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
);
167 void find_msb_to_float_cast(ir_expression
*ir
);
168 void imul_high_to_mul(ir_expression
*ir
);
170 ir_expression
*_carry(operand a
, operand b
);
173 } /* anonymous namespace */
176 * Determine if a particular type of lowering should occur
178 #define lowering(x) (this->lower & x)
181 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
183 lower_instructions_visitor
v(what_to_lower
);
185 visit_list_elements(&v
, instructions
);
190 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
192 ir
->operation
= ir_binop_add
;
193 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
194 ir
->operands
[1], NULL
);
195 this->progress
= true;
199 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
201 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
203 /* New expression for the 1.0 / op1 */
205 expr
= new(ir
) ir_expression(ir_unop_rcp
,
206 ir
->operands
[1]->type
,
209 /* op0 / op1 -> op0 * (1.0 / op1) */
210 ir
->operation
= ir_binop_mul
;
211 ir
->operands
[1] = expr
;
213 this->progress
= true;
217 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
219 assert(ir
->operands
[1]->type
->is_integer());
221 /* Be careful with integer division -- we need to do it as a
222 * float and re-truncate, since rcp(n > 1) of an integer would
225 ir_rvalue
*op0
, *op1
;
226 const struct glsl_type
*vec_type
;
228 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
229 ir
->operands
[1]->type
->vector_elements
,
230 ir
->operands
[1]->type
->matrix_columns
);
232 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
233 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
235 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
237 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
239 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
240 ir
->operands
[0]->type
->vector_elements
,
241 ir
->operands
[0]->type
->matrix_columns
);
243 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
244 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
246 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
248 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
249 ir
->type
->vector_elements
,
250 ir
->type
->matrix_columns
);
252 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
254 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
255 ir
->operation
= ir_unop_f2i
;
256 ir
->operands
[0] = op0
;
258 ir
->operation
= ir_unop_i2u
;
259 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
261 ir
->operands
[1] = NULL
;
263 this->progress
= true;
267 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
269 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
271 ir
->operation
= ir_unop_exp2
;
272 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
273 ir
->operands
[0], log2_e
);
274 this->progress
= true;
278 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
280 ir_expression
*const log2_x
=
281 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
284 ir
->operation
= ir_unop_exp2
;
285 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
286 ir
->operands
[1], log2_x
);
287 ir
->operands
[1] = NULL
;
288 this->progress
= true;
292 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
294 ir
->operation
= ir_binop_mul
;
295 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
296 ir
->operands
[0], NULL
);
297 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
298 this->progress
= true;
302 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
304 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
306 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
308 this->base_ir
->insert_before(x
);
309 this->base_ir
->insert_before(y
);
311 ir_assignment
*const assign_x
=
312 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
313 ir
->operands
[0], NULL
);
314 ir_assignment
*const assign_y
=
315 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
316 ir
->operands
[1], NULL
);
318 this->base_ir
->insert_before(assign_x
);
319 this->base_ir
->insert_before(assign_y
);
321 ir_expression
*const div_expr
=
322 new(ir
) ir_expression(ir_binop_div
, x
->type
,
323 new(ir
) ir_dereference_variable(x
),
324 new(ir
) ir_dereference_variable(y
));
326 /* Don't generate new IR that would need to be lowered in an additional
329 if (lowering(DIV_TO_MUL_RCP
) && (ir
->type
->is_float() || ir
->type
->is_double()))
330 div_to_mul_rcp(div_expr
);
332 ir_expression
*const floor_expr
=
333 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
335 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
336 dfloor_to_dfrac(floor_expr
);
338 ir_expression
*const mul_expr
=
339 new(ir
) ir_expression(ir_binop_mul
,
340 new(ir
) ir_dereference_variable(y
),
343 ir
->operation
= ir_binop_sub
;
344 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
345 ir
->operands
[1] = mul_expr
;
346 this->progress
= true;
350 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
353 * ir_binop_ldexp x exp
356 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
357 * resulting_biased_exp = extracted_biased_exp + exp;
359 * if (resulting_biased_exp < 1 || x == 0.0f) {
360 * return copysign(0.0, x);
363 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
364 * lshift(i2u(resulting_biased_exp), exp_shift));
366 * which we can't actually implement as such, since the GLSL IR doesn't
367 * have vectorized if-statements. We actually implement it without branches
368 * using conditional-select:
370 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
371 * resulting_biased_exp = extracted_biased_exp + exp;
373 * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
374 * gequal(resulting_biased_exp, 1);
375 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
376 * resulting_biased_exp = csel(is_not_zero_or_underflow,
377 * resulting_biased_exp, 0);
379 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
380 * lshift(i2u(resulting_biased_exp), exp_shift));
383 const unsigned vec_elem
= ir
->type
->vector_elements
;
386 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
387 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
390 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
392 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
