• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2015 Intel 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 "vtn_private.h"
25 #include "spirv_info.h"
26 #include "nir/nir_vla.h"
27 #include "util/debug.h"
28 
29 static struct vtn_block *
vtn_block(struct vtn_builder * b,uint32_t value_id)30 vtn_block(struct vtn_builder *b, uint32_t value_id)
31 {
32    return vtn_value(b, value_id, vtn_value_type_block)->block;
33 }
34 
35 static unsigned
glsl_type_count_function_params(const struct glsl_type * type)36 glsl_type_count_function_params(const struct glsl_type *type)
37 {
38    if (glsl_type_is_vector_or_scalar(type)) {
39       return 1;
40    } else if (glsl_type_is_array_or_matrix(type)) {
41       return glsl_get_length(type) *
42              glsl_type_count_function_params(glsl_get_array_element(type));
43    } else {
44       assert(glsl_type_is_struct_or_ifc(type));
45       unsigned count = 0;
46       unsigned elems = glsl_get_length(type);
47       for (unsigned i = 0; i < elems; i++) {
48          const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
49          count += glsl_type_count_function_params(elem_type);
50       }
51       return count;
52    }
53 }
54 
55 static void
glsl_type_add_to_function_params(const struct glsl_type * type,nir_function * func,unsigned * param_idx)56 glsl_type_add_to_function_params(const struct glsl_type *type,
57                                  nir_function *func,
58                                  unsigned *param_idx)
59 {
60    if (glsl_type_is_vector_or_scalar(type)) {
61       func->params[(*param_idx)++] = (nir_parameter) {
62          .num_components = glsl_get_vector_elements(type),
63          .bit_size = glsl_get_bit_size(type),
64       };
65    } else if (glsl_type_is_array_or_matrix(type)) {
66       unsigned elems = glsl_get_length(type);
67       const struct glsl_type *elem_type = glsl_get_array_element(type);
68       for (unsigned i = 0; i < elems; i++)
69          glsl_type_add_to_function_params(elem_type,func, param_idx);
70    } else {
71       assert(glsl_type_is_struct_or_ifc(type));
72       unsigned elems = glsl_get_length(type);
73       for (unsigned i = 0; i < elems; i++) {
74          const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
75          glsl_type_add_to_function_params(elem_type, func, param_idx);
76       }
77    }
78 }
79 
80 static void
vtn_ssa_value_add_to_call_params(struct vtn_builder * b,struct vtn_ssa_value * value,nir_call_instr * call,unsigned * param_idx)81 vtn_ssa_value_add_to_call_params(struct vtn_builder *b,
82                                  struct vtn_ssa_value *value,
83                                  nir_call_instr *call,
84                                  unsigned *param_idx)
85 {
86    if (glsl_type_is_vector_or_scalar(value->type)) {
87       call->params[(*param_idx)++] = nir_src_for_ssa(value->def);
88    } else {
89       unsigned elems = glsl_get_length(value->type);
90       for (unsigned i = 0; i < elems; i++) {
91          vtn_ssa_value_add_to_call_params(b, value->elems[i],
92                                           call, param_idx);
93       }
94    }
95 }
96 
97 static void
vtn_ssa_value_load_function_param(struct vtn_builder * b,struct vtn_ssa_value * value,unsigned * param_idx)98 vtn_ssa_value_load_function_param(struct vtn_builder *b,
99                                   struct vtn_ssa_value *value,
100                                   unsigned *param_idx)
101 {
102    if (glsl_type_is_vector_or_scalar(value->type)) {
103       value->def = nir_load_param(&b->nb, (*param_idx)++);
104    } else {
105       unsigned elems = glsl_get_length(value->type);
106       for (unsigned i = 0; i < elems; i++)
107          vtn_ssa_value_load_function_param(b, value->elems[i], param_idx);
108    }
109 }
110 
111 void
vtn_handle_function_call(struct vtn_builder * b,SpvOp opcode,const uint32_t * w,unsigned count)112 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
113                          const uint32_t *w, unsigned count)
114 {
115    struct vtn_function *vtn_callee =
116       vtn_value(b, w[3], vtn_value_type_function)->func;
117 
118    vtn_callee->referenced = true;
119 
120    nir_call_instr *call = nir_call_instr_create(b->nb.shader,
121                                                 vtn_callee->nir_func);
122 
123    unsigned param_idx = 0;
124 
125    nir_deref_instr *ret_deref = NULL;
126    struct vtn_type *ret_type = vtn_callee->type->return_type;
127    if (ret_type->base_type != vtn_base_type_void) {
128       nir_variable *ret_tmp =
129          nir_local_variable_create(b->nb.impl,
130                                    glsl_get_bare_type(ret_type->type),
131                                    "return_tmp");
132       ret_deref = nir_build_deref_var(&b->nb, ret_tmp);
133       call->params[param_idx++] = nir_src_for_ssa(&ret_deref->dest.ssa);
134    }
135 
136    for (unsigned i = 0; i < vtn_callee->type->length; i++) {
137       vtn_ssa_value_add_to_call_params(b, vtn_ssa_value(b, w[4 + i]),
138                                        call, &param_idx);
139    }
140    assert(param_idx == call->num_params);
141 
142    nir_builder_instr_insert(&b->nb, &call->instr);
143 
144    if (ret_type->base_type == vtn_base_type_void) {
145       vtn_push_value(b, w[2], vtn_value_type_undef);
146    } else {
147       vtn_push_ssa_value(b, w[2], vtn_local_load(b, ret_deref, 0));
148    }
149 }
150 
151 static void
function_decoration_cb(struct vtn_builder * b,struct vtn_value * val,int member,const struct vtn_decoration * dec,void * void_func)152 function_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
153                        const struct vtn_decoration *dec, void *void_func)
154 {
155    struct vtn_function *func = void_func;
156 
157    switch (dec->decoration) {
158    case SpvDecorationLinkageAttributes: {
159       unsigned name_words;
160       const char *name =
161          vtn_string_literal(b, dec->operands, dec->num_operands, &name_words);
162       vtn_fail_if(name_words >= dec->num_operands,
163                   "Malformed LinkageAttributes decoration");
164       (void)name; /* TODO: What is this? */
165       func->linkage = dec->operands[name_words];
166       break;
167    }
168 
169    default:
170       break;
171    }
172 }
173 
174 static bool
vtn_cfg_handle_prepass_instruction(struct vtn_builder * b,SpvOp opcode,const uint32_t * w,unsigned count)175 vtn_cfg_handle_prepass_instruction(struct vtn_builder *b, SpvOp opcode,
176                                    const uint32_t *w, unsigned count)
177 {
178    switch (opcode) {
179    case SpvOpFunction: {
180       vtn_assert(b->func == NULL);
181       b->func = rzalloc(b, struct vtn_function);
182 
183       b->func->node.type = vtn_cf_node_type_function;
184       b->func->node.parent = NULL;
185       list_inithead(&b->func->body);
186       b->func->linkage = SpvLinkageTypeMax;
187       b->func->control = w[3];
188 
189       UNUSED const struct glsl_type *result_type = vtn_get_type(b, w[1])->type;
190       struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_function);
191       val->func = b->func;
192 
193       vtn_foreach_decoration(b, val, function_decoration_cb, b->func);
194 
195       b->func->type = vtn_get_type(b, w[4]);
196       const struct vtn_type *func_type = b->func->type;
197 
198       vtn_assert(func_type->return_type->type == result_type);
199 
200       nir_function *func =
201          nir_function_create(b->shader, ralloc_strdup(b->shader, val->name));
202 
203       unsigned num_params = 0;
204       for (unsigned i = 0; i < func_type->length; i++)
205          num_params += glsl_type_count_function_params(func_type->params[i]->type);
206 
207       /* Add one parameter for the function return value */
208       if (func_type->return_type->base_type != vtn_base_type_void)
209          num_params++;
210 
211       func->num_params = num_params;
212       func->params = ralloc_array(b->shader, nir_parameter, num_params);
213 
214       unsigned idx = 0;
215       if (func_type->return_type->base_type != vtn_base_type_void) {
216          nir_address_format addr_format =
217             vtn_mode_to_address_format(b, vtn_variable_mode_function);
218          /* The return value is a regular pointer */
219          func->params[idx++] = (nir_parameter) {
220             .num_components = nir_address_format_num_components(addr_format),
221             .bit_size = nir_address_format_bit_size(addr_format),
222          };
223       }
224 
225       for (unsigned i = 0; i < func_type->length; i++)
226          glsl_type_add_to_function_params(func_type->params[i]->type, func, &idx);
227       assert(idx == num_params);
228 
229       b->func->nir_func = func;
230 
231       /* Set up a nir_function_impl and the builder so we can load arguments
232        * directly in our OpFunctionParameter handler.
