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
);
172 ir_expression
*_carry(operand a
, operand b
);
175 } /* anonymous namespace */
178 * Determine if a particular type of lowering should occur
180 #define lowering(x) (this->lower & x)
183 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
185 lower_instructions_visitor
v(what_to_lower
);
187 visit_list_elements(&v
, instructions
);
192 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
194 ir
->operation
= ir_binop_add
;
195 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
196 ir
->operands
[1], NULL
);
197 this->progress
= true;
201 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
203 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
205 /* New expression for the 1.0 / op1 */
207 expr
= new(ir
) ir_expression(ir_unop_rcp
,
208 ir
->operands
[1]->type
,
211 /* op0 / op1 -> op0 * (1.0 / op1) */
212 ir
->operation
= ir_binop_mul
;
213 ir
->operands
[1] = expr
;
215 this->progress
= true;
219 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
221 assert(ir
->operands
[1]->type
->is_integer());
223 /* Be careful with integer division -- we need to do it as a
224 * float and re-truncate, since rcp(n > 1) of an integer would
227 ir_rvalue
*op0
, *op1
;
228 const struct glsl_type
*vec_type
;
230 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
231 ir
->operands
[1]->type
->vector_elements
,
232 ir
->operands
[1]->type
->matrix_columns
);
234 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
235 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
237 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
239 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
241 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
242 ir
->operands
[0]->type
->vector_elements
,
243 ir
->operands
[0]->type
->matrix_columns
);
245 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
246 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
248 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
250 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
251 ir
->type
->vector_elements
,
252 ir
->type
->matrix_columns
);
254 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
256 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
257 ir
->operation
= ir_unop_f2i
;
258 ir
->operands
[0] = op0
;
260 ir
->operation
= ir_unop_i2u
;
261 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
263 ir
->operands
[1] = NULL
;
265 this->progress
= true;
269 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
271 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
273 ir
->operation
= ir_unop_exp2
;
274 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
275 ir
->operands
[0], log2_e
);
276 this->progress
= true;
280 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
282 ir_expression
*const log2_x
=
283 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
286 ir
->operation
= ir_unop_exp2
;
287 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
288 ir
->operands
[1], log2_x
);
289 ir
->operands
[1] = NULL
;
290 this->progress
= true;
294 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
296 ir
->operation
= ir_binop_mul
;
297 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
298 ir
->operands
[0], NULL
);
299 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
300 this->progress
= true;
304 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
306 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
308 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
310 this->base_ir
->insert_before(x
);
311 this->base_ir
->insert_before(y
);
313 ir_assignment
*const assign_x
=
314 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
315 ir
->operands
[0], NULL
);
316 ir_assignment
*const assign_y
=
317 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
318 ir
->operands
[1], NULL
);
320 this->base_ir
->insert_before(assign_x
);
321 this->base_ir
->insert_before(assign_y
);
323 ir_expression
*const div_expr
=
324 new(ir
) ir_expression(ir_binop_div
, x
->type
,
325 new(ir
) ir_dereference_variable(x
),
326 new(ir
) ir_dereference_variable(y
));
328 /* Don't generate new IR that would need to be lowered in an additional
331 if ((lowering(FDIV_TO_MUL_RCP
) && ir
->type
->is_float()) ||
332 (lowering(DDIV_TO_MUL_RCP
) && ir
->type
->is_double()))
333 div_to_mul_rcp(div_expr
);
335 ir_expression
*const floor_expr
=
336 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
338 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
339 dfloor_to_dfrac(floor_expr
);
341 ir_expression
*const mul_expr
=
342 new(ir
) ir_expression(ir_binop_mul
,
343 new(ir
) ir_dereference_variable(y
),
346 ir
->operation
= ir_binop_sub
;
347 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
348 ir
->operands
[1] = mul_expr
;
349 this->progress
= true;
353 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
356 * ir_binop_ldexp x exp
359 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
360 * resulting_biased_exp = extracted_biased_exp + exp;
362 * if (resulting_biased_exp < 1 || x == 0.0f) {
363 * return copysign(0.0, x);
366 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
367 * lshift(i2u(resulting_biased_exp), exp_shift));
369 * which we can't actually implement as such, since the GLSL IR doesn't
370 * have vectorized if-statements. We actually implement it without branches
371 * using conditional-select:
373 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
374 * resulting_biased_exp = extracted_biased_exp + exp;
376 * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
377 * gequal(resulting_biased_exp, 1);
378 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
379 * resulting_biased_exp = csel(is_not_zero_or_underflow,
380 * resulting_biased_exp, 0);
382 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
383 * lshift(i2u(resulting_biased_exp), exp_shift));
386 const unsigned vec_elem
= ir
->type
->vector_elements
;
389 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
390 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
393 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
395 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
