1 /*
2 * Copyright © Microsoft Corporation
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 "dxil_nir.h"
25 #include "dxil_module.h"
26
27 #include "nir_builder.h"
28 #include "nir_deref.h"
29 #include "nir_worklist.h"
30 #include "nir_to_dxil.h"
31 #include "util/u_math.h"
32 #include "vulkan/vulkan_core.h"
33
34 static void
cl_type_size_align(const struct glsl_type * type,unsigned * size,unsigned * align)35 cl_type_size_align(const struct glsl_type *type, unsigned *size,
36 unsigned *align)
37 {
38 *size = glsl_get_cl_size(type);
39 *align = glsl_get_cl_alignment(type);
40 }
41
42 static nir_def *
load_comps_to_vec(nir_builder * b,unsigned src_bit_size,nir_def ** src_comps,unsigned num_src_comps,unsigned dst_bit_size)43 load_comps_to_vec(nir_builder *b, unsigned src_bit_size,
44 nir_def **src_comps, unsigned num_src_comps,
45 unsigned dst_bit_size)
46 {
47 if (src_bit_size == dst_bit_size)
48 return nir_vec(b, src_comps, num_src_comps);
49 else if (src_bit_size > dst_bit_size)
50 return nir_extract_bits(b, src_comps, num_src_comps, 0, src_bit_size * num_src_comps / dst_bit_size, dst_bit_size);
51
52 unsigned num_dst_comps = DIV_ROUND_UP(num_src_comps * src_bit_size, dst_bit_size);
53 unsigned comps_per_dst = dst_bit_size / src_bit_size;
54 nir_def *dst_comps[4];
55
56 for (unsigned i = 0; i < num_dst_comps; i++) {
57 unsigned src_offs = i * comps_per_dst;
58
59 dst_comps[i] = nir_u2uN(b, src_comps[src_offs], dst_bit_size);
60 for (unsigned j = 1; j < comps_per_dst && src_offs + j < num_src_comps; j++) {
61 nir_def *tmp = nir_ishl_imm(b, nir_u2uN(b, src_comps[src_offs + j], dst_bit_size),
62 j * src_bit_size);
63 dst_comps[i] = nir_ior(b, dst_comps[i], tmp);
64 }
65 }
66
67 return nir_vec(b, dst_comps, num_dst_comps);
68 }
69
70 static bool
lower_32b_offset_load(nir_builder * b,nir_intrinsic_instr * intr,nir_variable * var)71 lower_32b_offset_load(nir_builder *b, nir_intrinsic_instr *intr, nir_variable *var)
72 {
73 unsigned bit_size = intr->def.bit_size;
74 unsigned num_components = intr->def.num_components;
75 unsigned num_bits = num_components * bit_size;
76
77 b->cursor = nir_before_instr(&intr->instr);
78
79 nir_def *offset = intr->src[0].ssa;
80 if (intr->intrinsic == nir_intrinsic_load_shared)
81 offset = nir_iadd_imm(b, offset, nir_intrinsic_base(intr));
82 else
83 offset = nir_u2u32(b, offset);
84 nir_def *index = nir_ushr_imm(b, offset, 2);
85 nir_def *comps[NIR_MAX_VEC_COMPONENTS];
86 nir_def *comps_32bit[NIR_MAX_VEC_COMPONENTS * 2];
87
88 /* We need to split loads in 32-bit accesses because the buffer
89 * is an i32 array and DXIL does not support type casts.
90 */
91 unsigned num_32bit_comps = DIV_ROUND_UP(num_bits, 32);
92 for (unsigned i = 0; i < num_32bit_comps; i++)
93 comps_32bit[i] = nir_load_array_var(b, var, nir_iadd_imm(b, index, i));
94 unsigned num_comps_per_pass = MIN2(num_32bit_comps, 4);
95
96 for (unsigned i = 0; i < num_32bit_comps; i += num_comps_per_pass) {
97 unsigned num_vec32_comps = MIN2(num_32bit_comps - i, 4);
98 unsigned num_dest_comps = num_vec32_comps * 32 / bit_size;
99 nir_def *vec32 = nir_vec(b, &comps_32bit[i], num_vec32_comps);
100
101 /* If we have 16 bits or less to load we need to adjust the u32 value so
102 * we can always extract the LSB.
103 */
104 if (num_bits <= 16) {
105 nir_def *shift =
106 nir_imul_imm(b, nir_iand_imm(b, offset, 3), 8);
107 vec32 = nir_ushr(b, vec32, shift);
108 }
109
110 /* And now comes the pack/unpack step to match the original type. */
111 unsigned dest_index = i * 32 / bit_size;
112 nir_def *temp_vec = nir_extract_bits(b, &vec32, 1, 0, num_dest_comps, bit_size);
113 for (unsigned comp = 0; comp < num_dest_comps; ++comp, ++dest_index)
114 comps[dest_index] = nir_channel(b, temp_vec, comp);
115 }
116
117 nir_def *result = nir_vec(b, comps, num_components);
118 nir_def_replace(&intr->def, result);
119
120 return true;
121 }
122
123 static void
lower_masked_store_vec32(nir_builder * b,nir_def * offset,nir_def * index,nir_def * vec32,unsigned num_bits,nir_variable * var,unsigned alignment)124 lower_masked_store_vec32(nir_builder *b, nir_def *offset, nir_def *index,
125 nir_def *vec32, unsigned num_bits, nir_variable *var, unsigned alignment)
126 {
127 nir_def *mask = nir_imm_int(b, (1 << num_bits) - 1);
128
129 /* If we have small alignments, we need to place them correctly in the u32 component. */
130 if (alignment <= 2) {
131 nir_def *shift =
132 nir_imul_imm(b, nir_iand_imm(b, offset, 3), 8);
133
134 vec32 = nir_ishl(b, vec32, shift);
135 mask = nir_ishl(b, mask, shift);
136 }
137
138 if (var->data.mode == nir_var_mem_shared) {
139 /* Use the dedicated masked intrinsic */
140 nir_deref_instr *deref = nir_build_deref_array(b, nir_build_deref_var(b, var), index);
141 nir_deref_atomic(b, 32, &deref->def, nir_inot(b, mask), .atomic_op = nir_atomic_op_iand);
142 nir_deref_atomic(b, 32, &deref->def, vec32, .atomic_op = nir_atomic_op_ior);
143 } else {
144 /* For scratch, since we don't need atomics, just generate the read-modify-write in NIR */
145 nir_def *load = nir_load_array_var(b, var, index);
146
147 nir_def *new_val = nir_ior(b, vec32,
148 nir_iand(b,
149 nir_inot(b, mask),
150 load));
151
152 nir_store_array_var(b, var, index, new_val, 1);
153 }
154 }
155
156 static bool
lower_32b_offset_store(nir_builder * b,nir_intrinsic_instr * intr,nir_variable * var)157 lower_32b_offset_store(nir_builder *b, nir_intrinsic_instr *intr, nir_variable *var)
158 {
159 unsigned num_components = nir_src_num_components(intr->src[0]);
160 unsigned bit_size = nir_src_bit_size(intr->src[0]);
161 unsigned num_bits = num_components * bit_size;
162
163 b->cursor = nir_before_instr(&intr->instr);
164
165 nir_def *offset = intr->src[1].ssa;
166 if (intr->intrinsic == nir_intrinsic_store_shared)
167 offset = nir_iadd_imm(b, offset, nir_intrinsic_base(intr));
168 else
169 offset = nir_u2u32(b, offset);
170 nir_def *comps[NIR_MAX_VEC_COMPONENTS];
171
172 unsigned comp_idx = 0;
173 for (unsigned i = 0; i < num_components; i++)
174 comps[i] = nir_channel(b, intr->src[0].ssa, i);
175
176 unsigned step = MAX2(bit_size, 32);
177 for (unsigned i = 0; i < num_bits; i += step) {
178 /* For each 4byte chunk (or smaller) we generate a 32bit scalar store. */
179 unsigned substore_num_bits = MIN2(num_bits - i, step);
180 nir_def *local_offset = nir_iadd_imm(b, offset, i / 8);
181 nir_def *vec32 = load_comps_to_vec(b, bit_size, &comps[comp_idx],
182 substore_num_bits / bit_size, 32);
183 nir_def *index = nir_ushr_imm(b, local_offset, 2);
184
185 /* For anything less than 32bits we need to use the masked version of the
186 * intrinsic to preserve data living in the same 32bit slot. */
187 if (substore_num_bits < 32) {
188 lower_masked_store_vec32(b, local_offset, index, vec32, num_bits, var, nir_intrinsic_align(intr));
189 } else {
190 for (unsigned i = 0; i < vec32->num_components; ++i)
191 nir_store_array_var(b, var, nir_iadd_imm(b, index, i), nir_channel(b, vec32, i), 1);
192 }
193
194 comp_idx += substore_num_bits / bit_size;
195 }
196
197 nir_instr_remove(&intr->instr);
198
199 return true;
200 }
201
202 #define CONSTANT_LOCATION_UNVISITED 0
203 #define CONSTANT_LOCATION_VALID 1
204 #define CONSTANT_LOCATION_INVALID 2
205
206 bool
dxil_nir_lower_constant_to_temp(nir_shader * nir)207 dxil_nir_lower_constant_to_temp(nir_shader *nir)
208 {
209 bool progress = false;
210 nir_foreach_variable_with_modes(var, nir, nir_var_mem_constant)
211 var->data.location = var->constant_initializer ?
212 CONSTANT_LOCATION_UNVISITED : CONSTANT_LOCATION_INVALID;
213
214 /* First pass: collect all UBO accesses that could be turned into
215 * shader temp accesses.
216 */
217 nir_foreach_function(func, nir) {
218 if (!func->is_entrypoint)
219 continue;
220 assert(func->impl);
221
222 nir_foreach_block(block, func->impl) {
223 nir_foreach_instr_safe(instr, block) {
224 if (instr->type != nir_instr_type_deref)
225 continue;
226
227 nir_deref_instr *deref = nir_instr_as_deref(instr);
228 if (!nir_deref_mode_is(deref, nir_var_mem_constant) ||
229 deref->deref_type != nir_deref_type_var ||
230 deref->var->data.location == CONSTANT_LOCATION_INVALID)
231 continue;
232
233 deref->var->data.location = nir_deref_instr_has_complex_use(deref, 0) ?
234 CONSTANT_LOCATION_INVALID : CONSTANT_LOCATION_VALID;
235 }
236 }
237 }
238
239 nir_foreach_variable_with_modes(var, nir, nir_var_mem_constant) {
240 if (var->data.location != CONSTANT_LOCATION_VALID)
241 continue;
242
243 /* Change the variable mode. */
244 var->data.mode = nir_var_shader_temp;
245
246 progress = true;
247 }
248
249 /* Second pass: patch all derefs that were accessing the converted UBOs
250 * variables.
251 */
252 nir_foreach_function(func, nir) {
253 if (!func->is_entrypoint)
254 continue;
255 assert(func->impl);
256
257 nir_builder b = nir_builder_create(func->impl);
258 nir_foreach_block(block, func->impl) {
259 nir_foreach_instr_safe(instr, block) {
260 if (instr->type != nir_instr_type_deref)
261 continue;
262
263 nir_deref_instr *deref = nir_instr_as_deref(instr);
264 if (nir_deref_mode_is(deref, nir_var_mem_constant)) {
265 nir_deref_instr *parent = deref;
266 while (parent && parent->deref_type != nir_deref_type_var)
267 parent = nir_src_as_deref(parent->parent);
268 if (parent && parent->var->data.mode != nir_var_mem_constant) {
269 deref->modes = parent->var->data.mode;
270 /* Also change "pointer" size to 32-bit since this is now a logical pointer */
271 deref->def.bit_size = 32;
272 if (deref->deref_type == nir_deref_type_array) {
273 b.cursor = nir_before_instr(instr);
274 nir_src_rewrite(&deref->arr.index, nir_u2u32(&b, deref->arr.index.ssa));
275 }
276 }
277 }
278 }
279 }
280 }
281
282 return progress;
283 }
284
285 static bool
flatten_var_arrays(nir_builder * b,nir_intrinsic_instr * intr,void * data)286 flatten_var_arrays(nir_builder *b, nir_intrinsic_instr *intr, void *data)
287 {
288 switch (intr->intrinsic) {
289 case nir_intrinsic_load_deref:
290 case nir_intrinsic_store_deref:
291 case nir_intrinsic_deref_atomic:
292 case nir_intrinsic_deref_atomic_swap:
293 break;
294 default:
295 return false;
296 }
297
298 nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
299 nir_variable *var = NULL;
300 for (nir_deref_instr *d = deref; d; d = nir_deref_instr_parent(d)) {
301 if (d->deref_type == nir_deref_type_cast)
302 return false;
303 if (d->deref_type == nir_deref_type_var) {
304 var = d->var;
305 if (d->type == var->type)
306 return false;
307 }
308 }
309 if (!var)
310 return false;
311
312 nir_deref_path path;
313 nir_deref_path_init(&path, deref, NULL);
314
315 assert(path.path[0]->deref_type == nir_deref_type_var);
316 b->cursor = nir_before_instr(&path.path[0]->instr);
317 nir_deref_instr *new_var_deref = nir_build_deref_var(b, var);
318 nir_def *index = NULL;
319 for (unsigned level = 1; path.path[level]; ++level) {
320 nir_deref_instr *arr_deref = path.path[level];
321 assert(arr_deref->deref_type == nir_deref_type_array);
322 b->cursor = nir_before_instr(&arr_deref->instr);
323 nir_def *val = nir_imul_imm(b, arr_deref->arr.index.ssa,
324 glsl_get_component_slots(arr_deref->type));
325 if (index) {
326 index = nir_iadd(b, index, val);
327 } else {
328 index = val;
329 }
330 }
331
332 unsigned vector_comps = intr->num_components;
333 if (vector_comps > 1) {
334 b->cursor = nir_before_instr(&intr->instr);
335 if (intr->intrinsic == nir_intrinsic_load_deref) {
336 nir_def *components[NIR_MAX_VEC_COMPONENTS];
337 for (unsigned i = 0; i < vector_comps; ++i) {
338 nir_def *final_index = index ? nir_iadd_imm(b, index, i) : nir_imm_int(b, i);
339 nir_deref_instr *comp_deref = nir_build_deref_array(b, new_var_deref, final_index);
340 components[i] = nir_load_deref(b, comp_deref);
341 }
342 nir_def_rewrite_uses(&intr->def, nir_vec(b, components, vector_comps));
343 } else if (intr->intrinsic == nir_intrinsic_store_deref) {
344 for (unsigned i = 0; i < vector_comps; ++i) {
345 if (((1 << i) & nir_intrinsic_write_mask(intr)) == 0)
346 continue;
347 nir_def *final_index = index ? nir_iadd_imm(b, index, i) : nir_imm_int(b, i);
348 nir_deref_instr *comp_deref = nir_build_deref_array(b, new_var_deref, final_index);
349 nir_store_deref(b, comp_deref, nir_channel(b, intr->src[1].ssa, i), 1);
350 }
351 }
352 nir_instr_remove(&intr->instr);
353 } else {
354 nir_src_rewrite(&intr->src[0], &nir_build_deref_array(b, new_var_deref, index)->def);
355 }
356
357 nir_deref_path_finish(&path);
358 return true;
359 }
360
361 static void
flatten_constant_initializer(nir_variable * var,nir_constant * src,nir_constant *** dest,unsigned vector_elements)362 flatten_constant_initializer(nir_variable *var, nir_constant *src, nir_constant ***dest, unsigned vector_elements)
363 {
364 if (src->num_elements == 0) {
365 for (unsigned i = 0; i < vector_elements; ++i) {
366 nir_constant *new_scalar = rzalloc(var, nir_constant);
367 memcpy(&new_scalar->values[0], &src->values[i], sizeof(src->values[0]));
368 new_scalar->is_null_constant = src->values[i].u64 == 0;
369
370 nir_constant **array_entry = (*dest)++;
371 *array_entry = new_scalar;
372 }
373 } else {
374 for (unsigned i = 0; i < src->num_elements; ++i)
375 flatten_constant_initializer(var, src->elements[i], dest, vector_elements);
376 }
377 }
378
379 static bool
flatten_var_array_types(nir_variable * var)380 flatten_var_array_types(nir_variable *var)
381 {
382 assert(!glsl_type_is_struct(glsl_without_array(var->type)));
383 const struct glsl_type *matrix_type = glsl_without_array(var->type);
384 if (!glsl_type_is_array_of_arrays(var->type) && glsl_get_components(matrix_type) == 1)
385 return false;
386
387 enum glsl_base_type base_type = glsl_get_base_type(matrix_type);
388 const struct glsl_type *flattened_type = glsl_array_type(glsl_scalar_type(base_type),
389 glsl_get_component_slots(var->type), 0);
390 var->type = flattened_type;
391 if (var->constant_initializer) {
392 nir_constant **new_elements = ralloc_array(var, nir_constant *, glsl_get_length(flattened_type));
393 nir_constant **temp = new_elements;
394 flatten_constant_initializer(var, var->constant_initializer, &temp, glsl_get_vector_elements(matrix_type));