394 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
396 /* Temporary variables */
397 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
398 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
400 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
403 ir_variable
*extracted_biased_exp
=
404 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
405 ir_variable
*resulting_biased_exp
=
406 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
408 ir_variable
*is_not_zero_or_underflow
=
409 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
411 ir_instruction
&i
= *base_ir
;
413 /* Copy <x> and <exp> arguments. */
415 i
.insert_before(assign(x
, ir
->operands
[0]));
416 i
.insert_before(exp
);
417 i
.insert_before(assign(exp
, ir
->operands
[1]));
419 /* Extract the biased exponent from <x>. */
420 i
.insert_before(extracted_biased_exp
);
421 i
.insert_before(assign(extracted_biased_exp
,
422 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
424 i
.insert_before(resulting_biased_exp
);
425 i
.insert_before(assign(resulting_biased_exp
,
426 add(extracted_biased_exp
, exp
)));
428 /* Test if result is ±0.0, subnormal, or underflow by checking if the
429 * resulting biased exponent would be less than 0x1. If so, the result is
430 * 0.0 with the sign of x. (Actually, invert the conditions so that
431 * immediate values are the second arguments, which is better for i965)
433 i
.insert_before(zero_sign_x
);
434 i
.insert_before(assign(zero_sign_x
,
435 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
437 i
.insert_before(is_not_zero_or_underflow
);
438 i
.insert_before(assign(is_not_zero_or_underflow
,
439 logic_and(nequal(x
, new(ir
) ir_constant(0.0f
, vec_elem
)),
440 gequal(resulting_biased_exp
,
441 new(ir
) ir_constant(0x1, vec_elem
)))));
442 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
444 i
.insert_before(assign(resulting_biased_exp
,
445 csel(is_not_zero_or_underflow
,
446 resulting_biased_exp
, zeroi
)));
448 /* We could test for overflows by checking if the resulting biased exponent
449 * would be greater than 0xFE. Turns out we don't need to because the GLSL
452 * "If this product is too large to be represented in the
453 * floating-point type, the result is undefined."
456 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
458 /* Don't generate new IR that would need to be lowered in an additional
461 if (!lowering(INSERT_TO_SHIFTS
)) {
462 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
463 ir
->operation
= ir_unop_bitcast_i2f
;
464 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
465 exp_shift_clone
, exp_width
);
466 ir
->operands
[1] = NULL
;
468 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x807fffffu
, vec_elem
);
469 ir
->operation
= ir_unop_bitcast_u2f
;
470 ir
->operands
[0] = bit_or(bit_and(bitcast_f2u(x
), sign_mantissa_mask
),
471 lshift(i2u(resulting_biased_exp
), exp_shift_clone
));
474 this->progress
= true;
478 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
480 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
481 * from the significand.
484 const unsigned vec_elem
= ir
->type
->vector_elements
;
487 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
488 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
491 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
493 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
495 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
496 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
497 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
499 /* Temporary variables */
500 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
501 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
503 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
506 ir_variable
*extracted_biased_exp
=
507 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
508 ir_variable
*resulting_biased_exp
=
509 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
511 ir_variable
*is_not_zero_or_underflow
=
512 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
514 ir_instruction
&i
= *base_ir
;
516 /* Copy <x> and <exp> arguments. */
518 i
.insert_before(assign(x
, ir
->operands
[0]));
519 i
.insert_before(exp
);
520 i
.insert_before(assign(exp
, ir
->operands
[1]));
522 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
523 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
524 dfrexp_exp_to_arith(frexp_exp
);
526 /* Extract the biased exponent from <x>. */
527 i
.insert_before(extracted_biased_exp
);
528 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
530 i
.insert_before(resulting_biased_exp
);
531 i
.insert_before(assign(resulting_biased_exp
,
532 add(extracted_biased_exp
, exp
)));
534 /* Test if result is ±0.0, subnormal, or underflow by checking if the
535 * resulting biased exponent would be less than 0x1. If so, the result is
536 * 0.0 with the sign of x. (Actually, invert the conditions so that
537 * immediate values are the second arguments, which is better for i965)
538 * TODO: Implement in a vector fashion.
540 i
.insert_before(zero_sign_x
);
541 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
542 ir_variable
*unpacked
=
543 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
544 i
.insert_before(unpacked
);
547 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
548 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
550 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
551 i
.insert_before(assign(zero_sign_x
,
552 expr(ir_unop_pack_double_2x32
, unpacked
),
555 i
.insert_before(is_not_zero_or_underflow
);
556 i
.insert_before(assign(is_not_zero_or_underflow
,
557 gequal(resulting_biased_exp
,
558 new(ir
) ir_constant(0x1, vec_elem
))));
559 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
561 i
.insert_before(assign(resulting_biased_exp
,
562 csel(is_not_zero_or_underflow
,
563 resulting_biased_exp
, zeroi
)));
565 /* We could test for overflows by checking if the resulting biased exponent
566 * would be greater than 0xFE. Turns out we don't need to because the GLSL
569 * "If this product is too large to be represented in the
570 * floating-point type, the result is undefined."
