case nir_intrinsic_load_view_index:
4328 result = ctx->nctx->view_index ? ctx->nctx->view_index : ctx->ac.i32_0;
4329 break;
4330 case nir_intrinsic_load_invocation_id:
4331 if (ctx->stage == MESA_SHADER_TESS_CTRL)
4332 result = unpack_param(&ctx->ac, ctx->abi->tcs_rel_ids, 8, 5);
4333 else
4334 result = ctx->abi->gs_invocation_id;
4335 break;
4336 case nir_intrinsic_load_primitive_id:
4337 if (ctx->stage == MESA_SHADER_GEOMETRY) {
4338 result = ctx->abi->gs_prim_id;
4339 } else if (ctx->stage == MESA_SHADER_TESS_CTRL) {
4340 result = ctx->abi->tcs_patch_id;
4341 } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
4342 result = ctx->abi->tes_patch_id;
4343 } else
4344 fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
4345 break;
4346 case nir_intrinsic_load_sample_id:
4347 result = unpack_param(&ctx->ac, ctx->abi->ancillary, 8, 4);
4348 break;
4349 case nir_intrinsic_load_sample_pos:
4350 result = load_sample_pos(ctx);
4351 break;
4352 case nir_intrinsic_load_sample_mask_in:
4353 if (ctx->nctx)
4354 result = load_sample_mask_in(ctx);
4355 else
4356 result = ctx->abi->sample_coverage;
4357 break;
4358 case nir_intrinsic_load_frag_coord: {
4359 LLVMValueRef values[4] = {
4360 ctx->abi->frag_pos[0],
4361 ctx->abi->frag_pos[1],
4362 ctx->abi->frag_pos[2],
4363 ac_build_fdiv(&ctx->ac, ctx->ac.f32_1, ctx->abi->frag_pos[3])
4364 };
4365 result = ac_build_gather_values(&ctx->ac, values, 4);
4366 break;
4367 }
4368 case nir_intrinsic_load_front_face:
4369 result = ctx->abi->front_face;
4370 break;
4371 case nir_intrinsic_load_helper_invocation:
4372 result = visit_load_helper_invocation(ctx);
4373 break;
4374 case nir_intrinsic_load_instance_id:
4375 result = ctx->abi->instance_id;
4376 break;
4377 case nir_intrinsic_load_num_work_groups:
4378 result = ctx->abi->num_work_groups;
4379 break;
4380 case nir_intrinsic_load_local_invocation_index:
4381 result = visit_load_local_invocation_index(ctx->nctx);
4382 break;
4383 case nir_intrinsic_load_push_constant:
4384 result = visit_load_push_constant(ctx->nctx, instr);
4385 break;
4386 case nir_intrinsic_vulkan_resource_index:
4387 result = visit_vulkan_resource_index(ctx->nctx, instr);
4388 break;
4389 case nir_intrinsic_vulkan_resource_reindex:
4390 result = visit_vulkan_resource_reindex(ctx->nctx, instr);
4391 break;
4392 case nir_intrinsic_store_ssbo:
4393 visit_store_ssbo(ctx, instr);
4394 break;
4395 case nir_intrinsic_load_ssbo:
4396 result = visit_load_buffer(ctx, instr);
4397 break;
4398 case nir_intrinsic_ssbo_atomic_add:
4399 case nir_intrinsic_ssbo_atomic_imin:
4400 case nir_intrinsic_ssbo_atomic_umin:
4401 case nir_intrinsic_ssbo_atomic_imax:
4402 case nir_intrinsic_ssbo_atomic_umax:
4403 case nir_intrinsic_ssbo_atomic_and:
4404 case nir_intrinsic_ssbo_atomic_or:
4405 case nir_intrinsic_ssbo_atomic_xor:
4406 case nir_intrinsic_ssbo_atomic_exchange:
4407 case nir_intrinsic_ssbo_atomic_comp_swap:
4408 result = visit_atomic_ssbo(ctx, instr);
4409 break;
4410 case nir_intrinsic_load_ubo:
4411 result = visit_load_ubo_buffer(ctx, instr);
4412 break;
4413 case nir_intrinsic_get_buffer_size:
4414 result = visit_get_buffer_size(ctx, instr);
4415 break;
4416 case nir_intrinsic_load_var:
4417 result = visit_load_var(ctx, instr);
4418 break;
4419 case nir_intrinsic_store_var:
4420 visit_store_var(ctx, instr);
4421 break;
4422 case nir_intrinsic_image_load:
4423 result = visit_image_load(ctx, instr);
4424 break;
4425 case nir_intrinsic_image_store:
4426 visit_image_store(ctx, instr);
4427 break;
4428 case nir_intrinsic_image_atomic_add:
4429 case nir_intrinsic_image_atomic_min:
4430 case nir_intrinsic_image_atomic_max:
4431 case nir_intrinsic_image_atomic_and:
4432 case nir_intrinsic_image_atomic_or:
4433 case nir_intrinsic_image_atomic_xor:
4434 case nir_intrinsic_image_atomic_exchange:
4435 case nir_intrinsic_image_atomic_comp_swap:
4436 result = visit_image_atomic(ctx, instr);
4437 break;
4438 case nir_intrinsic_image_size:
4439 result = visit_image_size(ctx, instr);
4440 break;
4441 case nir_intrinsic_shader_clock:
4442 result = ac_build_shader_clock(&ctx->ac);
4443 break;
4444 case nir_intrinsic_discard:
4445 case nir_intrinsic_discard_if:
4446 emit_discard(ctx, instr);
4447 break;
4448 case nir_intrinsic_memory_barrier:
4449 case nir_intrinsic_group_memory_barrier:
4450 case nir_intrinsic_memory_barrier_atomic_counter:
4451 case nir_intrinsic_memory_barrier_buffer:
4452 case nir_intrinsic_memory_barrier_image:
4453 case nir_intrinsic_memory_barrier_shared:
4454 emit_membar(ctx->nctx, instr);
4455 break;
4456 case nir_intrinsic_barrier:
4457 emit_barrier(&ctx->ac, ctx->stage);
4458 break;
4459 case nir_intrinsic_var_atomic_add:
4460 case nir_intrinsic_var_atomic_imin:
4461 case nir_intrinsic_var_atomic_umin:
4462 case nir_intrinsic_var_atomic_imax:
4463 case nir_intrinsic_var_atomic_umax:
4464 case nir_intrinsic_var_atomic_and:
4465 case nir_intrinsic_var_atomic_or:
4466 case nir_intrinsic_var_atomic_xor:
4467 case nir_intrinsic_var_atomic_exchange:
4468 case nir_intrinsic_var_atomic_comp_swap:
4469 result = visit_var_atomic(ctx->nctx, instr);
4470 break;
4471 case nir_intrinsic_interp_var_at_centroid:
4472 case nir_intrinsic_interp_var_at_sample:
4473 case nir_intrinsic_interp_var_at_offset:
4474 result = visit_interp(ctx, instr);
4475 break;
4476 case nir_intrinsic_emit_vertex:
4477 ctx->abi->emit_vertex(ctx->abi, nir_intrinsic_stream_id(instr), ctx->outputs);
4478 break;
4479 case nir_intrinsic_end_primitive:
4480 ctx->abi->emit_primitive(ctx->abi, nir_intrinsic_stream_id(instr));
4481 break;
4482 case nir_intrinsic_load_tess_coord: {
4483 LLVMTypeRef type = ctx->nctx ?
4484 get_def_type(ctx->nctx->nir, &instr->dest.ssa) :
4485 NULL;
4486 result = ctx->abi->load_tess_coord(ctx->abi, type, instr->num_components);
4487 break;
4488 }
4489 case nir_intrinsic_load_tess_level_outer:
4490 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER);
4491 break;
4492 case nir_intrinsic_load_tess_level_inner:
4493 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER);
4494 break;
4495 case nir_intrinsic_load_patch_vertices_in:
4496 result = ctx->abi->load_patch_vertices_in(ctx->abi);
4497 break;
4498 case nir_intrinsic_vote_all: {
4499 LLVMValueRef tmp = ac_build_vote_all(&ctx->ac, get_src(ctx, instr->src[0]));
4500 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
4501 break;
4502 }
4503 case nir_intrinsic_vote_any: {
4504 LLVMValueRef tmp = ac_build_vote_any(&ctx->ac, get_src(ctx, instr->src[0]));
4505 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
4506 break;
4507 }
4508 case nir_intrinsic_vote_eq: {
4509 LLVMValueRef tmp = ac_build_vote_eq(&ctx->ac, get_src(ctx, instr->src[0]));
4510 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
4511 break;
4512 }
4513 default:
4514 fprintf(stderr, "Unknown intrinsic: ");
4515 nir_print_instr(&instr->instr, stderr);
4516 fprintf(stderr, "\n");
4517 break;
4518 }
4519 if (result) {
4520 _mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
4521 }
4522 }
4523
4524 static LLVMValueRef radv_load_ssbo(struct ac_shader_abi *abi,
4525 LLVMValueRef buffer_ptr, bool write)
4526 {
4527 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
4528 LLVMValueRef result;
4529
4530 LLVMSetMetadata(buffer_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
4531
4532 result = LLVMBuildLoad(ctx->builder, buffer_ptr, "");
4533 LLVMSetMetadata(result, ctx->ac.invariant_load_md_kind, ctx->ac.empty_md);
4534
4535 return result;
4536 }
4537
4538 static LLVMValueRef radv_load_ubo(struct ac_shader_abi *abi, LLVMValueRef buffer_ptr)
4539 {
4540 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
4541 LLVMValueRef result;
4542
4543 LLVMSetMetadata(buffer_ptr, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
4544
4545 result = LLVMBuildLoad(ctx->builder, buffer_ptr, "");
4546 LLVMSetMetadata(result, ctx->ac.invariant_load_md_kind, ctx->ac.empty_md);
4547
4548 return result;
4549 }
4550
4551 static LLVMValueRef radv_get_sampler_desc(struct ac_shader_abi *abi,
4552 unsigned descriptor_set,
4553 unsigned base_index,
4554 unsigned constant_index,
4555 LLVMValueRef index,
4556 enum ac_descriptor_type desc_type,
4557 bool image, bool write)
4558 {
4559 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
4560 LLVMValueRef list = ctx->descriptor_sets[descriptor_set];
4561 struct radv_descriptor_set_layout *layout = ctx->options->layout->set[descriptor_set].layout;
4562 struct radv_descriptor_set_binding_layout *binding = layout->binding + base_index;
4563 unsigned offset = binding->offset;
4564 unsigned stride = binding->size;
4565 unsigned type_size;
4566 LLVMBuilderRef builder = ctx->builder;
4567 LLVMTypeRef type;
4568
4569 assert(base_index < layout->binding_count);
4570
4571 switch (desc_type) {
4572 case AC_DESC_IMAGE:
4573 type = ctx->ac.v8i32;
4574 type_size = 32;
4575 break;
4576 case AC_DESC_FMASK:
4577 type = ctx->ac.v8i32;
4578 offset += 32;
4579 type_size = 32;
4580 break;
4581 case AC_DESC_SAMPLER:
4582 type = ctx->ac.v4i32;
4583 if (binding->type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
4584 offset += 64;
4585
4586 type_size = 16;
4587 break;
4588 case AC_DESC_BUFFER:
4589 type = ctx->ac.v4i32;
4590 type_size = 16;
4591 break;
4592 default:
4593 unreachable("invalid desc_type\n");
4594 }
4595
4596 offset += constant_index * stride;
4597
4598 if (desc_type == AC_DESC_SAMPLER && binding->immutable_samplers_offset &&
4599 (!index || binding->immutable_samplers_equal)) {
4600 if (binding->immutable_samplers_equal)
4601 constant_index = 0;
4602
4603 const uint32_t *samplers = radv_immutable_samplers(layout, binding);
4604
4605 LLVMValueRef constants[] = {
4606 LLVMConstInt(ctx->ac.i32, samplers[constant_index * 4 + 0], 0),
4607 LLVMConstInt(ctx->ac.i32, samplers[constant_index * 4 + 1], 0),
4608 LLVMConstInt(ctx->ac.i32, samplers[constant_index * 4 + 2], 0),
4609 LLVMConstInt(ctx->ac.i32, samplers[constant_index * 4 + 3], 0),
4610 };
4611 return ac_build_gather_values(&ctx->ac, constants, 4);
4612 }
4613
4614 assert(stride % type_size == 0);
4615
4616 if (!index)
4617 index = ctx->ac.i32_0;
4618
4619 index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->ac.i32, stride / type_size, 0), "");
4620
4621 list = ac_build_gep0(&ctx->ac, list, LLVMConstInt(ctx->ac.i32, offset, 0));
4622 list = LLVMBuildPointerCast(builder, list, ac_array_in_const_addr_space(type), "");
4623
4624 return ac_build_load_to_sgpr(&ctx->ac, list, index);
4625 }
4626
4627 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
4628 const nir_deref_var *deref,
4629 enum ac_descriptor_type desc_type,
4630 const nir_tex_instr *tex_instr,
4631 bool image, bool write)
4632 {
4633 LLVMValueRef index = NULL;
4634 unsigned constant_index = 0;
4635 unsigned descriptor_set;
4636 unsigned base_index;
4637
4638 if (!deref) {
4639 assert(tex_instr && !image);
4640 descriptor_set = 0;
4641 base_index = tex_instr->sampler_index;
4642 } else {
4643 const nir_deref *tail = &deref->deref;
4644 while (tail->child) {
4645 const nir_deref_array *child = nir_deref_as_array(tail->child);
4646 unsigned array_size = glsl_get_aoa_size(tail->child->type);
4647
4648 if (!array_size)
4649 array_size = 1;
4650
4651 assert(child->deref_array_type != nir_deref_array_type_wildcard);
4652
4653 if (child->deref_array_type == nir_deref_array_type_indirect) {
4654 LLVMValueRef indirect = get_src(ctx, child->indirect);
4655
4656 indirect = LLVMBuildMul(ctx->ac.builder, indirect,
4657 LLVMConstInt(ctx->ac.i32, array_size, false), "");
4658
4659 if (!index)
4660 index = indirect;
4661 else
4662 index = LLVMBuildAdd(ctx->ac.builder, index, indirect, "");
4663 }
4664
4665 constant_index += child->base_offset * array_size;
4666
4667 tail = &child->deref;
4668 }
4669 descriptor_set = deref->var->data.descriptor_set;
4670 base_index = deref->var->data.binding;
4671 }
4672
4673 return ctx->abi->load_sampler_desc(ctx->abi,
4674 descriptor_set,
4675 base_index,
4676 constant_index, index,
4677 desc_type, image, write);
4678 }
4679
4680 static void set_tex_fetch_args(struct ac_llvm_context *ctx,
4681 struct ac_image_args *args,
4682 const nir_tex_instr *instr,
4683 nir_texop op,
4684 LLVMValueRef res_ptr, LLVMValueRef samp_ptr,
4685 LLVMValueRef *param, unsigned count,
4686 unsigned dmask)
4687 {
4688 unsigned is_rect = 0;
4689 bool da = instr->is_array || instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE;
4690
4691 if (op == nir_texop_lod)
4692 da = false;
4693 /* Pad to power of two vector */
4694 while (count < util_next_power_of_two(count))
4695 param[count++] = LLVMGetUndef(ctx->i32);
4696
4697 if (count > 1)
4698 args->addr = ac_build_gather_values(ctx, param, count);
4699 else
4700 args->addr = param[0];
4701
4702 args->resource = res_ptr;
4703 args->sampler = samp_ptr;
4704
4705 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF && op == nir_texop_txf) {
4706 args->addr = param[0];
4707 return;
4708 }
4709
4710 args->dmask = dmask;
4711 args->unorm = is_rect;
4712 args->da = da;
4713 }
4714
4715 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4716 *
4717 * SI-CI:
4718 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4719 * filtering manually. The driver sets img7 to a mask clearing
4720 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4721 * s_and_b32 samp0, samp0, img7
