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
->init_num_operands();
197 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
198 ir
->operands
[1], NULL
);
199 this->progress
= true;
203 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
205 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
207 /* New expression for the 1.0 / op1 */
209 expr
= new(ir
) ir_expression(ir_unop_rcp
,
210 ir
->operands
[1]->type
,
213 /* op0 / op1 -> op0 * (1.0 / op1) */
214 ir
->operation
= ir_binop_mul
;
215 ir
->init_num_operands();
216 ir
->operands
[1] = expr
;
218 this->progress
= true;
222 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
224 assert(ir
->operands
[1]->type
->is_integer());
226 /* Be careful with integer division -- we need to do it as a
227 * float and re-truncate, since rcp(n > 1) of an integer would
230 ir_rvalue
*op0
, *op1
;
231 const struct glsl_type
*vec_type
;
233 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
234 ir
->operands
[1]->type
->vector_elements
,
235 ir
->operands
[1]->type
->matrix_columns
);
237 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
238 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
240 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
242 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
244 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
245 ir
->operands
[0]->type
->vector_elements
,
246 ir
->operands
[0]->type
->matrix_columns
);
248 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
249 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
251 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
253 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
254 ir
->type
->vector_elements
,
255 ir
->type
->matrix_columns
);
257 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
259 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
260 ir
->operation
= ir_unop_f2i
;
261 ir
->operands
[0] = op0
;
263 ir
->operation
= ir_unop_i2u
;
264 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
266 ir
->init_num_operands();
267 ir
->operands
[1] = NULL
;
269 this->progress
= true;
273 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
275 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
277 ir
->operation
= ir_unop_exp2
;
278 ir
->init_num_operands();
279 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
280 ir
->operands
[0], log2_e
);
281 this->progress
= true;
285 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
287 ir_expression
*const log2_x
=
288 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
291 ir
->operation
= ir_unop_exp2
;
292 ir
->init_num_operands();
293 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
294 ir
->operands
[1], log2_x
);
295 ir
->operands
[1] = NULL
;
296 this->progress
= true;
300 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
302 ir
->operation
= ir_binop_mul
;
303 ir
->init_num_operands();
304 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
305 ir
->operands
[0], NULL
);
306 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
307 this->progress
= true;
311 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
313 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
315 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
317 this->base_ir
->insert_before(x
);
318 this->base_ir
->insert_before(y
);
320 ir_assignment
*const assign_x
=
321 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
322 ir
->operands
[0], NULL
);
323 ir_assignment
*const assign_y
=
324 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
325 ir
->operands
[1], NULL
);
327 this->base_ir
->insert_before(assign_x
);
328 this->base_ir
->insert_before(assign_y
);
330 ir_expression
*const div_expr
=
331 new(ir
) ir_expression(ir_binop_div
, x
->type
,
332 new(ir
) ir_dereference_variable(x
),
333 new(ir
) ir_dereference_variable(y
));
335 /* Don't generate new IR that would need to be lowered in an additional
338 if ((lowering(FDIV_TO_MUL_RCP
) && ir
->type
->is_float()) ||
339 (lowering(DDIV_TO_MUL_RCP
) && ir
->type
->is_double()))
340 div_to_mul_rcp(div_expr
);
342 ir_expression
*const floor_expr
=
343 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
345 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
346 dfloor_to_dfrac(floor_expr
);
348 ir_expression
*const mul_expr
=
349 new(ir
) ir_expression(ir_binop_mul
,
350 new(ir
) ir_dereference_variable(y
),
353 ir
->operation
= ir_binop_sub
;
354 ir
->init_num_operands();
355 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
356 ir
->operands
[1] = mul_expr
;
357 this->progress
= true;
361 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
364 * ir_binop_ldexp x exp
367 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
368 * resulting_biased_exp = min(extracted_biased_exp + exp, 255);
370 * if (extracted_biased_exp >= 255)
371 * return x; // +/-inf, NaN
373 * sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
375 * if (min(resulting_biased_exp, extracted_biased_exp) < 1)
376 * resulting_biased_exp = 0;
377 * if (resulting_biased_exp >= 255 ||
378 * min(resulting_biased_exp, extracted_biased_exp) < 1) {
379 * sign_mantissa &= sign_mask;
382 * return bitcast_u2f(sign_mantissa |
383 * lshift(i2u(resulting_biased_exp), exp_shift));
385 * which we can't actually implement as such, since the GLSL IR doesn't
386 * have vectorized if-statements. We actually implement it without branches
387 * using conditional-select:
389 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
390 * resulting_biased_exp = min(extracted_biased_exp + exp, 255);
392 * sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
394 * flush_to_zero = lequal(min(resulting_biased_exp, extracted_biased_exp), 0);
395 * resulting_biased_exp = csel(flush_to_zero, 0, resulting_biased_exp)
396 * zero_mantissa = logic_or(flush_to_zero,
397 * gequal(resulting_biased_exp, 255));
398 * sign_mantissa = csel(zero_mantissa, sign_mantissa & sign_mask, sign_mantissa);
400 * result = sign_mantissa |
401 * lshift(i2u(resulting_biased_exp), exp_shift));
403 * return csel(extracted_biased_exp >= 255, x, bitcast_u2f(result));
405 * The definition of ldexp in the GLSL spec says:
407 * "If this product is too large to be represented in the
408 * floating-point type, the result is undefined."
