2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file lower_instructions.cpp
27 * Many GPUs lack native instructions for certain expression operations, and
28 * must replace them with some other expression tree. This pass lowers some
29 * of the most common cases, allowing the lowering code to be implemented once
30 * rather than in each driver backend.
32 * Currently supported transformations:
35 * - INT_DIV_TO_MUL_RCP
49 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
51 * This simplifies expression reassociation, and for many backends
52 * there is no subtract operation separate from adding the negation.
53 * For backends with native subtract operations, they will probably
54 * want to recognize add(op0, neg(op1)) or the other way around to
55 * produce a subtract anyway.
57 * FDIV_TO_MUL_RCP, DDIV_TO_MUL_RCP, and INT_DIV_TO_MUL_RCP:
58 * ---------------------------------------------------------
59 * Breaks an ir_binop_div expression down to op0 * (rcp(op1)).
61 * Many GPUs don't have a divide instruction (945 and 965 included),
62 * but they do have an RCP instruction to compute an approximate
63 * reciprocal. By breaking the operation down, constant reciprocals
64 * can get constant folded.
66 * FDIV_TO_MUL_RCP only lowers single-precision floating point division;
67 * DDIV_TO_MUL_RCP only lowers double-precision floating point division.
68 * DIV_TO_MUL_RCP is a convenience macro that sets both flags.
69 * INT_DIV_TO_MUL_RCP handles the integer case, converting to and from floating
70 * point so that RCP is possible.
72 * EXP_TO_EXP2 and LOG_TO_LOG2:
73 * ----------------------------
74 * Many GPUs don't have a base e log or exponent instruction, but they
75 * do have base 2 versions, so this pass converts exp and log to exp2
76 * and log2 operations.
80 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
81 * x**y to 2**(y * log2(x)).
85 * Breaks an ir_binop_mod expression down to (op0 - op1 * floor(op0 / op1))
87 * Many GPUs don't have a MOD instruction (945 and 965 included), and
88 * if we have to break it down like this anyway, it gives an
89 * opportunity to do things like constant fold the (1.0 / op1) easily.
91 * Note: before we used to implement this as op1 * fract(op / op1) but this
92 * implementation had significant precision errors.
96 * Converts ir_binop_ldexp to arithmetic and bit operations for float sources.
98 * DFREXP_DLDEXP_TO_ARITH:
100 * Converts ir_binop_ldexp, ir_unop_frexp_sig, and ir_unop_frexp_exp to
101 * arithmetic and bit ops for double arguments.
105 * Converts ir_carry into (x + y) < x.
109 * Converts ir_borrow into (x < y).
113 * Converts ir_unop_saturate into min(max(x, 0.0), 1.0)
117 * Converts double trunc, ceil, floor, round to fract
120 #include "c99_math.h"
121 #include "program/prog_instruction.h" /* for swizzle */
122 #include "compiler/glsl_types.h"
124 #include "ir_builder.h"
125 #include "ir_optimization.h"
127 using namespace ir_builder
;
131 class lower_instructions_visitor
: public ir_hierarchical_visitor
{
133 lower_instructions_visitor(unsigned lower
)
134 : progress(false), lower(lower
) { }
136 ir_visitor_status
visit_leave(ir_expression
*);
141 unsigned lower
; /** Bitfield of which operations to lower */
143 void sub_to_add_neg(ir_expression
*);
144 void div_to_mul_rcp(ir_expression
*);
145 void int_div_to_mul_rcp(ir_expression
*);
146 void mod_to_floor(ir_expression
*);
147 void exp_to_exp2(ir_expression
*);
148 void pow_to_exp2(ir_expression
*);
149 void log_to_log2(ir_expression
*);
150 void ldexp_to_arith(ir_expression
*);
151 void dldexp_to_arith(ir_expression
*);
152 void dfrexp_sig_to_arith(ir_expression
*);
153 void dfrexp_exp_to_arith(ir_expression
*);
154 void carry_to_arith(ir_expression
*);
155 void borrow_to_arith(ir_expression
*);
156 void sat_to_clamp(ir_expression
*);
157 void double_dot_to_fma(ir_expression
*);
158 void double_lrp(ir_expression
*);
159 void dceil_to_dfrac(ir_expression
*);
160 void dfloor_to_dfrac(ir_expression
*);
161 void dround_even_to_dfrac(ir_expression
*);
162 void dtrunc_to_dfrac(ir_expression
*);
163 void dsign_to_csel(ir_expression
*);
164 void bit_count_to_math(ir_expression
*);
165 void extract_to_shifts(ir_expression
*);
166 void insert_to_shifts(ir_expression
*);
167 void reverse_to_shifts(ir_expression
*ir
);
168 void find_lsb_to_float_cast(ir_expression
*ir
);
169 void find_msb_to_float_cast(ir_expression
*ir
);
170 void imul_high_to_mul(ir_expression
*ir
);
171 void sqrt_to_abs_sqrt(ir_expression
*ir
);
173 ir_expression
*_carry(operand a
, operand b
);
176 } /* anonymous namespace */
179 * Determine if a particular type of lowering should occur
181 #define lowering(x) (this->lower & x)
184 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
186 lower_instructions_visitor
v(what_to_lower
);
188 visit_list_elements(&v
, instructions
);
193 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
195 ir
->operation
= ir_binop_add
;
196 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
197 ir
->operands
[1], NULL
);
198 this->progress
= true;
202 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
204 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
206 /* New expression for the 1.0 / op1 */
208 expr
= new(ir
) ir_expression(ir_unop_rcp
,
209 ir
->operands
[1]->type
,
212 /* op0 / op1 -> op0 * (1.0 / op1) */
213 ir
->operation
= ir_binop_mul
;
214 ir
->operands
[1] = expr
;
216 this->progress
= true;
220 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
222 assert(ir
->operands
[1]->type
->is_integer());
224 /* Be careful with integer division -- we need to do it as a
225 * float and re-truncate, since rcp(n > 1) of an integer would
228 ir_rvalue
*op0
, *op1
;
229 const struct glsl_type
*vec_type
;
231 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
232 ir
->operands
[1]->type
->vector_elements
,
233 ir
->operands
[1]->type
->matrix_columns
);
235 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
236 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
238 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
240 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
242 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
243 ir
->operands
[0]->type
->vector_elements
,
244 ir
->operands
[0]->type
->matrix_columns
);
246 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
247 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