4722 *
4723 * VI:
4724 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4725 */
4726 static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx,
4727 LLVMValueRef res, LLVMValueRef samp)
4728 {
4729 LLVMBuilderRef builder = ctx->ac.builder;
4730 LLVMValueRef img7, samp0;
4731
4732 if (ctx->ac.chip_class >= VI)
4733 return samp;
4734
4735 img7 = LLVMBuildExtractElement(builder, res,
4736 LLVMConstInt(ctx->ac.i32, 7, 0), "");
4737 samp0 = LLVMBuildExtractElement(builder, samp,
4738 LLVMConstInt(ctx->ac.i32, 0, 0), "");
4739 samp0 = LLVMBuildAnd(builder, samp0, img7, "");
4740 return LLVMBuildInsertElement(builder, samp, samp0,
4741 LLVMConstInt(ctx->ac.i32, 0, 0), "");
4742 }
4743
4744 static void tex_fetch_ptrs(struct ac_nir_context *ctx,
4745 nir_tex_instr *instr,
4746 LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr,
4747 LLVMValueRef *fmask_ptr)
4748 {
4749 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF)
4750 *res_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_BUFFER, instr, false, false);
4751 else
4752 *res_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_IMAGE, instr, false, false);
4753 if (samp_ptr) {
4754 if (instr->sampler)
4755 *samp_ptr = get_sampler_desc(ctx, instr->sampler, AC_DESC_SAMPLER, instr, false, false);
4756 else
4757 *samp_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_SAMPLER, instr, false, false);
4758 if (instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
4759 *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
4760 }
4761 if (fmask_ptr && !instr->sampler && (instr->op == nir_texop_txf_ms ||
4762 instr->op == nir_texop_samples_identical))
4763 *fmask_ptr = get_sampler_desc(ctx, instr->texture, AC_DESC_FMASK, instr, false, false);
4764 }
4765
4766 static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
4767 LLVMValueRef coord)
4768 {
4769 coord = ac_to_float(ctx, coord);
4770 coord = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &coord, 1, 0);
4771 coord = ac_to_integer(ctx, coord);
4772 return coord;
4773 }
4774
4775 static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
4776 {
4777 LLVMValueRef result = NULL;
4778 struct ac_image_args args = { 0 };
4779 unsigned dmask = 0xf;
4780 LLVMValueRef address[16];
4781 LLVMValueRef coords[5];
4782 LLVMValueRef coord = NULL, lod = NULL, comparator = NULL;
4783 LLVMValueRef bias = NULL, offsets = NULL;
4784 LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL, sample_index = NULL;
4785 LLVMValueRef ddx = NULL, ddy = NULL;
4786 LLVMValueRef derivs[6];
4787 unsigned chan, count = 0;
4788 unsigned const_src = 0, num_deriv_comp = 0;
4789 bool lod_is_zero = false;
4790
4791 tex_fetch_ptrs(ctx, instr, &res_ptr, &samp_ptr, &fmask_ptr);
4792
4793 for (unsigned i = 0; i < instr->num_srcs; i++) {
4794 switch (instr->src[i].src_type) {
4795 case nir_tex_src_coord:
4796 coord = get_src(ctx, instr->src[i].src);
4797 break;
4798 case nir_tex_src_projector:
4799 break;
4800 case nir_tex_src_comparator:
4801 comparator = get_src(ctx, instr->src[i].src);
4802 break;
4803 case nir_tex_src_offset:
4804 offsets = get_src(ctx, instr->src[i].src);
4805 const_src = i;
4806 break;
4807 case nir_tex_src_bias:
4808 bias = get_src(ctx, instr->src[i].src);
4809 break;
4810 case nir_tex_src_lod: {
4811 nir_const_value *val = nir_src_as_const_value(instr->src[i].src);
4812
4813 if (val && val->i32[0] == 0)
4814 lod_is_zero = true;
4815 lod = get_src(ctx, instr->src[i].src);
4816 break;
4817 }
4818 case nir_tex_src_ms_index:
4819 sample_index = get_src(ctx, instr->src[i].src);
4820 break;
4821 case nir_tex_src_ms_mcs:
4822 break;
4823 case nir_tex_src_ddx:
4824 ddx = get_src(ctx, instr->src[i].src);
4825 num_deriv_comp = instr->src[i].src.ssa->num_components;
4826 break;
4827 case nir_tex_src_ddy:
4828 ddy = get_src(ctx, instr->src[i].src);
4829 break;
4830 case nir_tex_src_texture_offset:
4831 case nir_tex_src_sampler_offset:
4832 case nir_tex_src_plane:
4833 default:
4834 break;
4835 }
4836 }
4837
4838 if (instr->op == nir_texop_txs && instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
4839 result = get_buffer_size(ctx, res_ptr, true);
4840 goto write_result;
4841 }
4842
4843 if (instr->op == nir_texop_texture_samples) {
4844 LLVMValueRef res, samples, is_msaa;
4845 res = LLVMBuildBitCast(ctx->ac.builder, res_ptr, ctx->ac.v8i32, "");
4846 samples = LLVMBuildExtractElement(ctx->ac.builder, res,
4847 LLVMConstInt(ctx->ac.i32, 3, false), "");
4848 is_msaa = LLVMBuildLShr(ctx->ac.builder, samples,
4849 LLVMConstInt(ctx->ac.i32, 28, false), "");
4850 is_msaa = LLVMBuildAnd(ctx->ac.builder, is_msaa,
4851 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
4852 is_msaa = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, is_msaa,
4853 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
4854
4855 samples = LLVMBuildLShr(ctx->ac.builder, samples,
4856 LLVMConstInt(ctx->ac.i32, 16, false), "");
4857 samples = LLVMBuildAnd(ctx->ac.builder, samples,
4858 LLVMConstInt(ctx->ac.i32, 0xf, false), "");
4859 samples = LLVMBuildShl(ctx->ac.builder, ctx->ac.i32_1,
4860 samples, "");
4861 samples = LLVMBuildSelect(ctx->ac.builder, is_msaa, samples,
4862 ctx->ac.i32_1, "");
4863 result = samples;
4864 goto write_result;
4865 }
4866
4867 if (coord)
4868 for (chan = 0; chan < instr->coord_components; chan++)
4869 coords[chan] = ac_llvm_extract_elem(&ctx->ac, coord, chan);
4870
4871 if (offsets && instr->op != nir_texop_txf) {
4872 LLVMValueRef offset[3], pack;
4873 for (chan = 0; chan < 3; ++chan)
4874 offset[chan] = ctx->ac.i32_0;
4875
4876 args.offset = true;
4877 for (chan = 0; chan < ac_get_llvm_num_components(offsets); chan++) {
4878 offset[chan] = ac_llvm_extract_elem(&ctx->ac, offsets, chan);
4879 offset[chan] = LLVMBuildAnd(ctx->ac.builder, offset[chan],
4880 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
4881 if (chan)
4882 offset[chan] = LLVMBuildShl(ctx->ac.builder, offset[chan],
4883 LLVMConstInt(ctx->ac.i32, chan * 8, false), "");
4884 }
4885 pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
4886 pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
4887 address[count++] = pack;
4888
4889 }
4890 /* pack LOD bias value */
4891 if (instr->op == nir_texop_txb && bias) {
4892 address[count++] = bias;
4893 }
4894
4895 /* Pack depth comparison value */
4896 if (instr->is_shadow && comparator) {
4897 LLVMValueRef z = ac_to_float(&ctx->ac,
4898 ac_llvm_extract_elem(&ctx->ac, comparator, 0));
4899
4900 /* TC-compatible HTILE on radeonsi promotes Z16 and Z24 to Z32_FLOAT,
4901 * so the depth comparison value isn't clamped for Z16 and
4902 * Z24 anymore. Do it manually here.
4903 *
4904 * It's unnecessary if the original texture format was
4905 * Z32_FLOAT, but we don't know that here.
4906 */
4907 if (ctx->ac.chip_class == VI && ctx->abi->clamp_shadow_reference)
4908 z = ac_build_clamp(&ctx->ac, z);
4909
4910 address[count++] = z;
4911 }
4912
4913 /* pack derivatives */
4914 if (ddx || ddy) {
4915 int num_src_deriv_channels, num_dest_deriv_channels;
4916 switch (instr->sampler_dim) {
4917 case GLSL_SAMPLER_DIM_3D:
4918 case GLSL_SAMPLER_DIM_CUBE:
4919 num_deriv_comp = 3;
4920 num_src_deriv_channels = 3;
4921 num_dest_deriv_channels = 3;
4922 break;
4923 case GLSL_SAMPLER_DIM_2D:
4924 default:
4925 num_src_deriv_channels = 2;
4926 num_dest_deriv_channels = 2;
4927 num_deriv_comp = 2;
4928 break;
4929 case GLSL_SAMPLER_DIM_1D:
4930 num_src_deriv_channels = 1;
4931 if (ctx->ac.chip_class >= GFX9) {
4932 num_dest_deriv_channels = 2;
4933 num_deriv_comp = 2;
4934 } else {
4935 num_dest_deriv_channels = 1;
4936 num_deriv_comp = 1;
4937 }
4938 break;
4939 }
4940
4941 for (unsigned i = 0; i < num_src_deriv_channels; i++) {
4942 derivs[i] = ac_to_float(&ctx->ac, ac_llvm_extract_elem(&ctx->ac, ddx, i));
4943 derivs[num_dest_deriv_channels + i] = ac_to_float(&ctx->ac, ac_llvm_extract_elem(&ctx->ac, ddy, i));
4944 }
4945 for (unsigned i = num_src_deriv_channels; i < num_dest_deriv_channels; i++) {
4946 derivs[i] = ctx->ac.f32_0;
4947 derivs[num_dest_deriv_channels + i] = ctx->ac.f32_0;
4948 }
4949 }
4950
4951 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && coord) {
4952 for (chan = 0; chan < instr->coord_components; chan++)
4953 coords[chan] = ac_to_float(&ctx->ac, coords[chan]);
4954 if (instr->coord_components == 3)
4955 coords[3] = LLVMGetUndef(ctx->ac.f32);
4956 ac_prepare_cube_coords(&ctx->ac,
4957 instr->op == nir_texop_txd, instr->is_array,
4958 instr->op == nir_texop_lod, coords, derivs);
4959 if (num_deriv_comp)
4960 num_deriv_comp--;
4961 }
4962
4963 if (ddx || ddy) {
4964 for (unsigned i = 0; i < num_deriv_comp * 2; i++)
4965 address[count++] = derivs[i];
4966 }
4967
4968 /* Pack texture coordinates */
4969 if (coord) {
4970 address[count++] = coords[0];
4971 if (instr->coord_components > 1) {
4972 if (instr->sampler_dim == GLSL_SAMPLER_DIM_1D && instr->is_array && instr->op != nir_texop_txf) {
4973 coords[1] = apply_round_slice(&ctx->ac, coords[1]);
4974 }
4975 address[count++] = coords[1];
4976 }
4977 if (instr->coord_components > 2) {
4978 /* This seems like a bit of a hack - but it passes Vulkan CTS with it */
4979 if (instr->sampler_dim != GLSL_SAMPLER_DIM_3D &&
4980 instr->sampler_dim != GLSL_SAMPLER_DIM_CUBE &&
4981 instr->op != nir_texop_txf) {
4982 coords[2] = apply_round_slice(&ctx->ac, coords[2]);
4983 }
4984 address[count++] = coords[2];
4985 }
4986
4987 if (ctx->ac.chip_class >= GFX9) {
4988 LLVMValueRef filler;
4989 if (instr->op == nir_texop_txf)
4990 filler = ctx->ac.i32_0;
4991 else
4992 filler = LLVMConstReal(ctx->ac.f32, 0.5);
4993
4994 if (instr->sampler_dim == GLSL_SAMPLER_DIM_1D) {
4995 /* No nir_texop_lod, because it does not take a slice
4996 * even with array textures. */
4997 if (instr->is_array && instr->op != nir_texop_lod ) {
4998 address[count] = address[count - 1];
4999 address[count - 1] = filler;
5000 count++;
5001 } else
5002 address[count++] = filler;
5003 }
5004 }
5005 }
5006
5007 /* Pack LOD */
5008 if (lod && ((instr->op == nir_texop_txl && !lod_is_zero) ||
5009 instr->op == nir_texop_txf)) {
5010 address[count++] = lod;
5011 } else if (instr->op == nir_texop_txf_ms && sample_index) {
5012 address[count++] = sample_index;
5013 } else if(instr->op == nir_texop_txs) {
5014 count = 0;
5015 if (lod)
5016 address[count++] = lod;
5017 else
5018 address[count++] = ctx->ac.i32_0;
5019 }
5020
5021 for (chan = 0; chan < count; chan++) {
5022 address[chan] = LLVMBuildBitCast(ctx->ac.builder,
5023 address[chan], ctx->ac.i32, "");
5024 }
5025
5026 if (instr->op == nir_texop_samples_identical) {
5027 LLVMValueRef txf_address[4];
5028 struct ac_image_args txf_args = { 0 };
5029 unsigned txf_count = count;
5030 memcpy(txf_address, address, sizeof(txf_address));
5031
5032 if (!instr->is_array)
5033 txf_address[2] = ctx->ac.i32_0;
5034 txf_address[3] = ctx->ac.i32_0;
5035
5036 set_tex_fetch_args(&ctx->ac, &txf_args, instr, nir_texop_txf,
5037 fmask_ptr, NULL,
5038 txf_address, txf_count, 0xf);
5039
5040 result = build_tex_intrinsic(ctx, instr, false, &txf_args);
5041
5042 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
5043 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, result, ctx->ac.i32_0);
5044 goto write_result;
5045 }
5046
5047 if (instr->sampler_dim == GLSL_SAMPLER_DIM_MS &&
5048 instr->op != nir_texop_txs) {
5049 unsigned sample_chan = instr->is_array ? 3 : 2;
5050 address[sample_chan] = adjust_sample_index_using_fmask(&ctx->ac,
5051 address[0],
5052 address[1],
5053 instr->is_array ? address[2] : NULL,
5054 address[sample_chan],
5055 fmask_ptr);
5056 }
5057
5058 if (offsets && instr->op == nir_texop_txf) {
5059 nir_const_value *const_offset =
5060 nir_src_as_const_value(instr->src[const_src].src);
5061 int num_offsets = instr->src[const_src].src.ssa->num_components;
5062 assert(const_offset);
5063 num_offsets = MIN2(num_offsets, instr->coord_components);
5064 if (num_offsets > 2)
5065 address[2] = LLVMBuildAdd(ctx->ac.builder,
5066 address[2], LLVMConstInt(ctx->ac.i32, const_offset->i32[2], false), "");
5067 if (num_offsets > 1)
5068 address[1] = LLVMBuildAdd(ctx->ac.builder,
5069 address[1], LLVMConstInt(ctx->ac.i32, const_offset->i32[1], false), "");
5070 address[0] = LLVMBuildAdd(ctx->ac.builder,
5071 address[0], LLVMConstInt(ctx->ac.i32, const_offset->i32[0], false), "");
5072
5073 }
5074
5075 /* TODO TG4 support */
5076 if (instr->op == nir_texop_tg4) {
5077 if (instr->is_shadow)
5078 dmask = 1;
5079 else
5080 dmask = 1 << instr->component;
5081 }
5082 set_tex_fetch_args(&ctx->ac, &args, instr, instr->op,
5083 res_ptr, samp_ptr, address, count, dmask);
5084
5085 result = build_tex_intrinsic(ctx, instr, lod_is_zero, &args);
5086
5087 if (instr->op == nir_texop_query_levels)
5088 result = LLVMBuildExtractElement(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 3, false), "");
5089 else if (instr->is_shadow && instr->is_new_style_shadow &&
5090 instr->op != nir_texop_txs && instr->op != nir_texop_lod &&
5091 instr->op != nir_texop_tg4)
5092 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
5093 else if (instr->op == nir_texop_txs &&
5094 instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
5095 instr->is_array) {
5096 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
5097 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