233        */
234       nir_function_impl *impl = nir_function_impl_create(func);
235       nir_builder_init(&b->nb, impl);
236       b->nb.cursor = nir_before_cf_list(&impl->body);
237       b->nb.exact = b->exact;
238 
239       b->func_param_idx = 0;
240 
241       /* The return value is the first parameter */
242       if (func_type->return_type->base_type != vtn_base_type_void)
243          b->func_param_idx++;
244       break;
245    }
246 
247    case SpvOpFunctionEnd:
248       b->func->end = w;
249       if (b->func->start_block == NULL) {
250          vtn_fail_if(b->func->linkage != SpvLinkageTypeImport,
251                      "A function declaration (an OpFunction with no basic "
252                      "blocks), must have a Linkage Attributes Decoration "
253                      "with the Import Linkage Type.");
254 
255          /* In this case, the function didn't have any actual blocks.  It's
256           * just a prototype so delete the function_impl.
257           */
258          b->func->nir_func->impl = NULL;
259       } else {
260          vtn_fail_if(b->func->linkage == SpvLinkageTypeImport,
261                      "A function definition (an OpFunction with basic blocks) "
262                      "cannot be decorated with the Import Linkage Type.");
263       }
264       b->func = NULL;
265       break;
266 
267    case SpvOpFunctionParameter: {
268       vtn_assert(b->func_param_idx < b->func->nir_func->num_params);
269       struct vtn_type *type = vtn_get_type(b, w[1]);
270       struct vtn_ssa_value *value = vtn_create_ssa_value(b, type->type);
271       vtn_ssa_value_load_function_param(b, value, &b->func_param_idx);
272       vtn_push_ssa_value(b, w[2], value);
273       break;
274    }
275 
276    case SpvOpLabel: {
277       vtn_assert(b->block == NULL);
278       b->block = rzalloc(b, struct vtn_block);
279       b->block->node.type = vtn_cf_node_type_block;
280       b->block->label = w;
281       vtn_push_value(b, w[1], vtn_value_type_block)->block = b->block;
282 
283       if (b->func->start_block == NULL) {
284          /* This is the first block encountered for this function.  In this
285           * case, we set the start block and add it to the list of
286           * implemented functions that we'll walk later.
287           */
288          b->func->start_block = b->block;
289          list_addtail(&b->func->node.link, &b->functions);
290       }
291       break;
292    }
293 
294    case SpvOpSelectionMerge:
295    case SpvOpLoopMerge:
296       vtn_assert(b->block && b->block->merge == NULL);
297       b->block->merge = w;
298       break;
299 
300    case SpvOpBranch:
301    case SpvOpBranchConditional:
302    case SpvOpSwitch:
303    case SpvOpKill:
304    case SpvOpTerminateInvocation:
305    case SpvOpIgnoreIntersectionKHR:
306    case SpvOpTerminateRayKHR:
307    case SpvOpReturn:
308    case SpvOpReturnValue:
309    case SpvOpUnreachable:
310       vtn_assert(b->block && b->block->branch == NULL);
311       b->block->branch = w;
312       b->block = NULL;
313       break;
314 
315    default:
316       /* Continue on as per normal */
317       return true;
318    }
319 
320    return true;
321 }
322 
323 /* This function performs a depth-first search of the cases and puts them
324  * in fall-through order.
325  */
326 static void
vtn_order_case(struct vtn_switch * swtch,struct vtn_case * cse)327 vtn_order_case(struct vtn_switch *swtch, struct vtn_case *cse)
328 {
329    if (cse->visited)
330       return;
331 
332    cse->visited = true;
333 
334    list_del(&cse->node.link);
335 
336    if (cse->fallthrough) {
337       vtn_order_case(swtch, cse->fallthrough);
338 
339       /* If we have a fall-through, place this case right before the case it
340        * falls through to.  This ensures that fallthroughs come one after
341        * the other.  These two can never get separated because that would
342        * imply something else falling through to the same case.  Also, this
343        * can't break ordering because the DFS ensures that this case is
344        * visited before anything that falls through to it.
345        */
346       list_addtail(&cse->node.link, &cse->fallthrough->node.link);
347    } else {
348       list_add(&cse->node.link, &swtch->cases);
349    }
350 }
351 
352 static void
vtn_switch_order_cases(struct vtn_switch * swtch)353 vtn_switch_order_cases(struct vtn_switch *swtch)
354 {
355    struct list_head cases;
356    list_replace(&swtch->cases, &cases);
357    list_inithead(&swtch->cases);
358    while (!list_is_empty(&cases)) {
359       struct vtn_case *cse =
360          list_first_entry(&cases, struct vtn_case, node.link);
361       vtn_order_case(swtch, cse);
362    }
363 }
364 
365 static void
vtn_block_set_merge_cf_node(struct vtn_builder * b,struct vtn_block * block,struct vtn_cf_node * cf_node)366 vtn_block_set_merge_cf_node(struct vtn_builder *b, struct vtn_block *block,
367                             struct vtn_cf_node *cf_node)
368 {
369    vtn_fail_if(block->merge_cf_node != NULL,
370                "The merge block declared by a header block cannot be a "
371                "merge block declared by any other header block.");
372 
373    block->merge_cf_node = cf_node;
374 }
375 
376 #define VTN_DECL_CF_NODE_FIND(_type)                        \
377 static inline struct vtn_##_type *                          \
378 vtn_cf_node_find_##_type(struct vtn_cf_node *node)          \
379 {                                                           \
380    while (node && node->type != vtn_cf_node_type_##_type)   \
381       node = node->parent;                                  \
382    return (struct vtn_##_type *)node;                       \
383 }
384 
385 VTN_DECL_CF_NODE_FIND(if)
VTN_DECL_CF_NODE_FIND(loop)386 VTN_DECL_CF_NODE_FIND(loop)
387 VTN_DECL_CF_NODE_FIND(case)
388 VTN_DECL_CF_NODE_FIND(switch)
389 VTN_DECL_CF_NODE_FIND(function)
390 
391 static enum vtn_branch_type
392 vtn_handle_branch(struct vtn_builder *b,
393                   struct vtn_cf_node *cf_parent,
394                   struct vtn_block *target_block)
395 {
396    struct vtn_loop *loop = vtn_cf_node_find_loop(cf_parent);
397 
398    /* Detect a loop back-edge first.  That way none of the code below
399     * accidentally operates on a loop back-edge.
