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 * DIV_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 * DIV_TO_MUL_RCP only lowers floating point division; INT_DIV_TO_MUL_RCP
67 * handles the integer case, converting to and from floating point so that
70 * EXP_TO_EXP2 and LOG_TO_LOG2:
71 * ----------------------------
72 * Many GPUs don't have a base e log or exponent instruction, but they
73 * do have base 2 versions, so this pass converts exp and log to exp2
74 * and log2 operations.
78 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
79 * x**y to 2**(y * log2(x)).
83 * Breaks an ir_binop_mod expression down to (op0 - op1 * floor(op0 / op1))
85 * Many GPUs don't have a MOD instruction (945 and 965 included), and
86 * if we have to break it down like this anyway, it gives an
87 * opportunity to do things like constant fold the (1.0 / op1) easily.
89 * Note: before we used to implement this as op1 * fract(op / op1) but this
90 * implementation had significant precision errors.
94 * Converts ir_binop_ldexp to arithmetic and bit operations for float sources.
96 * DFREXP_DLDEXP_TO_ARITH:
98 * Converts ir_binop_ldexp, ir_unop_frexp_sig, and ir_unop_frexp_exp to
99 * arithmetic and bit ops for double arguments.
103 * Converts ir_carry into (x + y) < x.
107 * Converts ir_borrow into (x < y).
111 * Converts ir_unop_saturate into min(max(x, 0.0), 1.0)
115 * Converts double trunc, ceil, floor, round to fract
118 #include "c99_math.h"
119 #include "program/prog_instruction.h" /* for swizzle */
120 #include "compiler/glsl_types.h"
122 #include "ir_builder.h"
123 #include "ir_optimization.h"
125 using namespace ir_builder
;
129 class lower_instructions_visitor
: public ir_hierarchical_visitor
{
131 lower_instructions_visitor(unsigned lower
)
132 : progress(false), lower(lower
) { }
134 ir_visitor_status
visit_leave(ir_expression
*);
139 unsigned lower
; /** Bitfield of which operations to lower */
141 void sub_to_add_neg(ir_expression
*);
142 void div_to_mul_rcp(ir_expression
*);
143 void int_div_to_mul_rcp(ir_expression
*);
144 void mod_to_floor(ir_expression
*);
145 void exp_to_exp2(ir_expression
*);
146 void pow_to_exp2(ir_expression
*);
147 void log_to_log2(ir_expression
*);
148 void ldexp_to_arith(ir_expression
*);
149 void dldexp_to_arith(ir_expression
*);
150 void dfrexp_sig_to_arith(ir_expression
*);
151 void dfrexp_exp_to_arith(ir_expression
*);
152 void carry_to_arith(ir_expression
*);
153 void borrow_to_arith(ir_expression
*);
154 void sat_to_clamp(ir_expression
*);
155 void double_dot_to_fma(ir_expression
*);
156 void double_lrp(ir_expression
*);
157 void dceil_to_dfrac(ir_expression
*);
158 void dfloor_to_dfrac(ir_expression
*);
159 void dround_even_to_dfrac(ir_expression
*);
160 void dtrunc_to_dfrac(ir_expression
*);
161 void dsign_to_csel(ir_expression
*);
162 void bit_count_to_math(ir_expression
*);
163 void extract_to_shifts(ir_expression
*);
164 void insert_to_shifts(ir_expression
*);
165 void reverse_to_shifts(ir_expression
*ir
);
166 void find_lsb_to_float_cast(ir_expression
*ir
);
169 } /* anonymous namespace */
172 * Determine if a particular type of lowering should occur
174 #define lowering(x) (this->lower & x)
177 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
179 lower_instructions_visitor
v(what_to_lower
);
181 visit_list_elements(&v
, instructions
);
186 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
188 ir
->operation
= ir_binop_add
;
189 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
190 ir
->operands
[1], NULL
);
191 this->progress
= true;
195 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
197 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
199 /* New expression for the 1.0 / op1 */
201 expr
= new(ir
) ir_expression(ir_unop_rcp
,
202 ir
->operands
[1]->type
,
205 /* op0 / op1 -> op0 * (1.0 / op1) */
206 ir
->operation
= ir_binop_mul
;
207 ir
->operands
[1] = expr
;
209 this->progress
= true;
213 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
215 assert(ir
->operands
[1]->type
->is_integer());
217 /* Be careful with integer division -- we need to do it as a
218 * float and re-truncate, since rcp(n > 1) of an integer would
221 ir_rvalue
*op0
, *op1
;
222 const struct glsl_type
*vec_type
;
224 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
225 ir
->operands
[1]->type
->vector_elements
,
226 ir
->operands
[1]->type
->matrix_columns
);
228 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
229 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
231 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
