2 * Copyright © 2010 Intel Corporation
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:
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
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.
25 * \file lower_instructions.cpp
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.
32 * Currently supported transformations:
35 * - INT_DIV_TO_MUL_RCP
49 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
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.
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)).
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.
66 * FDIV_TO_MUL_RCP only lowers single-precision floating point division;
67 * DDIV_TO_MUL_RCP only lowers double-precision floating point division.
68 * DIV_TO_MUL_RCP is a convenience macro that sets both flags.
69 * INT_DIV_TO_MUL_RCP handles the integer case, converting to and from floating
70 * point so that RCP is possible.
72 * EXP_TO_EXP2 and LOG_TO_LOG2:
73 * ----------------------------
74 * Many GPUs don't have a base e log or exponent instruction, but they
75 * do have base 2 versions, so this pass converts exp and log to exp2
76 * and log2 operations.
80 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
81 * x**y to 2**(y * log2(x)).
85 * Breaks an ir_binop_mod expression down to (op0 - op1 * floor(op0 / op1))
87 * Many GPUs don't have a MOD instruction (945 and 965 included), and
88 * if we have to break it down like this anyway, it gives an
89 * opportunity to do things like constant fold the (1.0 / op1) easily.
91 * Note: before we used to implement this as op1 * fract(op / op1) but this
92 * implementation had significant precision errors.
96 * Converts ir_binop_ldexp to arithmetic and bit operations for float sources.
98 * DFREXP_DLDEXP_TO_ARITH:
100 * Converts ir_binop_ldexp, ir_unop_frexp_sig, and ir_unop_frexp_exp to
101 * arithmetic and bit ops for double arguments.
105 * Converts ir_carry into (x + y) < x.
109 * Converts ir_borrow into (x < y).
113 * Converts ir_unop_saturate into min(max(x, 0.0), 1.0)
117 * Converts double trunc, ceil, floor, round to fract
120 #include "c99_math.h"
121 #include "program/prog_instruction.h" /* for swizzle */
122 #include "compiler/glsl_types.h"
124 #include "ir_builder.h"
125 #include "ir_optimization.h"
127 using namespace ir_builder
;
131 class lower_instructions_visitor
: public ir_hierarchical_visitor
{
133 lower_instructions_visitor(unsigned lower
)
134 : progress(false), lower(lower
) { }
136 ir_visitor_status
visit_leave(ir_expression
*);
141 unsigned lower
; /** Bitfield of which operations to lower */
143 void sub_to_add_neg(ir_expression
*);
144 void div_to_mul_rcp(ir_expression
*);
145 void int_div_to_mul_rcp(ir_expression
*);
146 void mod_to_floor(ir_expression
*);
147 void exp_to_exp2(ir_expression
*);
148 void pow_to_exp2(ir_expression
*);
149 void log_to_log2(ir_expression
*);
150 void ldexp_to_arith(ir_expression
*);
151 void dldexp_to_arith(ir_expression
*);
152 void dfrexp_sig_to_arith(ir_expression
*);
153 void dfrexp_exp_to_arith(ir_expression
*);
154 void carry_to_arith(ir_expression
*);
155 void borrow_to_arith(ir_expression
*);
156 void sat_to_clamp(ir_expression
*);
157 void double_dot_to_fma(ir_expression
*);
158 void double_lrp(ir_expression
*);
159 void dceil_to_dfrac(ir_expression
*);
160 void dfloor_to_dfrac(ir_expression
*);
161 void dround_even_to_dfrac(ir_expression
*);
162 void dtrunc_to_dfrac(ir_expression
*);
163 void dsign_to_csel(ir_expression
*);
164 void bit_count_to_math(ir_expression
*);
165 void extract_to_shifts(ir_expression
*);
166 void insert_to_shifts(ir_expression
*);
167 void reverse_to_shifts(ir_expression
*ir
);
168 void find_lsb_to_float_cast(ir_expression
*ir
);
169 void find_msb_to_float_cast(ir_expression
*ir
);
170 void imul_high_to_mul(ir_expression
*ir
);
171 void sqrt_to_abs_sqrt(ir_expression
*ir
);
173 ir_expression
*_carry(operand a
, operand b
);
176 } /* anonymous namespace */
179 * Determine if a particular type of lowering should occur
181 #define lowering(x) (this->lower & x)
184 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
186 lower_instructions_visitor
v(what_to_lower
);
188 visit_list_elements(&v
, instructions
);
193 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
195 ir
->operation
= ir_binop_add
;
196 ir
->init_num_operands();
197 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
198 ir
->operands
[1], NULL
);
199 this->progress
= true;
203 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
205 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
207 /* New expression for the 1.0 / op1 */
209 expr
= new(ir
) ir_expression(ir_unop_rcp
,
210 ir
->operands
[1]->type
,
213 /* op0 / op1 -> op0 * (1.0 / op1) */
214 ir
->operation
= ir_binop_mul
;
215 ir
->init_num_operands();
216 ir
->operands
[1] = expr
;
218 this->progress
= true;
222 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
224 assert(ir
->operands
[1]->type
->is_integer());
226 /* Be careful with integer division -- we need to do it as a
227 * float and re-truncate, since rcp(n > 1) of an integer would
230 ir_rvalue
*op0
, *op1
;
231 const struct glsl_type
*vec_type
;
233 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
234 ir
->operands
[1]->type
->vector_elements
,
235 ir
->operands
[1]->type
->matrix_columns
);
237 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
238 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
240 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
242 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
244 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
245 ir
->operands
[0]->type
->vector_elements
,
246 ir
->operands
[0]->type
->matrix_columns
);
248 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
249 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
251 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
253 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
254 ir
->type
->vector_elements
,
255 ir
->type
->matrix_columns
);
257 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
259 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
260 ir
->operation
= ir_unop_f2i
;
261 ir
->operands
[0] = op0
;
263 ir
->operation
= ir_unop_i2u