249 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
251 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
252 ir
->type
->vector_elements
,
253 ir
->type
->matrix_columns
);
255 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
257 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
258 ir
->operation
= ir_unop_f2i
;
259 ir
->operands
[0] = op0
;
261 ir
->operation
= ir_unop_i2u
;
262 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
264 ir
->operands
[1] = NULL
;
266 this->progress
= true;
270 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
272 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
274 ir
->operation
= ir_unop_exp2
;
275 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
276 ir
->operands
[0], log2_e
);
277 this->progress
= true;
281 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
283 ir_expression
*const log2_x
=
284 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
287 ir
->operation
= ir_unop_exp2
;
288 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
289 ir
->operands
[1], log2_x
);
290 ir
->operands
[1] = NULL
;
291 this->progress
= true;
295 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
297 ir
->operation
= ir_binop_mul
;
298 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
299 ir
->operands
[0], NULL
);
300 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
301 this->progress
= true;
305 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
307 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
309 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
311 this->base_ir
->insert_before(x
);
312 this->base_ir
->insert_before(y
);
314 ir_assignment
*const assign_x
=
315 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
316 ir
->operands
[0], NULL
);
317 ir_assignment
*const assign_y
=
318 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
319 ir
->operands
[1], NULL
);
321 this->base_ir
->insert_before(assign_x
);
322 this->base_ir
->insert_before(assign_y
);
324 ir_expression
*const div_expr
=
325 new(ir
) ir_expression(ir_binop_div
, x
->type
,
326 new(ir
) ir_dereference_variable(x
),
327 new(ir
) ir_dereference_variable(y
));
329 /* Don't generate new IR that would need to be lowered in an additional
332 if ((lowering(FDIV_TO_MUL_RCP
) && ir
->type
->is_float()) ||
333 (lowering(DDIV_TO_MUL_RCP
) && ir
->type
->is_double()))
334 div_to_mul_rcp(div_expr
);
336 ir_expression
*const floor_expr
=
337 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
339 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
340 dfloor_to_dfrac(floor_expr
);
342 ir_expression
*const mul_expr
=
343 new(ir
) ir_expression(ir_binop_mul
,
344 new(ir
) ir_dereference_variable(y
),
347 ir
->operation
= ir_binop_sub
;
348 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
349 ir
->operands
[1] = mul_expr
;
350 this->progress
= true;
354 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
357 * ir_binop_ldexp x exp
360 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
361 * resulting_biased_exp = extracted_biased_exp + exp;
363 * if (resulting_biased_exp < 1 || x == 0.0f) {
364 * return copysign(0.0, x);
367 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
368 * lshift(i2u(resulting_biased_exp), exp_shift));
370 * which we can't actually implement as such, since the GLSL IR doesn't
371 * have vectorized if-statements. We actually implement it without branches
372 * using conditional-select:
374 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
375 * resulting_biased_exp = extracted_biased_exp + exp;
377 * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
378 * gequal(resulting_biased_exp, 1);
379 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
380 * resulting_biased_exp = csel(is_not_zero_or_underflow,
381 * resulting_biased_exp, 0);
383 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
384 * lshift(i2u(resulting_biased_exp), exp_shift));
387 const unsigned vec_elem
= ir
->type
->vector_elements
;
390 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
391 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
394 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
396 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
398 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
400 /* Temporary variables */
401 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
402 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
404 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
407 ir_variable
*extracted_biased_exp
=
408 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
409 ir_variable
*resulting_biased_exp
=
410 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
412 ir_variable
*is_not_zero_or_underflow
=
413 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
415 ir_instruction
&i
= *base_ir
;
417 /* Copy <x> and <exp> arguments. */
419 i
.insert_before(assign(x
, ir
->operands
[0]));
420 i
.insert_before(exp
);
421 i
.insert_before(assign(exp
, ir
->operands
[1]));
423 /* Extract the biased exponent from <x>. */
424 i
.insert_before(extracted_biased_exp
);
425 i
.insert_before(assign(extracted_biased_exp
,
426 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
428 i
.insert_before(resulting_biased_exp
);
429 i
.insert_before(assign(resulting_biased_exp
,
430 add(extracted_biased_exp
, exp
)));
432 /* Test if result is ±0.0, subnormal, or underflow by checking if the
433 * resulting biased exponent would be less than 0x1. If so, the result is
434 * 0.0 with the sign of x. (Actually, invert the conditions so that
435 * immediate values are the second arguments, which is better for i965)
437 i
.insert_before(zero_sign_x
);
438 i
.insert_before(assign(zero_sign_x
,
439 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
441 i
.insert_before(is_not_zero_or_underflow
);
442 i
.insert_before(assign(is_not_zero_or_underflow
,
443 logic_and(nequal(x
, new(ir
) ir_constant(0.0f
, vec_elem
)),
444 gequal(resulting_biased_exp
,
445 new(ir
) ir_constant(0x1, vec_elem
)))));
446 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
448 i
.insert_before(assign(resulting_biased_exp
,
449 csel(is_not_zero_or_underflow
,
450 resulting_biased_exp
, zeroi
)));
452 /* We could test for overflows by checking if the resulting biased exponent
453 * would be greater than 0xFE. Turns out we don't need to because the GLSL
456 * "If this product is too large to be represented in the
457 * floating-point type, the result is undefined."