395 var->constant_initializer->num_elements = glsl_get_length(flattened_type);
396 var->constant_initializer->elements = new_elements;
397 }
398 return true;
399 }
400
401 bool
dxil_nir_flatten_var_arrays(nir_shader * shader,nir_variable_mode modes)402 dxil_nir_flatten_var_arrays(nir_shader *shader, nir_variable_mode modes)
403 {
404 bool progress = false;
405 nir_foreach_variable_with_modes(var, shader, modes & ~nir_var_function_temp)
406 progress |= flatten_var_array_types(var);
407
408 if (modes & nir_var_function_temp) {
409 nir_foreach_function_impl(impl, shader) {
410 nir_foreach_function_temp_variable(var, impl)
411 progress |= flatten_var_array_types(var);
412 }
413 }
414
415 if (!progress)
416 return false;
417
418 nir_shader_intrinsics_pass(shader, flatten_var_arrays,
419 nir_metadata_control_flow |
420 nir_metadata_loop_analysis,
421 NULL);
422 nir_remove_dead_derefs(shader);
423 return true;
424 }
425
426 static bool
lower_deref_bit_size(nir_builder * b,nir_intrinsic_instr * intr,void * data)427 lower_deref_bit_size(nir_builder *b, nir_intrinsic_instr *intr, void *data)
428 {
429 switch (intr->intrinsic) {
430 case nir_intrinsic_load_deref:
431 case nir_intrinsic_store_deref:
432 break;
433 default:
434 /* Atomics can't be smaller than 32-bit */
435 return false;
436 }
437
438 nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
439 nir_variable *var = nir_deref_instr_get_variable(deref);
440 /* Only interested in full deref chains */
441 if (!var)
442 return false;
443
444 const struct glsl_type *var_scalar_type = glsl_without_array(var->type);
445 if (deref->type == var_scalar_type || !glsl_type_is_scalar(var_scalar_type))
446 return false;
447
448 assert(deref->deref_type == nir_deref_type_var || deref->deref_type == nir_deref_type_array);
449 const struct glsl_type *old_glsl_type = deref->type;
450 nir_alu_type old_type = nir_get_nir_type_for_glsl_type(old_glsl_type);
451 nir_alu_type new_type = nir_get_nir_type_for_glsl_type(var_scalar_type);
452 if (glsl_get_bit_size(old_glsl_type) < glsl_get_bit_size(var_scalar_type)) {
453 deref->type = var_scalar_type;
454 if (intr->intrinsic == nir_intrinsic_load_deref) {
455 intr->def.bit_size = glsl_get_bit_size(var_scalar_type);
456 b->cursor = nir_after_instr(&intr->instr);
457 nir_def *downcast = nir_type_convert(b, &intr->def, new_type, old_type, nir_rounding_mode_undef);
458 nir_def_rewrite_uses_after(&intr->def, downcast, downcast->parent_instr);
459 }
460 else {
461 b->cursor = nir_before_instr(&intr->instr);
462 nir_def *upcast = nir_type_convert(b, intr->src[1].ssa, old_type, new_type, nir_rounding_mode_undef);
463 nir_src_rewrite(&intr->src[1], upcast);
464 }
465
466 while (deref->deref_type == nir_deref_type_array) {
467 nir_deref_instr *parent = nir_deref_instr_parent(deref);
468 parent->type = glsl_type_wrap_in_arrays(deref->type, parent->type);
469 deref = parent;
470 }
471 } else {
472 /* Assumed arrays are already flattened */
473 b->cursor = nir_before_instr(&deref->instr);
474 nir_deref_instr *parent = nir_build_deref_var(b, var);
475 if (deref->deref_type == nir_deref_type_array)
476 deref = nir_build_deref_array(b, parent, nir_imul_imm(b, deref->arr.index.ssa, 2));
477 else
478 deref = nir_build_deref_array_imm(b, parent, 0);
479 nir_deref_instr *deref2 = nir_build_deref_array(b, parent,
480 nir_iadd_imm(b, deref->arr.index.ssa, 1));
481 b->cursor = nir_before_instr(&intr->instr);
482 if (intr->intrinsic == nir_intrinsic_load_deref) {
483 nir_def *src1 = nir_load_deref(b, deref);
484 nir_def *src2 = nir_load_deref(b, deref2);
485 nir_def_rewrite_uses(&intr->def, nir_pack_64_2x32_split(b, src1, src2));
486 } else {
487 nir_def *src1 = nir_unpack_64_2x32_split_x(b, intr->src[1].ssa);
488 nir_def *src2 = nir_unpack_64_2x32_split_y(b, intr->src[1].ssa);
489 nir_store_deref(b, deref, src1, 1);
490 nir_store_deref(b, deref, src2, 1);
491 }
492 nir_instr_remove(&intr->instr);
493 }
494 return true;
495 }
496
497 static bool
lower_var_bit_size_types(nir_variable * var,unsigned min_bit_size,unsigned max_bit_size)498 lower_var_bit_size_types(nir_variable *var, unsigned min_bit_size, unsigned max_bit_size)
499 {
500 assert(!glsl_type_is_array_of_arrays(var->type) && !glsl_type_is_struct(var->type));
501 const struct glsl_type *type = glsl_without_array(var->type);
502 assert(glsl_type_is_scalar(type));
503 enum glsl_base_type base_type = glsl_get_base_type(type);
504 if (glsl_base_type_get_bit_size(base_type) < min_bit_size) {
505 switch (min_bit_size) {
506 case 16:
507 switch (base_type) {
508 case GLSL_TYPE_BOOL:
509 base_type = GLSL_TYPE_UINT16;
510 for (unsigned i = 0; i < (var->constant_initializer ? var->constant_initializer->num_elements : 0); ++i)
511 var->constant_initializer->elements[i]->values[0].u16 = var->constant_initializer->elements[i]->values[0].b ? 0xffff : 0;
512 break;
513 case GLSL_TYPE_INT8:
514 base_type = GLSL_TYPE_INT16;
515 for (unsigned i = 0; i < (var->constant_initializer ? var->constant_initializer->num_elements : 0); ++i)
516 var->constant_initializer->elements[i]->values[0].i16 = var->constant_initializer->elements[i]->values[0].i8;
517 break;
518 case GLSL_TYPE_UINT8: base_type = GLSL_TYPE_UINT16; break;
519 default: unreachable("Unexpected base type");
520 }
521 break;
522 case 32:
523 switch (base_type) {
524 case GLSL_TYPE_BOOL:
525 base_type = GLSL_TYPE_UINT;
526 for (unsigned i = 0; i < (var->constant_initializer ? var->constant_initializer->num_elements : 0); ++i)
527 var->constant_initializer->elements[i]->values[0].u32 = var->constant_initializer->elements[i]->values[0].b ? 0xffffffff : 0;
528 break;
529 case GLSL_TYPE_INT8:
530 base_type = GLSL_TYPE_INT;
531 for (unsigned i = 0; i < (var->constant_initializer ? var->constant_initializer->num_elements : 0); ++i)
532 var->constant_initializer->elements[i]->values[0].i32 = var->constant_initializer->elements[i]->values[0].i8;
533 break;
534 case GLSL_TYPE_INT16:
535 base_type = GLSL_TYPE_INT;
536 for (unsigned i = 0; i < (var->constant_initializer ? var->constant_initializer->num_elements : 0); ++i)
537 var->constant_initializer->elements[i]->values[0].i32 = var->constant_initializer->elements[i]->values[0].i16;
538 break;
539 case GLSL_TYPE_FLOAT16:
540 base_type = GLSL_TYPE_FLOAT;
541 for (unsigned i = 0; i < (var->constant_initializer ? var->constant_initializer->num_elements : 0); ++i)
542 var->constant_initializer->elements[i]->values[0].f32 = _mesa_half_to_float(var->constant_initializer->elements[i]->values[0].u16);
543 break;
544 case GLSL_TYPE_UINT8: base_type = GLSL_TYPE_UINT; break;
545 case GLSL_TYPE_UINT16: base_type = GLSL_TYPE_UINT; break;
546 default: unreachable("Unexpected base type");
547 }
548 break;
549 default: unreachable("Unexpected min bit size");
550 }
551 var->type = glsl_type_wrap_in_arrays(glsl_scalar_type(base_type), var->type);
552 return true;
553 }
554 if (glsl_base_type_bit_size(base_type) > max_bit_size) {
555 assert(!glsl_type_is_array_of_arrays(var->type));
556 var->type = glsl_array_type(glsl_scalar_type(GLSL_TYPE_UINT),
557 glsl_type_is_array(var->type) ? glsl_get_length(var->type) * 2 : 2,
558 0);
559 if (var->constant_initializer) {
560 unsigned num_elements = var->constant_initializer->num_elements ?
561 var->constant_initializer->num_elements * 2 : 2;
562 nir_constant **element_arr = ralloc_array(var, nir_constant *, num_elements);
563 nir_constant *elements = rzalloc_array(var, nir_constant, num_elements);
564 for (unsigned i = 0; i < var->constant_initializer->num_elements; ++i) {
565 element_arr[i*2] = &elements[i*2];
566 element_arr[i*2+1] = &elements[i*2+1];
567 const nir_const_value *src = var->constant_initializer->num_elements ?
568 var->constant_initializer->elements[i]->values : var->constant_initializer->values;
569 elements[i*2].values[0].u32 = (uint32_t)src->u64;
570 elements[i*2].is_null_constant = (uint32_t)src->u64 == 0;
571 elements[i*2+1].values[0].u32 = (uint32_t)(src->u64 >> 32);
572 elements[i*2+1].is_null_constant = (uint32_t)(src->u64 >> 32) == 0;
573 }
574 var->constant_initializer->num_elements = num_elements;
575 var->constant_initializer->elements = element_arr;
576 }
577 return true;
578 }
579 return false;
580 }
581
582 bool
dxil_nir_lower_var_bit_size(nir_shader * shader,nir_variable_mode modes,unsigned min_bit_size,unsigned max_bit_size)583 dxil_nir_lower_var_bit_size(nir_shader *shader, nir_variable_mode modes,
584 unsigned min_bit_size, unsigned max_bit_size)
585 {
586 bool progress = false;
587 nir_foreach_variable_with_modes(var, shader, modes & ~nir_var_function_temp)
588 progress |= lower_var_bit_size_types(var, min_bit_size, max_bit_size);
589
590 if (modes & nir_var_function_temp) {
591 nir_foreach_function_impl(impl, shader) {
592 nir_foreach_function_temp_variable(var, impl)
593 progress |= lower_var_bit_size_types(var, min_bit_size, max_bit_size);
594 }
595 }
596
597 if (!progress)
598 return false;
599
600 nir_shader_intrinsics_pass(shader, lower_deref_bit_size,
601 nir_metadata_control_flow |
602 nir_metadata_loop_analysis,
603 NULL);
604 nir_remove_dead_derefs(shader);
605 return true;
606 }
607
608 static bool
remove_oob_array_access(nir_builder * b,nir_intrinsic_instr * intr,void * data)609 remove_oob_array_access(nir_builder *b, nir_intrinsic_instr *intr, void *data)
610 {
611 uint32_t num_derefs = 1;
612
613 switch (intr->intrinsic) {
614 case nir_intrinsic_copy_deref:
615 num_derefs = 2;
616 FALLTHROUGH;
617 case nir_intrinsic_load_deref:
618 case nir_intrinsic_store_deref:
619 case nir_intrinsic_deref_atomic:
620 case nir_intrinsic_deref_atomic_swap:
621 break;
622 default:
623 return false;
624 }
625
626 for (uint32_t i = 0; i < num_derefs; ++i) {
627 if (nir_deref_instr_is_known_out_of_bounds(nir_src_as_deref(intr->src[i]))) {
628 switch (intr->intrinsic) {
629 case nir_intrinsic_load_deref:
630 case nir_intrinsic_deref_atomic:
631 case nir_intrinsic_deref_atomic_swap:
632 b->cursor = nir_before_instr(&intr->instr);
633 nir_def *undef = nir_undef(b, intr->def.num_components, intr->def.bit_size);
634 nir_def_rewrite_uses(&intr->def, undef);
635 break;
636 default:
637 break;
638 }
639 nir_instr_remove(&intr->instr);
640 return true;
641 }
642 }
643
644 return false;
645 }
646
647 bool
dxil_nir_remove_oob_array_accesses(nir_shader * shader)648 dxil_nir_remove_oob_array_accesses(nir_shader *shader)
649 {
650 return nir_shader_intrinsics_pass(shader, remove_oob_array_access,
651 nir_metadata_control_flow |
652 nir_metadata_loop_analysis,
653 NULL);
654 }
655
656 static bool
lower_shared_atomic(nir_builder * b,nir_intrinsic_instr * intr,nir_variable * var)657 lower_shared_atomic(nir_builder *b, nir_intrinsic_instr *intr, nir_variable *var)
658 {
659 b->cursor = nir_before_instr(&intr->instr);
660
661 nir_def *offset =
662 nir_iadd_imm(b, intr->src[0].ssa, nir_intrinsic_base(intr));
663 nir_def *index = nir_ushr_imm(b, offset, 2);
664
665 nir_deref_instr *deref = nir_build_deref_array(b, nir_build_deref_var(b, var), index);
666 nir_def *result;
667 if (intr->intrinsic == nir_intrinsic_shared_atomic_swap)
668 result = nir_deref_atomic_swap(b, 32, &deref->def, intr->src[1].ssa, intr->src[2].ssa,
669 .atomic_op = nir_intrinsic_atomic_op(intr));
670 else
671 result = nir_deref_atomic(b, 32, &deref->def, intr->src[1].ssa,
672 .atomic_op = nir_intrinsic_atomic_op(intr));
673
674 nir_def_replace(&intr->def, result);
675 return true;
676 }
677
678 bool
dxil_nir_lower_loads_stores_to_dxil(nir_shader * nir,const struct dxil_nir_lower_loads_stores_options * options)679 dxil_nir_lower_loads_stores_to_dxil(nir_shader *nir,
680 const struct dxil_nir_lower_loads_stores_options *options)
681 {
682 bool progress = nir_remove_dead_variables(nir, nir_var_function_temp | nir_var_mem_shared, NULL);
683 nir_variable *shared_var = NULL;
684 if (nir->info.shared_size) {
685 shared_var = nir_variable_create(nir, nir_var_mem_shared,
686 glsl_array_type(glsl_uint_type(), DIV_ROUND_UP(nir->info.shared_size, 4), 4),
687 "lowered_shared_mem");
688 }
689
690 unsigned ptr_size = nir->info.cs.ptr_size;
691 if (nir->info.stage == MESA_SHADER_KERNEL) {
692 /* All the derefs created here will be used as GEP indices so force 32-bit */
693 nir->info.cs.ptr_size = 32;
694 }
695 nir_foreach_function_impl(impl, nir) {
696 nir_builder b = nir_builder_create(impl);
697
698 nir_variable *scratch_var = NULL;
699 if (nir->scratch_size) {
700 const struct glsl_type *scratch_type = glsl_array_type(glsl_uint_type(), DIV_ROUND_UP(nir->scratch_size, 4), 4);
701 scratch_var = nir_local_variable_create(impl, scratch_type, "lowered_scratch_mem");
702 }
703
704 nir_foreach_block(block, impl) {
705 nir_foreach_instr_safe(instr, block) {
706 if (instr->type != nir_instr_type_intrinsic)
707 continue;
708 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
709
710 switch (intr->intrinsic) {
711 case nir_intrinsic_load_shared:
712 progress |= lower_32b_offset_load(&b, intr, shared_var);
713 break;
714 case nir_intrinsic_load_scratch:
715 progress |= lower_32b_offset_load(&b, intr, scratch_var);
716 break;
717 case nir_intrinsic_store_shared:
718 progress |= lower_32b_offset_store(&b, intr, shared_var);
719 break;
720 case nir_intrinsic_store_scratch:
721 progress |= lower_32b_offset_store(&b, intr, scratch_var);
722 break;
723 case nir_intrinsic_shared_atomic:
724 case nir_intrinsic_shared_atomic_swap:
725 progress |= lower_shared_atomic(&b, intr, shared_var);
726 break;
727 default:
728 break;
729 }
730 }
731 }
732 }
733 if (nir->info.stage == MESA_SHADER_KERNEL) {
734 nir->info.cs.ptr_size = ptr_size;
735 }
736
737 return progress;
738 }
739
740 static bool
lower_deref_ssbo(nir_builder * b,nir_deref_instr * deref)741 lower_deref_ssbo(nir_builder *b, nir_deref_instr *deref)
742 {
743 assert(nir_deref_mode_is(deref, nir_var_mem_ssbo));
744 assert(deref->deref_type == nir_deref_type_var ||
745 deref->deref_type == nir_deref_type_cast);
746 nir_variable *var = deref->var;
747
748 b->cursor = nir_before_instr(&deref->instr);
749
750 if (deref->deref_type == nir_deref_type_var) {
751 /* We turn all deref_var into deref_cast and build a pointer value based on
752 * the var binding which encodes the UAV id.