410 * However, the definition of ldexp in the GLSL ES spec does not contain
411 * this sentence, so we do need to handle overflow correctly.
413 * There is additional language limiting the defined range of exp, but this
414 * is merely to allow implementations that store 2^exp in a temporary
418 const unsigned vec_elem
= ir
->type
->vector_elements
;
421 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
422 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
423 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
425 /* Temporary variables */
426 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
427 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
428 ir_variable
*result
= new(ir
) ir_variable(uvec
, "result", ir_var_temporary
);
430 ir_variable
*extracted_biased_exp
=
431 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
432 ir_variable
*resulting_biased_exp
=
433 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
435 ir_variable
*sign_mantissa
=
436 new(ir
) ir_variable(uvec
, "sign_mantissa", ir_var_temporary
);
438 ir_variable
*flush_to_zero
=
439 new(ir
) ir_variable(bvec
, "flush_to_zero", ir_var_temporary
);
440 ir_variable
*zero_mantissa
=
441 new(ir
) ir_variable(bvec
, "zero_mantissa", ir_var_temporary
);
443 ir_instruction
&i
= *base_ir
;
445 /* Copy <x> and <exp> arguments. */
447 i
.insert_before(assign(x
, ir
->operands
[0]));
448 i
.insert_before(exp
);
449 i
.insert_before(assign(exp
, ir
->operands
[1]));
451 /* Extract the biased exponent from <x>. */
452 i
.insert_before(extracted_biased_exp
);
453 i
.insert_before(assign(extracted_biased_exp
,
454 rshift(bitcast_f2i(abs(x
)),
455 new(ir
) ir_constant(23, vec_elem
))));
457 /* The definition of ldexp in the GLSL 4.60 spec says:
459 * "If exp is greater than +128 (single-precision) or +1024
460 * (double-precision), the value returned is undefined. If exp is less
461 * than -126 (single-precision) or -1022 (double-precision), the value
462 * returned may be flushed to zero."
464 * So we do not have to guard against the possibility of addition overflow,
465 * which could happen when exp is close to INT_MAX. Addition underflow
466 * cannot happen (the worst case is 0 + (-INT_MAX)).
468 i
.insert_before(resulting_biased_exp
);
469 i
.insert_before(assign(resulting_biased_exp
,
470 min2(add(extracted_biased_exp
, exp
),
471 new(ir
) ir_constant(255, vec_elem
))));
473 i
.insert_before(sign_mantissa
);
474 i
.insert_before(assign(sign_mantissa
,
475 bit_and(bitcast_f2u(x
),
476 new(ir
) ir_constant(0x807fffffu
, vec_elem
))));
478 /* We flush to zero if the original or resulting biased exponent is 0,
479 * indicating a +/-0.0 or subnormal input or output.
481 * The mantissa is set to 0 if the resulting biased exponent is 255, since
482 * an overflow should produce a +/-inf result.
484 * Note that NaN inputs are handled separately.
486 i
.insert_before(flush_to_zero
);
487 i
.insert_before(assign(flush_to_zero
,
488 lequal(min2(resulting_biased_exp
,
489 extracted_biased_exp
),
490 ir_constant::zero(ir
, ivec
))));
491 i
.insert_before(assign(resulting_biased_exp
,
493 ir_constant::zero(ir
, ivec
),
494 resulting_biased_exp
)));
496 i
.insert_before(zero_mantissa
);
497 i
.insert_before(assign(zero_mantissa
,
498 logic_or(flush_to_zero
,
499 equal(resulting_biased_exp
,
500 new(ir
) ir_constant(255, vec_elem
)))));
501 i
.insert_before(assign(sign_mantissa
,
503 bit_and(sign_mantissa
,
504 new(ir
) ir_constant(0x80000000u
, vec_elem
)),
507 /* Don't generate new IR that would need to be lowered in an additional
510 i
.insert_before(result
);
511 if (!lowering(INSERT_TO_SHIFTS
)) {
512 i
.insert_before(assign(result
,
513 bitfield_insert(sign_mantissa
,
514 i2u(resulting_biased_exp
),
515 new(ir
) ir_constant(23u, vec_elem
),
516 new(ir
) ir_constant(8u, vec_elem
))));
518 i
.insert_before(assign(result
,
519 bit_or(sign_mantissa
,
520 lshift(i2u(resulting_biased_exp
),
521 new(ir
) ir_constant(23, vec_elem
)))));
524 ir
->operation
= ir_triop_csel
;
525 ir
->init_num_operands();
526 ir
->operands
[0] = gequal(extracted_biased_exp
,
527 new(ir
) ir_constant(255, vec_elem
));
528 ir
->operands
[1] = new(ir
) ir_dereference_variable(x
);
529 ir
->operands
[2] = bitcast_u2f(result
);
531 this->progress
= true;
535 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
537 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
538 * from the significand.
