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
);
171 } /* anonymous namespace */
174 * Determine if a particular type of lowering should occur
176 #define lowering(x) (this->lower & x)
179 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
181 lower_instructions_visitor
v(what_to_lower
);
183 visit_list_elements(&v
, instructions
);
188 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
190 ir
->operation
= ir_binop_add
;
191 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
192 ir
->operands
[1], NULL
);
193 this->progress
= true;
197 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
199 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
201 /* New expression for the 1.0 / op1 */
203 expr
= new(ir
) ir_expression(ir_unop_rcp
,
204 ir
->operands
[1]->type
,
207 /* op0 / op1 -> op0 * (1.0 / op1) */
208 ir
->operation
= ir_binop_mul
;
209 ir
->operands
[1] = expr
;
211 this->progress
= true;
215 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
217 assert(ir
->operands
[1]->type
->is_integer());
219 /* Be careful with integer division -- we need to do it as a
220 * float and re-truncate, since rcp(n > 1) of an integer would
223 ir_rvalue
*op0
, *op1
;
224 const struct glsl_type
*vec_type
;
226 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
227 ir
->operands
[1]->type
->vector_elements
,
228 ir
->operands
[1]->type
->matrix_columns
);
230 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
231 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
233 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
235 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
237 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
238 ir
->operands
[0]->type
->vector_elements
,
239 ir
->operands
[0]->type
->matrix_columns
);
241 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
242 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
244 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
246 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
247 ir
->type
->vector_elements
,
248 ir
->type
->matrix_columns
);
250 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
252 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
253 ir
->operation
= ir_unop_f2i
;
254 ir
->operands
[0] = op0
;
256 ir
->operation
= ir_unop_i2u
;
257 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
259 ir
->operands
[1] = NULL
;
261 this->progress
= true;
265 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
267 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
269 ir
->operation
= ir_unop_exp2
;
270 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
271 ir
->operands
[0], log2_e
);
272 this->progress
= true;
276 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
278 ir_expression
*const log2_x
=
279 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
282 ir
->operation
= ir_unop_exp2
;
283 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
284 ir
->operands
[1], log2_x
);
285 ir
->operands
[1] = NULL
;
286 this->progress
= true;
290 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
292 ir
->operation
= ir_binop_mul
;
293 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
294 ir
->operands
[0], NULL
);
295 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
296 this->progress
= true;
300 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
302 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
304 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
306 this->base_ir
->insert_before(x
);
307 this->base_ir
->insert_before(y
);
309 ir_assignment
*const assign_x
=
310 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
311 ir
->operands
[0], NULL
);
312 ir_assignment
*const assign_y
=
313 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
314 ir
->operands
[1], NULL
);
316 this->base_ir
->insert_before(assign_x
);
317 this->base_ir
->insert_before(assign_y
);
319 ir_expression
*const div_expr
=
320 new(ir
) ir_expression(ir_binop_div
, x
->type
,
321 new(ir
) ir_dereference_variable(x
),
322 new(ir
) ir_dereference_variable(y
));
324 /* Don't generate new IR that would need to be lowered in an additional
327 if (lowering(DIV_TO_MUL_RCP
) && (ir
->type
->is_float() || ir
->type
->is_double()))
328 div_to_mul_rcp(div_expr
);
330 ir_expression
*const floor_expr
=
331 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
333 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
334 dfloor_to_dfrac(floor_expr
);
336 ir_expression
*const mul_expr
=
337 new(ir
) ir_expression(ir_binop_mul
,
338 new(ir
) ir_dereference_variable(y
),
341 ir
->operation
= ir_binop_sub
;
342 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
343 ir
->operands
[1] = mul_expr
;
344 this->progress
= true;
348 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
351 * ir_binop_ldexp x exp
354 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
355 * resulting_biased_exp = extracted_biased_exp + exp;
357 * if (resulting_biased_exp < 1 || x == 0.0f) {
358 * return copysign(0.0, x);
361 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
362 * lshift(i2u(resulting_biased_exp), exp_shift));
364 * which we can't actually implement as such, since the GLSL IR doesn't
365 * have vectorized if-statements. We actually implement it without branches
366 * using conditional-select:
368 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
369 * resulting_biased_exp = extracted_biased_exp + exp;
371 * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
372 * gequal(resulting_biased_exp, 1);
373 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
374 * resulting_biased_exp = csel(is_not_zero_or_underflow,
375 * resulting_biased_exp, 0);
377 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
378 * lshift(i2u(resulting_biased_exp), exp_shift));
381 const unsigned vec_elem
= ir
->type
->vector_elements
;
384 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
385 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