460 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
462 /* Don't generate new IR that would need to be lowered in an additional
465 if (!lowering(INSERT_TO_SHIFTS
)) {
466 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
467 ir
->operation
= ir_unop_bitcast_i2f
;
468 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
469 exp_shift_clone
, exp_width
);
470 ir
->operands
[1] = NULL
;
472 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x807fffffu
, vec_elem
);
473 ir
->operation
= ir_unop_bitcast_u2f
;
474 ir
->operands
[0] = bit_or(bit_and(bitcast_f2u(x
), sign_mantissa_mask
),
475 lshift(i2u(resulting_biased_exp
), exp_shift_clone
));
478 this->progress
= true;
482 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
484 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
485 * from the significand.
488 const unsigned vec_elem
= ir
->type
->vector_elements
;
491 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
492 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
495 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
497 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
499 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
500 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
501 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
503 /* Temporary variables */
504 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
505 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
507 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
510 ir_variable
*extracted_biased_exp
=
511 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
512 ir_variable
*resulting_biased_exp
=
513 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
515 ir_variable
*is_not_zero_or_underflow
=
516 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
518 ir_instruction
&i
= *base_ir
;
520 /* Copy <x> and <exp> arguments. */
522 i
.insert_before(assign(x
, ir
->operands
[0]));
523 i
.insert_before(exp
);
524 i
.insert_before(assign(exp
, ir
->operands
[1]));
526 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
527 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
528 dfrexp_exp_to_arith(frexp_exp
);
530 /* Extract the biased exponent from <x>. */
531 i
.insert_before(extracted_biased_exp
);
532 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
534 i
.insert_before(resulting_biased_exp
);
535 i
.insert_before(assign(resulting_biased_exp
,
536 add(extracted_biased_exp
, exp
)));
538 /* Test if result is ±0.0, subnormal, or underflow by checking if the
539 * resulting biased exponent would be less than 0x1. If so, the result is
540 * 0.0 with the sign of x. (Actually, invert the conditions so that
541 * immediate values are the second arguments, which is better for i965)
542 * TODO: Implement in a vector fashion.
544 i
.insert_before(zero_sign_x
);
545 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
546 ir_variable
*unpacked
=
547 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
548 i
.insert_before(unpacked
);
551 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
552 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
554 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
555 i
.insert_before(assign(zero_sign_x
,
556 expr(ir_unop_pack_double_2x32
, unpacked
),
559 i
.insert_before(is_not_zero_or_underflow
);
560 i
.insert_before(assign(is_not_zero_or_underflow
,
561 gequal(resulting_biased_exp
,
562 new(ir
) ir_constant(0x1, vec_elem
))));
563 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
565 i
.insert_before(assign(resulting_biased_exp
,
566 csel(is_not_zero_or_underflow
,
567 resulting_biased_exp
, zeroi
)));
569 /* We could test for overflows by checking if the resulting biased exponent
570 * would be greater than 0xFE. Turns out we don't need to because the GLSL
573 * "If this product is too large to be represented in the
574 * floating-point type, the result is undefined."
577 ir_rvalue
*results
[4] = {NULL
};
578 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
579 ir_variable
*unpacked
=
580 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
581 i
.insert_before(unpacked
);
584 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
586 ir_expression
*bfi
= bitfield_insert(
588 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
589 exp_shift
->clone(ir
, NULL
),
590 exp_width
->clone(ir
, NULL
));
592 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
594 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
597 ir
->operation
= ir_quadop_vector
;
598 ir
->operands
[0] = results
[0];
599 ir
->operands
[1] = results
[1];
600 ir
->operands
[2] = results
[2];
601 ir
->operands
[3] = results
[3];
603 /* Don't generate new IR that would need to be lowered in an additional
607 this->progress
= true;
611 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
613 const unsigned vec_elem
= ir
->type
->vector_elements
;
614 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
616 /* Double-precision floating-point values are stored as
621 * We're just extracting the significand here, so we only need to modify
622 * the upper 32-bit uint. Unfortunately we must extract each double
623 * independently as there is no vector version of unpackDouble.
626 ir_instruction
&i
= *base_ir
;
628 ir_variable
*is_not_zero
=
629 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
630 ir_rvalue
*results
[4] = {NULL
};
632 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
633 i
.insert_before(is_not_zero
);
636 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
638 /* TODO: Remake this as more vector-friendly when int64 support is
641 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
642 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
643 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
645 /* Exponent of double floating-point values in the range [0.5, 1.0). */
646 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
649 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
650 ir_variable
*unpacked
=
651 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
653 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
655 i
.insert_before(bits
);
656 i
.insert_before(unpacked
);
657 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
659 /* Manipulate the high uint to remove the exponent and replace it with
660 * either the default exponent or zero.
