lower_instructions.cpp revision 7ec681f3
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 * - INT_DIV_TO_MUL_RCP
36 * - EXP_TO_EXP2
37 * - POW_TO_EXP2
38 * - LOG_TO_LOG2
39 * - MOD_TO_FLOOR
40 * - LDEXP_TO_ARITH
41 * - DFREXP_TO_ARITH
42 * - CARRY_TO_ARITH
43 * - BORROW_TO_ARITH
44 * - SAT_TO_CLAMP
45 * - DOPS_TO_DFRAC
46 *
47 * SUB_TO_ADD_NEG:
48 * ---------------
49 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
50 *
51 * This simplifies expression reassociation, and for many backends
52 * there is no subtract operation separate from adding the negation.
53 * For backends with native subtract operations, they will probably
54 * want to recognize add(op0, neg(op1)) or the other way around to
55 * produce a subtract anyway.
56 *
57 * FDIV_TO_MUL_RCP, DDIV_TO_MUL_RCP, and INT_DIV_TO_MUL_RCP:
58 * ---------------------------------------------------------
59 * Breaks an ir_binop_div expression down to op0 * (rcp(op1)).
60 *
61 * Many GPUs don't have a divide instruction (945 and 965 included),
62 * but they do have an RCP instruction to compute an approximate
63 * reciprocal.  By breaking the operation down, constant reciprocals
64 * can get constant folded.
65 *
66 * FDIV_TO_MUL_RCP lowers single-precision and half-precision
67 * floating point division;
68 * DDIV_TO_MUL_RCP only lowers double-precision floating point division.
69 * DIV_TO_MUL_RCP is a convenience macro that sets both flags.
70 * INT_DIV_TO_MUL_RCP handles the integer case, converting to and from floating
71 * point so that RCP is possible.
72 *
73 * EXP_TO_EXP2 and LOG_TO_LOG2:
74 * ----------------------------
75 * Many GPUs don't have a base e log or exponent instruction, but they
76 * do have base 2 versions, so this pass converts exp and log to exp2
77 * and log2 operations.
78 *
79 * POW_TO_EXP2:
80 * -----------
81 * Many older GPUs don't have an x**y instruction.  For these GPUs, convert
82 * x**y to 2**(y * log2(x)).
83 *
84 * MOD_TO_FLOOR:
85 * -------------
86 * Breaks an ir_binop_mod expression down to (op0 - op1 * floor(op0 / op1))
87 *
88 * Many GPUs don't have a MOD instruction (945 and 965 included), and
89 * if we have to break it down like this anyway, it gives an
90 * opportunity to do things like constant fold the (1.0 / op1) easily.
91 *
92 * Note: before we used to implement this as op1 * fract(op / op1) but this
93 * implementation had significant precision errors.
94 *
95 * LDEXP_TO_ARITH:
96 * -------------
97 * Converts ir_binop_ldexp to arithmetic and bit operations for float sources.
98 *
99 * DFREXP_DLDEXP_TO_ARITH:
100 * ---------------
101 * Converts ir_binop_ldexp, ir_unop_frexp_sig, and ir_unop_frexp_exp to
102 * arithmetic and bit ops for double arguments.
103 *
104 * CARRY_TO_ARITH:
105 * ---------------
106 * Converts ir_carry into (x + y) < x.
107 *
108 * BORROW_TO_ARITH:
109 * ----------------
110 * Converts ir_borrow into (x < y).
111 *
112 * SAT_TO_CLAMP:
113 * -------------
114 * Converts ir_unop_saturate into min(max(x, 0.0), 1.0)
115 *
116 * DOPS_TO_DFRAC:
117 * --------------
118 * Converts double trunc, ceil, floor, round to fract
119 */
120
121#include "c99_math.h"
122#include "program/prog_instruction.h" /* for swizzle */
123#include "compiler/glsl_types.h"
124#include "ir.h"
125#include "ir_builder.h"
126#include "ir_optimization.h"
127#include "util/half_float.h"
128
129using namespace ir_builder;
130
131namespace {
132
133class lower_instructions_visitor : public ir_hierarchical_visitor {
134public:
135   lower_instructions_visitor(unsigned lower)
136      : progress(false), lower(lower) { }
137
138   ir_visitor_status visit_leave(ir_expression *);
139
140   bool progress;
141
142private:
143   unsigned lower; /** Bitfield of which operations to lower */
144
145   void sub_to_add_neg(ir_expression *);
146   void div_to_mul_rcp(ir_expression *);
147   void int_div_to_mul_rcp(ir_expression *);
148   void mod_to_floor(ir_expression *);
149   void exp_to_exp2(ir_expression *);
150   void pow_to_exp2(ir_expression *);
151   void log_to_log2(ir_expression *);
152   void ldexp_to_arith(ir_expression *);
153   void dldexp_to_arith(ir_expression *);
154   void dfrexp_sig_to_arith(ir_expression *);
155   void dfrexp_exp_to_arith(ir_expression *);
156   void carry_to_arith(ir_expression *);
157   void borrow_to_arith(ir_expression *);
158   void sat_to_clamp(ir_expression *);
159   void double_dot_to_fma(ir_expression *);
160   void double_lrp(ir_expression *);
161   void dceil_to_dfrac(ir_expression *);
162   void dfloor_to_dfrac(ir_expression *);
163   void dround_even_to_dfrac(ir_expression *);
164   void dtrunc_to_dfrac(ir_expression *);
165   void dsign_to_csel(ir_expression *);
166   void bit_count_to_math(ir_expression *);
167   void extract_to_shifts(ir_expression *);
168   void insert_to_shifts(ir_expression *);
169   void reverse_to_shifts(ir_expression *ir);
170   void find_lsb_to_float_cast(ir_expression *ir);
171   void find_msb_to_float_cast(ir_expression *ir);
172   void imul_high_to_mul(ir_expression *ir);
173   void sqrt_to_abs_sqrt(ir_expression *ir);
174   void mul64_to_mul_and_mul_high(ir_expression *ir);
175
176   ir_expression *_carry(operand a, operand b);
177
178   static ir_constant *_imm_fp(void *mem_ctx,
179                               const glsl_type *type,
180                               double f,
181                               unsigned vector_elements=1);
182};
183
184} /* anonymous namespace */
185
186/**
187 * Determine if a particular type of lowering should occur
188 */
189#define lowering(x) (this->lower & x)
190
191bool
192lower_instructions(exec_list *instructions, unsigned what_to_lower)
193{
194   lower_instructions_visitor v(what_to_lower);
195
196   visit_list_elements(&v, instructions);
197   return v.progress;
198}
199
200void
201lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
202{
203   ir->operation = ir_binop_add;
204   ir->init_num_operands();
205   ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type,
206					   ir->operands[1], NULL);
207   this->progress = true;
208}
209
210void
211lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
212{
213   assert(ir->operands[1]->type->is_float_16_32_64());
214
215   /* New expression for the 1.0 / op1 */
216   ir_rvalue *expr;
217   expr = new(ir) ir_expression(ir_unop_rcp,
218				ir->operands[1]->type,
219				ir->operands[1]);
220
221   /* op0 / op1 -> op0 * (1.0 / op1) */
222   ir->operation = ir_binop_mul;
223   ir->init_num_operands();
224   ir->operands[1] = expr;
225
226   this->progress = true;
227}
228
229void
230lower_instructions_visitor::int_div_to_mul_rcp(ir_expression *ir)
231{
232   assert(ir->operands[1]->type->is_integer_32());
233
234   /* Be careful with integer division -- we need to do it as a
235    * float and re-truncate, since rcp(n > 1) of an integer would
236    * just be 0.
237    */
238   ir_rvalue *op0, *op1;
239   const struct glsl_type *vec_type;
240
241   vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
242				      ir->operands[1]->type->vector_elements,
243				      ir->operands[1]->type->matrix_columns);
244
245   if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
246      op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
247   else
248      op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
249
250   op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
251
252   vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
253				      ir->operands[0]->type->vector_elements,
254				      ir->operands[0]->type->matrix_columns);
255
256   if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
257      op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
258   else
259      op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
260
261   vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
262				      ir->type->vector_elements,
263				      ir->type->matrix_columns);
264
265   op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
266
267   if (ir->operands[1]->type->base_type == GLSL_TYPE_INT) {
268      ir->operation = ir_unop_f2i;
269      ir->operands[0] = op0;
270   } else {
271      ir->operation = ir_unop_i2u;
272      ir->operands[0] = new(ir) ir_expression(ir_unop_f2i, op0);
273   }
274   ir->init_num_operands();
275   ir->operands[1] = NULL;
276
277   this->progress = true;
278}
279
280void
281lower_instructions_visitor::exp_to_exp2(ir_expression *ir)
282{
283   ir_constant *log2_e = _imm_fp(ir, ir->type, M_LOG2E);
284
285   ir->operation = ir_unop_exp2;
286   ir->init_num_operands();
287   ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type,
288					   ir->operands[0], log2_e);
289   this->progress = true;
290}
291
292void
293lower_instructions_visitor::pow_to_exp2(ir_expression *ir)
294{
295   ir_expression *const log2_x =
296      new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
297			    ir->operands[0]);
298
299   ir->operation = ir_unop_exp2;
300   ir->init_num_operands();
301   ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type,
302					   ir->operands[1], log2_x);
303   ir->operands[1] = NULL;
304   this->progress = true;
305}
306
307void
308lower_instructions_visitor::log_to_log2(ir_expression *ir)
309{
310   ir->operation = ir_binop_mul;
311   ir->init_num_operands();
312   ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type,
313					   ir->operands[0], NULL);
314   ir->operands[1] = _imm_fp(ir, ir->operands[0]->type, 1.0 / M_LOG2E);
315   this->progress = true;
316}
317
318void
319lower_instructions_visitor::mod_to_floor(ir_expression *ir)
320{
321   ir_variable *x = new(ir) ir_variable(ir->operands[0]->type, "mod_x",
322                                         ir_var_temporary);
323   ir_variable *y = new(ir) ir_variable(ir->operands[1]->type, "mod_y",
324                                         ir_var_temporary);
325   this->base_ir->insert_before(x);
326   this->base_ir->insert_before(y);
327
328   ir_assignment *const assign_x =
329      new(ir) ir_assignment(new(ir) ir_dereference_variable(x),
330                            ir->operands[0]);
331   ir_assignment *const assign_y =
332      new(ir) ir_assignment(new(ir) ir_dereference_variable(y),
333                            ir->operands[1]);
334
335   this->base_ir->insert_before(assign_x);
336   this->base_ir->insert_before(assign_y);
337
338   ir_expression *const div_expr =
339      new(ir) ir_expression(ir_binop_div, x->type,
340                            new(ir) ir_dereference_variable(x),
341                            new(ir) ir_dereference_variable(y));
342
343   /* Don't generate new IR that would need to be lowered in an additional
344    * pass.