573 ir_rvalue
*results
[4] = {NULL
};
574 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
575 ir_variable
*unpacked
=
576 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
577 i
.insert_before(unpacked
);
580 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
582 ir_expression
*bfi
= bitfield_insert(
584 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
585 exp_shift
->clone(ir
, NULL
),
586 exp_width
->clone(ir
, NULL
));
588 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
590 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
593 ir
->operation
= ir_quadop_vector
;
594 ir
->operands
[0] = results
[0];
595 ir
->operands
[1] = results
[1];
596 ir
->operands
[2] = results
[2];
597 ir
->operands
[3] = results
[3];
599 /* Don't generate new IR that would need to be lowered in an additional
603 this->progress
= true;
607 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
609 const unsigned vec_elem
= ir
->type
->vector_elements
;
610 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
612 /* Double-precision floating-point values are stored as
617 * We're just extracting the significand here, so we only need to modify
618 * the upper 32-bit uint. Unfortunately we must extract each double
619 * independently as there is no vector version of unpackDouble.
622 ir_instruction
&i
= *base_ir
;
624 ir_variable
*is_not_zero
=
625 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
626 ir_rvalue
*results
[4] = {NULL
};
628 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
629 i
.insert_before(is_not_zero
);
632 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
634 /* TODO: Remake this as more vector-friendly when int64 support is
637 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
638 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
639 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
641 /* Exponent of double floating-point values in the range [0.5, 1.0). */
642 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
645 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
646 ir_variable
*unpacked
=
647 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
649 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
651 i
.insert_before(bits
);
652 i
.insert_before(unpacked
);
653 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
655 /* Manipulate the high uint to remove the exponent and replace it with
656 * either the default exponent or zero.
658 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
659 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
660 i
.insert_before(assign(bits
, bit_or(bits
,
661 csel(swizzle(is_not_zero
, elem
, 1),
664 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
665 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
668 /* Put the dvec back together */
669 ir
->operation
= ir_quadop_vector
;
670 ir
->operands
[0] = results
[0];
671 ir
->operands
[1] = results
[1];
672 ir
->operands
[2] = results
[2];
673 ir
->operands
[3] = results
[3];
675 this->progress
= true;
679 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
681 const unsigned vec_elem
= ir
->type
->vector_elements
;
682 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
683 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
685 /* Double-precision floating-point values are stored as
690 * We're just extracting the exponent here, so we only care about the upper
694 ir_instruction
&i
= *base_ir
;
696 ir_variable
*is_not_zero
=
697 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
698 ir_variable
*high_words
=
699 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
700 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
701 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
703 ir_rvalue
*absval
= abs(ir
->operands
[0]);
705 i
.insert_before(is_not_zero
);
706 i
.insert_before(high_words
);
707 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
709 /* Extract all of the upper uints. */
710 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
711 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
713 i
.insert_before(assign(high_words
,
714 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
718 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
719 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
721 /* For non-zero inputs, shift the exponent down and apply bias. */
722 ir
->operation
= ir_triop_csel
;
723 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
724 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
725 ir
->operands
[2] = izero
;
727 this->progress
= true;
731 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
736 * sum = ir_binop_add x y
737 * bcarry = ir_binop_less sum x
738 * carry = ir_unop_b2i bcarry
741 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
742 ir
->operation
= ir_unop_i2u
;
743 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
744 ir
->operands
[1] = NULL
;
746 this->progress
= true;
750 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
753 * ir_binop_borrow x y
755 * bcarry = ir_binop_less x y
756 * carry = ir_unop_b2i bcarry
759 ir
->operation
= ir_unop_i2u
;
760 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
761 ir
->operands
[1] = NULL
;
763 this->progress
= true;
767 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
772 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
775 ir
->operation
= ir_binop_min
;
776 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
778 new(ir
) ir_constant(0.0f
));
779 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
781 this->progress
= true;
785 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
787 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
789 this->base_ir
->insert_before(temp
);
791 int nc
= ir
->operands
[0]->type
->components();
792 for (int i
= nc
- 1; i
>= 1; i
--) {
793 ir_assignment
*assig
;
795 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
796 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
798 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
799 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
802 this->base_ir
->insert_before(assig
);
805 ir
->operation
= ir_triop_fma
;
806 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
807 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
808 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
810 this->progress
= true;
815 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
818 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
819 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
821 switch (op2
->type
->vector_elements
) {
823 swizval
= SWIZZLE_XXXX
;
826 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
827 swizval
= SWIZZLE_XYZW
;
831 ir
->operation
= ir_triop_fma
;
832 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
833 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
835 this->progress
= true;
839 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
843 * temp = sub(x, frtemp);
844 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
846 ir_instruction
&i
= *base_ir
;
847 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
848 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
849 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
852 i
.insert_before(frtemp
);
853 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
855 ir
->operation
= ir_binop_add
;
856 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
857 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
859 this->progress
= true;
863 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
867 * result = sub(x, frtemp);
869 ir
->operation
