1/*
2 * Copyright © 2010 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
21 * DEALINGS IN THE SOFTWARE.
22 */
23
24/**
25 * \file opt_dead_code_local.cpp
26 *
27 * Eliminates local dead assignments from the code.
28 *
29 * This operates on basic blocks, tracking assignments and finding if
30 * they're used before the variable is completely reassigned.
31 *
32 * Compare this to ir_dead_code.cpp, which operates globally looking
33 * for assignments to variables that are never read.
34 */
35
36#include "ir.h"
37#include "ir_basic_block.h"
38#include "ir_optimization.h"
39#include "compiler/glsl_types.h"
40
41static bool debug = false;
42
43namespace {
44
45class assignment_entry : public exec_node
46{
47public:
48   /* override operator new from exec_node */
49   DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(assignment_entry)
50
51   assignment_entry(ir_variable *lhs, ir_assignment *ir)
52   {
53      assert(lhs);
54      assert(ir);
55      this->lhs = lhs;
56      this->ir = ir;
57      this->unused = ir->write_mask;
58   }
59
60   ir_variable *lhs;
61   ir_assignment *ir;
62
63   /* bitmask of xyzw channels written that haven't been used so far. */
64   int unused;
65};
66
67class kill_for_derefs_visitor : public ir_hierarchical_visitor {
68public:
69   using ir_hierarchical_visitor::visit;
70
71   kill_for_derefs_visitor(exec_list *assignments)
72   {
73      this->assignments = assignments;
74   }
75
76   void use_channels(ir_variable *const var, int used)
77   {
78      foreach_in_list_safe(assignment_entry, entry, this->assignments) {
79	 if (entry->lhs == var) {
80	    if (var->type->is_scalar() || var->type->is_vector()) {
81	       if (debug)
82		  printf("used %s (0x%01x - 0x%01x)\n", entry->lhs->name,
83			 entry->unused, used & 0xf);
84	       entry->unused &= ~used;
85	       if (!entry->unused)
86		  entry->remove();
87	    } else {
88	       if (debug)
89		  printf("used %s\n", entry->lhs->name);
90	       entry->remove();
91	    }
92	 }
93      }
94   }
95
96   virtual ir_visitor_status visit(ir_dereference_variable *ir)
97   {
98      use_channels(ir->var, ~0);
99
100      return visit_continue;
101   }
102
103   virtual ir_visitor_status visit(ir_swizzle *ir)
104   {
105      ir_dereference_variable *deref = ir->val->as_dereference_variable();
106      if (!deref)
107	 return visit_continue;
108
109      int used = 0;
110      used |= 1 << ir->mask.x;
111      if (ir->mask.num_components > 1)
112         used |= 1 << ir->mask.y;
113      if (ir->mask.num_components > 2)
114         used |= 1 << ir->mask.z;
115      if (ir->mask.num_components > 3)
116         used |= 1 << ir->mask.w;
117
118      use_channels(deref->var, used);
119
120      return visit_continue_with_parent;
121   }
122
123   virtual ir_visitor_status visit_leave(ir_emit_vertex *)
124   {
125      /* For the purpose of dead code elimination, emitting a vertex counts as
126       * "reading" all of the currently assigned output variables.
127       */
128      foreach_in_list_safe(assignment_entry, entry, this->assignments) {
129         if (entry->lhs->data.mode == ir_var_shader_out) {
130            if (debug)
131               printf("kill %s\n", entry->lhs->name);
132            entry->remove();
133         }
134      }
135
136      return visit_continue;
137   }
138
139private:
140   exec_list *assignments;
141};
142
143class array_index_visit : public ir_hierarchical_visitor {
144public:
145   array_index_visit(ir_hierarchical_visitor *v)
146   {
147      this->visitor = v;
148   }
149
150   virtual ir_visitor_status visit_enter(class ir_dereference_array *ir)
151   {
152      ir->array_index->accept(visitor);
153      return visit_continue;
154   }
155
156   static void run(ir_instruction *ir, ir_hierarchical_visitor *v)
157   {
158      array_index_visit top_visit(v);
159      ir->accept(& top_visit);
160   }
161
162   ir_hierarchical_visitor *visitor;
163};
164
165} /* unnamed namespace */
166
167/**
168 * Adds an entry to the available copy list if it's a plain assignment
169 * of a variable to a variable.
170 */
171static bool
172process_assignment(void *lin_ctx, ir_assignment *ir, exec_list *assignments)
173{
174   ir_variable *var = NULL;
175   bool progress = false;
176   kill_for_derefs_visitor v(assignments);
177
178   /* If this is an assignment of the form "foo = foo;", remove the whole
179    * instruction and be done with it.
180    */
181   const ir_variable *const lhs_var = ir->whole_variable_written();
182   if (lhs_var != NULL && lhs_var == ir->rhs->whole_variable_referenced()) {
183      ir->remove();
184      return true;
185   }
186
187   /* Kill assignment entries for things used to produce this assignment. */
188   ir->rhs->accept(&v);
189
190   /* Kill assignment enties used as array indices.