233 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
235 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
236 ir
->operands
[0]->type
->vector_elements
,
237 ir
->operands
[0]->type
->matrix_columns
);
239 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
240 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
242 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
244 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
245 ir
->type
->vector_elements
,
246 ir
->type
->matrix_columns
);
248 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
250 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
251 ir
->operation
= ir_unop_f2i
;
252 ir
->operands
[0] = op0
;
254 ir
->operation
= ir_unop_i2u
;
255 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
257 ir
->operands
[1] = NULL
;
259 this->progress
= true;
263 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
265 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
267 ir
->operation
= ir_unop_exp2
;
268 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
269 ir
->operands
[0], log2_e
);
270 this->progress
= true;
274 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
276 ir_expression
*const log2_x
=
277 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
280 ir
->operation
= ir_unop_exp2
;
281 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
282 ir
->operands
[1], log2_x
);
283 ir
->operands
[1] = NULL
;
284 this->progress
= true;
288 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
290 ir
->operation
= ir_binop_mul
;
291 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
292 ir
->operands
[0], NULL
);
293 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
294 this->progress
= true;
298 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
300 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
302 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
304 this->base_ir
->insert_before(x
);
305 this->base_ir
->insert_before(y
);
307 ir_assignment
*const assign_x
=
308 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
309 ir
->operands
[0], NULL
);
310 ir_assignment
*const assign_y
=
311 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
312 ir
->operands
[1], NULL
);
314 this->base_ir
->insert_before(assign_x
);
315 this->base_ir
->insert_before(assign_y
);
317 ir_expression
*const div_expr
=
318 new(ir
) ir_expression(ir_binop_div
, x
->type
,
319 new(ir
) ir_dereference_variable(x
),
320 new(ir
) ir_dereference_variable(y
));
322 /* Don't generate new IR that would need to be lowered in an additional
325 if (lowering(DIV_TO_MUL_RCP
) && (ir
->type
->is_float() || ir
->type
->is_double()))
326 div_to_mul_rcp(div_expr
);
328 ir_expression
*const floor_expr
=
329 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
331 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
332 dfloor_to_dfrac(floor_expr
);
334 ir_expression
*const mul_expr
=
335 new(ir
) ir_expression(ir_binop_mul
,
336 new(ir
) ir_dereference_variable(y
),
339 ir
->operation
= ir_binop_sub
;
340 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
341 ir
->operands
[1] = mul_expr
;
342 this->progress
= true;
346 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
349 * ir_binop_ldexp x exp
352 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
353 * resulting_biased_exp = extracted_biased_exp + exp;
355 * if (resulting_biased_exp < 1 || x == 0.0f) {
356 * return copysign(0.0, x);
359 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
360 * lshift(i2u(resulting_biased_exp), exp_shift));
362 * which we can't actually implement as such, since the GLSL IR doesn't
363 * have vectorized if-statements. We actually implement it without branches
364 * using conditional-select:
366 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
367 * resulting_biased_exp = extracted_biased_exp + exp;
369 * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
370 * gequal(resulting_biased_exp, 1);
371 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
372 * resulting_biased_exp = csel(is_not_zero_or_underflow,
373 * resulting_biased_exp, 0);
375 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
376 * lshift(i2u(resulting_biased_exp), exp_shift));
379 const unsigned vec_elem
= ir
->type
->vector_elements
;
382 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
383 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
386 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
388 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
390 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