388 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
390 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
392 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
393 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
395 /* Temporary variables */
396 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
397 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
399 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
402 ir_variable
*extracted_biased_exp
=
403 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
404 ir_variable
*resulting_biased_exp
=
405 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
407 ir_variable
*is_not_zero_or_underflow
=
408 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
410 ir_instruction
&i
= *base_ir
;
412 /* Copy <x> and <exp> arguments. */
414 i
.insert_before(assign(x
, ir
->operands
[0]));
415 i
.insert_before(exp
);
416 i
.insert_before(assign(exp
, ir
->operands
[1]));
418 /* Extract the biased exponent from <x>. */
419 i
.insert_before(extracted_biased_exp
);
420 i
.insert_before(assign(extracted_biased_exp
,
421 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
423 i
.insert_before(resulting_biased_exp
);
424 i
.insert_before(assign(resulting_biased_exp
,
425 add(extracted_biased_exp
, exp
)));
427 /* Test if result is ±0.0, subnormal, or underflow by checking if the
428 * resulting biased exponent would be less than 0x1. If so, the result is
429 * 0.0 with the sign of x. (Actually, invert the conditions so that
430 * immediate values are the second arguments, which is better for i965)
432 i
.insert_before(zero_sign_x
);
433 i
.insert_before(assign(zero_sign_x
,
434 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
436 i
.insert_before(is_not_zero_or_underflow
);
437 i
.insert_before(assign(is_not_zero_or_underflow
,
438 logic_and(nequal(x
, new(ir
) ir_constant(0.0f
, vec_elem
)),
439 gequal(resulting_biased_exp
,
440 new(ir
) ir_constant(0x1, vec_elem
)))));
441 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
443 i
.insert_before(assign(resulting_biased_exp
,
444 csel(is_not_zero_or_underflow
,
445 resulting_biased_exp
, zeroi
)));
447 /* We could test for overflows by checking if the resulting biased exponent
448 * would be greater than 0xFE. Turns out we don't need to because the GLSL
451 * "If this product is too large to be represented in the
452 * floating-point type, the result is undefined."
455 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
456 ir
->operation
= ir_unop_bitcast_i2f
;
457 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
458 exp_shift_clone
, exp_width
);
459 ir
->operands
[1] = NULL
;
461 this->progress
= true;
465 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
467 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
468 * from the significand.
471 const unsigned vec_elem
= ir
->type
->vector_elements
;
474 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
475 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
478 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
480 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
482 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
483 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
484 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
486 /* Temporary variables */
487 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
488 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
490 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
493 ir_variable
*extracted_biased_exp
=
494 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
495 ir_variable
*resulting_biased_exp
=
496 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
498 ir_variable
*is_not_zero_or_underflow
=
499 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
501 ir_instruction
&i
= *base_ir
;
503 /* Copy <x> and <exp> arguments. */
505 i
.insert_before(assign(x
, ir
->operands
[0]));
506 i
.insert_before(exp
);
507 i
.insert_before(assign(exp
, ir
->operands
[1]));
509 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
510 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
511 dfrexp_exp_to_arith(frexp_exp
);
513 /* Extract the biased exponent from <x>. */
514 i
.insert_before(extracted_biased_exp
);
515 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
517 i
.insert_before(resulting_biased_exp
);
518 i
.insert_before(assign(resulting_biased_exp
,
519 add(extracted_biased_exp
, exp
)));
521 /* Test if result is ±0.0, subnormal, or underflow by checking if the
522 * resulting biased exponent would be less than 0x1. If so, the result is
523 * 0.0 with the sign of x. (Actually, invert the conditions so that
524 * immediate values are the second arguments, which is better for i965)
525 * TODO: Implement in a vector fashion.
527 i
.insert_before(zero_sign_x
);
528 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
529 ir_variable
*unpacked
=
530 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
531 i
.insert_before(unpacked
);
534 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
535 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
537 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
538 i
.insert_before(assign(zero_sign_x
,
539 expr(ir_unop_pack_double_2x32
, unpacked
),
542 i
.insert_before(is_not_zero_or_underflow
);
543 i
.insert_before(assign(is_not_zero_or_underflow
,
544 gequal(resulting_biased_exp
,
545 new(ir
) ir_constant(0x1, vec_elem
))));
546 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
548 i
.insert_before(assign(resulting_biased_exp
,
549 csel(is_not_zero_or_underflow
,
550 resulting_biased_exp
, zeroi
)));
552 /* We could test for overflows by checking if the resulting biased exponent
553 * would be greater than 0xFE. Turns out we don't need to because the GLSL
556 * "If this product is too large to be represented in the
557 * floating-point type, the result is undefined."