541 const unsigned vec_elem
= ir
->type
->vector_elements
;
544 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
545 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
548 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
550 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
552 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
553 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
554 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
556 /* Temporary variables */
557 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
558 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
560 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
563 ir_variable
*extracted_biased_exp
=
564 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
565 ir_variable
*resulting_biased_exp
=
566 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
568 ir_variable
*is_not_zero_or_underflow
=
569 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
571 ir_instruction
&i
= *base_ir
;
573 /* Copy <x> and <exp> arguments. */
575 i
.insert_before(assign(x
, ir
->operands
[0]));
576 i
.insert_before(exp
);
577 i
.insert_before(assign(exp
, ir
->operands
[1]));
579 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
580 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
581 dfrexp_exp_to_arith(frexp_exp
);
583 /* Extract the biased exponent from <x>. */
584 i
.insert_before(extracted_biased_exp
);
585 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
587 i
.insert_before(resulting_biased_exp
);
588 i
.insert_before(assign(resulting_biased_exp
,
589 add(extracted_biased_exp
, exp
)));
591 /* Test if result is ±0.0, subnormal, or underflow by checking if the
592 * resulting biased exponent would be less than 0x1. If so, the result is
593 * 0.0 with the sign of x. (Actually, invert the conditions so that
594 * immediate values are the second arguments, which is better for i965)
595 * TODO: Implement in a vector fashion.
597 i
.insert_before(zero_sign_x
);
598 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
599 ir_variable
*unpacked
=
600 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
601 i
.insert_before(unpacked
);
604 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
605 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
607 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
608 i
.insert_before(assign(zero_sign_x
,
609 expr(ir_unop_pack_double_2x32
, unpacked
),
612 i
.insert_before(is_not_zero_or_underflow
);
613 i
.insert_before(assign(is_not_zero_or_underflow
,
614 gequal(resulting_biased_exp
,
615 new(ir
) ir_constant(0x1, vec_elem
))));
616 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
618 i
.insert_before(assign(resulting_biased_exp
,
619 csel(is_not_zero_or_underflow
,
620 resulting_biased_exp
, zeroi
)));
622 /* We could test for overflows by checking if the resulting biased exponent
623 * would be greater than 0xFE. Turns out we don't need to because the GLSL
626 * "If this product is too large to be represented in the
627 * floating-point type, the result is undefined."
630 ir_rvalue
*results
[4] = {NULL
};
631 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
632 ir_variable
*unpacked
=
633 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
634 i
.insert_before(unpacked
);
637 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
639 ir_expression
*bfi
= bitfield_insert(
641 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
642 exp_shift
->clone(ir
, NULL
),
643 exp_width
->clone(ir
, NULL
));
645 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
647 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
650 ir
->operation
= ir_quadop_vector
;
651 ir
->init_num_operands();
652 ir
->operands
[0] = results
[0];
653 ir
->operands
[1] = results
[1];
654 ir
->operands
[2] = results
[2];
655 ir
->operands
[3] = results
[3];
657 /* Don't generate new IR that would need to be lowered in an additional
661 this->progress
= true;
665 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
667 const unsigned vec_elem
= ir
->type
->vector_elements
;
668 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
670 /* Double-precision floating-point values are stored as
675 * We're just extracting the significand here, so we only need to modify
676 * the upper 32-bit uint. Unfortunately we must extract each double
677 * independently as there is no vector version of unpackDouble.
680 ir_instruction
&i
= *base_ir
;
682 ir_variable
*is_not_zero
=
683 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
684 ir_rvalue
*results
[4] = {NULL
};
686 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
687 i
.insert_before(is_not_zero
);
690 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
692 /* TODO: Remake this as more vector-friendly when int64 support is
695 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
696 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
697 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
699 /* Exponent of double floating-point values in the range [0.5, 1.0). */
700 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
703 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
704 ir_variable
*unpacked
=
705 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
707 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
709 i
.insert_before(bits
);
710 i
.insert_before(unpacked
);
711 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
713 /* Manipulate the high uint to remove the exponent and replace it with
714 * either the default exponent or zero.