5098 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
5099 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
5100 result = LLVMBuildInsertElement(ctx->ac.builder, result, z, two, "");
5101 } else if (ctx->ac.chip_class >= GFX9 &&
5102 instr->op == nir_texop_txs &&
5103 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
5104 instr->is_array) {
5105 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
5106 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
5107 result = LLVMBuildInsertElement(ctx->ac.builder, result, layers,
5108 ctx->ac.i32_1, "");
5109 } else if (instr->dest.ssa.num_components != 4)
5110 result = trim_vector(&ctx->ac, result, instr->dest.ssa.num_components);
5111
5112 write_result:
5113 if (result) {
5114 assert(instr->dest.is_ssa);
5115 result = ac_to_integer(&ctx->ac, result);
5116 _mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
5117 }
5118 }
5119
5120
5121 static void visit_phi(struct ac_nir_context *ctx, nir_phi_instr *instr)
5122 {
5123 LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
5124 LLVMValueRef result = LLVMBuildPhi(ctx->ac.builder, type, "");
5125
5126 _mesa_hash_table_insert(ctx->defs, &instr->dest.ssa, result);
5127 _mesa_hash_table_insert(ctx->phis, instr, result);
5128 }
5129
5130 static void visit_post_phi(struct ac_nir_context *ctx,
5131 nir_phi_instr *instr,
5132 LLVMValueRef llvm_phi)
5133 {
5134 nir_foreach_phi_src(src, instr) {
5135 LLVMBasicBlockRef block = get_block(ctx, src->pred);
5136 LLVMValueRef llvm_src = get_src(ctx, src->src);
5137
5138 LLVMAddIncoming(llvm_phi, &llvm_src, &block, 1);
5139 }
5140 }
5141
5142 static void phi_post_pass(struct ac_nir_context *ctx)
5143 {
5144 struct hash_entry *entry;
5145 hash_table_foreach(ctx->phis, entry) {
5146 visit_post_phi(ctx, (nir_phi_instr*)entry->key,
5147 (LLVMValueRef)entry->data);
5148 }
5149 }
5150
5151
5152 static void visit_ssa_undef(struct ac_nir_context *ctx,
5153 const nir_ssa_undef_instr *instr)
5154 {
5155 unsigned num_components = instr->def.num_components;
5156 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
5157 LLVMValueRef undef;
5158
5159 if (num_components == 1)
5160 undef = LLVMGetUndef(type);
5161 else {
5162 undef = LLVMGetUndef(LLVMVectorType(type, num_components));
5163 }
5164 _mesa_hash_table_insert(ctx->defs, &instr->def, undef);
5165 }
5166
5167 static void visit_jump(struct ac_nir_context *ctx,
5168 const nir_jump_instr *instr)
5169 {
5170 switch (instr->type) {
5171 case nir_jump_break:
5172 LLVMBuildBr(ctx->ac.builder, ctx->break_block);
5173 LLVMClearInsertionPosition(ctx->ac.builder);
5174 break;
5175 case nir_jump_continue:
5176 LLVMBuildBr(ctx->ac.builder, ctx->continue_block);
5177 LLVMClearInsertionPosition(ctx->ac.builder);
5178 break;
5179 default:
5180 fprintf(stderr, "Unknown NIR jump instr: ");
5181 nir_print_instr(&instr->instr, stderr);
5182 fprintf(stderr, "\n");
5183 abort();
5184 }
5185 }
5186
5187 static void visit_cf_list(struct ac_nir_context *ctx,
5188 struct exec_list *list);
5189
5190 static void visit_block(struct ac_nir_context *ctx, nir_block *block)
5191 {
5192 LLVMBasicBlockRef llvm_block = LLVMGetInsertBlock(ctx->ac.builder);
5193 nir_foreach_instr(instr, block)
5194 {
5195 switch (instr->type) {
5196 case nir_instr_type_alu:
5197 visit_alu(ctx, nir_instr_as_alu(instr));
5198 break;
5199 case nir_instr_type_load_const:
5200 visit_load_const(ctx, nir_instr_as_load_const(instr));
5201 break;
5202 case nir_instr_type_intrinsic:
5203 visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
5204 break;
5205 case nir_instr_type_tex:
5206 visit_tex(ctx, nir_instr_as_tex(instr));
5207 break;
5208 case nir_instr_type_phi:
5209 visit_phi(ctx, nir_instr_as_phi(instr));
5210 break;
5211 case nir_instr_type_ssa_undef:
5212 visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
5213 break;
5214 case nir_instr_type_jump:
5215 visit_jump(ctx, nir_instr_as_jump(instr));
5216 break;
5217 default:
5218 fprintf(stderr, "Unknown NIR instr type: ");
5219 nir_print_instr(instr, stderr);
5220 fprintf(stderr, "\n");
5221 abort();
5222 }
5223 }
5224
5225 _mesa_hash_table_insert(ctx->defs, block, llvm_block);
5226 }
5227
5228 static void visit_if(struct ac_nir_context *ctx, nir_if *if_stmt)
5229 {
5230 LLVMValueRef value = get_src(ctx, if_stmt->condition);
5231
5232 LLVMValueRef fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->ac.builder));
5233 LLVMBasicBlockRef merge_block =
5234 LLVMAppendBasicBlockInContext(ctx->ac.context, fn, "");
5235 LLVMBasicBlockRef if_block =
5236 LLVMAppendBasicBlockInContext(ctx->ac.context, fn, "");
5237 LLVMBasicBlockRef else_block = merge_block;
5238 if (!exec_list_is_empty(&if_stmt->else_list))
5239 else_block = LLVMAppendBasicBlockInContext(
5240 ctx->ac.context, fn, "");
5241
5242 LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, value,
5243 ctx->ac.i32_0, "");
5244 LLVMBuildCondBr(ctx->ac.builder, cond, if_block, else_block);
5245
5246 LLVMPositionBuilderAtEnd(ctx->ac.builder, if_block);
5247 visit_cf_list(ctx, &if_stmt->then_list);
5248 if (LLVMGetInsertBlock(ctx->ac.builder))
5249 LLVMBuildBr(ctx->ac.builder, merge_block);
5250
5251 if (!exec_list_is_empty(&if_stmt->else_list)) {
5252 LLVMPositionBuilderAtEnd(ctx->ac.builder, else_block);
5253 visit_cf_list(ctx, &if_stmt->else_list);
5254 if (LLVMGetInsertBlock(ctx->ac.builder))
5255 LLVMBuildBr(ctx->ac.builder, merge_block);
5256 }
5257
5258 LLVMPositionBuilderAtEnd(ctx->ac.builder, merge_block);
5259 }
5260
5261 static void visit_loop(struct ac_nir_context *ctx, nir_loop *loop)
5262 {
5263 LLVMValueRef fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->ac.builder));
5264 LLVMBasicBlockRef continue_parent = ctx->continue_block;
5265 LLVMBasicBlockRef break_parent = ctx->break_block;
5266
5267 ctx->continue_block =
5268 LLVMAppendBasicBlockInContext(ctx->ac.context, fn, "");
5269 ctx->break_block =
5270 LLVMAppendBasicBlockInContext(ctx->ac.context, fn, "");
5271
5272 LLVMBuildBr(ctx->ac.builder, ctx->continue_block);
5273 LLVMPositionBuilderAtEnd(ctx->ac.builder, ctx->continue_block);
5274 visit_cf_list(ctx, &loop->body);
5275
5276 if (LLVMGetInsertBlock(ctx->ac.builder))
5277 LLVMBuildBr(ctx->ac.builder, ctx->continue_block);
5278 LLVMPositionBuilderAtEnd(ctx->ac.builder, ctx->break_block);
5279
5280 ctx->continue_block = continue_parent;
5281 ctx->break_block = break_parent;
5282 }
5283
5284 static void visit_cf_list(struct ac_nir_context *ctx,
5285 struct exec_list *list)
5286 {
5287 foreach_list_typed(nir_cf_node, node, node, list)
5288 {
5289 switch (node->type) {
5290 case nir_cf_node_block:
5291 visit_block(ctx, nir_cf_node_as_block(node));
5292 break;
5293
5294 case nir_cf_node_if:
5295 visit_if(ctx, nir_cf_node_as_if(node));
5296 break;
5297
5298 case nir_cf_node_loop:
5299 visit_loop(ctx, nir_cf_node_as_loop(node));
5300 break;
5301
5302 default:
5303 assert(0);
5304 }
5305 }
5306 }
5307
5308 static void
5309 handle_vs_input_decl(struct nir_to_llvm_context *ctx,
5310 struct nir_variable *variable)
5311 {
5312 LLVMValueRef t_list_ptr = ctx->vertex_buffers;
5313 LLVMValueRef t_offset;
5314 LLVMValueRef t_list;
5315 LLVMValueRef input;
5316 LLVMValueRef buffer_index;
5317 int index = variable->data.location - VERT_ATTRIB_GENERIC0;
5318 int idx = variable->data.location;
5319 unsigned attrib_count = glsl_count_attribute_slots(variable->type, true);
5320 uint8_t input_usage_mask =
5321 ctx->shader_info->info.vs.input_usage_mask[variable->data.location];
5322 unsigned num_channels = util_last_bit(input_usage_mask);
5323
5324 variable->data.driver_location = idx * 4;
5325
5326 for (unsigned i = 0; i < attrib_count; ++i, ++idx) {
5327 if (ctx->options->key.vs.instance_rate_inputs & (1u << (index + i))) {
5328 buffer_index = LLVMBuildAdd(ctx->builder, ctx->abi.instance_id,
5329 ctx->abi.start_instance, "");
5330 if (ctx->options->key.vs.as_ls) {
5331 ctx->shader_info->vs.vgpr_comp_cnt =
5332 MAX2(2, ctx->shader_info->vs.vgpr_comp_cnt);
5333 } else {
5334 ctx->shader_info->vs.vgpr_comp_cnt =
5335 MAX2(1, ctx->shader_info->vs.vgpr_comp_cnt);
5336 }
5337 } else
5338 buffer_index = LLVMBuildAdd(ctx->builder, ctx->abi.vertex_id,
5339 ctx->abi.base_vertex, "");
5340 t_offset = LLVMConstInt(ctx->ac.i32, index + i, false);
5341
5342 t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
5343
5344 input = ac_build_buffer_load_format(&ctx->ac, t_list,
5345 buffer_index,
5346 ctx->ac.i32_0,
5347 num_channels, false, true);
5348
5349 input = ac_build_expand_to_vec4(&ctx->ac, input, num_channels);
5350
5351 for (unsigned chan = 0; chan < 4; chan++) {
5352 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
5353 ctx->inputs[radeon_llvm_reg_index_soa(idx, chan)] =
5354 ac_to_integer(&ctx->ac, LLVMBuildExtractElement(ctx->builder,
5355 input, llvm_chan, ""));
5356 }
5357 }
5358 }
5359
5360 static void interp_fs_input(struct nir_to_llvm_context *ctx,
5361 unsigned attr,
5362 LLVMValueRef interp_param,
5363 LLVMValueRef prim_mask,
5364 LLVMValueRef result[4])
5365 {
5366 LLVMValueRef attr_number;
5367 unsigned chan;
5368 LLVMValueRef i, j;
5369 bool interp = interp_param != NULL;
5370
5371 attr_number = LLVMConstInt(ctx->ac.i32, attr, false);
5372
5373 /* fs.constant returns the param from the middle vertex, so it's not
5374 * really useful for flat shading. It's meant to be used for custom
5375 * interpolation (but the intrinsic can't fetch from the other two
5376 * vertices).
5377 *
5378 * Luckily, it doesn't matter, because we rely on the FLAT_SHADE state
5379 * to do the right thing. The only reason we use fs.constant is that
5380 * fs.interp cannot be used on integers, because they can be equal
5381 * to NaN.
5382 */
5383 if (interp) {
5384 interp_param = LLVMBuildBitCast(ctx->builder, interp_param,
5385 ctx->ac.v2f32, "");
5386
5387 i = LLVMBuildExtractElement(ctx->builder, interp_param,
5388 ctx->ac.i32_0, "");
5389 j = LLVMBuildExtractElement(ctx->builder, interp_param,
5390 ctx->ac.i32_1, "");
5391 }
5392
5393 for (chan = 0; chan < 4; chan++) {
5394 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
5395
5396 if (interp) {
5397 result[chan] = ac_build_fs_interp(&ctx->ac,
5398 llvm_chan,
5399 attr_number,
5400 prim_mask, i, j);
5401 } else {
5402 result[chan] = ac_build_fs_interp_mov(&ctx->ac,
5403 LLVMConstInt(ctx->ac.i32, 2, false),
5404 llvm_chan,
5405 attr_number,
5406 prim_mask);
5407 }
5408 }
5409 }
5410
5411 static void
5412 handle_fs_input_decl(struct nir_to_llvm_context *ctx,
5413 struct nir_variable *variable)
5414 {
5415 int idx = variable->data.location;
5416 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5417 LLVMValueRef interp;
5418
5419 variable->data.driver_location = idx * 4;
5420 ctx->input_mask |= ((1ull << attrib_count) - 1) << variable->data.location;
5421
5422 if (glsl_get_base_type(glsl_without_array(variable->type)) == GLSL_TYPE_FLOAT) {
5423 unsigned interp_type;
5424 if (variable->data.sample) {
5425 interp_type = INTERP_SAMPLE;
5426 ctx->shader_info->info.ps.force_persample = true;
5427 } else if (variable->data.centroid)
5428 interp_type = INTERP_CENTROID;
5429 else
5430 interp_type = INTERP_CENTER;
5431
5432 interp = lookup_interp_param(&ctx->abi, variable->data.interpolation, interp_type);
5433 } else
5434 interp = NULL;
5435
5436 for (unsigned i = 0; i < attrib_count; ++i)
5437 ctx->inputs[radeon_llvm_reg_index_soa(idx + i, 0)] = interp;
5438
5439 }
5440
5441 static void
5442 handle_vs_inputs(struct nir_to_llvm_context *ctx,
5443 struct nir_shader *nir) {
5444 nir_foreach_variable(variable, &nir->inputs)
5445 handle_vs_input_decl(ctx, variable);
5446 }
5447
5448 static void
5449 prepare_interp_optimize(struct nir_to_llvm_context *ctx,
5450 struct nir_shader *nir)
5451 {
5452 if (!ctx->options->key.fs.multisample)
5453 return;
5454
5455 bool uses_center = false;
5456 bool uses_centroid = false;
5457 nir_foreach_variable(variable, &nir->inputs) {
5458 if (glsl_get_base_type(glsl_without_array(variable->type)) != GLSL_TYPE_FLOAT ||
5459 variable->data.sample)
5460 continue;
5461
5462 if (variable->data.centroid)
5463 uses_centroid = true;
5464 else
5465 uses_center = true;
5466 }
5467
5468 if (uses_center && uses_centroid) {
5469 LLVMValueRef sel = LLVMBuildICmp(ctx->builder, LLVMIntSLT, ctx->abi.prim_mask, ctx->ac.i32_0, "");
5470 ctx->persp_centroid = LLVMBuildSelect(ctx->builder, sel, ctx->persp_center, ctx->persp_centroid, "");
5471 ctx->linear_centroid = LLVMBuildSelect(ctx->builder, sel, ctx->linear_center, ctx->linear_centroid, "");
5472 }
5473 }
5474
5475 static void
5476 handle_fs_inputs(struct nir_to_llvm_context *ctx,
5477 struct nir_shader *nir)
5478 {
5479 prepare_interp_optimize(ctx, nir);
5480
5481 nir_foreach_variable(variable, &nir->inputs)
5482 handle_fs_input_decl(ctx, variable);
5483
5484 unsigned index = 0;
5485
5486 if (ctx->shader_info->info.ps.uses_input_attachments ||
5487 ctx->shader_info->info.needs_multiview_view_index)
5488 ctx->input_mask |= 1ull << VARYING_SLOT_LAYER;
5489
5490 for (unsigned i = 0; i < RADEON_LLVM_MAX_INPUTS; ++i) {
5491 LLVMValueRef interp_param;
5492 LLVMValueRef *inputs = ctx->inputs +radeon_llvm_reg_index_soa(i, 0);
5493
5494 if (!(ctx->input_mask & (1ull << i)))
5495 continue;
5496
5497 if (i >= VARYING_SLOT_VAR0 || i == VARYING_SLOT_PNTC ||