560 ir_rvalue
*results
[4] = {NULL
};
561 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
562 ir_variable
*unpacked
=
563 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
564 i
.insert_before(unpacked
);
567 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
569 ir_expression
*bfi
= bitfield_insert(
571 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
572 exp_shift
->clone(ir
, NULL
),
573 exp_width
->clone(ir
, NULL
));
575 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
577 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
580 ir
->operation
= ir_quadop_vector
;
581 ir
->operands
[0] = results
[0];
582 ir
->operands
[1] = results
[1];
583 ir
->operands
[2] = results
[2];
584 ir
->operands
[3] = results
[3];
586 /* Don't generate new IR that would need to be lowered in an additional
590 this->progress
= true;
594 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
596 const unsigned vec_elem
= ir
->type
->vector_elements
;
597 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
599 /* Double-precision floating-point values are stored as
604 * We're just extracting the significand here, so we only need to modify
605 * the upper 32-bit uint. Unfortunately we must extract each double
606 * independently as there is no vector version of unpackDouble.
609 ir_instruction
&i
= *base_ir
;
611 ir_variable
*is_not_zero
=
612 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
613 ir_rvalue
*results
[4] = {NULL
};
615 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
616 i
.insert_before(is_not_zero
);
619 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
621 /* TODO: Remake this as more vector-friendly when int64 support is
624 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
625 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
626 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
628 /* Exponent of double floating-point values in the range [0.5, 1.0). */
629 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
632 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
633 ir_variable
*unpacked
=
634 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
636 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
638 i
.insert_before(bits
);
639 i
.insert_before(unpacked
);
640 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
642 /* Manipulate the high uint to remove the exponent and replace it with
643 * either the default exponent or zero.
645 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
646 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
647 i
.insert_before(assign(bits
, bit_or(bits
,
648 csel(swizzle(is_not_zero
, elem
, 1),
651 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
652 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
655 /* Put the dvec back together */
656 ir
->operation
= ir_quadop_vector
;
657 ir
->operands
[0] = results
[0];
658 ir
->operands
[1] = results
[1];
659 ir
->operands
[2] = results
[2];
660 ir
->operands
[3] = results
[3];
662 this->progress
= true;
666 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
668 const unsigned vec_elem
= ir
->type
->vector_elements
;
669 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
670 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
672 /* Double-precision floating-point values are stored as
677 * We're just extracting the exponent here, so we only care about the upper
681 ir_instruction
&i
= *base_ir
;
683 ir_variable
*is_not_zero
=
684 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
685 ir_variable
*high_words
=
686 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
687 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
688 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
690 ir_rvalue
*absval
= abs(ir
->operands
[0]);
692 i
.insert_before(is_not_zero
);
693 i
.insert_before(high_words
);
694 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
696 /* Extract all of the upper uints. */
697 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
698 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
700 i
.insert_before(assign(high_words
,
701 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
705 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
706 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
708 /* For non-zero inputs, shift the exponent down and apply bias. */
709 ir
->operation
= ir_triop_csel
;
710 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
711 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
712 ir
->operands
[2] = izero
;
714 this->progress
= true;
718 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
723 * sum = ir_binop_add x y
724 * bcarry = ir_binop_less sum x
725 * carry = ir_unop_b2i bcarry
728 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
729 ir
->operation
= ir_unop_i2u
;
730 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
731 ir
->operands
[1] = NULL
;
733 this->progress
= true;
737 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
740 * ir_binop_borrow x y
742 * bcarry = ir_binop_less x y
743 * carry = ir_unop_b2i bcarry
746 ir
->operation
= ir_unop_i2u
;
747 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
748 ir
->operands
[1] = NULL
;
750 this->progress
= true;
754 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
759 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
762 ir
->operation
= ir_binop_min
;
763 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
765 new(ir
) ir_constant(0.0f
));
766 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
768 this->progress
= true;
772 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
774 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
776 this->base_ir
->insert_before(temp
);
778 int nc
= ir
->operands
[0]->type
->components();
779 for (int i
= nc
- 1; i
>= 1; i
--) {
780 ir_assignment
*assig
;
782 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
783 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
785 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
786 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
789 this->base_ir
->insert_before(assig
);
792 ir
->operation
= ir_triop_fma
;
793 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
794 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
795 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
797 this->progress
= true;
802 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
805 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
806 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
808 switch (op2
->type
->vector_elements
) {
810 swizval
= SWIZZLE_XXXX
;
813 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
814 swizval
= SWIZZLE_XYZW
;
818 ir
->operation
= ir_triop_fma
;
819 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
820 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
822 this->progress
= true;
826 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
830 * temp = sub(x, frtemp);
831 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
833 ir_instruction
&i
= *base_ir
;
834 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
835 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
836 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
839 i
.insert_before(frtemp
);
840 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
842 ir
->operation
= ir_binop_add
;
843 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
844 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
846 this->progress
= true;
850 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
854 * result = sub(x, frtemp);
856 ir
->operation