716 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
717 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
718 i
.insert_before(assign(bits
, bit_or(bits
,
719 csel(swizzle(is_not_zero
, elem
, 1),
722 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
723 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
726 /* Put the dvec back together */
727 ir
->operation
= ir_quadop_vector
;
728 ir
->init_num_operands();
729 ir
->operands
[0] = results
[0];
730 ir
->operands
[1] = results
[1];
731 ir
->operands
[2] = results
[2];
732 ir
->operands
[3] = results
[3];
734 this->progress
= true;
738 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
740 const unsigned vec_elem
= ir
->type
->vector_elements
;
741 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
742 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
744 /* Double-precision floating-point values are stored as
749 * We're just extracting the exponent here, so we only care about the upper
753 ir_instruction
&i
= *base_ir
;
755 ir_variable
*is_not_zero
=
756 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
757 ir_variable
*high_words
=
758 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
759 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
760 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
762 ir_rvalue
*absval
= abs(ir
->operands
[0]);
764 i
.insert_before(is_not_zero
);
765 i
.insert_before(high_words
);
766 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
768 /* Extract all of the upper uints. */
769 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
770 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
772 i
.insert_before(assign(high_words
,
773 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
777 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
778 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
780 /* For non-zero inputs, shift the exponent down and apply bias. */
781 ir
->operation
= ir_triop_csel
;
782 ir
->init_num_operands();
783 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
784 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
785 ir
->operands
[2] = izero
;
787 this->progress
= true;
791 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
796 * sum = ir_binop_add x y
797 * bcarry = ir_binop_less sum x
798 * carry = ir_unop_b2i bcarry
801 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
802 ir
->operation
= ir_unop_i2u
;
803 ir
->init_num_operands();
804 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
805 ir
->operands
[1] = NULL
;
807 this->progress
= true;
811 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
814 * ir_binop_borrow x y
816 * bcarry = ir_binop_less x y
817 * carry = ir_unop_b2i bcarry
820 ir
->operation
= ir_unop_i2u
;
821 ir
->init_num_operands();
822 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
823 ir
->operands
[1] = NULL
;
825 this->progress
= true;
829 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
834 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
837 ir
->operation
= ir_binop_min
;
838 ir
->init_num_operands();
839 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
841 new(ir
) ir_constant(0.0f
));
842 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
844 this->progress
= true;
848 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
850 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
852 this->base_ir
->insert_before(temp
);
854 int nc
= ir
->operands
[0]->type
->components();
855 for (int i
= nc
- 1; i
>= 1; i
--) {
856 ir_assignment
*assig
;
858 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
859 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
861 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
862 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
865 this->base_ir
->insert_before(assig
);
868 ir
->operation
= ir_triop_fma
;
869 ir
->init_num_operands();
870 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
871 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
872 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
874 this->progress
= true;
879 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
882 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
883 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
885 switch (op2
->type
->vector_elements
) {
887 swizval
= SWIZZLE_XXXX
;
890 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
891 swizval
= SWIZZLE_XYZW
;
895 ir
->operation
= ir_triop_fma
;
896 ir
->init_num_operands();
897 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
898 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
900 this->progress
= true;
904 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
908 * temp = sub(x, frtemp);
909 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
911 ir_instruction
&i
= *base_ir
;
912 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
913 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
914 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
917 i
.insert_before(frtemp
);
918 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
920 ir
->operation
= ir_binop_add
;
921 ir
->init_num_operands();
922 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
923 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
925 this->progress
= true;
929 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
933 * result = sub(x, frtemp);
935 ir
->operation
= ir_binop_sub
;
936 ir
->init_num_operands();
937 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
939 this->progress
= true;
942 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
947 * frtemp = frac(temp);
948 * t2 = sub(temp, frtemp);
949 * if (frac(x) == 0.5)
950 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
955 ir_instruction
&i
= *base_ir
;
956 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
958 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
960 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
962 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
963 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
964 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
966 i
.insert_before(temp
);
967 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
969 i
.insert_before(frtemp
);
970 i
.insert_before(assign(frtemp
, fract(temp
)));
973 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
975 ir
->operation
= ir_triop_csel
;
976 ir
->init_num_operands();
977 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
978 p5
->clone(ir
, NULL
));
979 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
983 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
985 this->progress