5498 i == VARYING_SLOT_PRIMITIVE_ID || i == VARYING_SLOT_LAYER) {
5499 interp_param = *inputs;
5500 interp_fs_input(ctx, index, interp_param, ctx->abi.prim_mask,
5501 inputs);
5502
5503 if (!interp_param)
5504 ctx->shader_info->fs.flat_shaded_mask |= 1u << index;
5505 ++index;
5506 } else if (i == VARYING_SLOT_POS) {
5507 for(int i = 0; i < 3; ++i)
5508 inputs[i] = ctx->abi.frag_pos[i];
5509
5510 inputs[3] = ac_build_fdiv(&ctx->ac, ctx->ac.f32_1,
5511 ctx->abi.frag_pos[3]);
5512 }
5513 }
5514 ctx->shader_info->fs.num_interp = index;
5515 if (ctx->input_mask & (1 << VARYING_SLOT_PNTC))
5516 ctx->shader_info->fs.has_pcoord = true;
5517 if (ctx->input_mask & (1 << VARYING_SLOT_PRIMITIVE_ID))
5518 ctx->shader_info->fs.prim_id_input = true;
5519 if (ctx->input_mask & (1 << VARYING_SLOT_LAYER))
5520 ctx->shader_info->fs.layer_input = true;
5521 ctx->shader_info->fs.input_mask = ctx->input_mask >> VARYING_SLOT_VAR0;
5522
5523 if (ctx->shader_info->info.needs_multiview_view_index)
5524 ctx->view_index = ctx->inputs[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)];
5525 }
5526
5527 static LLVMValueRef
5528 ac_build_alloca(struct ac_llvm_context *ac,
5529 LLVMTypeRef type,
5530 const char *name)
5531 {
5532 LLVMBuilderRef builder = ac->builder;
5533 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
5534 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
5535 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
5536 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
5537 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
5538 LLVMValueRef res;
5539
5540 if (first_instr) {
5541 LLVMPositionBuilderBefore(first_builder, first_instr);
5542 } else {
5543 LLVMPositionBuilderAtEnd(first_builder, first_block);
5544 }
5545
5546 res = LLVMBuildAlloca(first_builder, type, name);
5547 LLVMBuildStore(builder, LLVMConstNull(type), res);
5548
5549 LLVMDisposeBuilder(first_builder);
5550
5551 return res;
5552 }
5553
5554 static LLVMValueRef si_build_alloca_undef(struct ac_llvm_context *ac,
5555 LLVMTypeRef type,
5556 const char *name)
5557 {
5558 LLVMValueRef ptr = ac_build_alloca(ac, type, name);
5559 LLVMBuildStore(ac->builder, LLVMGetUndef(type), ptr);
5560 return ptr;
5561 }
5562
5563 static void
5564 scan_shader_output_decl(struct nir_to_llvm_context *ctx,
5565 struct nir_variable *variable,
5566 struct nir_shader *shader,
5567 gl_shader_stage stage)
5568 {
5569 int idx = variable->data.location + variable->data.index;
5570 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5571 uint64_t mask_attribs;
5572
5573 variable->data.driver_location = idx * 4;
5574
5575 /* tess ctrl has it's own load/store paths for outputs */
5576 if (stage == MESA_SHADER_TESS_CTRL)
5577 return;
5578
5579 mask_attribs = ((1ull << attrib_count) - 1) << idx;
5580 if (stage == MESA_SHADER_VERTEX ||
5581 stage == MESA_SHADER_TESS_EVAL ||
5582 stage == MESA_SHADER_GEOMETRY) {
5583 if (idx == VARYING_SLOT_CLIP_DIST0) {
5584 int length = shader->info.clip_distance_array_size +
5585 shader->info.cull_distance_array_size;
5586 if (stage == MESA_SHADER_VERTEX) {
5587 ctx->shader_info->vs.outinfo.clip_dist_mask = (1 << shader->info.clip_distance_array_size) - 1;
5588 ctx->shader_info->vs.outinfo.cull_dist_mask = (1 << shader->info.cull_distance_array_size) - 1;
5589 }
5590 if (stage == MESA_SHADER_TESS_EVAL) {
5591 ctx->shader_info->tes.outinfo.clip_dist_mask = (1 << shader->info.clip_distance_array_size) - 1;
5592 ctx->shader_info->tes.outinfo.cull_dist_mask = (1 << shader->info.cull_distance_array_size) - 1;
5593 }
5594
5595 if (length > 4)
5596 attrib_count = 2;
5597 else
5598 attrib_count = 1;
5599 mask_attribs = 1ull << idx;
5600 }
5601 }
5602
5603 ctx->output_mask |= mask_attribs;
5604 }
5605
5606 static void
5607 handle_shader_output_decl(struct ac_nir_context *ctx,
5608 struct nir_shader *nir,
5609 struct nir_variable *variable)
5610 {
5611 unsigned output_loc = variable->data.driver_location / 4;
5612 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5613
5614 /* tess ctrl has it's own load/store paths for outputs */
5615 if (ctx->stage == MESA_SHADER_TESS_CTRL)
5616 return;
5617
5618 if (ctx->stage == MESA_SHADER_VERTEX ||
5619 ctx->stage == MESA_SHADER_TESS_EVAL ||
5620 ctx->stage == MESA_SHADER_GEOMETRY) {
5621 int idx = variable->data.location + variable->data.index;
5622 if (idx == VARYING_SLOT_CLIP_DIST0) {
5623 int length = nir->info.clip_distance_array_size +
5624 nir->info.cull_distance_array_size;
5625
5626 if (length > 4)
5627 attrib_count = 2;
5628 else
5629 attrib_count = 1;
5630 }
5631 }
5632
5633 for (unsigned i = 0; i < attrib_count; ++i) {
5634 for (unsigned chan = 0; chan < 4; chan++) {
5635 ctx->outputs[radeon_llvm_reg_index_soa(output_loc + i, chan)] =
5636 si_build_alloca_undef(&ctx->ac, ctx->ac.f32, "");
5637 }
5638 }
5639 }
5640
5641 static LLVMTypeRef
5642 glsl_base_to_llvm_type(struct nir_to_llvm_context *ctx,
5643 enum glsl_base_type type)
5644 {
5645 switch (type) {
5646 case GLSL_TYPE_INT:
5647 case GLSL_TYPE_UINT:
5648 case GLSL_TYPE_BOOL:
5649 case GLSL_TYPE_SUBROUTINE:
5650 return ctx->ac.i32;
5651 case GLSL_TYPE_FLOAT: /* TODO handle mediump */
5652 return ctx->ac.f32;
5653 case GLSL_TYPE_INT64:
5654 case GLSL_TYPE_UINT64:
5655 return ctx->ac.i64;
5656 case GLSL_TYPE_DOUBLE:
5657 return ctx->ac.f64;
5658 default:
5659 unreachable("unknown GLSL type");
5660 }
5661 }
5662
5663 static LLVMTypeRef
5664 glsl_to_llvm_type(struct nir_to_llvm_context *ctx,
5665 const struct glsl_type *type)
5666 {
5667 if (glsl_type_is_scalar(type)) {
5668 return glsl_base_to_llvm_type(ctx, glsl_get_base_type(type));
5669 }
5670
5671 if (glsl_type_is_vector(type)) {
5672 return LLVMVectorType(
5673 glsl_base_to_llvm_type(ctx, glsl_get_base_type(type)),
5674 glsl_get_vector_elements(type));
5675 }
5676
5677 if (glsl_type_is_matrix(type)) {
5678 return LLVMArrayType(
5679 glsl_to_llvm_type(ctx, glsl_get_column_type(type)),
5680 glsl_get_matrix_columns(type));
5681 }
5682
5683 if (glsl_type_is_array(type)) {
5684 return LLVMArrayType(
5685 glsl_to_llvm_type(ctx, glsl_get_array_element(type)),
5686 glsl_get_length(type));
5687 }
5688
5689 assert(glsl_type_is_struct(type));
5690
5691 LLVMTypeRef member_types[glsl_get_length(type)];
5692
5693 for (unsigned i = 0; i < glsl_get_length(type); i++) {
5694 member_types[i] =
5695 glsl_to_llvm_type(ctx,
5696 glsl_get_struct_field(type, i));
5697 }
5698
5699 return LLVMStructTypeInContext(ctx->context, member_types,
5700 glsl_get_length(type), false);
5701 }
5702
5703 static void
5704 setup_locals(struct ac_nir_context *ctx,
5705 struct nir_function *func)
5706 {
5707 int i, j;
5708 ctx->num_locals = 0;
5709 nir_foreach_variable(variable, &func->impl->locals) {
5710 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5711 variable->data.driver_location = ctx->num_locals * 4;
5712 variable->data.location_frac = 0;
5713 ctx->num_locals += attrib_count;
5714 }
5715 ctx->locals = malloc(4 * ctx->num_locals * sizeof(LLVMValueRef));
5716 if (!ctx->locals)
5717 return;
5718
5719 for (i = 0; i < ctx->num_locals; i++) {
5720 for (j = 0; j < 4; j++) {
5721 ctx->locals[i * 4 + j] =
5722 si_build_alloca_undef(&ctx->ac, ctx->ac.f32, "temp");
5723 }
5724 }
5725 }
5726
5727 static void
5728 setup_shared(struct ac_nir_context *ctx,
5729 struct nir_shader *nir)
5730 {
5731 nir_foreach_variable(variable, &nir->shared) {
5732 LLVMValueRef shared =
5733 LLVMAddGlobalInAddressSpace(
5734 ctx->ac.module, glsl_to_llvm_type(ctx->nctx, variable->type),
5735 variable->name ? variable->name : "",
5736 AC_LOCAL_ADDR_SPACE);
5737 _mesa_hash_table_insert(ctx->vars, variable, shared);
5738 }
5739 }
5740
5741 static LLVMValueRef
5742 emit_float_saturate(struct ac_llvm_context *ctx, LLVMValueRef v, float lo, float hi)
5743 {
5744 v = ac_to_float(ctx, v);
5745 v = emit_intrin_2f_param(ctx, "llvm.maxnum", ctx->f32, v, LLVMConstReal(ctx->f32, lo));
5746 return emit_intrin_2f_param(ctx, "llvm.minnum", ctx->f32, v, LLVMConstReal(ctx->f32, hi));
5747 }
5748
5749
5750 static LLVMValueRef emit_pack_int16(struct nir_to_llvm_context *ctx,
5751 LLVMValueRef src0, LLVMValueRef src1)
5752 {
5753 LLVMValueRef const16 = LLVMConstInt(ctx->ac.i32, 16, false);
5754 LLVMValueRef comp[2];
5755
5756 comp[0] = LLVMBuildAnd(ctx->builder, src0, LLVMConstInt(ctx->ac.i32, 65535, 0), "");
5757 comp[1] = LLVMBuildAnd(ctx->builder, src1, LLVMConstInt(ctx->ac.i32, 65535, 0), "");
5758 comp[1] = LLVMBuildShl(ctx->builder, comp[1], const16, "");
5759 return LLVMBuildOr(ctx->builder, comp[0], comp[1], "");
5760 }
5761
5762 /* Initialize arguments for the shader export intrinsic */
5763 static void
5764 si_llvm_init_export_args(struct nir_to_llvm_context *ctx,
5765 LLVMValueRef *values,
5766 unsigned target,
5767 struct ac_export_args *args)
5768 {
5769 /* Default is 0xf. Adjusted below depending on the format. */
5770 args->enabled_channels = 0xf;
5771
5772 /* Specify whether the EXEC mask represents the valid mask */
5773 args->valid_mask = 0;
5774
5775 /* Specify whether this is the last export */
5776 args->done = 0;
5777
5778 /* Specify the target we are exporting */
5779 args->target = target;
5780
5781 args->compr = false;
5782 args->out[0] = LLVMGetUndef(ctx->ac.f32);
5783 args->out[1] = LLVMGetUndef(ctx->ac.f32);
5784 args->out[2] = LLVMGetUndef(ctx->ac.f32);
5785 args->out[3] = LLVMGetUndef(ctx->ac.f32);
5786
5787 if (!values)
5788 return;
5789
5790 if (ctx->stage == MESA_SHADER_FRAGMENT && target >= V_008DFC_SQ_EXP_MRT) {
5791 LLVMValueRef val[4];
5792 unsigned index = target - V_008DFC_SQ_EXP_MRT;
5793 unsigned col_format = (ctx->options->key.fs.col_format >> (4 * index)) & 0xf;
5794 bool is_int8 = (ctx->options->key.fs.is_int8 >> index) & 1;
5795 bool is_int10 = (ctx->options->key.fs.is_int10 >> index) & 1;
5796
5797 switch(col_format) {
5798 case V_028714_SPI_SHADER_ZERO:
5799 args->enabled_channels = 0; /* writemask */
5800 args->target = V_008DFC_SQ_EXP_NULL;
5801 break;
5802
5803 case V_028714_SPI_SHADER_32_R:
5804 args->enabled_channels = 1;
5805 args->out[0] = values[0];
5806 break;
5807
5808 case V_028714_SPI_SHADER_32_GR:
5809 args->enabled_channels = 0x3;
5810 args->out[0] = values[0];
5811 args->out[1] = values[1];
5812 break;
5813
5814 case V_028714_SPI_SHADER_32_AR:
5815 args->enabled_channels = 0x9;
5816 args->out[0] = values[0];
5817 args->out[3] = values[3];
5818 break;
5819
5820 case V_028714_SPI_SHADER_FP16_ABGR:
5821 args->compr = 1;
5822
5823 for (unsigned chan = 0; chan < 2; chan++) {
5824 LLVMValueRef pack_args[2] = {
5825 values[2 * chan],
5826 values[2 * chan + 1]
5827 };
5828 LLVMValueRef packed;
5829
5830 packed = ac_build_cvt_pkrtz_f16(&ctx->ac, pack_args);
5831 args->out[chan] = packed;
5832 }
5833 break;
5834
5835 case V_028714_SPI_SHADER_UNORM16_ABGR:
5836 for (unsigned chan = 0; chan < 4; chan++) {
5837 val[chan] = ac_build_clamp(&ctx->ac, values[chan]);
5838 val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
5839 LLVMConstReal(ctx->ac.f32, 65535), "");
5840 val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
5841 LLVMConstReal(ctx->ac.f32, 0.5), "");
5842 val[chan] = LLVMBuildFPToUI(ctx->builder, val[chan],
5843 ctx->ac.i32, "");
5844 }
5845
5846 args->compr = 1;
5847 args->out[0] = emit_pack_int16(ctx, val[0], val[1]);
5848 args->out[1] = emit_pack_int16(ctx, val[2], val[3]);
5849 break;
5850
5851 case V_028714_SPI_SHADER_SNORM16_ABGR:
5852 for (unsigned chan = 0; chan < 4; chan++) {
5853 val[chan] = emit_float_saturate(&ctx->ac, values[chan], -1, 1);
5854 val[chan] = LLVMBuildFMul(ctx->builder, val[chan],
5855 LLVMConstReal(ctx->ac.f32, 32767), "");
5856
5857 /* If positive, add 0.5, else add -0.5. */
5858 val[chan] = LLVMBuildFAdd(ctx->builder, val[chan],
5859 LLVMBuildSelect(ctx->builder,
5860 LLVMBuildFCmp(ctx->builder, LLVMRealOGE,
5861 val[chan], ctx->ac.f32_0, ""),
5862 LLVMConstReal(ctx->ac.f32, 0.5),
5863 LLVMConstReal(ctx->ac.f32, -0.5), ""), "");
5864 val[chan] = LLVMBuildFPToSI(ctx->builder, val[chan], ctx->ac.i32, "");
5865 }
5866
5867 args->compr = 1;
5868 args->out[0] = emit_pack_int16(ctx, val[0], val[1]);
5869 args->out[1] = emit_pack_int16(ctx, val[2], val[3]);
5870 break;
5871
5872 case V_028714_SPI_SHADER_UINT16_ABGR: {
5873 LLVMValueRef max_rgb = LLVMConstInt(ctx->ac.i32,
5874 is_int8 ? 255 : is_int10 ? 1023 : 65535, 0);
5875 LLVMValueRef max_alpha = !is_int10 ? max_rgb : LLVMConstInt(ctx->ac.i32, 3, 0);
5876
5877 for (unsigned chan = 0; chan < 4; chan++) {
5878 val[chan] = ac_to_integer(&ctx->ac, values[chan]);
5879 val[chan] = emit_minmax_int(&ctx->ac, LLVMIntULT, val[chan], chan == 3 ? max_alpha : max_rgb);
5880 }
5881
5882 args->compr = 1;
5883 args->out[0] = emit_pack_int16(ctx, val[0], val[1]);
5884 args->out[1] = emit_pack_int16(ctx, val[2], val[3]);
5885 break;
5886 }
5887
5888 case V_028714_SPI_SHADER_SINT16_ABGR: {
5889 LLVMValueRef max_rgb = LLVMConstInt(ctx->ac.i32,
5890 is_int8 ? 127 : is_int10 ? 511 : 32767, 0);
5891 LLVMValueRef min_rgb = LLVMConstInt(ctx->ac.i32,
5892 is_int8 ? -128 : is_int10 ? -512 : -32768, 0);
5893 LLVMValueRef max_alpha = !is_int10 ? max_rgb : ctx->ac.i32_1;
5894 LLVMValueRef min_alpha = !is_int10 ? min_rgb : LLVMConstInt(ctx->ac.i32, -2, 0);
5895
5896 /* Clamp. */
5897 for (unsigned chan = 0; chan < 4; chan++) {
5898 val[chan] = ac_to_integer(&ctx->ac, values[chan]);