397 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
399 /* Temporary variables */
400 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
401 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
403 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
406 ir_variable
*extracted_biased_exp
=
407 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
408 ir_variable
*resulting_biased_exp
=
409 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
411 ir_variable
*is_not_zero_or_underflow
=
412 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
414 ir_instruction
&i
= *base_ir
;
416 /* Copy <x> and <exp> arguments. */
418 i
.insert_before(assign(x
, ir
->operands
[0]));
419 i
.insert_before(exp
);
420 i
.insert_before(assign(exp
, ir
->operands
[1]));
422 /* Extract the biased exponent from <x>. */
423 i
.insert_before(extracted_biased_exp
);
424 i
.insert_before(assign(extracted_biased_exp
,
425 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
427 i
.insert_before(resulting_biased_exp
);
428 i
.insert_before(assign(resulting_biased_exp
,
429 add(extracted_biased_exp
, exp
)));
431 /* Test if result is ±0.0, subnormal, or underflow by checking if the
432 * resulting biased exponent would be less than 0x1. If so, the result is
433 * 0.0 with the sign of x. (Actually, invert the conditions so that
434 * immediate values are the second arguments, which is better for i965)
436 i
.insert_before(zero_sign_x
);
437 i
.insert_before(assign(zero_sign_x
,
438 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
440 i
.insert_before(is_not_zero_or_underflow
);
441 i
.insert_before(assign(is_not_zero_or_underflow
,
442 logic_and(nequal(x
, new(ir
) ir_constant(0.0f
, vec_elem
)),
443 gequal(resulting_biased_exp
,
444 new(ir
) ir_constant(0x1, vec_elem
)))));
445 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
447 i
.insert_before(assign(resulting_biased_exp
,
448 csel(is_not_zero_or_underflow
,
449 resulting_biased_exp
, zeroi
)));
451 /* We could test for overflows by checking if the resulting biased exponent
452 * would be greater than 0xFE. Turns out we don't need to because the GLSL
455 * "If this product is too large to be represented in the
456 * floating-point type, the result is undefined."
459 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
461 /* Don't generate new IR that would need to be lowered in an additional
464 if (!lowering(INSERT_TO_SHIFTS
)) {
465 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
466 ir
->operation
= ir_unop_bitcast_i2f
;
467 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
468 exp_shift_clone
, exp_width
);
469 ir
->operands
[1] = NULL
;
471 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x807fffffu
, vec_elem
);
472 ir
->operation
= ir_unop_bitcast_u2f
;
473 ir
->operands
[0] = bit_or(bit_and(bitcast_f2u(x
), sign_mantissa_mask
),
474 lshift(i2u(resulting_biased_exp
), exp_shift_clone
));
477 this->progress
= true;
481 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
483 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
484 * from the significand.
487 const unsigned vec_elem
= ir
->type
->vector_elements
;
490 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
491 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
494 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
496 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
498 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
499 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
500 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
502 /* Temporary variables */
503 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
504 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
506 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
509 ir_variable
*extracted_biased_exp
=
510 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
511 ir_variable
*resulting_biased_exp
=
512 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
514 ir_variable
*is_not_zero_or_underflow
=
515 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
517 ir_instruction
&i
= *base_ir
;
519 /* Copy <x> and <exp> arguments. */
521 i
.insert_before(assign(x
, ir
->operands
[0]));
522 i
.insert_before(exp
);
523 i
.insert_before(assign(exp
, ir
->operands
[1]));
525 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
526 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
527 dfrexp_exp_to_arith(frexp_exp
);
529 /* Extract the biased exponent from <x>. */
530 i
.insert_before(extracted_biased_exp
);
531 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
533 i
.insert_before(resulting_biased_exp
);
534 i
.insert_before(assign(resulting_biased_exp
,
535 add(extracted_biased_exp
, exp
)));
537 /* Test if result is ±0.0, subnormal, or underflow by checking if the
538 * resulting biased exponent would be less than 0x1. If so, the result is
539 * 0.0 with the sign of x. (Actually, invert the conditions so that
540 * immediate values are the second arguments, which is better for i965)
541 * TODO: Implement in a vector fashion.
543 i
.insert_before(zero_sign_x
);
544 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
545 ir_variable
*unpacked
=
546 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
547 i
.insert_before(unpacked
);
550 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
551 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
553 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
554 i
.insert_before(assign(zero_sign_x
,
555 expr(ir_unop_pack_double_2x32
, unpacked
),
558 i
.insert_before(is_not_zero_or_underflow
);
559 i
.insert_before(assign(is_not_zero_or_underflow
,
560 gequal(resulting_biased_exp
,
561 new(ir
) ir_constant(0x1, vec_elem
))));
562 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
564 i
.insert_before(assign(resulting_biased_exp
,
565 csel(is_not_zero_or_underflow
,
566 resulting_biased_exp
, zeroi
)));
568 /* We could test for overflows by checking if the resulting biased exponent
569 * would be greater than 0xFE. Turns out we don't need to because the GLSL
572 * "If this product is too large to be represented in the
573 * floating-point type, the result is undefined."