= ir_binop_sub
;
870 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
872 this->progress
= true;
875 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
880 * frtemp = frac(temp);
881 * t2 = sub(temp, frtemp);
882 * if (frac(x) == 0.5)
883 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
888 ir_instruction
&i
= *base_ir
;
889 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
891 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
893 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
895 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
896 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
897 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
899 i
.insert_before(temp
);
900 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
902 i
.insert_before(frtemp
);
903 i
.insert_before(assign(frtemp
, fract(temp
)));
906 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
908 ir
->operation
= ir_triop_csel
;
909 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
910 p5
->clone(ir
, NULL
));
911 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
915 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
917 this->progress
= true;
921 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
925 * temp = sub(x, frtemp);
926 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
928 ir_rvalue
*arg
= ir
->operands
[0];
929 ir_instruction
&i
= *base_ir
;
931 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
932 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
933 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
935 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
938 i
.insert_before(frtemp
);
939 i
.insert_before(assign(frtemp
, fract(arg
)));
940 i
.insert_before(temp
);
941 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
943 ir
->operation
= ir_triop_csel
;
944 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
945 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
946 ir
->operands
[2] = add(temp
,
947 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
948 zero
->clone(ir
, NULL
),
951 this->progress
= true;
955 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
958 * temp = x > 0.0 ? 1.0 : 0.0;
959 * result = x < 0.0 ? -1.0 : temp;
961 ir_rvalue
*arg
= ir
->operands
[0];
962 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
963 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
964 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
966 ir
->operation
= ir_triop_csel
;
967 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
968 zero
->clone(ir
, NULL
));
969 ir
->operands
[1] = neg_one
;
970 ir
->operands
[2] = csel(greater(arg
, zero
),
972 zero
->clone(ir
, NULL
));
974 this->progress
= true;
978 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
980 /* For more details, see:
982 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
984 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
985 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
987 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
988 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
989 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
990 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
991 ir_constant
*c1
= new(ir
) ir_constant(1u);
992 ir_constant
*c2
= new(ir
) ir_constant(2u);
993 ir_constant
*c4
= new(ir
) ir_constant(4u);
994 ir_constant
*c24
= new(ir
) ir_constant(24u);
996 base_ir
->insert_before(temp
);
998 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
999 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
1001 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1002 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
1005 /* temp = temp - ((temp >> 1) & 0x55555555u); */
1006 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
1009 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
1010 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
1011 bit_and(rshift(temp
, c2
),
1012 c33333333
->clone(ir
, NULL
)))));
1014 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1015 ir
->operation
= ir_unop_u2i
;
1016 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1020 this->progress
= true;
1024 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1027 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1029 base_ir
->insert_before(bits
);
1030 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1032 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1034 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1036 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1037 ir_constant
*cFFFFFFFF
=
1038 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1040 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1041 * we'd get a mask of 0 when bits is 32. Special case it.
1043 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1045 ir_expression
*mask
= csel(equal(bits
, c32
),
1047 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1049 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1051 * If bits is zero, the result will be zero.
1053 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1054 * select as in the signed integer case.
1056 * (value >> offset) & mask;
1058 ir
->operation
= ir_binop_bit_and
;
1059 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1060 ir
->operands
[1] = mask
;
1061 ir
->operands
[2] = NULL
;
1064 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1066 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1068 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1070 /* temp = 32 - bits; */
1071 base_ir
->insert_before(temp
);
1072 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1074 /* expr = value << (temp - offset)) >> temp; */
1075 ir_expression
*expr
=
1076 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1078 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1080 * If bits is zero, the result will be zero.
1082 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1083 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1086 * (bits == 0) ? 0 : e;
1088 ir
->operation
= ir_triop_csel
;
1089 ir
->operands
[0] = equal(c0
, bits
);
1090 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1091 ir
->operands
[2] = expr
;
1094 this->progress
= true;
1098 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1102 ir_constant
*cFFFFFFFF
;
1103 ir_variable
*offset
=
1104 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1106 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1108 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1110 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1111 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1112 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1113 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1115 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1117 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1118 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1119 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1122 base_ir
->insert_before(offset
);
1123 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1125 base_ir
->insert_before(bits
);
1126 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1128 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1129 * we'd get a mask of 0 when bits is 32. Special case it.
1131 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1133 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1135 * The result will be undefined if offset or bits is negative, or if the
1136 * sum of offset and bits is greater than the number of bits used to
1137 * store the operand.