;
264 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
266 ir
->init_num_operands();
267 ir
->operands
[1] = NULL
;
269 this->progress
= true;
273 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
275 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
277 ir
->operation
= ir_unop_exp2
;
278 ir
->init_num_operands();
279 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
280 ir
->operands
[0], log2_e
);
281 this->progress
= true;
285 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
287 ir_expression
*const log2_x
=
288 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
291 ir
->operation
= ir_unop_exp2
;
292 ir
->init_num_operands();
293 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
294 ir
->operands
[1], log2_x
);
295 ir
->operands
[1] = NULL
;
296 this->progress
= true;
300 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
302 ir
->operation
= ir_binop_mul
;
303 ir
->init_num_operands();
304 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
305 ir
->operands
[0], NULL
);
306 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
307 this->progress
= true;
311 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
313 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
315 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
317 this->base_ir
->insert_before(x
);
318 this->base_ir
->insert_before(y
);
320 ir_assignment
*const assign_x
=
321 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
322 ir
->operands
[0], NULL
);
323 ir_assignment
*const assign_y
=
324 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
325 ir
->operands
[1], NULL
);
327 this->base_ir
->insert_before(assign_x
);
328 this->base_ir
->insert_before(assign_y
);
330 ir_expression
*const div_expr
=
331 new(ir
) ir_expression(ir_binop_div
, x
->type
,
332 new(ir
) ir_dereference_variable(x
),
333 new(ir
) ir_dereference_variable(y
));
335 /* Don't generate new IR that would need to be lowered in an additional
338 if ((lowering(FDIV_TO_MUL_RCP
) && ir
->type
->is_float()) ||
339 (lowering(DDIV_TO_MUL_RCP
) && ir
->type
->is_double()))
340 div_to_mul_rcp(div_expr
);
342 ir_expression
*const floor_expr
=
343 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
345 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
346 dfloor_to_dfrac(floor_expr
);
348 ir_expression
*const mul_expr
=
349 new(ir
) ir_expression(ir_binop_mul
,
350 new(ir
) ir_dereference_variable(y
),
353 ir
->operation
= ir_binop_sub
;
354 ir
->init_num_operands();
355 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
356 ir
->operands
[1] = mul_expr
;
357 this->progress
= true;
361 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
364 * ir_binop_ldexp x exp
367 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
368 * resulting_biased_exp = extracted_biased_exp + exp;
370 * if (resulting_biased_exp < 1 || x == 0.0f) {
371 * return copysign(0.0, x);
374 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
375 * lshift(i2u(resulting_biased_exp), exp_shift));
377 * which we can't actually implement as such, since the GLSL IR doesn't
378 * have vectorized if-statements. We actually implement it without branches
379 * using conditional-select:
381 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
382 * resulting_biased_exp = extracted_biased_exp + exp;
384 * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
385 * gequal(resulting_biased_exp, 1);
386 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
387 * resulting_biased_exp = csel(is_not_zero_or_underflow,
388 * resulting_biased_exp, 0);
390 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
391 * lshift(i2u(resulting_biased_exp), exp_shift));
394 const unsigned vec_elem
= ir
->type
->vector_elements
;
397 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
398 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
401 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
403 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
405 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
407 /* Temporary variables */
408 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
409 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
411 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
414 ir_variable
*extracted_biased_exp
=
415 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
416 ir_variable
*resulting_biased_exp
=
417 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
419 ir_variable
*is_not_zero_or_underflow
=
420 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
422 ir_instruction
&i
= *base_ir
;
424 /* Copy <x> and <exp> arguments. */
426 i
.insert_before(assign(x
, ir
->operands
[0]));
427 i
.insert_before(exp
);
428 i
.insert_before(assign(exp
, ir
->operands
[1]));
430 /* Extract the biased exponent from <x>. */
431 i
.insert_before(extracted_biased_exp
);
432 i
.insert_before(assign(extracted_biased_exp
,
433 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
435 i
.insert_before(resulting_biased_exp
);
436 i
.insert_before(assign(resulting_biased_exp
,
437 add(extracted_biased_exp
, exp
)));
439 /* Test if result is ±0.0, subnormal, or underflow by checking if the
440 * resulting biased exponent would be less than 0x1. If so, the result is
441 * 0.0 with the sign of x. (Actually, invert the conditions so that
442 * immediate values are the second arguments, which is better for i965)
444 i
.insert_before(zero_sign_x
);
445 i
.insert_before(assign(zero_sign_x
,
446 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
448 i
.insert_before(is_not_zero_or_underflow
);
449 i
.insert_before(assign(is_not_zero_or_underflow
,
450 logic_and(nequal(x
, new(ir
) ir_constant(0.0f
, vec_elem
)),
451 gequal(resulting_biased_exp
,
452 new(ir
) ir_constant(0x1, vec_elem
)))));
453 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
455 i
.insert_before(assign(resulting_biased_exp
,
456 csel(is_not_zero_or_underflow
,
457 resulting_biased_exp
, zeroi
)));
459 /* We could test for overflows by checking if the resulting biased exponent
460 * would be greater than 0xFE. Turns out we don't need to because the GLSL
463 * "If this product is too large to be represented in the
464 * floating-point type, the result is undefined."