345    */
346   if ((lowering(FDIV_TO_MUL_RCP) && ir->type->is_float_16_32()) ||
347       (lowering(DDIV_TO_MUL_RCP) && ir->type->is_double()))
348      div_to_mul_rcp(div_expr);
349
350   ir_expression *const floor_expr =
351      new(ir) ir_expression(ir_unop_floor, x->type, div_expr);
352
353   if (lowering(DOPS_TO_DFRAC) && ir->type->is_double())
354      dfloor_to_dfrac(floor_expr);
355
356   ir_expression *const mul_expr =
357      new(ir) ir_expression(ir_binop_mul,
358                            new(ir) ir_dereference_variable(y),
359                            floor_expr);
360
361   ir->operation = ir_binop_sub;
362   ir->init_num_operands();
363   ir->operands[0] = new(ir) ir_dereference_variable(x);
364   ir->operands[1] = mul_expr;
365   this->progress = true;
366}
367
368void
369lower_instructions_visitor::ldexp_to_arith(ir_expression *ir)
370{
371   /* Translates
372    *    ir_binop_ldexp x exp
373    * into
374    *
375    *    extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
376    *    resulting_biased_exp = min(extracted_biased_exp + exp, 255);
377    *
378    *    if (extracted_biased_exp >= 255)
379    *       return x; // +/-inf, NaN
380    *
381    *    sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
382    *
383    *    if (min(resulting_biased_exp, extracted_biased_exp) < 1)
384    *       resulting_biased_exp = 0;
385    *    if (resulting_biased_exp >= 255 ||
386    *        min(resulting_biased_exp, extracted_biased_exp) < 1) {
387    *       sign_mantissa &= sign_mask;
388    *    }
389    *
390    *    return bitcast_u2f(sign_mantissa |
391    *                       lshift(i2u(resulting_biased_exp), exp_shift));
392    *
393    * which we can't actually implement as such, since the GLSL IR doesn't
394    * have vectorized if-statements. We actually implement it without branches
395    * using conditional-select:
396    *
397    *    extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
398    *    resulting_biased_exp = min(extracted_biased_exp + exp, 255);
399    *
400    *    sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
401    *
402    *    flush_to_zero = lequal(min(resulting_biased_exp, extracted_biased_exp), 0);
403    *    resulting_biased_exp = csel(flush_to_zero, 0, resulting_biased_exp)
404    *    zero_mantissa = logic_or(flush_to_zero,
405    *                             gequal(resulting_biased_exp, 255));
406    *    sign_mantissa = csel(zero_mantissa, sign_mantissa & sign_mask, sign_mantissa);
407    *
408    *    result = sign_mantissa |
409    *             lshift(i2u(resulting_biased_exp), exp_shift));
410    *
411    *    return csel(extracted_biased_exp >= 255, x, bitcast_u2f(result));
412    *
413    * The definition of ldexp in the GLSL spec says:
414    *
415    *    "If this product is too large to be represented in the
416    *     floating-point type, the result is undefined."
417    *
418    * However, the definition of ldexp in the GLSL ES spec does not contain
419    * this sentence, so we do need to handle overflow correctly.
420    *
421    * There is additional language limiting the defined range of exp, but this
422    * is merely to allow implementations that store 2^exp in a temporary
423    * variable.
424    */
425
426   const unsigned vec_elem = ir->type->vector_elements;
427
428   /* Types */
429   const glsl_type *ivec = glsl_type::get_instance(GLSL_TYPE_INT, vec_elem, 1);
430   const glsl_type *uvec = glsl_type::get_instance(GLSL_TYPE_UINT, vec_elem, 1);
431   const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1);
432
433   /* Temporary variables */
434   ir_variable *x = new(ir) ir_variable(ir->type, "x", ir_var_temporary);
435   ir_variable *exp = new(ir) ir_variable(ivec, "exp", ir_var_temporary);
436   ir_variable *result = new(ir) ir_variable(uvec, "result", ir_var_temporary);
437
438   ir_variable *extracted_biased_exp =
439      new(ir) ir_variable(ivec, "extracted_biased_exp", ir_var_temporary);
440   ir_variable *resulting_biased_exp =
441      new(ir) ir_variable(ivec, "resulting_biased_exp", ir_var_temporary);
442
443   ir_variable *sign_mantissa =
444      new(ir) ir_variable(uvec, "sign_mantissa", ir_var_temporary);
445
446   ir_variable *flush_to_zero =
447      new(ir) ir_variable(bvec, "flush_to_zero", ir_var_temporary);
448   ir_variable *zero_mantissa =
449      new(ir) ir_variable(bvec, "zero_mantissa", ir_var_temporary);
450
451   ir_instruction &i = *base_ir;
452
453   /* Copy <x> and <exp> arguments. */
454   i.insert_before(x);
455   i.insert_before(assign(x, ir->operands[0]));
456   i.insert_before(exp);
457   i.insert_before(assign(exp, ir->operands[1]));
458
459   /* Extract the biased exponent from <x>. */
460   i.insert_before(extracted_biased_exp);
461   i.insert_before(assign(extracted_biased_exp,
462                          rshift(bitcast_f2i(abs(x)),
463                                 new(ir) ir_constant(23, vec_elem))));
464
465   /* The definition of ldexp in the GLSL 4.60 spec says:
466    *
467    *    "If exp is greater than +128 (single-precision) or +1024
468    *     (double-precision), the value returned is undefined. If exp is less
469    *     than -126 (single-precision) or -1022 (double-precision), the value
470    *     returned may be flushed to zero."
471    *
472    * So we do not have to guard against the possibility of addition overflow,
473    * which could happen when exp is close to INT_MAX. Addition underflow
474    * cannot happen (the worst case is 0 + (-INT_MAX)).
475    */
476   i.insert_before(resulting_biased_exp);
477   i.insert_before(assign(resulting_biased_exp,
478                          min2(add(extracted_biased_exp, exp),
479                               new(ir) ir_constant(255, vec_elem))));
480
481   i.insert_before(sign_mantissa);
482   i.insert_before(assign(sign_mantissa,
483                          bit_and(bitcast_f2u(x),
484                                  new(ir) ir_constant(0x807fffffu, vec_elem))));
485
486   /* We flush to zero if the original or resulting biased exponent is 0,
487    * indicating a +/-0.0 or subnormal input or output.
488    *
489    * The mantissa is set to 0 if the resulting biased exponent is 255, since
490    * an overflow should produce a +/-inf result.
491    *
492    * Note that NaN inputs are handled separately.
493    */
494   i.insert_before(flush_to_zero);
495   i.insert_before(assign(flush_to_zero,
496                          lequal(min2(resulting_biased_exp,
497                                      extracted_biased_exp),
498                                 ir_constant::zero(ir, ivec))));
499   i.insert_before(assign(resulting_biased_exp,
500                          csel(flush_to_zero,
501                               ir_constant::zero(ir, ivec),
502                               resulting_biased_exp)));
503
504   i.insert_before(zero_mantissa);
505   i.insert_before(assign(zero_mantissa,
506                          logic_or(flush_to_zero,
507                                   equal(resulting_biased_exp,
508                                         new(ir) ir_constant(255, vec_elem)))));
509   i.insert_before(assign(sign_mantissa,
510                          csel(zero_mantissa,
511                               bit_and(sign_mantissa,
512                                       new(ir) ir_constant(0x80000000u, vec_elem)),
513                               sign_mantissa)));
514
515   /* Don't generate new IR that would need to be lowered in an additional
516    * pass.
517    */
518   i.insert_before(result);
519   if (!lowering(INSERT_TO_SHIFTS)) {
520      i.insert_before(assign(result,
521                             bitfield_insert(sign_mantissa,
522                                             i2u(resulting_biased_exp),
523                                             new(ir) ir_constant(23u, vec_elem),
524                                             new(ir) ir_constant(8u, vec_elem))));
525   } else {
526      i.insert_before(assign(result,
527                             bit_or(sign_mantissa,
528                                    lshift(i2u(resulting_biased_exp),
529                                           new(ir) ir_constant(23, vec_elem)))));
530   }
531
532   ir->operation = ir_triop_csel;
533   ir->init_num_operands();
534   ir->operands[0] = gequal(extracted_biased_exp,
535                            new(ir) ir_constant(255, vec_elem));
536   ir->operands[1] = new(ir) ir_dereference_variable(x);
537   ir->operands[2] = bitcast_u2f(result);
538
539   this->progress = true;
540}
541
542void
543lower_instructions_visitor::dldexp_to_arith(ir_expression *ir)
544{
545   /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
546    * from the significand.