391 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
393 /* Temporary variables */
394 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
395 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
397 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
400 ir_variable
*extracted_biased_exp
=
401 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
402 ir_variable
*resulting_biased_exp
=
403 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
405 ir_variable
*is_not_zero_or_underflow
=
406 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
408 ir_instruction
&i
= *base_ir
;
410 /* Copy <x> and <exp> arguments. */
412 i
.insert_before(assign(x
, ir
->operands
[0]));
413 i
.insert_before(exp
);
414 i
.insert_before(assign(exp
, ir
->operands
[1]));
416 /* Extract the biased exponent from <x>. */
417 i
.insert_before(extracted_biased_exp
);
418 i
.insert_before(assign(extracted_biased_exp
,
419 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
421 i
.insert_before(resulting_biased_exp
);
422 i
.insert_before(assign(resulting_biased_exp
,
423 add(extracted_biased_exp
, exp
)));
425 /* Test if result is ±0.0, subnormal, or underflow by checking if the
426 * resulting biased exponent would be less than 0x1. If so, the result is
427 * 0.0 with the sign of x. (Actually, invert the conditions so that
428 * immediate values are the second arguments, which is better for i965)
430 i
.insert_before(zero_sign_x
);
431 i
.insert_before(assign(zero_sign_x
,
432 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
434 i
.insert_before(is_not_zero_or_underflow
);
435 i
.insert_before(assign(is_not_zero_or_underflow
,
436 logic_and(nequal(x
, new(ir
) ir_constant(0.0f
, vec_elem
)),
437 gequal(resulting_biased_exp
,
438 new(ir
) ir_constant(0x1, vec_elem
)))));
439 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
441 i
.insert_before(assign(resulting_biased_exp
,
442 csel(is_not_zero_or_underflow
,
443 resulting_biased_exp
, zeroi
)));
445 /* We could test for overflows by checking if the resulting biased exponent
446 * would be greater than 0xFE. Turns out we don't need to because the GLSL
449 * "If this product is too large to be represented in the
450 * floating-point type, the result is undefined."
453 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
454 ir
->operation
= ir_unop_bitcast_i2f
;
455 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
456 exp_shift_clone
, exp_width
);
457 ir
->operands
[1] = NULL
;
459 this->progress
= true;
463 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
465 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
466 * from the significand.
469 const unsigned vec_elem
= ir
->type
->vector_elements
;
472 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
473 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
476 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
478 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
480 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
481 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
482 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
484 /* Temporary variables */
485 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
486 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
488 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
491 ir_variable
*extracted_biased_exp
=
492 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
493 ir_variable
*resulting_biased_exp
=
494 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
496 ir_variable
*is_not_zero_or_underflow
=
497 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
499 ir_instruction
&i
= *base_ir
;
501 /* Copy <x> and <exp> arguments. */
503 i
.insert_before(assign(x
, ir
->operands
[0]));
504 i
.insert_before(exp
);
505 i
.insert_before(assign(exp
, ir
->operands
[1]));
507 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
508 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
509 dfrexp_exp_to_arith(frexp_exp
);
511 /* Extract the biased exponent from <x>. */
512 i
.insert_before(extracted_biased_exp
);
513 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
515 i
.insert_before(resulting_biased_exp
);
516 i
.insert_before(assign(resulting_biased_exp
,
517 add(extracted_biased_exp
, exp
)));
519 /* Test if result is ±0.0, subnormal, or underflow by checking if the
520 * resulting biased exponent would be less than 0x1. If so, the result is
521 * 0.0 with the sign of x. (Actually, invert the conditions so that
522 * immediate values are the second arguments, which is better for i965)
523 * TODO: Implement in a vector fashion.