662 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
663 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
664 i
.insert_before(assign(bits
, bit_or(bits
,
665 csel(swizzle(is_not_zero
, elem
, 1),
668 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
669 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
672 /* Put the dvec back together */
673 ir
->operation
= ir_quadop_vector
;
674 ir
->operands
[0] = results
[0];
675 ir
->operands
[1] = results
[1];
676 ir
->operands
[2] = results
[2];
677 ir
->operands
[3] = results
[3];
679 this->progress
= true;
683 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
685 const unsigned vec_elem
= ir
->type
->vector_elements
;
686 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
687 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
689 /* Double-precision floating-point values are stored as
694 * We're just extracting the exponent here, so we only care about the upper
698 ir_instruction
&i
= *base_ir
;
700 ir_variable
*is_not_zero
=
701 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
702 ir_variable
*high_words
=
703 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
704 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
705 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
707 ir_rvalue
*absval
= abs(ir
->operands
[0]);
709 i
.insert_before(is_not_zero
);
710 i
.insert_before(high_words
);
711 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
713 /* Extract all of the upper uints. */
714 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
715 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
717 i
.insert_before(assign(high_words
,
718 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
722 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
723 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
725 /* For non-zero inputs, shift the exponent down and apply bias. */
726 ir
->operation
= ir_triop_csel
;
727 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
728 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
729 ir
->operands
[2] = izero
;
731 this->progress
= true;
735 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
740 * sum = ir_binop_add x y
741 * bcarry = ir_binop_less sum x
742 * carry = ir_unop_b2i bcarry
745 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
746 ir
->operation
= ir_unop_i2u
;
747 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
748 ir
->operands
[1] = NULL
;
750 this->progress
= true;
754 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
757 * ir_binop_borrow x y
759 * bcarry = ir_binop_less x y
760 * carry = ir_unop_b2i bcarry
763 ir
->operation
= ir_unop_i2u
;
764 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
765 ir
->operands
[1] = NULL
;
767 this->progress
= true;
771 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
776 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
779 ir
->operation
= ir_binop_min
;
780 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
782 new(ir
) ir_constant(0.0f
));
783 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
785 this->progress
= true;
789 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
791 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
793 this->base_ir
->insert_before(temp
);
795 int nc
= ir
->operands
[0]->type
->components();
796 for (int i
= nc
- 1; i
>= 1; i
--) {
797 ir_assignment
*assig
;
799 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
800 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
802 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
803 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
806 this->base_ir
->insert_before(assig
);
809 ir
->operation
= ir_triop_fma
;
810 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
811 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
812 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
814 this->progress
= true;
819 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
822 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
823 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
825 switch (op2
->type
->vector_elements
) {
827 swizval
= SWIZZLE_XXXX
;
830 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
831 swizval
= SWIZZLE_XYZW
;
835 ir
->operation
= ir_triop_fma
;
836 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
837 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
839 this->progress
= true;
843 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
847 * temp = sub(x, frtemp);
848 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
850 ir_instruction
&i
= *base_ir
;
851 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
852 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
853 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
856 i
.insert_before(frtemp
);
857 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
859 ir
->operation
= ir_binop_add
;
860 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
861 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
863 this->progress
= true;
867 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
871 * result = sub(x, frtemp);
873 ir
->operation
= ir_binop_sub
;
874 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
876 this->progress
= true;
879 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
884 * frtemp = frac(temp);
885 * t2 = sub(temp, frtemp);
886 * if (frac(x) == 0.5)
887 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
892 ir_instruction
&i
= *base_ir
;
893 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
895 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
897 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
899 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
900 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
901 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
903 i
.insert_before(temp
);
904 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
906 i
.insert_before(frtemp
);
907 i
.insert_before(assign(frtemp
, fract(temp
)));
910 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
912 ir
->operation
= ir_triop_csel
;
913 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
914 p5
->clone(ir
, NULL
));
915 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
919 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
921 this->progress
= true;
925 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
929 * temp = sub(x, frtemp);
930 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
932 ir_rvalue
*arg
= ir
->operands
[0];
933 ir_instruction
&i
= *base_ir
;
935 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
936 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
937 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
939 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
942 i
.insert_before(frtemp
);
943 i
.insert_before(assign(frtemp
, fract(arg
)));
944 i
.insert_before(temp
);
945 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
947 ir
->operation
= ir_triop_csel
;
948 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
949 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
950 ir
->operands
[2] = add(temp
,
951 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
952 zero
->clone(ir
, NULL
),
955 this->progress
= true;
959 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
962 * temp = x > 0.0 ? 1.0 : 0.0;
963 * result = x < 0.0 ? -1.0 : temp;
965 ir_rvalue
*arg
= ir
->operands
[0];
966 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
967 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
968 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
970 ir
->operation
= ir_triop_csel
;
971 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
972 zero
->clone(ir
, NULL
));
973 ir
->operands
[1] = neg_one
;
974 ir
->operands
[2] = csel(greater(arg
, zero
),
976 zero
->clone(ir
, NULL
));
978 this->progress
= true;
982 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
984 /* For more details, see:
986 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
988 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
989 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
991 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
992 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
993 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
994 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
995 ir_constant
*c1
= new(ir
) ir_constant(1u);
996 ir_constant
*c2
= new(ir
) ir_constant(2u);
997 ir_constant
*c4
= new(ir
) ir_constant(4u);
998 ir_constant
*c24
= new(ir
) ir_constant(24u);
1000 base_ir
->insert_before(temp
);
1002 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1003 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
1005 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1006 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
1009 /* temp = temp - ((temp >> 1) & 0x55555555u); */
1010 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
1013 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
1014 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
1015 bit_and(rshift(temp
, c2
),
1016 c33333333
->clone(ir
, NULL
)))));
1018 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1019 ir
->operation
= ir_unop_u2i
;
1020 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1024 this->progress
= true;
1028 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1031 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1033 base_ir
->insert_before(bits
);
1034 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1036 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1038 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1040 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1041 ir_constant
*cFFFFFFFF
=
1042 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1044 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1045 * we'd get a mask of 0 when bits is 32. Special case it.