= true;
989 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
993 * temp = sub(x, frtemp);
994 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
996 ir_rvalue
*arg
= ir
->operands
[0];
997 ir_instruction
&i
= *base_ir
;
999 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
1000 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
1001 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
1003 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
1006 i
.insert_before(frtemp
);
1007 i
.insert_before(assign(frtemp
, fract(arg
)));
1008 i
.insert_before(temp
);
1009 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
1011 ir
->operation
= ir_triop_csel
;
1012 ir
->init_num_operands();
1013 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
1014 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
1015 ir
->operands
[2] = add(temp
,
1016 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
1017 zero
->clone(ir
, NULL
),
1020 this->progress
= true;
1024 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
1027 * temp = x > 0.0 ? 1.0 : 0.0;
1028 * result = x < 0.0 ? -1.0 : temp;
1030 ir_rvalue
*arg
= ir
->operands
[0];
1031 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
1032 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
1033 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
1035 ir
->operation
= ir_triop_csel
;
1036 ir
->init_num_operands();
1037 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
1038 zero
->clone(ir
, NULL
));
1039 ir
->operands
[1] = neg_one
;
1040 ir
->operands
[2] = csel(greater(arg
, zero
),
1042 zero
->clone(ir
, NULL
));
1044 this->progress
= true;
1048 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
1050 /* For more details, see:
1052 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
1054 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1055 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
1057 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
1058 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
1059 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
1060 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
1061 ir_constant
*c1
= new(ir
) ir_constant(1u);
1062 ir_constant
*c2
= new(ir
) ir_constant(2u);
1063 ir_constant
*c4
= new(ir
) ir_constant(4u);
1064 ir_constant
*c24
= new(ir
) ir_constant(24u);
1066 base_ir
->insert_before(temp
);
1068 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1069 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
1071 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1072 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
1075 /* temp = temp - ((temp >> 1) & 0x55555555u); */
1076 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
1079 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
1080 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
1081 bit_and(rshift(temp
, c2
),
1082 c33333333
->clone(ir
, NULL
)))));
1084 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1085 ir
->operation
= ir_unop_u2i
;
1086 ir
->init_num_operands();
1087 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1091 this->progress
= true;
1095 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1098 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1100 base_ir
->insert_before(bits
);
1101 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1103 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1105 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1107 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1108 ir_constant
*cFFFFFFFF
=
1109 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1111 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1112 * we'd get a mask of 0 when bits is 32. Special case it.
1114 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1116 ir_expression
*mask
= csel(equal(bits
, c32
),
1118 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1120 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1122 * If bits is zero, the result will be zero.
1124 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1125 * select as in the signed integer case.
1127 * (value >> offset) & mask;
1129 ir
->operation
= ir_binop_bit_and
;
1130 ir
->init_num_operands();
1131 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1132 ir
->operands
[1] = mask
;
1133 ir
->operands
[2] = NULL
;
1136 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1138 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1140 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1142 /* temp = 32 - bits; */
1143 base_ir
->insert_before(temp
);
1144 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1146 /* expr = value << (temp - offset)) >> temp; */
1147 ir_expression
*expr
=
1148 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1150 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1152 * If bits is zero, the result will be zero.
1154 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1155 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1158 * (bits == 0) ? 0 : e;
1160 ir
->operation
= ir_triop_csel
;
1161 ir
->init_num_operands();
1162 ir
->operands
[0] = equal(c0
, bits
);
1163 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1164 ir
->operands
[2] = expr
;
1167 this->progress
= true;
1171 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1175 ir_constant
*cFFFFFFFF
;
1176 ir_variable
*offset
=
1177 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1179 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1181 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1183 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1184 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1185 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1186 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1188 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1190 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1191 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1192 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1195 base_ir
->insert_before(offset
);
1196 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1198 base_ir
->insert_before(bits
);
1199 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1201 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1202 * we'd get a mask of 0 when bits is 32. Special case it.
1204 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1206 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1208 * The result will be undefined if offset or bits is negative, or if the
1209 * sum of offset and bits is greater than the number of bits used to
1210 * store the operand.