467 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
469 /* Don't generate new IR that would need to be lowered in an additional
472 if (!lowering(INSERT_TO_SHIFTS
)) {
473 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
474 ir
->operation
= ir_unop_bitcast_i2f
;
475 ir
->init_num_operands();
476 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
477 exp_shift_clone
, exp_width
);
478 ir
->operands
[1] = NULL
;
480 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x807fffffu
, vec_elem
);
481 ir
->operation
= ir_unop_bitcast_u2f
;
482 ir
->init_num_operands();
483 ir
->operands
[0] = bit_or(bit_and(bitcast_f2u(x
), sign_mantissa_mask
),
484 lshift(i2u(resulting_biased_exp
), exp_shift_clone
));
487 this->progress
= true;
491 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
493 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
494 * from the significand.
497 const unsigned vec_elem
= ir
->type
->vector_elements
;
500 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
501 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
504 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
506 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
508 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
509 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
510 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
512 /* Temporary variables */
513 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
514 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
516 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
519 ir_variable
*extracted_biased_exp
=
520 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
521 ir_variable
*resulting_biased_exp
=
522 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
524 ir_variable
*is_not_zero_or_underflow
=
525 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
527 ir_instruction
&i
= *base_ir
;
529 /* Copy <x> and <exp> arguments. */
531 i
.insert_before(assign(x
, ir
->operands
[0]));
532 i
.insert_before(exp
);
533 i
.insert_before(assign(exp
, ir
->operands
[1]));
535 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
536 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
537 dfrexp_exp_to_arith(frexp_exp
);
539 /* Extract the biased exponent from <x>. */
540 i
.insert_before(extracted_biased_exp
);
541 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
543 i
.insert_before(resulting_biased_exp
);
544 i
.insert_before(assign(resulting_biased_exp
,
545 add(extracted_biased_exp
, exp
)));
547 /* Test if result is ±0.0, subnormal, or underflow by checking if the
548 * resulting biased exponent would be less than 0x1. If so, the result is
549 * 0.0 with the sign of x. (Actually, invert the conditions so that
550 * immediate values are the second arguments, which is better for i965)
551 * TODO: Implement in a vector fashion.
553 i
.insert_before(zero_sign_x
);
554 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
555 ir_variable
*unpacked
=
556 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
557 i
.insert_before(unpacked
);
560 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
561 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
563 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
564 i
.insert_before(assign(zero_sign_x
,
565 expr(ir_unop_pack_double_2x32
, unpacked
),
568 i
.insert_before(is_not_zero_or_underflow
);
569 i
.insert_before(assign(is_not_zero_or_underflow
,
570 gequal(resulting_biased_exp
,
571 new(ir
) ir_constant(0x1, vec_elem
))));
572 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
574 i
.insert_before(assign(resulting_biased_exp
,
575 csel(is_not_zero_or_underflow
,
576 resulting_biased_exp
, zeroi
)));
578 /* We could test for overflows by checking if the resulting biased exponent
579 * would be greater than 0xFE. Turns out we don't need to because the GLSL
582 * "If this product is too large to be represented in the
583 * floating-point type, the result is undefined."
586 ir_rvalue
*results
[4] = {NULL
};
587 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
588 ir_variable
*unpacked
=
589 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
590 i
.insert_before(unpacked
);
593 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
595 ir_expression
*bfi
= bitfield_insert(
597 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
598 exp_shift
->clone(ir
, NULL
),
599 exp_width
->clone(ir
, NULL
));
601 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
603 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
606 ir
->operation
= ir_quadop_vector
;
607 ir
->init_num_operands();
608 ir
->operands
[0] = results
[0];
609 ir
->operands
[1] = results
[1];
610 ir
->operands
[2] = results
[2];
611 ir
->operands
[3] = results
[3];
613 /* Don't generate new IR that would need to be lowered in an additional
617 this->progress
= true;
621 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
623 const unsigned vec_elem
= ir
->type
->vector_elements
;
624 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
626 /* Double-precision floating-point values are stored as
631 * We're just extracting the significand here, so we only need to modify
632 * the upper 32-bit uint. Unfortunately we must extract each double
633 * independently as there is no vector version of unpackDouble.
636 ir_instruction
&i
= *base_ir
;
638 ir_variable
*is_not_zero
=
639 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
640 ir_rvalue
*results
[4] = {NULL
};
642 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
643 i
.insert_before(is_not_zero
);
646 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
648 /* TODO: Remake this as more vector-friendly when int64 support is
651 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
652 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
653 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
655 /* Exponent of double floating-point values in the range [0.5, 1.0). */
656 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
659 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
660 ir_variable
*unpacked
=
661 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
663 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
665 i
.insert_before(bits
);
666 i
.insert_before(unpacked
);
667 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
669 /* Manipulate the high uint to remove the exponent and replace it with
670 * either the default exponent or zero.
