Merge branch 'lp-offset-twoside'
[mesa.git] / src / glsl / lower_instructions.cpp
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 lower_instructions.cpp
26 *
27 * Many GPUs lack native instructions for certain expression operations, and
28 * must replace them with some other expression tree. This pass lowers some
29 * of the most common cases, allowing the lowering code to be implemented once
30 * rather than in each driver backend.
31 *
32 * Currently supported transformations:
33 * - SUB_TO_ADD_NEG
34 * - DIV_TO_MUL_RCP
35 * - EXP_TO_EXP2
36 * - LOG_TO_LOG2
37 * - MOD_TO_FRACT
38 *
39 * SUB_TO_ADD_NEG:
40 * ---------------
41 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
42 *
43 * This simplifies expression reassociation, and for many backends
44 * there is no subtract operation separate from adding the negation.
45 * For backends with native subtract operations, they will probably
46 * want to recognize add(op0, neg(op1)) or the other way around to
47 * produce a subtract anyway.
48 *
49 * DIV_TO_MUL_RCP:
50 * ---------------
51 * Breaks an ir_unop_div expression down to op0 * (rcp(op1)).
52 *
53 * Many GPUs don't have a divide instruction (945 and 965 included),
54 * but they do have an RCP instruction to compute an approximate
55 * reciprocal. By breaking the operation down, constant reciprocals
56 * can get constant folded.
57 *
58 * EXP_TO_EXP2 and LOG_TO_LOG2:
59 * ----------------------------
60 * Many GPUs don't have a base e log or exponent instruction, but they
61 * do have base 2 versions, so this pass converts exp and log to exp2
62 * and log2 operations.
63 *
64 * MOD_TO_FRACT:
65 * -------------
66 * Breaks an ir_unop_mod expression down to (op1 * fract(op0 / op1))
67 *
68 * Many GPUs don't have a MOD instruction (945 and 965 included), and
69 * if we have to break it down like this anyway, it gives an
70 * opportunity to do things like constant fold the (1.0 / op1) easily.
71 */
72
73 #include "main/core.h" /* for M_E */
74 #include "glsl_types.h"
75 #include "ir.h"
76 #include "ir_optimization.h"
77
78 class lower_instructions_visitor : public ir_hierarchical_visitor {
79 public:
80 lower_instructions_visitor(unsigned lower)
81 : progress(false), lower(lower) { }
82
83 ir_visitor_status visit_leave(ir_expression *);
84
85 bool progress;
86
87 private:
88 unsigned lower; /** Bitfield of which operations to lower */
89
90 void sub_to_add_neg(ir_expression *);
91 void div_to_mul_rcp(ir_expression *);
92 void mod_to_fract(ir_expression *);
93 void exp_to_exp2(ir_expression *);
94 void log_to_log2(ir_expression *);
95 };
96
97 /**
98 * Determine if a particular type of lowering should occur
99 */
100 #define lowering(x) (this->lower & x)
101
102 bool
103 lower_instructions(exec_list *instructions, unsigned what_to_lower)
104 {
105 lower_instructions_visitor v(what_to_lower);
106
107 visit_list_elements(&v, instructions);
108 return v.progress;
109 }
110
111 void
112 lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
113 {
114 ir->operation = ir_binop_add;
115 ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type,
116 ir->operands[1], NULL);
117 this->progress = true;
118 }
119
120 void
121 lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
122 {
123 if (!ir->operands[1]->type->is_integer()) {
124 /* New expression for the 1.0 / op1 */
125 ir_rvalue *expr;
126 expr = new(ir) ir_expression(ir_unop_rcp,
127 ir->operands[1]->type,
128 ir->operands[1],
129 NULL);
130
131 /* op0 / op1 -> op0 * (1.0 / op1) */
132 ir->operation = ir_binop_mul;
133 ir->operands[1] = expr;
134 } else {
135 /* Be careful with integer division -- we need to do it as a
136 * float and re-truncate, since rcp(n > 1) of an integer would
137 * just be 0.
138 */
139 ir_rvalue *op0, *op1;
140 const struct glsl_type *vec_type;
141
142 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
143 ir->operands[1]->type->vector_elements,
144 ir->operands[1]->type->matrix_columns);
145
146 if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
147 op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
148 else
149 op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
150
151 op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
152
153 vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
154 ir->operands[0]->type->vector_elements,
155 ir->operands[0]->type->matrix_columns);
156
157 if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
158 op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
159 else
160 op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
161
162 op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
163
164 ir->operation = ir_unop_f2i;
165 ir->operands[0] = op0;
166 ir->operands[1] = NULL;
167 }
168
169 this->progress = true;
170 }
171
172 void
173 lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
174 {
175 ir_constant *log2_e = new(ir) ir_constant(log2f(M_E));
176
177 ir->operation = ir_unop_exp2;
178 ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type,
179 ir->operands[0], log2_e);
180 this->progress = true;
181 }
182
183 void
184 lower_instructions_visitor::log_to_log2(ir_expression *ir)
185 {
186 ir->operation = ir_binop_mul;
187 ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
188 ir->operands[0], NULL);
189 ir->operands[1] = new(ir) ir_constant(1.0f / log2f(M_E));
190 this->progress = true;
191 }
192
193 void
194 lower_instructions_visitor::mod_to_fract(ir_expression *ir)
195 {
196 ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b",
197 ir_var_temporary);
198 this->base_ir->insert_before(temp);
199
200 ir_assignment *const assign =
201 new(ir) ir_assignment(new(ir) ir_dereference_variable(temp),
202 ir->operands[1], NULL);
203
204 this->base_ir->insert_before(assign);
205
206 ir_expression *const div_expr =
207 new(ir) ir_expression(ir_binop_div, ir->operands[0]->type,
208 ir->operands[0],
209 new(ir) ir_dereference_variable(temp));
210
211 /* Don't generate new IR that would need to be lowered in an additional
212 * pass.
213 */
214 if (lowering(DIV_TO_MUL_RCP))
215 div_to_mul_rcp(div_expr);
216
217 ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract,
218 ir->operands[0]->type,
219 div_expr,
220 NULL);
221
222 ir->operation = ir_binop_mul;
223 ir->operands[0] = new(ir) ir_dereference_variable(temp);
224 ir->operands[1] = expr;
225 this->progress = true;
226 }
227
228 ir_visitor_status
229 lower_instructions_visitor::visit_leave(ir_expression *ir)
230 {
231 switch (ir->operation) {
232 case ir_binop_sub:
233 if (lowering(SUB_TO_ADD_NEG))
234 sub_to_add_neg(ir);
235 break;
236
237 case ir_binop_div:
238 if (lowering(DIV_TO_MUL_RCP))
239 div_to_mul_rcp(ir);
240 break;
241
242 case ir_unop_exp:
243 if (lowering(EXP_TO_EXP2))
244 exp_to_exp2(ir);
245 break;
246
247 case ir_unop_log:
248 if (lowering(LOG_TO_LOG2))
249 log_to_log2(ir);
250 break;
251
252 case ir_binop_mod:
253 if (lowering(MOD_TO_FRACT))
254 mod_to_fract(ir);
255 break;
256
257 default:
258 return visit_continue;
259 }
260
261 return visit_continue;
262 }