525 i
.insert_before(zero_sign_x
);
526 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
527 ir_variable
*unpacked
=
528 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
529 i
.insert_before(unpacked
);
532 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
533 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
535 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
536 i
.insert_before(assign(zero_sign_x
,
537 expr(ir_unop_pack_double_2x32
, unpacked
),
540 i
.insert_before(is_not_zero_or_underflow
);
541 i
.insert_before(assign(is_not_zero_or_underflow
,
542 gequal(resulting_biased_exp
,
543 new(ir
) ir_constant(0x1, vec_elem
))));
544 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
546 i
.insert_before(assign(resulting_biased_exp
,
547 csel(is_not_zero_or_underflow
,
548 resulting_biased_exp
, zeroi
)));
550 /* We could test for overflows by checking if the resulting biased exponent
551 * would be greater than 0xFE. Turns out we don't need to because the GLSL
554 * "If this product is too large to be represented in the
555 * floating-point type, the result is undefined."
558 ir_rvalue
*results
[4] = {NULL
};
559 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
560 ir_variable
*unpacked
=
561 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
562 i
.insert_before(unpacked
);
565 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
567 ir_expression
*bfi
= bitfield_insert(
569 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
570 exp_shift
->clone(ir
, NULL
),
571 exp_width
->clone(ir
, NULL
));
573 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
575 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
578 ir
->operation
= ir_quadop_vector
;
579 ir
->operands
[0] = results
[0];
580 ir
->operands
[1] = results
[1];
581 ir
->operands
[2] = results
[2];
582 ir
->operands
[3] = results
[3];
584 /* Don't generate new IR that would need to be lowered in an additional
588 this->progress
= true;
592 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
594 const unsigned vec_elem
= ir
->type
->vector_elements
;
595 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
597 /* Double-precision floating-point values are stored as
602 * We're just extracting the significand here, so we only need to modify
603 * the upper 32-bit uint. Unfortunately we must extract each double
604 * independently as there is no vector version of unpackDouble.
607 ir_instruction
&i
= *base_ir
;
609 ir_variable
*is_not_zero
=
610 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
611 ir_rvalue
*results
[4] = {NULL
};
613 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
614 i
.insert_before(is_not_zero
);
617 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
619 /* TODO: Remake this as more vector-friendly when int64 support is
622 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
623 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
624 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
626 /* Exponent of double floating-point values in the range [0.5, 1.0). */
627 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
630 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
631 ir_variable
*unpacked
=
632 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
634 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
636 i
.insert_before(bits
);
637 i
.insert_before(unpacked
);
638 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
640 /* Manipulate the high uint to remove the exponent and replace it with
641 * either the default exponent or zero.
643 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
644 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
645 i
.insert_before(assign(bits
, bit_or(bits
,
646 csel(swizzle(is_not_zero
, elem
, 1),
649 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
650 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
653 /* Put the dvec back together */
654 ir
->operation
= ir_quadop_vector
;
655 ir
->operands
[0] = results
[0];
656 ir
->operands
[1] = results
[1];
657 ir
->operands
[2] = results
[2];
658 ir
->operands
[3] = results
[3];
660 this->progress
= true;
664 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
666 const unsigned vec_elem
= ir
->type
->vector_elements
;
667 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
668 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
670 /* Double-precision floating-point values are stored as
675 * We're just extracting the exponent here, so we only care about the upper
679 ir_instruction
&i
= *base_ir
;
681 ir_variable
*is_not_zero
=
682 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
683 ir_variable
*high_words
=
684 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
685 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
686 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
688 ir_rvalue
*absval
= abs(ir
->operands
[0]);
690 i
.insert_before(is_not_zero
);
691 i
.insert_before(high_words
);
692 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
694 /* Extract all of the upper uints. */
695 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
696 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
698 i
.insert_before(assign(high_words
,
699 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
703 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
704 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
706 /* For non-zero inputs, shift the exponent down and apply bias. */
707 ir
->operation
= ir_triop_csel
;
708 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
709 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
710 ir
->operands
[2] = izero
;
712 this->progress
= true;
716 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
721 * sum = ir_binop_add x y
722 * bcarry = ir_binop_less sum x
723 * carry = ir_unop_b2i bcarry
726 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
727 ir
->operation
= ir_unop_i2u
;
728 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
729 ir
->operands
[1] = NULL
;
731 this->progress
= true;
735 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
738 * ir_binop_borrow x y
740 * bcarry = ir_binop_less x y
741 * carry = ir_unop_b2i bcarry
744 ir
->operation
= ir_unop_i2u
;
745 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
746 ir
->operands
[1] = NULL
;
748 this->progress
= true;
752 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
757 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
760 ir
->operation
= ir_binop_min
;
761 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
763 new(ir
) ir_constant(0.0f
));
764 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
766 this->progress
= true;
770 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