400     */
401    if (loop && target_block == loop->header_block)
402       return vtn_branch_type_loop_back_edge;
403 
404    /* Try to detect fall-through */
405    if (target_block->switch_case) {
406       /* When it comes to handling switch cases, we can break calls to
407        * vtn_handle_branch into two cases: calls from within a case construct
408        * and calls for the jump to each case construct.  In the second case,
409        * cf_parent is the vtn_switch itself and vtn_cf_node_find_case() will
410        * return the outer switch case in which this switch is contained.  It's
411        * fine if the target block is a switch case from an outer switch as
412        * long as it is also the switch break for this switch.
413        */
414       struct vtn_case *switch_case = vtn_cf_node_find_case(cf_parent);
415 
416       /* This doesn't get called for the OpSwitch */
417       vtn_fail_if(switch_case == NULL,
418                   "A switch case can only be entered through an OpSwitch or "
419                   "falling through from another switch case.");
420 
421       /* Because block->switch_case is only set on the entry block for a given
422        * switch case, we only ever get here if we're jumping to the start of a
423        * switch case.  It's possible, however, that a switch case could jump
424        * to itself via a back-edge.  That *should* get caught by the loop
425        * handling case above but if we have a back edge without a loop merge,
426        * we could en up here.
427        */
428       vtn_fail_if(target_block->switch_case == switch_case,
429                   "A switch cannot fall-through to itself.  Likely, there is "
430                   "a back-edge which is not to a loop header.");
431 
432       vtn_fail_if(target_block->switch_case->node.parent !=
433                      switch_case->node.parent,
434                   "A switch case fall-through must come from the same "
435                   "OpSwitch construct");
436 
437       vtn_fail_if(switch_case->fallthrough != NULL &&
438                   switch_case->fallthrough != target_block->switch_case,
439                   "Each case construct can have at most one branch to "
440                   "another case construct");
441 
442       switch_case->fallthrough = target_block->switch_case;
443 
444       /* We don't immediately return vtn_branch_type_switch_fallthrough
445        * because it may also be a loop or switch break for an inner loop or
446        * switch and that takes precedence.
447        */
448    }
449 
450    if (loop && target_block == loop->cont_block)
451       return vtn_branch_type_loop_continue;
452 
453    /* We walk blocks as a breadth-first search on the control-flow construct
454     * tree where, when we find a construct, we add the vtn_cf_node for that
455     * construct and continue iterating at the merge target block (if any).
456     * Therefore, we want merges whose with parent == cf_parent to be treated
457     * as regular branches.  We only want to consider merges if they break out
458     * of the current CF construct.
459     */
460    if (target_block->merge_cf_node != NULL &&
461        target_block->merge_cf_node->parent != cf_parent) {
462       switch (target_block->merge_cf_node->type) {
463       case vtn_cf_node_type_if:
464          for (struct vtn_cf_node *node = cf_parent;
465               node != target_block->merge_cf_node; node = node->parent) {
466             vtn_fail_if(node == NULL || node->type != vtn_cf_node_type_if,
467                         "Branching to the merge block of a selection "
468                         "construct can only be used to break out of a "
469                         "selection construct");
470 
471             struct vtn_if *if_stmt = vtn_cf_node_as_if(node);
472 
473             /* This should be guaranteed by our iteration */
474             assert(if_stmt->merge_block != target_block);
475 
476             vtn_fail_if(if_stmt->merge_block != NULL,
477                         "Branching to the merge block of a selection "
478                         "construct can only be used to break out of the "
479                         "inner most nested selection level");
480          }
481          return vtn_branch_type_if_merge;
482 
483       case vtn_cf_node_type_loop:
484          vtn_fail_if(target_block->merge_cf_node != &loop->node,
485                      "Loop breaks can only break out of the inner most "
486                      "nested loop level");
487          return vtn_branch_type_loop_break;
488 
489       case vtn_cf_node_type_switch: {
490          struct vtn_switch *swtch = vtn_cf_node_find_switch(cf_parent);
491          vtn_fail_if(target_block->merge_cf_node != &swtch->node,
492                      "Switch breaks can only break out of the inner most "
493                      "nested switch level");
494          return vtn_branch_type_switch_break;
495       }
496 
497       default:
498          unreachable("Invalid CF node type for a merge");
499       }
500    }
501 
502    if (target_block->switch_case)
503       return vtn_branch_type_switch_fallthrough;
504 
505    return vtn_branch_type_none;
506 }
507 
508 struct vtn_cfg_work_item {
509    struct list_head link;
510 
511    struct vtn_cf_node *cf_parent;
512    struct list_head *cf_list;
513    struct vtn_block *start_block;
514 };
515 
516 static void
vtn_add_cfg_work_item(struct vtn_builder * b,struct list_head * work_list,struct vtn_cf_node * cf_parent,struct list_head * cf_list,struct vtn_block * start_block)517 vtn_add_cfg_work_item(struct vtn_builder *b,
518                       struct list_head *work_list,
519                       struct vtn_cf_node *cf_parent,
520                       struct list_head *cf_list,
521                       struct vtn_block *start_block)
522 {
523    struct vtn_cfg_work_item *work = ralloc(b, struct vtn_cfg_work_item);
524    work->cf_parent = cf_parent;
525    work->cf_list = cf_list;
526    work->start_block = start_block;
527    list_addtail(&work->link, work_list);
528 }
529 
530 /* returns the default block */
531 static void
vtn_parse_switch(struct vtn_builder * b,struct vtn_switch * swtch,const uint32_t * branch,struct list_head * case_list)532 vtn_parse_switch(struct vtn_builder *b,
533                  struct vtn_switch *swtch,
534                  const uint32_t *branch,
535                  struct list_head *case_list)
536 {
537    const uint32_t *branch_end = branch + (branch[0] >> SpvWordCountShift);
538 
539    struct vtn_value *sel_val = vtn_untyped_value(b, branch[1]);
540    vtn_fail_if(!sel_val->type ||
541                sel_val->type->base_type != vtn_base_type_scalar,
542                "Selector of OpSwitch must have a type of OpTypeInt");
543 
544    nir_alu_type sel_type =
545       nir_get_nir_type_for_glsl_type(sel_val->type->type);
546    vtn_fail_if(nir_alu_type_get_base_type(sel_type) != nir_type_int &&
547                nir_alu_type_get_base_type(sel_type) != nir_type_uint,
548                "Selector of OpSwitch must have a type of OpTypeInt");
549 
550    struct hash_table *block_to_case = _mesa_pointer_hash_table_create(b);
551 
552    bool is_default = true;
553    const unsigned bitsize = nir_alu_type_get_type_size(sel_type);
554    for (const uint32_t *w = branch + 2; w < branch_end;) {
555       uint64_t literal = 0;
556       if (!is_default) {
557          if (bitsize <= 32) {
558             literal = *(w++);
559          } else {
560             assert(bitsize == 64);
561             literal = vtn_u64_literal(w);
562             w += 2;
563          }
564       }
565       struct vtn_block *case_block = vtn_block(b, *(w++));
566 
567       struct hash_entry *case_entry =
568          _mesa_hash_table_search(block_to_case, case_block);
569 
570       struct vtn_case *cse;
571       if (case_entry) {
572          cse = case_entry->data;
573       } else {
574          cse = rzalloc(b, struct vtn_case);
575 
576          cse->node.type = vtn_cf_node_type_case;
577          cse->node.parent = swtch ? &swtch->node : NULL;
578          cse->block = case_block;
579          list_inithead(&cse->body);
580          util_dynarray_init(&cse->values, b);
581 
582          list_addtail(&cse->node.link, case_list);
583          _mesa_hash_table_insert(block_to_case, case_block, cse);
584       }
585 
586       if (is_default) {
587          cse->is_default = true;
588       } else {
589          util_dynarray_append(&cse->values, uint64_t, literal);
590       }
591 
592       is_default = false;
593    }
594 
595    _mesa_hash_table_destroy(block_to_case, NULL);
596 }
597 
598 /* Processes a block and returns the next block to process or NULL if we've
599  * reached the end of the construct.