753 */
754 nir_def *ptr = nir_imm_int64(b, (uint64_t)var->data.binding << 32);
755 nir_deref_instr *deref_cast =
756 nir_build_deref_cast(b, ptr, nir_var_mem_ssbo, deref->type,
757 glsl_get_explicit_stride(var->type));
758 nir_def_replace(&deref->def, &deref_cast->def);
759
760 deref = deref_cast;
761 return true;
762 }
763 return false;
764 }
765
766 bool
dxil_nir_lower_deref_ssbo(nir_shader * nir)767 dxil_nir_lower_deref_ssbo(nir_shader *nir)
768 {
769 bool progress = false;
770
771 foreach_list_typed(nir_function, func, node, &nir->functions) {
772 if (!func->is_entrypoint)
773 continue;
774 assert(func->impl);
775
776 nir_builder b = nir_builder_create(func->impl);
777
778 nir_foreach_block(block, func->impl) {
779 nir_foreach_instr_safe(instr, block) {
780 if (instr->type != nir_instr_type_deref)
781 continue;
782
783 nir_deref_instr *deref = nir_instr_as_deref(instr);
784
785 if (!nir_deref_mode_is(deref, nir_var_mem_ssbo) ||
786 (deref->deref_type != nir_deref_type_var &&
787 deref->deref_type != nir_deref_type_cast))
788 continue;
789
790 progress |= lower_deref_ssbo(&b, deref);
791 }
792 }
793 }
794
795 return progress;
796 }
797
798 static bool
lower_alu_deref_srcs(nir_builder * b,nir_alu_instr * alu)799 lower_alu_deref_srcs(nir_builder *b, nir_alu_instr *alu)
800 {
801 const nir_op_info *info = &nir_op_infos[alu->op];
802 bool progress = false;
803
804 b->cursor = nir_before_instr(&alu->instr);
805
806 for (unsigned i = 0; i < info->num_inputs; i++) {
807 nir_deref_instr *deref = nir_src_as_deref(alu->src[i].src);
808
809 if (!deref)
810 continue;
811
812 nir_deref_path path;
813 nir_deref_path_init(&path, deref, NULL);
814 nir_deref_instr *root_deref = path.path[0];
815 nir_deref_path_finish(&path);
816
817 if (root_deref->deref_type != nir_deref_type_cast)
818 continue;
819
820 nir_def *ptr =
821 nir_iadd(b, root_deref->parent.ssa,
822 nir_build_deref_offset(b, deref, cl_type_size_align));
823 nir_src_rewrite(&alu->src[i].src, ptr);
824 progress = true;
825 }
826
827 return progress;
828 }
829
830 bool
dxil_nir_opt_alu_deref_srcs(nir_shader * nir)831 dxil_nir_opt_alu_deref_srcs(nir_shader *nir)
832 {
833 bool progress = false;
834
835 foreach_list_typed(nir_function, func, node, &nir->functions) {
836 if (!func->is_entrypoint)
837 continue;
838 assert(func->impl);
839
840 nir_builder b = nir_builder_create(func->impl);
841
842 nir_foreach_block(block, func->impl) {
843 nir_foreach_instr_safe(instr, block) {
844 if (instr->type != nir_instr_type_alu)
845 continue;
846
847 nir_alu_instr *alu = nir_instr_as_alu(instr);
848 progress |= lower_alu_deref_srcs(&b, alu);
849 }
850 }
851 }
852
853 return progress;
854 }
855
856 static void
cast_phi(nir_builder * b,nir_phi_instr * phi,unsigned new_bit_size)857 cast_phi(nir_builder *b, nir_phi_instr *phi, unsigned new_bit_size)
858 {
859 nir_phi_instr *lowered = nir_phi_instr_create(b->shader);
860 int num_components = 0;
861 int old_bit_size = phi->def.bit_size;
862
863 nir_foreach_phi_src(src, phi) {
864 assert(num_components == 0 || num_components == src->src.ssa->num_components);
865 num_components = src->src.ssa->num_components;
866
867 b->cursor = nir_after_instr_and_phis(src->src.ssa->parent_instr);
868
869 nir_def *cast = nir_u2uN(b, src->src.ssa, new_bit_size);
870
871 nir_phi_instr_add_src(lowered, src->pred, cast);
872 }
873
874 nir_def_init(&lowered->instr, &lowered->def, num_components,
875 new_bit_size);
876
877 b->cursor = nir_before_instr(&phi->instr);
878 nir_builder_instr_insert(b, &lowered->instr);
879
880 b->cursor = nir_after_phis(nir_cursor_current_block(b->cursor));
881 nir_def *result = nir_u2uN(b, &lowered->def, old_bit_size);
882
883 nir_def_replace(&phi->def, result);
884 }
885
886 static bool
upcast_phi_impl(nir_function_impl * impl,unsigned min_bit_size)887 upcast_phi_impl(nir_function_impl *impl, unsigned min_bit_size)
888 {
889 nir_builder b = nir_builder_create(impl);
890 bool progress = false;
891
892 nir_foreach_block_reverse(block, impl) {
893 nir_foreach_phi_safe(phi, block) {
894 if (phi->def.bit_size == 1 ||
895 phi->def.bit_size >= min_bit_size)
896 continue;
897
898 cast_phi(&b, phi, min_bit_size);
899 progress = true;
900 }
901 }
902
903 if (progress) {
904 nir_metadata_preserve(impl, nir_metadata_control_flow);
905 } else {
906 nir_metadata_preserve(impl, nir_metadata_all);
907 }
908
909 return progress;
910 }
911
912 bool
dxil_nir_lower_upcast_phis(nir_shader * shader,unsigned min_bit_size)913 dxil_nir_lower_upcast_phis(nir_shader *shader, unsigned min_bit_size)
914 {
915 bool progress = false;
916
917 nir_foreach_function_impl(impl, shader) {
918 progress |= upcast_phi_impl(impl, min_bit_size);
919 }
920
921 return progress;
922 }
923
924 struct dxil_nir_split_clip_cull_distance_params {
925 nir_variable *new_var[2];
926 nir_shader *shader;
927 };
928
929 /* In GLSL and SPIR-V, clip and cull distance are arrays of floats (with a limit of 8).
930 * In DXIL, clip and cull distances are up to 2 float4s combined.
931 * Coming from GLSL, we can request this 2 float4 format, but coming from SPIR-V,
932 * we can't, and have to accept a "compact" array of scalar floats.
933 *
934 * To help emitting a valid input signature for this case, split the variables so that they
935 * match what we need to put in the signature (e.g. { float clip[4]; float clip1; float cull[3]; })
936 */
937 static bool
dxil_nir_split_clip_cull_distance_instr(nir_builder * b,nir_instr * instr,void * cb_data)938 dxil_nir_split_clip_cull_distance_instr(nir_builder *b,
939 nir_instr *instr,
940 void *cb_data)
941 {
942 struct dxil_nir_split_clip_cull_distance_params *params = cb_data;
943
944 if (instr->type != nir_instr_type_deref)
945 return false;
946
947 nir_deref_instr *deref = nir_instr_as_deref(instr);
948 nir_variable *var = nir_deref_instr_get_variable(deref);
949 if (!var ||
950 var->data.location < VARYING_SLOT_CLIP_DIST0 ||
951 var->data.location > VARYING_SLOT_CULL_DIST1 ||
952 !var->data.compact)
953 return false;
954
955 unsigned new_var_idx = var->data.mode == nir_var_shader_in ? 0 : 1;
956 nir_variable *new_var = params->new_var[new_var_idx];
957
958 /* The location should only be inside clip distance, because clip
959 * and cull should've been merged by nir_lower_clip_cull_distance_arrays()
960 */
961 assert(var->data.location == VARYING_SLOT_CLIP_DIST0 ||
962 var->data.location == VARYING_SLOT_CLIP_DIST1);
963
964 /* The deref chain to the clip/cull variables should be simple, just the
965 * var and an array with a constant index, otherwise more lowering/optimization
966 * might be needed before this pass, e.g. copy prop, lower_io_to_temporaries,
967 * split_var_copies, and/or lower_var_copies. In the case of arrayed I/O like
968 * inputs to the tessellation or geometry stages, there might be a second level
969 * of array index.
970 */
971 assert(deref->deref_type == nir_deref_type_var ||
972 deref->deref_type == nir_deref_type_array);
973
974 b->cursor = nir_before_instr(instr);
975 unsigned arrayed_io_length = 0;
976 const struct glsl_type *old_type = var->type;
977 if (nir_is_arrayed_io(var, b->shader->info.stage)) {
978 arrayed_io_length = glsl_array_size(old_type);
979 old_type = glsl_get_array_element(old_type);
980 }
981 if (!new_var) {
982 /* Update lengths for new and old vars */
983 int old_length = glsl_array_size(old_type);
984 int new_length = (old_length + var->data.location_frac) - 4;
985 old_length -= new_length;
986
987 /* The existing variable fits in the float4 */
988 if (new_length <= 0)
989 return false;
990
991 new_var = nir_variable_clone(var, params->shader);
992 nir_shader_add_variable(params->shader, new_var);
993 assert(glsl_get_base_type(glsl_get_array_element(old_type)) == GLSL_TYPE_FLOAT);
994 var->type = glsl_array_type(glsl_float_type(), old_length, 0);
995 new_var->type = glsl_array_type(glsl_float_type(), new_length, 0);
996 if (arrayed_io_length) {
997 var->type = glsl_array_type(var->type, arrayed_io_length, 0);
998 new_var->type = glsl_array_type(new_var->type, arrayed_io_length, 0);
999 }
1000 new_var->data.location++;
1001 new_var->data.location_frac = 0;
1002 params->new_var[new_var_idx] = new_var;
1003 }
1004
1005 /* Update the type for derefs of the old var */
1006 if (deref->deref_type == nir_deref_type_var) {
1007 deref->type = var->type;
1008 return false;
1009 }
1010
1011 if (glsl_type_is_array(deref->type)) {
1012 assert(arrayed_io_length > 0);
1013 deref->type = glsl_get_array_element(var->type);
1014 return false;
1015 }
1016
1017 assert(glsl_get_base_type(deref->type) == GLSL_TYPE_FLOAT);
1018
1019 nir_const_value *index = nir_src_as_const_value(deref->arr.index);
1020 assert(index);
1021
1022 /* Treat this array as a vector starting at the component index in location_frac,
1023 * so if location_frac is 1 and index is 0, then it's accessing the 'y' component
1024 * of the vector. If index + location_frac is >= 4, there's no component there,
1025 * so we need to add a new variable and adjust the index.
1026 */
1027 unsigned total_index = index->u32 + var->data.location_frac;
1028 if (total_index < 4)
1029 return false;
1030
1031 nir_deref_instr *new_var_deref = nir_build_deref_var(b, new_var);
1032 nir_deref_instr *new_intermediate_deref = new_var_deref;
1033 if (arrayed_io_length) {
1034 nir_deref_instr *parent = nir_src_as_deref(deref->parent);
1035 assert(parent->deref_type == nir_deref_type_array);
1036 new_intermediate_deref = nir_build_deref_array(b, new_intermediate_deref, parent->arr.index.ssa);
1037 }
1038 nir_deref_instr *new_array_deref = nir_build_deref_array(b, new_intermediate_deref, nir_imm_int(b, total_index % 4));
1039 nir_def_rewrite_uses(&deref->def, &new_array_deref->def);
1040 return true;
1041 }
1042
1043 bool
dxil_nir_split_clip_cull_distance(nir_shader * shader)1044 dxil_nir_split_clip_cull_distance(nir_shader *shader)
1045 {
1046 struct dxil_nir_split_clip_cull_distance_params params = {
1047 .new_var = { NULL, NULL },
1048 .shader = shader,
1049 };
1050 nir_shader_instructions_pass(shader,
1051 dxil_nir_split_clip_cull_distance_instr,
1052 nir_metadata_control_flow |
1053 nir_metadata_loop_analysis,
1054 ¶ms);
1055 return params.new_var[0] != NULL || params.new_var[1] != NULL;
1056 }
1057
1058 static bool
dxil_nir_lower_double_math_instr(nir_builder * b,nir_instr * instr,UNUSED void * cb_data)1059 dxil_nir_lower_double_math_instr(nir_builder *b,
1060 nir_instr *instr,
1061 UNUSED void *cb_data)
1062 {
1063 if (instr->type == nir_instr_type_intrinsic) {
1064 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
1065 switch (intr->intrinsic) {
1066 case nir_intrinsic_reduce:
1067 case nir_intrinsic_exclusive_scan:
1068 case nir_intrinsic_inclusive_scan:
1069 break;
1070 default:
1071 return false;
1072 }
1073 if (intr->def.bit_size != 64)
1074 return false;
1075 nir_op reduction = nir_intrinsic_reduction_op(intr);
1076 switch (reduction) {
1077 case nir_op_fmul:
1078 case nir_op_fadd:
1079 case nir_op_fmin:
1080 case nir_op_fmax:
1081 break;
1082 default:
1083 return false;
1084 }
1085 b->cursor = nir_before_instr(instr);
1086 nir_src_rewrite(&intr->src[0], nir_pack_double_2x32_dxil(b, nir_unpack_64_2x32(b, intr->src[0].ssa)));
1087 b->cursor = nir_after_instr(instr);
1088 nir_def *result = nir_pack_64_2x32(b, nir_unpack_double_2x32_dxil(b, &intr->def));
1089 nir_def_rewrite_uses_after(&intr->def, result, result->parent_instr);
1090 return true;
1091 }
1092
1093 if (instr->type != nir_instr_type_alu)
1094 return false;
1095
1096 nir_alu_instr *alu = nir_instr_as_alu(instr);
1097
1098 /* TODO: See if we can apply this explicitly to packs/unpacks that are then
1099 * used as a double. As-is, if we had an app explicitly do a 64bit integer op,
1100 * then try to bitcast to double (not expressible in HLSL, but it is in other
1101 * source languages), this would unpack the integer and repack as a double, when
1102 * we probably want to just send the bitcast through to the backend.