547    */
548
549   const unsigned vec_elem = ir->type->vector_elements;
550
551   /* Types */
552   const glsl_type *ivec = glsl_type::get_instance(GLSL_TYPE_INT, vec_elem, 1);
553   const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1);
554
555   /* Constants */
556   ir_constant *zeroi = ir_constant::zero(ir, ivec);
557
558   ir_constant *sign_mask = new(ir) ir_constant(0x80000000u);
559
560   ir_constant *exp_shift = new(ir) ir_constant(20u);
561   ir_constant *exp_width = new(ir) ir_constant(11u);
562   ir_constant *exp_bias = new(ir) ir_constant(1022, vec_elem);
563
564   /* Temporary variables */
565   ir_variable *x = new(ir) ir_variable(ir->type, "x", ir_var_temporary);
566   ir_variable *exp = new(ir) ir_variable(ivec, "exp", ir_var_temporary);
567
568   ir_variable *zero_sign_x = new(ir) ir_variable(ir->type, "zero_sign_x",
569                                                  ir_var_temporary);
570
571   ir_variable *extracted_biased_exp =
572      new(ir) ir_variable(ivec, "extracted_biased_exp", ir_var_temporary);
573   ir_variable *resulting_biased_exp =
574      new(ir) ir_variable(ivec, "resulting_biased_exp", ir_var_temporary);
575
576   ir_variable *is_not_zero_or_underflow =
577      new(ir) ir_variable(bvec, "is_not_zero_or_underflow", ir_var_temporary);
578
579   ir_instruction &i = *base_ir;
580
581   /* Copy <x> and <exp> arguments. */
582   i.insert_before(x);
583   i.insert_before(assign(x, ir->operands[0]));
584   i.insert_before(exp);
585   i.insert_before(assign(exp, ir->operands[1]));
586
587   ir_expression *frexp_exp = expr(ir_unop_frexp_exp, x);
588   if (lowering(DFREXP_DLDEXP_TO_ARITH))
589      dfrexp_exp_to_arith(frexp_exp);
590
591   /* Extract the biased exponent from <x>. */
592   i.insert_before(extracted_biased_exp);
593   i.insert_before(assign(extracted_biased_exp, add(frexp_exp, exp_bias)));
594
595   i.insert_before(resulting_biased_exp);
596   i.insert_before(assign(resulting_biased_exp,
597                          add(extracted_biased_exp, exp)));
598
599   /* Test if result is ±0.0, subnormal, or underflow by checking if the
600    * resulting biased exponent would be less than 0x1. If so, the result is
601    * 0.0 with the sign of x. (Actually, invert the conditions so that
602    * immediate values are the second arguments, which is better for i965)
603    * TODO: Implement in a vector fashion.
604    */
605   i.insert_before(zero_sign_x);
606   for (unsigned elem = 0; elem < vec_elem; elem++) {
607      ir_variable *unpacked =
608         new(ir) ir_variable(glsl_type::uvec2_type, "unpacked", ir_var_temporary);
609      i.insert_before(unpacked);
610      i.insert_before(
611            assign(unpacked,
612                   expr(ir_unop_unpack_double_2x32, swizzle(x, elem, 1))));
613      i.insert_before(assign(unpacked, bit_and(swizzle_y(unpacked), sign_mask->clone(ir, NULL)),
614                             WRITEMASK_Y));
615      i.insert_before(assign(unpacked, ir_constant::zero(ir, glsl_type::uint_type), WRITEMASK_X));
616      i.insert_before(assign(zero_sign_x,
617                             expr(ir_unop_pack_double_2x32, unpacked),
618                             1 << elem));
619   }
620   i.insert_before(is_not_zero_or_underflow);
621   i.insert_before(assign(is_not_zero_or_underflow,
622                          gequal(resulting_biased_exp,
623                                  new(ir) ir_constant(0x1, vec_elem))));
624   i.insert_before(assign(x, csel(is_not_zero_or_underflow,
625                                  x, zero_sign_x)));
626   i.insert_before(assign(resulting_biased_exp,
627                          csel(is_not_zero_or_underflow,
628                               resulting_biased_exp, zeroi)));
629
630   /* We could test for overflows by checking if the resulting biased exponent
631    * would be greater than 0xFE. Turns out we don't need to because the GLSL
632    * spec says:
633    *
634    *    "If this product is too large to be represented in the
635    *     floating-point type, the result is undefined."
636    */
637
638   ir_rvalue *results[4] = {NULL};
639   for (unsigned elem = 0; elem < vec_elem; elem++) {
640      ir_variable *unpacked =
641         new(ir) ir_variable(glsl_type::uvec2_type, "unpacked", ir_var_temporary);
642      i.insert_before(unpacked);
643      i.insert_before(
644            assign(unpacked,
645                   expr(ir_unop_unpack_double_2x32, swizzle(x, elem, 1))));
646
647      ir_expression *bfi = bitfield_insert(
648            swizzle_y(unpacked),
649            i2u(swizzle(resulting_biased_exp, elem, 1)),
650            exp_shift->clone(ir, NULL),
651            exp_width->clone(ir, NULL));
652
653      i.insert_before(assign(unpacked, bfi, WRITEMASK_Y));
654
655      results[elem] = expr(ir_unop_pack_double_2x32, unpacked);
656   }
657
658   ir->operation = ir_quadop_vector;
659   ir->init_num_operands();
660   ir->operands[0] = results[0];
661   ir->operands[1] = results[1];
662   ir->operands[2] = results[2];
663   ir->operands[3] = results[3];
664
665   /* Don't generate new IR that would need to be lowered in an additional
666    * pass.
667    */
668
669   this->progress = true;
670}
671
672void
673lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression *ir)
674{
675   const unsigned vec_elem = ir->type->vector_elements;
676   const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1);
677
678   /* Double-precision floating-point values are stored as
679    *   1 sign bit;
680    *   11 exponent bits;
681    *   52 mantissa bits.
682    *
683    * We're just extracting the significand here, so we only need to modify
684    * the upper 32-bit uint. Unfortunately we must extract each double
685    * independently as there is no vector version of unpackDouble.
686    */
687
688   ir_instruction &i = *base_ir;
689
690   ir_variable *is_not_zero =
691      new(ir) ir_variable(bvec, "is_not_zero", ir_var_temporary);
692   ir_rvalue *results[4] = {NULL};
693
694   ir_constant *dzero = new(ir) ir_constant(0.0, vec_elem);
695   i.insert_before(is_not_zero);
696   i.insert_before(
697         assign(is_not_zero,
698                nequal(abs(ir->operands[0]->clone(ir, NULL)), dzero)));
699
700   /* TODO: Remake this as more vector-friendly when int64 support is
701    * available.
702    */
703   for (unsigned elem = 0; elem < vec_elem; elem++) {
704      ir_constant *zero = new(ir) ir_constant(0u, 1);
705      ir_constant *sign_mantissa_mask = new(ir) ir_constant(0x800fffffu, 1);
706
707      /* Exponent of double floating-point values in the range [0.5, 1.0). */
708      ir_constant *exponent_value = new(ir) ir_constant(0x3fe00000u, 1);
709
710      ir_variable *bits =
711         new(ir) ir_variable(glsl_type::uint_type, "bits", ir_var_temporary);
712      ir_variable *unpacked =
713         new(ir) ir_variable(glsl_type::uvec2_type, "unpacked", ir_var_temporary);
714
715      ir_rvalue *x = swizzle(ir->operands[0]->clone(ir, NULL), elem, 1);
716
717      i.insert_before(bits);
718      i.insert_before(unpacked);
719      i.insert_before(assign(unpacked, expr(ir_unop_unpack_double_2x32, x)));
720
721      /* Manipulate the high uint to remove the exponent and replace it with
722       * either the default exponent or zero.
723       */
724      i.insert_before(assign(bits, swizzle_y(unpacked)));
725      i.insert_before(assign(bits, bit_and(bits, sign_mantissa_mask)));
726      i.insert_before(assign(bits, bit_or(bits,
727                                          csel(swizzle(is_not_zero, elem, 1),
728                                               exponent_value,
729                                               zero))));
730      i.insert_before(assign(unpacked, bits, WRITEMASK_Y));
731      results[elem] = expr(ir_unop_pack_double_2x32, unpacked);
732   }
733
734   /* Put the dvec back together */
735   ir->operation = ir_quadop_vector;
736   ir->init_num_operands();
737   ir->operands[0] = results[0];
738   ir->operands[1] = results[1];
739   ir->operands[2] = results[2];
740   ir->operands[3] = results[3];
741
742   this->progress = true;
743}
744
745void
746lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression *ir)
747{
748   const unsigned vec_elem = ir->type->vector_elements;
749   const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1);
750   const glsl_type *uvec = glsl_type::get_instance(GLSL_TYPE_UINT, vec_elem, 1);
751
752   /* Double-precision floating-point values are stored as
753    *   1 sign bit;
754    *   11 exponent bits;
755    *   52 mantissa bits.
756    *
757    * We're just extracting the exponent here, so we only care about the upper
758    * 32-bit uint.