600  */
601 static struct vtn_block *
vtn_process_block(struct vtn_builder * b,struct list_head * work_list,struct vtn_cf_node * cf_parent,struct list_head * cf_list,struct vtn_block * block)602 vtn_process_block(struct vtn_builder *b,
603                   struct list_head *work_list,
604                   struct vtn_cf_node *cf_parent,
605                   struct list_head *cf_list,
606                   struct vtn_block *block)
607 {
608    if (!list_is_empty(cf_list)) {
609       /* vtn_process_block() acts like an iterator: it processes the given
610        * block and then returns the next block to process.  For a given
611        * control-flow construct, vtn_build_cfg() calls vtn_process_block()
612        * repeatedly until it finally returns NULL.  Therefore, we know that
613        * the only blocks on which vtn_process_block() can be called are either
614        * the first block in a construct or a block that vtn_process_block()
615        * returned for the current construct.  If cf_list is empty then we know
616        * that we're processing the first block in the construct and we have to
617        * add it to the list.
618        *
619        * If cf_list is not empty, then it must be the block returned by the
620        * previous call to vtn_process_block().  We know a priori that
621        * vtn_process_block only returns either normal branches
622        * (vtn_branch_type_none) or merge target blocks.
623        */
624       switch (vtn_handle_branch(b, cf_parent, block)) {
625       case vtn_branch_type_none:
626          /* For normal branches, we want to process them and add them to the
627           * current construct.  Merge target blocks also look like normal
628           * branches from the perspective of this construct.  See also
629           * vtn_handle_branch().
630           */
631          break;
632 
633       case vtn_branch_type_loop_continue:
634       case vtn_branch_type_switch_fallthrough:
635          /* The two cases where we can get early exits from a construct that
636           * are not to that construct's merge target are loop continues and
637           * switch fall-throughs.  In these cases, we need to break out of the
638           * current construct by returning NULL.
639           */
640          return NULL;
641 
642       default:
643          /* The only way we can get here is if something was used as two kinds
644           * of merges at the same time and that's illegal.
645           */
646          vtn_fail("A block was used as a merge target from two or more "
647                   "structured control-flow constructs");
648       }
649    }
650 
651    /* Once a block has been processed, it is placed into and the list link
652     * will point to something non-null.  If we see a node we've already
653     * processed here, it either exists in multiple functions or it's an
654     * invalid back-edge.
655     */
656    if (block->node.parent != NULL) {
657       vtn_fail_if(vtn_cf_node_find_function(&block->node) !=
658                   vtn_cf_node_find_function(cf_parent),
659                   "A block cannot exist in two functions at the "
660                   "same time");
661 
662       vtn_fail("Invalid back or cross-edge in the CFG");
663    }
664 
665    if (block->merge && (*block->merge & SpvOpCodeMask) == SpvOpLoopMerge &&
666        block->loop == NULL) {
667       vtn_fail_if((*block->branch & SpvOpCodeMask) != SpvOpBranch &&
668                   (*block->branch & SpvOpCodeMask) != SpvOpBranchConditional,
669                   "An OpLoopMerge instruction must immediately precede "
670                   "either an OpBranch or OpBranchConditional instruction.");
671 
672       struct vtn_loop *loop = rzalloc(b, struct vtn_loop);
673 
674       loop->node.type = vtn_cf_node_type_loop;
675       loop->node.parent = cf_parent;
676       list_inithead(&loop->body);
677       list_inithead(&loop->cont_body);
678       loop->header_block = block;
679       loop->break_block = vtn_block(b, block->merge[1]);
680       loop->cont_block = vtn_block(b, block->merge[2]);
681       loop->control = block->merge[3];
682 
683       list_addtail(&loop->node.link, cf_list);
684       block->loop = loop;
685 
686       /* Note: The work item for the main loop body will start with the
687        * current block as its start block.  If we weren't careful, we would
688        * get here again and end up in an infinite loop.  This is why we set
689        * block->loop above and check for it before creating one.  This way,
690        * we only create the loop once and the second iteration that tries to
691        * handle this loop goes to the cases below and gets handled as a
692        * regular block.
693        */
694       vtn_add_cfg_work_item(b, work_list, &loop->node,
695                             &loop->body, loop->header_block);
696 
697       /* For continue targets, SPIR-V guarantees the following:
698        *
699        *  - the Continue Target must dominate the back-edge block
700        *  - the back-edge block must post dominate the Continue Target
701        *
702        * If the header block is the same as the continue target, this
703        * condition is trivially satisfied and there is no real continue
704        * section.
705        */
706       if (loop->cont_block != loop->header_block) {
707          vtn_add_cfg_work_item(b, work_list, &loop->node,
708                                &loop->cont_body, loop->cont_block);
709       }
710 
711       vtn_block_set_merge_cf_node(b, loop->break_block, &loop->node);
712 
713       return loop->break_block;
714    }
715 
716    /* Add the block to the CF list */
717    block->node.parent = cf_parent;
718    list_addtail(&block->node.link, cf_list);
719 
720    switch (*block->branch & SpvOpCodeMask) {
721    case SpvOpBranch: {
722       struct vtn_block *branch_block = vtn_block(b, block->branch[1]);
723 
724       block->branch_type = vtn_handle_branch(b, cf_parent, branch_block);
725 
726       if (block->branch_type == vtn_branch_type_none)
727          return branch_block;
728       else
729          return NULL;
730    }
731 
732    case SpvOpReturn:
733    case SpvOpReturnValue:
734       block->branch_type = vtn_branch_type_return;
735       return NULL;
736 
737    case SpvOpKill:
738       block->branch_type = vtn_branch_type_discard;
739       return NULL;
740 
741    case SpvOpTerminateInvocation:
742       block->branch_type = vtn_branch_type_terminate_invocation;
743       return NULL;
744 
745    case SpvOpIgnoreIntersectionKHR:
746       block->branch_type = vtn_branch_type_ignore_intersection;
747       return NULL;
748 
749    case SpvOpTerminateRayKHR:
750       block->branch_type = vtn_branch_type_terminate_ray;
751       return NULL;
752 
753    case SpvOpBranchConditional: {
754       struct vtn_value *cond_val = vtn_untyped_value(b, block->branch[1]);
755       vtn_fail_if(!cond_val->type ||
756                   cond_val->type->base_type != vtn_base_type_scalar ||
757                   cond_val->type->type != glsl_bool_type(),
758                   "Condition must be a Boolean type scalar");
759 
760       struct vtn_if *if_stmt = rzalloc(b, struct vtn_if);
761 
762       if_stmt->node.type = vtn_cf_node_type_if;
763       if_stmt->node.parent = cf_parent;
764       if_stmt->header_block = block;
765       list_inithead(&if_stmt->then_body);
766       list_inithead(&if_stmt->else_body);
767 
768       list_addtail(&if_stmt->node.link, cf_list);
769 
770       if (block->merge &&
771           (*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge) {
772          /* We may not always have a merge block and that merge doesn't
773           * technically have to be an OpSelectionMerge.  We could have a block
774           * with an OpLoopMerge which ends in an OpBranchConditional.