1103 */
1104
1105 b->cursor = nir_before_instr(&alu->instr);
1106
1107 bool progress = false;
1108 for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; ++i) {
1109 if (nir_alu_type_get_base_type(nir_op_infos[alu->op].input_types[i]) == nir_type_float &&
1110 alu->src[i].src.ssa->bit_size == 64) {
1111 unsigned num_components = nir_op_infos[alu->op].input_sizes[i];
1112 if (!num_components)
1113 num_components = alu->def.num_components;
1114 nir_def *components[NIR_MAX_VEC_COMPONENTS];
1115 for (unsigned c = 0; c < num_components; ++c) {
1116 nir_def *packed_double = nir_channel(b, alu->src[i].src.ssa, alu->src[i].swizzle[c]);
1117 nir_def *unpacked_double = nir_unpack_64_2x32(b, packed_double);
1118 components[c] = nir_pack_double_2x32_dxil(b, unpacked_double);
1119 alu->src[i].swizzle[c] = c;
1120 }
1121 nir_src_rewrite(&alu->src[i].src,
1122 nir_vec(b, components, num_components));
1123 progress = true;
1124 }
1125 }
1126
1127 if (nir_alu_type_get_base_type(nir_op_infos[alu->op].output_type) == nir_type_float &&
1128 alu->def.bit_size == 64) {
1129 b->cursor = nir_after_instr(&alu->instr);
1130 nir_def *components[NIR_MAX_VEC_COMPONENTS];
1131 for (unsigned c = 0; c < alu->def.num_components; ++c) {
1132 nir_def *packed_double = nir_channel(b, &alu->def, c);
1133 nir_def *unpacked_double = nir_unpack_double_2x32_dxil(b, packed_double);
1134 components[c] = nir_pack_64_2x32(b, unpacked_double);
1135 }
1136 nir_def *repacked_dvec = nir_vec(b, components, alu->def.num_components);
1137 nir_def_rewrite_uses_after(&alu->def, repacked_dvec, repacked_dvec->parent_instr);
1138 progress = true;
1139 }
1140
1141 return progress;
1142 }
1143
1144 bool
dxil_nir_lower_double_math(nir_shader * shader)1145 dxil_nir_lower_double_math(nir_shader *shader)
1146 {
1147 return nir_shader_instructions_pass(shader,
1148 dxil_nir_lower_double_math_instr,
1149 nir_metadata_control_flow |
1150 nir_metadata_loop_analysis,
1151 NULL);
1152 }
1153
1154 typedef struct {
1155 gl_system_value *values;
1156 uint32_t count;
1157 } zero_system_values_state;
1158
1159 static bool
lower_system_value_to_zero_filter(const nir_instr * instr,const void * cb_state)1160 lower_system_value_to_zero_filter(const nir_instr* instr, const void* cb_state)
1161 {
1162 if (instr->type != nir_instr_type_intrinsic) {
1163 return false;
1164 }
1165
1166 nir_intrinsic_instr* intrin = nir_instr_as_intrinsic(instr);
1167
1168 /* All the intrinsics we care about are loads */
1169 if (!nir_intrinsic_infos[intrin->intrinsic].has_dest)
1170 return false;
1171
1172 zero_system_values_state* state = (zero_system_values_state*)cb_state;
1173 for (uint32_t i = 0; i < state->count; ++i) {
1174 gl_system_value value = state->values[i];
1175 nir_intrinsic_op value_op = nir_intrinsic_from_system_value(value);
1176
1177 if (intrin->intrinsic == value_op) {
1178 return true;
1179 } else if (intrin->intrinsic == nir_intrinsic_load_deref) {
1180 nir_deref_instr* deref = nir_src_as_deref(intrin->src[0]);
1181 if (!nir_deref_mode_is(deref, nir_var_system_value))
1182 return false;
1183
1184 nir_variable* var = deref->var;
1185 if (var->data.location == value) {
1186 return true;
1187 }
1188 }
1189 }
1190
1191 return false;
1192 }
1193
1194 static nir_def*
lower_system_value_to_zero_instr(nir_builder * b,nir_instr * instr,void * _state)1195 lower_system_value_to_zero_instr(nir_builder* b, nir_instr* instr, void* _state)
1196 {
1197 return nir_imm_int(b, 0);
1198 }
1199
1200 bool
dxil_nir_lower_system_values_to_zero(nir_shader * shader,gl_system_value * system_values,uint32_t count)1201 dxil_nir_lower_system_values_to_zero(nir_shader* shader,
1202 gl_system_value* system_values,
1203 uint32_t count)
1204 {
1205 zero_system_values_state state = { system_values, count };
1206 return nir_shader_lower_instructions(shader,
1207 lower_system_value_to_zero_filter,
1208 lower_system_value_to_zero_instr,
1209 &state);
1210 }
1211
1212 static void
lower_load_local_group_size(nir_builder * b,nir_intrinsic_instr * intr)1213 lower_load_local_group_size(nir_builder *b, nir_intrinsic_instr *intr)
1214 {
1215 b->cursor = nir_after_instr(&intr->instr);
1216
1217 nir_const_value v[3] = {
1218 nir_const_value_for_int(b->shader->info.workgroup_size[0], 32),
1219 nir_const_value_for_int(b->shader->info.workgroup_size[1], 32),
1220 nir_const_value_for_int(b->shader->info.workgroup_size[2], 32)
1221 };
1222 nir_def *size = nir_build_imm(b, 3, 32, v);
1223 nir_def_replace(&intr->def, size);
1224 }
1225
1226 static bool
lower_system_values_impl(nir_builder * b,nir_intrinsic_instr * intr,void * _state)1227 lower_system_values_impl(nir_builder *b, nir_intrinsic_instr *intr,
1228 void *_state)
1229 {
1230 switch (intr->intrinsic) {
1231 case nir_intrinsic_load_workgroup_size:
1232 lower_load_local_group_size(b, intr);
1233 return true;
1234 default:
1235 return false;
1236 }
1237 }
1238
1239 bool
dxil_nir_lower_system_values(nir_shader * shader)1240 dxil_nir_lower_system_values(nir_shader *shader)
1241 {
1242 return nir_shader_intrinsics_pass(shader, lower_system_values_impl,
1243 nir_metadata_control_flow | nir_metadata_loop_analysis,
1244 NULL);
1245 }
1246
1247 static const struct glsl_type *
get_bare_samplers_for_type(const struct glsl_type * type,bool is_shadow)1248 get_bare_samplers_for_type(const struct glsl_type *type, bool is_shadow)
1249 {
1250 const struct glsl_type *base_sampler_type =
1251 is_shadow ?
1252 glsl_bare_shadow_sampler_type() : glsl_bare_sampler_type();
1253 return glsl_type_wrap_in_arrays(base_sampler_type, type);
1254 }
1255
1256 static const struct glsl_type *
get_textures_for_sampler_type(const struct glsl_type * type)1257 get_textures_for_sampler_type(const struct glsl_type *type)
1258 {
1259 return glsl_type_wrap_in_arrays(
1260 glsl_sampler_type_to_texture(
1261 glsl_without_array(type)), type);
1262 }
1263
1264 static bool
redirect_sampler_derefs(struct nir_builder * b,nir_instr * instr,void * data)1265 redirect_sampler_derefs(struct nir_builder *b, nir_instr *instr, void *data)
1266 {
1267 if (instr->type != nir_instr_type_tex)
1268 return false;
1269
1270 nir_tex_instr *tex = nir_instr_as_tex(instr);
1271
1272 int sampler_idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
1273 if (sampler_idx == -1) {
1274 /* No sampler deref - does this instruction even need a sampler? If not,
1275 * sampler_index doesn't necessarily point to a sampler, so early-out.
1276 */
1277 if (!nir_tex_instr_need_sampler(tex))
1278 return false;
1279
1280 /* No derefs but needs a sampler, must be using indices */
1281 nir_variable *bare_sampler = _mesa_hash_table_u64_search(data, tex->sampler_index);
1282
1283 /* Already have a bare sampler here */
1284 if (bare_sampler)
1285 return false;
1286
1287 nir_variable *old_sampler = NULL;
1288 nir_foreach_variable_with_modes(var, b->shader, nir_var_uniform) {
1289 if (var->data.binding <= tex->sampler_index &&
1290 var->data.binding + glsl_type_get_sampler_count(var->type) >
1291 tex->sampler_index) {
1292
1293 /* Already have a bare sampler for this binding and it is of the
1294 * correct type, add it to the table */
1295 if (glsl_type_is_bare_sampler(glsl_without_array(var->type)) &&
1296 glsl_sampler_type_is_shadow(glsl_without_array(var->type)) ==
1297 tex->is_shadow) {
1298 _mesa_hash_table_u64_insert(data, tex->sampler_index, var);
1299 return false;
1300 }
1301
1302 old_sampler = var;
1303 }
1304 }
1305
1306 assert(old_sampler);
1307
1308 /* Clone the original sampler to a bare sampler of the correct type */
1309 bare_sampler = nir_variable_clone(old_sampler, b->shader);
1310 nir_shader_add_variable(b->shader, bare_sampler);
1311
1312 bare_sampler->type =
1313 get_bare_samplers_for_type(old_sampler->type, tex->is_shadow);
1314 _mesa_hash_table_u64_insert(data, tex->sampler_index, bare_sampler);
1315 return true;
1316 }
1317
1318 /* Using derefs, means we have to rewrite the deref chain in addition to cloning */
1319 nir_deref_instr *final_deref = nir_src_as_deref(tex->src[sampler_idx].src);
1320 nir_deref_path path;
1321 nir_deref_path_init(&path, final_deref, NULL);
1322
1323 nir_deref_instr *old_tail = path.path[0];
1324 assert(old_tail->deref_type == nir_deref_type_var);
1325 nir_variable *old_var = old_tail->var;
1326 if (glsl_type_is_bare_sampler(glsl_without_array(old_var->type)) &&
1327 glsl_sampler_type_is_shadow(glsl_without_array(old_var->type)) ==
1328 tex->is_shadow) {
1329 nir_deref_path_finish(&path);
1330 return false;
1331 }
1332
1333 uint64_t var_key = ((uint64_t)old_var->data.descriptor_set << 32) |
1334 old_var->data.binding;
1335 nir_variable *new_var = _mesa_hash_table_u64_search(data, var_key);
1336 if (!new_var) {
1337 new_var = nir_variable_clone(old_var, b->shader);
1338 nir_shader_add_variable(b->shader, new_var);
1339 new_var->type =
1340 get_bare_samplers_for_type(old_var->type, tex->is_shadow);
1341 _mesa_hash_table_u64_insert(data, var_key, new_var);
1342 }
1343
1344 b->cursor = nir_after_instr(&old_tail->instr);
1345 nir_deref_instr *new_tail = nir_build_deref_var(b, new_var);
1346
1347 for (unsigned i = 1; path.path[i]; ++i) {
1348 b->cursor = nir_after_instr(&path.path[i]->instr);
1349 new_tail = nir_build_deref_follower(b, new_tail, path.path[i]);
1350 }
1351
1352 nir_deref_path_finish(&path);
1353 nir_src_rewrite(&tex->src[sampler_idx].src, &new_tail->def);
1354 return true;
1355 }
1356
1357 static bool
redirect_texture_derefs(struct nir_builder * b,nir_instr * instr,void * data)1358 redirect_texture_derefs(struct nir_builder *b, nir_instr *instr, void *data)
1359 {
1360 if (instr->type != nir_instr_type_tex)
1361 return false;
1362
1363 nir_tex_instr *tex = nir_instr_as_tex(instr);
1364
1365 int texture_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
1366 if (texture_idx == -1) {
1367 /* No derefs, must be using indices */
1368 nir_variable *bare_sampler = _mesa_hash_table_u64_search(data, tex->texture_index);
1369
1370 /* Already have a texture here */
1371 if (bare_sampler)
1372 return false;
1373
1374 nir_variable *typed_sampler = NULL;
1375 nir_foreach_variable_with_modes(var, b->shader, nir_var_uniform) {
1376 if (var->data.binding <= tex->texture_index &&
1377 var->data.binding + glsl_type_get_texture_count(var->type) > tex->texture_index) {
1378 /* Already have a texture for this binding, add it to the table */
1379 _mesa_hash_table_u64_insert(data, tex->texture_index, var);
1380 return false;
1381 }
1382
1383 if (var->data.binding <= tex->texture_index &&
1384 var->data.binding + glsl_type_get_sampler_count(var->type) > tex->texture_index &&
1385 !glsl_type_is_bare_sampler(glsl_without_array(var->type))) {
1386 typed_sampler = var;
1387 }
1388 }
1389
1390 /* Clone the typed sampler to a texture and we're done */
1391 assert(typed_sampler);
1392 bare_sampler = nir_variable_clone(typed_sampler, b->shader);
1393 bare_sampler->type = get_textures_for_sampler_type(typed_sampler->type);
1394 nir_shader_add_variable(b->shader, bare_sampler);
1395 _mesa_hash_table_u64_insert(data, tex->texture_index, bare_sampler);
1396 return true;
1397 }
1398
1399 /* Using derefs, means we have to rewrite the deref chain in addition to cloning */
1400 nir_deref_instr *final_deref = nir_src_as_deref(tex->src[texture_idx].src);
1401 nir_deref_path path;
1402 nir_deref_path_init(&path, final_deref, NULL);
1403
1404 nir_deref_instr *old_tail = path.path[0];
1405 assert(old_tail->deref_type == nir_deref_type_var);
1406 nir_variable *old_var = old_tail->var;
1407 if (glsl_type_is_texture(glsl_without_array(old_var->type)) ||
1408 glsl_type_is_image(glsl_without_array(old_var->type))) {
1409 nir_deref_path_finish(&path);
1410 return false;
1411 }
1412
1413 uint64_t var_key = ((uint64_t)old_var->data.descriptor_set << 32) |
1414 old_var->data.binding;
1415 nir_variable *new_var = _mesa_hash_table_u64_search(data, var_key);
1416 if (!new_var) {
1417 new_var = nir_variable_clone(old_var, b->shader);
1418 new_var->type = get_textures_for_sampler_type(old_var->type);
1419 nir_shader_add_variable(b->shader, new_var);
1420 _mesa_hash_table_u64_insert(data, var_key, new_var);
1421 }
1422
1423 b->cursor = nir_after_instr(&old_tail->instr);
1424 nir_deref_instr *new_tail = nir_build_deref_var(b, new_var);
1425
1426 for (unsigned i = 1; path.path[i]; ++i) {
1427 b->cursor = nir_after_instr(&path.path[i]->instr);
1428 new_tail = nir_build_deref_follower(b, new_tail, path.path[i]);
1429 }
1430
1431 nir_deref_path_finish(&path);
1432 nir_src_rewrite(&tex->src[texture_idx].src, &new_tail->def);
1433
1434 return true;
1435 }
1436
1437 bool
dxil_nir_split_typed_samplers(nir_shader * nir)1438 dxil_nir_split_typed_samplers(nir_shader *nir)
1439 {
1440 struct hash_table_u64 *hash_table = _mesa_hash_table_u64_create(NULL);
1441
1442 bool progress = nir_shader_instructions_pass(nir, redirect_sampler_derefs,
1443 nir_metadata_control_flow | nir_metadata_loop_analysis, hash_table);
1444
1445 _mesa_hash_table_u64_clear(hash_table);
1446
1447 progress |= nir_shader_instructions_pass(nir, redirect_texture_derefs,
1448 nir_metadata_control_flow | nir_metadata_loop_analysis, hash_table);
1449
1450 _mesa_hash_table_u64_destroy(hash_table);
1451 return progress;
1452 }
1453
1454
1455 static bool
lower_sysval_to_load_input_impl(nir_builder * b,nir_intrinsic_instr * intr,void * data)1456 lower_sysval_to_load_input_impl(nir_builder *b, nir_intrinsic_instr *intr,
1457 void *data)
1458 {
1459 gl_system_value sysval = SYSTEM_VALUE_MAX;
1460 switch (intr->intrinsic) {
1461 case nir_intrinsic_load_instance_id:
1462 sysval = SYSTEM_VALUE_INSTANCE_ID;
1463 break;
1464 case nir_intrinsic_load_vertex_id_zero_base:
1465 sysval = SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
1466 break;
1467 default:
1468 return false;
1469 }
1470
1471 nir_variable **sysval_vars = (nir_variable **)data;
1472 nir_variable *var = sysval_vars[sysval];
1473 assert(var);
1474
1475 const nir_alu_type dest_type = nir_get_nir_type_for_glsl_type(var->type);
1476 const unsigned bit_size = intr->def.bit_size;
1477
1478 b->cursor = nir_before_instr(&intr->instr);
1479 nir_def *result = nir_load_input(b, intr->def.num_components, bit_size, nir_imm_int(b, 0),
1480 .base = var->data.driver_location, .dest_type = dest_type);
1481
1482 nir_def_rewrite_uses(&intr->def, result);
1483 return true;
1484 }
1485
1486 bool
dxil_nir_lower_sysval_to_load_input(nir_shader * s,nir_variable ** sysval_vars)1487 dxil_nir_lower_sysval_to_load_input(nir_shader *s, nir_variable **sysval_vars)
1488 {
1489 return nir_shader_intrinsics_pass(s, lower_sysval_to_load_input_impl,
1490 nir_metadata_control_flow,
1491 sysval_vars);
1492 }
1493
1494 /* Comparison function to sort io values so that first come normal varyings,
1495 * then system values, and then system generated values.
1496 */
1497 static int
variable_location_cmp(const nir_variable * a,const nir_variable * b)1498 variable_location_cmp(const nir_variable* a, const nir_variable* b)
1499 {
1500 // Sort by stream, driver_location, location, location_frac, then index
1501 // If all else is equal, sort full vectors before partial ones
1502 unsigned a_location = a->data.location;
1503 if (a_location >= VARYING_SLOT_PATCH0)
1504 a_location -= VARYING_SLOT_PATCH0;
1505 unsigned b_location = b->data.location;
1506 if (b_location >= VARYING_SLOT_PATCH0)
1507 b_location -= VARYING_SLOT_PATCH0;
1508 unsigned a_stream = a->data.stream & ~NIR_STREAM_PACKED;
1509 unsigned b_stream = b->data.stream & ~NIR_STREAM_PACKED;
1510 return a_stream != b_stream ?
1511 a_stream - b_stream :
1512 a->data.driver_location != b->data.driver_location ?
1513 a->data.driver_location - b->data.driver_location :
1514 a_location != b_location ?
1515 a_location - b_location :
1516 a->data.location_frac != b->data.location_frac ?
1517 a->data.location_frac - b->data.location_frac :
1518 a->data.index != b->data.index ?
1519 a->data.index - b->data.index :
1520 glsl_get_component_slots(b->type) - glsl_get_component_slots(a->type);
1521 }
1522
1523 /* Order varyings according to driver location */
1524 void
dxil_sort_by_driver_location(nir_shader * s,nir_variable_mode modes)1525 dxil_sort_by_driver_location(nir_shader* s, nir_variable_mode modes)
1526 {
1527 nir_sort_variables_with_modes(s, variable_location_cmp, modes);
1528 }
1529
1530 /* Sort PS outputs so that color outputs come first */
1531 void
dxil_sort_ps_outputs(nir_shader * s)1532 dxil_sort_ps_outputs(nir_shader* s)
1533 {
1534 nir_foreach_variable_with_modes_safe(var, s, nir_var_shader_out) {
1535 /* We use the driver_location here to avoid introducing a new
1536 * struct or member variable here. The true, updated driver location
1537 * will be written below, after sorting */
1538 switch (var->data.location) {
1539 case FRAG_RESULT_DEPTH:
1540 var->data.driver_location = 1;
1541 break;
1542 case FRAG_RESULT_STENCIL:
1543 var->data.driver_location = 2;
1544 break;
1545 case FRAG_RESULT_SAMPLE_MASK:
1546 var->data.driver_location = 3;
1547 break;
1548 default:
1549 var->data.driver_location = 0;
1550 }
1551 }
1552
1553 nir_sort_variables_with_modes(s, variable_location_cmp,
1554 nir_var_shader_out);
1555
1556 unsigned driver_loc = 0;
1557 nir_foreach_variable_with_modes(var, s, nir_var_shader_out) {
1558 /* Fractional vars should use the same driver_location as the base. These will
1559 * get fully merged during signature processing.