759    */
760
761   ir_instruction &i = *base_ir;
762
763   ir_variable *is_not_zero =
764      new(ir) ir_variable(bvec, "is_not_zero", ir_var_temporary);
765   ir_variable *high_words =
766      new(ir) ir_variable(uvec, "high_words", ir_var_temporary);
767   ir_constant *dzero = new(ir) ir_constant(0.0, vec_elem);
768   ir_constant *izero = new(ir) ir_constant(0, vec_elem);
769
770   ir_rvalue *absval = abs(ir->operands[0]);
771
772   i.insert_before(is_not_zero);
773   i.insert_before(high_words);
774   i.insert_before(assign(is_not_zero, nequal(absval->clone(ir, NULL), dzero)));
775
776   /* Extract all of the upper uints. */
777   for (unsigned elem = 0; elem < vec_elem; elem++) {
778      ir_rvalue *x = swizzle(absval->clone(ir, NULL), elem, 1);
779
780      i.insert_before(assign(high_words,
781                             swizzle_y(expr(ir_unop_unpack_double_2x32, x)),
782                             1 << elem));
783
784   }
785   ir_constant *exponent_shift = new(ir) ir_constant(20, vec_elem);
786   ir_constant *exponent_bias = new(ir) ir_constant(-1022, vec_elem);
787
788   /* For non-zero inputs, shift the exponent down and apply bias. */
789   ir->operation = ir_triop_csel;
790   ir->init_num_operands();
791   ir->operands[0] = new(ir) ir_dereference_variable(is_not_zero);
792   ir->operands[1] = add(exponent_bias, u2i(rshift(high_words, exponent_shift)));
793   ir->operands[2] = izero;
794
795   this->progress = true;
796}
797
798void
799lower_instructions_visitor::carry_to_arith(ir_expression *ir)
800{
801   /* Translates
802    *   ir_binop_carry x y
803    * into
804    *   sum = ir_binop_add x y
805    *   bcarry = ir_binop_less sum x
806    *   carry = ir_unop_b2i bcarry
807    */
808
809   ir_rvalue *x_clone = ir->operands[0]->clone(ir, NULL);
810   ir->operation = ir_unop_i2u;
811   ir->init_num_operands();
812   ir->operands[0] = b2i(less(add(ir->operands[0], ir->operands[1]), x_clone));
813   ir->operands[1] = NULL;
814
815   this->progress = true;
816}
817
818void
819lower_instructions_visitor::borrow_to_arith(ir_expression *ir)
820{
821   /* Translates
822    *   ir_binop_borrow x y
823    * into
824    *   bcarry = ir_binop_less x y
825    *   carry = ir_unop_b2i bcarry
826    */
827
828   ir->operation = ir_unop_i2u;
829   ir->init_num_operands();
830   ir->operands[0] = b2i(less(ir->operands[0], ir->operands[1]));
831   ir->operands[1] = NULL;
832
833   this->progress = true;
834}
835
836void
837lower_instructions_visitor::sat_to_clamp(ir_expression *ir)
838{
839   /* Translates
840    *   ir_unop_saturate x
841    * into
842    *   ir_binop_min (ir_binop_max(x, 0.0), 1.0)
843    */
844
845   ir->operation = ir_binop_min;
846   ir->init_num_operands();
847
848   ir_constant *zero = _imm_fp(ir, ir->operands[0]->type, 0.0);
849   ir->operands[0] = new(ir) ir_expression(ir_binop_max, ir->operands[0]->type,
850                                           ir->operands[0], zero);
851   ir->operands[1] = _imm_fp(ir, ir->operands[0]->type, 1.0);
852
853   this->progress = true;
854}
855
856void
857lower_instructions_visitor::double_dot_to_fma(ir_expression *ir)
858{
859   ir_variable *temp = new(ir) ir_variable(ir->operands[0]->type->get_base_type(), "dot_res",
860					   ir_var_temporary);
861   this->base_ir->insert_before(temp);
862
863   int nc = ir->operands[0]->type->components();
864   for (int i = nc - 1; i >= 1; i--) {
865      ir_assignment *assig;
866      if (i == (nc - 1)) {
867         assig = assign(temp, mul(swizzle(ir->operands[0]->clone(ir, NULL), i, 1),
868                                  swizzle(ir->operands[1]->clone(ir, NULL), i, 1)));
869      } else {
870         assig = assign(temp, fma(swizzle(ir->operands[0]->clone(ir, NULL), i, 1),
871                                  swizzle(ir->operands[1]->clone(ir, NULL), i, 1),
872                                  temp));
873      }
874      this->base_ir->insert_before(assig);
875   }
876
877   ir->operation = ir_triop_fma;
878   ir->init_num_operands();
879   ir->operands[0] = swizzle(ir->operands[0], 0, 1);
880   ir->operands[1] = swizzle(ir->operands[1], 0, 1);
881   ir->operands[2] = new(ir) ir_dereference_variable(temp);
882
883   this->progress = true;
884
885}
886
887void
888lower_instructions_visitor::double_lrp(ir_expression *ir)
889{
890   int swizval;
891   ir_rvalue *op0 = ir->operands[0], *op2 = ir->operands[2];
892   ir_constant *one = new(ir) ir_constant(1.0, op2->type->vector_elements);
893
894   switch (op2->type->vector_elements) {
895   case 1:
896      swizval = SWIZZLE_XXXX;
897      break;
898   default:
899      assert(op0->type->vector_elements == op2->type->vector_elements);
900      swizval = SWIZZLE_XYZW;
901      break;
902   }
903
904   ir->operation = ir_triop_fma;
905   ir->init_num_operands();
906   ir->operands[0] = swizzle(op2, swizval, op0->type->vector_elements);
907   ir->operands[2] = mul(sub(one, op2->clone(ir, NULL)), op0);
908
909   this->progress = true;
910}
911
912void
913lower_instructions_visitor::dceil_to_dfrac(ir_expression *ir)
914{
915   /*
916    * frtemp = frac(x);
917    * temp = sub(x, frtemp);
918    * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
919    */
920   ir_instruction &i = *base_ir;
921   ir_constant *zero = new(ir) ir_constant(0.0, ir->operands[0]->type->vector_elements);
922   ir_constant *one = new(ir) ir_constant(1.0, ir->operands[0]->type->vector_elements);
923   ir_variable *frtemp = new(ir) ir_variable(ir->operands[0]->type, "frtemp",
924                                             ir_var_temporary);
925
926   i.insert_before(frtemp);
927   i.insert_before(assign(frtemp, fract(ir->operands[0])));
928
929   ir->operation = ir_binop_add;
930   ir->init_num_operands();
931   ir->operands[0] = sub(ir->operands[0]->clone(ir, NULL), frtemp);
932   ir->operands[1] = csel(nequal(frtemp, zero), one, zero->clone(ir, NULL));
933
934   this->progress = true;
935}
936
937void
938lower_instructions_visitor::dfloor_to_dfrac(ir_expression *ir)
939{
940   /*
941    * frtemp = frac(x);
942    * result = sub(x, frtemp);
943    */
944   ir->operation = ir_binop_sub;
945   ir->init_num_operands();
946   ir->operands[1] = fract(ir->operands[0]->clone(ir, NULL));
947
948   this->progress = true;
949}
950void
951lower_instructions_visitor::dround_even_to_dfrac(ir_expression *ir)
952{
953   /*
954    * insane but works
955    * temp = x + 0.5;
956    * frtemp = frac(temp);
957    * t2 = sub(temp, frtemp);
958    * if (frac(x) == 0.5)
959    *     result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
960    *  else
961    *     result = t2;
962
963    */
964   ir_instruction &i = *base_ir;
965   ir_variable *frtemp = new(ir) ir_variable(ir->operands[0]->type, "frtemp",
966                                             ir_var_temporary);
967   ir_variable *temp = new(ir) ir_variable(ir->operands[0]->type, "temp",
968                                           ir_var_temporary);
969   ir_variable *t2 = new(ir) ir_variable(ir->operands[0]->type, "t2",
970                                           ir_var_temporary);
971   ir_constant *p5 = new(ir) ir_constant(0.5, ir->operands[0]->type->vector_elements);
972   ir_constant *one = new(ir) ir_constant(1.0, ir->operands[0]->type->vector_elements);
973   ir_constant *zero = new(ir) ir_constant(0.0, ir->operands[0]->type->vector_elements);
974
975   i.insert_before(temp);
976   i.insert_before(assign(temp, add(ir->operands[0], p5)));
977
978   i.insert_before(frtemp);
979   i.insert_before(assign(frtemp, fract(temp)));
980
981   i.insert_before(t2);
982   i.insert_before(assign(t2, sub(temp, frtemp)));
983
984   ir->operation = ir_triop_csel;
985   ir->init_num_operands();
986   ir->operands[0] = equal(fract(ir->operands[0]->clone(ir, NULL)),
987                           p5->clone(ir, NULL));
988   ir->operands[1] = csel(equal(fract(mul(t2, p5->clone(ir, NULL))),
989                                zero),
990                          t2,
991                          sub(t2, one));
992   ir->operands[2] = new(ir) ir_dereference_variable(t2);
993
994   this->progress = true;
995}
996
997void
998lower_instructions_visitor::dtrunc_to_dfrac(ir_expression *ir)
999{
1000   /*
1001    * frtemp = frac(x);
1002    * temp = sub(x, frtemp);
1003    * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
1004    */
1005   ir_rvalue *arg = ir->operands[0];
1006   ir_instruction &i = *base_ir;
1007
1008   ir_constant *zero = new(ir) ir_constant(0.0, arg->type->vector_elements);
1009   ir_constant *one = new(ir) ir_constant(1.0, arg->type->vector_elements);
1010   ir_variable *frtemp = new(ir) ir_variable(arg->type, "frtemp",