775           */
776          if_stmt->merge_block = vtn_block(b, block->merge[1]);
777          vtn_block_set_merge_cf_node(b, if_stmt->merge_block, &if_stmt->node);
778 
779          if_stmt->control = block->merge[2];
780       }
781 
782       struct vtn_block *then_block = vtn_block(b, block->branch[2]);
783       if_stmt->then_type = vtn_handle_branch(b, &if_stmt->node, then_block);
784       if (if_stmt->then_type == vtn_branch_type_none) {
785          vtn_add_cfg_work_item(b, work_list, &if_stmt->node,
786                                &if_stmt->then_body, then_block);
787       }
788 
789       struct vtn_block *else_block = vtn_block(b, block->branch[3]);
790       if (then_block != else_block) {
791          if_stmt->else_type = vtn_handle_branch(b, &if_stmt->node, else_block);
792          if (if_stmt->else_type == vtn_branch_type_none) {
793             vtn_add_cfg_work_item(b, work_list, &if_stmt->node,
794                                   &if_stmt->else_body, else_block);
795          }
796       }
797 
798       return if_stmt->merge_block;
799    }
800 
801    case SpvOpSwitch: {
802       struct vtn_switch *swtch = rzalloc(b, struct vtn_switch);
803 
804       swtch->node.type = vtn_cf_node_type_switch;
805       swtch->node.parent = cf_parent;
806       swtch->selector = block->branch[1];
807       list_inithead(&swtch->cases);
808 
809       list_addtail(&swtch->node.link, cf_list);
810 
811       /* We may not always have a merge block */
812       if (block->merge) {
813          vtn_fail_if((*block->merge & SpvOpCodeMask) != SpvOpSelectionMerge,
814                      "An OpLoopMerge instruction must immediately precede "
815                      "either an OpBranch or OpBranchConditional "
816                      "instruction.");
817          swtch->break_block = vtn_block(b, block->merge[1]);
818          vtn_block_set_merge_cf_node(b, swtch->break_block, &swtch->node);
819       }
820 
821       /* First, we go through and record all of the cases. */
822       vtn_parse_switch(b, swtch, block->branch, &swtch->cases);
823 
824       /* Gather the branch types for the switch */
825       vtn_foreach_cf_node(case_node, &swtch->cases) {
826          struct vtn_case *cse = vtn_cf_node_as_case(case_node);
827 
828          cse->type = vtn_handle_branch(b, &swtch->node, cse->block);
829          switch (cse->type) {
830          case vtn_branch_type_none:
831             /* This is a "real" cases which has stuff in it */
832             vtn_fail_if(cse->block->switch_case != NULL,
833                         "OpSwitch has a case which is also in another "
834                         "OpSwitch construct");
835             cse->block->switch_case = cse;
836             vtn_add_cfg_work_item(b, work_list, &cse->node,
837                                   &cse->body, cse->block);
838             break;
839 
840          case vtn_branch_type_switch_break:
841          case vtn_branch_type_loop_break:
842          case vtn_branch_type_loop_continue:
843             /* Switch breaks as well as loop breaks and continues can be
844              * used to break out of a switch construct or as direct targets
845              * of the OpSwitch.
846              */
847             break;
848 
849          default:
850             vtn_fail("Target of OpSwitch is not a valid structured exit "
851                      "from the switch construct.");
852          }
853       }
854 
855       return swtch->break_block;
856    }
857 
858    case SpvOpUnreachable:
859       return NULL;
860 
861    default:
862       vtn_fail("Block did not end with a valid branch instruction");
863    }
864 }
865 
866 void
vtn_build_cfg(struct vtn_builder * b,const uint32_t * words,const uint32_t * end)867 vtn_build_cfg(struct vtn_builder *b, const uint32_t *words, const uint32_t *end)
868 {
869    vtn_foreach_instruction(b, words, end,
870                            vtn_cfg_handle_prepass_instruction);
871 
872    if (b->shader->info.stage == MESA_SHADER_KERNEL)
873       return;
874 
875    vtn_foreach_cf_node(func_node, &b->functions) {
876       struct vtn_function *func = vtn_cf_node_as_function(func_node);
877 
878       /* We build the CFG for each function by doing a breadth-first search on
879        * the control-flow graph.  We keep track of our state using a worklist.
880        * Doing a BFS ensures that we visit each structured control-flow
881        * construct and its merge node before we visit the stuff inside the
882        * construct.
883        */
884       struct list_head work_list;
885       list_inithead(&work_list);
886       vtn_add_cfg_work_item(b, &work_list, &func->node, &func->body,
887                             func->start_block);
888 
889       while (!list_is_empty(&work_list)) {
890          struct vtn_cfg_work_item *work =
891             list_first_entry(&work_list, struct vtn_cfg_work_item, link);
892          list_del(&work->link);
893 
894          for (struct vtn_block *block = work->start_block; block; ) {
895             block = vtn_process_block(b, &work_list, work->cf_parent,
896                                       work->cf_list, block);
897          }
898       }
899    }
900 }
901 
902 static bool
vtn_handle_phis_first_pass(struct vtn_builder * b,SpvOp opcode,const uint32_t * w,unsigned count)903 vtn_handle_phis_first_pass(struct vtn_builder *b, SpvOp opcode,
904                            const uint32_t *w, unsigned count)
905 {
906    if (opcode == SpvOpLabel)
907       return true; /* Nothing to do */
908 
909    /* If this isn't a phi node, stop. */
910    if (opcode != SpvOpPhi)
911       return false;
912 
913    /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
914     * For each phi, we create a variable with the appropreate type and
915     * do a load from that variable.  Then, in a second pass, we add
916     * stores to that variable to each of the predecessor blocks.
917     *
918     * We could do something more intelligent here.  However, in order to
919     * handle loops and things properly, we really need dominance
920     * information.  It would end up basically being the into-SSA
921     * algorithm all over again.  It's easier if we just let
922     * lower_vars_to_ssa do that for us instead of repeating it here.
923     */
924    struct vtn_type *type = vtn_get_type(b, w[1]);
925    nir_variable *phi_var =
926       nir_local_variable_create(b->nb.impl, type->type, "phi");
927 
928    struct vtn_value *phi_val = vtn_untyped_value(b, w[2]);
929    if (vtn_value_is_relaxed_precision(b, phi_val))
930       phi_var->data.precision = GLSL_PRECISION_MEDIUM;
931 
932    _mesa_hash_table_insert(b->phi_table, w, phi_var);
933 
934    vtn_push_ssa_value(b, w[2],
935       vtn_local_load(b, nir_build_deref_var(&b->nb, phi_var), 0));
936 
937    return true;
938 }
939 
940 static bool
vtn_handle_phi_second_pass(struct vtn_builder * b,SpvOp opcode,const uint32_t * w,unsigned count)941 vtn_handle_phi_second_pass(struct vtn_builder *b, SpvOp opcode,
942                            const uint32_t *w, unsigned count)
943 {
944    if (opcode != SpvOpPhi)
945       return true;
946 
947    struct hash_entry *phi_entry = _mesa_hash_table_search(b->phi_table, w);
948 
949    /* It's possible that this phi is in an unreachable block in which case it
950     * may never have been emitted and therefore may not be in the hash table.