1560 */
1561 var->data.driver_location = var->data.location_frac ? driver_loc - 1 : driver_loc++;
1562 }
1563 }
1564
1565 enum dxil_sysvalue_type {
1566 DXIL_NO_SYSVALUE = 0,
1567 DXIL_USED_SYSVALUE,
1568 DXIL_UNUSED_NO_SYSVALUE,
1569 DXIL_SYSVALUE,
1570 DXIL_GENERATED_SYSVALUE,
1571 };
1572
1573 static enum dxil_sysvalue_type
nir_var_to_dxil_sysvalue_type(nir_variable * var,uint64_t other_stage_mask,const BITSET_WORD * other_stage_frac_mask)1574 nir_var_to_dxil_sysvalue_type(nir_variable *var, uint64_t other_stage_mask,
1575 const BITSET_WORD *other_stage_frac_mask)
1576 {
1577 switch (var->data.location) {
1578 case VARYING_SLOT_FACE:
1579 return DXIL_GENERATED_SYSVALUE;
1580 case VARYING_SLOT_POS:
1581 case VARYING_SLOT_PRIMITIVE_ID:
1582 case VARYING_SLOT_CLIP_DIST0:
1583 case VARYING_SLOT_CLIP_DIST1:
1584 case VARYING_SLOT_PSIZ:
1585 case VARYING_SLOT_TESS_LEVEL_INNER:
1586 case VARYING_SLOT_TESS_LEVEL_OUTER:
1587 case VARYING_SLOT_VIEWPORT:
1588 case VARYING_SLOT_LAYER:
1589 case VARYING_SLOT_VIEW_INDEX:
1590 if (!((1ull << var->data.location) & other_stage_mask))
1591 return DXIL_SYSVALUE;
1592 return DXIL_USED_SYSVALUE;
1593 default:
1594 if (var->data.location < VARYING_SLOT_PATCH0 &&
1595 !((1ull << var->data.location) & other_stage_mask))
1596 return DXIL_UNUSED_NO_SYSVALUE;
1597 if (var->data.location_frac && other_stage_frac_mask &&
1598 var->data.location >= VARYING_SLOT_VAR0 &&
1599 !BITSET_TEST(other_stage_frac_mask, ((var->data.location - VARYING_SLOT_VAR0) * 4 + var->data.location_frac)))
1600 return DXIL_UNUSED_NO_SYSVALUE;
1601 return DXIL_NO_SYSVALUE;
1602 }
1603 }
1604
1605 /* Order between stage values so that normal varyings come first,
1606 * then sysvalues and then system generated values.
1607 */
1608 void
dxil_reassign_driver_locations(nir_shader * s,nir_variable_mode modes,uint64_t other_stage_mask,const BITSET_WORD * other_stage_frac_mask)1609 dxil_reassign_driver_locations(nir_shader* s, nir_variable_mode modes,
1610 uint64_t other_stage_mask, const BITSET_WORD *other_stage_frac_mask)
1611 {
1612 nir_foreach_variable_with_modes_safe(var, s, modes) {
1613 /* We use the driver_location here to avoid introducing a new
1614 * struct or member variable here. The true, updated driver location
1615 * will be written below, after sorting */
1616 var->data.driver_location = nir_var_to_dxil_sysvalue_type(var, other_stage_mask, other_stage_frac_mask);
1617 }
1618
1619 nir_sort_variables_with_modes(s, variable_location_cmp, modes);
1620
1621 unsigned driver_loc = 0, driver_patch_loc = 0;
1622 nir_foreach_variable_with_modes(var, s, modes) {
1623 /* Overlap patches with non-patch */
1624 var->data.driver_location = var->data.patch ?
1625 driver_patch_loc++ : driver_loc++;
1626 }
1627 }
1628
1629 static bool
lower_ubo_array_one_to_static(struct nir_builder * b,nir_intrinsic_instr * intrin,void * cb_data)1630 lower_ubo_array_one_to_static(struct nir_builder *b,
1631 nir_intrinsic_instr *intrin,
1632 void *cb_data)
1633 {
1634 if (intrin->intrinsic != nir_intrinsic_load_vulkan_descriptor)
1635 return false;
1636
1637 nir_variable *var =
1638 nir_get_binding_variable(b->shader, nir_chase_binding(intrin->src[0]));
1639
1640 if (!var)
1641 return false;
1642
1643 if (!glsl_type_is_array(var->type) || glsl_array_size(var->type) != 1)
1644 return false;
1645
1646 nir_intrinsic_instr *index = nir_src_as_intrinsic(intrin->src[0]);
1647 /* We currently do not support reindex */
1648 assert(index && index->intrinsic == nir_intrinsic_vulkan_resource_index);
1649
1650 if (nir_src_is_const(index->src[0]) && nir_src_as_uint(index->src[0]) == 0)
1651 return false;
1652
1653 if (nir_intrinsic_desc_type(index) != VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER)
1654 return false;
1655
1656 b->cursor = nir_instr_remove(&index->instr);
1657
1658 // Indexing out of bounds on array of UBOs is considered undefined
1659 // behavior. Therefore, we just hardcode all the index to 0.
1660 uint8_t bit_size = index->def.bit_size;
1661 nir_def *zero = nir_imm_intN_t(b, 0, bit_size);
1662 nir_def *dest =
1663 nir_vulkan_resource_index(b, index->num_components, bit_size, zero,
1664 .desc_set = nir_intrinsic_desc_set(index),
1665 .binding = nir_intrinsic_binding(index),
1666 .desc_type = nir_intrinsic_desc_type(index));
1667
1668 nir_def_rewrite_uses(&index->def, dest);
1669
1670 return true;
1671 }
1672
1673 bool
dxil_nir_lower_ubo_array_one_to_static(nir_shader * s)1674 dxil_nir_lower_ubo_array_one_to_static(nir_shader *s)
1675 {
1676 bool progress = nir_shader_intrinsics_pass(s,
1677 lower_ubo_array_one_to_static,
1678 nir_metadata_none, NULL);
1679
1680 return progress;
1681 }
1682
1683 static bool
is_fquantize2f16(const nir_instr * instr,const void * data)1684 is_fquantize2f16(const nir_instr *instr, const void *data)
1685 {
1686 if (instr->type != nir_instr_type_alu)
1687 return false;
1688
1689 nir_alu_instr *alu = nir_instr_as_alu(instr);
1690 return alu->op == nir_op_fquantize2f16;
1691 }
1692
1693 static nir_def *
lower_fquantize2f16(struct nir_builder * b,nir_instr * instr,void * data)1694 lower_fquantize2f16(struct nir_builder *b, nir_instr *instr, void *data)
1695 {
1696 /*
1697 * SpvOpQuantizeToF16 documentation says:
1698 *
1699 * "
1700 * If Value is an infinity, the result is the same infinity.
1701 * If Value is a NaN, the result is a NaN, but not necessarily the same NaN.
1702 * If Value is positive with a magnitude too large to represent as a 16-bit
1703 * floating-point value, the result is positive infinity. If Value is negative
1704 * with a magnitude too large to represent as a 16-bit floating-point value,
1705 * the result is negative infinity. If the magnitude of Value is too small to
1706 * represent as a normalized 16-bit floating-point value, the result may be
1707 * either +0 or -0.
1708 * "
1709 *
1710 * which we turn into:
1711 *
1712 * if (val < MIN_FLOAT16)
1713 * return -INFINITY;
1714 * else if (val > MAX_FLOAT16)
1715 * return -INFINITY;
1716 * else if (fabs(val) < SMALLEST_NORMALIZED_FLOAT16 && sign(val) != 0)
1717 * return -0.0f;
1718 * else if (fabs(val) < SMALLEST_NORMALIZED_FLOAT16 && sign(val) == 0)
1719 * return +0.0f;
1720 * else
1721 * return round(val);
1722 */
1723 nir_alu_instr *alu = nir_instr_as_alu(instr);
1724 nir_def *src =
1725 alu->src[0].src.ssa;
1726
1727 nir_def *neg_inf_cond =
1728 nir_flt_imm(b, src, -65504.0f);
1729 nir_def *pos_inf_cond =
1730 nir_fgt_imm(b, src, 65504.0f);
1731 nir_def *zero_cond =
1732 nir_flt_imm(b, nir_fabs(b, src), ldexpf(1.0, -14));
1733 nir_def *zero = nir_iand_imm(b, src, 1 << 31);
1734 nir_def *round = nir_iand_imm(b, src, ~BITFIELD_MASK(13));
1735
1736 nir_def *res =
1737 nir_bcsel(b, neg_inf_cond, nir_imm_float(b, -INFINITY), round);
1738 res = nir_bcsel(b, pos_inf_cond, nir_imm_float(b, INFINITY), res);
1739 res = nir_bcsel(b, zero_cond, zero, res);
1740 return res;
1741 }
1742
1743 bool
dxil_nir_lower_fquantize2f16(nir_shader * s)1744 dxil_nir_lower_fquantize2f16(nir_shader *s)
1745 {
1746 return nir_shader_lower_instructions(s, is_fquantize2f16, lower_fquantize2f16, NULL);
1747 }
1748
1749 static bool
fix_io_uint_deref_types(struct nir_builder * builder,nir_instr * instr,void * data)1750 fix_io_uint_deref_types(struct nir_builder *builder, nir_instr *instr, void *data)
1751 {
1752 if (instr->type != nir_instr_type_deref)
1753 return false;
1754
1755 nir_deref_instr *deref = nir_instr_as_deref(instr);
1756 nir_variable *var = nir_deref_instr_get_variable(deref);
1757
1758 if (var == data) {
1759 deref->type = glsl_type_wrap_in_arrays(glsl_uint_type(), deref->type);
1760 return true;
1761 }
1762
1763 return false;
1764 }
1765
1766 static bool
fix_io_uint_type(nir_shader * s,nir_variable_mode modes,int slot)1767 fix_io_uint_type(nir_shader *s, nir_variable_mode modes, int slot)
1768 {
1769 nir_variable *fixed_var = NULL;
1770 nir_foreach_variable_with_modes(var, s, modes) {
1771 if (var->data.location == slot) {
1772 const struct glsl_type *plain_type = glsl_without_array(var->type);
1773 if (plain_type == glsl_uint_type())
1774 return false;
1775
1776 assert(plain_type == glsl_int_type());
1777 var->type = glsl_type_wrap_in_arrays(glsl_uint_type(), var->type);
1778 fixed_var = var;
1779 break;
1780 }
1781 }
1782
1783 assert(fixed_var);
1784
1785 return nir_shader_instructions_pass(s, fix_io_uint_deref_types,
1786 nir_metadata_all, fixed_var);
1787 }
1788
1789 bool
dxil_nir_fix_io_uint_type(nir_shader * s,uint64_t in_mask,uint64_t out_mask)1790 dxil_nir_fix_io_uint_type(nir_shader *s, uint64_t in_mask, uint64_t out_mask)
1791 {
1792 if (!(s->info.outputs_written & out_mask) &&
1793 !(s->info.inputs_read & in_mask))
1794 return false;
1795
1796 bool progress = false;
1797
1798 while (in_mask) {
1799 int slot = u_bit_scan64(&in_mask);
1800 progress |= (s->info.inputs_read & (1ull << slot)) &&
1801 fix_io_uint_type(s, nir_var_shader_in, slot);
1802 }
1803
1804 while (out_mask) {
1805 int slot = u_bit_scan64(&out_mask);
1806 progress |= (s->info.outputs_written & (1ull << slot)) &&
1807 fix_io_uint_type(s, nir_var_shader_out, slot);
1808 }
1809
1810 return progress;
1811 }
1812
1813 static bool
lower_kill(struct nir_builder * builder,nir_intrinsic_instr * intr,void * _cb_data)1814 lower_kill(struct nir_builder *builder, nir_intrinsic_instr *intr,
1815 void *_cb_data)
1816 {
1817 if (intr->intrinsic != nir_intrinsic_terminate &&
1818 intr->intrinsic != nir_intrinsic_terminate_if)
1819 return false;
1820
1821 builder->cursor = nir_instr_remove(&intr->instr);
1822 nir_def *condition;
1823
1824 if (intr->intrinsic == nir_intrinsic_terminate) {
1825 nir_demote(builder);
1826 condition = nir_imm_true(builder);
1827 } else {
1828 nir_demote_if(builder, intr->src[0].ssa);
1829 condition = intr->src[0].ssa;
1830 }
1831
1832 /* Create a new block by branching on the discard condition so that this return
1833 * is definitely the last instruction in its own block */
1834 nir_if *nif = nir_push_if(builder, condition);
1835 nir_jump(builder, nir_jump_return);
1836 nir_pop_if(builder, nif);
1837
1838 return true;
1839 }
1840
1841 bool
dxil_nir_lower_discard_and_terminate(nir_shader * s)1842 dxil_nir_lower_discard_and_terminate(nir_shader *s)
1843 {
1844 if (s->info.stage != MESA_SHADER_FRAGMENT)
1845 return false;
1846
1847 // This pass only works if all functions have been inlined
1848 assert(exec_list_length(&s->functions) == 1);
1849 return nir_shader_intrinsics_pass(s, lower_kill, nir_metadata_none, NULL);
1850 }
1851
1852 static bool
update_writes(struct nir_builder * b,nir_intrinsic_instr * intr,void * _state)1853 update_writes(struct nir_builder *b, nir_intrinsic_instr *intr, void *_state)
1854 {
1855 if (intr->intrinsic != nir_intrinsic_store_output)
1856 return false;
1857
1858 nir_io_semantics io = nir_intrinsic_io_semantics(intr);
1859 if (io.location != VARYING_SLOT_POS)
1860 return false;
1861
1862 nir_def *src = intr->src[0].ssa;
1863 unsigned write_mask = nir_intrinsic_write_mask(intr);
1864 if (src->num_components == 4 && write_mask == 0xf)
1865 return false;
1866
1867 b->cursor = nir_before_instr(&intr->instr);
1868 unsigned first_comp = nir_intrinsic_component(intr);
1869 nir_def *channels[4] = { NULL, NULL, NULL, NULL };
1870 assert(first_comp + src->num_components <= ARRAY_SIZE(channels));
1871 for (unsigned i = 0; i < src->num_components; ++i)
1872 if (write_mask & (1 << i))
1873 channels[i + first_comp] = nir_channel(b, src, i);
1874 for (unsigned i = 0; i < 4; ++i)
1875 if (!channels[i])
1876 channels[i] = nir_imm_intN_t(b, 0, src->bit_size);
1877
1878 intr->num_components = 4;
1879 nir_src_rewrite(&intr->src[0], nir_vec(b, channels, 4));
1880 nir_intrinsic_set_component(intr, 0);
1881 nir_intrinsic_set_write_mask(intr, 0xf);
1882 return true;
1883 }
1884
1885 bool
dxil_nir_ensure_position_writes(nir_shader * s)1886 dxil_nir_ensure_position_writes(nir_shader *s)
1887 {
1888 if (s->info.stage != MESA_SHADER_VERTEX &&
1889 s->info.stage != MESA_SHADER_GEOMETRY &&
1890 s->info.stage != MESA_SHADER_TESS_EVAL)
1891 return false;
1892 if ((s->info.outputs_written & VARYING_BIT_POS) == 0)
1893 return false;
1894
1895 return nir_shader_intrinsics_pass(s, update_writes,
1896 nir_metadata_control_flow,
1897 NULL);
1898 }
1899
1900 static bool
is_sample_pos(const nir_instr * instr,const void * _data)1901 is_sample_pos(const nir_instr *instr, const void *_data)
1902 {
1903 if (instr->type != nir_instr_type_intrinsic)
1904 return false;
1905 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
1906 return intr->intrinsic == nir_intrinsic_load_sample_pos;
1907 }
1908
1909 static nir_def *
lower_sample_pos(nir_builder * b,nir_instr * instr,void * _data)1910 lower_sample_pos(nir_builder *b, nir_instr *instr, void *_data)
1911 {
1912 return nir_load_sample_pos_from_id(b, 32, nir_load_sample_id(b));
1913 }
1914
1915 bool
dxil_nir_lower_sample_pos(nir_shader * s)1916 dxil_nir_lower_sample_pos(nir_shader *s)
1917 {
1918 return nir_shader_lower_instructions(s, is_sample_pos, lower_sample_pos, NULL);
1919 }
1920
1921 static bool
lower_subgroup_id(nir_builder * b,nir_intrinsic_instr * intr,void * data)1922 lower_subgroup_id(nir_builder *b, nir_intrinsic_instr *intr, void *data)
1923 {
1924 if (intr->intrinsic != nir_intrinsic_load_subgroup_id)
1925 return false;
1926
1927 b->cursor = nir_before_impl(b->impl);
1928 if (b->shader->info.stage == MESA_SHADER_COMPUTE &&
1929 b->shader->info.workgroup_size[1] == 1 &&
1930 b->shader->info.workgroup_size[2] == 1) {
1931 /* When using Nx1x1 groups, use a simple stable algorithm
1932 * which is almost guaranteed to be correct. */