5899 val[chan] = emit_minmax_int(&ctx->ac, LLVMIntSLT, val[chan], chan == 3 ? max_alpha : max_rgb);
5900 val[chan] = emit_minmax_int(&ctx->ac, LLVMIntSGT, val[chan], chan == 3 ? min_alpha : min_rgb);
5901 }
5902
5903 args->compr = 1;
5904 args->out[0] = emit_pack_int16(ctx, val[0], val[1]);
5905 args->out[1] = emit_pack_int16(ctx, val[2], val[3]);
5906 break;
5907 }
5908
5909 default:
5910 case V_028714_SPI_SHADER_32_ABGR:
5911 memcpy(&args->out[0], values, sizeof(values[0]) * 4);
5912 break;
5913 }
5914 } else
5915 memcpy(&args->out[0], values, sizeof(values[0]) * 4);
5916
5917 for (unsigned i = 0; i < 4; ++i)
5918 args->out[i] = ac_to_float(&ctx->ac, args->out[i]);
5919 }
5920
5921 static void
5922 handle_vs_outputs_post(struct nir_to_llvm_context *ctx,
5923 bool export_prim_id,
5924 struct ac_vs_output_info *outinfo)
5925 {
5926 uint32_t param_count = 0;
5927 unsigned target;
5928 unsigned pos_idx, num_pos_exports = 0;
5929 struct ac_export_args args, pos_args[4] = {};
5930 LLVMValueRef psize_value = NULL, layer_value = NULL, viewport_index_value = NULL;
5931 int i;
5932
5933 if (ctx->options->key.has_multiview_view_index) {
5934 LLVMValueRef* tmp_out = &ctx->nir->outputs[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)];
5935 if(!*tmp_out) {
5936 for(unsigned i = 0; i < 4; ++i)
5937 ctx->nir->outputs[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER, i)] =
5938 si_build_alloca_undef(&ctx->ac, ctx->ac.f32, "");
5939 }
5940
5941 LLVMBuildStore(ctx->builder, ac_to_float(&ctx->ac, ctx->view_index), *tmp_out);
5942 ctx->output_mask |= 1ull << VARYING_SLOT_LAYER;
5943 }
5944
5945 memset(outinfo->vs_output_param_offset, AC_EXP_PARAM_UNDEFINED,
5946 sizeof(outinfo->vs_output_param_offset));
5947
5948 if (ctx->output_mask & (1ull << VARYING_SLOT_CLIP_DIST0)) {
5949 LLVMValueRef slots[8];
5950 unsigned j;
5951
5952 if (outinfo->cull_dist_mask)
5953 outinfo->cull_dist_mask <<= ctx->num_output_clips;
5954
5955 i = VARYING_SLOT_CLIP_DIST0;
5956 for (j = 0; j < ctx->num_output_clips + ctx->num_output_culls; j++)
5957 slots[j] = ac_to_float(&ctx->ac, LLVMBuildLoad(ctx->builder,
5958 ctx->nir->outputs[radeon_llvm_reg_index_soa(i, j)], ""));
5959
5960 for (i = ctx->num_output_clips + ctx->num_output_culls; i < 8; i++)
5961 slots[i] = LLVMGetUndef(ctx->ac.f32);
5962
5963 if (ctx->num_output_clips + ctx->num_output_culls > 4) {
5964 target = V_008DFC_SQ_EXP_POS + 3;
5965 si_llvm_init_export_args(ctx, &slots[4], target, &args);
5966 memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
5967 &args, sizeof(args));
5968 }
5969
5970 target = V_008DFC_SQ_EXP_POS + 2;
5971 si_llvm_init_export_args(ctx, &slots[0], target, &args);
5972 memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
5973 &args, sizeof(args));
5974
5975 }
5976
5977 LLVMValueRef pos_values[4] = {ctx->ac.f32_0, ctx->ac.f32_0, ctx->ac.f32_0, ctx->ac.f32_1};
5978 if (ctx->output_mask & (1ull << VARYING_SLOT_POS)) {
5979 for (unsigned j = 0; j < 4; j++)
5980 pos_values[j] = LLVMBuildLoad(ctx->builder,
5981 ctx->nir->outputs[radeon_llvm_reg_index_soa(VARYING_SLOT_POS, j)], "");
5982 }
5983 si_llvm_init_export_args(ctx, pos_values, V_008DFC_SQ_EXP_POS, &pos_args[0]);
5984
5985 if (ctx->output_mask & (1ull << VARYING_SLOT_PSIZ)) {
5986 outinfo->writes_pointsize = true;
5987 psize_value = LLVMBuildLoad(ctx->builder,
5988 ctx->nir->outputs[radeon_llvm_reg_index_soa(VARYING_SLOT_PSIZ, 0)], "");
5989 }
5990
5991 if (ctx->output_mask & (1ull << VARYING_SLOT_LAYER)) {
5992 outinfo->writes_layer = true;
5993 layer_value = LLVMBuildLoad(ctx->builder,
5994 ctx->nir->outputs[radeon_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)], "");
5995 }
5996
5997 if (ctx->output_mask & (1ull << VARYING_SLOT_VIEWPORT)) {
5998 outinfo->writes_viewport_index = true;
5999 viewport_index_value = LLVMBuildLoad(ctx->builder,
6000 ctx->nir->outputs[radeon_llvm_reg_index_soa(VARYING_SLOT_VIEWPORT, 0)], "");
6001 }
6002
6003 if (outinfo->writes_pointsize ||
6004 outinfo->writes_layer ||
6005 outinfo->writes_viewport_index) {
6006 pos_args[1].enabled_channels = ((outinfo->writes_pointsize == true ? 1 : 0) |
6007 (outinfo->writes_layer == true ? 4 : 0));
6008 pos_args[1].valid_mask = 0;
6009 pos_args[1].done = 0;
6010 pos_args[1].target = V_008DFC_SQ_EXP_POS + 1;
6011 pos_args[1].compr = 0;
6012 pos_args[1].out[0] = ctx->ac.f32_0; /* X */
6013 pos_args[1].out[1] = ctx->ac.f32_0; /* Y */
6014 pos_args[1].out[2] = ctx->ac.f32_0; /* Z */
6015 pos_args[1].out[3] = ctx->ac.f32_0; /* W */
6016
6017 if (outinfo->writes_pointsize == true)
6018 pos_args[1].out[0] = psize_value;
6019 if (outinfo->writes_layer == true)
6020 pos_args[1].out[2] = layer_value;
6021 if (outinfo->writes_viewport_index == true) {
6022 if (ctx->options->chip_class >= GFX9) {
6023 /* GFX9 has the layer in out.z[10:0] and the viewport
6024 * index in out.z[19:16].
6025 */
6026 LLVMValueRef v = viewport_index_value;
6027 v = ac_to_integer(&ctx->ac, v);
6028 v = LLVMBuildShl(ctx->builder, v,
6029 LLVMConstInt(ctx->ac.i32, 16, false),
6030 "");
6031 v = LLVMBuildOr(ctx->builder, v,
6032 ac_to_integer(&ctx->ac, pos_args[1].out[2]), "");
6033
6034 pos_args[1].out[2] = ac_to_float(&ctx->ac, v);
6035 pos_args[1].enabled_channels |= 1 << 2;
6036 } else {
6037 pos_args[1].out[3] = viewport_index_value;
6038 pos_args[1].enabled_channels |= 1 << 3;
6039 }
6040 }
6041 }
6042 for (i = 0; i < 4; i++) {
6043 if (pos_args[i].out[0])
6044 num_pos_exports++;
6045 }
6046
6047 pos_idx = 0;
6048 for (i = 0; i < 4; i++) {
6049 if (!pos_args[i].out[0])
6050 continue;
6051
6052 /* Specify the target we are exporting */
6053 pos_args[i].target = V_008DFC_SQ_EXP_POS + pos_idx++;
6054 if (pos_idx == num_pos_exports)
6055 pos_args[i].done = 1;
6056 ac_build_export(&ctx->ac, &pos_args[i]);
6057 }
6058
6059 for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
6060 LLVMValueRef values[4];
6061 if (!(ctx->output_mask & (1ull << i)))
6062 continue;
6063
6064 for (unsigned j = 0; j < 4; j++)
6065 values[j] = ac_to_float(&ctx->ac, LLVMBuildLoad(ctx->builder,
6066 ctx->nir->outputs[radeon_llvm_reg_index_soa(i, j)], ""));
6067
6068 if (i == VARYING_SLOT_LAYER) {
6069 target = V_008DFC_SQ_EXP_PARAM + param_count;
6070 outinfo->vs_output_param_offset[VARYING_SLOT_LAYER] = param_count;
6071 param_count++;
6072 } else if (i == VARYING_SLOT_PRIMITIVE_ID) {
6073 target = V_008DFC_SQ_EXP_PARAM + param_count;
6074 outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID] = param_count;
6075 param_count++;
6076 } else if (i >= VARYING_SLOT_VAR0) {
6077 outinfo->export_mask |= 1u << (i - VARYING_SLOT_VAR0);
6078 target = V_008DFC_SQ_EXP_PARAM + param_count;
6079 outinfo->vs_output_param_offset[i] = param_count;
6080 param_count++;
6081 } else
6082 continue;
6083
6084 si_llvm_init_export_args(ctx, values, target, &args);
6085
6086 if (target >= V_008DFC_SQ_EXP_POS &&
6087 target <= (V_008DFC_SQ_EXP_POS + 3)) {
6088 memcpy(&pos_args[target - V_008DFC_SQ_EXP_POS],
6089 &args, sizeof(args));
6090 } else {
6091 ac_build_export(&ctx->ac, &args);
6092 }
6093 }
6094
6095 if (export_prim_id) {
6096 LLVMValueRef values[4];
6097 target = V_008DFC_SQ_EXP_PARAM + param_count;
6098 outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID] = param_count;
6099 param_count++;
6100
6101 values[0] = ctx->vs_prim_id;
6102 ctx->shader_info->vs.vgpr_comp_cnt = MAX2(2,
6103 ctx->shader_info->vs.vgpr_comp_cnt);
6104 for (unsigned j = 1; j < 4; j++)
6105 values[j] = ctx->ac.f32_0;
6106 si_llvm_init_export_args(ctx, values, target, &args);
6107 ac_build_export(&ctx->ac, &args);
6108 outinfo->export_prim_id = true;
6109 }
6110
6111 outinfo->pos_exports = num_pos_exports;
6112 outinfo->param_exports = param_count;
6113 }
6114
6115 static void
6116 handle_es_outputs_post(struct nir_to_llvm_context *ctx,
6117 struct ac_es_output_info *outinfo)
6118 {
6119 int j;
6120 uint64_t max_output_written = 0;
6121 LLVMValueRef lds_base = NULL;
6122
6123 for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
6124 int param_index;
6125 int length = 4;
6126
6127 if (!(ctx->output_mask & (1ull << i)))
6128 continue;
6129
6130 if (i == VARYING_SLOT_CLIP_DIST0)
6131 length = ctx->num_output_clips + ctx->num_output_culls;
6132
6133 param_index = shader_io_get_unique_index(i);
6134
6135 max_output_written = MAX2(param_index + (length > 4), max_output_written);
6136 }
6137
6138 outinfo->esgs_itemsize = (max_output_written + 1) * 16;
6139
6140 if (ctx->ac.chip_class >= GFX9) {
6141 unsigned itemsize_dw = outinfo->esgs_itemsize / 4;
6142 LLVMValueRef vertex_idx = ac_get_thread_id(&ctx->ac);
6143 LLVMValueRef wave_idx = ac_build_bfe(&ctx->ac, ctx->merged_wave_info,
6144 LLVMConstInt(ctx->ac.i32, 24, false),
6145 LLVMConstInt(ctx->ac.i32, 4, false), false);
6146 vertex_idx = LLVMBuildOr(ctx->ac.builder, vertex_idx,
6147 LLVMBuildMul(ctx->ac.builder, wave_idx,
6148 LLVMConstInt(ctx->ac.i32, 64, false), ""), "");
6149 lds_base = LLVMBuildMul(ctx->ac.builder, vertex_idx,
6150 LLVMConstInt(ctx->ac.i32, itemsize_dw, 0), "");
6151 }
6152
6153 for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
6154 LLVMValueRef dw_addr;
6155 LLVMValueRef *out_ptr = &ctx->nir->outputs[i * 4];
6156 int param_index;
6157 int length = 4;
6158
6159 if (!(ctx->output_mask & (1ull << i)))
6160 continue;
6161
6162 if (i == VARYING_SLOT_CLIP_DIST0)
6163 length = ctx->num_output_clips + ctx->num_output_culls;
6164
6165 param_index = shader_io_get_unique_index(i);
6166
6167 if (lds_base) {
6168 dw_addr = LLVMBuildAdd(ctx->builder, lds_base,
6169 LLVMConstInt(ctx->ac.i32, param_index * 4, false),
6170 "");
6171 }
6172 for (j = 0; j < length; j++) {
6173 LLVMValueRef out_val = LLVMBuildLoad(ctx->builder, out_ptr[j], "");
6174 out_val = LLVMBuildBitCast(ctx->builder, out_val, ctx->ac.i32, "");
6175
6176 if (ctx->ac.chip_class >= GFX9) {
6177 ac_lds_store(&ctx->ac, dw_addr,
6178 LLVMBuildLoad(ctx->builder, out_ptr[j], ""));
6179 dw_addr = LLVMBuildAdd(ctx->builder, dw_addr, ctx->ac.i32_1, "");
6180 } else {
6181 ac_build_buffer_store_dword(&ctx->ac,
6182 ctx->esgs_ring,
6183 out_val, 1,
6184 NULL, ctx->es2gs_offset,
6185 (4 * param_index + j) * 4,
6186 1, 1, true, true);
6187 }
6188 }
6189 }
6190 }
6191
6192 static void
6193 handle_ls_outputs_post(struct nir_to_llvm_context *ctx)
6194 {
6195 LLVMValueRef vertex_id = ctx->rel_auto_id;
6196 LLVMValueRef vertex_dw_stride = unpack_param(&ctx->ac, ctx->ls_out_layout, 13, 8);
6197 LLVMValueRef base_dw_addr = LLVMBuildMul(ctx->builder, vertex_id,
6198 vertex_dw_stride, "");
6199
6200 for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
6201 LLVMValueRef *out_ptr = &ctx->nir->outputs[i * 4];
6202 int length = 4;
6203
6204 if (!(ctx->output_mask & (1ull << i)))
6205 continue;
6206
6207 if (i == VARYING_SLOT_CLIP_DIST0)
6208 length = ctx->num_output_clips + ctx->num_output_culls;
6209 int param = shader_io_get_unique_index(i);
6210 mark_tess_output(ctx, false, param);
6211 if (length > 4)
6212 mark_tess_output(ctx, false, param + 1);
6213 LLVMValueRef dw_addr = LLVMBuildAdd(ctx->builder, base_dw_addr,
6214 LLVMConstInt(ctx->ac.i32, param * 4, false),
6215 "");
6216 for (unsigned j = 0; j < length; j++) {
6217 ac_lds_store(&ctx->ac, dw_addr,
6218 LLVMBuildLoad(ctx->builder, out_ptr[j], ""));
6219 dw_addr = LLVMBuildAdd(ctx->builder, dw_addr, ctx->ac.i32_1, "");
6220 }
6221 }
6222 }
6223
6224 struct ac_build_if_state
6225 {
6226 struct nir_to_llvm_context *ctx;
6227 LLVMValueRef condition;
6228 LLVMBasicBlockRef entry_block;
6229 LLVMBasicBlockRef true_block;
6230 LLVMBasicBlockRef false_block;
6231 LLVMBasicBlockRef merge_block;
6232 };
6233
6234 static LLVMBasicBlockRef
6235 ac_build_insert_new_block(struct nir_to_llvm_context *ctx, const char *name)
6236 {
6237 LLVMBasicBlockRef current_block;
6238 LLVMBasicBlockRef next_block;
6239 LLVMBasicBlockRef new_block;
6240
6241 /* get current basic block */
6242 current_block = LLVMGetInsertBlock(ctx->builder);
6243
6244 /* chqeck if there's another block after this one */
6245 next_block = LLVMGetNextBasicBlock(current_block);
6246 if (next_block) {
6247 /* insert the new block before the next block */
6248 new_block = LLVMInsertBasicBlockInContext(ctx->context, next_block, name);
6249 }
6250 else {
6251 /* append new block after current block */
6252 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
6253 new_block = LLVMAppendBasicBlockInContext(ctx->context, function, name);
6254 }
6255 return new_block;
6256 }
6257
6258 static void
6259 ac_nir_build_if(struct ac_build_if_state *ifthen,
6260 struct nir_to_llvm_context *ctx,
6261 LLVMValueRef condition)
6262 {
6263 LLVMBasicBlockRef block = LLVMGetInsertBlock(ctx->builder);
6264
6265 memset(ifthen, 0, sizeof *ifthen);
6266 ifthen->ctx = ctx;
6267 ifthen->condition = condition;
6268 ifthen->entry_block = block;
6269
6270 /* create endif/merge basic block for the phi functions */
6271 ifthen->merge_block = ac_build_insert_new_block(ctx, "endif-block");
6272
6273 /* create/insert true_block before merge_block */
6274 ifthen->true_block =
6275 LLVMInsertBasicBlockInContext(ctx->context,
6276 ifthen->merge_block,
6277 "if-true-block");
6278
6279 /* successive code goes into the true block */
6280 LLVMPositionBuilderAtEnd(ctx->builder, ifthen->true_block);
6281 }
6282
6283 /**
6284 * End a conditional.