951     * In this case, there's no var for it and it's safe to just bail.
952     */
953    if (phi_entry == NULL)
954       return true;
955 
956    nir_variable *phi_var = phi_entry->data;
957 
958    for (unsigned i = 3; i < count; i += 2) {
959       struct vtn_block *pred = vtn_block(b, w[i + 1]);
960 
961       /* If block does not have end_nop, that is because it is an unreacheable
962        * block, and hence it is not worth to handle it */
963       if (!pred->end_nop)
964          continue;
965 
966       b->nb.cursor = nir_after_instr(&pred->end_nop->instr);
967 
968       struct vtn_ssa_value *src = vtn_ssa_value(b, w[i]);
969 
970       vtn_local_store(b, src, nir_build_deref_var(&b->nb, phi_var), 0);
971    }
972 
973    return true;
974 }
975 
976 static void
vtn_emit_branch(struct vtn_builder * b,enum vtn_branch_type branch_type,nir_variable * switch_fall_var,bool * has_switch_break)977 vtn_emit_branch(struct vtn_builder *b, enum vtn_branch_type branch_type,
978                 nir_variable *switch_fall_var, bool *has_switch_break)
979 {
980    switch (branch_type) {
981    case vtn_branch_type_if_merge:
982       break; /* Nothing to do */
983    case vtn_branch_type_switch_break:
984       nir_store_var(&b->nb, switch_fall_var, nir_imm_false(&b->nb), 1);
985       *has_switch_break = true;
986       break;
987    case vtn_branch_type_switch_fallthrough:
988       break; /* Nothing to do */
989    case vtn_branch_type_loop_break:
990       nir_jump(&b->nb, nir_jump_break);
991       break;
992    case vtn_branch_type_loop_continue:
993       nir_jump(&b->nb, nir_jump_continue);
994       break;
995    case vtn_branch_type_loop_back_edge:
996       break;
997    case vtn_branch_type_return:
998       nir_jump(&b->nb, nir_jump_return);
999       break;
1000    case vtn_branch_type_discard:
1001       if (b->convert_discard_to_demote)
1002          nir_demote(&b->nb);
1003       else
1004          nir_discard(&b->nb);
1005       break;
1006    case vtn_branch_type_terminate_invocation:
1007       nir_terminate(&b->nb);
1008       break;
1009    case vtn_branch_type_ignore_intersection:
1010       nir_ignore_ray_intersection(&b->nb);
1011       nir_jump(&b->nb, nir_jump_halt);
1012       break;
1013    case vtn_branch_type_terminate_ray:
1014       nir_terminate_ray(&b->nb);
1015       nir_jump(&b->nb, nir_jump_halt);
1016       break;
1017    default:
1018       vtn_fail("Invalid branch type");
1019    }
1020 }
1021 
1022 static nir_ssa_def *
vtn_switch_case_condition(struct vtn_builder * b,struct vtn_switch * swtch,nir_ssa_def * sel,struct vtn_case * cse)1023 vtn_switch_case_condition(struct vtn_builder *b, struct vtn_switch *swtch,
1024                           nir_ssa_def *sel, struct vtn_case *cse)
1025 {
1026    if (cse->is_default) {
1027       nir_ssa_def *any = nir_imm_false(&b->nb);
1028       vtn_foreach_cf_node(other_node, &swtch->cases) {
1029          struct vtn_case *other = vtn_cf_node_as_case(other_node);
1030          if (other->is_default)
1031             continue;
1032 
1033          any = nir_ior(&b->nb, any,
1034                        vtn_switch_case_condition(b, swtch, sel, other));
1035       }
1036       return nir_inot(&b->nb, any);
1037    } else {
1038       nir_ssa_def *cond = nir_imm_false(&b->nb);
1039       util_dynarray_foreach(&cse->values, uint64_t, val)
1040          cond = nir_ior(&b->nb, cond, nir_ieq_imm(&b->nb, sel, *val));
1041       return cond;
1042    }
1043 }
1044 
1045 static nir_loop_control
vtn_loop_control(struct vtn_builder * b,struct vtn_loop * vtn_loop)1046 vtn_loop_control(struct vtn_builder *b, struct vtn_loop *vtn_loop)
1047 {
1048    if (vtn_loop->control == SpvLoopControlMaskNone)
1049       return nir_loop_control_none;
1050    else if (vtn_loop->control & SpvLoopControlDontUnrollMask)
1051       return nir_loop_control_dont_unroll;
1052    else if (vtn_loop->control & SpvLoopControlUnrollMask)
1053       return nir_loop_control_unroll;
1054    else if (vtn_loop->control & SpvLoopControlDependencyInfiniteMask ||
1055             vtn_loop->control & SpvLoopControlDependencyLengthMask ||
1056             vtn_loop->control & SpvLoopControlMinIterationsMask ||
1057             vtn_loop->control & SpvLoopControlMaxIterationsMask ||
1058             vtn_loop->control & SpvLoopControlIterationMultipleMask ||
1059             vtn_loop->control & SpvLoopControlPeelCountMask ||
1060             vtn_loop->control & SpvLoopControlPartialCountMask) {
1061       /* We do not do anything special with these yet. */
1062       return nir_loop_control_none;
1063    } else {
1064       vtn_fail("Invalid loop control");
1065    }
1066 }
1067 
1068 static nir_selection_control
vtn_selection_control(struct vtn_builder * b,struct vtn_if * vtn_if)1069 vtn_selection_control(struct vtn_builder *b, struct vtn_if *vtn_if)
1070 {
1071    if (vtn_if->control == SpvSelectionControlMaskNone)
1072       return nir_selection_control_none;
1073    else if (vtn_if->control & SpvSelectionControlDontFlattenMask)
1074       return nir_selection_control_dont_flatten;
1075    else if (vtn_if->control & SpvSelectionControlFlattenMask)
1076       return nir_selection_control_flatten;
1077    else
1078       vtn_fail("Invalid selection control");
1079 }
1080 
1081 static void
vtn_emit_ret_store(struct vtn_builder * b,struct vtn_block * block)1082 vtn_emit_ret_store(struct vtn_builder *b, struct vtn_block *block)
1083 {
1084    if ((*block->branch & SpvOpCodeMask) != SpvOpReturnValue)
1085       return;
1086 
1087    vtn_fail_if(b->func->type->return_type->base_type == vtn_base_type_void,
1088                "Return with a value from a function returning void");
1089    struct vtn_ssa_value *src = vtn_ssa_value(b, block->branch[1]);
1090    const struct glsl_type *ret_type =
1091       glsl_get_bare_type(b->func->type->return_type->type);
1092    nir_deref_instr *ret_deref =
1093       nir_build_deref_cast(&b->nb, nir_load_param(&b->nb, 0),
1094                            nir_var_function_temp, ret_type, 0);
1095    vtn_local_store(b, src, ret_deref, 0);
1096 }
1097 
1098 static void
vtn_emit_cf_list_structured(struct vtn_builder * b,struct list_head * cf_list,nir_variable * switch_fall_var,bool * has_switch_break,vtn_instruction_handler handler)1099 vtn_emit_cf_list_structured(struct vtn_builder *b, struct list_head *cf_list,
1100                             nir_variable *switch_fall_var,
1101                             bool *has_switch_break,
1102                             vtn_instruction_handler handler)
1103 {
1104    vtn_foreach_cf_node(node, cf_list) {
1105       switch (node->type) {
1106       case vtn_cf_node_type_block: {
1107          struct vtn_block *block = vtn_cf_node_as_block(node);
1108 
1109          const uint32_t *block_start = block->label;
1110          const uint32_t *block_end = block->merge ? block->merge :
1111                                                     block->branch;
1112 
1113          block_start = vtn_foreach_instruction(b, block_start, block_end,
1114                                                vtn_handle_phis_first_pass);
1115 
1116          vtn_foreach_instruction(b, block_start, block_end, handler);
1117 
1118          block->end_nop = nir_nop(&b->nb);
1119 
1120          vtn_emit_ret_store(b, block);
1121 
1122          if (block->branch_type != vtn_branch_type_none) {
1123             vtn_emit_branch(b, block->branch_type,
1124                             switch_fall_var, has_switch_break);
1125             return;
1126          }
1127 
1128          break;
1129       }
1130 
1131       case vtn_cf_node_type_if: {
1132          struct vtn_if *vtn_if = vtn_cf_node_as_if(node);
1133          const uint32_t *branch = vtn_if->header_block->branch;
1134          vtn_assert((branch[0] & SpvOpCodeMask) == SpvOpBranchConditional);
1135 
1136          bool sw_break = false;
1137          /* If both branches are the same, just emit the first block, which is
1138           * the only one we filled when building the CFG.