= ir_binop_sub
;
857 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
859 this->progress
= true;
862 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
867 * frtemp = frac(temp);
868 * t2 = sub(temp, frtemp);
869 * if (frac(x) == 0.5)
870 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
875 ir_instruction
&i
= *base_ir
;
876 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
878 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
880 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
882 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
883 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
884 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
886 i
.insert_before(temp
);
887 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
889 i
.insert_before(frtemp
);
890 i
.insert_before(assign(frtemp
, fract(temp
)));
893 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
895 ir
->operation
= ir_triop_csel
;
896 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
897 p5
->clone(ir
, NULL
));
898 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
902 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
904 this->progress
= true;
908 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
912 * temp = sub(x, frtemp);
913 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
915 ir_rvalue
*arg
= ir
->operands
[0];
916 ir_instruction
&i
= *base_ir
;
918 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
919 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
920 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
922 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
925 i
.insert_before(frtemp
);
926 i
.insert_before(assign(frtemp
, fract(arg
)));
927 i
.insert_before(temp
);
928 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
930 ir
->operation
= ir_triop_csel
;
931 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
932 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
933 ir
->operands
[2] = add(temp
,
934 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
935 zero
->clone(ir
, NULL
),
938 this->progress
= true;
942 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
945 * temp = x > 0.0 ? 1.0 : 0.0;
946 * result = x < 0.0 ? -1.0 : temp;
948 ir_rvalue
*arg
= ir
->operands
[0];
949 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
950 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
951 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
953 ir
->operation
= ir_triop_csel
;
954 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
955 zero
->clone(ir
, NULL
));
956 ir
->operands
[1] = neg_one
;
957 ir
->operands
[2] = csel(greater(arg
, zero
),
959 zero
->clone(ir
, NULL
));
961 this->progress
= true;
965 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
967 /* For more details, see:
969 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
971 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
972 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
974 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
975 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
976 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
977 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
978 ir_constant
*c1
= new(ir
) ir_constant(1u);
979 ir_constant
*c2
= new(ir
) ir_constant(2u);
980 ir_constant
*c4
= new(ir
) ir_constant(4u);
981 ir_constant
*c24
= new(ir
) ir_constant(24u);
983 base_ir
->insert_before(temp
);
985 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
986 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
988 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
989 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
992 /* temp = temp - ((temp >> 1) & 0x55555555u); */
993 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
996 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
997 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
998 bit_and(rshift(temp
, c2
),
999 c33333333
->clone(ir
, NULL
)))));
1001 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1002 ir
->operation
= ir_unop_u2i
;
1003 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1007 this->progress
= true;
1011 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1014 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1016 base_ir
->insert_before(bits
);
1017 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1019 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1021 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1023 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1024 ir_constant
*cFFFFFFFF
=
1025 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1027 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1028 * we'd get a mask of 0 when bits is 32. Special case it.
1030 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1032 ir_expression
*mask
= csel(equal(bits
, c32
),
1034 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1036 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1038 * If bits is zero, the result will be zero.
1040 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1041 * select as in the signed integer case.
1043 * (value >> offset) & mask;
1045 ir
->operation
= ir_binop_bit_and
;
1046 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1047 ir
->operands
[1] = mask
;
1048 ir
->operands
[2] = NULL
;
1051 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1053 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1055 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1057 /* temp = 32 - bits; */
1058 base_ir
->insert_before(temp
);
1059 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1061 /* expr = value << (temp - offset)) >> temp; */
1062 ir_expression
*expr
=
1063 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1065 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1067 * If bits is zero, the result will be zero.
1069 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1070 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1073 * (bits == 0) ? 0 : e;
1075 ir
->operation
= ir_triop_csel
;
1076 ir
->operands
[0] = equal(c0
, bits
);
1077 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1078 ir
->operands
[2] = expr
;
1081 this->progress
= true;
1085 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1089 ir_constant
*cFFFFFFFF
;
1090 ir_variable
*offset
=
1091 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1093 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1095 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1097 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1098 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1099 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1100 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1102 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1104 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1105 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1106 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1109 base_ir
->insert_before(offset
);
1110 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1112 base_ir
->insert_before(bits
);
1113 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1115 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1116 * we'd get a mask of 0 when bits is 32. Special case it.
1118 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1120 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1122 * The result will be undefined if offset or bits is negative, or if the
1123 * sum of offset and bits is greater than the number of bits used to
1124 * store the operand.