576 ir_rvalue
*results
[4] = {NULL
};
577 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
578 ir_variable
*unpacked
=
579 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
580 i
.insert_before(unpacked
);
583 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
585 ir_expression
*bfi
= bitfield_insert(
587 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
588 exp_shift
->clone(ir
, NULL
),
589 exp_width
->clone(ir
, NULL
));
591 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
593 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
596 ir
->operation
= ir_quadop_vector
;
597 ir
->operands
[0] = results
[0];
598 ir
->operands
[1] = results
[1];
599 ir
->operands
[2] = results
[2];
600 ir
->operands
[3] = results
[3];
602 /* Don't generate new IR that would need to be lowered in an additional
606 this->progress
= true;
610 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
612 const unsigned vec_elem
= ir
->type
->vector_elements
;
613 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
615 /* Double-precision floating-point values are stored as
620 * We're just extracting the significand here, so we only need to modify
621 * the upper 32-bit uint. Unfortunately we must extract each double
622 * independently as there is no vector version of unpackDouble.
625 ir_instruction
&i
= *base_ir
;
627 ir_variable
*is_not_zero
=
628 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
629 ir_rvalue
*results
[4] = {NULL
};
631 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
632 i
.insert_before(is_not_zero
);
635 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
637 /* TODO: Remake this as more vector-friendly when int64 support is
640 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
641 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
642 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
644 /* Exponent of double floating-point values in the range [0.5, 1.0). */
645 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
648 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
649 ir_variable
*unpacked
=
650 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
652 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
654 i
.insert_before(bits
);
655 i
.insert_before(unpacked
);
656 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
658 /* Manipulate the high uint to remove the exponent and replace it with
659 * either the default exponent or zero.
661 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
662 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
663 i
.insert_before(assign(bits
, bit_or(bits
,
664 csel(swizzle(is_not_zero
, elem
, 1),
667 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
668 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
671 /* Put the dvec back together */
672 ir
->operation
= ir_quadop_vector
;
673 ir
->operands
[0] = results
[0];
674 ir
->operands
[1] = results
[1];
675 ir
->operands
[2] = results
[2];
676 ir
->operands
[3] = results
[3];
678 this->progress
= true;
682 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
684 const unsigned vec_elem
= ir
->type
->vector_elements
;
685 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
686 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
688 /* Double-precision floating-point values are stored as
693 * We're just extracting the exponent here, so we only care about the upper
697 ir_instruction
&i
= *base_ir
;
699 ir_variable
*is_not_zero
=
700 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
701 ir_variable
*high_words
=
702 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
703 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
704 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
706 ir_rvalue
*absval
= abs(ir
->operands
[0]);
708 i
.insert_before(is_not_zero
);
709 i
.insert_before(high_words
);
710 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
712 /* Extract all of the upper uints. */
713 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
714 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
716 i
.insert_before(assign(high_words
,
717 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
721 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
722 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
724 /* For non-zero inputs, shift the exponent down and apply bias. */
725 ir
->operation
= ir_triop_csel
;
726 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
727 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
728 ir
->operands
[2] = izero
;
730 this->progress
= true;
734 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
739 * sum = ir_binop_add x y
740 * bcarry = ir_binop_less sum x
741 * carry = ir_unop_b2i bcarry
744 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
745 ir
->operation
= ir_unop_i2u
;
746 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
747 ir
->operands
[1] = NULL
;
749 this->progress
= true;
753 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
756 * ir_binop_borrow x y
758 * bcarry = ir_binop_less x y
759 * carry = ir_unop_b2i bcarry
762 ir
->operation
= ir_unop_i2u
;
763 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
764 ir
->operands
[1] = NULL
;
766 this->progress
= true;
770 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
775 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
778 ir
->operation
= ir_binop_min
;
779 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
781 new(ir
) ir_constant(0.0f
));
782 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
784 this->progress
= true;
788 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
790 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
792 this->base_ir
->insert_before(temp
);
794 int nc
= ir
->operands
[0]->type
->components();
795 for (int i
= nc
- 1; i
>= 1; i
--) {
796 ir_assignment
*assig
;
798 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
799 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
801 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
802 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
805 this->base_ir
->insert_before(assig
);
808 ir
->operation
= ir_triop_fma
;
809 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
810 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
811 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
813 this->progress
= true;
818 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
821 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
822 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
824 switch (op2
->type
->vector_elements
) {
826 swizval
= SWIZZLE_XXXX
;
829 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
830 swizval
= SWIZZLE_XYZW
;
834 ir
->operation
= ir_triop_fma
;
835 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
836 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
838 this->progress
= true;
842 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
846 * temp = sub(x, frtemp);
847 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
849 ir_instruction
&i
= *base_ir
;
850 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
851 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
852 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
855 i
.insert_before(frtemp
);
856 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
858 ir
->operation
= ir_binop_add
;
859 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
860 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
862 this->progress
= true;
866 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
870 * result = sub(x, frtemp);
872 ir
->operation