1933 nir_def *subgroup_id = nir_udiv(b, nir_load_local_invocation_index(b), nir_load_subgroup_size(b));
1934 nir_def_rewrite_uses(&intr->def, subgroup_id);
1935 return true;
1936 }
1937
1938 nir_def **subgroup_id = (nir_def **)data;
1939 if (*subgroup_id == NULL) {
1940 nir_variable *subgroup_id_counter = nir_variable_create(b->shader, nir_var_mem_shared, glsl_uint_type(), "dxil_SubgroupID_counter");
1941 nir_variable *subgroup_id_local = nir_local_variable_create(b->impl, glsl_uint_type(), "dxil_SubgroupID_local");
1942 nir_store_var(b, subgroup_id_local, nir_imm_int(b, 0), 1);
1943
1944 nir_deref_instr *counter_deref = nir_build_deref_var(b, subgroup_id_counter);
1945 nir_def *tid = nir_load_local_invocation_index(b);
1946 nir_if *nif = nir_push_if(b, nir_ieq_imm(b, tid, 0));
1947 nir_store_deref(b, counter_deref, nir_imm_int(b, 0), 1);
1948 nir_pop_if(b, nif);
1949
1950 nir_barrier(b,
1951 .execution_scope = SCOPE_WORKGROUP,
1952 .memory_scope = SCOPE_WORKGROUP,
1953 .memory_semantics = NIR_MEMORY_ACQ_REL,
1954 .memory_modes = nir_var_mem_shared);
1955
1956 nif = nir_push_if(b, nir_elect(b, 1));
1957 nir_def *subgroup_id_first_thread = nir_deref_atomic(b, 32, &counter_deref->def, nir_imm_int(b, 1),
1958 .atomic_op = nir_atomic_op_iadd);
1959 nir_store_var(b, subgroup_id_local, subgroup_id_first_thread, 1);
1960 nir_pop_if(b, nif);
1961
1962 nir_def *subgroup_id_loaded = nir_load_var(b, subgroup_id_local);
1963 *subgroup_id = nir_read_first_invocation(b, subgroup_id_loaded);
1964 }
1965 nir_def_rewrite_uses(&intr->def, *subgroup_id);
1966 return true;
1967 }
1968
1969 bool
dxil_nir_lower_subgroup_id(nir_shader * s)1970 dxil_nir_lower_subgroup_id(nir_shader *s)
1971 {
1972 nir_def *subgroup_id = NULL;
1973 return nir_shader_intrinsics_pass(s, lower_subgroup_id, nir_metadata_none,
1974 &subgroup_id);
1975 }
1976
1977 static bool
lower_num_subgroups(nir_builder * b,nir_intrinsic_instr * intr,void * data)1978 lower_num_subgroups(nir_builder *b, nir_intrinsic_instr *intr, void *data)
1979 {
1980 if (intr->intrinsic != nir_intrinsic_load_num_subgroups)
1981 return false;
1982
1983 b->cursor = nir_before_instr(&intr->instr);
1984 nir_def *subgroup_size = nir_load_subgroup_size(b);
1985 nir_def *size_minus_one = nir_iadd_imm(b, subgroup_size, -1);
1986 nir_def *workgroup_size_vec = nir_load_workgroup_size(b);
1987 nir_def *workgroup_size = nir_imul(b, nir_channel(b, workgroup_size_vec, 0),
1988 nir_imul(b, nir_channel(b, workgroup_size_vec, 1),
1989 nir_channel(b, workgroup_size_vec, 2)));
1990 nir_def *ret = nir_idiv(b, nir_iadd(b, workgroup_size, size_minus_one), subgroup_size);
1991 nir_def_rewrite_uses(&intr->def, ret);
1992 return true;
1993 }
1994
1995 bool
dxil_nir_lower_num_subgroups(nir_shader * s)1996 dxil_nir_lower_num_subgroups(nir_shader *s)
1997 {
1998 return nir_shader_intrinsics_pass(s, lower_num_subgroups,
1999 nir_metadata_control_flow |
2000 nir_metadata_loop_analysis, NULL);
2001 }
2002
2003
2004 static const struct glsl_type *
get_cast_type(unsigned bit_size)2005 get_cast_type(unsigned bit_size)
2006 {
2007 switch (bit_size) {
2008 case 64:
2009 return glsl_int64_t_type();
2010 case 32:
2011 return glsl_int_type();
2012 case 16:
2013 return glsl_int16_t_type();
2014 case 8:
2015 return glsl_int8_t_type();
2016 }
2017 unreachable("Invalid bit_size");
2018 }
2019
2020 static void
split_unaligned_load(nir_builder * b,nir_intrinsic_instr * intrin,unsigned alignment)2021 split_unaligned_load(nir_builder *b, nir_intrinsic_instr *intrin, unsigned alignment)
2022 {
2023 enum gl_access_qualifier access = nir_intrinsic_access(intrin);
2024 nir_def *srcs[NIR_MAX_VEC_COMPONENTS * NIR_MAX_VEC_COMPONENTS * sizeof(int64_t) / 8];
2025 unsigned comp_size = intrin->def.bit_size / 8;
2026 unsigned num_comps = intrin->def.num_components;
2027
2028 b->cursor = nir_before_instr(&intrin->instr);
2029
2030 nir_deref_instr *ptr = nir_src_as_deref(intrin->src[0]);
2031
2032 const struct glsl_type *cast_type = get_cast_type(alignment * 8);
2033 nir_deref_instr *cast = nir_build_deref_cast(b, &ptr->def, ptr->modes, cast_type, alignment);
2034
2035 unsigned num_loads = DIV_ROUND_UP(comp_size * num_comps, alignment);
2036 for (unsigned i = 0; i < num_loads; ++i) {
2037 nir_deref_instr *elem = nir_build_deref_ptr_as_array(b, cast, nir_imm_intN_t(b, i, cast->def.bit_size));
2038 srcs[i] = nir_load_deref_with_access(b, elem, access);
2039 }
2040
2041 nir_def *new_dest = nir_extract_bits(b, srcs, num_loads, 0, num_comps, intrin->def.bit_size);
2042 nir_def_replace(&intrin->def, new_dest);
2043 }
2044
2045 static void
split_unaligned_store(nir_builder * b,nir_intrinsic_instr * intrin,unsigned alignment)2046 split_unaligned_store(nir_builder *b, nir_intrinsic_instr *intrin, unsigned alignment)
2047 {
2048 enum gl_access_qualifier access = nir_intrinsic_access(intrin);
2049
2050 nir_def *value = intrin->src[1].ssa;
2051 unsigned comp_size = value->bit_size / 8;
2052 unsigned num_comps = value->num_components;
2053
2054 b->cursor = nir_before_instr(&intrin->instr);
2055
2056 nir_deref_instr *ptr = nir_src_as_deref(intrin->src[0]);
2057
2058 const struct glsl_type *cast_type = get_cast_type(alignment * 8);
2059 nir_deref_instr *cast = nir_build_deref_cast(b, &ptr->def, ptr->modes, cast_type, alignment);
2060
2061 unsigned num_stores = DIV_ROUND_UP(comp_size * num_comps, alignment);
2062 for (unsigned i = 0; i < num_stores; ++i) {
2063 nir_def *substore_val = nir_extract_bits(b, &value, 1, i * alignment * 8, 1, alignment * 8);
2064 nir_deref_instr *elem = nir_build_deref_ptr_as_array(b, cast, nir_imm_intN_t(b, i, cast->def.bit_size));
2065 nir_store_deref_with_access(b, elem, substore_val, ~0, access);
2066 }
2067
2068 nir_instr_remove(&intrin->instr);
2069 }
2070
2071 bool
dxil_nir_split_unaligned_loads_stores(nir_shader * shader,nir_variable_mode modes)2072 dxil_nir_split_unaligned_loads_stores(nir_shader *shader, nir_variable_mode modes)
2073 {
2074 bool progress = false;
2075
2076 nir_foreach_function_impl(impl, shader) {
2077 nir_builder b = nir_builder_create(impl);
2078
2079 nir_foreach_block(block, impl) {
2080 nir_foreach_instr_safe(instr, block) {
2081 if (instr->type != nir_instr_type_intrinsic)
2082 continue;
2083 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
2084 if (intrin->intrinsic != nir_intrinsic_load_deref &&
2085 intrin->intrinsic != nir_intrinsic_store_deref)
2086 continue;
2087 nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
2088 if (!nir_deref_mode_may_be(deref, modes))
2089 continue;
2090
2091 unsigned align_mul = 0, align_offset = 0;
2092 nir_get_explicit_deref_align(deref, true, &align_mul, &align_offset);
2093
2094 unsigned alignment = align_offset ? 1 << (ffs(align_offset) - 1) : align_mul;
2095
2096 /* We can load anything at 4-byte alignment, except for
2097 * UBOs (AKA CBs where the granularity is 16 bytes).
2098 */
2099 unsigned req_align = (nir_deref_mode_is_one_of(deref, nir_var_mem_ubo | nir_var_mem_push_const) ? 16 : 4);
2100 if (alignment >= req_align)
2101 continue;
2102
2103 nir_def *val;
2104 if (intrin->intrinsic == nir_intrinsic_load_deref) {
2105 val = &intrin->def;
2106 } else {
2107 val = intrin->src[1].ssa;
2108 }
2109
2110 unsigned scalar_byte_size = glsl_type_is_boolean(deref->type) ? 4 : glsl_get_bit_size(deref->type) / 8;
2111 unsigned num_components =
2112 /* If the vector stride is larger than the scalar size, lower_explicit_io will
2113 * turn this into multiple scalar loads anyway, so we don't have to split it here. */
2114 glsl_get_explicit_stride(deref->type) > scalar_byte_size ? 1 :
2115 (val->num_components == 3 ? 4 : val->num_components);
2116 unsigned natural_alignment = scalar_byte_size * num_components;
2117
2118 if (alignment >= natural_alignment)
2119 continue;
2120
2121 if (intrin->intrinsic == nir_intrinsic_load_deref)
2122 split_unaligned_load(&b, intrin, alignment);
2123 else
2124 split_unaligned_store(&b, intrin, alignment);
2125 progress = true;
2126 }
2127 }
2128 }
2129
2130 return progress;
2131 }
2132
2133 static void
lower_inclusive_to_exclusive(nir_builder * b,nir_intrinsic_instr * intr)2134 lower_inclusive_to_exclusive(nir_builder *b, nir_intrinsic_instr *intr)
2135 {
2136 b->cursor = nir_after_instr(&intr->instr);
2137
2138 nir_op op = nir_intrinsic_reduction_op(intr);
2139 intr->intrinsic = nir_intrinsic_exclusive_scan;
2140 nir_intrinsic_set_reduction_op(intr, op);
2141
2142 nir_def *final_val = nir_build_alu2(b, nir_intrinsic_reduction_op(intr),
2143 &intr->def, intr->src[0].ssa);
2144 nir_def_rewrite_uses_after(&intr->def, final_val, final_val->parent_instr);
2145 }
2146
2147 static bool
lower_subgroup_scan(nir_builder * b,nir_intrinsic_instr * intr,void * data)2148 lower_subgroup_scan(nir_builder *b, nir_intrinsic_instr *intr, void *data)
2149 {
2150 switch (intr->intrinsic) {
2151 case nir_intrinsic_exclusive_scan:
2152 case nir_intrinsic_inclusive_scan:
2153 switch ((nir_op)nir_intrinsic_reduction_op(intr)) {
2154 case nir_op_iadd:
2155 case nir_op_fadd:
2156 case nir_op_imul:
2157 case nir_op_fmul:
2158 if (intr->intrinsic == nir_intrinsic_exclusive_scan)
2159 return false;
2160 lower_inclusive_to_exclusive(b, intr);
2161 return true;
2162 default:
2163 break;
2164 }
2165 break;
2166 default:
2167 return false;
2168 }
2169
2170 b->cursor = nir_before_instr(&intr->instr);
2171 nir_op op = nir_intrinsic_reduction_op(intr);
2172 nir_def *subgroup_id = nir_load_subgroup_invocation(b);
2173 nir_def *subgroup_size = nir_load_subgroup_size(b);
2174 nir_def *active_threads = nir_ballot(b, 4, 32, nir_imm_true(b));
2175 nir_def *base_value;
2176 uint32_t bit_size = intr->def.bit_size;
2177 if (op == nir_op_iand || op == nir_op_umin)
2178 base_value = nir_imm_intN_t(b, ~0ull, bit_size);
2179 else if (op == nir_op_imin)
2180 base_value = nir_imm_intN_t(b, (1ull << (bit_size - 1)) - 1, bit_size);
2181 else if (op == nir_op_imax)
2182 base_value = nir_imm_intN_t(b, 1ull << (bit_size - 1), bit_size);
2183 else if (op == nir_op_fmax)
2184 base_value = nir_imm_floatN_t(b, -INFINITY, bit_size);
2185 else if (op == nir_op_fmin)
2186 base_value = nir_imm_floatN_t(b, INFINITY, bit_size);
2187 else
2188 base_value = nir_imm_intN_t(b, 0, bit_size);
2189
2190 nir_variable *loop_counter_var = nir_local_variable_create(b->impl, glsl_uint_type(), "subgroup_loop_counter");
2191 nir_variable *result_var = nir_local_variable_create(b->impl,
2192 glsl_vector_type(nir_get_glsl_base_type_for_nir_type(
2193 nir_op_infos[op].input_types[0] | bit_size), 1),
2194 "subgroup_loop_result");
2195 nir_store_var(b, loop_counter_var, nir_imm_int(b, 0), 1);
2196 nir_store_var(b, result_var, base_value, 1);
2197 nir_loop *loop = nir_push_loop(b);
2198 nir_def *loop_counter = nir_load_var(b, loop_counter_var);
2199
2200 nir_if *nif = nir_push_if(b, nir_ilt(b, loop_counter, subgroup_size));
2201 nir_def *other_thread_val = nir_read_invocation(b, intr->src[0].ssa, loop_counter);
2202 nir_def *thread_in_range = intr->intrinsic == nir_intrinsic_inclusive_scan ?
2203 nir_ige(b, subgroup_id, loop_counter) :
2204 nir_ilt(b, loop_counter, subgroup_id);
2205 nir_def *thread_active = nir_ballot_bitfield_extract(b, 1, active_threads, loop_counter);
2206
2207 nir_if *if_active_thread = nir_push_if(b, nir_iand(b, thread_in_range, thread_active));
2208 nir_def *result = nir_build_alu2(b, op, nir_load_var(b, result_var), other_thread_val);
2209 nir_store_var(b, result_var, result, 1);
2210 nir_pop_if(b, if_active_thread);
2211
2212 nir_store_var(b, loop_counter_var, nir_iadd_imm(b, loop_counter, 1), 1);
2213 nir_jump(b, nir_jump_continue);
2214 nir_pop_if(b, nif);
2215
2216 nir_jump(b, nir_jump_break);
2217 nir_pop_loop(b, loop);
2218
2219 result = nir_load_var(b, result_var);
2220 nir_def_rewrite_uses(&intr->def, result);
2221 return true;
2222 }
2223
2224 bool
dxil_nir_lower_unsupported_subgroup_scan(nir_shader * s)2225 dxil_nir_lower_unsupported_subgroup_scan(nir_shader *s)
2226 {
2227 bool ret = nir_shader_intrinsics_pass(s, lower_subgroup_scan,
2228 nir_metadata_none, NULL);
2229 if (ret) {
2230 /* Lower the ballot bitfield tests */
2231 nir_lower_subgroups_options options = { .ballot_bit_size = 32, .ballot_components = 4 };
2232 nir_lower_subgroups(s, &options);
2233 }
2234 return ret;
2235 }
2236
2237 bool
dxil_nir_forward_front_face(nir_shader * nir)2238 dxil_nir_forward_front_face(nir_shader *nir)
2239 {
2240 assert(nir->info.stage == MESA_SHADER_FRAGMENT);
2241
2242 nir_variable *var = nir_find_variable_with_location(nir, nir_var_shader_in, VARYING_SLOT_FACE);
2243 if (var) {
2244 var->data.location = VARYING_SLOT_VAR12;
2245 return true;
2246 }
2247 return false;
2248 }
2249
2250 static bool
move_consts(nir_builder * b,nir_instr * instr,void * data)2251 move_consts(nir_builder *b, nir_instr *instr, void *data)
2252 {
2253 bool progress = false;
2254 switch (instr->type) {
2255 case nir_instr_type_load_const: {
2256 /* Sink load_const to their uses if there's multiple */
2257 nir_load_const_instr *load_const = nir_instr_as_load_const(instr);
2258 if (!list_is_singular(&load_const->def.uses)) {
2259 nir_foreach_use_safe(src, &load_const->def) {
2260 b->cursor = nir_before_src(src);
2261 nir_load_const_instr *new_load = nir_load_const_instr_create(b->shader,
2262 load_const->def.num_components,
2263 load_const->def.bit_size);
2264 memcpy(new_load->value, load_const->value, sizeof(load_const->value[0]) * load_const->def.num_components);
2265 nir_builder_instr_insert(b, &new_load->instr);
2266 nir_src_rewrite(src, &new_load->def);
2267 progress = true;
2268 }
2269 }
2270 return progress;
2271 }
2272 default:
2273 return false;
2274 }
2275 }
2276
2277 /* Sink all consts so that they have only have a single use.