672 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
673 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
674 i
.insert_before(assign(bits
, bit_or(bits
,
675 csel(swizzle(is_not_zero
, elem
, 1),
678 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
679 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
682 /* Put the dvec back together */
683 ir
->operation
= ir_quadop_vector
;
684 ir
->init_num_operands();
685 ir
->operands
[0] = results
[0];
686 ir
->operands
[1] = results
[1];
687 ir
->operands
[2] = results
[2];
688 ir
->operands
[3] = results
[3];
690 this->progress
= true;
694 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
696 const unsigned vec_elem
= ir
->type
->vector_elements
;
697 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
698 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
700 /* Double-precision floating-point values are stored as
705 * We're just extracting the exponent here, so we only care about the upper
709 ir_instruction
&i
= *base_ir
;
711 ir_variable
*is_not_zero
=
712 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
713 ir_variable
*high_words
=
714 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
715 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
716 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
718 ir_rvalue
*absval
= abs(ir
->operands
[0]);
720 i
.insert_before(is_not_zero
);
721 i
.insert_before(high_words
);
722 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
724 /* Extract all of the upper uints. */
725 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
726 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
728 i
.insert_before(assign(high_words
,
729 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
733 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
734 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
736 /* For non-zero inputs, shift the exponent down and apply bias. */
737 ir
->operation
= ir_triop_csel
;
738 ir
->init_num_operands();
739 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
740 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
741 ir
->operands
[2] = izero
;
743 this->progress
= true;
747 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
752 * sum = ir_binop_add x y
753 * bcarry = ir_binop_less sum x
754 * carry = ir_unop_b2i bcarry
757 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
758 ir
->operation
= ir_unop_i2u
;
759 ir
->init_num_operands();
760 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
761 ir
->operands
[1] = NULL
;
763 this->progress
= true;
767 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
770 * ir_binop_borrow x y
772 * bcarry = ir_binop_less x y
773 * carry = ir_unop_b2i bcarry
776 ir
->operation
= ir_unop_i2u
;
777 ir
->init_num_operands();
778 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
779 ir
->operands
[1] = NULL
;
781 this->progress
= true;
785 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
790 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
793 ir
->operation
= ir_binop_min
;
794 ir
->init_num_operands();
795 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
797 new(ir
) ir_constant(0.0f
));
798 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
800 this->progress
= true;
804 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
806 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
808 this->base_ir
->insert_before(temp
);
810 int nc
= ir
->operands
[0]->type
->components();
811 for (int i
= nc
- 1; i
>= 1; i
--) {
812 ir_assignment
*assig
;
814 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
815 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
817 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
818 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
821 this->base_ir
->insert_before(assig
);
824 ir
->operation
= ir_triop_fma
;
825 ir
->init_num_operands();
826 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
827 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
828 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
830 this->progress
= true;
835 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
838 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
839 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
841 switch (op2
->type
->vector_elements
) {
843 swizval
= SWIZZLE_XXXX
;
846 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
847 swizval
= SWIZZLE_XYZW
;
851 ir
->operation
= ir_triop_fma
;
852 ir
->init_num_operands();
853 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
854 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
856 this->progress
= true;
860 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
864 * temp = sub(x, frtemp);
865 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
867 ir_instruction
&i
= *base_ir
;
868 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
869 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
870 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
873 i
.insert_before(frtemp
);
874 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
876 ir
->operation
= ir_binop_add
;
877 ir
->init_num_operands();
878 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
879 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
881 this->progress
= true;
885 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
889 * result = sub(x, frtemp);
891 ir
->operation
= ir_binop_sub
;
892 ir
->init_num_operands();
893 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
895 this->progress
= true;
898 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
903 * frtemp = frac(temp);
904 * t2 = sub(temp, frtemp);
905 * if (frac(x) == 0.5)
906 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
911 ir_instruction
&i
= *base_ir
;
912 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
914 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
916 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
918 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
919 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
920 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
922 i
.insert_before(temp
);
923 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
925 i
.insert_before(frtemp
);
926 i
.insert_before(assign(frtemp
, fract(temp
)));
929 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
931 ir
->operation
= ir_triop_csel
;
932 ir
->init_num_operands();
933 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
934 p5
->clone(ir
, NULL
));
935 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
939 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
941 this->progress
= true;
945 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
949 * temp = sub(x, frtemp);
950 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
952 ir_rvalue
*arg
= ir
->operands
[0];
953 ir_instruction
&i
= *base_ir
;
955 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
956 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
957 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
959 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
962 i
.insert_before(frtemp
);
963 i
.insert_before(assign(frtemp
, fract(arg
)));
964 i
.insert_before(temp
);
965 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
967 ir
->operation
= ir_triop_csel
;
968 ir
->init_num_operands();
969 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
970 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
971 ir
->operands
[2] = add(temp
,
972 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
973 zero
->clone(ir
, NULL
),
976 this->progress
= true;
980 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
983 * temp = x > 0.0 ? 1.0 : 0.0;
984 * result = x < 0.0 ? -1.0 : temp;
986 ir_rvalue
*arg
= ir
->operands
[0];
987 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
988 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
989 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
991 ir
->operation
= ir_triop_csel
;
992 ir
->init_num_operands();
993 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
994 zero
->clone(ir
, NULL
));
995 ir
->operands
[1] = neg_one
;
996 ir
->operands
[2] = csel(greater(arg
, zero
),
998 zero
->clone(ir
, NULL
));
1000 this->progress
= true;
1004 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
1006 /* For more details, see:
1008 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
1010 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1011 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
1013 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
1014 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
1015 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
1016 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
1017 ir_constant
*c1
= new(ir
) ir_constant(1u);
1018 ir_constant
*c2
= new(ir
) ir_constant(2u);
1019 ir_constant
*c4
= new(ir
) ir_constant(4u);
1020 ir_constant
*c24
= new(ir
) ir_constant(24u);
1022 base_ir
->insert_before(temp
);
1024 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1025 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
1027 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1028 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
1031 /* temp = temp - ((temp >> 1) & 0x55555555u); */
1032 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
1035 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
1036 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
1037 bit_and(rshift(temp
, c2
),
1038 c33333333
->clone(ir
, NULL
)))));
1040 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1041 ir
->operation
= ir_unop_u2i
;
1042 ir
->init_num_operands();
1043 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1047 this->progress
= true;
1051 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1054 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1056 base_ir
->insert_before(bits
);
1057 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1059 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1061 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1063 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1064 ir_constant
*cFFFFFFFF
=
1065 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1067 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1068 * we'd get a mask of 0 when bits is 32. Special case it.