= ir_binop_sub
;
873 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
875 this->progress
= true;
878 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
883 * frtemp = frac(temp);
884 * t2 = sub(temp, frtemp);
885 * if (frac(x) == 0.5)
886 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
891 ir_instruction
&i
= *base_ir
;
892 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
894 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
896 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
898 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
899 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
900 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
902 i
.insert_before(temp
);
903 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
905 i
.insert_before(frtemp
);
906 i
.insert_before(assign(frtemp
, fract(temp
)));
909 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
911 ir
->operation
= ir_triop_csel
;
912 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
913 p5
->clone(ir
, NULL
));
914 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
918 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
920 this->progress
= true;
924 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
928 * temp = sub(x, frtemp);
929 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
931 ir_rvalue
*arg
= ir
->operands
[0];
932 ir_instruction
&i
= *base_ir
;
934 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
935 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
936 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
938 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
941 i
.insert_before(frtemp
);
942 i
.insert_before(assign(frtemp
, fract(arg
)));
943 i
.insert_before(temp
);
944 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
946 ir
->operation
= ir_triop_csel
;
947 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
948 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
949 ir
->operands
[2] = add(temp
,
950 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
951 zero
->clone(ir
, NULL
),
954 this->progress
= true;
958 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
961 * temp = x > 0.0 ? 1.0 : 0.0;
962 * result = x < 0.0 ? -1.0 : temp;
964 ir_rvalue
*arg
= ir
->operands
[0];
965 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
966 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
967 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
969 ir
->operation
= ir_triop_csel
;
970 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
971 zero
->clone(ir
, NULL
));
972 ir
->operands
[1] = neg_one
;
973 ir
->operands
[2] = csel(greater(arg
, zero
),
975 zero
->clone(ir
, NULL
));
977 this->progress
= true;
981 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
983 /* For more details, see:
985 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
987 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
988 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
990 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
991 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
992 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
993 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
994 ir_constant
*c1
= new(ir
) ir_constant(1u);
995 ir_constant
*c2
= new(ir
) ir_constant(2u);
996 ir_constant
*c4
= new(ir
) ir_constant(4u);
997 ir_constant
*c24
= new(ir
) ir_constant(24u);
999 base_ir
->insert_before(temp
);
1001 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1002 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
1004 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1005 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
1008 /* temp = temp - ((temp >> 1) & 0x55555555u); */
1009 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
1012 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
1013 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
1014 bit_and(rshift(temp
, c2
),
1015 c33333333
->clone(ir
, NULL
)))));
1017 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1018 ir
->operation
= ir_unop_u2i
;
1019 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1023 this->progress
= true;
1027 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1030 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1032 base_ir
->insert_before(bits
);
1033 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1035 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1037 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1039 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1040 ir_constant
*cFFFFFFFF
=
1041 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1043 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1044 * we'd get a mask of 0 when bits is 32. Special case it.
1046 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1048 ir_expression
*mask
= csel(equal(bits
, c32
),
1050 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1052 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1054 * If bits is zero, the result will be zero.
1056 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1057 * select as in the signed integer case.
1059 * (value >> offset) & mask;
1061 ir
->operation
= ir_binop_bit_and
;
1062 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1063 ir
->operands
[1] = mask
;
1064 ir
->operands
[2] = NULL
;
1067 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1069 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1071 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1073 /* temp = 32 - bits; */
1074 base_ir
->insert_before(temp
);
1075 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1077 /* expr = value << (temp - offset)) >> temp; */
1078 ir_expression
*expr
=
1079 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1081 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1083 * If bits is zero, the result will be zero.
1085 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1086 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1089 * (bits == 0) ? 0 : e;
1091 ir
->operation
= ir_triop_csel
;
1092 ir
->operands
[0] = equal(c0
, bits
);
1093 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1094 ir
->operands
[2] = expr
;
1097 this->progress
= true;
1101 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1105 ir_constant
*cFFFFFFFF
;
1106 ir_variable
*offset
=
1107 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1109 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1111 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1113 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1114 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1115 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1116 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1118 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1120 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1121 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1122 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1125 base_ir
->insert_before(offset
);
1126 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1128 base_ir
->insert_before(bits
);
1129 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1131 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1132 * we'd get a mask of 0 when bits is 32. Special case it.
1134 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1136 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1138 * The result will be undefined if offset or bits is negative, or if the
1139 * sum of offset and bits is greater than the number of bits used to
1140 * store the operand.