1047 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1049 ir_expression
*mask
= csel(equal(bits
, c32
),
1051 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1053 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1055 * If bits is zero, the result will be zero.
1057 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1058 * select as in the signed integer case.
1060 * (value >> offset) & mask;
1062 ir
->operation
= ir_binop_bit_and
;
1063 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1064 ir
->operands
[1] = mask
;
1065 ir
->operands
[2] = NULL
;
1068 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1070 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1072 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1074 /* temp = 32 - bits; */
1075 base_ir
->insert_before(temp
);
1076 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1078 /* expr = value << (temp - offset)) >> temp; */
1079 ir_expression
*expr
=
1080 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1082 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1084 * If bits is zero, the result will be zero.
1086 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1087 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1090 * (bits == 0) ? 0 : e;
1092 ir
->operation
= ir_triop_csel
;
1093 ir
->operands
[0] = equal(c0
, bits
);
1094 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1095 ir
->operands
[2] = expr
;
1098 this->progress
= true;
1102 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1106 ir_constant
*cFFFFFFFF
;
1107 ir_variable
*offset
=
1108 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1110 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1112 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1114 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1115 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1116 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1117 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1119 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1121 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1122 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1123 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1126 base_ir
->insert_before(offset
);
1127 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1129 base_ir
->insert_before(bits
);
1130 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1132 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1133 * we'd get a mask of 0 when bits is 32. Special case it.
1135 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1137 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1139 * The result will be undefined if offset or bits is negative, or if the
1140 * sum of offset and bits is greater than the number of bits used to
1141 * store the operand.
1143 * Since it's undefined, there are a couple other ways this could be
1144 * implemented. The other way that was considered was to put the csel
1145 * around the whole thing:
1147 * final_result = bits == 32 ? insert : ... ;
1149 base_ir
->insert_before(mask
);
1151 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1153 lshift(sub(lshift(c1
, bits
),
1154 c1
->clone(ir
, NULL
)),
1157 /* (base & ~mask) | ((insert << offset) & mask) */
1158 ir
->operation
= ir_binop_bit_or
;
1159 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1160 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1161 ir
->operands
[2] = NULL
;
1162 ir
->operands
[3] = NULL
;
1164 this->progress
= true;
1168 lower_instructions_visitor::reverse_to_shifts(ir_expression
*ir
)
1170 /* For more details, see:
1172 * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1175 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1177 new(ir
) ir_constant(2u, ir
->operands
[0]->type
->vector_elements
);
1179 new(ir
) ir_constant(4u, ir
->operands
[0]->type
->vector_elements
);
1181 new(ir
) ir_constant(8u, ir
->operands
[0]->type
->vector_elements
);
1183 new(ir
) ir_constant(16u, ir
->operands
[0]->type
->vector_elements
);
1184 ir_constant
*c33333333
=
1185 new(ir
) ir_constant(0x33333333u
, ir
->operands
[0]->type
->vector_elements
);
1186 ir_constant
*c55555555
=
1187 new(ir
) ir_constant(0x55555555u
, ir
->operands
[0]->type
->vector_elements
);
1188 ir_constant
*c0F0F0F0F
=
1189 new(ir
) ir_constant(0x0F0F0F0Fu
, ir
->operands
[0]->type
->vector_elements
);
1190 ir_constant
*c00FF00FF
=
1191 new(ir
) ir_constant(0x00FF00FFu
, ir
->operands
[0]->type
->vector_elements
);
1193 new(ir
) ir_variable(glsl_type::uvec(ir
->operands
[0]->type
->vector_elements
),
1194 "temp", ir_var_temporary
);
1195 ir_instruction
&i
= *base_ir
;
1197 i
.insert_before(temp
);
1199 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1200 i
.insert_before(assign(temp
, ir
->operands
[0]));
1202 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1203 i
.insert_before(assign(temp
, i2u(ir
->operands
[0])));
1206 /* Swap odd and even bits.
1208 * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1210 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c1
), c55555555
),
1211 lshift(bit_and(temp
, c55555555
->clone(ir
, NULL
)),
1212 c1
->clone(ir
, NULL
)))));
1213 /* Swap consecutive pairs.
1215 * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1217 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c2
), c33333333
),
1218 lshift(bit_and(temp
, c33333333
->clone(ir
, NULL
)),
1219 c2
->clone(ir
, NULL
)))));
1223 * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1225 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c4
), c0F0F0F0F
),
1226 lshift(bit_and(temp
, c0F0F0F0F
->clone(ir
, NULL
)),
1227 c4
->clone(ir
, NULL
)))));
1229 /* The last step is, basically, bswap. Swap the bytes, then swap the
1230 * words. When this code is run through GCC on x86, it does generate a
1231 * bswap instruction.