1070 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1072 ir_expression
*mask
= csel(equal(bits
, c32
),
1074 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1076 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1078 * If bits is zero, the result will be zero.
1080 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1081 * select as in the signed integer case.
1083 * (value >> offset) & mask;
1085 ir
->operation
= ir_binop_bit_and
;
1086 ir
->init_num_operands();
1087 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1088 ir
->operands
[1] = mask
;
1089 ir
->operands
[2] = NULL
;
1092 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1094 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1096 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1098 /* temp = 32 - bits; */
1099 base_ir
->insert_before(temp
);
1100 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1102 /* expr = value << (temp - offset)) >> temp; */
1103 ir_expression
*expr
=
1104 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1106 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1108 * If bits is zero, the result will be zero.
1110 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1111 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1114 * (bits == 0) ? 0 : e;
1116 ir
->operation
= ir_triop_csel
;
1117 ir
->init_num_operands();
1118 ir
->operands
[0] = equal(c0
, bits
);
1119 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1120 ir
->operands
[2] = expr
;
1123 this->progress
= true;
1127 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1131 ir_constant
*cFFFFFFFF
;
1132 ir_variable
*offset
=
1133 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1135 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1137 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1139 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1140 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1141 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1142 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1144 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1146 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1147 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1148 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1151 base_ir
->insert_before(offset
);
1152 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1154 base_ir
->insert_before(bits
);
1155 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1157 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1158 * we'd get a mask of 0 when bits is 32. Special case it.
1160 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1162 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1164 * The result will be undefined if offset or bits is negative, or if the
1165 * sum of offset and bits is greater than the number of bits used to
1166 * store the operand.
1168 * Since it's undefined, there are a couple other ways this could be
1169 * implemented. The other way that was considered was to put the csel
1170 * around the whole thing:
1172 * final_result = bits == 32 ? insert : ... ;
1174 base_ir
->insert_before(mask
);
1176 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1178 lshift(sub(lshift(c1
, bits
),
1179 c1
->clone(ir
, NULL
)),
1182 /* (base & ~mask) | ((insert << offset) & mask) */
1183 ir
->operation
= ir_binop_bit_or
;
1184 ir
->init_num_operands();
1185 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1186 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1187 ir
->operands
[2] = NULL
;
1188 ir
->operands
[3] = NULL
;
1190 this->progress
= true;
1194 lower_instructions_visitor::reverse_to_shifts(ir_expression
*ir
)
1196 /* For more details, see:
1198 * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1201 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1203 new(ir
) ir_constant(2u, ir
->operands
[0]->type
->vector_elements
);
1205 new(ir
) ir_constant(4u, ir
->operands
[0]->type
->vector_elements
);
1207 new(ir
) ir_constant(8u, ir
->operands
[0]->type
->vector_elements
);
1209 new(ir
) ir_constant(16u, ir
->operands
[0]->type
->vector_elements
);
1210 ir_constant
*c33333333
=
1211 new(ir
) ir_constant(0x33333333u
, ir
->operands
[0]->type
->vector_elements
);
1212 ir_constant
*c55555555
=
1213 new(ir
) ir_constant(0x55555555u
, ir
->operands
[0]->type
->vector_elements
);
1214 ir_constant
*c0F0F0F0F
=
1215 new(ir
) ir_constant(0x0F0F0F0Fu
, ir
->operands
[0]->type
->vector_elements
);
1216 ir_constant
*c00FF00FF
=
1217 new(ir
) ir_constant(0x00FF00FFu
, ir
->operands
[0]->type
->vector_elements
);
1219 new(ir
) ir_variable(glsl_type::uvec(ir
->operands
[0]->type
->vector_elements
),
1220 "temp", ir_var_temporary
);
1221 ir_instruction
&i
= *base_ir
;
1223 i
.insert_before(temp
);
1225 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1226 i
.insert_before(assign(temp
, ir
->operands
[0]));
1228 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1229 i
.insert_before(assign(temp
, i2u(ir
->operands
[0])));
1232 /* Swap odd and even bits.
1234 * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1236 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c1
), c55555555
),
1237 lshift(bit_and(temp
, c55555555
->clone(ir
, NULL
)),
1238 c1
->clone(ir
, NULL
)))));
1239 /* Swap consecutive pairs.
1241 * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1243 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c2
), c33333333
),
1244 lshift(bit_and(temp
, c33333333
->clone(ir
, NULL
)),
1245 c2
->clone(ir
, NULL
)))));
1249 * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1251 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c4
), c0F0F0F0F
),
1252 lshift(bit_and(temp
, c0F0F0F0F
->clone(ir
, NULL
)),
1253 c4
->clone(ir
, NULL
)))));
1255 /* The last step is, basically, bswap. Swap the bytes, then swap the
1256 * words. When this code is run through GCC on x86, it does generate a
1257 * bswap instruction.