1139 * Since it's undefined, there are a couple other ways this could be
1140 * implemented. The other way that was considered was to put the csel
1141 * around the whole thing:
1143 * final_result = bits == 32 ? insert : ... ;
1145 base_ir
->insert_before(mask
);
1147 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1149 lshift(sub(lshift(c1
, bits
),
1150 c1
->clone(ir
, NULL
)),
1153 /* (base & ~mask) | ((insert << offset) & mask) */
1154 ir
->operation
= ir_binop_bit_or
;
1155 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1156 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1157 ir
->operands
[2] = NULL
;
1158 ir
->operands
[3] = NULL
;
1160 this->progress
= true;
1164 lower_instructions_visitor::reverse_to_shifts(ir_expression
*ir
)
1166 /* For more details, see:
1168 * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1171 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1173 new(ir
) ir_constant(2u, ir
->operands
[0]->type
->vector_elements
);
1175 new(ir
) ir_constant(4u, ir
->operands
[0]->type
->vector_elements
);
1177 new(ir
) ir_constant(8u, ir
->operands
[0]->type
->vector_elements
);
1179 new(ir
) ir_constant(16u, ir
->operands
[0]->type
->vector_elements
);
1180 ir_constant
*c33333333
=
1181 new(ir
) ir_constant(0x33333333u
, ir
->operands
[0]->type
->vector_elements
);
1182 ir_constant
*c55555555
=
1183 new(ir
) ir_constant(0x55555555u
, ir
->operands
[0]->type
->vector_elements
);
1184 ir_constant
*c0F0F0F0F
=
1185 new(ir
) ir_constant(0x0F0F0F0Fu
, ir
->operands
[0]->type
->vector_elements
);
1186 ir_constant
*c00FF00FF
=
1187 new(ir
) ir_constant(0x00FF00FFu
, ir
->operands
[0]->type
->vector_elements
);
1189 new(ir
) ir_variable(glsl_type::uvec(ir
->operands
[0]->type
->vector_elements
),
1190 "temp", ir_var_temporary
);
1191 ir_instruction
&i
= *base_ir
;
1193 i
.insert_before(temp
);
1195 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1196 i
.insert_before(assign(temp
, ir
->operands
[0]));
1198 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1199 i
.insert_before(assign(temp
, i2u(ir
->operands
[0])));
1202 /* Swap odd and even bits.
1204 * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1206 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c1
), c55555555
),
1207 lshift(bit_and(temp
, c55555555
->clone(ir
, NULL
)),
1208 c1
->clone(ir
, NULL
)))));
1209 /* Swap consecutive pairs.
1211 * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1213 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c2
), c33333333
),
1214 lshift(bit_and(temp
, c33333333
->clone(ir
, NULL
)),
1215 c2
->clone(ir
, NULL
)))));
1219 * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1221 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c4
), c0F0F0F0F
),
1222 lshift(bit_and(temp
, c0F0F0F0F
->clone(ir
, NULL
)),
1223 c4
->clone(ir
, NULL
)))));
1225 /* The last step is, basically, bswap. Swap the bytes, then swap the
1226 * words. When this code is run through GCC on x86, it does generate a
1227 * bswap instruction.
1229 * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1230 * temp = ( temp >> 16 ) | ( temp << 16);
1232 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c8
), c00FF00FF
),
1233 lshift(bit_and(temp
, c00FF00FF
->clone(ir
, NULL
)),
1234 c8
->clone(ir
, NULL
)))));
1236 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1237 ir
->operation
= ir_binop_bit_or
;
1238 ir
->operands
[0] = rshift(temp
, c16
);
1239 ir
->operands
[1] = lshift(temp
, c16
->clone(ir
, NULL
));
1241 ir
->operation
= ir_unop_u2i
;
1242 ir
->operands
[0] = bit_or(rshift(temp
, c16
),
1243 lshift(temp
, c16
->clone(ir
, NULL
)));
1246 this->progress
= true;
1250 lower_instructions_visitor::find_lsb_to_float_cast(ir_expression
*ir
)
1252 /* For more details, see:
1254 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1256 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1257 ir_constant
*c0
= new(ir
) ir_constant(unsigned(0), elements
);
1258 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1259 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1260 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1262 new(ir
) ir_variable(glsl_type::ivec(elements
), "temp", ir_var_temporary
);
1263 ir_variable
*lsb_only
=
1264 new(ir
) ir_variable(glsl_type::uvec(elements
), "lsb_only", ir_var_temporary
);
1265 ir_variable
*as_float
=
1266 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1268 new(ir
) ir_variable(glsl_type::ivec(elements
), "lsb", ir_var_temporary
);
1270 ir_instruction
&i
= *base_ir
;
1272 i
.insert_before(temp
);
1274 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1275 i
.insert_before(assign(temp
, ir
->operands
[0]));
1277 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1278 i
.insert_before(assign(temp
, u2i(ir
->operands
[0])));
1281 /* The int-to-float conversion is lossless because (value & -value) is
1282 * either a power of two or zero. We don't use the result in the zero
1283 * case. The uint() cast is necessary so that 0x80000000 does not
1284 * generate a negative value.