1233 * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1234 * temp = ( temp >> 16 ) | ( temp << 16);
1236 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c8
), c00FF00FF
),
1237 lshift(bit_and(temp
, c00FF00FF
->clone(ir
, NULL
)),
1238 c8
->clone(ir
, NULL
)))));
1240 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1241 ir
->operation
= ir_binop_bit_or
;
1242 ir
->operands
[0] = rshift(temp
, c16
);
1243 ir
->operands
[1] = lshift(temp
, c16
->clone(ir
, NULL
));
1245 ir
->operation
= ir_unop_u2i
;
1246 ir
->operands
[0] = bit_or(rshift(temp
, c16
),
1247 lshift(temp
, c16
->clone(ir
, NULL
)));
1250 this->progress
= true;
1254 lower_instructions_visitor::find_lsb_to_float_cast(ir_expression
*ir
)
1256 /* For more details, see:
1258 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1260 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1261 ir_constant
*c0
= new(ir
) ir_constant(unsigned(0), elements
);
1262 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1263 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1264 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1266 new(ir
) ir_variable(glsl_type::ivec(elements
), "temp", ir_var_temporary
);
1267 ir_variable
*lsb_only
=
1268 new(ir
) ir_variable(glsl_type::uvec(elements
), "lsb_only", ir_var_temporary
);
1269 ir_variable
*as_float
=
1270 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1272 new(ir
) ir_variable(glsl_type::ivec(elements
), "lsb", ir_var_temporary
);
1274 ir_instruction
&i
= *base_ir
;
1276 i
.insert_before(temp
);
1278 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1279 i
.insert_before(assign(temp
, ir
->operands
[0]));
1281 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1282 i
.insert_before(assign(temp
, u2i(ir
->operands
[0])));
1285 /* The int-to-float conversion is lossless because (value & -value) is
1286 * either a power of two or zero. We don't use the result in the zero
1287 * case. The uint() cast is necessary so that 0x80000000 does not
1288 * generate a negative value.
1290 * uint lsb_only = uint(value & -value);
1291 * float as_float = float(lsb_only);
1293 i
.insert_before(lsb_only
);
1294 i
.insert_before(assign(lsb_only
, i2u(bit_and(temp
, neg(temp
)))));
1296 i
.insert_before(as_float
);
1297 i
.insert_before(assign(as_float
, u2f(lsb_only
)));
1299 /* This is basically an open-coded frexp. Implementations that have a
1300 * native frexp instruction would be better served by that. This is
1301 * optimized versus a full-featured open-coded implementation in two ways:
1303 * - We don't care about a correct result from subnormal numbers (including
1304 * 0.0), so the raw exponent can always be safely unbiased.
1306 * - The value cannot be negative, so it does not need to be masked off to
1307 * extract the exponent.
1309 * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1311 i
.insert_before(lsb
);
1312 i
.insert_before(assign(lsb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1314 /* Use lsb_only in the comparison instead of temp so that the & (far above)
1315 * can possibly generate the result without an explicit comparison.
1317 * (lsb_only == 0) ? -1 : lsb;
1319 * Since our input values are all integers, the unbiased exponent must not
1320 * be negative. It will only be negative (-0x7f, in fact) if lsb_only is
1321 * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is
1322 * better is likely GPU dependent. Either way, the difference should be
1325 ir
->operation
= ir_triop_csel
;
1326 ir
->operands
[0] = equal(lsb_only
, c0
);
1327 ir
->operands
[1] = cminus1
;
1328 ir
->operands
[2] = new(ir
) ir_dereference_variable(lsb
);
1330 this->progress
= true;
1334 lower_instructions_visitor::find_msb_to_float_cast(ir_expression
*ir
)
1336 /* For more details, see:
1338 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1340 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1341 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1342 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1343 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1344 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1345 ir_constant
*c000000FF
= new(ir
) ir_constant(0x000000FFu
, elements
);
1346 ir_constant
*cFFFFFF00
= new(ir
) ir_constant(0xFFFFFF00u
, elements
);
1348 new(ir
) ir_variable(glsl_type::uvec(elements
), "temp", ir_var_temporary
);
1349 ir_variable
*as_float
=
1350 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1352 new(ir
) ir_variable(glsl_type::ivec(elements
), "msb", ir_var_temporary
);
1354 ir_instruction
&i
= *base_ir
;
1356 i
.insert_before(temp
);
1358 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1359 i
.insert_before(assign(temp
, ir
->operands
[0]));
1361 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1363 /* findMSB(uint(abs(some_int))) almost always does the right thing.
1364 * There are two problem values:
1366 * * 0x80000000. Since abs(0x80000000) == 0x80000000, findMSB returns
1367 * 31. However, findMSB(int(0x80000000)) == 30.
1369 * * 0xffffffff. Since abs(0xffffffff) == 1, findMSB returns
1370 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1372 * For a value of zero or negative one, -1 will be returned.
1374 * For all negative number cases, including 0x80000000 and 0xffffffff,
1375 * the correct value is obtained from findMSB if instead of negating the
1376 * (already negative) value the logical-not is used. A conditonal
1377 * logical-not can be achieved in two instructions.
1379 ir_variable
*as_int
=
1380 new(ir
) ir_variable(glsl_type::ivec(elements
), "as_int", ir_var_temporary
);
1381 ir_constant
*c31
= new(ir
) ir_constant(int(31), elements
);
1383 i
.insert_before(as_int
);
1384 i
.insert_before(assign(as_int
, ir
->operands
[0]));
1385 i
.insert_before(assign(temp
, i2u(expr(ir_binop_bit_xor
,
1387 rshift(as_int
, c31
)))));
1390 /* The int-to-float conversion is lossless because bits are conditionally
1391 * masked off the bottom of temp to ensure the value has at most 24 bits of
1392 * data or is zero. We don't use the result in the zero case. The uint()
1393 * cast is necessary so that 0x80000000 does not generate a negative value.