1259 * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1260 * temp = ( temp >> 16 ) | ( temp << 16);
1262 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c8
), c00FF00FF
),
1263 lshift(bit_and(temp
, c00FF00FF
->clone(ir
, NULL
)),
1264 c8
->clone(ir
, NULL
)))));
1266 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1267 ir
->operation
= ir_binop_bit_or
;
1268 ir
->init_num_operands();
1269 ir
->operands
[0] = rshift(temp
, c16
);
1270 ir
->operands
[1] = lshift(temp
, c16
->clone(ir
, NULL
));
1272 ir
->operation
= ir_unop_u2i
;
1273 ir
->init_num_operands();
1274 ir
->operands
[0] = bit_or(rshift(temp
, c16
),
1275 lshift(temp
, c16
->clone(ir
, NULL
)));
1278 this->progress
= true;
1282 lower_instructions_visitor::find_lsb_to_float_cast(ir_expression
*ir
)
1284 /* For more details, see:
1286 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1288 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1289 ir_constant
*c0
= new(ir
) ir_constant(unsigned(0), elements
);
1290 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1291 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1292 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1294 new(ir
) ir_variable(glsl_type::ivec(elements
), "temp", ir_var_temporary
);
1295 ir_variable
*lsb_only
=
1296 new(ir
) ir_variable(glsl_type::uvec(elements
), "lsb_only", ir_var_temporary
);
1297 ir_variable
*as_float
=
1298 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1300 new(ir
) ir_variable(glsl_type::ivec(elements
), "lsb", ir_var_temporary
);
1302 ir_instruction
&i
= *base_ir
;
1304 i
.insert_before(temp
);
1306 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1307 i
.insert_before(assign(temp
, ir
->operands
[0]));
1309 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1310 i
.insert_before(assign(temp
, u2i(ir
->operands
[0])));
1313 /* The int-to-float conversion is lossless because (value & -value) is
1314 * either a power of two or zero. We don't use the result in the zero
1315 * case. The uint() cast is necessary so that 0x80000000 does not
1316 * generate a negative value.
1318 * uint lsb_only = uint(value & -value);
1319 * float as_float = float(lsb_only);
1321 i
.insert_before(lsb_only
);
1322 i
.insert_before(assign(lsb_only
, i2u(bit_and(temp
, neg(temp
)))));
1324 i
.insert_before(as_float
);
1325 i
.insert_before(assign(as_float
, u2f(lsb_only
)));
1327 /* This is basically an open-coded frexp. Implementations that have a
1328 * native frexp instruction would be better served by that. This is
1329 * optimized versus a full-featured open-coded implementation in two ways:
1331 * - We don't care about a correct result from subnormal numbers (including
1332 * 0.0), so the raw exponent can always be safely unbiased.
1334 * - The value cannot be negative, so it does not need to be masked off to
1335 * extract the exponent.
1337 * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1339 i
.insert_before(lsb
);
1340 i
.insert_before(assign(lsb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1342 /* Use lsb_only in the comparison instead of temp so that the & (far above)
1343 * can possibly generate the result without an explicit comparison.
1345 * (lsb_only == 0) ? -1 : lsb;
1347 * Since our input values are all integers, the unbiased exponent must not
1348 * be negative. It will only be negative (-0x7f, in fact) if lsb_only is
1349 * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is
1350 * better is likely GPU dependent. Either way, the difference should be
1353 ir
->operation
= ir_triop_csel
;
1354 ir
->init_num_operands();
1355 ir
->operands
[0] = equal(lsb_only
, c0
);
1356 ir
->operands
[1] = cminus1
;
1357 ir
->operands
[2] = new(ir
) ir_dereference_variable(lsb
);
1359 this->progress
= true;
1363 lower_instructions_visitor::find_msb_to_float_cast(ir_expression
*ir
)
1365 /* For more details, see:
1367 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1369 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1370 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1371 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1372 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1373 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1374 ir_constant
*c000000FF
= new(ir
) ir_constant(0x000000FFu
, elements
);
1375 ir_constant
*cFFFFFF00
= new(ir
) ir_constant(0xFFFFFF00u
, elements
);
1377 new(ir
) ir_variable(glsl_type::uvec(elements
), "temp", ir_var_temporary
);
1378 ir_variable
*as_float
=
1379 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1381 new(ir
) ir_variable(glsl_type::ivec(elements
), "msb", ir_var_temporary
);
1383 ir_instruction
&i
= *base_ir
;
1385 i
.insert_before(temp
);
1387 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1388 i
.insert_before(assign(temp
, ir
->operands
[0]));
1390 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1392 /* findMSB(uint(abs(some_int))) almost always does the right thing.
1393 * There are two problem values:
1395 * * 0x80000000. Since abs(0x80000000) == 0x80000000, findMSB returns
1396 * 31. However, findMSB(int(0x80000000)) == 30.
1398 * * 0xffffffff. Since abs(0xffffffff) == 1, findMSB returns
1399 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1401 * For a value of zero or negative one, -1 will be returned.
1403 * For all negative number cases, including 0x80000000 and 0xffffffff,
1404 * the correct value is obtained from findMSB if instead of negating the
1405 * (already negative) value the logical-not is used. A conditonal
1406 * logical-not can be achieved in two instructions.
1408 ir_variable
*as_int
=
1409 new(ir
) ir_variable(glsl_type::ivec(elements
), "as_int", ir_var_temporary
);
1410 ir_constant
*c31
= new(ir
) ir_constant(int(31), elements
);
1412 i
.insert_before(as_int
);
1413 i
.insert_before(assign(as_int
, ir
->operands
[0]));
1414 i
.insert_before(assign(temp
, i2u(expr(ir_binop_bit_xor
,
1416 rshift(as_int
, c31
)))));
1419 /* The int-to-float conversion is lossless because bits are conditionally
1420 * masked off the bottom of temp to ensure the value has at most 24 bits of
1421 * data or is zero. We don't use the result in the zero case. The uint()
1422 * cast is necessary so that 0x80000000 does not generate a negative value.