6285 */
6286 static void
6287 ac_nir_build_endif(struct ac_build_if_state *ifthen)
6288 {
6289 LLVMBuilderRef builder = ifthen->ctx->builder;
6290
6291 /* Insert branch to the merge block from current block */
6292 LLVMBuildBr(builder, ifthen->merge_block);
6293
6294 /*
6295 * Now patch in the various branch instructions.
6296 */
6297
6298 /* Insert the conditional branch instruction at the end of entry_block */
6299 LLVMPositionBuilderAtEnd(builder, ifthen->entry_block);
6300 if (ifthen->false_block) {
6301 /* we have an else clause */
6302 LLVMBuildCondBr(builder, ifthen->condition,
6303 ifthen->true_block, ifthen->false_block);
6304 }
6305 else {
6306 /* no else clause */
6307 LLVMBuildCondBr(builder, ifthen->condition,
6308 ifthen->true_block, ifthen->merge_block);
6309 }
6310
6311 /* Resume building code at end of the ifthen->merge_block */
6312 LLVMPositionBuilderAtEnd(builder, ifthen->merge_block);
6313 }
6314
6315 static void
6316 write_tess_factors(struct nir_to_llvm_context *ctx)
6317 {
6318 unsigned stride, outer_comps, inner_comps;
6319 struct ac_build_if_state if_ctx, inner_if_ctx;
6320 LLVMValueRef invocation_id = unpack_param(&ctx->ac, ctx->abi.tcs_rel_ids, 8, 5);
6321 LLVMValueRef rel_patch_id = unpack_param(&ctx->ac, ctx->abi.tcs_rel_ids, 0, 8);
6322 unsigned tess_inner_index, tess_outer_index;
6323 LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
6324 LLVMValueRef out[6], vec0, vec1, tf_base, inner[4], outer[4];
6325 int i;
6326 emit_barrier(&ctx->ac, ctx->stage);
6327
6328 switch (ctx->options->key.tcs.primitive_mode) {
6329 case GL_ISOLINES:
6330 stride = 2;
6331 outer_comps = 2;
6332 inner_comps = 0;
6333 break;
6334 case GL_TRIANGLES:
6335 stride = 4;
6336 outer_comps = 3;
6337 inner_comps = 1;
6338 break;
6339 case GL_QUADS:
6340 stride = 6;
6341 outer_comps = 4;
6342 inner_comps = 2;
6343 break;
6344 default:
6345 return;
6346 }
6347
6348 ac_nir_build_if(&if_ctx, ctx,
6349 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
6350 invocation_id, ctx->ac.i32_0, ""));
6351
6352 tess_inner_index = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER);
6353 tess_outer_index = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER);
6354
6355 mark_tess_output(ctx, true, tess_inner_index);
6356 mark_tess_output(ctx, true, tess_outer_index);
6357 lds_base = get_tcs_out_current_patch_data_offset(ctx);
6358 lds_inner = LLVMBuildAdd(ctx->builder, lds_base,
6359 LLVMConstInt(ctx->ac.i32, tess_inner_index * 4, false), "");
6360 lds_outer = LLVMBuildAdd(ctx->builder, lds_base,
6361 LLVMConstInt(ctx->ac.i32, tess_outer_index * 4, false), "");
6362
6363 for (i = 0; i < 4; i++) {
6364 inner[i] = LLVMGetUndef(ctx->ac.i32);
6365 outer[i] = LLVMGetUndef(ctx->ac.i32);
6366 }
6367
6368 // LINES reverseal
6369 if (ctx->options->key.tcs.primitive_mode == GL_ISOLINES) {
6370 outer[0] = out[1] = ac_lds_load(&ctx->ac, lds_outer);
6371 lds_outer = LLVMBuildAdd(ctx->builder, lds_outer,
6372 ctx->ac.i32_1, "");
6373 outer[1] = out[0] = ac_lds_load(&ctx->ac, lds_outer);
6374 } else {
6375 for (i = 0; i < outer_comps; i++) {
6376 outer[i] = out[i] =
6377 ac_lds_load(&ctx->ac, lds_outer);
6378 lds_outer = LLVMBuildAdd(ctx->builder, lds_outer,
6379 ctx->ac.i32_1, "");
6380 }
6381 for (i = 0; i < inner_comps; i++) {
6382 inner[i] = out[outer_comps+i] =
6383 ac_lds_load(&ctx->ac, lds_inner);
6384 lds_inner = LLVMBuildAdd(ctx->builder, lds_inner,
6385 ctx->ac.i32_1, "");
6386 }
6387 }
6388
6389 /* Convert the outputs to vectors for stores. */
6390 vec0 = ac_build_gather_values(&ctx->ac, out, MIN2(stride, 4));
6391 vec1 = NULL;
6392
6393 if (stride > 4)
6394 vec1 = ac_build_gather_values(&ctx->ac, out + 4, stride - 4);
6395
6396
6397 buffer = ctx->hs_ring_tess_factor;
6398 tf_base = ctx->tess_factor_offset;
6399 byteoffset = LLVMBuildMul(ctx->builder, rel_patch_id,
6400 LLVMConstInt(ctx->ac.i32, 4 * stride, false), "");
6401 unsigned tf_offset = 0;
6402
6403 if (ctx->options->chip_class <= VI) {
6404 ac_nir_build_if(&inner_if_ctx, ctx,
6405 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
6406 rel_patch_id, ctx->ac.i32_0, ""));
6407
6408 /* Store the dynamic HS control word. */
6409 ac_build_buffer_store_dword(&ctx->ac, buffer,
6410 LLVMConstInt(ctx->ac.i32, 0x80000000, false),
6411 1, ctx->ac.i32_0, tf_base,
6412 0, 1, 0, true, false);
6413 tf_offset += 4;
6414
6415 ac_nir_build_endif(&inner_if_ctx);
6416 }
6417
6418 /* Store the tessellation factors. */
6419 ac_build_buffer_store_dword(&ctx->ac, buffer, vec0,
6420 MIN2(stride, 4), byteoffset, tf_base,
6421 tf_offset, 1, 0, true, false);
6422 if (vec1)
6423 ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
6424 stride - 4, byteoffset, tf_base,
6425 16 + tf_offset, 1, 0, true, false);
6426
6427 //store to offchip for TES to read - only if TES reads them
6428 if (ctx->options->key.tcs.tes_reads_tess_factors) {
6429 LLVMValueRef inner_vec, outer_vec, tf_outer_offset;
6430 LLVMValueRef tf_inner_offset;
6431 unsigned param_outer, param_inner;
6432
6433 param_outer = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_OUTER);
6434 tf_outer_offset = get_tcs_tes_buffer_address(ctx, NULL,
6435 LLVMConstInt(ctx->ac.i32, param_outer, 0));
6436
6437 outer_vec = ac_build_gather_values(&ctx->ac, outer,
6438 util_next_power_of_two(outer_comps));
6439
6440 ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, outer_vec,
6441 outer_comps, tf_outer_offset,
6442 ctx->oc_lds, 0, 1, 0, true, false);
6443 if (inner_comps) {
6444 param_inner = shader_io_get_unique_index(VARYING_SLOT_TESS_LEVEL_INNER);
6445 tf_inner_offset = get_tcs_tes_buffer_address(ctx, NULL,
6446 LLVMConstInt(ctx->ac.i32, param_inner, 0));
6447
6448 inner_vec = inner_comps == 1 ? inner[0] :
6449 ac_build_gather_values(&ctx->ac, inner, inner_comps);
6450 ac_build_buffer_store_dword(&ctx->ac, ctx->hs_ring_tess_offchip, inner_vec,
6451 inner_comps, tf_inner_offset,
6452 ctx->oc_lds, 0, 1, 0, true, false);
6453 }
6454 }
6455 ac_nir_build_endif(&if_ctx);
6456 }
6457
6458 static void
6459 handle_tcs_outputs_post(struct nir_to_llvm_context *ctx)
6460 {
6461 write_tess_factors(ctx);
6462 }
6463
6464 static bool
6465 si_export_mrt_color(struct nir_to_llvm_context *ctx,
6466 LLVMValueRef *color, unsigned param, bool is_last,
6467 struct ac_export_args *args)
6468 {
6469 /* Export */
6470 si_llvm_init_export_args(ctx, color, param,
6471 args);
6472
6473 if (is_last) {
6474 args->valid_mask = 1; /* whether the EXEC mask is valid */
6475 args->done = 1; /* DONE bit */
6476 } else if (!args->enabled_channels)
6477 return false; /* unnecessary NULL export */
6478
6479 return true;
6480 }
6481
6482 static void
6483 radv_export_mrt_z(struct nir_to_llvm_context *ctx,
6484 LLVMValueRef depth, LLVMValueRef stencil,
6485 LLVMValueRef samplemask)
6486 {
6487 struct ac_export_args args;
6488
6489 ac_export_mrt_z(&ctx->ac, depth, stencil, samplemask, &args);
6490
6491 ac_build_export(&ctx->ac, &args);
6492 }
6493
6494 static void
6495 handle_fs_outputs_post(struct nir_to_llvm_context *ctx)
6496 {
6497 unsigned index = 0;
6498 LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
6499 struct ac_export_args color_args[8];
6500
6501 for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
6502 LLVMValueRef values[4];
6503
6504 if (!(ctx->output_mask & (1ull << i)))
6505 continue;
6506
6507 if (i == FRAG_RESULT_DEPTH) {
6508 ctx->shader_info->fs.writes_z = true;
6509 depth = ac_to_float(&ctx->ac, LLVMBuildLoad(ctx->builder,
6510 ctx->nir->outputs[radeon_llvm_reg_index_soa(i, 0)], ""));
6511 } else if (i == FRAG_RESULT_STENCIL) {
6512 ctx->shader_info->fs.writes_stencil = true;
6513 stencil = ac_to_float(&ctx->ac, LLVMBuildLoad(ctx->builder,
6514 ctx->nir->outputs[radeon_llvm_reg_index_soa(i, 0)], ""));
6515 } else if (i == FRAG_RESULT_SAMPLE_MASK) {
6516 ctx->shader_info->fs.writes_sample_mask = true;
6517 samplemask = ac_to_float(&ctx->ac, LLVMBuildLoad(ctx->builder,
6518 ctx->nir->outputs[radeon_llvm_reg_index_soa(i, 0)], ""));
6519 } else {
6520 bool last = false;
6521 for (unsigned j = 0; j < 4; j++)
6522 values[j] = ac_to_float(&ctx->ac, LLVMBuildLoad(ctx->builder,
6523 ctx->nir->outputs[radeon_llvm_reg_index_soa(i, j)], ""));
6524
6525 if (!ctx->shader_info->fs.writes_z && !ctx->shader_info->fs.writes_stencil && !ctx->shader_info->fs.writes_sample_mask)
6526 last = ctx->output_mask <= ((1ull << (i + 1)) - 1);
6527
6528 bool ret = si_export_mrt_color(ctx, values, V_008DFC_SQ_EXP_MRT + (i - FRAG_RESULT_DATA0), last, &color_args[index]);
6529 if (ret)
6530 index++;
6531 }
6532 }
6533
6534 for (unsigned i = 0; i < index; i++)
6535 ac_build_export(&ctx->ac, &color_args[i]);
6536 if (depth || stencil || samplemask)
6537 radv_export_mrt_z(ctx, depth, stencil, samplemask);
6538 else if (!index) {
6539 si_export_mrt_color(ctx, NULL, V_008DFC_SQ_EXP_NULL, true, &color_args[0]);
6540 ac_build_export(&ctx->ac, &color_args[0]);
6541 }
6542 }
6543
6544 static void
6545 emit_gs_epilogue(struct nir_to_llvm_context *ctx)
6546 {
6547 ac_build_sendmsg(&ctx->ac, AC_SENDMSG_GS_OP_NOP | AC_SENDMSG_GS_DONE, ctx->gs_wave_id);
6548 }
6549
6550 static void
6551 handle_shader_outputs_post(struct ac_shader_abi *abi, unsigned max_outputs,
6552 LLVMValueRef *addrs)
6553 {
6554 struct nir_to_llvm_context *ctx = nir_to_llvm_context_from_abi(abi);
6555
6556 switch (ctx->stage) {
6557 case MESA_SHADER_VERTEX:
6558 if (ctx->options->key.vs.as_ls)
6559 handle_ls_outputs_post(ctx);
6560 else if (ctx->options->key.vs.as_es)
6561 handle_es_outputs_post(ctx, &ctx->shader_info->vs.es_info);
6562 else
6563 handle_vs_outputs_post(ctx, ctx->options->key.vs.export_prim_id,
6564 &ctx->shader_info->vs.outinfo);
6565 break;
6566 case MESA_SHADER_FRAGMENT:
6567 handle_fs_outputs_post(ctx);
6568 break;
6569 case MESA_SHADER_GEOMETRY:
6570 emit_gs_epilogue(ctx);
6571 break;
6572 case MESA_SHADER_TESS_CTRL:
6573 handle_tcs_outputs_post(ctx);
6574 break;
6575 case MESA_SHADER_TESS_EVAL:
6576 if (ctx->options->key.tes.as_es)
6577 handle_es_outputs_post(ctx, &ctx->shader_info->tes.es_info);
6578 else
6579 handle_vs_outputs_post(ctx, ctx->options->key.tes.export_prim_id,
6580 &ctx->shader_info->tes.outinfo);
6581 break;
6582 default:
6583 break;
6584 }
6585 }
6586
6587 static void ac_llvm_finalize_module(struct nir_to_llvm_context * ctx)
6588 {
6589 LLVMPassManagerRef passmgr;
6590 /* Create the pass manager */
6591 passmgr = LLVMCreateFunctionPassManagerForModule(
6592 ctx->module);
6593
6594 /* This pass should eliminate all the load and store instructions */
6595 LLVMAddPromoteMemoryToRegisterPass(passmgr);
6596
6597 /* Add some optimization passes */
6598 LLVMAddScalarReplAggregatesPass(passmgr);
6599 LLVMAddLICMPass(passmgr);
6600 LLVMAddAggressiveDCEPass(passmgr);
6601 LLVMAddCFGSimplificationPass(passmgr);
6602 LLVMAddInstructionCombiningPass(passmgr);
6603
6604 /* Run the pass */
6605 LLVMInitializeFunctionPassManager(passmgr);
6606 LLVMRunFunctionPassManager(passmgr, ctx->main_function);
6607 LLVMFinalizeFunctionPassManager(passmgr);
6608
6609 LLVMDisposeBuilder(ctx->builder);
6610 LLVMDisposePassManager(passmgr);
6611 }
6612
6613 static void
6614 ac_nir_eliminate_const_vs_outputs(struct nir_to_llvm_context *ctx)
6615 {
6616 struct ac_vs_output_info *outinfo;
6617
6618 switch (ctx->stage) {
6619 case MESA_SHADER_FRAGMENT:
6620 case MESA_SHADER_COMPUTE:
6621 case MESA_SHADER_TESS_CTRL:
6622 case MESA_SHADER_GEOMETRY:
6623 return;
6624 case MESA_SHADER_VERTEX:
6625 if (ctx->options->key.vs.as_ls ||
6626 ctx->options->key.vs.as_es)
6627 return;