1142 * Since it's undefined, there are a couple other ways this could be
1143 * implemented. The other way that was considered was to put the csel
1144 * around the whole thing:
1146 * final_result = bits == 32 ? insert : ... ;
1148 base_ir
->insert_before(mask
);
1150 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1152 lshift(sub(lshift(c1
, bits
),
1153 c1
->clone(ir
, NULL
)),
1156 /* (base & ~mask) | ((insert << offset) & mask) */
1157 ir
->operation
= ir_binop_bit_or
;
1158 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1159 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1160 ir
->operands
[2] = NULL
;
1161 ir
->operands
[3] = NULL
;
1163 this->progress
= true;
1167 lower_instructions_visitor::reverse_to_shifts(ir_expression
*ir
)
1169 /* For more details, see:
1171 * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1174 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1176 new(ir
) ir_constant(2u, ir
->operands
[0]->type
->vector_elements
);
1178 new(ir
) ir_constant(4u, ir
->operands
[0]->type
->vector_elements
);
1180 new(ir
) ir_constant(8u, ir
->operands
[0]->type
->vector_elements
);
1182 new(ir
) ir_constant(16u, ir
->operands
[0]->type
->vector_elements
);
1183 ir_constant
*c33333333
=
1184 new(ir
) ir_constant(0x33333333u
, ir
->operands
[0]->type
->vector_elements
);
1185 ir_constant
*c55555555
=
1186 new(ir
) ir_constant(0x55555555u
, ir
->operands
[0]->type
->vector_elements
);
1187 ir_constant
*c0F0F0F0F
=
1188 new(ir
) ir_constant(0x0F0F0F0Fu
, ir
->operands
[0]->type
->vector_elements
);
1189 ir_constant
*c00FF00FF
=
1190 new(ir
) ir_constant(0x00FF00FFu
, ir
->operands
[0]->type
->vector_elements
);
1192 new(ir
) ir_variable(glsl_type::uvec(ir
->operands
[0]->type
->vector_elements
),
1193 "temp", ir_var_temporary
);
1194 ir_instruction
&i
= *base_ir
;
1196 i
.insert_before(temp
);
1198 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1199 i
.insert_before(assign(temp
, ir
->operands
[0]));
1201 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1202 i
.insert_before(assign(temp
, i2u(ir
->operands
[0])));
1205 /* Swap odd and even bits.
1207 * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1209 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c1
), c55555555
),
1210 lshift(bit_and(temp
, c55555555
->clone(ir
, NULL
)),
1211 c1
->clone(ir
, NULL
)))));
1212 /* Swap consecutive pairs.
1214 * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1216 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c2
), c33333333
),
1217 lshift(bit_and(temp
, c33333333
->clone(ir
, NULL
)),
1218 c2
->clone(ir
, NULL
)))));
1222 * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1224 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c4
), c0F0F0F0F
),
1225 lshift(bit_and(temp
, c0F0F0F0F
->clone(ir
, NULL
)),
1226 c4
->clone(ir
, NULL
)))));
1228 /* The last step is, basically, bswap. Swap the bytes, then swap the
1229 * words. When this code is run through GCC on x86, it does generate a
1230 * bswap instruction.
1232 * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1233 * temp = ( temp >> 16 ) | ( temp << 16);
1235 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c8
), c00FF00FF
),
1236 lshift(bit_and(temp
, c00FF00FF
->clone(ir
, NULL
)),
1237 c8
->clone(ir
, NULL
)))));
1239 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1240 ir
->operation
= ir_binop_bit_or
;
1241 ir
->operands
[0] = rshift(temp
, c16
);
1242 ir
->operands
[1] = lshift(temp
, c16
->clone(ir
, NULL
));
1244 ir
->operation
= ir_unop_u2i
;
1245 ir
->operands
[0] = bit_or(rshift(temp
, c16
),
1246 lshift(temp
, c16
->clone(ir
, NULL
)));
1249 this->progress
= true;
1253 lower_instructions_visitor::find_lsb_to_float_cast(ir_expression
*ir
)
1255 /* For more details, see:
1257 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1259 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1260 ir_constant
*c0
= new(ir
) ir_constant(unsigned(0), elements
);
1261 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1262 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1263 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1265 new(ir
) ir_variable(glsl_type::ivec(elements
), "temp", ir_var_temporary
);
1266 ir_variable
*lsb_only
=
1267 new(ir
) ir_variable(glsl_type::uvec(elements
), "lsb_only", ir_var_temporary
);
1268 ir_variable
*as_float
=
1269 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1271 new(ir
) ir_variable(glsl_type::ivec(elements
), "lsb", ir_var_temporary
);
1273 ir_instruction
&i
= *base_ir
;
1275 i
.insert_before(temp
);
1277 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1278 i
.insert_before(assign(temp
, ir
->operands
[0]));
1280 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1281 i
.insert_before(assign(temp
, u2i(ir
->operands
[0])));
1284 /* The int-to-float conversion is lossless because (value & -value) is
1285 * either a power of two or zero. We don't use the result in the zero
1286 * case. The uint() cast is necessary so that 0x80000000 does not
1287 * generate a negative value.
