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 * - POW_TO_EXP2
37 * - LOG_TO_LOG2
38 * - MOD_TO_FRACT
39 *
40 * SUB_TO_ADD_NEG:
41 * ---------------
42 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
43 *
44 * This simplifies expression reassociation, and for many backends
45 * there is no subtract operation separate from adding the negation.
46 * For backends with native subtract operations, they will probably
47 * want to recognize add(op0, neg(op1)) or the other way around to
48 * produce a subtract anyway.
49 *
50 * DIV_TO_MUL_RCP:
51 * ---------------
52 * Breaks an ir_unop_div expression down to op0 * (rcp(op1)).
53 *
54 * Many GPUs don't have a divide instruction (945 and 965 included),
55 * but they do have an RCP instruction to compute an approximate
56 * reciprocal.  By breaking the operation down, constant reciprocals
57 * can get constant folded.
58 *
59 * EXP_TO_EXP2 and LOG_TO_LOG2:
60 * ----------------------------
61 * Many GPUs don't have a base e log or exponent instruction, but they
62 * do have base 2 versions, so this pass converts exp and log to exp2
63 * and log2 operations.
64 *
65 * POW_TO_EXP2:
66 * -----------
67 * Many older GPUs don't have an x**y instruction.  For these GPUs, convert
68 * x**y to 2**(y * log2(x)).
69 *
70 * MOD_TO_FRACT:
71 * -------------
72 * Breaks an ir_unop_mod expression down to (op1 * fract(op0 / op1))
73 *
74 * Many GPUs don't have a MOD instruction (945 and 965 included), and
75 * if we have to break it down like this anyway, it gives an
76 * opportunity to do things like constant fold the (1.0 / op1) easily.
77 */
78
79#include "main/core.h" /* for M_LOG2E */
80#include "glsl_types.h"
81#include "ir.h"
82#include "ir_optimization.h"
83
84class lower_instructions_visitor : public ir_hierarchical_visitor {
85public:
86   lower_instructions_visitor(unsigned lower)
87      : progress(false), lower(lower) { }
88
89   ir_visitor_status visit_leave(ir_expression *);
90
91   bool progress;
92
93private:
94   unsigned lower; /** Bitfield of which operations to lower */
95
96   void sub_to_add_neg(ir_expression *);
97   void div_to_mul_rcp(ir_expression *);
98   void mod_to_fract(ir_expression *);
99   void exp_to_exp2(ir_expression *);
100   void pow_to_exp2(ir_expression *);
101   void log_to_log2(ir_expression *);
102};
103
104/**
105 * Determine if a particular type of lowering should occur
106 */
107#define lowering(x) (this->lower & x)
108
109bool
110lower_instructions(exec_list *instructions, unsigned what_to_lower)
111{
112   lower_instructions_visitor v(what_to_lower);
113
114   visit_list_elements(&v, instructions);
115   return v.progress;
116}
117
118void
119lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
120{
121   ir->operation = ir_binop_add;
122   ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type,
123					   ir->operands[1], NULL);
124   this->progress = true;
125}
126
127void
128lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
129{
130   if (!ir->operands[1]->type->is_integer()) {
131      /* New expression for the 1.0 / op1 */
132      ir_rvalue *expr;
133      expr = new(ir) ir_expression(ir_unop_rcp,
134				   ir->operands[1]->type,
135				   ir->operands[1],
136				   NULL);
137
138      /* op0 / op1 -> op0 * (1.0 / op1) */
139      ir->operation = ir_binop_mul;
140      ir->operands[1] = expr;
141   } else {
142      /* Be careful with integer division -- we need to do it as a
143       * float and re-truncate, since rcp(n > 1) of an integer would
144       * just be 0.
145       */
146      ir_rvalue *op0, *op1;
147      const struct glsl_type *vec_type;
148
149      vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
150					 ir->operands[1]->type->vector_elements,
151					 ir->operands[1]->type->matrix_columns);
152
153      if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
154	 op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
155      else
156	 op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
157
158      op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
159
160      vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
161					 ir->operands[0]->type->vector_elements,
162					 ir->operands[0]->type->matrix_columns);
163
164      if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
165	 op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
166      else
167	 op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
168
169      op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
170
171      ir->operation = ir_unop_f2i;
172      ir->operands[0] = op0;
173      ir->operands[1] = NULL;
174   }
175
176   this->progress = true;
177}
178
179void
180lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
181{
182   ir_constant *log2_e = new(ir) ir_constant(float(M_LOG2E));
183
184   ir->operation = ir_unop_exp2;
185   ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type,
186					   ir->operands[0], log2_e);
187   this->progress = true;
188}
189
190void
191lower_instructions_visitor::pow_to_exp2(ir_expression *ir)
192{
193   ir_expression *const log2_x =
194      new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
195			    ir->operands[0]);
196
197   ir->operation = ir_unop_exp2;
198   ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type,
199					   ir->operands[1], log2_x);
200   ir->operands[1] = NULL;
201   this->progress = true;
202}
203
204void
205lower_instructions_visitor::log_to_log2(ir_expression *ir)
206{
207   ir->operation = ir_binop_mul;
208   ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
209					   ir->operands[0], NULL);
210   ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E));
211   this->progress = true;
212}
213
214void
215lower_instructions_visitor::mod_to_fract(ir_expression *ir)
216{
217   ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b",
218					   ir_var_temporary);
219   this->base_ir->insert_before(temp);
220
221   ir_assignment *const assign =
222      new(ir) ir_assignment(new(ir) ir_dereference_variable(temp),
223			    ir->operands[1], NULL);
224
225   this->base_ir->insert_before(assign);
226
227   ir_expression *const div_expr =
228      new(ir) ir_expression(ir_binop_div, ir->operands[0]->type,
229			    ir->operands[0],
230			    new(ir) ir_dereference_variable(temp));
231
232   /* Don't generate new IR that would need to be lowered in an additional
233    * pass.
234    */
235   if (lowering(DIV_TO_MUL_RCP))
236      div_to_mul_rcp(div_expr);
237
238   ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract,
239					   ir->operands[0]->type,
240					   div_expr,
241					   NULL);
242
243   ir->operation = ir_binop_mul;
244   ir->operands[0] = new(ir) ir_dereference_variable(temp);
245   ir->operands[1] = expr;
246   this->progress = true;
247}
248
249ir_visitor_status
250lower_instructions_visitor::visit_leave(ir_expression *ir)
251{
252   switch (ir->operation) {
253   case ir_binop_sub:
254      if (lowering(SUB_TO_ADD_NEG))
255	 sub_to_add_neg(ir);
256      break;
257
258   case ir_binop_div:
259      if (lowering(DIV_TO_MUL_RCP))
260	 div_to_mul_rcp(ir);
261      break;
262
263   case ir_unop_exp:
264      if (lowering(EXP_TO_EXP2))
265	 exp_to_exp2(ir);
266      break;
267
268   case ir_unop_log:
269      if (lowering(LOG_TO_LOG2))
270	 log_to_log2(ir);
271      break;
272
273   case ir_binop_mod:
274      if (lowering(MOD_TO_FRACT))
275	 mod_to_fract(ir);
276      break;
277
278   case ir_binop_pow:
279      if (lowering(POW_TO_EXP2))
280	 pow_to_exp2(ir);
281      break;
282
283   default:
284      return visit_continue;
285   }
286
287   return visit_continue;
288}
289