1286 * uint lsb_only = uint(value & -value);
1287 * float as_float = float(lsb_only);
1289 i
.insert_before(lsb_only
);
1290 i
.insert_before(assign(lsb_only
, i2u(bit_and(temp
, neg(temp
)))));
1292 i
.insert_before(as_float
);
1293 i
.insert_before(assign(as_float
, u2f(lsb_only
)));
1295 /* This is basically an open-coded frexp. Implementations that have a
1296 * native frexp instruction would be better served by that. This is
1297 * optimized versus a full-featured open-coded implementation in two ways:
1299 * - We don't care about a correct result from subnormal numbers (including
1300 * 0.0), so the raw exponent can always be safely unbiased.
1302 * - The value cannot be negative, so it does not need to be masked off to
1303 * extract the exponent.
1305 * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1307 i
.insert_before(lsb
);
1308 i
.insert_before(assign(lsb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1310 /* Use lsb_only in the comparison instead of temp so that the & (far above)
1311 * can possibly generate the result without an explicit comparison.
1313 * (lsb_only == 0) ? -1 : lsb;
1315 * Since our input values are all integers, the unbiased exponent must not
1316 * be negative. It will only be negative (-0x7f, in fact) if lsb_only is
1317 * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is
1318 * better is likely GPU dependent. Either way, the difference should be
1321 ir
->operation
= ir_triop_csel
;
1322 ir
->operands
[0] = equal(lsb_only
, c0
);
1323 ir
->operands
[1] = cminus1
;
1324 ir
->operands
[2] = new(ir
) ir_dereference_variable(lsb
);
1326 this->progress
= true;
1330 lower_instructions_visitor::find_msb_to_float_cast(ir_expression
*ir
)
1332 /* For more details, see:
1334 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1336 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1337 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1338 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1339 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1340 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1341 ir_constant
*c000000FF
= new(ir
) ir_constant(0x000000FFu
, elements
);
1342 ir_constant
*cFFFFFF00
= new(ir
) ir_constant(0xFFFFFF00u
, elements
);
1344 new(ir
) ir_variable(glsl_type::uvec(elements
), "temp", ir_var_temporary
);
1345 ir_variable
*as_float
=
1346 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1348 new(ir
) ir_variable(glsl_type::ivec(elements
), "msb", ir_var_temporary
);
1350 ir_instruction
&i
= *base_ir
;
1352 i
.insert_before(temp
);
1354 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1355 i
.insert_before(assign(temp
, ir
->operands
[0]));
1357 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1359 /* findMSB(uint(abs(some_int))) almost always does the right thing.
1360 * There are two problem values:
1362 * * 0x80000000. Since abs(0x80000000) == 0x80000000, findMSB returns
1363 * 31. However, findMSB(int(0x80000000)) == 30.
1365 * * 0xffffffff. Since abs(0xffffffff) == 1, findMSB returns
1366 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1368 * For a value of zero or negative one, -1 will be returned.
1370 * For all negative number cases, including 0x80000000 and 0xffffffff,
1371 * the correct value is obtained from findMSB if instead of negating the
1372 * (already negative) value the logical-not is used. A conditonal
1373 * logical-not can be achieved in two instructions.
1375 ir_variable
*as_int
=
1376 new(ir
) ir_variable(glsl_type::ivec(elements
), "as_int", ir_var_temporary
);
1377 ir_constant
*c31
= new(ir
) ir_constant(int(31), elements
);
1379 i
.insert_before(as_int
);
1380 i
.insert_before(assign(as_int
, ir
->operands
[0]));
1381 i
.insert_before(assign(temp
, i2u(expr(ir_binop_bit_xor
,
1383 rshift(as_int
, c31
)))));
1386 /* The int-to-float conversion is lossless because bits are conditionally
1387 * masked off the bottom of temp to ensure the value has at most 24 bits of
1388 * data or is zero. We don't use the result in the zero case. The uint()
1389 * cast is necessary so that 0x80000000 does not generate a negative value.
1391 * float as_float = float(temp > 255 ? temp & ~255 : temp);
1393 i
.insert_before(as_float
);
1394 i
.insert_before(assign(as_float
, u2f(csel(greater(temp
, c000000FF
),
1395 bit_and(temp
, cFFFFFF00
),
1398 /* This is basically an open-coded frexp. Implementations that have a
1399 * native frexp instruction would be better served by that. This is
1400 * optimized versus a full-featured open-coded implementation in two ways:
1402 * - We don't care about a correct result from subnormal numbers (including
1403 * 0.0), so the raw exponent can always be safely unbiased.
1405 * - The value cannot be negative, so it does not need to be masked off to
1406 * extract the exponent.
1408 * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1410 i
.insert_before(msb
);
1411 i
.insert_before(assign(msb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1413 /* Use msb in the comparison instead of temp so that the subtract can
1414 * possibly generate the result without an explicit comparison.