1126 * Since it's undefined, there are a couple other ways this could be
1127 * implemented. The other way that was considered was to put the csel
1128 * around the whole thing:
1130 * final_result = bits == 32 ? insert : ... ;
1132 base_ir
->insert_before(mask
);
1134 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1136 lshift(sub(lshift(c1
, bits
),
1137 c1
->clone(ir
, NULL
)),
1140 /* (base & ~mask) | ((insert << offset) & mask) */
1141 ir
->operation
= ir_binop_bit_or
;
1142 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1143 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1144 ir
->operands
[2] = NULL
;
1145 ir
->operands
[3] = NULL
;
1147 this->progress
= true;
1151 lower_instructions_visitor::reverse_to_shifts(ir_expression
*ir
)
1153 /* For more details, see:
1155 * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1158 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1160 new(ir
) ir_constant(2u, ir
->operands
[0]->type
->vector_elements
);
1162 new(ir
) ir_constant(4u, ir
->operands
[0]->type
->vector_elements
);
1164 new(ir
) ir_constant(8u, ir
->operands
[0]->type
->vector_elements
);
1166 new(ir
) ir_constant(16u, ir
->operands
[0]->type
->vector_elements
);
1167 ir_constant
*c33333333
=
1168 new(ir
) ir_constant(0x33333333u
, ir
->operands
[0]->type
->vector_elements
);
1169 ir_constant
*c55555555
=
1170 new(ir
) ir_constant(0x55555555u
, ir
->operands
[0]->type
->vector_elements
);
1171 ir_constant
*c0F0F0F0F
=
1172 new(ir
) ir_constant(0x0F0F0F0Fu
, ir
->operands
[0]->type
->vector_elements
);
1173 ir_constant
*c00FF00FF
=
1174 new(ir
) ir_constant(0x00FF00FFu
, ir
->operands
[0]->type
->vector_elements
);
1176 new(ir
) ir_variable(glsl_type::uvec(ir
->operands
[0]->type
->vector_elements
),
1177 "temp", ir_var_temporary
);
1178 ir_instruction
&i
= *base_ir
;
1180 i
.insert_before(temp
);
1182 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1183 i
.insert_before(assign(temp
, ir
->operands
[0]));
1185 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1186 i
.insert_before(assign(temp
, i2u(ir
->operands
[0])));
1189 /* Swap odd and even bits.
1191 * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1193 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c1
), c55555555
),
1194 lshift(bit_and(temp
, c55555555
->clone(ir
, NULL
)),
1195 c1
->clone(ir
, NULL
)))));
1196 /* Swap consecutive pairs.
1198 * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1200 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c2
), c33333333
),
1201 lshift(bit_and(temp
, c33333333
->clone(ir
, NULL
)),
1202 c2
->clone(ir
, NULL
)))));
1206 * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1208 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c4
), c0F0F0F0F
),
1209 lshift(bit_and(temp
, c0F0F0F0F
->clone(ir
, NULL
)),
1210 c4
->clone(ir
, NULL
)))));
1212 /* The last step is, basically, bswap. Swap the bytes, then swap the
1213 * words. When this code is run through GCC on x86, it does generate a
1214 * bswap instruction.
1216 * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1217 * temp = ( temp >> 16 ) | ( temp << 16);
1219 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c8
), c00FF00FF
),
1220 lshift(bit_and(temp
, c00FF00FF
->clone(ir
, NULL
)),
1221 c8
->clone(ir
, NULL
)))));
1223 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1224 ir
->operation
= ir_binop_bit_or
;
1225 ir
->operands
[0] = rshift(temp
, c16
);
1226 ir
->operands
[1] = lshift(temp
, c16
->clone(ir
, NULL
));
1228 ir
->operation
= ir_unop_u2i
;
1229 ir
->operands
[0] = bit_or(rshift(temp
, c16
),
1230 lshift(temp
, c16
->clone(ir
, NULL
)));
1233 this->progress
= true;
1237 lower_instructions_visitor::find_lsb_to_float_cast(ir_expression
*ir
)
1239 /* For more details, see:
1241 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1243 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1244 ir_constant
*c0
= new(ir
) ir_constant(unsigned(0), elements
);
1245 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1246 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1247 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1249 new(ir
) ir_variable(glsl_type::ivec(elements
), "temp", ir_var_temporary
);
1250 ir_variable
*lsb_only
=
1251 new(ir
) ir_variable(glsl_type::uvec(elements
), "lsb_only", ir_var_temporary
);
1252 ir_variable
*as_float
=
1253 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1255 new(ir
) ir_variable(glsl_type::ivec(elements
), "lsb", ir_var_temporary
);
1257 ir_instruction
&i
= *base_ir
;
1259 i
.insert_before(temp
);
1261 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1262 i
.insert_before(assign(temp
, ir
->operands
[0]));
1264 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1265 i
.insert_before(assign(temp
, u2i(ir
->operands
[0])));
1268 /* The int-to-float conversion is lossless because (value & -value) is
1269 * either a power of two or zero. We don't use the result in the zero
1270 * case. The uint() cast is necessary so that 0x80000000 does not
1271 * generate a negative value.