1011                                             ir_var_temporary);
1012   ir_variable *temp = new(ir) ir_variable(ir->operands[0]->type, "temp",
1013                                           ir_var_temporary);
1014
1015   i.insert_before(frtemp);
1016   i.insert_before(assign(frtemp, fract(arg)));
1017   i.insert_before(temp);
1018   i.insert_before(assign(temp, sub(arg->clone(ir, NULL), frtemp)));
1019
1020   ir->operation = ir_triop_csel;
1021   ir->init_num_operands();
1022   ir->operands[0] = gequal(arg->clone(ir, NULL), zero);
1023   ir->operands[1] = new (ir) ir_dereference_variable(temp);
1024   ir->operands[2] = add(temp,
1025                         csel(equal(frtemp, zero->clone(ir, NULL)),
1026                              zero->clone(ir, NULL),
1027                              one));
1028
1029   this->progress = true;
1030}
1031
1032void
1033lower_instructions_visitor::dsign_to_csel(ir_expression *ir)
1034{
1035   /*
1036    * temp = x > 0.0 ? 1.0 : 0.0;
1037    * result = x < 0.0 ? -1.0 : temp;
1038    */
1039   ir_rvalue *arg = ir->operands[0];
1040   ir_constant *zero = new(ir) ir_constant(0.0, arg->type->vector_elements);
1041   ir_constant *one = new(ir) ir_constant(1.0, arg->type->vector_elements);
1042   ir_constant *neg_one = new(ir) ir_constant(-1.0, arg->type->vector_elements);
1043
1044   ir->operation = ir_triop_csel;
1045   ir->init_num_operands();
1046   ir->operands[0] = less(arg->clone(ir, NULL),
1047                          zero->clone(ir, NULL));
1048   ir->operands[1] = neg_one;
1049   ir->operands[2] = csel(greater(arg, zero),
1050                          one,
1051                          zero->clone(ir, NULL));
1052
1053   this->progress = true;
1054}
1055
1056void
1057lower_instructions_visitor::bit_count_to_math(ir_expression *ir)
1058{
1059   /* For more details, see:
1060    *
1061    * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
1062    */
1063   const unsigned elements = ir->operands[0]->type->vector_elements;
1064   ir_variable *temp = new(ir) ir_variable(glsl_type::uvec(elements), "temp",
1065                                           ir_var_temporary);
1066   ir_constant *c55555555 = new(ir) ir_constant(0x55555555u);
1067   ir_constant *c33333333 = new(ir) ir_constant(0x33333333u);
1068   ir_constant *c0F0F0F0F = new(ir) ir_constant(0x0F0F0F0Fu);
1069   ir_constant *c01010101 = new(ir) ir_constant(0x01010101u);
1070   ir_constant *c1 = new(ir) ir_constant(1u);
1071   ir_constant *c2 = new(ir) ir_constant(2u);
1072   ir_constant *c4 = new(ir) ir_constant(4u);
1073   ir_constant *c24 = new(ir) ir_constant(24u);
1074
1075   base_ir->insert_before(temp);
1076
1077   if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) {
1078      base_ir->insert_before(assign(temp, ir->operands[0]));
1079   } else {
1080      assert(ir->operands[0]->type->base_type == GLSL_TYPE_INT);
1081      base_ir->insert_before(assign(temp, i2u(ir->operands[0])));
1082   }
1083
1084   /* temp = temp - ((temp >> 1) & 0x55555555u); */
1085   base_ir->insert_before(assign(temp, sub(temp, bit_and(rshift(temp, c1),
1086                                                         c55555555))));
1087
1088   /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
1089   base_ir->insert_before(assign(temp, add(bit_and(temp, c33333333),
1090                                           bit_and(rshift(temp, c2),
1091                                                   c33333333->clone(ir, NULL)))));
1092
1093   /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1094   ir->operation = ir_unop_u2i;
1095   ir->init_num_operands();
1096   ir->operands[0] = rshift(mul(bit_and(add(temp, rshift(temp, c4)), c0F0F0F0F),
1097                                c01010101),
1098                            c24);
1099
1100   this->progress = true;
1101}
1102
1103void
1104lower_instructions_visitor::extract_to_shifts(ir_expression *ir)
1105{
1106   ir_variable *bits =
1107      new(ir) ir_variable(ir->operands[0]->type, "bits", ir_var_temporary);
1108
1109   base_ir->insert_before(bits);
1110   base_ir->insert_before(assign(bits, ir->operands[2]));
1111
1112   if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) {
1113      ir_constant *c1 =
1114         new(ir) ir_constant(1u, ir->operands[0]->type->vector_elements);
1115      ir_constant *c32 =
1116         new(ir) ir_constant(32u, ir->operands[0]->type->vector_elements);
1117      ir_constant *cFFFFFFFF =
1118         new(ir) ir_constant(0xFFFFFFFFu, ir->operands[0]->type->vector_elements);
1119
1120      /* At least some hardware treats (x << y) as (x << (y%32)).  This means
1121       * we'd get a mask of 0 when bits is 32.  Special case it.
1122       *
1123       * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1124       */
1125      ir_expression *mask = csel(equal(bits, c32),
1126                                 cFFFFFFFF,
1127                                 sub(lshift(c1, bits), c1->clone(ir, NULL)));
1128
1129      /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1130       *
1131       *    If bits is zero, the result will be zero.
1132       *
1133       * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1134       * select as in the signed integer case.
1135       *
1136       * (value >> offset) & mask;
1137       */
1138      ir->operation = ir_binop_bit_and;
1139      ir->init_num_operands();
1140      ir->operands[0] = rshift(ir->operands[0], ir->operands[1]);
1141      ir->operands[1] = mask;
1142      ir->operands[2] = NULL;
1143   } else {
1144      ir_constant *c0 =
1145         new(ir) ir_constant(int(0), ir->operands[0]->type->vector_elements);
1146      ir_constant *c32 =
1147         new(ir) ir_constant(int(32), ir->operands[0]->type->vector_elements);
1148      ir_variable *temp =
1149         new(ir) ir_variable(ir->operands[0]->type, "temp", ir_var_temporary);
1150
1151      /* temp = 32 - bits; */
1152      base_ir->insert_before(temp);
1153      base_ir->insert_before(assign(temp, sub(c32, bits)));
1154
1155      /* expr = value << (temp - offset)) >> temp; */
1156      ir_expression *expr =
1157         rshift(lshift(ir->operands[0], sub(temp, ir->operands[1])), temp);
1158
1159      /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1160       *
1161       *    If bits is zero, the result will be zero.
1162       *
1163       * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1164       * (32-0)) doesn't "erase" all of the data as we would like, so finish
1165       * up with:
1166       *
1167       * (bits == 0) ? 0 : e;
1168       */
1169      ir->operation = ir_triop_csel;
1170      ir->init_num_operands();
1171      ir->operands[0] = equal(c0, bits);
1172      ir->operands[1] = c0->clone(ir, NULL);
1173      ir->operands[2] = expr;
1174   }
1175
1176   this->progress = true;
1177}
1178
1179void
1180lower_instructions_visitor::insert_to_shifts(ir_expression *ir)
1181{
1182   ir_constant *c1;
1183   ir_constant *c32;
1184   ir_constant *cFFFFFFFF;
1185   ir_variable *offset =
1186      new(ir) ir_variable(ir->operands[0]->type, "offset", ir_var_temporary);
1187   ir_variable *bits =
1188      new(ir) ir_variable(ir->operands[0]->type, "bits", ir_var_temporary);
1189   ir_variable *mask =
1190      new(ir) ir_variable(ir->operands[0]->type, "mask", ir_var_temporary);
1191
1192   if (ir->operands[0]->type->base_type == GLSL_TYPE_INT) {
1193      c1 = new(ir) ir_constant(int(1), ir->operands[0]->type->vector_elements);
1194      c32 = new(ir) ir_constant(int(32), ir->operands[0]->type->vector_elements);
1195      cFFFFFFFF = new(ir) ir_constant(int(0xFFFFFFFF), ir->operands[0]->type->vector_elements);
1196   } else {
1197      assert(ir->operands[0]->type->base_type == GLSL_TYPE_UINT);
1198
1199      c1 = new(ir) ir_constant(1u, ir->operands[0]->type->vector_elements);
1200      c32 = new(ir) ir_constant(32u, ir->operands[0]->type->vector_elements);
1201      cFFFFFFFF = new(ir) ir_constant(0xFFFFFFFFu, ir->operands[0]->type->vector_elements);
1202   }
1203
1204   base_ir->insert_before(offset);
1205   base_ir->insert_before(assign(offset, ir->operands[2]));
1206
1207   base_ir->insert_before(bits);
1208   base_ir->insert_before(assign(bits, ir->operands[3]));
1209
1210   /* At least some hardware treats (x << y) as (x << (y%32)).  This means
1211    * we'd get a mask of 0 when bits is 32.  Special case it.
1212    *
1213    * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1214    *
1215    * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1216    *
1217    *    The result will be undefined if offset or bits is negative, or if the
1218    *    sum of offset and bits is greater than the number of bits used to
1219    *    store the operand.