191    */
192   array_index_visit::run(ir->lhs, &v);
193   var = ir->lhs->variable_referenced();
194   assert(var);
195
196   /* Now, check if we did a whole-variable assignment. */
197   ir_dereference_variable *deref_var = ir->lhs->as_dereference_variable();
198
199   /* If it's a vector type, we can do per-channel elimination of
200    * use of the RHS.
201    */
202   if (deref_var && (deref_var->var->type->is_scalar() ||
203                     deref_var->var->type->is_vector())) {
204
205      if (debug)
206         printf("looking for %s.0x%01x to remove\n", var->name,
207                ir->write_mask);
208
209      foreach_in_list_safe(assignment_entry, entry, assignments) {
210         if (entry->lhs != var)
211            continue;
212
213         /* Skip if the assignment we're trying to eliminate isn't a plain
214          * variable deref. */
215         if (entry->ir->lhs->ir_type != ir_type_dereference_variable)
216            continue;
217
218         int remove = entry->unused & ir->write_mask;
219         if (debug) {
220            printf("%s 0x%01x - 0x%01x = 0x%01x\n",
221                   var->name,
222                   entry->ir->write_mask,
223                   remove, entry->ir->write_mask & ~remove);
224         }
225         if (remove) {
226            progress = true;
227
228            if (debug) {
229               printf("rewriting:\n  ");
230               entry->ir->print();
231               printf("\n");
232            }
233
234            entry->ir->write_mask &= ~remove;
235            entry->unused &= ~remove;
236            if (entry->ir->write_mask == 0) {
237               /* Delete the dead assignment. */
238               entry->ir->remove();
239               entry->remove();
240            } else {
241               void *mem_ctx = ralloc_parent(entry->ir);
242               /* Reswizzle the RHS arguments according to the new
243                * write_mask.
244                */
245               unsigned components[4];
246               unsigned channels = 0;
247               unsigned next = 0;
248
249               for (int i = 0; i < 4; i++) {
250                  if ((entry->ir->write_mask | remove) & (1 << i)) {
251                     if (!(remove & (1 << i)))
252                        components[channels++] = next;
253                     next++;
254                  }
255               }
256
257               entry->ir->rhs = new(mem_ctx) ir_swizzle(entry->ir->rhs,
258                                                        components,
259                                                        channels);
260               if (debug) {
261                  printf("to:\n  ");
262                  entry->ir->print();
263                  printf("\n");
264               }
265            }
266         }
267      }
268   } else if (ir->whole_variable_written() != NULL) {
269      /* We did a whole-variable assignment.  So, any instruction in
270       * the assignment list with the same LHS is dead.
271       */
272      if (debug)
273         printf("looking for %s to remove\n", var->name);
274      foreach_in_list_safe(assignment_entry, entry, assignments) {
275         if (entry->lhs == var) {
276            if (debug)
277               printf("removing %s\n", var->name);
278            entry->ir->remove();
279            entry->remove();
280            progress = true;
281         }
282      }
283   }
284
285   /* Add this instruction to the assignment list available to be removed. */
286   assignment_entry *entry = new(lin_ctx) assignment_entry(var, ir);
287   assignments->push_tail(entry);
288
289   if (debug) {
290      printf("add %s\n", var->name);
291
292      printf("current entries\n");
293      foreach_in_list(assignment_entry, entry, assignments) {
294	 printf("    %s (0x%01x)\n", entry->lhs->name, entry->unused);
295      }
296   }
297
298   return progress;
299}
300
301static void
302dead_code_local_basic_block(ir_instruction *first,
303			     ir_instruction *last,
304			     void *data)
305{
306   ir_instruction *ir, *ir_next;
307   /* List of avaialble_copy */
308   exec_list assignments;
309   bool *out_progress = (bool *)data;
310   bool progress = false;
311
312   void *ctx = ralloc_context(NULL);
313   void *lin_ctx = linear_alloc_parent(ctx, 0);
314
315   /* Safe looping, since process_assignment */
316   for (ir = first, ir_next = (ir_instruction *)first->next;;
317	ir = ir_next, ir_next = (ir_instruction *)ir->next) {
318      ir_assignment *ir_assign = ir->as_assignment();
319
320      if (debug) {
321	 ir->print();
322	 printf("\n");
323      }
324
325      if (ir_assign) {
326	 progress = process_assignment(lin_ctx, ir_assign, &assignments) ||
327                    progress;
328      } else {
329	 kill_for_derefs_visitor kill(&assignments);
330	 ir->accept(&kill);
331      }
332
333      if (ir == last)
334	 break;
335   }
336   *out_progress = progress;
337   ralloc_free(ctx);
338}
339
340/**
341 * Does a copy propagation pass on the code present in the instruction stream.
342 */
343bool
344do_dead_code_local(exec_list *instructions)
345{
346   bool progress = false;
347
348   call_for_basic_blocks(instructions, dead_code_local_basic_block, &progress);
349
350   return progress;
351}
352