1273 * uint lsb_only = uint(value & -value);
1274 * float as_float = float(lsb_only);
1276 i
.insert_before(lsb_only
);
1277 i
.insert_before(assign(lsb_only
, i2u(bit_and(temp
, neg(temp
)))));
1279 i
.insert_before(as_float
);
1280 i
.insert_before(assign(as_float
, u2f(lsb_only
)));
1282 /* This is basically an open-coded frexp. Implementations that have a
1283 * native frexp instruction would be better served by that. This is
1284 * optimized versus a full-featured open-coded implementation in two ways:
1286 * - We don't care about a correct result from subnormal numbers (including
1287 * 0.0), so the raw exponent can always be safely unbiased.
1289 * - The value cannot be negative, so it does not need to be masked off to
1290 * extract the exponent.
1292 * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1294 i
.insert_before(lsb
);
1295 i
.insert_before(assign(lsb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1297 /* Use lsb_only in the comparison instead of temp so that the & (far above)
1298 * can possibly generate the result without an explicit comparison.
1300 * (lsb_only == 0) ? -1 : lsb;
1302 * Since our input values are all integers, the unbiased exponent must not
1303 * be negative. It will only be negative (-0x7f, in fact) if lsb_only is
1304 * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is
1305 * better is likely GPU dependent. Either way, the difference should be
1308 ir
->operation
= ir_triop_csel
;
1309 ir
->operands
[0] = equal(lsb_only
, c0
);
1310 ir
->operands
[1] = cminus1
;
1311 ir
->operands
[2] = new(ir
) ir_dereference_variable(lsb
);
1313 this->progress
= true;
1317 lower_instructions_visitor::find_msb_to_float_cast(ir_expression
*ir
)
1319 /* For more details, see:
1321 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1323 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1324 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1325 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1326 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1327 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1328 ir_constant
*c000000FF
= new(ir
) ir_constant(0x000000FFu
, elements
);
1329 ir_constant
*cFFFFFF00
= new(ir
) ir_constant(0xFFFFFF00u
, elements
);
1331 new(ir
) ir_variable(glsl_type::uvec(elements
), "temp", ir_var_temporary
);
1332 ir_variable
*as_float
=
1333 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1335 new(ir
) ir_variable(glsl_type::ivec(elements
), "msb", ir_var_temporary
);
1337 ir_instruction
&i
= *base_ir
;
1339 i
.insert_before(temp
);
1341 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1342 i
.insert_before(assign(temp
, ir
->operands
[0]));
1344 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1346 /* findMSB(uint(abs(some_int))) almost always does the right thing.
1347 * There are two problem values:
1349 * * 0x80000000. Since abs(0x80000000) == 0x80000000, findMSB returns
1350 * 31. However, findMSB(int(0x80000000)) == 30.
1352 * * 0xffffffff. Since abs(0xffffffff) == 1, findMSB returns
1353 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1355 * For a value of zero or negative one, -1 will be returned.
1357 * For all negative number cases, including 0x80000000 and 0xffffffff,
1358 * the correct value is obtained from findMSB if instead of negating the
1359 * (already negative) value the logical-not is used. A conditonal
1360 * logical-not can be achieved in two instructions.
1362 ir_variable
*as_int
=
1363 new(ir
) ir_variable(glsl_type::ivec(elements
), "as_int", ir_var_temporary
);
1364 ir_constant
*c31
= new(ir
) ir_constant(int(31), elements
);
1366 i
.insert_before(as_int
);
1367 i
.insert_before(assign(as_int
, ir
->operands
[0]));
1368 i
.insert_before(assign(temp
, i2u(expr(ir_binop_bit_xor
,
1370 rshift(as_int
, c31
)))));
1373 /* The int-to-float conversion is lossless because bits are conditionally
1374 * masked off the bottom of temp to ensure the value has at most 24 bits of
1375 * data or is zero. We don't use the result in the zero case. The uint()
1376 * cast is necessary so that 0x80000000 does not generate a negative value.
1378 * float as_float = float(temp > 255 ? temp & ~255 : temp);
1380 i
.insert_before(as_float
);
1381 i
.insert_before(assign(as_float
, u2f(csel(greater(temp
, c000000FF
),
1382 bit_and(temp
, cFFFFFF00
),
1385 /* This is basically an open-coded frexp. Implementations that have a
1386 * native frexp instruction would be better served by that. This is
1387 * optimized versus a full-featured open-coded implementation in two ways:
1389 * - We don't care about a correct result from subnormal numbers (including
1390 * 0.0), so the raw exponent can always be safely unbiased.
1392 * - The value cannot be negative, so it does not need to be masked off to
1393 * extract the exponent.
1395 * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1397 i
.insert_before(msb
);
1398 i
.insert_before(assign(msb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1400 /* Use msb in the comparison instead of temp so that the subtract can
1401 * possibly generate the result without an explicit comparison.