1212 * Since it's undefined, there are a couple other ways this could be
1213 * implemented. The other way that was considered was to put the csel
1214 * around the whole thing:
1216 * final_result = bits == 32 ? insert : ... ;
1218 base_ir
->insert_before(mask
);
1220 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1222 lshift(sub(lshift(c1
, bits
),
1223 c1
->clone(ir
, NULL
)),
1226 /* (base & ~mask) | ((insert << offset) & mask) */
1227 ir
->operation
= ir_binop_bit_or
;
1228 ir
->init_num_operands();
1229 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1230 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1231 ir
->operands
[2] = NULL
;
1232 ir
->operands
[3] = NULL
;
1234 this->progress
= true;
1238 lower_instructions_visitor::reverse_to_shifts(ir_expression
*ir
)
1240 /* For more details, see:
1242 * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1245 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1247 new(ir
) ir_constant(2u, ir
->operands
[0]->type
->vector_elements
);
1249 new(ir
) ir_constant(4u, ir
->operands
[0]->type
->vector_elements
);
1251 new(ir
) ir_constant(8u, ir
->operands
[0]->type
->vector_elements
);
1253 new(ir
) ir_constant(16u, ir
->operands
[0]->type
->vector_elements
);
1254 ir_constant
*c33333333
=
1255 new(ir
) ir_constant(0x33333333u
, ir
->operands
[0]->type
->vector_elements
);
1256 ir_constant
*c55555555
=
1257 new(ir
) ir_constant(0x55555555u
, ir
->operands
[0]->type
->vector_elements
);
1258 ir_constant
*c0F0F0F0F
=
1259 new(ir
) ir_constant(0x0F0F0F0Fu
, ir
->operands
[0]->type
->vector_elements
);
1260 ir_constant
*c00FF00FF
=
1261 new(ir
) ir_constant(0x00FF00FFu
, ir
->operands
[0]->type
->vector_elements
);
1263 new(ir
) ir_variable(glsl_type::uvec(ir
->operands
[0]->type
->vector_elements
),
1264 "temp", ir_var_temporary
);
1265 ir_instruction
&i
= *base_ir
;
1267 i
.insert_before(temp
);
1269 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1270 i
.insert_before(assign(temp
, ir
->operands
[0]));
1272 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1273 i
.insert_before(assign(temp
, i2u(ir
->operands
[0])));
1276 /* Swap odd and even bits.
1278 * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1280 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c1
), c55555555
),
1281 lshift(bit_and(temp
, c55555555
->clone(ir
, NULL
)),
1282 c1
->clone(ir
, NULL
)))));
1283 /* Swap consecutive pairs.
1285 * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1287 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c2
), c33333333
),
1288 lshift(bit_and(temp
, c33333333
->clone(ir
, NULL
)),
1289 c2
->clone(ir
, NULL
)))));
1293 * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1295 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c4
), c0F0F0F0F
),
1296 lshift(bit_and(temp
, c0F0F0F0F
->clone(ir
, NULL
)),
1297 c4
->clone(ir
, NULL
)))));
1299 /* The last step is, basically, bswap. Swap the bytes, then swap the
1300 * words. When this code is run through GCC on x86, it does generate a
1301 * bswap instruction.
1303 * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1304 * temp = ( temp >> 16 ) | ( temp << 16);
1306 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c8
), c00FF00FF
),
1307 lshift(bit_and(temp
, c00FF00FF
->clone(ir
, NULL
)),
1308 c8
->clone(ir
, NULL
)))));
1310 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1311 ir
->operation
= ir_binop_bit_or
;
1312 ir
->init_num_operands();
1313 ir
->operands
[0] = rshift(temp
, c16
);
1314 ir
->operands
[1] = lshift(temp
, c16
->clone(ir
, NULL
));
1316 ir
->operation
= ir_unop_u2i
;
1317 ir
->init_num_operands();
1318 ir
->operands
[0] = bit_or(rshift(temp
, c16
),
1319 lshift(temp
, c16
->clone(ir
, NULL
)));
1322 this->progress
= true;
1326 lower_instructions_visitor::find_lsb_to_float_cast(ir_expression
*ir
)
1328 /* For more details, see:
1330 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1332 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1333 ir_constant
*c0
= new(ir
) ir_constant(unsigned(0), elements
);
1334 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1335 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1336 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1338 new(ir
) ir_variable(glsl_type::ivec(elements
), "temp", ir_var_temporary
);
1339 ir_variable
*lsb_only
=
1340 new(ir
) ir_variable(glsl_type::uvec(elements
), "lsb_only", ir_var_temporary
);
1341 ir_variable
*as_float
=
1342 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1344 new(ir
) ir_variable(glsl_type::ivec(elements
), "lsb", ir_var_temporary
);
1346 ir_instruction
&i
= *base_ir
;
1348 i
.insert_before(temp
);
1350 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1351 i
.insert_before(assign(temp
, ir
->operands
[0]));
1353 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1354 i
.insert_before(assign(temp
, u2i(ir
->operands
[0])));
1357 /* The int-to-float conversion is lossless because (value & -value) is
1358 * either a power of two or zero. We don't use the result in the zero
1359 * case. The uint() cast is necessary so that 0x80000000 does not
1360 * generate a negative value.
1362 * uint lsb_only = uint(value & -value);
1363 * float as_float = float(lsb_only);
1365 i
.insert_before(lsb_only
);
1366 i
.insert_before(assign(lsb_only
, i2u(bit_and(temp
, neg(temp
)))));
1368 i
.insert_before(as_float
);
1369 i
.insert_before(assign(as_float
, u2f(lsb_only
)));
1371 /* This is basically an open-coded frexp. Implementations that have a
1372 * native frexp instruction would be better served by that. This is
1373 * optimized versus a full-featured open-coded implementation in two ways:
1375 * - We don't care about a correct result from subnormal numbers (including
1376 * 0.0), so the raw exponent can always be safely unbiased.