1424 * float as_float = float(temp > 255 ? temp & ~255 : temp);
1426 i
.insert_before(as_float
);
1427 i
.insert_before(assign(as_float
, u2f(csel(greater(temp
, c000000FF
),
1428 bit_and(temp
, cFFFFFF00
),
1431 /* This is basically an open-coded frexp. Implementations that have a
1432 * native frexp instruction would be better served by that. This is
1433 * optimized versus a full-featured open-coded implementation in two ways:
1435 * - We don't care about a correct result from subnormal numbers (including
1436 * 0.0), so the raw exponent can always be safely unbiased.
1438 * - The value cannot be negative, so it does not need to be masked off to
1439 * extract the exponent.
1441 * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1443 i
.insert_before(msb
);
1444 i
.insert_before(assign(msb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1446 /* Use msb in the comparison instead of temp so that the subtract can
1447 * possibly generate the result without an explicit comparison.
1449 * (msb < 0) ? -1 : msb;
1451 * Since our input values are all integers, the unbiased exponent must not
1452 * be negative. It will only be negative (-0x7f, in fact) if temp is 0.
1454 ir
->operation
= ir_triop_csel
;
1455 ir
->init_num_operands();
1456 ir
->operands
[0] = less(msb
, c0
);
1457 ir
->operands
[1] = cminus1
;
1458 ir
->operands
[2] = new(ir
) ir_dereference_variable(msb
);
1460 this->progress
= true;
1464 lower_instructions_visitor::_carry(operand a
, operand b
)
1466 if (lowering(CARRY_TO_ARITH
))
1467 return i2u(b2i(less(add(a
, b
),
1468 a
.val
->clone(ralloc_parent(a
.val
), NULL
))));
1474 lower_instructions_visitor::imul_high_to_mul(ir_expression
*ir
)
1479 * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
1481 * In GLSL, (a * b) becomes
1483 * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu);
1484 * uint m2 = (a & 0x0000ffffu) * (b >> 16);
1485 * uint m3 = (a >> 16) * (b & 0x0000ffffu);
1486 * uint m4 = (a >> 16) * (b >> 16);
1493 * lo_result = uaddCarry(m1, m2 << 16, c1);
1494 * hi_result = m4 + c1;
1495 * lo_result = uaddCarry(lo_result, m3 << 16, c2);
1496 * hi_result = hi_result + c2;
1497 * hi_result = hi_result + (m2 >> 16) + (m3 >> 16);
1499 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1501 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1", ir_var_temporary
);
1502 ir_variable
*src1h
=
1503 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1h", ir_var_temporary
);
1504 ir_variable
*src1l
=
1505 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1l", ir_var_temporary
);
1507 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2", ir_var_temporary
);
1508 ir_variable
*src2h
=
1509 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2h", ir_var_temporary
);
1510 ir_variable
*src2l
=
1511 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2l", ir_var_temporary
);
1513 new(ir
) ir_variable(glsl_type::uvec(elements
), "t1", ir_var_temporary
);
1515 new(ir
) ir_variable(glsl_type::uvec(elements
), "t2", ir_var_temporary
);
1517 new(ir
) ir_variable(glsl_type::uvec(elements
), "lo", ir_var_temporary
);
1519 new(ir
) ir_variable(glsl_type::uvec(elements
), "hi", ir_var_temporary
);
1520 ir_variable
*different_signs
= NULL
;
1521 ir_constant
*c0000FFFF
= new(ir
) ir_constant(0x0000FFFFu
, elements
);
1522 ir_constant
*c16
= new(ir
) ir_constant(16u, elements
);
1524 ir_instruction
&i
= *base_ir
;
1526 i
.insert_before(src1
);
1527 i
.insert_before(src2
);
1528 i
.insert_before(src1h
);
1529 i
.insert_before(src2h
);
1530 i
.insert_before(src1l
);
1531 i
.insert_before(src2l
);
1533 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1534 i
.insert_before(assign(src1
, ir
->operands
[0]));
1535 i
.insert_before(assign(src2
, ir
->operands
[1]));
1537 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1539 ir_variable
*itmp1
=
1540 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp1", ir_var_temporary
);
1541 ir_variable
*itmp2
=
1542 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp2", ir_var_temporary
);
1543 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1545 i
.insert_before(itmp1
);
1546 i
.insert_before(itmp2
);
1547 i
.insert_before(assign(itmp1
, ir
->operands
[0]));
1548 i
.insert_before(assign(itmp2
, ir
->operands
[1]));
1551 new(ir
) ir_variable(glsl_type::bvec(elements
), "different_signs",
1554 i
.insert_before(different_signs
);
1555 i
.insert_before(assign(different_signs
, expr(ir_binop_logic_xor
,
1557 less(itmp2
, c0
->clone(ir
, NULL
)))));
1559 i
.insert_before(assign(src1
, i2u(abs(itmp1
))));
1560 i
.insert_before(assign(src2
, i2u(abs(itmp2
))));
1563 i
.insert_before(assign(src1l
, bit_and(src1
, c0000FFFF
)));
1564 i
.insert_before(assign(src2l
, bit_and(src2
, c0000FFFF
->clone(ir
, NULL
))));
1565 i
.insert_before(assign(src1h
, rshift(src1
, c16
)));
1566 i
.insert_before(assign(src2h
, rshift(src2
, c16
->clone(ir
, NULL
))));
1568 i
.insert_before(lo
);
1569 i
.insert_before(hi
);
1570 i
.insert_before(t1
);
1571 i
.insert_before(t2
);
1573 i
.insert_before(assign(lo
, mul(src1l
, src2l
)));
1574 i
.insert_before(assign(t1
, mul(src1l
, src2h
)));
1575 i
.insert_before(assign(t2
, mul(src1h
, src2l
)));
1576 i
.insert_before(assign(hi
, mul(src1h
, src2h
)));
1578 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t1
, c16
->clone(ir
, NULL
))))));
1579 i
.insert_before(assign(lo
, add(lo
, lshift(t1
, c16
->clone(ir
, NULL
)))));
1581 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t2
, c16
->clone(ir
, NULL
))))));
1582 i
.insert_before(assign(lo
, add(lo
, lshift(t2
, c16
->clone(ir
, NULL
)))));
1584 if (different_signs
== NULL
) {
1585 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1587 ir
->operation
= ir_binop_add
;
1588 ir
->init_num_operands();
1589 ir
->operands
[0] = add(hi
, rshift(t1
, c16
->clone(ir
, NULL
)));
1590 ir
->operands
[1] = rshift(t2
, c16
->clone(ir
, NULL
));
1592 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1594 i
.insert_before(assign(hi
, add(add(hi
, rshift(t1
, c16
->clone(ir
, NULL
))),
1595 rshift(t2
, c16
->clone(ir
, NULL
)))));
1597 /* For channels where different_signs is set we have to perform a 64-bit
1598 * negation. This is *not* the same as just negating the high 32-bits.