1139           */
1140          if (branch[2] == branch[3]) {
1141             if (vtn_if->then_type == vtn_branch_type_none) {
1142                vtn_emit_cf_list_structured(b, &vtn_if->then_body,
1143                                            switch_fall_var, &sw_break, handler);
1144             } else {
1145                vtn_emit_branch(b, vtn_if->then_type, switch_fall_var, &sw_break);
1146             }
1147             break;
1148          }
1149 
1150          nir_if *nif =
1151             nir_push_if(&b->nb, vtn_get_nir_ssa(b, branch[1]));
1152 
1153          nif->control = vtn_selection_control(b, vtn_if);
1154 
1155          if (vtn_if->then_type == vtn_branch_type_none) {
1156             vtn_emit_cf_list_structured(b, &vtn_if->then_body,
1157                                         switch_fall_var, &sw_break, handler);
1158          } else {
1159             vtn_emit_branch(b, vtn_if->then_type, switch_fall_var, &sw_break);
1160          }
1161 
1162          nir_push_else(&b->nb, nif);
1163          if (vtn_if->else_type == vtn_branch_type_none) {
1164             vtn_emit_cf_list_structured(b, &vtn_if->else_body,
1165                                         switch_fall_var, &sw_break, handler);
1166          } else {
1167             vtn_emit_branch(b, vtn_if->else_type, switch_fall_var, &sw_break);
1168          }
1169 
1170          nir_pop_if(&b->nb, nif);
1171 
1172          /* If we encountered a switch break somewhere inside of the if,
1173           * then it would have been handled correctly by calling
1174           * emit_cf_list or emit_branch for the interrior.  However, we
1175           * need to predicate everything following on wether or not we're
1176           * still going.
1177           */
1178          if (sw_break) {
1179             *has_switch_break = true;
1180             nir_push_if(&b->nb, nir_load_var(&b->nb, switch_fall_var));
1181          }
1182          break;
1183       }
1184 
1185       case vtn_cf_node_type_loop: {
1186          struct vtn_loop *vtn_loop = vtn_cf_node_as_loop(node);
1187 
1188          nir_loop *loop = nir_push_loop(&b->nb);
1189          loop->control = vtn_loop_control(b, vtn_loop);
1190 
1191          vtn_emit_cf_list_structured(b, &vtn_loop->body, NULL, NULL, handler);
1192 
1193          if (!list_is_empty(&vtn_loop->cont_body)) {
1194             /* If we have a non-trivial continue body then we need to put
1195              * it at the beginning of the loop with a flag to ensure that
1196              * it doesn't get executed in the first iteration.
1197              */
1198             nir_variable *do_cont =
1199                nir_local_variable_create(b->nb.impl, glsl_bool_type(), "cont");
1200 
1201             b->nb.cursor = nir_before_cf_node(&loop->cf_node);
1202             nir_store_var(&b->nb, do_cont, nir_imm_false(&b->nb), 1);
1203 
1204             b->nb.cursor = nir_before_cf_list(&loop->body);
1205 
1206             nir_if *cont_if =
1207                nir_push_if(&b->nb, nir_load_var(&b->nb, do_cont));
1208 
1209             vtn_emit_cf_list_structured(b, &vtn_loop->cont_body, NULL, NULL,
1210                                         handler);
1211 
1212             nir_pop_if(&b->nb, cont_if);
1213 
1214             nir_store_var(&b->nb, do_cont, nir_imm_true(&b->nb), 1);
1215          }
1216 
1217          nir_pop_loop(&b->nb, loop);
1218          break;
1219       }
1220 
1221       case vtn_cf_node_type_switch: {
1222          struct vtn_switch *vtn_switch = vtn_cf_node_as_switch(node);
1223 
1224          /* Before we can emit anything, we need to sort the list of cases in
1225           * fall-through order.
1226           */
1227          vtn_switch_order_cases(vtn_switch);
1228 
1229          /* First, we create a variable to keep track of whether or not the
1230           * switch is still going at any given point.  Any switch breaks
1231           * will set this variable to false.
1232           */
1233          nir_variable *fall_var =
1234             nir_local_variable_create(b->nb.impl, glsl_bool_type(), "fall");
1235          nir_store_var(&b->nb, fall_var, nir_imm_false(&b->nb), 1);
1236 
1237          nir_ssa_def *sel = vtn_get_nir_ssa(b, vtn_switch->selector);
1238 
1239          /* Now we can walk the list of cases and actually emit code */
1240          vtn_foreach_cf_node(case_node, &vtn_switch->cases) {
1241             struct vtn_case *cse = vtn_cf_node_as_case(case_node);
1242 
1243             /* If this case jumps directly to the break block, we don't have
1244              * to handle the case as the body is empty and doesn't fall
1245              * through.