1403 * (msb < 0) ? -1 : msb;
1405 * Since our input values are all integers, the unbiased exponent must not
1406 * be negative. It will only be negative (-0x7f, in fact) if temp is 0.
1408 ir
->operation
= ir_triop_csel
;
1409 ir
->operands
[0] = less(msb
, c0
);
1410 ir
->operands
[1] = cminus1
;
1411 ir
->operands
[2] = new(ir
) ir_dereference_variable(msb
);
1413 this->progress
= true;
1417 lower_instructions_visitor::imul_high_to_mul(ir_expression
*ir
)
1422 * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
1424 * In GLSL, (a * b) becomes
1426 * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu);
1427 * uint m2 = (a & 0x0000ffffu) * (b >> 16);
1428 * uint m3 = (a >> 16) * (b & 0x0000ffffu);
1429 * uint m4 = (a >> 16) * (b >> 16);
1436 * lo_result = uaddCarry(m1, m2 << 16, c1);
1437 * hi_result = m4 + c1;
1438 * lo_result = uaddCarry(lo_result, m3 << 16, c2);
1439 * hi_result = hi_result + c2;
1440 * hi_result = hi_result + (m2 >> 16) + (m3 >> 16);
1442 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1444 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1", ir_var_temporary
);
1445 ir_variable
*src1h
=
1446 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1h", ir_var_temporary
);
1447 ir_variable
*src1l
=
1448 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1l", ir_var_temporary
);
1450 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2", ir_var_temporary
);
1451 ir_variable
*src2h
=
1452 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2h", ir_var_temporary
);
1453 ir_variable
*src2l
=
1454 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2l", ir_var_temporary
);
1456 new(ir
) ir_variable(glsl_type::uvec(elements
), "t1", ir_var_temporary
);
1458 new(ir
) ir_variable(glsl_type::uvec(elements
), "t2", ir_var_temporary
);
1460 new(ir
) ir_variable(glsl_type::uvec(elements
), "lo", ir_var_temporary
);
1462 new(ir
) ir_variable(glsl_type::uvec(elements
), "hi", ir_var_temporary
);
1463 ir_variable
*different_signs
= NULL
;
1464 ir_constant
*c0000FFFF
= new(ir
) ir_constant(0x0000FFFFu
, elements
);
1465 ir_constant
*c16
= new(ir
) ir_constant(16u, elements
);
1467 ir_instruction
&i
= *base_ir
;
1469 i
.insert_before(src1
);
1470 i
.insert_before(src2
);
1471 i
.insert_before(src1h
);
1472 i
.insert_before(src2h
);
1473 i
.insert_before(src1l
);
1474 i
.insert_before(src2l
);
1476 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1477 i
.insert_before(assign(src1
, ir
->operands
[0]));
1478 i
.insert_before(assign(src2
, ir
->operands
[1]));
1480 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1482 ir_variable
*itmp1
=
1483 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp1", ir_var_temporary
);
1484 ir_variable
*itmp2
=
1485 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp2", ir_var_temporary
);
1486 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1488 i
.insert_before(itmp1
);
1489 i
.insert_before(itmp2
);
1490 i
.insert_before(assign(itmp1
, ir
->operands
[0]));
1491 i
.insert_before(assign(itmp2
, ir
->operands
[1]));
1494 new(ir
) ir_variable(glsl_type::bvec(elements
), "different_signs",
1497 i
.insert_before(different_signs
);
1498 i
.insert_before(assign(different_signs
, expr(ir_binop_logic_xor
,
1500 less(itmp2
, c0
->clone(ir
, NULL
)))));
1502 i
.insert_before(assign(src1
, i2u(abs(itmp1
))));
1503 i
.insert_before(assign(src2
, i2u(abs(itmp2
))));
1506 i
.insert_before(assign(src1l
, bit_and(src1
, c0000FFFF
)));
1507 i
.insert_before(assign(src2l
, bit_and(src2
, c0000FFFF
->clone(ir
, NULL
))));
1508 i
.insert_before(assign(src1h
, rshift(src1
, c16
)));
1509 i
.insert_before(assign(src2h
, rshift(src2
, c16
->clone(ir
, NULL
))));
1511 i
.insert_before(lo
);
1512 i
.insert_before(hi
);
1513 i
.insert_before(t1
);
1514 i
.insert_before(t2
);
1516 i
.insert_before(assign(lo
, mul(src1l
, src2l
)));
1517 i
.insert_before(assign(t1
, mul(src1l
, src2h
)));
1518 i
.insert_before(assign(t2
, mul(src1h
, src2l
)));
1519 i
.insert_before(assign(hi
, mul(src1h
, src2h
)));
1521 i
.insert_before(assign(hi
, add(hi
, carry(lo
, lshift(t1
, c16
->clone(ir
, NULL
))))));
1522 i
.insert_before(assign(lo
, add(lo
, lshift(t1
, c16
->clone(ir
, NULL
)))));
1524 i
.insert_before(assign(hi
, add(hi
, carry(lo
, lshift(t2
, c16
->clone(ir
, NULL
))))));
1525 i
.insert_before(assign(lo
, add(lo
, lshift(t2
, c16
->clone(ir
, NULL
)))));
1527 if (different_signs
== NULL
) {
1528 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1530 ir
->operation
= ir_binop_add
;
1531 ir
->operands
[0] = add(hi
, rshift(t1
, c16
->clone(ir
, NULL
)));
1532 ir
->operands
[1] = rshift(t2
, c16
->clone(ir
, NULL
));
1534 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1536 i
.insert_before(assign(hi
, add(add(hi
, rshift(t1
, c16
->clone(ir
, NULL
))),
1537 rshift(t2
, c16
->clone(ir
, NULL
)))));
1539 /* For channels where different_signs is set we have to perform a 64-bit
1540 * negation. This is *not* the same as just negating the high 32-bits.