6628 outinfo = &ctx->shader_info->vs.outinfo;
6629 break;
6630 case MESA_SHADER_TESS_EVAL:
6631 if (ctx->options->key.vs.as_es)
6632 return;
6633 outinfo = &ctx->shader_info->tes.outinfo;
6634 break;
6635 default:
6636 unreachable("Unhandled shader type");
6637 }
6638
6639 ac_optimize_vs_outputs(&ctx->ac,
6640 ctx->main_function,
6641 outinfo->vs_output_param_offset,
6642 VARYING_SLOT_MAX,
6643 &outinfo->param_exports);
6644 }
6645
6646 static void
6647 ac_setup_rings(struct nir_to_llvm_context *ctx)
6648 {
6649 if ((ctx->stage == MESA_SHADER_VERTEX && ctx->options->key.vs.as_es) ||
6650 (ctx->stage == MESA_SHADER_TESS_EVAL && ctx->options->key.tes.as_es)) {
6651 ctx->esgs_ring = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_ESGS_VS, false));
6652 }
6653
6654 if (ctx->is_gs_copy_shader) {
6655 ctx->gsvs_ring = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_GSVS_VS, false));
6656 }
6657 if (ctx->stage == MESA_SHADER_GEOMETRY) {
6658 LLVMValueRef tmp;
6659 ctx->esgs_ring = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_ESGS_GS, false));
6660 ctx->gsvs_ring = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_GSVS_GS, false));
6661
6662 ctx->gsvs_ring = LLVMBuildBitCast(ctx->builder, ctx->gsvs_ring, ctx->ac.v4i32, "");
6663
6664 ctx->gsvs_ring = LLVMBuildInsertElement(ctx->builder, ctx->gsvs_ring, ctx->gsvs_num_entries, LLVMConstInt(ctx->ac.i32, 2, false), "");
6665 tmp = LLVMBuildExtractElement(ctx->builder, ctx->gsvs_ring, ctx->ac.i32_1, "");
6666 tmp = LLVMBuildOr(ctx->builder, tmp, ctx->gsvs_ring_stride, "");
6667 ctx->gsvs_ring = LLVMBuildInsertElement(ctx->builder, ctx->gsvs_ring, tmp, ctx->ac.i32_1, "");
6668 }
6669
6670 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
6671 ctx->stage == MESA_SHADER_TESS_EVAL) {
6672 ctx->hs_ring_tess_offchip = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_HS_TESS_OFFCHIP, false));
6673 ctx->hs_ring_tess_factor = ac_build_load_to_sgpr(&ctx->ac, ctx->ring_offsets, LLVMConstInt(ctx->ac.i32, RING_HS_TESS_FACTOR, false));
6674 }
6675 }
6676
6677 static unsigned
6678 ac_nir_get_max_workgroup_size(enum chip_class chip_class,
6679 const struct nir_shader *nir)
6680 {
6681 switch (nir->info.stage) {
6682 case MESA_SHADER_TESS_CTRL:
6683 return chip_class >= CIK ? 128 : 64;
6684 case MESA_SHADER_GEOMETRY:
6685 return chip_class >= GFX9 ? 128 : 64;
6686 case MESA_SHADER_COMPUTE:
6687 break;
6688 default:
6689 return 0;
6690 }
6691
6692 unsigned max_workgroup_size = nir->info.cs.local_size[0] *
6693 nir->info.cs.local_size[1] *
6694 nir->info.cs.local_size[2];
6695 return max_workgroup_size;
6696 }
6697
6698 /* Fixup the HW not emitting the TCS regs if there are no HS threads. */
6699 static void ac_nir_fixup_ls_hs_input_vgprs(struct nir_to_llvm_context *ctx)
6700 {
6701 LLVMValueRef count = ac_build_bfe(&ctx->ac, ctx->merged_wave_info,
6702 LLVMConstInt(ctx->ac.i32, 8, false),
6703 LLVMConstInt(ctx->ac.i32, 8, false), false);
6704 LLVMValueRef hs_empty = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, count,
6705 ctx->ac.i32_0, "");
6706 ctx->abi.instance_id = LLVMBuildSelect(ctx->ac.builder, hs_empty, ctx->rel_auto_id, ctx->abi.instance_id, "");
6707 ctx->vs_prim_id = LLVMBuildSelect(ctx->ac.builder, hs_empty, ctx->abi.vertex_id, ctx->vs_prim_id, "");
6708 ctx->rel_auto_id = LLVMBuildSelect(ctx->ac.builder, hs_empty, ctx->abi.tcs_rel_ids, ctx->rel_auto_id, "");
6709 ctx->abi.vertex_id = LLVMBuildSelect(ctx->ac.builder, hs_empty, ctx->abi.tcs_patch_id, ctx->abi.vertex_id, "");
6710 }
6711
6712 static void prepare_gs_input_vgprs(struct nir_to_llvm_context *ctx)
6713 {
6714 for(int i = 5; i >= 0; --i) {
6715 ctx->gs_vtx_offset[i] = ac_build_bfe(&ctx->ac, ctx->gs_vtx_offset[i & ~1],
6716 LLVMConstInt(ctx->ac.i32, (i & 1) * 16, false),
6717 LLVMConstInt(ctx->ac.i32, 16, false), false);
6718 }
6719
6720 ctx->gs_wave_id = ac_build_bfe(&ctx->ac, ctx->merged_wave_info,
6721 LLVMConstInt(ctx->ac.i32, 16, false),
6722 LLVMConstInt(ctx->ac.i32, 8, false), false);
6723 }
6724
6725 void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
6726 struct nir_shader *nir, struct nir_to_llvm_context *nctx)
6727 {
6728 struct ac_nir_context ctx = {};
6729 struct nir_function *func;
6730
6731 ctx.ac = *ac;
6732 ctx.abi = abi;
6733
6734 ctx.nctx = nctx;
6735 if (nctx)
6736 nctx->nir = &ctx;
6737
6738 ctx.stage = nir->info.stage;
6739
6740 ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
6741
6742 nir_foreach_variable(variable, &nir->outputs)
6743 handle_shader_output_decl(&ctx, nir, variable);
6744
6745 ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
6746 _mesa_key_pointer_equal);
6747 ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
6748 _mesa_key_pointer_equal);
6749 ctx.vars = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
6750 _mesa_key_pointer_equal);
6751
6752 func = (struct nir_function *)exec_list_get_head(&nir->functions);
6753
6754 setup_locals(&ctx, func);
6755
6756 if (nir->info.stage == MESA_SHADER_COMPUTE)
6757 setup_shared(&ctx, nir);
6758
6759 visit_cf_list(&ctx, &func->impl->body);
6760 phi_post_pass(&ctx);
6761
6762 if (nir->info.stage != MESA_SHADER_COMPUTE)
6763 ctx.abi->emit_outputs(ctx.abi, RADEON_LLVM_MAX_OUTPUTS,
6764 ctx.outputs);
6765
6766 free(ctx.locals);
6767 ralloc_free(ctx.defs);
6768 ralloc_free(ctx.phis);
6769 ralloc_free(ctx.vars);
6770
6771 if (nctx)
6772 nctx->nir = NULL;
6773 }
6774
6775 static
6776 LLVMModuleRef ac_translate_nir_to_llvm(LLVMTargetMachineRef tm,
6777 struct nir_shader *const *shaders,
6778 int shader_count,
6779 struct ac_shader_variant_info *shader_info,
6780 const struct ac_nir_compiler_options *options)
6781 {
6782 struct nir_to_llvm_context ctx = {0};
6783 unsigned i;
6784 ctx.options = options;
6785 ctx.shader_info = shader_info;
6786 ctx.context = LLVMContextCreate();
6787 ctx.module = LLVMModuleCreateWithNameInContext("shader", ctx.context);
6788
6789 ac_llvm_context_init(&ctx.ac, ctx.context, options->chip_class,
6790 options->family);
6791 ctx.ac.module = ctx.module;
6792 LLVMSetTarget(ctx.module, options->supports_spill ? "amdgcn-mesa-mesa3d" : "amdgcn--");
6793
6794 LLVMTargetDataRef data_layout = LLVMCreateTargetDataLayout(tm);
6795 char *data_layout_str = LLVMCopyStringRepOfTargetData(data_layout);
6796 LLVMSetDataLayout(ctx.module, data_layout_str);
6797 LLVMDisposeTargetData(data_layout);
6798 LLVMDisposeMessage(data_layout_str);
6799
6800 enum ac_float_mode float_mode =
6801 options->unsafe_math ? AC_FLOAT_MODE_UNSAFE_FP_MATH :
6802 AC_FLOAT_MODE_DEFAULT;
6803
6804 ctx.builder = ac_create_builder(ctx.context, float_mode);
6805 ctx.ac.builder = ctx.builder;
6806
6807 memset(shader_info, 0, sizeof(*shader_info));
6808
6809 for(int i = 0; i < shader_count; ++i)
6810 ac_nir_shader_info_pass(shaders[i], options, &shader_info->info);
6811
6812 for (i = 0; i < AC_UD_MAX_SETS; i++)
6813 shader_info->user_sgprs_locs.descriptor_sets[i].sgpr_idx = -1;
6814 for (i = 0; i < AC_UD_MAX_UD; i++)
6815 shader_info->user_sgprs_locs.shader_data[i].sgpr_idx = -1;
6816
6817 ctx.max_workgroup_size = 0;
6818 for (int i = 0; i < shader_count; ++i) {
6819 ctx.max_workgroup_size = MAX2(ctx.max_workgroup_size,
6820 ac_nir_get_max_workgroup_size(ctx.options->chip_class,
6821 shaders[i]));
6822 }
6823
6824 create_function(&ctx, shaders[shader_count - 1]->info.stage, shader_count >= 2,
6825 shader_count >= 2 ? shaders[shader_count - 2]->info.stage : MESA_SHADER_VERTEX);
6826
6827 ctx.abi.inputs = &ctx.inputs[0];
6828 ctx.abi.emit_outputs = handle_shader_outputs_post;
6829 ctx.abi.emit_vertex = visit_emit_vertex;
6830 ctx.abi.load_ubo = radv_load_ubo;
6831 ctx.abi.load_ssbo = radv_load_ssbo;
6832 ctx.abi.load_sampler_desc = radv_get_sampler_desc;
6833 ctx.abi.clamp_shadow_reference = false;
6834
6835 if (shader_count >= 2)
6836 ac_init_exec_full_mask(&ctx.ac);
6837
6838 if (ctx.ac.chip_class == GFX9 &&
6839 shaders[shader_count - 1]->info.stage == MESA_SHADER_TESS_CTRL)
6840 ac_nir_fixup_ls_hs_input_vgprs(&ctx);
6841
6842 for(int i = 0; i < shader_count; ++i) {
6843 ctx.stage = shaders[i]->info.stage;
6844 ctx.output_mask = 0;
6845 ctx.tess_outputs_written = 0;
6846 ctx.num_output_clips = shaders[i]->info.clip_distance_array_size;
6847 ctx.num_output_culls = shaders[i]->info.cull_distance_array_size;
6848
6849 if (shaders[i]->info.stage == MESA_SHADER_GEOMETRY) {
6850 ctx.gs_next_vertex = ac_build_alloca(&ctx.ac, ctx.ac.i32, "gs_next_vertex");
6851 ctx.gs_max_out_vertices = shaders[i]->info.gs.vertices_out;
6852 ctx.abi.load_inputs = load_gs_input;
6853 ctx.abi.emit_primitive = visit_end_primitive;
6854 } else if (shaders[i]->info.stage == MESA_SHADER_TESS_CTRL) {
6855 ctx.tcs_outputs_read = shaders[i]->info.outputs_read;
6856 ctx.tcs_patch_outputs_read = shaders[i]->info.patch_outputs_read;
6857 ctx.abi.load_tess_varyings = load_tcs_varyings;
6858 ctx.abi.load_patch_vertices_in = load_patch_vertices_in;
6859 ctx.abi.store_tcs_outputs = store_tcs_output;
6860 } else if (shaders[i]->info.stage == MESA_SHADER_TESS_EVAL) {
6861 ctx.tes_primitive_mode = shaders[i]->info.tess.primitive_mode;
6862 ctx.abi.load_tess_varyings = load_tes_input;
6863 ctx.abi.load_tess_coord = load_tess_coord;
6864 ctx.abi.load_patch_vertices_in = load_patch_vertices_in;
6865 } else if (shaders[i]->info.stage == MESA_SHADER_VERTEX) {
6866 if (shader_info->info.vs.needs_instance_id) {
6867 if (ctx.options->key.vs.as_ls) {
6868 ctx.shader_info->vs.vgpr_comp_cnt =
6869 MAX2(2, ctx.shader_info->vs.vgpr_comp_cnt);
6870 } else {
6871 ctx.shader_info->vs.vgpr_comp_cnt =
6872 MAX2(1, ctx.shader_info->vs.vgpr_comp_cnt);
6873 }
6874 }
6875 } else if (shaders[i]->info.stage == MESA_SHADER_FRAGMENT) {
6876 shader_info->fs.can_discard = shaders[i]->info.fs.uses_discard;
6877 ctx.abi.lookup_interp_param = lookup_interp_param;
6878 ctx.abi.load_sample_position = load_sample_position;
6879 }
6880
6881 if (i)
6882 emit_barrier(&ctx.ac, ctx.stage);
6883
6884 ac_setup_rings(&ctx);
6885
6886 LLVMBasicBlockRef merge_block;
6887 if (shader_count >= 2) {
6888 LLVMValueRef fn = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
6889 LLVMBasicBlockRef then_block = LLVMAppendBasicBlockInContext(ctx.ac.context, fn, "");
6890 merge_block = LLVMAppendBasicBlockInContext(ctx.ac.context, fn, "");
6891
6892 LLVMValueRef count = ac_build_bfe(&ctx.ac, ctx.merged_wave_info,
6893 LLVMConstInt(ctx.ac.i32, 8 * i, false),
6894 LLVMConstInt(ctx.ac.i32, 8, false), false);
6895 LLVMValueRef thread_id = ac_get_thread_id(&ctx.ac);
6896 LLVMValueRef cond = LLVMBuildICmp(ctx.ac.builder, LLVMIntULT,
6897 thread_id, count, "");
6898 LLVMBuildCondBr(ctx.ac.builder, cond, then_block, merge_block);
6899
6900 LLVMPositionBuilderAtEnd(ctx.ac.builder, then_block);
6901 }
6902
6903 if (shaders[i]->info.stage == MESA_SHADER_FRAGMENT)
6904 handle_fs_inputs(&ctx, shaders[i]);
6905 else if(shaders[i]->info.stage == MESA_SHADER_VERTEX)
6906 handle_vs_inputs(&ctx, shaders[i]);
6907 else if(shader_count >= 2 && shaders[i]->info.stage == MESA_SHADER_GEOMETRY)
6908 prepare_gs_input_vgprs(&ctx);
6909
6910 nir_foreach_variable(variable, &shaders[i]->outputs)
6911 scan_shader_output_decl(&ctx, variable, shaders[i], shaders[i]->info.stage);
6912
6913 ac_nir_translate(&ctx.ac, &ctx.abi, shaders[i], &ctx);
6914
6915 if (shader_count >= 2) {
6916 LLVMBuildBr(ctx.ac.builder, merge_block);
6917 LLVMPositionBuilderAtEnd(ctx.ac.builder, merge_block);
6918 }
6919
6920 if (shaders[i]->info.stage == MESA_SHADER_GEOMETRY) {
6921 unsigned addclip = shaders[i]->info.clip_distance_array_size +
6922 shaders[i]->info.cull_distance_array_size > 4;