2278 * The DXIL backend will already de-dupe the constants to the
2279 * same dxil_value if they have the same type, but this allows a single constant
2280 * to have different types without bitcasts. */
2281 bool
dxil_nir_move_consts(nir_shader * s)2282 dxil_nir_move_consts(nir_shader *s)
2283 {
2284 return nir_shader_instructions_pass(s, move_consts,
2285 nir_metadata_control_flow,
2286 NULL);
2287 }
2288
2289 static void
clear_pass_flags(nir_function_impl * impl)2290 clear_pass_flags(nir_function_impl *impl)
2291 {
2292 nir_foreach_block(block, impl) {
2293 nir_foreach_instr(instr, block) {
2294 instr->pass_flags = 0;
2295 }
2296 }
2297 }
2298
2299 static bool
add_def_to_worklist(nir_def * def,void * state)2300 add_def_to_worklist(nir_def *def, void *state)
2301 {
2302 nir_foreach_use_including_if(src, def) {
2303 if (nir_src_is_if(src)) {
2304 nir_if *nif = nir_src_parent_if(src);
2305 nir_foreach_block_in_cf_node(block, &nif->cf_node) {
2306 nir_foreach_instr(instr, block)
2307 nir_instr_worklist_push_tail(state, instr);
2308 }
2309 } else
2310 nir_instr_worklist_push_tail(state, nir_src_parent_instr(src));
2311 }
2312 return true;
2313 }
2314
2315 static bool
set_input_bits(struct dxil_module * mod,nir_intrinsic_instr * intr,BITSET_WORD * input_bits,uint32_t *** tables,const uint32_t ** table_sizes)2316 set_input_bits(struct dxil_module *mod, nir_intrinsic_instr *intr, BITSET_WORD *input_bits, uint32_t ***tables, const uint32_t **table_sizes)
2317 {
2318 if (intr->intrinsic == nir_intrinsic_load_view_index) {
2319 BITSET_SET(input_bits, 0);
2320 return true;
2321 }
2322
2323 bool any_bits_set = false;
2324 nir_src *row_src = intr->intrinsic == nir_intrinsic_load_per_vertex_input ? &intr->src[1] : &intr->src[0];
2325 bool is_patch_constant = mod->shader_kind == DXIL_DOMAIN_SHADER && intr->intrinsic == nir_intrinsic_load_input;
2326 const struct dxil_signature_record *sig_rec = is_patch_constant ?
2327 &mod->patch_consts[nir_intrinsic_base(intr)] :
2328 &mod->inputs[mod->input_mappings[nir_intrinsic_base(intr)]];
2329 if (is_patch_constant) {
2330 /* Redirect to the second I/O table */
2331 *tables = *tables + 1;
2332 *table_sizes = *table_sizes + 1;
2333 }
2334 for (uint32_t component = 0; component < intr->num_components; ++component) {
2335 uint32_t base_element = 0;
2336 uint32_t num_elements = sig_rec->num_elements;
2337 if (nir_src_is_const(*row_src)) {
2338 base_element = (uint32_t)nir_src_as_uint(*row_src);
2339 num_elements = 1;
2340 }
2341 for (uint32_t element = 0; element < num_elements; ++element) {
2342 uint32_t row = sig_rec->elements[element + base_element].reg;
2343 if (row == 0xffffffff)
2344 continue;
2345 BITSET_SET(input_bits, row * 4 + component + nir_intrinsic_component(intr));
2346 any_bits_set = true;
2347 }
2348 }
2349 return any_bits_set;
2350 }
2351
2352 static bool
set_output_bits(struct dxil_module * mod,nir_intrinsic_instr * intr,BITSET_WORD * input_bits,uint32_t ** tables,const uint32_t * table_sizes)2353 set_output_bits(struct dxil_module *mod, nir_intrinsic_instr *intr, BITSET_WORD *input_bits, uint32_t **tables, const uint32_t *table_sizes)
2354 {
2355 bool any_bits_set = false;
2356 nir_src *row_src = intr->intrinsic == nir_intrinsic_store_per_vertex_output ? &intr->src[2] : &intr->src[1];
2357 bool is_patch_constant = mod->shader_kind == DXIL_HULL_SHADER && intr->intrinsic == nir_intrinsic_store_output;
2358 const struct dxil_signature_record *sig_rec = is_patch_constant ?
2359 &mod->patch_consts[nir_intrinsic_base(intr)] :
2360 &mod->outputs[nir_intrinsic_base(intr)];
2361 for (uint32_t component = 0; component < intr->num_components; ++component) {
2362 uint32_t base_element = 0;
2363 uint32_t num_elements = sig_rec->num_elements;
2364 if (nir_src_is_const(*row_src)) {
2365 base_element = (uint32_t)nir_src_as_uint(*row_src);
2366 num_elements = 1;
2367 }
2368 for (uint32_t element = 0; element < num_elements; ++element) {
2369 uint32_t row = sig_rec->elements[element + base_element].reg;
2370 if (row == 0xffffffff)
2371 continue;
2372 uint32_t stream = sig_rec->elements[element + base_element].stream;
2373 uint32_t table_idx = is_patch_constant ? 1 : stream;
2374 uint32_t *table = tables[table_idx];
2375 uint32_t output_component = component + nir_intrinsic_component(intr);
2376 uint32_t input_component;
2377 BITSET_FOREACH_SET(input_component, input_bits, 32 * 4) {
2378 uint32_t *table_for_input_component = table + table_sizes[table_idx] * input_component;
2379 BITSET_SET(table_for_input_component, row * 4 + output_component);
2380 any_bits_set = true;
2381 }
2382 }
2383 }
2384 return any_bits_set;
2385 }
2386
2387 static bool
propagate_input_to_output_dependencies(struct dxil_module * mod,nir_intrinsic_instr * load_intr,uint32_t ** tables,const uint32_t * table_sizes)2388 propagate_input_to_output_dependencies(struct dxil_module *mod, nir_intrinsic_instr *load_intr, uint32_t **tables, const uint32_t *table_sizes)
2389 {
2390 /* Which input components are being loaded by this instruction */
2391 BITSET_DECLARE(input_bits, 32 * 4) = { 0 };
2392 if (!set_input_bits(mod, load_intr, input_bits, &tables, &table_sizes))
2393 return false;
2394
2395 nir_instr_worklist *worklist = nir_instr_worklist_create();
2396 nir_instr_worklist_push_tail(worklist, &load_intr->instr);
2397 bool any_bits_set = false;
2398 nir_foreach_instr_in_worklist(instr, worklist) {
2399 if (instr->pass_flags)
2400 continue;
2401
2402 instr->pass_flags = 1;
2403 nir_foreach_def(instr, add_def_to_worklist, worklist);
2404 switch (instr->type) {
2405 case nir_instr_type_jump: {
2406 nir_jump_instr *jump = nir_instr_as_jump(instr);
2407 switch (jump->type) {
2408 case nir_jump_break:
2409 case nir_jump_continue: {
2410 nir_cf_node *parent = &instr->block->cf_node;
2411 while (parent->type != nir_cf_node_loop)
2412 parent = parent->parent;
2413 nir_foreach_block_in_cf_node(block, parent)
2414 nir_foreach_instr(i, block)
2415 nir_instr_worklist_push_tail(worklist, i);
2416 }
2417 break;
2418 default:
2419 unreachable("Don't expect any other jumps");
2420 }
2421 break;
2422 }
2423 case nir_instr_type_intrinsic: {
2424 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
2425 switch (intr->intrinsic) {
2426 case nir_intrinsic_store_output:
2427 case nir_intrinsic_store_per_vertex_output:
2428 any_bits_set |= set_output_bits(mod, intr, input_bits, tables, table_sizes);
2429 break;
2430 /* TODO: Memory writes */
2431 default:
2432 break;
2433 }
2434 break;
2435 }
2436 default:
2437 break;
2438 }
2439 }
2440
2441 nir_instr_worklist_destroy(worklist);
2442 return any_bits_set;
2443 }
2444
2445 /* For every input load, compute the set of output stores that it can contribute to.
2446 * If it contributes to a store to memory, If it's used for control flow, then any
2447 * instruction in the CFG that it impacts is considered to contribute.
2448 * Ideally, we should also handle stores to outputs/memory and then loads from that
2449 * output/memory, but this is non-trivial and unclear how much impact that would have. */
2450 bool
dxil_nir_analyze_io_dependencies(struct dxil_module * mod,nir_shader * s)2451 dxil_nir_analyze_io_dependencies(struct dxil_module *mod, nir_shader *s)
2452 {
2453 bool any_outputs = false;
2454 for (uint32_t i = 0; i < 4; ++i)
2455 any_outputs |= mod->num_psv_outputs[i] > 0;
2456 if (mod->shader_kind == DXIL_HULL_SHADER)
2457 any_outputs |= mod->num_psv_patch_consts > 0;
2458 if (!any_outputs)
2459 return false;
2460
2461 bool any_bits_set = false;
2462 nir_foreach_function(func, s) {
2463 assert(func->impl);
2464 /* Hull shaders have a patch constant function */
2465 assert(func->is_entrypoint || s->info.stage == MESA_SHADER_TESS_CTRL);
2466
2467 /* Pass 1: input/view ID -> output dependencies */
2468 nir_foreach_block(block, func->impl) {
2469 nir_foreach_instr(instr, block) {
2470 if (instr->type != nir_instr_type_intrinsic)
2471 continue;
2472 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
2473 uint32_t **tables = mod->io_dependency_table;
2474 const uint32_t *table_sizes = mod->dependency_table_dwords_per_input;
2475 switch (intr->intrinsic) {
2476 case nir_intrinsic_load_view_index:
2477 tables = mod->viewid_dependency_table;
2478 FALLTHROUGH;
2479 case nir_intrinsic_load_input:
2480 case nir_intrinsic_load_per_vertex_input:
2481 case nir_intrinsic_load_interpolated_input:
2482 break;
2483 default:
2484 continue;
2485 }
2486
2487 clear_pass_flags(func->impl);
2488 any_bits_set |= propagate_input_to_output_dependencies(mod, intr, tables, table_sizes);
2489 }
2490 }
2491
2492 /* Pass 2: output -> output dependencies */
2493 /* TODO */
2494 }
2495 return any_bits_set;
2496 }
2497
2498 static enum pipe_format
get_format_for_var(unsigned num_comps,enum glsl_base_type sampled_type)2499 get_format_for_var(unsigned num_comps, enum glsl_base_type sampled_type)
2500 {
2501 switch (sampled_type) {
2502 case GLSL_TYPE_INT:
2503 case GLSL_TYPE_INT64:
2504 case GLSL_TYPE_INT16:
2505 switch (num_comps) {
2506 case 1: return PIPE_FORMAT_R32_SINT;
2507 case 2: return PIPE_FORMAT_R32G32_SINT;
2508 case 3: return PIPE_FORMAT_R32G32B32_SINT;
2509 case 4: return PIPE_FORMAT_R32G32B32A32_SINT;
2510 default: unreachable("Invalid num_comps");
2511 }
2512 case GLSL_TYPE_UINT:
2513 case GLSL_TYPE_UINT64:
2514 case GLSL_TYPE_UINT16:
2515 switch (num_comps) {
2516 case 1: return PIPE_FORMAT_R32_UINT;
2517 case 2: return PIPE_FORMAT_R32G32_UINT;
2518 case 3: return PIPE_FORMAT_R32G32B32_UINT;
2519 case 4: return PIPE_FORMAT_R32G32B32A32_UINT;
2520 default: unreachable("Invalid num_comps");
2521 }
2522 case GLSL_TYPE_FLOAT:
2523 case GLSL_TYPE_FLOAT16:
2524 case GLSL_TYPE_DOUBLE:
2525 switch (num_comps) {
2526 case 1: return PIPE_FORMAT_R32_FLOAT;
2527 case 2: return PIPE_FORMAT_R32G32_FLOAT;
2528 case 3: return PIPE_FORMAT_R32G32B32_FLOAT;
2529 case 4: return PIPE_FORMAT_R32G32B32A32_FLOAT;
2530 default: unreachable("Invalid num_comps");
2531 }
2532 default: unreachable("Invalid sampler return type");
2533 }
2534 }
2535
2536 static unsigned
aoa_size(const struct glsl_type * type)2537 aoa_size(const struct glsl_type *type)
2538 {
2539 return glsl_type_is_array(type) ? glsl_get_aoa_size(type) : 1;
2540 }
2541
2542 static bool
guess_image_format_for_var(nir_shader * s,nir_variable * var)2543 guess_image_format_for_var(nir_shader *s, nir_variable *var)
2544 {
2545 const struct glsl_type *base_type = glsl_without_array(var->type);
2546 if (!glsl_type_is_image(base_type))
2547 return false;
2548 if (var->data.image.format != PIPE_FORMAT_NONE)
2549 return false;
2550
2551 nir_foreach_function_impl(impl, s) {
2552 nir_foreach_block(block, impl) {
2553 nir_foreach_instr(instr, block) {
2554 if (instr->type != nir_instr_type_intrinsic)
2555 continue;
2556 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
2557 switch (intr->intrinsic) {
2558 case nir_intrinsic_image_deref_load:
2559 case nir_intrinsic_image_deref_store:
2560 case nir_intrinsic_image_deref_atomic:
2561 case nir_intrinsic_image_deref_atomic_swap:
2562 if (nir_intrinsic_get_var(intr, 0) != var)
2563 continue;
2564 break;
2565 case nir_intrinsic_image_load:
2566 case nir_intrinsic_image_store:
2567 case nir_intrinsic_image_atomic:
2568 case nir_intrinsic_image_atomic_swap: {
2569 unsigned binding = nir_src_as_uint(intr->src[0]);
2570 if (binding < var->data.binding ||
2571 binding >= var->data.binding + aoa_size(var->type))
2572 continue;
2573 break;
2574 }
2575 default:
2576 continue;
2577 }
2578 break;
2579
2580 switch (intr->intrinsic) {
2581 case nir_intrinsic_image_deref_load:
2582 case nir_intrinsic_image_load:
2583 case nir_intrinsic_image_deref_store:
2584 case nir_intrinsic_image_store:
2585 /* Increase unknown formats up to 4 components if a 4-component accessor is used */
2586 if (intr->num_components > util_format_get_nr_components(var->data.image.format))
2587 var->data.image.format = get_format_for_var(intr->num_components, glsl_get_sampler_result_type(base_type));
2588 break;
2589 default:
2590 /* If an atomic is used, the image format must be 1-component; return immediately */
2591 var->data.image.format = get_format_for_var(1, glsl_get_sampler_result_type(base_type));
2592 return true;
2593 }
2594 }
2595 }
2596 }
2597 /* Dunno what it is, assume 4-component */
2598 if (var->data.image.format == PIPE_FORMAT_NONE)
2599 var->data.image.format = get_format_for_var(4, glsl_get_sampler_result_type(base_type));
2600 return true;
2601 }
2602
2603 static void
update_intrinsic_format_and_type(nir_intrinsic_instr * intr,nir_variable * var)2604 update_intrinsic_format_and_type(nir_intrinsic_instr *intr, nir_variable *var)
2605 {
2606 nir_intrinsic_set_format(intr, var->data.image.format);
2607 nir_alu_type alu_type =
2608 nir_get_nir_type_for_glsl_base_type(glsl_get_sampler_result_type(glsl_without_array(var->type)));
2609 if (nir_intrinsic_has_src_type(intr))
2610 nir_intrinsic_set_src_type(intr, alu_type);
2611 else if (nir_intrinsic_has_dest_type(intr))
2612 nir_intrinsic_set_dest_type(intr, alu_type);
2613 }
2614
2615 static bool
update_intrinsic_formats(nir_builder * b,nir_intrinsic_instr * intr,void * data)2616 update_intrinsic_formats(nir_builder *b, nir_intrinsic_instr *intr,
2617 void *data)
2618 {
2619 if (!nir_intrinsic_has_format(intr))
2620 return false;
2621 nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
2622 if (deref) {
2623 nir_variable *var = nir_deref_instr_get_variable(deref);
2624 if (var)
2625 update_intrinsic_format_and_type(intr, var);
2626 return var != NULL;
2627 }
2628
2629 if (!nir_intrinsic_has_range_base(intr))
2630 return false;
2631
2632 unsigned binding = nir_src_as_uint(intr->src[0]);
2633 nir_foreach_variable_with_modes(var, b->shader, nir_var_image) {
2634 if (var->data.binding <= binding &&
2635 var->data.binding + aoa_size(var->type) > binding) {
2636 update_intrinsic_format_and_type(intr, var);
2637 return true;
2638 }
2639 }
2640 return false;
2641 }
2642
2643 bool