1416 * (msb < 0) ? -1 : msb;
1418 * Since our input values are all integers, the unbiased exponent must not
1419 * be negative. It will only be negative (-0x7f, in fact) if temp is 0.
1421 ir
->operation
= ir_triop_csel
;
1422 ir
->operands
[0] = less(msb
, c0
);
1423 ir
->operands
[1] = cminus1
;
1424 ir
->operands
[2] = new(ir
) ir_dereference_variable(msb
);
1426 this->progress
= true;
1430 lower_instructions_visitor::_carry(operand a
, operand b
)
1432 if (lowering(CARRY_TO_ARITH
))
1433 return i2u(b2i(less(add(a
, b
),
1434 a
.val
->clone(ralloc_parent(a
.val
), NULL
))));
1440 lower_instructions_visitor::imul_high_to_mul(ir_expression
*ir
)
1445 * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
1447 * In GLSL, (a * b) becomes
1449 * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu);
1450 * uint m2 = (a & 0x0000ffffu) * (b >> 16);
1451 * uint m3 = (a >> 16) * (b & 0x0000ffffu);
1452 * uint m4 = (a >> 16) * (b >> 16);
1459 * lo_result = uaddCarry(m1, m2 << 16, c1);
1460 * hi_result = m4 + c1;
1461 * lo_result = uaddCarry(lo_result, m3 << 16, c2);
1462 * hi_result = hi_result + c2;
1463 * hi_result = hi_result + (m2 >> 16) + (m3 >> 16);
1465 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1467 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1", ir_var_temporary
);
1468 ir_variable
*src1h
=
1469 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1h", ir_var_temporary
);
1470 ir_variable
*src1l
=
1471 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1l", ir_var_temporary
);
1473 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2", ir_var_temporary
);
1474 ir_variable
*src2h
=
1475 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2h", ir_var_temporary
);
1476 ir_variable
*src2l
=
1477 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2l", ir_var_temporary
);
1479 new(ir
) ir_variable(glsl_type::uvec(elements
), "t1", ir_var_temporary
);
1481 new(ir
) ir_variable(glsl_type::uvec(elements
), "t2", ir_var_temporary
);
1483 new(ir
) ir_variable(glsl_type::uvec(elements
), "lo", ir_var_temporary
);
1485 new(ir
) ir_variable(glsl_type::uvec(elements
), "hi", ir_var_temporary
);
1486 ir_variable
*different_signs
= NULL
;
1487 ir_constant
*c0000FFFF
= new(ir
) ir_constant(0x0000FFFFu
, elements
);
1488 ir_constant
*c16
= new(ir
) ir_constant(16u, elements
);
1490 ir_instruction
&i
= *base_ir
;
1492 i
.insert_before(src1
);
1493 i
.insert_before(src2
);
1494 i
.insert_before(src1h
);
1495 i
.insert_before(src2h
);
1496 i
.insert_before(src1l
);
1497 i
.insert_before(src2l
);
1499 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1500 i
.insert_before(assign(src1
, ir
->operands
[0]));
1501 i
.insert_before(assign(src2
, ir
->operands
[1]));
1503 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1505 ir_variable
*itmp1
=
1506 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp1", ir_var_temporary
);
1507 ir_variable
*itmp2
=
1508 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp2", ir_var_temporary
);
1509 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1511 i
.insert_before(itmp1
);
1512 i
.insert_before(itmp2
);
1513 i
.insert_before(assign(itmp1
, ir
->operands
[0]));
1514 i
.insert_before(assign(itmp2
, ir
->operands
[1]));
1517 new(ir
) ir_variable(glsl_type::bvec(elements
), "different_signs",
1520 i
.insert_before(different_signs
);
1521 i
.insert_before(assign(different_signs
, expr(ir_binop_logic_xor
,
1523 less(itmp2
, c0
->clone(ir
, NULL
)))));
1525 i
.insert_before(assign(src1
, i2u(abs(itmp1
))));
1526 i
.insert_before(assign(src2
, i2u(abs(itmp2
))));
1529 i
.insert_before(assign(src1l
, bit_and(src1
, c0000FFFF
)));
1530 i
.insert_before(assign(src2l
, bit_and(src2
, c0000FFFF
->clone(ir
, NULL
))));
1531 i
.insert_before(assign(src1h
, rshift(src1
, c16
)));
1532 i
.insert_before(assign(src2h
, rshift(src2
, c16
->clone(ir
, NULL
))));
1534 i
.insert_before(lo
);
1535 i
.insert_before(hi
);
1536 i
.insert_before(t1
);
1537 i
.insert_before(t2
);
1539 i
.insert_before(assign(lo
, mul(src1l
, src2l
)));
1540 i
.insert_before(assign(t1
, mul(src1l
, src2h
)));
1541 i
.insert_before(assign(t2
, mul(src1h
, src2l
)));
1542 i
.insert_before(assign(hi
, mul(src1h
, src2h
)));
1544 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t1
, c16
->clone(ir
, NULL
))))));
1545 i
.insert_before(assign(lo
, add(lo
, lshift(t1
, c16
->clone(ir
, NULL
)))));
1547 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t2
, c16
->clone(ir
, NULL
))))));
1548 i
.insert_before(assign(lo
, add(lo
, lshift(t2
, c16
->clone(ir
, NULL
)))));
1550 if (different_signs
== NULL
) {
1551 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1553 ir
->operation
= ir_binop_add
;
1554 ir
->operands
[0] = add(hi
, rshift(t1
, c16
->clone(ir
, NULL
)));
1555 ir
->operands
[1] = rshift(t2
, c16
->clone(ir
, NULL
));
1557 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1559 i
.insert_before(assign(hi
, add(add(hi
, rshift(t1
, c16
->clone(ir
, NULL
))),
1560 rshift(t2
, c16
->clone(ir
, NULL
)))));
1562 /* For channels where different_signs is set we have to perform a 64-bit
1563 * negation. This is *not* the same as just negating the high 32-bits.