772 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
774 this->base_ir
->insert_before(temp
);
776 int nc
= ir
->operands
[0]->type
->components();
777 for (int i
= nc
- 1; i
>= 1; i
--) {
778 ir_assignment
*assig
;
780 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
781 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
783 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
784 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
787 this->base_ir
->insert_before(assig
);
790 ir
->operation
= ir_triop_fma
;
791 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
792 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
793 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
795 this->progress
= true;
800 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
803 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
804 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
806 switch (op2
->type
->vector_elements
) {
808 swizval
= SWIZZLE_XXXX
;
811 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
812 swizval
= SWIZZLE_XYZW
;
816 ir
->operation
= ir_triop_fma
;
817 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
818 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
820 this->progress
= true;
824 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
828 * temp = sub(x, frtemp);
829 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
831 ir_instruction
&i
= *base_ir
;
832 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
833 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
834 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
837 i
.insert_before(frtemp
);
838 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
840 ir
->operation
= ir_binop_add
;
841 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
842 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
844 this->progress
= true;
848 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
852 * result = sub(x, frtemp);
854 ir
->operation
= ir_binop_sub
;
855 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
857 this->progress
= true;
860 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
865 * frtemp = frac(temp);
866 * t2 = sub(temp, frtemp);
867 * if (frac(x) == 0.5)
868 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
873 ir_instruction
&i
= *base_ir
;
874 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
876 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
878 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
880 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
881 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
882 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
884 i
.insert_before(temp
);
885 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
887 i
.insert_before(frtemp
);
888 i
.insert_before(assign(frtemp
, fract(temp
)));
891 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
893 ir
->operation
= ir_triop_csel
;
894 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
895 p5
->clone(ir
, NULL
));
896 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
900 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
902 this->progress
= true;
906 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
910 * temp = sub(x, frtemp);
911 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
913 ir_rvalue
*arg
= ir
->operands
[0];
914 ir_instruction
&i
= *base_ir
;
916 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
917 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
918 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
920 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
923 i
.insert_before(frtemp
);
924 i
.insert_before(assign(frtemp
, fract(arg
)));
925 i
.insert_before(temp
);
926 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
928 ir
->operation
= ir_triop_csel
;
929 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
930 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
931 ir
->operands
[2] = add(temp
,
932 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
933 zero
->clone(ir
, NULL
),
936 this->progress
= true;
940 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
943 * temp = x > 0.0 ? 1.0 : 0.0;
944 * result = x < 0.0 ? -1.0 : temp;
946 ir_rvalue
*arg
= ir
->operands
[0];
947 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
948 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
949 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
951 ir
->operation
= ir_triop_csel
;
952 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
953 zero
->clone(ir
, NULL
));
954 ir
->operands
[1] = neg_one
;
955 ir
->operands
[2] = csel(greater(arg
, zero
),
957 zero
->clone(ir
, NULL
));
959 this->progress
= true;
963 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
965 /* For more details, see:
967 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
969 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
970 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
972 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
973 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
974 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
975 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
976 ir_constant
*c1
= new(ir
) ir_constant(1u);
977 ir_constant
*c2
= new(ir
) ir_constant(2u);
978 ir_constant
*c4
= new(ir
) ir_constant(4u);
979 ir_constant
*c24
= new(ir
) ir_constant(24u);
981 base_ir
->insert_before(temp
);
983 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
984 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
986 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
987 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
990 /* temp = temp - ((temp >> 1) & 0x55555555u); */
991 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
994 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
995 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
996 bit_and(rshift(temp
, c2
),
997 c33333333
->clone(ir
, NULL
)))));
999 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1000 ir
->operation
= ir_unop_u2i
;
1001 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1005 this->progress
= true;
1009 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1012 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1014 base_ir
->insert_before(bits
);
1015 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1017 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1019 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1021 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1022 ir_constant
*cFFFFFFFF
=
1023 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1025 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1026 * we'd get a mask of 0 when bits is 32. Special case it.
