1/*
2 * Copyright © 2013 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#include "ir.h"
24#include "ir_builder.h"
25#include "ir_rvalue_visitor.h"
26#include "ir_optimization.h"
27#include "main/shader_types.h"
28
29using namespace ir_builder;
30
31namespace {
32
33class vector_deref_visitor : public ir_rvalue_enter_visitor {
34public:
35   vector_deref_visitor(void *mem_ctx, gl_shader_stage shader_stage)
36      : progress(false), shader_stage(shader_stage),
37        factory(&factory_instructions, mem_ctx)
38   {
39   }
40
41   virtual ~vector_deref_visitor()
42   {
43   }
44
45   virtual void handle_rvalue(ir_rvalue **rv);
46   virtual ir_visitor_status visit_enter(ir_assignment *ir);
47
48   bool progress;
49   gl_shader_stage shader_stage;
50   exec_list factory_instructions;
51   ir_factory factory;
52};
53
54} /* anonymous namespace */
55
56ir_visitor_status
57vector_deref_visitor::visit_enter(ir_assignment *ir)
58{
59   if (!ir->lhs || ir->lhs->ir_type != ir_type_dereference_array)
60      return ir_rvalue_enter_visitor::visit_enter(ir);
61
62   ir_dereference_array *const deref = (ir_dereference_array *) ir->lhs;
63   if (!deref->array->type->is_vector())
64      return ir_rvalue_enter_visitor::visit_enter(ir);
65
66   /* SSBOs and shared variables are backed by memory and may be accessed by
67    * multiple threads simultaneously.  It's not safe to lower a single
68    * component store to a load-vec-store because it may race with writes to
69    * other components.
70    */
71   ir_variable *var = deref->variable_referenced();
72   if (var->data.mode == ir_var_shader_storage ||
73       var->data.mode == ir_var_shader_shared)
74      return ir_rvalue_enter_visitor::visit_enter(ir);
75
76   ir_rvalue *const new_lhs = deref->array;
77
78   void *mem_ctx = ralloc_parent(ir);
79   ir_constant *old_index_constant =
80      deref->array_index->constant_expression_value(mem_ctx);
81   if (!old_index_constant) {
82      if (shader_stage == MESA_SHADER_TESS_CTRL &&
83          deref->variable_referenced()->data.mode == ir_var_shader_out) {
84         /* Tessellation control shader outputs act as if they have memory
85          * backing them and if we have writes from multiple threads
86          * targeting the same vec4 (this can happen for patch outputs), the
87          * load-vec-store pattern of ir_triop_vector_insert doesn't work.
88          * Instead, we have to lower to a series of conditional write-masked
89          * assignments.
90          */
91         ir_variable *const src_temp =
92            factory.make_temp(ir->rhs->type, "scalar_tmp");
93
94         /* The newly created variable declaration goes before the assignment
95          * because we're going to set it as the new LHS.
96          */
97         ir->insert_before(factory.instructions);
98         ir->set_lhs(new(mem_ctx) ir_dereference_variable(src_temp));
99
100         ir_variable *const arr_index =
101            factory.make_temp(deref->array_index->type, "index_tmp");
102         factory.emit(assign(arr_index, deref->array_index));
103
104         for (unsigned i = 0; i < new_lhs->type->vector_elements; i++) {
105            ir_constant *const cmp_index =
106               ir_constant::zero(factory.mem_ctx, deref->array_index->type);
107            cmp_index->value.u[0] = i;
108
109            ir_rvalue *const lhs_clone = new_lhs->clone(factory.mem_ctx, NULL);
110            ir_dereference_variable *const src_temp_deref =
111               new(mem_ctx) ir_dereference_variable(src_temp);
112
113            if (new_lhs->ir_type != ir_type_swizzle) {
114               assert(lhs_clone->as_dereference());
115
116               factory.emit(if_tree(equal(arr_index, cmp_index),
117                                    assign(lhs_clone->as_dereference(),
118                                           src_temp_deref,
119                                           WRITEMASK_X << i)));
120            } else {
121               ir_assignment *cond_assign =
122                  new(mem_ctx) ir_assignment(swizzle(lhs_clone, i, 1),
123                                             src_temp_deref);
124
125               factory.emit(if_tree(equal(arr_index, cmp_index), cond_assign));
126            }
127         }
128         ir->insert_after(factory.instructions);
129      } else {
130         ir->rhs = new(mem_ctx) ir_expression(ir_triop_vector_insert,
131                                              new_lhs->type,
132                                              new_lhs->clone(mem_ctx, NULL),
133                                              ir->rhs,
134                                              deref->array_index);
135         ir->write_mask = (1 << new_lhs->type->vector_elements) - 1;
136         ir->set_lhs(new_lhs);
137      }
138   } else {
139      unsigned index = old_index_constant->get_uint_component(0);
140
141      if (index >= new_lhs->type->vector_elements) {
142         /* Section 5.11 (Out-of-Bounds Accesses) of the GLSL 4.60 spec says:
143          *
144          *  In the subsections described above for array, vector, matrix and
145          *  structure accesses, any out-of-bounds access produced undefined
146          *  behavior.... Out-of-bounds writes may be discarded or overwrite
147          *  other variables of the active program.
148          */
149         ir->remove();
150         return visit_continue;
151      }
152
153      if (new_lhs->ir_type != ir_type_swizzle) {
154         ir->set_lhs(new_lhs);
155         ir->write_mask = 1 << index;
156      } else {
157         /* If the "new" LHS is a swizzle, use the set_lhs helper to instead
158          * swizzle the RHS.
159          */
160         unsigned component[1] = { index };
161         ir->set_lhs(new(mem_ctx) ir_swizzle(new_lhs, component, 1));
162      }
163   }
164
165   return ir_rvalue_enter_visitor::visit_enter(ir);
166}
167
168void
169vector_deref_visitor::handle_rvalue(ir_rvalue **rv)
170{
171   if (*rv == NULL || (*rv)->ir_type != ir_type_dereference_array)
172      return;
173
174   ir_dereference_array *const deref = (ir_dereference_array *) *rv;
175   if (!deref->array->type->is_vector())
176      return;
177
178   /* Back-ends need to be able to handle derefs on vectors for SSBOs, UBOs,
179    * and shared variables.  They have to handle it for writes anyway so we
180    * may as well require it for reads.
181    */
182   ir_variable *var = deref->variable_referenced();
183   if (var && (var->data.mode == ir_var_shader_storage ||
184               var->data.mode == ir_var_shader_shared ||
185               (var->data.mode == ir_var_uniform &&
186                var->get_interface_type())))
187      return;
188
189   void *mem_ctx = ralloc_parent(deref);
190   *rv = new(mem_ctx) ir_expression(ir_binop_vector_extract,
191                                    deref->array,
192                                    deref->array_index);
193}
194
195bool
196lower_vector_derefs(gl_linked_shader *shader)
197{
198   vector_deref_visitor v(shader->ir, shader->Stage);
199
200   visit_list_elements(&v, shader->ir);
201
202   return v.progress;
203}
204