1599 * Consider -3 * 2. The high 32-bits is 0, but the desired result is
1600 * -1, not -0! Recall -x == ~x + 1.
1602 ir_variable
*neg_hi
=
1603 new(ir
) ir_variable(glsl_type::ivec(elements
), "neg_hi", ir_var_temporary
);
1604 ir_constant
*c1
= new(ir
) ir_constant(1u, elements
);
1606 i
.insert_before(neg_hi
);
1607 i
.insert_before(assign(neg_hi
, add(bit_not(u2i(hi
)),
1608 u2i(_carry(bit_not(lo
), c1
)))));
1610 ir
->operation
= ir_triop_csel
;
1611 ir
->init_num_operands();
1612 ir
->operands
[0] = new(ir
) ir_dereference_variable(different_signs
);
1613 ir
->operands
[1] = new(ir
) ir_dereference_variable(neg_hi
);
1614 ir
->operands
[2] = u2i(hi
);
1619 lower_instructions_visitor::sqrt_to_abs_sqrt(ir_expression
*ir
)
1621 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_abs
, ir
->operands
[0]);
1622 this->progress
= true;
1626 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1628 switch (ir
->operation
) {
1630 if (ir
->operands
[0]->type
->is_double())
1631 double_dot_to_fma(ir
);
1634 if (ir
->operands
[0]->type
->is_double())
1638 if (lowering(SUB_TO_ADD_NEG
))
1643 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1644 int_div_to_mul_rcp(ir
);
1645 else if ((ir
->operands
[1]->type
->is_float() && lowering(FDIV_TO_MUL_RCP
)) ||
1646 (ir
->operands
[1]->type
->is_double() && lowering(DDIV_TO_MUL_RCP
)))
1651 if (lowering(EXP_TO_EXP2
))
1656 if (lowering(LOG_TO_LOG2
))
1661 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1666 if (lowering(POW_TO_EXP2
))
1670 case ir_binop_ldexp
:
1671 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1673 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1674 dldexp_to_arith(ir
);
1677 case ir_unop_frexp_exp
:
1678 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1679 dfrexp_exp_to_arith(ir
);
1682 case ir_unop_frexp_sig
:
1683 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1684 dfrexp_sig_to_arith(ir
);
1687 case ir_binop_carry
:
1688 if (lowering(CARRY_TO_ARITH
))
1692 case ir_binop_borrow
:
1693 if (lowering(BORROW_TO_ARITH
))
1694 borrow_to_arith(ir
);
1697 case ir_unop_saturate
:
1698 if (lowering(SAT_TO_CLAMP
))
1703 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1704 dtrunc_to_dfrac(ir
);
1708 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1713 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1714 dfloor_to_dfrac(ir
);
1717 case ir_unop_round_even
:
1718 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1719 dround_even_to_dfrac(ir
);
1723 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1727 case ir_unop_bit_count
:
1728 if (lowering(BIT_COUNT_TO_MATH
))
1729 bit_count_to_math(ir
);
1732 case ir_triop_bitfield_extract
:
1733 if (lowering(EXTRACT_TO_SHIFTS
))
1734 extract_to_shifts(ir
);
1737 case ir_quadop_bitfield_insert
:
1738 if (lowering(INSERT_TO_SHIFTS
))
1739 insert_to_shifts(ir
);
1742 case ir_unop_bitfield_reverse
:
1743 if (lowering(REVERSE_TO_SHIFTS
))
1744 reverse_to_shifts(ir
);
1747 case ir_unop_find_lsb
:
1748 if (lowering(FIND_LSB_TO_FLOAT_CAST
))
1749 find_lsb_to_float_cast(ir
);
1752 case ir_unop_find_msb
:
1753 if (lowering(FIND_MSB_TO_FLOAT_CAST
))
1754 find_msb_to_float_cast(ir
);
1757 case ir_binop_imul_high
:
1758 if (lowering(IMUL_HIGH_TO_MUL
))
1759 imul_high_to_mul(ir
);
1764 if (lowering(SQRT_TO_ABS_SQRT
))
1765 sqrt_to_abs_sqrt(ir
);
1769 return visit_continue
;
1772 return visit_continue
;