1378 * - The value cannot be negative, so it does not need to be masked off to
1379 * extract the exponent.
1381 * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1383 i
.insert_before(lsb
);
1384 i
.insert_before(assign(lsb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1386 /* Use lsb_only in the comparison instead of temp so that the & (far above)
1387 * can possibly generate the result without an explicit comparison.
1389 * (lsb_only == 0) ? -1 : lsb;
1391 * Since our input values are all integers, the unbiased exponent must not
1392 * be negative. It will only be negative (-0x7f, in fact) if lsb_only is
1393 * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is
1394 * better is likely GPU dependent. Either way, the difference should be
1397 ir
->operation
= ir_triop_csel
;
1398 ir
->init_num_operands();
1399 ir
->operands
[0] = equal(lsb_only
, c0
);
1400 ir
->operands
[1] = cminus1
;
1401 ir
->operands
[2] = new(ir
) ir_dereference_variable(lsb
);
1403 this->progress
= true;
1407 lower_instructions_visitor::find_msb_to_float_cast(ir_expression
*ir
)
1409 /* For more details, see:
1411 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1413 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1414 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1415 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1416 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1417 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1418 ir_constant
*c000000FF
= new(ir
) ir_constant(0x000000FFu
, elements
);
1419 ir_constant
*cFFFFFF00
= new(ir
) ir_constant(0xFFFFFF00u
, elements
);
1421 new(ir
) ir_variable(glsl_type::uvec(elements
), "temp", ir_var_temporary
);
1422 ir_variable
*as_float
=
1423 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1425 new(ir
) ir_variable(glsl_type::ivec(elements
), "msb", ir_var_temporary
);
1427 ir_instruction
&i
= *base_ir
;
1429 i
.insert_before(temp
);
1431 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1432 i
.insert_before(assign(temp
, ir
->operands
[0]));
1434 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1436 /* findMSB(uint(abs(some_int))) almost always does the right thing.
1437 * There are two problem values:
1439 * * 0x80000000. Since abs(0x80000000) == 0x80000000, findMSB returns
1440 * 31. However, findMSB(int(0x80000000)) == 30.
1442 * * 0xffffffff. Since abs(0xffffffff) == 1, findMSB returns
1443 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1445 * For a value of zero or negative one, -1 will be returned.
1447 * For all negative number cases, including 0x80000000 and 0xffffffff,
1448 * the correct value is obtained from findMSB if instead of negating the
1449 * (already negative) value the logical-not is used. A conditonal
1450 * logical-not can be achieved in two instructions.
1452 ir_variable
*as_int
=
1453 new(ir
) ir_variable(glsl_type::ivec(elements
), "as_int", ir_var_temporary
);
1454 ir_constant
*c31
= new(ir
) ir_constant(int(31), elements
);
1456 i
.insert_before(as_int
);
1457 i
.insert_before(assign(as_int
, ir
->operands
[0]));
1458 i
.insert_before(assign(temp
, i2u(expr(ir_binop_bit_xor
,
1460 rshift(as_int
, c31
)))));
1463 /* The int-to-float conversion is lossless because bits are conditionally
1464 * masked off the bottom of temp to ensure the value has at most 24 bits of
1465 * data or is zero. We don't use the result in the zero case. The uint()
1466 * cast is necessary so that 0x80000000 does not generate a negative value.
1468 * float as_float = float(temp > 255 ? temp & ~255 : temp);
1470 i
.insert_before(as_float
);
1471 i
.insert_before(assign(as_float
, u2f(csel(greater(temp
, c000000FF
),
1472 bit_and(temp
, cFFFFFF00
),
1475 /* This is basically an open-coded frexp. Implementations that have a
1476 * native frexp instruction would be better served by that. This is
1477 * optimized versus a full-featured open-coded implementation in two ways:
1479 * - We don't care about a correct result from subnormal numbers (including
1480 * 0.0), so the raw exponent can always be safely unbiased.
1482 * - The value cannot be negative, so it does not need to be masked off to
1483 * extract the exponent.
1485 * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1487 i
.insert_before(msb
);
1488 i
.insert_before(assign(msb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1490 /* Use msb in the comparison instead of temp so that the subtract can
1491 * possibly generate the result without an explicit comparison.
1493 * (msb < 0) ? -1 : msb;
1495 * Since our input values are all integers, the unbiased exponent must not
1496 * be negative. It will only be negative (-0x7f, in fact) if temp is 0.