1289 * uint lsb_only = uint(value & -value);
1290 * float as_float = float(lsb_only);
1292 i
.insert_before(lsb_only
);
1293 i
.insert_before(assign(lsb_only
, i2u(bit_and(temp
, neg(temp
)))));
1295 i
.insert_before(as_float
);
1296 i
.insert_before(assign(as_float
, u2f(lsb_only
)));
1298 /* This is basically an open-coded frexp. Implementations that have a
1299 * native frexp instruction would be better served by that. This is
1300 * optimized versus a full-featured open-coded implementation in two ways:
1302 * - We don't care about a correct result from subnormal numbers (including
1303 * 0.0), so the raw exponent can always be safely unbiased.
1305 * - The value cannot be negative, so it does not need to be masked off to
1306 * extract the exponent.
1308 * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1310 i
.insert_before(lsb
);
1311 i
.insert_before(assign(lsb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1313 /* Use lsb_only in the comparison instead of temp so that the & (far above)
1314 * can possibly generate the result without an explicit comparison.
1316 * (lsb_only == 0) ? -1 : lsb;
1318 * Since our input values are all integers, the unbiased exponent must not
1319 * be negative. It will only be negative (-0x7f, in fact) if lsb_only is
1320 * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is
1321 * better is likely GPU dependent. Either way, the difference should be
1324 ir
->operation
= ir_triop_csel
;
1325 ir
->operands
[0] = equal(lsb_only
, c0
);
1326 ir
->operands
[1] = cminus1
;
1327 ir
->operands
[2] = new(ir
) ir_dereference_variable(lsb
);
1329 this->progress
= true;
1333 lower_instructions_visitor::find_msb_to_float_cast(ir_expression
*ir
)
1335 /* For more details, see:
1337 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1339 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1340 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1341 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1342 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1343 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1344 ir_constant
*c000000FF
= new(ir
) ir_constant(0x000000FFu
, elements
);
1345 ir_constant
*cFFFFFF00
= new(ir
) ir_constant(0xFFFFFF00u
, elements
);
1347 new(ir
) ir_variable(glsl_type::uvec(elements
), "temp", ir_var_temporary
);
1348 ir_variable
*as_float
=
1349 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1351 new(ir
) ir_variable(glsl_type::ivec(elements
), "msb", ir_var_temporary
);
1353 ir_instruction
&i
= *base_ir
;
1355 i
.insert_before(temp
);
1357 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1358 i
.insert_before(assign(temp
, ir
->operands
[0]));
1360 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1362 /* findMSB(uint(abs(some_int))) almost always does the right thing.
1363 * There are two problem values:
1365 * * 0x80000000. Since abs(0x80000000) == 0x80000000, findMSB returns
1366 * 31. However, findMSB(int(0x80000000)) == 30.
1368 * * 0xffffffff. Since abs(0xffffffff) == 1, findMSB returns
1369 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1371 * For a value of zero or negative one, -1 will be returned.
1373 * For all negative number cases, including 0x80000000 and 0xffffffff,
1374 * the correct value is obtained from findMSB if instead of negating the
1375 * (already negative) value the logical-not is used. A conditonal
1376 * logical-not can be achieved in two instructions.
1378 ir_variable
*as_int
=
1379 new(ir
) ir_variable(glsl_type::ivec(elements
), "as_int", ir_var_temporary
);
1380 ir_constant
*c31
= new(ir
) ir_constant(int(31), elements
);
1382 i
.insert_before(as_int
);
1383 i
.insert_before(assign(as_int
, ir
->operands
[0]));
1384 i
.insert_before(assign(temp
, i2u(expr(ir_binop_bit_xor
,
1386 rshift(as_int
, c31
)))));
1389 /* The int-to-float conversion is lossless because bits are conditionally
1390 * masked off the bottom of temp to ensure the value has at most 24 bits of
1391 * data or is zero. We don't use the result in the zero case. The uint()
1392 * cast is necessary so that 0x80000000 does not generate a negative value.
1394 * float as_float = float(temp > 255 ? temp & ~255 : temp);
1396 i
.insert_before(as_float
);
1397 i
.insert_before(assign(as_float
, u2f(csel(greater(temp
, c000000FF
),
1398 bit_and(temp
, cFFFFFF00
),
1401 /* This is basically an open-coded frexp. Implementations that have a
1402 * native frexp instruction would be better served by that. This is
1403 * optimized versus a full-featured open-coded implementation in two ways:
1405 * - We don't care about a correct result from subnormal numbers (including
1406 * 0.0), so the raw exponent can always be safely unbiased.
1408 * - The value cannot be negative, so it does not need to be masked off to
1409 * extract the exponent.
1411 * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1413 i
.insert_before(msb
);
1414 i
.insert_before(assign(msb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1416 /* Use msb in the comparison instead of temp so that the subtract can
1417 * possibly generate the result without an explicit comparison.