1541 * Consider -3 * 2. The high 32-bits is 0, but the desired result is
1542 * -1, not -0! Recall -x == ~x + 1.
1544 ir_variable
*neg_hi
=
1545 new(ir
) ir_variable(glsl_type::ivec(elements
), "neg_hi", ir_var_temporary
);
1546 ir_constant
*c1
= new(ir
) ir_constant(1u, elements
);
1548 i
.insert_before(neg_hi
);
1549 i
.insert_before(assign(neg_hi
, add(bit_not(u2i(hi
)),
1550 u2i(carry(bit_not(lo
), c1
)))));
1552 ir
->operation
= ir_triop_csel
;
1553 ir
->operands
[0] = new(ir
) ir_dereference_variable(different_signs
);
1554 ir
->operands
[1] = new(ir
) ir_dereference_variable(neg_hi
);
1555 ir
->operands
[2] = u2i(hi
);
1560 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1562 switch (ir
->operation
) {
1564 if (ir
->operands
[0]->type
->is_double())
1565 double_dot_to_fma(ir
);
1568 if (ir
->operands
[0]->type
->is_double())
1572 if (lowering(SUB_TO_ADD_NEG
))
1577 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1578 int_div_to_mul_rcp(ir
);
1579 else if ((ir
->operands
[1]->type
->is_float() ||
1580 ir
->operands
[1]->type
->is_double()) && lowering(DIV_TO_MUL_RCP
))
1585 if (lowering(EXP_TO_EXP2
))
1590 if (lowering(LOG_TO_LOG2
))
1595 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1600 if (lowering(POW_TO_EXP2
))
1604 case ir_binop_ldexp
:
1605 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1607 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1608 dldexp_to_arith(ir
);
1611 case ir_unop_frexp_exp
:
1612 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1613 dfrexp_exp_to_arith(ir
);
1616 case ir_unop_frexp_sig
:
1617 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1618 dfrexp_sig_to_arith(ir
);
1621 case ir_binop_carry
:
1622 if (lowering(CARRY_TO_ARITH
))
1626 case ir_binop_borrow
:
1627 if (lowering(BORROW_TO_ARITH
))
1628 borrow_to_arith(ir
);
1631 case ir_unop_saturate
:
1632 if (lowering(SAT_TO_CLAMP
))
1637 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1638 dtrunc_to_dfrac(ir
);
1642 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1647 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1648 dfloor_to_dfrac(ir
);
1651 case ir_unop_round_even
:
1652 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1653 dround_even_to_dfrac(ir
);
1657 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1661 case ir_unop_bit_count
:
1662 if (lowering(BIT_COUNT_TO_MATH
))
1663 bit_count_to_math(ir
);
1666 case ir_triop_bitfield_extract
:
1667 if (lowering(EXTRACT_TO_SHIFTS
))
1668 extract_to_shifts(ir
);
1671 case ir_quadop_bitfield_insert
:
1672 if (lowering(INSERT_TO_SHIFTS
))
1673 insert_to_shifts(ir
);
1676 case ir_unop_bitfield_reverse
:
1677 if (lowering(REVERSE_TO_SHIFTS
))
1678 reverse_to_shifts(ir
);
1681 case ir_unop_find_lsb
:
1682 if (lowering(FIND_LSB_TO_FLOAT_CAST
))
1683 find_lsb_to_float_cast(ir
);
1686 case ir_unop_find_msb
:
1687 if (lowering(FIND_MSB_TO_FLOAT_CAST
))
1688 find_msb_to_float_cast(ir
);
1691 case ir_binop_imul_high
:
1692 if (lowering(IMUL_HIGH_TO_MUL
))
1693 imul_high_to_mul(ir
);
1697 return visit_continue
;
1700 return visit_continue
;