1246              */
1247             if (cse->block == vtn_switch->break_block)
1248                continue;
1249 
1250             /* Figure out the condition */
1251             nir_ssa_def *cond =
1252                vtn_switch_case_condition(b, vtn_switch, sel, cse);
1253             /* Take fallthrough into account */
1254             cond = nir_ior(&b->nb, cond, nir_load_var(&b->nb, fall_var));
1255 
1256             nir_if *case_if = nir_push_if(&b->nb, cond);
1257 
1258             bool has_break = false;
1259             nir_store_var(&b->nb, fall_var, nir_imm_true(&b->nb), 1);
1260             vtn_emit_cf_list_structured(b, &cse->body, fall_var, &has_break,
1261                                         handler);
1262             (void)has_break; /* We don't care */
1263 
1264             nir_pop_if(&b->nb, case_if);
1265          }
1266 
1267          break;
1268       }
1269 
1270       default:
1271          vtn_fail("Invalid CF node type");
1272       }
1273    }
1274 }
1275 
1276 static struct nir_block *
vtn_new_unstructured_block(struct vtn_builder * b,struct vtn_function * func)1277 vtn_new_unstructured_block(struct vtn_builder *b, struct vtn_function *func)
1278 {
1279    struct nir_block *n = nir_block_create(b->shader);
1280    exec_list_push_tail(&func->nir_func->impl->body, &n->cf_node.node);
1281    n->cf_node.parent = &func->nir_func->impl->cf_node;
1282    return n;
1283 }
1284 
1285 static void
vtn_add_unstructured_block(struct vtn_builder * b,struct vtn_function * func,struct list_head * work_list,struct vtn_block * block)1286 vtn_add_unstructured_block(struct vtn_builder *b,
1287                            struct vtn_function *func,
1288                            struct list_head *work_list,
1289                            struct vtn_block *block)
1290 {
1291    if (!block->block) {
1292       block->block = vtn_new_unstructured_block(b, func);
1293       list_addtail(&block->node.link, work_list);
1294    }
1295 }
1296 
1297 static void
vtn_emit_cf_func_unstructured(struct vtn_builder * b,struct vtn_function * func,vtn_instruction_handler handler)1298 vtn_emit_cf_func_unstructured(struct vtn_builder *b, struct vtn_function *func,
1299                               vtn_instruction_handler handler)
1300 {
1301    struct list_head work_list;
1302    list_inithead(&work_list);
1303 
1304    func->start_block->block = nir_start_block(func->nir_func->impl);
1305    list_addtail(&func->start_block->node.link, &work_list);
1306    while (!list_is_empty(&work_list)) {
1307       struct vtn_block *block =
1308          list_first_entry(&work_list, struct vtn_block, node.link);
1309       list_del(&block->node.link);
1310 
1311       vtn_assert(block->block);
1312 
1313       const uint32_t *block_start = block->label;
1314       const uint32_t *block_end = block->branch;
1315 
1316       b->nb.cursor = nir_after_block(block->block);
1317       block_start = vtn_foreach_instruction(b, block_start, block_end,
1318                                             vtn_handle_phis_first_pass);
1319       vtn_foreach_instruction(b, block_start, block_end, handler);
1320       block->end_nop = nir_nop(&b->nb);
1321 
1322       SpvOp op = *block_end & SpvOpCodeMask;
1323       switch (op) {
1324       case SpvOpBranch: {
1325          struct vtn_block *branch_block = vtn_block(b, block->branch[1]);
1326          vtn_add_unstructured_block(b, func, &work_list, branch_block);
1327          nir_goto(&b->nb, branch_block->block);
1328          break;
1329       }
1330 
1331       case SpvOpBranchConditional: {
1332          nir_ssa_def *cond = vtn_ssa_value(b, block->branch[1])->def;
1333          struct vtn_block *then_block = vtn_block(b, block->branch[2]);
1334          struct vtn_block *else_block = vtn_block(b, block->branch[3]);
1335 
1336          vtn_add_unstructured_block(b, func, &work_list, then_block);
1337          if (then_block == else_block) {
1338             nir_goto(&b->nb, then_block->block);
1339          } else {
1340             vtn_add_unstructured_block(b, func, &work_list, else_block);
1341             nir_goto_if(&b->nb, then_block->block, nir_src_for_ssa(cond),
1342                                 else_block->block);
1343          }
1344 
1345          break;
1346       }
1347 
1348       case SpvOpSwitch: {
1349          struct list_head cases;
1350          list_inithead(&cases);
1351          vtn_parse_switch(b, NULL, block->branch, &cases);
1352 
1353          nir_ssa_def *sel = vtn_get_nir_ssa(b, block->branch[1]);
1354 
1355          struct vtn_case *def = NULL;
1356          vtn_foreach_cf_node(case_node, &cases) {
1357             struct vtn_case *cse = vtn_cf_node_as_case(case_node);
1358             if (cse->is_default) {
1359                assert(def == NULL);
1360                def = cse;
1361                continue;
1362             }
1363 
1364             nir_ssa_def *cond = nir_imm_false(&b->nb);
1365             util_dynarray_foreach(&cse->values, uint64_t, val)
1366                cond = nir_ior(&b->nb, cond, nir_ieq_imm(&b->nb, sel, *val));
1367 
1368             /* block for the next check */
1369             nir_block *e = vtn_new_unstructured_block(b, func);
1370             vtn_add_unstructured_block(b, func, &work_list, cse->block);
1371 
1372             /* add branching */
1373             nir_goto_if(&b->nb, cse->block->block, nir_src_for_ssa(cond), e);
1374             b->nb.cursor = nir_after_block(e);
1375          }
1376 
1377          vtn_assert(def != NULL);
1378          vtn_add_unstructured_block(b, func, &work_list, def->block);
1379 
1380          /* now that all cases are handled, branch into the default block */
1381          nir_goto(&b->nb, def->block->block);
1382          break;
1383       }
1384 
1385       case SpvOpKill: {
1386          nir_discard(&b->nb);
1387          nir_goto(&b->nb, b->func->nir_func->impl->end_block);
1388          break;
1389       }
1390 
1391       case SpvOpUnreachable:
1392       case SpvOpReturn:
1393       case SpvOpReturnValue: {
1394          vtn_emit_ret_store(b, block);
1395          nir_goto(&b->nb, b->func->nir_func->impl->end_block);
1396          break;
1397       }
1398 
1399       default:
1400          vtn_fail("Unhandled opcode %s", spirv_op_to_string(op));
1401       }
1402    }
1403 }
1404 
1405 void
vtn_function_emit(struct vtn_builder * b,struct vtn_function * func,vtn_instruction_handler instruction_handler)1406 vtn_function_emit(struct vtn_builder *b, struct vtn_function *func,
1407                   vtn_instruction_handler instruction_handler)
1408 {
1409    static int force_unstructured = -1;
1410    if (force_unstructured < 0) {
1411       force_unstructured =
1412          env_var_as_boolean("MESA_SPIRV_FORCE_UNSTRUCTURED", false);
1413    }
1414 
1415    nir_function_impl *impl = func->nir_func->impl;
1416    nir_builder_init(&b->nb, impl);
1417    b->func = func;
1418    b->nb.cursor = nir_after_cf_list(&impl->body);
1419    b->nb.exact = b->exact;
1420    b->phi_table = _mesa_pointer_hash_table_create(b);
1421 
1422    if (b->shader->info.stage == MESA_SHADER_KERNEL || force_unstructured) {
1423       impl->structured = false;
1424       vtn_emit_cf_func_unstructured(b, func, instruction_handler);
1425    } else {
1426       vtn_emit_cf_list_structured(b, &func->body, NULL, NULL,
1427                                   instruction_handler);
1428    }
1429 
1430    vtn_foreach_instruction(b, func->start_block->label, func->end,
1431                            vtn_handle_phi_second_pass);
1432 
1433    if (func->nir_func->impl->structured)
1434       nir_copy_prop_impl(impl);
1435    nir_rematerialize_derefs_in_use_blocks_impl(impl);
1436 
1437    /*
1438     * There are some cases where we need to repair SSA to insert
1439     * the needed phi nodes:
1440     *
1441     * - Continue blocks for loops get inserted before the body of the loop
1442     *   but instructions in the continue may use SSA defs in the loop body.
1443     *
1444     * - Early termination instructions `OpKill` and `OpTerminateInvocation`,
1445     *   in NIR. They're represented by regular intrinsics with no control-flow
1446     *   semantics. This means that the SSA form from the SPIR-V may not
1447     *   100% match NIR.
1448     *
1449     * - Switches with only default case may also define SSA which may
1450     *   subsequently be used out of the switch.
1451     */
1452    if (func->nir_func->impl->structured)
1453       nir_repair_ssa_impl(impl);
1454 
1455    func->emitted = true;
1456 }
1457