1419 * (msb < 0) ? -1 : msb;
1421 * Since our input values are all integers, the unbiased exponent must not
1422 * be negative. It will only be negative (-0x7f, in fact) if temp is 0.
1424 ir
->operation
= ir_triop_csel
;
1425 ir
->operands
[0] = less(msb
, c0
);
1426 ir
->operands
[1] = cminus1
;
1427 ir
->operands
[2] = new(ir
) ir_dereference_variable(msb
);
1429 this->progress
= true;
1433 lower_instructions_visitor::_carry(operand a
, operand b
)
1435 if (lowering(CARRY_TO_ARITH
))
1436 return i2u(b2i(less(add(a
, b
),
1437 a
.val
->clone(ralloc_parent(a
.val
), NULL
))));
1443 lower_instructions_visitor::imul_high_to_mul(ir_expression
*ir
)
1448 * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
1450 * In GLSL, (a * b) becomes
1452 * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu);
1453 * uint m2 = (a & 0x0000ffffu) * (b >> 16);
1454 * uint m3 = (a >> 16) * (b & 0x0000ffffu);
1455 * uint m4 = (a >> 16) * (b >> 16);
1462 * lo_result = uaddCarry(m1, m2 << 16, c1);
1463 * hi_result = m4 + c1;
1464 * lo_result = uaddCarry(lo_result, m3 << 16, c2);
1465 * hi_result = hi_result + c2;
1466 * hi_result = hi_result + (m2 >> 16) + (m3 >> 16);
1468 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1470 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1", ir_var_temporary
);
1471 ir_variable
*src1h
=
1472 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1h", ir_var_temporary
);
1473 ir_variable
*src1l
=
1474 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1l", ir_var_temporary
);
1476 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2", ir_var_temporary
);
1477 ir_variable
*src2h
=
1478 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2h", ir_var_temporary
);
1479 ir_variable
*src2l
=
1480 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2l", ir_var_temporary
);
1482 new(ir
) ir_variable(glsl_type::uvec(elements
), "t1", ir_var_temporary
);
1484 new(ir
) ir_variable(glsl_type::uvec(elements
), "t2", ir_var_temporary
);
1486 new(ir
) ir_variable(glsl_type::uvec(elements
), "lo", ir_var_temporary
);
1488 new(ir
) ir_variable(glsl_type::uvec(elements
), "hi", ir_var_temporary
);
1489 ir_variable
*different_signs
= NULL
;
1490 ir_constant
*c0000FFFF
= new(ir
) ir_constant(0x0000FFFFu
, elements
);
1491 ir_constant
*c16
= new(ir
) ir_constant(16u, elements
);
1493 ir_instruction
&i
= *base_ir
;
1495 i
.insert_before(src1
);
1496 i
.insert_before(src2
);
1497 i
.insert_before(src1h
);
1498 i
.insert_before(src2h
);
1499 i
.insert_before(src1l
);
1500 i
.insert_before(src2l
);
1502 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1503 i
.insert_before(assign(src1
, ir
->operands
[0]));
1504 i
.insert_before(assign(src2
, ir
->operands
[1]));
1506 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1508 ir_variable
*itmp1
=
1509 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp1", ir_var_temporary
);
1510 ir_variable
*itmp2
=
1511 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp2", ir_var_temporary
);
1512 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1514 i
.insert_before(itmp1
);
1515 i
.insert_before(itmp2
);
1516 i
.insert_before(assign(itmp1
, ir
->operands
[0]));
1517 i
.insert_before(assign(itmp2
, ir
->operands
[1]));
1520 new(ir
) ir_variable(glsl_type::bvec(elements
), "different_signs",
1523 i
.insert_before(different_signs
);
1524 i
.insert_before(assign(different_signs
, expr(ir_binop_logic_xor
,
1526 less(itmp2
, c0
->clone(ir
, NULL
)))));
1528 i
.insert_before(assign(src1
, i2u(abs(itmp1
))));
1529 i
.insert_before(assign(src2
, i2u(abs(itmp2
))));
1532 i
.insert_before(assign(src1l
, bit_and(src1
, c0000FFFF
)));
1533 i
.insert_before(assign(src2l
, bit_and(src2
, c0000FFFF
->clone(ir
, NULL
))));
1534 i
.insert_before(assign(src1h
, rshift(src1
, c16
)));
1535 i
.insert_before(assign(src2h
, rshift(src2
, c16
->clone(ir
, NULL
))));
1537 i
.insert_before(lo
);
1538 i
.insert_before(hi
);
1539 i
.insert_before(t1
);
1540 i
.insert_before(t2
);
1542 i
.insert_before(assign(lo
, mul(src1l
, src2l
)));
1543 i
.insert_before(assign(t1
, mul(src1l
, src2h
)));
1544 i
.insert_before(assign(t2
, mul(src1h
, src2l
)));
1545 i
.insert_before(assign(hi
, mul(src1h
, src2h
)));
1547 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t1
, c16
->clone(ir
, NULL
))))));
1548 i
.insert_before(assign(lo
, add(lo
, lshift(t1
, c16
->clone(ir
, NULL
)))));
1550 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t2
, c16
->clone(ir
, NULL
))))));
1551 i
.insert_before(assign(lo
, add(lo
, lshift(t2
, c16
->clone(ir
, NULL
)))));
1553 if (different_signs
== NULL
) {
1554 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1556 ir
->operation
= ir_binop_add
;
1557 ir
->operands
[0] = add(hi
, rshift(t1
, c16
->clone(ir
, NULL
)));
1558 ir
->operands
[1] = rshift(t2
, c16
->clone(ir
, NULL
));
1560 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1562 i
.insert_before(assign(hi
, add(add(hi
, rshift(t1
, c16
->clone(ir
, NULL
))),
1563 rshift(t2
, c16
->clone(ir
, NULL
)))));
1565 /* For channels where different_signs is set we have to perform a 64-bit
1566 * negation. This is *not* the same as just negating the high 32-bits.