1028 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1030 ir_expression
*mask
= csel(equal(bits
, c32
),
1032 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1034 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1036 * If bits is zero, the result will be zero.
1038 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1039 * select as in the signed integer case.
1041 * (value >> offset) & mask;
1043 ir
->operation
= ir_binop_bit_and
;
1044 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1045 ir
->operands
[1] = mask
;
1046 ir
->operands
[2] = NULL
;
1049 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1051 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1053 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1055 /* temp = 32 - bits; */
1056 base_ir
->insert_before(temp
);
1057 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1059 /* expr = value << (temp - offset)) >> temp; */
1060 ir_expression
*expr
=
1061 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1063 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1065 * If bits is zero, the result will be zero.
1067 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1068 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1071 * (bits == 0) ? 0 : e;
1073 ir
->operation
= ir_triop_csel
;
1074 ir
->operands
[0] = equal(c0
, bits
);
1075 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1076 ir
->operands
[2] = expr
;
1079 this->progress
= true;
1083 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1087 ir_constant
*cFFFFFFFF
;
1088 ir_variable
*offset
=
1089 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1091 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1093 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1095 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1096 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1097 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1098 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1100 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1102 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1103 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1104 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1107 base_ir
->insert_before(offset
);
1108 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1110 base_ir
->insert_before(bits
);
1111 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1113 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1114 * we'd get a mask of 0 when bits is 32. Special case it.
1116 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1118 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1120 * The result will be undefined if offset or bits is negative, or if the
1121 * sum of offset and bits is greater than the number of bits used to
1122 * store the operand.
1124 * Since it's undefined, there are a couple other ways this could be
1125 * implemented. The other way that was considered was to put the csel
1126 * around the whole thing:
1128 * final_result = bits == 32 ? insert : ... ;
1130 base_ir
->insert_before(mask
);
1132 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1134 lshift(sub(lshift(c1
, bits
),
1135 c1
->clone(ir
, NULL
)),
1138 /* (base & ~mask) | ((insert << offset) & mask) */
1139 ir
->operation
= ir_binop_bit_or
;
1140 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1141 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1142 ir
->operands
[2] = NULL
;
1143 ir
->operands
[3] = NULL
;
1145 this->progress
= true;
1149 lower_instructions_visitor::reverse_to_shifts(ir_expression
*ir
)
1151 /* For more details, see:
1153 * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1156 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1158 new(ir
) ir_constant(2u, ir
->operands
[0]->type
->vector_elements
);
1160 new(ir
) ir_constant(4u, ir
->operands
[0]->type
->vector_elements
);
1162 new(ir
) ir_constant(8u, ir
->operands
[0]->type
->vector_elements
);
1164 new(ir
) ir_constant(16u, ir
->operands
[0]->type
->vector_elements
);
1165 ir_constant
*c33333333
=
1166 new(ir
) ir_constant(0x33333333u
, ir
->operands
[0]->type
->vector_elements
);
1167 ir_constant
*c55555555
=
1168 new(ir
) ir_constant(0x55555555u
, ir
->operands
[0]->type
->vector_elements
);
1169 ir_constant
*c0F0F0F0F
=
1170 new(ir
) ir_constant(0x0F0F0F0Fu
, ir
->operands
[0]->type
->vector_elements
);
1171 ir_constant
*c00FF00FF
=
1172 new(ir
) ir_constant(0x00FF00FFu
, ir
->operands
[0]->type
->vector_elements
);
1174 new(ir
) ir_variable(glsl_type::uvec(ir
->operands
[0]->type
->vector_elements
),
1175 "temp", ir_var_temporary
);
1176 ir_instruction
&i
= *base_ir
;
1178 i
.insert_before(temp
);
1180 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1181 i
.insert_before(assign(temp
, ir
->operands
[0]));
1183 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1184 i
.insert_before(assign(temp
, i2u(ir
->operands
[0])));
1187 /* Swap odd and even bits.