1220    *
1221    * Since it's undefined, there are a couple other ways this could be
1222    * implemented.  The other way that was considered was to put the csel
1223    * around the whole thing:
1224    *
1225    *    final_result = bits == 32 ? insert : ... ;
1226    */
1227   base_ir->insert_before(mask);
1228
1229   base_ir->insert_before(assign(mask, csel(equal(bits, c32),
1230                                            cFFFFFFFF,
1231                                            lshift(sub(lshift(c1, bits),
1232                                                       c1->clone(ir, NULL)),
1233                                                   offset))));
1234
1235   /* (base & ~mask) | ((insert << offset) & mask) */
1236   ir->operation = ir_binop_bit_or;
1237   ir->init_num_operands();
1238   ir->operands[0] = bit_and(ir->operands[0], bit_not(mask));
1239   ir->operands[1] = bit_and(lshift(ir->operands[1], offset), mask);
1240   ir->operands[2] = NULL;
1241   ir->operands[3] = NULL;
1242
1243   this->progress = true;
1244}
1245
1246void
1247lower_instructions_visitor::reverse_to_shifts(ir_expression *ir)
1248{
1249   /* For more details, see:
1250    *
1251    * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1252    */
1253   ir_constant *c1 =
1254      new(ir) ir_constant(1u, ir->operands[0]->type->vector_elements);
1255   ir_constant *c2 =
1256      new(ir) ir_constant(2u, ir->operands[0]->type->vector_elements);
1257   ir_constant *c4 =
1258      new(ir) ir_constant(4u, ir->operands[0]->type->vector_elements);
1259   ir_constant *c8 =
1260      new(ir) ir_constant(8u, ir->operands[0]->type->vector_elements);
1261   ir_constant *c16 =
1262      new(ir) ir_constant(16u, ir->operands[0]->type->vector_elements);
1263   ir_constant *c33333333 =
1264      new(ir) ir_constant(0x33333333u, ir->operands[0]->type->vector_elements);
1265   ir_constant *c55555555 =
1266      new(ir) ir_constant(0x55555555u, ir->operands[0]->type->vector_elements);
1267   ir_constant *c0F0F0F0F =
1268      new(ir) ir_constant(0x0F0F0F0Fu, ir->operands[0]->type->vector_elements);
1269   ir_constant *c00FF00FF =
1270      new(ir) ir_constant(0x00FF00FFu, ir->operands[0]->type->vector_elements);
1271   ir_variable *temp =
1272      new(ir) ir_variable(glsl_type::uvec(ir->operands[0]->type->vector_elements),
1273                          "temp", ir_var_temporary);
1274   ir_instruction &i = *base_ir;
1275
1276   i.insert_before(temp);
1277
1278   if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) {
1279      i.insert_before(assign(temp, ir->operands[0]));
1280   } else {
1281      assert(ir->operands[0]->type->base_type == GLSL_TYPE_INT);
1282      i.insert_before(assign(temp, i2u(ir->operands[0])));
1283   }
1284
1285   /* Swap odd and even bits.
1286    *
1287    * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1288    */
1289   i.insert_before(assign(temp, bit_or(bit_and(rshift(temp, c1), c55555555),
1290                                       lshift(bit_and(temp, c55555555->clone(ir, NULL)),
1291                                              c1->clone(ir, NULL)))));
1292   /* Swap consecutive pairs.
1293    *
1294    * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1295    */
1296   i.insert_before(assign(temp, bit_or(bit_and(rshift(temp, c2), c33333333),
1297                                       lshift(bit_and(temp, c33333333->clone(ir, NULL)),
1298                                              c2->clone(ir, NULL)))));
1299
1300   /* Swap nibbles.
1301    *
1302    * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1303    */
1304   i.insert_before(assign(temp, bit_or(bit_and(rshift(temp, c4), c0F0F0F0F),
1305                                       lshift(bit_and(temp, c0F0F0F0F->clone(ir, NULL)),
1306                                              c4->clone(ir, NULL)))));
1307
1308   /* The last step is, basically, bswap.  Swap the bytes, then swap the
1309    * words.  When this code is run through GCC on x86, it does generate a
1310    * bswap instruction.
1311    *
1312    * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1313    * temp = ( temp >> 16              ) | ( temp                << 16);
1314    */
1315   i.insert_before(assign(temp, bit_or(bit_and(rshift(temp, c8), c00FF00FF),
1316                                       lshift(bit_and(temp, c00FF00FF->clone(ir, NULL)),
1317                                              c8->clone(ir, NULL)))));
1318
1319   if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) {
1320      ir->operation = ir_binop_bit_or;
1321      ir->init_num_operands();
1322      ir->operands[0] = rshift(temp, c16);
1323      ir->operands[1] = lshift(temp, c16->clone(ir, NULL));
1324   } else {
1325      ir->operation = ir_unop_u2i;
1326      ir->init_num_operands();
1327      ir->operands[0] = bit_or(rshift(temp, c16),
1328                               lshift(temp, c16->clone(ir, NULL)));
1329   }
1330
1331   this->progress = true;
1332}
1333
1334void
1335lower_instructions_visitor::find_lsb_to_float_cast(ir_expression *ir)
1336{
1337   /* For more details, see:
1338    *
1339    * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1340    */
1341   const unsigned elements = ir->operands[0]->type->vector_elements;
1342   ir_constant *c0 = new(ir) ir_constant(unsigned(0), elements);
1343   ir_constant *cminus1 = new(ir) ir_constant(int(-1), elements);
1344   ir_constant *c23 = new(ir) ir_constant(int(23), elements);
1345   ir_constant *c7F = new(ir) ir_constant(int(0x7F), elements);
1346   ir_variable *temp =
1347      new(ir) ir_variable(glsl_type::ivec(elements), "temp", ir_var_temporary);
1348   ir_variable *lsb_only =
1349      new(ir) ir_variable(glsl_type::uvec(elements), "lsb_only", ir_var_temporary);
1350   ir_variable *as_float =
1351      new(ir) ir_variable(glsl_type::vec(elements), "as_float", ir_var_temporary);
1352   ir_variable *lsb =
1353      new(ir) ir_variable(glsl_type::ivec(elements), "lsb", ir_var_temporary);
1354
1355   ir_instruction &i = *base_ir;
1356
1357   i.insert_before(temp);
1358
1359   if (ir->operands[0]->type->base_type == GLSL_TYPE_INT) {
1360      i.insert_before(assign(temp, ir->operands[0]));
1361   } else {
1362      assert(ir->operands[0]->type->base_type == GLSL_TYPE_UINT);
1363      i.insert_before(assign(temp, u2i(ir->operands[0])));
1364   }
1365
1366   /* The int-to-float conversion is lossless because (value & -value) is
1367    * either a power of two or zero.  We don't use the result in the zero
1368    * case.  The uint() cast is necessary so that 0x80000000 does not
1369    * generate a negative value.
1370    *
1371    * uint lsb_only = uint(value & -value);
1372    * float as_float = float(lsb_only);
1373    */
1374   i.insert_before(lsb_only);
1375   i.insert_before(assign(lsb_only, i2u(bit_and(temp, neg(temp)))));
1376
1377   i.insert_before(as_float);
1378   i.insert_before(assign(as_float, u2f(lsb_only)));
1379
1380   /* This is basically an open-coded frexp.  Implementations that have a
1381    * native frexp instruction would be better served by that.  This is
1382    * optimized versus a full-featured open-coded implementation in two ways:
1383    *
1384    * - We don't care about a correct result from subnormal numbers (including
1385    *   0.0), so the raw exponent can always be safely unbiased.
1386    *
1387    * - The value cannot be negative, so it does not need to be masked off to
1388    *   extract the exponent.
1389    *
1390    * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1391    */
1392   i.insert_before(lsb);
1393   i.insert_before(assign(lsb, sub(rshift(bitcast_f2i(as_float), c23), c7F)));
1394
1395   /* Use lsb_only in the comparison instead of temp so that the & (far above)
1396    * can possibly generate the result without an explicit comparison.
1397    *
1398    * (lsb_only == 0) ? -1 : lsb;
1399    *
1400    * Since our input values are all integers, the unbiased exponent must not
1401    * be negative.  It will only be negative (-0x7f, in fact) if lsb_only is
1402    * 0.  Instead of using (lsb_only == 0), we could use (lsb >= 0).  Which is
1403    * better is likely GPU dependent.  Either way, the difference should be
1404    * small.
1405    */
1406   ir->operation = ir_triop_csel;
1407   ir->init_num_operands();
1408   ir->operands[0] = equal(lsb_only, c0);
1409   ir->operands[1] = cminus1;
1410   ir->operands[2] = new(ir) ir_dereference_variable(lsb);
1411
1412   this->progress = true;
1413}
1414
1415void
1416lower_instructions_visitor::find_msb_to_float_cast(ir_expression *ir)
1417{
1418   /* For more details, see:
1419    *
1420    * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1421    */
1422   const unsigned elements = ir->operands[0]->type->vector_elements;
1423   ir_constant *c0 = new(ir) ir_constant(int(0), elements);
1424   ir_constant *cminus1 = new(ir) ir_constant(int(-1), elements);
1425   ir_constant *c23 = new(ir) ir_constant(int(23), elements);
1426   ir_constant *c7F = new(ir) ir_constant(int(0x7F), elements);
1427   ir_constant *c000000FF = new(ir) ir_constant(0x000000FFu, elements);
1428   ir_constant *cFFFFFF00 = new(ir) ir_constant(0xFFFFFF00u, elements);
1429   ir_variable *temp =
1430      new(ir) ir_variable(glsl_type::uvec(elements), "temp", ir_var_temporary);
1431   ir_variable *as_float =
1432      new(ir) ir_variable(glsl_type::vec(elements), "as_float", ir_var_temporary);
1433   ir_variable *msb =
1434      new(ir) ir_variable(glsl_type::ivec(elements), "msb", ir_var_temporary);
1435
1436   ir_instruction &i = *base_ir;
1437
1438   i.insert_before(temp);
1439
1440   if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) {
1441      i.insert_before(assign(temp, ir->operands[0]));
1442   } else {
1443      assert(ir->operands[0]->type->base_type == GLSL_TYPE_INT);
1444
1445      /* findMSB(uint(abs(some_int))) almost always does the right thing.
1446       * There are two problem values:
1447       *
1448       * * 0x80000000.  Since abs(0x80000000) == 0x80000000, findMSB returns
1449       *   31.  However, findMSB(int(0x80000000)) == 30.
1450       *
1451       * * 0xffffffff.  Since abs(0xffffffff) == 1, findMSB returns
1452       *   31.  Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1453       *
1454       *    For a value of zero or negative one, -1 will be returned.
1455       *
1456       * For all negative number cases, including 0x80000000 and 0xffffffff,
1457       * the correct value is obtained from findMSB if instead of negating the
1458       * (already negative) value the logical-not is used.  A conditonal
1459       * logical-not can be achieved in two instructions.