1395 * float as_float = float(temp > 255 ? temp & ~255 : temp);
1397 i
.insert_before(as_float
);
1398 i
.insert_before(assign(as_float
, u2f(csel(greater(temp
, c000000FF
),
1399 bit_and(temp
, cFFFFFF00
),
1402 /* This is basically an open-coded frexp. Implementations that have a
1403 * native frexp instruction would be better served by that. This is
1404 * optimized versus a full-featured open-coded implementation in two ways:
1406 * - We don't care about a correct result from subnormal numbers (including
1407 * 0.0), so the raw exponent can always be safely unbiased.
1409 * - The value cannot be negative, so it does not need to be masked off to
1410 * extract the exponent.
1412 * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1414 i
.insert_before(msb
);
1415 i
.insert_before(assign(msb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1417 /* Use msb in the comparison instead of temp so that the subtract can
1418 * possibly generate the result without an explicit comparison.
1420 * (msb < 0) ? -1 : msb;
1422 * Since our input values are all integers, the unbiased exponent must not
1423 * be negative. It will only be negative (-0x7f, in fact) if temp is 0.
1425 ir
->operation
= ir_triop_csel
;
1426 ir
->operands
[0] = less(msb
, c0
);
1427 ir
->operands
[1] = cminus1
;
1428 ir
->operands
[2] = new(ir
) ir_dereference_variable(msb
);
1430 this->progress
= true;
1434 lower_instructions_visitor::_carry(operand a
, operand b
)
1436 if (lowering(CARRY_TO_ARITH
))
1437 return i2u(b2i(less(add(a
, b
),
1438 a
.val
->clone(ralloc_parent(a
.val
), NULL
))));
1444 lower_instructions_visitor::imul_high_to_mul(ir_expression
*ir
)
1449 * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
1451 * In GLSL, (a * b) becomes
1453 * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu);
1454 * uint m2 = (a & 0x0000ffffu) * (b >> 16);
1455 * uint m3 = (a >> 16) * (b & 0x0000ffffu);
1456 * uint m4 = (a >> 16) * (b >> 16);
1463 * lo_result = uaddCarry(m1, m2 << 16, c1);
1464 * hi_result = m4 + c1;
1465 * lo_result = uaddCarry(lo_result, m3 << 16, c2);
1466 * hi_result = hi_result + c2;
1467 * hi_result = hi_result + (m2 >> 16) + (m3 >> 16);
1469 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1471 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1", ir_var_temporary
);
1472 ir_variable
*src1h
=
1473 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1h", ir_var_temporary
);
1474 ir_variable
*src1l
=
1475 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1l", ir_var_temporary
);
1477 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2", ir_var_temporary
);
1478 ir_variable
*src2h
=
1479 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2h", ir_var_temporary
);
1480 ir_variable
*src2l
=
1481 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2l", ir_var_temporary
);
1483 new(ir
) ir_variable(glsl_type::uvec(elements
), "t1", ir_var_temporary
);
1485 new(ir
) ir_variable(glsl_type::uvec(elements
), "t2", ir_var_temporary
);
1487 new(ir
) ir_variable(glsl_type::uvec(elements
), "lo", ir_var_temporary
);
1489 new(ir
) ir_variable(glsl_type::uvec(elements
), "hi", ir_var_temporary
);
1490 ir_variable
*different_signs
= NULL
;
1491 ir_constant
*c0000FFFF
= new(ir
) ir_constant(0x0000FFFFu
, elements
);
1492 ir_constant
*c16
= new(ir
) ir_constant(16u, elements
);
1494 ir_instruction
&i
= *base_ir
;
1496 i
.insert_before(src1
);
1497 i
.insert_before(src2
);
1498 i
.insert_before(src1h
);
1499 i
.insert_before(src2h
);
1500 i
.insert_before(src1l
);
1501 i
.insert_before(src2l
);
1503 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1504 i
.insert_before(assign(src1
, ir
->operands
[0]));
1505 i
.insert_before(assign(src2
, ir
->operands
[1]));
1507 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1509 ir_variable
*itmp1
=
1510 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp1", ir_var_temporary
);
1511 ir_variable
*itmp2
=
1512 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp2", ir_var_temporary
);
1513 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1515 i
.insert_before(itmp1
);
1516 i
.insert_before(itmp2
);
1517 i
.insert_before(assign(itmp1
, ir
->operands
[0]));
1518 i
.insert_before(assign(itmp2
, ir
->operands
[1]));
1521 new(ir
) ir_variable(glsl_type::bvec(elements
), "different_signs",
1524 i
.insert_before(different_signs
);
1525 i
.insert_before(assign(different_signs
, expr(ir_binop_logic_xor
,
1527 less(itmp2
, c0
->clone(ir
, NULL
)))));
1529 i
.insert_before(assign(src1
, i2u(abs(itmp1
))));
1530 i
.insert_before(assign(src2
, i2u(abs(itmp2
))));
1533 i
.insert_before(assign(src1l
, bit_and(src1
, c0000FFFF
)));
1534 i
.insert_before(assign(src2l
, bit_and(src2
, c0000FFFF
->clone(ir
, NULL
))));
1535 i
.insert_before(assign(src1h
, rshift(src1
, c16
)));
1536 i
.insert_before(assign(src2h
, rshift(src2
, c16
->clone(ir
, NULL
))));
1538 i
.insert_before(lo
);
1539 i
.insert_before(hi
);
1540 i
.insert_before(t1
);
1541 i
.insert_before(t2
);
1543 i
.insert_before(assign(lo
, mul(src1l
, src2l
)));
1544 i
.insert_before(assign(t1
, mul(src1l
, src2h
)));
1545 i
.insert_before(assign(t2
, mul(src1h
, src2l
)));
1546 i
.insert_before(assign(hi
, mul(src1h
, src2h
)));
1548 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t1
, c16
->clone(ir
, NULL
))))));
1549 i
.insert_before(assign(lo
, add(lo
, lshift(t1
, c16
->clone(ir
, NULL
)))));
1551 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t2
, c16
->clone(ir
, NULL
))))));
1552 i
.insert_before(assign(lo
, add(lo
, lshift(t2
, c16
->clone(ir
, NULL
)))));
1554 if (different_signs
== NULL
) {
1555 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1557 ir
->operation
= ir_binop_add
;
1558 ir
->operands
[0] = add(hi
, rshift(t1
, c16
->clone(ir
, NULL
)));
1559 ir
->operands
[1] = rshift(t2
, c16
->clone(ir
, NULL
));
1561 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1563 i
.insert_before(assign(hi
, add(add(hi
, rshift(t1
, c16
->clone(ir
, NULL
))),
1564 rshift(t2
, c16
->clone(ir
, NULL
)))));
1566 /* For channels where different_signs is set we have to perform a 64-bit
1567 * negation. This is *not* the same as just negating the high 32-bits.