1189 * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1191 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c1
), c55555555
),
1192 lshift(bit_and(temp
, c55555555
->clone(ir
, NULL
)),
1193 c1
->clone(ir
, NULL
)))));
1194 /* Swap consecutive pairs.
1196 * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1198 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c2
), c33333333
),
1199 lshift(bit_and(temp
, c33333333
->clone(ir
, NULL
)),
1200 c2
->clone(ir
, NULL
)))));
1204 * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1206 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c4
), c0F0F0F0F
),
1207 lshift(bit_and(temp
, c0F0F0F0F
->clone(ir
, NULL
)),
1208 c4
->clone(ir
, NULL
)))));
1210 /* The last step is, basically, bswap. Swap the bytes, then swap the
1211 * words. When this code is run through GCC on x86, it does generate a
1212 * bswap instruction.
1214 * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1215 * temp = ( temp >> 16 ) | ( temp << 16);
1217 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c8
), c00FF00FF
),
1218 lshift(bit_and(temp
, c00FF00FF
->clone(ir
, NULL
)),
1219 c8
->clone(ir
, NULL
)))));
1221 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1222 ir
->operation
= ir_binop_bit_or
;
1223 ir
->operands
[0] = rshift(temp
, c16
);
1224 ir
->operands
[1] = lshift(temp
, c16
->clone(ir
, NULL
));
1226 ir
->operation
= ir_unop_u2i
;
1227 ir
->operands
[0] = bit_or(rshift(temp
, c16
),
1228 lshift(temp
, c16
->clone(ir
, NULL
)));
1231 this->progress
= true;
1235 lower_instructions_visitor::find_lsb_to_float_cast(ir_expression
*ir
)
1237 /* For more details, see:
1239 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1241 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1242 ir_constant
*c0
= new(ir
) ir_constant(unsigned(0), elements
);
1243 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1244 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1245 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1247 new(ir
) ir_variable(glsl_type::ivec(elements
), "temp", ir_var_temporary
);
1248 ir_variable
*lsb_only
=
1249 new(ir
) ir_variable(glsl_type::uvec(elements
), "lsb_only", ir_var_temporary
);
1250 ir_variable
*as_float
=
1251 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1253 new(ir
) ir_variable(glsl_type::ivec(elements
), "lsb", ir_var_temporary
);
1255 ir_instruction
&i
= *base_ir
;
1257 i
.insert_before(temp
);
1259 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1260 i
.insert_before(assign(temp
, ir
->operands
[0]));
1262 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1263 i
.insert_before(assign(temp
, u2i(ir
->operands
[0])));
1266 /* The int-to-float conversion is lossless because (value & -value) is
1267 * either a power of two or zero. We don't use the result in the zero
1268 * case. The uint() cast is necessary so that 0x80000000 does not
1269 * generate a negative value.