1460       */
1461      ir_variable *as_int =
1462         new(ir) ir_variable(glsl_type::ivec(elements), "as_int", ir_var_temporary);
1463      ir_constant *c31 = new(ir) ir_constant(int(31), elements);
1464
1465      i.insert_before(as_int);
1466      i.insert_before(assign(as_int, ir->operands[0]));
1467      i.insert_before(assign(temp, i2u(expr(ir_binop_bit_xor,
1468                                            as_int,
1469                                            rshift(as_int, c31)))));
1470   }
1471
1472   /* The int-to-float conversion is lossless because bits are conditionally
1473    * masked off the bottom of temp to ensure the value has at most 24 bits of
1474    * data or is zero.  We don't use the result in the zero case.  The uint()
1475    * cast is necessary so that 0x80000000 does not generate a negative value.
1476    *
1477    * float as_float = float(temp > 255 ? temp & ~255 : temp);
1478    */
1479   i.insert_before(as_float);
1480   i.insert_before(assign(as_float, u2f(csel(greater(temp, c000000FF),
1481                                             bit_and(temp, cFFFFFF00),
1482                                             temp))));
1483
1484   /* This is basically an open-coded frexp.  Implementations that have a
1485    * native frexp instruction would be better served by that.  This is
1486    * optimized versus a full-featured open-coded implementation in two ways:
1487    *
1488    * - We don't care about a correct result from subnormal numbers (including
1489    *   0.0), so the raw exponent can always be safely unbiased.
1490    *
1491    * - The value cannot be negative, so it does not need to be masked off to
1492    *   extract the exponent.
1493    *
1494    * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1495    */
1496   i.insert_before(msb);
1497   i.insert_before(assign(msb, sub(rshift(bitcast_f2i(as_float), c23), c7F)));
1498
1499   /* Use msb in the comparison instead of temp so that the subtract can
1500    * possibly generate the result without an explicit comparison.
1501    *
1502    * (msb < 0) ? -1 : msb;
1503    *
1504    * Since our input values are all integers, the unbiased exponent must not
1505    * be negative.  It will only be negative (-0x7f, in fact) if temp is 0.
1506    */
1507   ir->operation = ir_triop_csel;
1508   ir->init_num_operands();
1509   ir->operands[0] = less(msb, c0);
1510   ir->operands[1] = cminus1;
1511   ir->operands[2] = new(ir) ir_dereference_variable(msb);
1512
1513   this->progress = true;
1514}
1515
1516ir_expression *
1517lower_instructions_visitor::_carry(operand a, operand b)
1518{
1519   if (lowering(CARRY_TO_ARITH))
1520      return i2u(b2i(less(add(a, b),
1521                          a.val->clone(ralloc_parent(a.val), NULL))));
1522   else
1523      return carry(a, b);
1524}
1525
1526ir_constant *
1527lower_instructions_visitor::_imm_fp(void *mem_ctx,
1528                                    const glsl_type *type,
1529                                    double f,
1530                                    unsigned vector_elements)
1531{
1532   switch (type->base_type) {
1533   case GLSL_TYPE_FLOAT:
1534      return new(mem_ctx) ir_constant((float) f, vector_elements);
1535   case GLSL_TYPE_DOUBLE:
1536      return new(mem_ctx) ir_constant((double) f, vector_elements);
1537   case GLSL_TYPE_FLOAT16:
1538      return new(mem_ctx) ir_constant(float16_t(f), vector_elements);
1539   default:
1540      assert(!"unknown float type for immediate");
1541      return NULL;
1542   }
1543}
1544
1545void
1546lower_instructions_visitor::imul_high_to_mul(ir_expression *ir)
1547{
1548   /*   ABCD
1549    * * EFGH
1550    * ======
1551    * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
1552    *
1553    * In GLSL, (a * b) becomes
1554    *
1555    * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu);
1556    * uint m2 = (a & 0x0000ffffu) * (b >> 16);
1557    * uint m3 = (a >> 16)         * (b & 0x0000ffffu);
1558    * uint m4 = (a >> 16)         * (b >> 16);
1559    *
1560    * uint c1;
1561    * uint c2;
1562    * uint lo_result;
1563    * uint hi_result;
1564    *
1565    * lo_result = uaddCarry(m1, m2 << 16, c1);
1566    * hi_result = m4 + c1;
1567    * lo_result = uaddCarry(lo_result, m3 << 16, c2);
1568    * hi_result = hi_result + c2;
1569    * hi_result = hi_result + (m2 >> 16) + (m3 >> 16);
1570    */
1571   const unsigned elements = ir->operands[0]->type->vector_elements;
1572   ir_variable *src1 =
1573      new(ir) ir_variable(glsl_type::uvec(elements), "src1", ir_var_temporary);
1574   ir_variable *src1h =
1575      new(ir) ir_variable(glsl_type::uvec(elements), "src1h", ir_var_temporary);
1576   ir_variable *src1l =
1577      new(ir) ir_variable(glsl_type::uvec(elements), "src1l", ir_var_temporary);
1578   ir_variable *src2 =
1579      new(ir) ir_variable(glsl_type::uvec(elements), "src2", ir_var_temporary);
1580   ir_variable *src2h =
1581      new(ir) ir_variable(glsl_type::uvec(elements), "src2h", ir_var_temporary);
1582   ir_variable *src2l =
1583      new(ir) ir_variable(glsl_type::uvec(elements), "src2l", ir_var_temporary);
1584   ir_variable *t1 =
1585      new(ir) ir_variable(glsl_type::uvec(elements), "t1", ir_var_temporary);
1586   ir_variable *t2 =
1587      new(ir) ir_variable(glsl_type::uvec(elements), "t2", ir_var_temporary);
1588   ir_variable *lo =
1589      new(ir) ir_variable(glsl_type::uvec(elements), "lo", ir_var_temporary);
1590   ir_variable *hi =
1591      new(ir) ir_variable(glsl_type::uvec(elements), "hi", ir_var_temporary);
1592   ir_variable *different_signs = NULL;
1593   ir_constant *c0000FFFF = new(ir) ir_constant(0x0000FFFFu, elements);
1594   ir_constant *c16 = new(ir) ir_constant(16u, elements);
1595
1596   ir_instruction &i = *base_ir;
1597
1598   i.insert_before(src1);
1599   i.insert_before(src2);
1600   i.insert_before(src1h);
1601   i.insert_before(src2h);
1602   i.insert_before(src1l);
1603   i.insert_before(src2l);
1604
1605   if (ir->operands[0]->type->base_type == GLSL_TYPE_UINT) {
1606      i.insert_before(assign(src1, ir->operands[0]));
1607      i.insert_before(assign(src2, ir->operands[1]));
1608   } else {
1609      assert(ir->operands[0]->type->base_type == GLSL_TYPE_INT);
1610
1611      ir_variable *itmp1 =
1612         new(ir) ir_variable(glsl_type::ivec(elements), "itmp1", ir_var_temporary);
1613      ir_variable *itmp2 =
1614         new(ir) ir_variable(glsl_type::ivec(elements), "itmp2", ir_var_temporary);
1615      ir_constant *c0 = new(ir) ir_constant(int(0), elements);
1616
1617      i.insert_before(itmp1);
1618      i.insert_before(itmp2);
1619      i.insert_before(assign(itmp1, ir->operands[0]));
1620      i.insert_before(assign(itmp2, ir->operands[1]));
1621
1622      different_signs =
1623         new(ir) ir_variable(glsl_type::bvec(elements), "different_signs",
1624                             ir_var_temporary);
1625
1626      i.insert_before(different_signs);
1627      i.insert_before(assign(different_signs, expr(ir_binop_logic_xor,
1628                                                   less(itmp1, c0),
1629                                                   less(itmp2, c0->clone(ir, NULL)))));
1630
1631      i.insert_before(assign(src1, i2u(abs(itmp1))));
1632      i.insert_before(assign(src2, i2u(abs(itmp2))));
1633   }
1634
1635   i.insert_before(assign(src1l, bit_and(src1, c0000FFFF)));
1636   i.insert_before(assign(src2l, bit_and(src2, c0000FFFF->clone(ir, NULL))));
1637   i.insert_before(assign(src1h, rshift(src1, c16)));
1638   i.insert_before(assign(src2h, rshift(src2, c16->clone(ir, NULL))));
1639
1640   i.insert_before(lo);
1641   i.insert_before(hi);
1642   i.insert_before(t1);
1643   i.insert_before(t2);
1644
1645   i.insert_before(assign(lo, mul(src1l, src2l)));
1646   i.insert_before(assign(t1, mul(src1l, src2h)));
1647   i.insert_before(assign(t2, mul(src1h, src2l)));
1648   i.insert_before(assign(hi, mul(src1h, src2h)));
1649
1650   i.insert_before(assign(hi, add(hi, _carry(lo, lshift(t1, c16->clone(ir, NULL))))));
1651   i.insert_before(assign(lo,            add(lo, lshift(t1, c16->clone(ir, NULL)))));
1652
1653   i.insert_before(assign(hi, add(hi, _carry(lo, lshift(t2, c16->clone(ir, NULL))))));
1654   i.insert_before(assign(lo,            add(lo, lshift(t2, c16->clone(ir, NULL)))));
1655
1656   if (different_signs == NULL) {
1657      assert(ir->operands[0]->type->base_type == GLSL_TYPE_UINT);
1658
1659      ir->operation = ir_binop_add;
1660      ir->init_num_operands();
1661      ir->operands[0] = add(hi, rshift(t1, c16->clone(ir, NULL)));
1662      ir->operands[1] = rshift(t2, c16->clone(ir, NULL));
1663   } else {
1664      assert(ir->operands[0]->type->base_type == GLSL_TYPE_INT);
1665
1666      i.insert_before(assign(hi, add(add(hi, rshift(t1, c16->clone(ir, NULL))),
1667                                     rshift(t2, c16->clone(ir, NULL)))));
1668
1669      /* For channels where different_signs is set we have to perform a 64-bit
1670       * negation.  This is *not* the same as just negating the high 32-bits.