dxil_nir_guess_image_formats(nir_shader * s)2644 dxil_nir_guess_image_formats(nir_shader *s)
2645 {
2646 bool progress = false;
2647 nir_foreach_variable_with_modes(var, s, nir_var_image) {
2648 progress |= guess_image_format_for_var(s, var);
2649 }
2650 nir_shader_intrinsics_pass(s, update_intrinsic_formats, nir_metadata_all,
2651 NULL);
2652 return progress;
2653 }
2654
2655 static void
set_binding_variables_coherent(nir_shader * s,nir_binding binding,nir_variable_mode modes)2656 set_binding_variables_coherent(nir_shader *s, nir_binding binding, nir_variable_mode modes)
2657 {
2658 nir_foreach_variable_with_modes(var, s, modes) {
2659 if (var->data.binding == binding.binding &&
2660 var->data.descriptor_set == binding.desc_set) {
2661 var->data.access |= ACCESS_COHERENT;
2662 }
2663 }
2664 }
2665
2666 static void
set_deref_variables_coherent(nir_shader * s,nir_deref_instr * deref)2667 set_deref_variables_coherent(nir_shader *s, nir_deref_instr *deref)
2668 {
2669 while (deref->deref_type != nir_deref_type_var &&
2670 deref->deref_type != nir_deref_type_cast) {
2671 deref = nir_deref_instr_parent(deref);
2672 }
2673 if (deref->deref_type == nir_deref_type_var) {
2674 deref->var->data.access |= ACCESS_COHERENT;
2675 return;
2676 }
2677
2678 /* For derefs with casts, we only support pre-lowered Vulkan accesses */
2679 assert(deref->deref_type == nir_deref_type_cast);
2680 nir_intrinsic_instr *cast_src = nir_instr_as_intrinsic(deref->parent.ssa->parent_instr);
2681 assert(cast_src->intrinsic == nir_intrinsic_load_vulkan_descriptor);
2682 nir_binding binding = nir_chase_binding(cast_src->src[0]);
2683 set_binding_variables_coherent(s, binding, nir_var_mem_ssbo);
2684 }
2685
2686 static nir_def *
get_atomic_for_load_store(nir_builder * b,nir_intrinsic_instr * intr,unsigned bit_size)2687 get_atomic_for_load_store(nir_builder *b, nir_intrinsic_instr *intr, unsigned bit_size)
2688 {
2689 nir_def *zero = nir_imm_intN_t(b, 0, bit_size);
2690 switch (intr->intrinsic) {
2691 case nir_intrinsic_load_deref:
2692 return nir_deref_atomic(b, bit_size, intr->src[0].ssa, zero, .atomic_op = nir_atomic_op_iadd);
2693 case nir_intrinsic_load_ssbo:
2694 return nir_ssbo_atomic(b, bit_size, intr->src[0].ssa, intr->src[1].ssa, zero, .atomic_op = nir_atomic_op_iadd);
2695 case nir_intrinsic_image_deref_load:
2696 return nir_image_deref_atomic(b, bit_size, intr->src[0].ssa, intr->src[1].ssa, intr->src[2].ssa, zero, .atomic_op = nir_atomic_op_iadd);
2697 case nir_intrinsic_image_load:
2698 return nir_image_atomic(b, bit_size, intr->src[0].ssa, intr->src[1].ssa, intr->src[2].ssa, zero, .atomic_op = nir_atomic_op_iadd);
2699 case nir_intrinsic_store_deref:
2700 return nir_deref_atomic(b, bit_size, intr->src[0].ssa, intr->src[1].ssa, .atomic_op = nir_atomic_op_xchg);
2701 case nir_intrinsic_store_ssbo:
2702 return nir_ssbo_atomic(b, bit_size, intr->src[1].ssa, intr->src[2].ssa, intr->src[0].ssa, .atomic_op = nir_atomic_op_xchg);
2703 case nir_intrinsic_image_deref_store:
2704 return nir_image_deref_atomic(b, bit_size, intr->src[0].ssa, intr->src[1].ssa, intr->src[2].ssa, intr->src[3].ssa, .atomic_op = nir_atomic_op_xchg);
2705 case nir_intrinsic_image_store:
2706 return nir_image_atomic(b, bit_size, intr->src[0].ssa, intr->src[1].ssa, intr->src[2].ssa, intr->src[3].ssa, .atomic_op = nir_atomic_op_xchg);
2707 default:
2708 return NULL;
2709 }
2710 }
2711
2712 static bool
lower_coherent_load_store(nir_builder * b,nir_intrinsic_instr * intr,void * context)2713 lower_coherent_load_store(nir_builder *b, nir_intrinsic_instr *intr, void *context)
2714 {
2715 if (!nir_intrinsic_has_access(intr) || (nir_intrinsic_access(intr) & ACCESS_COHERENT) == 0)
2716 return false;
2717
2718 nir_def *atomic_def = NULL;
2719 b->cursor = nir_before_instr(&intr->instr);
2720 switch (intr->intrinsic) {
2721 case nir_intrinsic_load_deref:
2722 case nir_intrinsic_load_ssbo:
2723 case nir_intrinsic_image_deref_load:
2724 case nir_intrinsic_image_load: {
2725 if (intr->def.bit_size < 32 || intr->def.num_components > 1) {
2726 if (intr->intrinsic == nir_intrinsic_load_deref)
2727 set_deref_variables_coherent(b->shader, nir_src_as_deref(intr->src[0]));
2728 else {
2729 nir_binding binding = {0};
2730 if (nir_src_is_const(intr->src[0]))
2731 binding.binding = nir_src_as_uint(intr->src[0]);
2732 set_binding_variables_coherent(b->shader, binding,
2733 intr->intrinsic == nir_intrinsic_load_ssbo ? nir_var_mem_ssbo : nir_var_image);
2734 }
2735 return false;
2736 }
2737
2738 atomic_def = get_atomic_for_load_store(b, intr, intr->def.bit_size);
2739 nir_def_rewrite_uses(&intr->def, atomic_def);
2740 break;
2741 }
2742 case nir_intrinsic_store_deref:
2743 case nir_intrinsic_store_ssbo:
2744 case nir_intrinsic_image_deref_store:
2745 case nir_intrinsic_image_store: {
2746 int resource_idx = intr->intrinsic == nir_intrinsic_store_ssbo ? 1 : 0;
2747 int value_idx = intr->intrinsic == nir_intrinsic_store_ssbo ? 0 :
2748 intr->intrinsic == nir_intrinsic_store_deref ? 1 : 3;
2749 unsigned num_components = nir_intrinsic_has_write_mask(intr) ?
2750 util_bitcount(nir_intrinsic_write_mask(intr)) : intr->src[value_idx].ssa->num_components;
2751 if (intr->src[value_idx].ssa->bit_size < 32 || num_components > 1) {
2752 if (intr->intrinsic == nir_intrinsic_store_deref)
2753 set_deref_variables_coherent(b->shader, nir_src_as_deref(intr->src[resource_idx]));
2754 else {
2755 nir_binding binding = {0};
2756 if (nir_src_is_const(intr->src[resource_idx]))
2757 binding.binding = nir_src_as_uint(intr->src[resource_idx]);
2758 set_binding_variables_coherent(b->shader, binding,
2759 intr->intrinsic == nir_intrinsic_store_ssbo ? nir_var_mem_ssbo : nir_var_image);
2760 }
2761 return false;
2762 }
2763
2764 atomic_def = get_atomic_for_load_store(b, intr, intr->src[value_idx].ssa->bit_size);
2765 break;
2766 }
2767 default:
2768 return false;
2769 }
2770
2771 nir_intrinsic_instr *atomic = nir_instr_as_intrinsic(atomic_def->parent_instr);
2772 nir_intrinsic_set_access(atomic, nir_intrinsic_access(intr));
2773 if (nir_intrinsic_has_image_dim(intr))
2774 nir_intrinsic_set_image_dim(atomic, nir_intrinsic_image_dim(intr));
2775 if (nir_intrinsic_has_image_array(intr))
2776 nir_intrinsic_set_image_array(atomic, nir_intrinsic_image_array(intr));
2777 if (nir_intrinsic_has_format(intr))
2778 nir_intrinsic_set_format(atomic, nir_intrinsic_format(intr));
2779 if (nir_intrinsic_has_range_base(intr))
2780 nir_intrinsic_set_range_base(atomic, nir_intrinsic_range_base(intr));
2781 nir_instr_remove(&intr->instr);
2782 return true;
2783 }
2784
2785 bool
dxil_nir_lower_coherent_loads_and_stores(nir_shader * s)2786 dxil_nir_lower_coherent_loads_and_stores(nir_shader *s)
2787 {
2788 return nir_shader_intrinsics_pass(s, lower_coherent_load_store,
2789 nir_metadata_control_flow | nir_metadata_loop_analysis,
2790 NULL);
2791 }
2792
2793 struct undefined_varying_masks {
2794 uint64_t io_mask;
2795 uint32_t patch_io_mask;
2796 const BITSET_WORD *frac_io_mask;
2797 };
2798
2799 static bool
is_dead_in_variable(nir_variable * var,void * data)2800 is_dead_in_variable(nir_variable *var, void *data)
2801 {
2802 switch (var->data.location) {
2803 /* Only these values can be system generated values in addition to varyings */
2804 case VARYING_SLOT_PRIMITIVE_ID:
2805 case VARYING_SLOT_FACE:
2806 case VARYING_SLOT_VIEW_INDEX:
2807 return false;
2808 /* Tessellation input vars must remain untouched */
2809 case VARYING_SLOT_TESS_LEVEL_INNER:
2810 case VARYING_SLOT_TESS_LEVEL_OUTER:
2811 return false;
2812 default:
2813 return true;
2814 }
2815 }
2816
2817 static bool
kill_undefined_varyings(struct nir_builder * b,nir_instr * instr,void * data)2818 kill_undefined_varyings(struct nir_builder *b,
2819 nir_instr *instr,
2820 void *data)
2821 {
2822 const struct undefined_varying_masks *masks = data;
2823
2824 if (instr->type != nir_instr_type_intrinsic)
2825 return false;
2826
2827 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
2828
2829 if (intr->intrinsic != nir_intrinsic_load_deref)
2830 return false;
2831
2832 nir_variable *var = nir_intrinsic_get_var(intr, 0);
2833 if (!var || var->data.mode != nir_var_shader_in)
2834 return false;
2835
2836 if (!is_dead_in_variable(var, NULL))
2837 return false;
2838
2839 uint32_t loc = var->data.patch && var->data.location >= VARYING_SLOT_PATCH0 ?
2840 var->data.location - VARYING_SLOT_PATCH0 :
2841 var->data.location;
2842 uint64_t written = var->data.patch && var->data.location >= VARYING_SLOT_PATCH0 ?
2843 masks->patch_io_mask : masks->io_mask;
2844 if (BITFIELD64_RANGE(loc, glsl_varying_count(var->type)) & written) {
2845 if (!masks->frac_io_mask || !var->data.location_frac ||
2846 var->data.location < VARYING_SLOT_VAR0 ||
2847 BITSET_TEST(masks->frac_io_mask, (var->data.location - VARYING_SLOT_VAR0) * 4 + var->data.location_frac))
2848 return false;
2849 }
2850
2851 b->cursor = nir_after_instr(instr);
2852 /* Note: zero is used instead of undef, because optimization is not run here, but is
2853 * run later on. If we load an undef here, and that undef ends up being used to store
2854 * to position later on, that can cause some or all of the components in that position
2855 * write to be removed, which is problematic especially in the case of all components,
2856 * since that would remove the store instruction, and would make it tricky to satisfy
2857 * the DXIL requirements of writing all position components.
2858 */
2859 nir_def *zero = nir_imm_zero(b, intr->def.num_components,
2860 intr->def.bit_size);
2861 nir_def_replace(&intr->def, zero);
2862 return true;
2863 }
2864
2865 bool
dxil_nir_kill_undefined_varyings(nir_shader * shader,uint64_t prev_stage_written_mask,uint32_t prev_stage_patch_written_mask,const BITSET_WORD * prev_stage_frac_output_mask)2866 dxil_nir_kill_undefined_varyings(nir_shader *shader, uint64_t prev_stage_written_mask, uint32_t prev_stage_patch_written_mask,
2867 const BITSET_WORD *prev_stage_frac_output_mask)
2868 {
2869 struct undefined_varying_masks masks = {
2870 .io_mask = prev_stage_written_mask,
2871 .patch_io_mask = prev_stage_patch_written_mask,
2872 .frac_io_mask = prev_stage_frac_output_mask
2873 };
2874 bool progress = nir_shader_instructions_pass(shader,
2875 kill_undefined_varyings,
2876 nir_metadata_control_flow |
2877 nir_metadata_loop_analysis,
2878 (void *)&masks);
2879 if (progress) {
2880 nir_opt_dce(shader);
2881 nir_remove_dead_derefs(shader);
2882 }
2883
2884 const struct nir_remove_dead_variables_options options = {
2885 .can_remove_var = is_dead_in_variable,
2886 .can_remove_var_data = &masks,
2887 };
2888 progress |= nir_remove_dead_variables(shader, nir_var_shader_in, &options);
2889 return progress;
2890 }
2891
2892 static bool
is_dead_out_variable(nir_variable * var,void * data)2893 is_dead_out_variable(nir_variable *var, void *data)
2894 {
2895 return !nir_slot_is_sysval_output(var->data.location, MESA_SHADER_NONE);
2896 }
2897
2898 static bool
kill_unused_outputs(struct nir_builder * b,nir_instr * instr,void * data)2899 kill_unused_outputs(struct nir_builder *b,
2900 nir_instr *instr,
2901 void *data)
2902 {
2903 const struct undefined_varying_masks *masks = data;
2904
2905 if (instr->type != nir_instr_type_intrinsic)
2906 return false;
2907
2908 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
2909
2910 if (intr->intrinsic != nir_intrinsic_store_deref &&
2911 intr->intrinsic != nir_intrinsic_load_deref)
2912 return false;
2913
2914 nir_variable *var = nir_intrinsic_get_var(intr, 0);
2915 if (!var || var->data.mode != nir_var_shader_out ||
2916 /* always_active_io can mean two things: xfb or GL separable shaders. We can't delete
2917 * varyings that are used for xfb (we'll just sort them last), but we must delete varyings
2918 * that are mismatching between TCS and TES. Fortunately TCS can't do xfb, so we can ignore
2919 the always_active_io bit for TCS outputs. */
2920 (b->shader->info.stage != MESA_SHADER_TESS_CTRL && var->data.always_active_io))
2921 return false;
2922
2923 if (!is_dead_out_variable(var, NULL))
2924 return false;
2925
2926 unsigned loc = var->data.patch && var->data.location >= VARYING_SLOT_PATCH0 ?
2927 var->data.location - VARYING_SLOT_PATCH0 :
2928 var->data.location;
2929 uint64_t read = var->data.patch && var->data.location >= VARYING_SLOT_PATCH0 ?
2930 masks->patch_io_mask : masks->io_mask;
2931 if (BITFIELD64_RANGE(loc, glsl_varying_count(var->type)) & read) {
2932 if (!masks->frac_io_mask || !var->data.location_frac ||
2933 var->data.location < VARYING_SLOT_VAR0 ||
2934 BITSET_TEST(masks->frac_io_mask, (var->data.location - VARYING_SLOT_VAR0) * 4 + var->data.location_frac))
2935 return false;
2936 }
2937
2938 if (intr->intrinsic == nir_intrinsic_load_deref) {
2939 b->cursor = nir_after_instr(&intr->instr);
2940 nir_def *zero = nir_imm_zero(b, intr->def.num_components, intr->def.bit_size);
2941 nir_def_rewrite_uses(&intr->def, zero);
2942 }
2943 nir_instr_remove(instr);
2944 return true;
2945 }
2946
2947 bool
dxil_nir_kill_unused_outputs(nir_shader * shader,uint64_t next_stage_read_mask,uint32_t next_stage_patch_read_mask,const BITSET_WORD * next_stage_frac_input_mask)2948 dxil_nir_kill_unused_outputs(nir_shader *shader, uint64_t next_stage_read_mask, uint32_t next_stage_patch_read_mask,
2949 const BITSET_WORD *next_stage_frac_input_mask)
2950 {
2951 struct undefined_varying_masks masks = {
2952 .io_mask = next_stage_read_mask,
2953 .patch_io_mask = next_stage_patch_read_mask,
2954 .frac_io_mask = next_stage_frac_input_mask
2955 };
2956
2957 bool progress = nir_shader_instructions_pass(shader,
2958 kill_unused_outputs,
2959 nir_metadata_control_flow |
2960 nir_metadata_loop_analysis,
2961 (void *)&masks);
2962
2963 if (progress) {
2964 nir_opt_dce(shader);
2965 nir_remove_dead_derefs(shader);
2966 }
2967 const struct nir_remove_dead_variables_options options = {
2968 .can_remove_var = is_dead_out_variable,
2969 .can_remove_var_data = &masks,
2970 };
2971 progress |= nir_remove_dead_variables(shader, nir_var_shader_out, &options);
2972 return progress;
2973 }
2974