1567 * Consider -3 * 2. The high 32-bits is 0, but the desired result is
1568 * -1, not -0! Recall -x == ~x + 1.
1570 ir_variable
*neg_hi
=
1571 new(ir
) ir_variable(glsl_type::ivec(elements
), "neg_hi", ir_var_temporary
);
1572 ir_constant
*c1
= new(ir
) ir_constant(1u, elements
);
1574 i
.insert_before(neg_hi
);
1575 i
.insert_before(assign(neg_hi
, add(bit_not(u2i(hi
)),
1576 u2i(_carry(bit_not(lo
), c1
)))));
1578 ir
->operation
= ir_triop_csel
;
1579 ir
->operands
[0] = new(ir
) ir_dereference_variable(different_signs
);
1580 ir
->operands
[1] = new(ir
) ir_dereference_variable(neg_hi
);
1581 ir
->operands
[2] = u2i(hi
);
1586 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1588 switch (ir
->operation
) {
1590 if (ir
->operands
[0]->type
->is_double())
1591 double_dot_to_fma(ir
);
1594 if (ir
->operands
[0]->type
->is_double())
1598 if (lowering(SUB_TO_ADD_NEG
))
1603 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1604 int_div_to_mul_rcp(ir
);
1605 else if ((ir
->operands
[1]->type
->is_float() && lowering(FDIV_TO_MUL_RCP
)) ||
1606 (ir
->operands
[1]->type
->is_double() && lowering(DDIV_TO_MUL_RCP
)))
1611 if (lowering(EXP_TO_EXP2
))
1616 if (lowering(LOG_TO_LOG2
))
1621 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1626 if (lowering(POW_TO_EXP2
))
1630 case ir_binop_ldexp
:
1631 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1633 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1634 dldexp_to_arith(ir
);
1637 case ir_unop_frexp_exp
:
1638 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1639 dfrexp_exp_to_arith(ir
);
1642 case ir_unop_frexp_sig
:
1643 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1644 dfrexp_sig_to_arith(ir
);
1647 case ir_binop_carry
:
1648 if (lowering(CARRY_TO_ARITH
))
1652 case ir_binop_borrow
:
1653 if (lowering(BORROW_TO_ARITH
))
1654 borrow_to_arith(ir
);
1657 case ir_unop_saturate
:
1658 if (lowering(SAT_TO_CLAMP
))
1663 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1664 dtrunc_to_dfrac(ir
);
1668 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1673 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1674 dfloor_to_dfrac(ir
);
1677 case ir_unop_round_even
:
1678 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1679 dround_even_to_dfrac(ir
);
1683 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1687 case ir_unop_bit_count
:
1688 if (lowering(BIT_COUNT_TO_MATH
))
1689 bit_count_to_math(ir
);
1692 case ir_triop_bitfield_extract
:
1693 if (lowering(EXTRACT_TO_SHIFTS
))
1694 extract_to_shifts(ir
);
1697 case ir_quadop_bitfield_insert
:
1698 if (lowering(INSERT_TO_SHIFTS
))
1699 insert_to_shifts(ir
);
1702 case ir_unop_bitfield_reverse
:
1703 if (lowering(REVERSE_TO_SHIFTS
))
1704 reverse_to_shifts(ir
);
1707 case ir_unop_find_lsb
:
1708 if (lowering(FIND_LSB_TO_FLOAT_CAST
))
1709 find_lsb_to_float_cast(ir
);
1712 case ir_unop_find_msb
:
1713 if (lowering(FIND_MSB_TO_FLOAT_CAST
))
1714 find_msb_to_float_cast(ir
);
1717 case ir_binop_imul_high
:
1718 if (lowering(IMUL_HIGH_TO_MUL
))
1719 imul_high_to_mul(ir
);
1723 return visit_continue
;
1726 return visit_continue
;