1564 * Consider -3 * 2. The high 32-bits is 0, but the desired result is
1565 * -1, not -0! Recall -x == ~x + 1.
1567 ir_variable
*neg_hi
=
1568 new(ir
) ir_variable(glsl_type::ivec(elements
), "neg_hi", ir_var_temporary
);
1569 ir_constant
*c1
= new(ir
) ir_constant(1u, elements
);
1571 i
.insert_before(neg_hi
);
1572 i
.insert_before(assign(neg_hi
, add(bit_not(u2i(hi
)),
1573 u2i(_carry(bit_not(lo
), c1
)))));
1575 ir
->operation
= ir_triop_csel
;
1576 ir
->operands
[0] = new(ir
) ir_dereference_variable(different_signs
);
1577 ir
->operands
[1] = new(ir
) ir_dereference_variable(neg_hi
);
1578 ir
->operands
[2] = u2i(hi
);
1583 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1585 switch (ir
->operation
) {
1587 if (ir
->operands
[0]->type
->is_double())
1588 double_dot_to_fma(ir
);
1591 if (ir
->operands
[0]->type
->is_double())
1595 if (lowering(SUB_TO_ADD_NEG
))
1600 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1601 int_div_to_mul_rcp(ir
);
1602 else if ((ir
->operands
[1]->type
->is_float() ||
1603 ir
->operands
[1]->type
->is_double()) && lowering(DIV_TO_MUL_RCP
))
1608 if (lowering(EXP_TO_EXP2
))
1613 if (lowering(LOG_TO_LOG2
))
1618 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1623 if (lowering(POW_TO_EXP2
))
1627 case ir_binop_ldexp
:
1628 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1630 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1631 dldexp_to_arith(ir
);
1634 case ir_unop_frexp_exp
:
1635 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1636 dfrexp_exp_to_arith(ir
);
1639 case ir_unop_frexp_sig
:
1640 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1641 dfrexp_sig_to_arith(ir
);
1644 case ir_binop_carry
:
1645 if (lowering(CARRY_TO_ARITH
))
1649 case ir_binop_borrow
:
1650 if (lowering(BORROW_TO_ARITH
))
1651 borrow_to_arith(ir
);
1654 case ir_unop_saturate
:
1655 if (lowering(SAT_TO_CLAMP
))
1660 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1661 dtrunc_to_dfrac(ir
);
1665 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1670 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1671 dfloor_to_dfrac(ir
);
1674 case ir_unop_round_even
:
1675 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1676 dround_even_to_dfrac(ir
);
1680 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1684 case ir_unop_bit_count
:
1685 if (lowering(BIT_COUNT_TO_MATH
))
1686 bit_count_to_math(ir
);
1689 case ir_triop_bitfield_extract
:
1690 if (lowering(EXTRACT_TO_SHIFTS
))
1691 extract_to_shifts(ir
);
1694 case ir_quadop_bitfield_insert
:
1695 if (lowering(INSERT_TO_SHIFTS
))
1696 insert_to_shifts(ir
);
1699 case ir_unop_bitfield_reverse
:
1700 if (lowering(REVERSE_TO_SHIFTS
))
1701 reverse_to_shifts(ir
);
1704 case ir_unop_find_lsb
:
1705 if (lowering(FIND_LSB_TO_FLOAT_CAST
))
1706 find_lsb_to_float_cast(ir
);
1709 case ir_unop_find_msb
:
1710 if (lowering(FIND_MSB_TO_FLOAT_CAST
))
1711 find_msb_to_float_cast(ir
);
1714 case ir_binop_imul_high
:
1715 if (lowering(IMUL_HIGH_TO_MUL
))
1716 imul_high_to_mul(ir
);
1720 return visit_continue
;
1723 return visit_continue
;