6923 shader_info->gs.gsvs_vertex_size = (util_bitcount64(ctx.output_mask) + addclip) * 16;
6924 shader_info->gs.max_gsvs_emit_size = shader_info->gs.gsvs_vertex_size *
6925 shaders[i]->info.gs.vertices_out;
6926 } else if (shaders[i]->info.stage == MESA_SHADER_TESS_CTRL) {
6927 shader_info->tcs.outputs_written = ctx.tess_outputs_written;
6928 shader_info->tcs.patch_outputs_written = ctx.tess_patch_outputs_written;
6929 } else if (shaders[i]->info.stage == MESA_SHADER_VERTEX && ctx.options->key.vs.as_ls) {
6930 shader_info->vs.outputs_written = ctx.tess_outputs_written;
6931 }
6932 }
6933
6934 LLVMBuildRetVoid(ctx.builder);
6935
6936 if (options->dump_preoptir)
6937 ac_dump_module(ctx.module);
6938
6939 ac_llvm_finalize_module(&ctx);
6940
6941 if (shader_count == 1)
6942 ac_nir_eliminate_const_vs_outputs(&ctx);
6943
6944 return ctx.module;
6945 }
6946
6947 static void ac_diagnostic_handler(LLVMDiagnosticInfoRef di, void *context)
6948 {
6949 unsigned *retval = (unsigned *)context;
6950 LLVMDiagnosticSeverity severity = LLVMGetDiagInfoSeverity(di);
6951 char *description = LLVMGetDiagInfoDescription(di);
6952
6953 if (severity == LLVMDSError) {
6954 *retval = 1;
6955 fprintf(stderr, "LLVM triggered Diagnostic Handler: %s\n",
6956 description);
6957 }
6958
6959 LLVMDisposeMessage(description);
6960 }
6961
6962 static unsigned ac_llvm_compile(LLVMModuleRef M,
6963 struct ac_shader_binary *binary,
6964 LLVMTargetMachineRef tm)
6965 {
6966 unsigned retval = 0;
6967 char *err;
6968 LLVMContextRef llvm_ctx;
6969 LLVMMemoryBufferRef out_buffer;
6970 unsigned buffer_size;
6971 const char *buffer_data;
6972 LLVMBool mem_err;
6973
6974 /* Setup Diagnostic Handler*/
6975 llvm_ctx = LLVMGetModuleContext(M);
6976
6977 LLVMContextSetDiagnosticHandler(llvm_ctx, ac_diagnostic_handler,
6978 &retval);
6979
6980 /* Compile IR*/
6981 mem_err = LLVMTargetMachineEmitToMemoryBuffer(tm, M, LLVMObjectFile,
6982 &err, &out_buffer);
6983
6984 /* Process Errors/Warnings */
6985 if (mem_err) {
6986 fprintf(stderr, "%s: %s", __FUNCTION__, err);
6987 free(err);
6988 retval = 1;
6989 goto out;
6990 }
6991
6992 /* Extract Shader Code*/
6993 buffer_size = LLVMGetBufferSize(out_buffer);
6994 buffer_data = LLVMGetBufferStart(out_buffer);
6995
6996 ac_elf_read(buffer_data, buffer_size, binary);
6997
6998 /* Clean up */
6999 LLVMDisposeMemoryBuffer(out_buffer);
7000
7001 out:
7002 return retval;
7003 }
7004
7005 static void ac_compile_llvm_module(LLVMTargetMachineRef tm,
7006 LLVMModuleRef llvm_module,
7007 struct ac_shader_binary *binary,
7008 struct ac_shader_config *config,
7009 struct ac_shader_variant_info *shader_info,
7010 gl_shader_stage stage,
7011 bool dump_shader, bool supports_spill)
7012 {
7013 if (dump_shader)
7014 ac_dump_module(llvm_module);
7015
7016 memset(binary, 0, sizeof(*binary));
7017 int v = ac_llvm_compile(llvm_module, binary, tm);
7018 if (v) {
7019 fprintf(stderr, "compile failed\n");
7020 }
7021
7022 if (dump_shader)
7023 fprintf(stderr, "disasm:\n%s\n", binary->disasm_string);
7024
7025 ac_shader_binary_read_config(binary, config, 0, supports_spill);
7026
7027 LLVMContextRef ctx = LLVMGetModuleContext(llvm_module);
7028 LLVMDisposeModule(llvm_module);
7029 LLVMContextDispose(ctx);
7030
7031 if (stage == MESA_SHADER_FRAGMENT) {
7032 shader_info->num_input_vgprs = 0;
7033 if (G_0286CC_PERSP_SAMPLE_ENA(config->spi_ps_input_addr))
7034 shader_info->num_input_vgprs += 2;
7035 if (G_0286CC_PERSP_CENTER_ENA(config->spi_ps_input_addr))
7036 shader_info->num_input_vgprs += 2;
7037 if (G_0286CC_PERSP_CENTROID_ENA(config->spi_ps_input_addr))
7038 shader_info->num_input_vgprs += 2;
7039 if (G_0286CC_PERSP_PULL_MODEL_ENA(config->spi_ps_input_addr))
7040 shader_info->num_input_vgprs += 3;
7041 if (G_0286CC_LINEAR_SAMPLE_ENA(config->spi_ps_input_addr))
7042 shader_info->num_input_vgprs += 2;
7043 if (G_0286CC_LINEAR_CENTER_ENA(config->spi_ps_input_addr))
7044 shader_info->num_input_vgprs += 2;
7045 if (G_0286CC_LINEAR_CENTROID_ENA(config->spi_ps_input_addr))
7046 shader_info->num_input_vgprs += 2;
7047 if (G_0286CC_LINE_STIPPLE_TEX_ENA(config->spi_ps_input_addr))
7048 shader_info->num_input_vgprs += 1;
7049 if (G_0286CC_POS_X_FLOAT_ENA(config->spi_ps_input_addr))
7050 shader_info->num_input_vgprs += 1;
7051 if (G_0286CC_POS_Y_FLOAT_ENA(config->spi_ps_input_addr))
7052 shader_info->num_input_vgprs += 1;
7053 if (G_0286CC_POS_Z_FLOAT_ENA(config->spi_ps_input_addr))
7054 shader_info->num_input_vgprs += 1;
7055 if (G_0286CC_POS_W_FLOAT_ENA(config->spi_ps_input_addr))
7056 shader_info->num_input_vgprs += 1;
7057 if (G_0286CC_FRONT_FACE_ENA(config->spi_ps_input_addr))
7058 shader_info->num_input_vgprs += 1;
7059 if (G_0286CC_ANCILLARY_ENA(config->spi_ps_input_addr))
7060 shader_info->num_input_vgprs += 1;
7061 if (G_0286CC_SAMPLE_COVERAGE_ENA(config->spi_ps_input_addr))
7062 shader_info->num_input_vgprs += 1;
7063 if (G_0286CC_POS_FIXED_PT_ENA(config->spi_ps_input_addr))
7064 shader_info->num_input_vgprs += 1;
7065 }
7066 config->num_vgprs = MAX2(config->num_vgprs, shader_info->num_input_vgprs);
7067
7068 /* +3 for scratch wave offset and VCC */
7069 config->num_sgprs = MAX2(config->num_sgprs,
7070 shader_info->num_input_sgprs + 3);
7071
7072 /* Enable 64-bit and 16-bit denormals, because there is no performance
7073 * cost.
7074 *
7075 * If denormals are enabled, all floating-point output modifiers are
7076 * ignored.
7077 *
7078 * Don't enable denormals for 32-bit floats, because:
7079 * - Floating-point output modifiers would be ignored by the hw.
7080 * - Some opcodes don't support denormals, such as v_mad_f32. We would
7081 * have to stop using those.
7082 * - SI & CI would be very slow.
7083 */
7084 config->float_mode |= V_00B028_FP_64_DENORMS;
7085 }
7086
7087 static void
7088 ac_fill_shader_info(struct ac_shader_variant_info *shader_info, struct nir_shader *nir, const struct ac_nir_compiler_options *options)
7089 {
7090 switch (nir->info.stage) {
7091 case MESA_SHADER_COMPUTE:
7092 for (int i = 0; i < 3; ++i)
7093 shader_info->cs.block_size[i] = nir->info.cs.local_size[i];
7094 break;
7095 case MESA_SHADER_FRAGMENT:
7096 shader_info->fs.early_fragment_test = nir->info.fs.early_fragment_tests;
7097 break;
7098 case MESA_SHADER_GEOMETRY:
7099 shader_info->gs.vertices_in = nir->info.gs.vertices_in;
7100 shader_info->gs.vertices_out = nir->info.gs.vertices_out;
7101 shader_info->gs.output_prim = nir->info.gs.output_primitive;
7102 shader_info->gs.invocations = nir->info.gs.invocations;
7103 break;
7104 case MESA_SHADER_TESS_EVAL:
7105 shader_info->tes.primitive_mode = nir->info.tess.primitive_mode;
7106 shader_info->tes.spacing = nir->info.tess.spacing;
7107 shader_info->tes.ccw = nir->info.tess.ccw;
7108 shader_info->tes.point_mode = nir->info.tess.point_mode;
7109 shader_info->tes.as_es = options->key.tes.as_es;
7110 break;
7111 case MESA_SHADER_TESS_CTRL:
7112 shader_info->tcs.tcs_vertices_out = nir->info.tess.tcs_vertices_out;
7113 break;
7114 case MESA_SHADER_VERTEX:
7115 shader_info->vs.as_es = options->key.vs.as_es;
7116 shader_info->vs.as_ls = options->key.vs.as_ls;
7117 /* in LS mode we need at least 1, invocation id needs 2, handled elsewhere */
7118 if (options->key.vs.as_ls)
7119 shader_info->vs.vgpr_comp_cnt = MAX2(1, shader_info->vs.vgpr_comp_cnt);
7120 break;
7121 default:
7122 break;
7123 }
7124 }
7125
7126 void ac_compile_nir_shader(LLVMTargetMachineRef tm,
7127 struct ac_shader_binary *binary,
7128 struct ac_shader_config *config,
7129 struct ac_shader_variant_info *shader_info,
7130 struct nir_shader *const *nir,
7131 int nir_count,
7132 const struct ac_nir_compiler_options *options,
7133 bool dump_shader)
7134 {
7135
7136 LLVMModuleRef llvm_module = ac_translate_nir_to_llvm(tm, nir, nir_count, shader_info,
7137 options);
7138
7139 ac_compile_llvm_module(tm, llvm_module, binary, config, shader_info, nir[0]->info.stage, dump_shader, options->supports_spill);
7140 for (int i = 0; i < nir_count; ++i)
7141 ac_fill_shader_info(shader_info, nir[i], options);
7142
7143 /* Determine the ES type (VS or TES) for the GS on GFX9. */
7144 if (options->chip_class == GFX9) {
7145 if (nir_count == 2 &&
7146 nir[1]->info.stage == MESA_SHADER_GEOMETRY) {
7147 shader_info->gs.es_type = nir[0]->info.stage;
7148 }
7149 }
7150 }
7151
7152 static void
7153 ac_gs_copy_shader_emit(struct nir_to_llvm_context *ctx)
7154 {
7155 LLVMValueRef vtx_offset =
7156 LLVMBuildMul(ctx->builder, ctx->abi.vertex_id,
7157 LLVMConstInt(ctx->ac.i32, 4, false), "");
7158 int idx = 0;
7159
7160 for (unsigned i = 0; i < RADEON_LLVM_MAX_OUTPUTS; ++i) {
7161 int length = 4;
7162 int slot = idx;
7163 int slot_inc = 1;
7164 if (!(ctx->output_mask & (1ull << i)))
7165 continue;
7166
7167 if (i == VARYING_SLOT_CLIP_DIST0) {
7168 /* unpack clip and cull from a single set of slots */
7169 length = ctx->num_output_clips + ctx->num_output_culls;
7170 if (length > 4)
7171 slot_inc = 2;
7172 }
7173
7174 for (unsigned j = 0; j < length; j++) {
7175 LLVMValueRef value, soffset;
7176
7177 soffset = LLVMConstInt(ctx->ac.i32,
7178 (slot * 4 + j) *
7179 ctx->gs_max_out_vertices * 16 * 4, false);
7180
7181 value = ac_build_buffer_load(&ctx->ac, ctx->gsvs_ring,
7182 1, ctx->ac.i32_0,
7183 vtx_offset, soffset,
7184 0, 1, 1, true, false);
7185
7186 LLVMBuildStore(ctx->builder,
7187 ac_to_float(&ctx->ac, value), ctx->nir->outputs[radeon_llvm_reg_index_soa(i, j)]);
7188 }
7189 idx += slot_inc;
7190 }
7191 handle_vs_outputs_post(ctx, false, &ctx->shader_info->vs.outinfo);
7192 }
7193
7194 void ac_create_gs_copy_shader(LLVMTargetMachineRef tm,
7195 struct nir_shader *geom_shader,
7196 struct ac_shader_binary *binary,
7197 struct ac_shader_config *config,
7198 struct ac_shader_variant_info *shader_info,
7199 const struct ac_nir_compiler_options *options,
7200 bool dump_shader)
7201 {
7202 struct nir_to_llvm_context ctx = {0};
7203 ctx.context = LLVMContextCreate();
7204 ctx.module = LLVMModuleCreateWithNameInContext("shader", ctx.context);
7205 ctx.options = options;
7206 ctx.shader_info = shader_info;
7207
7208 ac_llvm_context_init(&ctx.ac, ctx.context, options->chip_class,
7209 options->family);
7210 ctx.ac.module = ctx.module;
7211
7212 ctx.is_gs_copy_shader = true;
7213 LLVMSetTarget(ctx.module, "amdgcn--");
7214
7215 enum ac_float_mode float_mode =
7216 options->unsafe_math ? AC_FLOAT_MODE_UNSAFE_FP_MATH :
7217 AC_FLOAT_MODE_DEFAULT;
7218
7219 ctx.builder = ac_create_builder(ctx.context, float_mode);
7220 ctx.ac.builder = ctx.builder;
7221 ctx.stage = MESA_SHADER_VERTEX;
7222
7223 create_function(&ctx, MESA_SHADER_VERTEX, false, MESA_SHADER_VERTEX);
7224
7225 ctx.gs_max_out_vertices = geom_shader->info.gs.vertices_out;
7226 ac_setup_rings(&ctx);
7227
7228 ctx.num_output_clips = geom_shader->info.clip_distance_array_size;
7229 ctx.num_output_culls = geom_shader->info.cull_distance_array_size;
7230
7231 struct ac_nir_context nir_ctx = {};
7232 nir_ctx.ac = ctx.ac;
7233 nir_ctx.abi = &ctx.abi;
7234
7235 nir_ctx.nctx = &ctx;
7236 ctx.nir = &nir_ctx;
7237
7238 nir_foreach_variable(variable, &geom_shader->outputs) {
7239 scan_shader_output_decl(&ctx, variable, geom_shader, MESA_SHADER_VERTEX);
7240 handle_shader_output_decl(&nir_ctx, geom_shader, variable);
7241 }
7242
7243 ac_gs_copy_shader_emit(&ctx);
7244
7245 ctx.nir = NULL;
7246
7247 LLVMBuildRetVoid(ctx.builder);
7248
7249 ac_llvm_finalize_module(&ctx);
7250
7251 ac_compile_llvm_module(tm, ctx.module, binary, config, shader_info,
7252 MESA_SHADER_VERTEX,
7253 dump_shader, options->supports_spill);
7254 }