1271 * uint lsb_only = uint(value & -value);
1272 * float as_float = float(lsb_only);
1274 i
.insert_before(lsb_only
);
1275 i
.insert_before(assign(lsb_only
, i2u(bit_and(temp
, neg(temp
)))));
1277 i
.insert_before(as_float
);
1278 i
.insert_before(assign(as_float
, u2f(lsb_only
)));
1280 /* This is basically an open-coded frexp. Implementations that have a
1281 * native frexp instruction would be better served by that. This is
1282 * optimized versus a full-featured open-coded implementation in two ways:
1284 * - We don't care about a correct result from subnormal numbers (including
1285 * 0.0), so the raw exponent can always be safely unbiased.
1287 * - The value cannot be negative, so it does not need to be masked off to
1288 * extract the exponent.
1290 * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1292 i
.insert_before(lsb
);
1293 i
.insert_before(assign(lsb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1295 /* Use lsb_only in the comparison instead of temp so that the & (far above)
1296 * can possibly generate the result without an explicit comparison.
1298 * (lsb_only == 0) ? -1 : lsb;
1300 * Since our input values are all integers, the unbiased exponent must not
1301 * be negative. It will only be negative (-0x7f, in fact) if lsb_only is
1302 * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is
1303 * better is likely GPU dependent. Either way, the difference should be
1306 ir
->operation
= ir_triop_csel
;
1307 ir
->operands
[0] = equal(lsb_only
, c0
);
1308 ir
->operands
[1] = cminus1
;
1309 ir
->operands
[2] = new(ir
) ir_dereference_variable(lsb
);
1311 this->progress
= true;
1315 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1317 switch (ir
->operation
) {
1319 if (ir
->operands
[0]->type
->is_double())
1320 double_dot_to_fma(ir
);
1323 if (ir
->operands
[0]->type
->is_double())
1327 if (lowering(SUB_TO_ADD_NEG
))
1332 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1333 int_div_to_mul_rcp(ir
);
1334 else if ((ir
->operands
[1]->type
->is_float() ||
1335 ir
->operands
[1]->type
->is_double()) && lowering(DIV_TO_MUL_RCP
))
1340 if (lowering(EXP_TO_EXP2
))
1345 if (lowering(LOG_TO_LOG2
))
1350 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1355 if (lowering(POW_TO_EXP2
))
1359 case ir_binop_ldexp
:
1360 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1362 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1363 dldexp_to_arith(ir
);
1366 case ir_unop_frexp_exp
:
1367 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1368 dfrexp_exp_to_arith(ir
);
1371 case ir_unop_frexp_sig
:
1372 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1373 dfrexp_sig_to_arith(ir
);
1376 case ir_binop_carry
:
1377 if (lowering(CARRY_TO_ARITH
))
1381 case ir_binop_borrow
:
1382 if (lowering(BORROW_TO_ARITH
))
1383 borrow_to_arith(ir
);
1386 case ir_unop_saturate
:
1387 if (lowering(SAT_TO_CLAMP
))
1392 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1393 dtrunc_to_dfrac(ir
);
1397 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1402 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1403 dfloor_to_dfrac(ir
);
1406 case ir_unop_round_even
:
1407 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1408 dround_even_to_dfrac(ir
);
1412 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1416 case ir_unop_bit_count
:
1417 if (lowering(BIT_COUNT_TO_MATH
))
1418 bit_count_to_math(ir
);
1421 case ir_triop_bitfield_extract
:
1422 if (lowering(EXTRACT_TO_SHIFTS
))
1423 extract_to_shifts(ir
);
1426 case ir_quadop_bitfield_insert
:
1427 if (lowering(INSERT_TO_SHIFTS
))
1428 insert_to_shifts(ir
);
1431 case ir_unop_bitfield_reverse
:
1432 if (lowering(REVERSE_TO_SHIFTS
))
1433 reverse_to_shifts(ir
);
1436 case ir_unop_find_lsb
:
1437 if (lowering(FIND_LSB_TO_FLOAT_CAST
))
1438 find_lsb_to_float_cast(ir
);
1442 return visit_continue
;
1445 return visit_continue
;