1568 * Consider -3 * 2. The high 32-bits is 0, but the desired result is
1569 * -1, not -0! Recall -x == ~x + 1.
1571 ir_variable
*neg_hi
=
1572 new(ir
) ir_variable(glsl_type::ivec(elements
), "neg_hi", ir_var_temporary
);
1573 ir_constant
*c1
= new(ir
) ir_constant(1u, elements
);
1575 i
.insert_before(neg_hi
);
1576 i
.insert_before(assign(neg_hi
, add(bit_not(u2i(hi
)),
1577 u2i(_carry(bit_not(lo
), c1
)))));
1579 ir
->operation
= ir_triop_csel
;
1580 ir
->operands
[0] = new(ir
) ir_dereference_variable(different_signs
);
1581 ir
->operands
[1] = new(ir
) ir_dereference_variable(neg_hi
);
1582 ir
->operands
[2] = u2i(hi
);
1587 lower_instructions_visitor::sqrt_to_abs_sqrt(ir_expression
*ir
)
1589 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_abs
, ir
->operands
[0]);
1590 this->progress
= true;
1594 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1596 switch (ir
->operation
) {
1598 if (ir
->operands
[0]->type
->is_double())
1599 double_dot_to_fma(ir
);
1602 if (ir
->operands
[0]->type
->is_double())
1606 if (lowering(SUB_TO_ADD_NEG
))
1611 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1612 int_div_to_mul_rcp(ir
);
1613 else if ((ir
->operands
[1]->type
->is_float() && lowering(FDIV_TO_MUL_RCP
)) ||
1614 (ir
->operands
[1]->type
->is_double() && lowering(DDIV_TO_MUL_RCP
)))
1619 if (lowering(EXP_TO_EXP2
))
1624 if (lowering(LOG_TO_LOG2
))
1629 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1634 if (lowering(POW_TO_EXP2
))
1638 case ir_binop_ldexp
:
1639 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1641 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1642 dldexp_to_arith(ir
);
1645 case ir_unop_frexp_exp
:
1646 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1647 dfrexp_exp_to_arith(ir
);
1650 case ir_unop_frexp_sig
:
1651 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1652 dfrexp_sig_to_arith(ir
);
1655 case ir_binop_carry
:
1656 if (lowering(CARRY_TO_ARITH
))
1660 case ir_binop_borrow
:
1661 if (lowering(BORROW_TO_ARITH
))
1662 borrow_to_arith(ir
);
1665 case ir_unop_saturate
:
1666 if (lowering(SAT_TO_CLAMP
))
1671 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1672 dtrunc_to_dfrac(ir
);
1676 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1681 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1682 dfloor_to_dfrac(ir
);
1685 case ir_unop_round_even
:
1686 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1687 dround_even_to_dfrac(ir
);
1691 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1695 case ir_unop_bit_count
:
1696 if (lowering(BIT_COUNT_TO_MATH
))
1697 bit_count_to_math(ir
);
1700 case ir_triop_bitfield_extract
:
1701 if (lowering(EXTRACT_TO_SHIFTS
))
1702 extract_to_shifts(ir
);
1705 case ir_quadop_bitfield_insert
:
1706 if (lowering(INSERT_TO_SHIFTS
))
1707 insert_to_shifts(ir
);
1710 case ir_unop_bitfield_reverse
:
1711 if (lowering(REVERSE_TO_SHIFTS
))
1712 reverse_to_shifts(ir
);
1715 case ir_unop_find_lsb
:
1716 if (lowering(FIND_LSB_TO_FLOAT_CAST
))
1717 find_lsb_to_float_cast(ir
);
1720 case ir_unop_find_msb
:
1721 if (lowering(FIND_MSB_TO_FLOAT_CAST
))
1722 find_msb_to_float_cast(ir
);
1725 case ir_binop_imul_high
:
1726 if (lowering(IMUL_HIGH_TO_MUL
))
1727 imul_high_to_mul(ir
);
1732 if (lowering(SQRT_TO_ABS_SQRT
))
1733 sqrt_to_abs_sqrt(ir
);
1737 return visit_continue
;
1740 return visit_continue
;