1671       * Consider -3 * 2.  The high 32-bits is 0, but the desired result is
1672       * -1, not -0!  Recall -x == ~x + 1.
1673       */
1674      ir_variable *neg_hi =
1675         new(ir) ir_variable(glsl_type::ivec(elements), "neg_hi", ir_var_temporary);
1676      ir_constant *c1 = new(ir) ir_constant(1u, elements);
1677
1678      i.insert_before(neg_hi);
1679      i.insert_before(assign(neg_hi, add(bit_not(u2i(hi)),
1680                                         u2i(_carry(bit_not(lo), c1)))));
1681
1682      ir->operation = ir_triop_csel;
1683      ir->init_num_operands();
1684      ir->operands[0] = new(ir) ir_dereference_variable(different_signs);
1685      ir->operands[1] = new(ir) ir_dereference_variable(neg_hi);
1686      ir->operands[2] = u2i(hi);
1687   }
1688}
1689
1690void
1691lower_instructions_visitor::sqrt_to_abs_sqrt(ir_expression *ir)
1692{
1693   ir->operands[0] = new(ir) ir_expression(ir_unop_abs, ir->operands[0]);
1694   this->progress = true;
1695}
1696
1697void
1698lower_instructions_visitor::mul64_to_mul_and_mul_high(ir_expression *ir)
1699{
1700   /* Lower 32x32-> 64 to
1701    *    msb = imul_high(x_lo, y_lo)
1702    *    lsb = mul(x_lo, y_lo)
1703    */
1704   const unsigned elements = ir->operands[0]->type->vector_elements;
1705
1706   const ir_expression_operation operation =
1707      ir->type->base_type == GLSL_TYPE_UINT64 ? ir_unop_pack_uint_2x32
1708                                              : ir_unop_pack_int_2x32;
1709
1710   const glsl_type *var_type = ir->type->base_type == GLSL_TYPE_UINT64
1711                               ? glsl_type::uvec(elements)
1712                               : glsl_type::ivec(elements);
1713
1714   const glsl_type *ret_type = ir->type->base_type == GLSL_TYPE_UINT64
1715                               ? glsl_type::uvec2_type
1716                               : glsl_type::ivec2_type;
1717
1718   ir_instruction &i = *base_ir;
1719
1720   ir_variable *msb =
1721      new(ir) ir_variable(var_type, "msb", ir_var_temporary);
1722   ir_variable *lsb =
1723      new(ir) ir_variable(var_type, "lsb", ir_var_temporary);
1724   ir_variable *x =
1725      new(ir) ir_variable(var_type, "x", ir_var_temporary);
1726   ir_variable *y =
1727      new(ir) ir_variable(var_type, "y", ir_var_temporary);
1728
1729   i.insert_before(x);
1730   i.insert_before(assign(x, ir->operands[0]));
1731   i.insert_before(y);
1732   i.insert_before(assign(y, ir->operands[1]));
1733   i.insert_before(msb);
1734   i.insert_before(lsb);
1735
1736   i.insert_before(assign(msb, imul_high(x, y)));
1737   i.insert_before(assign(lsb, mul(x, y)));
1738
1739   ir_rvalue *result[4] = {NULL};
1740   for (unsigned elem = 0; elem < elements; elem++) {
1741      ir_rvalue *val = new(ir) ir_expression(ir_quadop_vector, ret_type,
1742                                             swizzle(lsb, elem, 1),
1743                                             swizzle(msb, elem, 1), NULL, NULL);
1744      result[elem] = expr(operation, val);
1745   }
1746
1747   ir->operation = ir_quadop_vector;
1748   ir->init_num_operands();
1749   ir->operands[0] = result[0];
1750   ir->operands[1] = result[1];
1751   ir->operands[2] = result[2];
1752   ir->operands[3] = result[3];
1753
1754   this->progress = true;
1755}
1756
1757ir_visitor_status
1758lower_instructions_visitor::visit_leave(ir_expression *ir)
1759{
1760   switch (ir->operation) {
1761   case ir_binop_dot:
1762      if (ir->operands[0]->type->is_double())
1763         double_dot_to_fma(ir);
1764      break;
1765   case ir_triop_lrp:
1766      if (ir->operands[0]->type->is_double())
1767         double_lrp(ir);
1768      break;
1769   case ir_binop_sub:
1770      if (lowering(SUB_TO_ADD_NEG))
1771	 sub_to_add_neg(ir);
1772      break;
1773
1774   case ir_binop_div:
1775      if (ir->operands[1]->type->is_integer_32() && lowering(INT_DIV_TO_MUL_RCP))
1776	 int_div_to_mul_rcp(ir);
1777      else if ((ir->operands[1]->type->is_float_16_32() && lowering(FDIV_TO_MUL_RCP)) ||
1778               (ir->operands[1]->type->is_double() && lowering(DDIV_TO_MUL_RCP)))
1779	 div_to_mul_rcp(ir);
1780      break;
1781
1782   case ir_unop_exp:
1783      if (lowering(EXP_TO_EXP2))
1784	 exp_to_exp2(ir);
1785      break;
1786
1787   case ir_unop_log:
1788      if (lowering(LOG_TO_LOG2))
1789	 log_to_log2(ir);
1790      break;
1791
1792   case ir_binop_mod:
1793      if (lowering(MOD_TO_FLOOR) && ir->type->is_float_16_32_64())
1794	 mod_to_floor(ir);
1795      break;
1796
1797   case ir_binop_pow:
1798      if (lowering(POW_TO_EXP2))
1799	 pow_to_exp2(ir);
1800      break;
1801
1802   case ir_binop_ldexp:
1803      if (lowering(LDEXP_TO_ARITH) && ir->type->is_float())
1804         ldexp_to_arith(ir);
1805      if (lowering(DFREXP_DLDEXP_TO_ARITH) && ir->type->is_double())
1806         dldexp_to_arith(ir);
1807      break;
1808
1809   case ir_unop_frexp_exp:
1810      if (lowering(DFREXP_DLDEXP_TO_ARITH) && ir->operands[0]->type->is_double())
1811         dfrexp_exp_to_arith(ir);
1812      break;
1813
1814   case ir_unop_frexp_sig:
1815      if (lowering(DFREXP_DLDEXP_TO_ARITH) && ir->operands[0]->type->is_double())
1816         dfrexp_sig_to_arith(ir);
1817      break;
1818
1819   case ir_binop_carry:
1820      if (lowering(CARRY_TO_ARITH))
1821         carry_to_arith(ir);
1822      break;
1823
1824   case ir_binop_borrow:
1825      if (lowering(BORROW_TO_ARITH))
1826         borrow_to_arith(ir);
1827      break;
1828
1829   case ir_unop_saturate:
1830      if (lowering(SAT_TO_CLAMP))
1831         sat_to_clamp(ir);
1832      break;
1833
1834   case ir_unop_trunc:
1835      if (lowering(DOPS_TO_DFRAC) && ir->type->is_double())
1836         dtrunc_to_dfrac(ir);
1837      break;
1838
1839   case ir_unop_ceil:
1840      if (lowering(DOPS_TO_DFRAC) && ir->type->is_double())
1841         dceil_to_dfrac(ir);
1842      break;
1843
1844   case ir_unop_floor:
1845      if (lowering(DOPS_TO_DFRAC) && ir->type->is_double())
1846         dfloor_to_dfrac(ir);
1847      break;
1848
1849   case ir_unop_round_even:
1850      if (lowering(DOPS_TO_DFRAC) && ir->type->is_double())
1851         dround_even_to_dfrac(ir);
1852      break;
1853
1854   case ir_unop_sign:
1855      if (lowering(DOPS_TO_DFRAC) && ir->type->is_double())
1856         dsign_to_csel(ir);
1857      break;
1858
1859   case ir_unop_bit_count:
1860      if (lowering(BIT_COUNT_TO_MATH))
1861         bit_count_to_math(ir);
1862      break;
1863
1864   case ir_triop_bitfield_extract:
1865      if (lowering(EXTRACT_TO_SHIFTS))
1866         extract_to_shifts(ir);
1867      break;
1868
1869   case ir_quadop_bitfield_insert:
1870      if (lowering(INSERT_TO_SHIFTS))
1871         insert_to_shifts(ir);
1872      break;
1873
1874   case ir_unop_bitfield_reverse:
1875      if (lowering(REVERSE_TO_SHIFTS))
1876         reverse_to_shifts(ir);
1877      break;
1878
1879   case ir_unop_find_lsb:
1880      if (lowering(FIND_LSB_TO_FLOAT_CAST))
1881         find_lsb_to_float_cast(ir);
1882      break;
1883
1884   case ir_unop_find_msb:
1885      if (lowering(FIND_MSB_TO_FLOAT_CAST))
1886         find_msb_to_float_cast(ir);
1887      break;
1888
1889   case ir_binop_imul_high:
1890      if (lowering(IMUL_HIGH_TO_MUL))
1891         imul_high_to_mul(ir);
1892      break;
1893
1894   case ir_binop_mul:
1895      if (lowering(MUL64_TO_MUL_AND_MUL_HIGH) &&
1896          (ir->type->base_type == GLSL_TYPE_INT64 ||
1897           ir->type->base_type == GLSL_TYPE_UINT64) &&
1898          (ir->operands[0]->type->base_type == GLSL_TYPE_INT ||
1899           ir->operands[1]->type->base_type == GLSL_TYPE_UINT))
1900         mul64_to_mul_and_mul_high(ir);
1901      break;
1902
1903   case ir_unop_rsq:
1904   case ir_unop_sqrt:
1905      if (lowering(SQRT_TO_ABS_SQRT))
1906         sqrt_to_abs_sqrt(ir);
1907      break;
1908
1909   default:
1910      return visit_continue;
1911   }
1912
1913   return visit_continue;
1914}
1915