1498 ir
->operation
= ir_triop_csel
;
1499 ir
->init_num_operands();
1500 ir
->operands
[0] = less(msb
, c0
);
1501 ir
->operands
[1] = cminus1
;
1502 ir
->operands
[2] = new(ir
) ir_dereference_variable(msb
);
1504 this->progress
= true;
1508 lower_instructions_visitor::_carry(operand a
, operand b
)
1510 if (lowering(CARRY_TO_ARITH
))
1511 return i2u(b2i(less(add(a
, b
),
1512 a
.val
->clone(ralloc_parent(a
.val
), NULL
))));
1518 lower_instructions_visitor::imul_high_to_mul(ir_expression
*ir
)
1523 * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
1525 * In GLSL, (a * b) becomes
1527 * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu);
1528 * uint m2 = (a & 0x0000ffffu) * (b >> 16);
1529 * uint m3 = (a >> 16) * (b & 0x0000ffffu);
1530 * uint m4 = (a >> 16) * (b >> 16);
1537 * lo_result = uaddCarry(m1, m2 << 16, c1);
1538 * hi_result = m4 + c1;
1539 * lo_result = uaddCarry(lo_result, m3 << 16, c2);
1540 * hi_result = hi_result + c2;
1541 * hi_result = hi_result + (m2 >> 16) + (m3 >> 16);
1543 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1545 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1", ir_var_temporary
);
1546 ir_variable
*src1h
=
1547 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1h", ir_var_temporary
);
1548 ir_variable
*src1l
=
1549 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1l", ir_var_temporary
);
1551 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2", ir_var_temporary
);
1552 ir_variable
*src2h
=
1553 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2h", ir_var_temporary
);
1554 ir_variable
*src2l
=
1555 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2l", ir_var_temporary
);
1557 new(ir
) ir_variable(glsl_type::uvec(elements
), "t1", ir_var_temporary
);
1559 new(ir
) ir_variable(glsl_type::uvec(elements
), "t2", ir_var_temporary
);
1561 new(ir
) ir_variable(glsl_type::uvec(elements
), "lo", ir_var_temporary
);
1563 new(ir
) ir_variable(glsl_type::uvec(elements
), "hi", ir_var_temporary
);
1564 ir_variable
*different_signs
= NULL
;
1565 ir_constant
*c0000FFFF
= new(ir
) ir_constant(0x0000FFFFu
, elements
);
1566 ir_constant
*c16
= new(ir
) ir_constant(16u, elements
);
1568 ir_instruction
&i
= *base_ir
;
1570 i
.insert_before(src1
);
1571 i
.insert_before(src2
);
1572 i
.insert_before(src1h
);
1573 i
.insert_before(src2h
);
1574 i
.insert_before(src1l
);
1575 i
.insert_before(src2l
);
1577 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1578 i
.insert_before(assign(src1
, ir
->operands
[0]));
1579 i
.insert_before(assign(src2
, ir
->operands
[1]));
1581 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1583 ir_variable
*itmp1
=
1584 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp1", ir_var_temporary
);
1585 ir_variable
*itmp2
=
1586 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp2", ir_var_temporary
);
1587 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1589 i
.insert_before(itmp1
);
1590 i
.insert_before(itmp2
);
1591 i
.insert_before(assign(itmp1
, ir
->operands
[0]));
1592 i
.insert_before(assign(itmp2
, ir
->operands
[1]));
1595 new(ir
) ir_variable(glsl_type::bvec(elements
), "different_signs",
1598 i
.insert_before(different_signs
);
1599 i
.insert_before(assign(different_signs
, expr(ir_binop_logic_xor
,
1601 less(itmp2
, c0
->clone(ir
, NULL
)))));
1603 i
.insert_before(assign(src1
, i2u(abs(itmp1
))));
1604 i
.insert_before(assign(src2
, i2u(abs(itmp2
))));
1607 i
.insert_before(assign(src1l
, bit_and(src1
, c0000FFFF
)));
1608 i
.insert_before(assign(src2l
, bit_and(src2
, c0000FFFF
->clone(ir
, NULL
))));
1609 i
.insert_before(assign(src1h
, rshift(src1
, c16
)));
1610 i
.insert_before(assign(src2h
, rshift(src2
, c16
->clone(ir
, NULL
))));
1612 i
.insert_before(lo
);
1613 i
.insert_before(hi
);
1614 i
.insert_before(t1
);
1615 i
.insert_before(t2
);
1617 i
.insert_before(assign(lo
, mul(src1l
, src2l
)));
1618 i
.insert_before(assign(t1
, mul(src1l
, src2h
)));
1619 i
.insert_before(assign(t2
, mul(src1h
, src2l
)));
1620 i
.insert_before(assign(hi
, mul(src1h
, src2h
)));
1622 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t1
, c16
->clone(ir
, NULL
))))));
1623 i
.insert_before(assign(lo
, add(lo
, lshift(t1
, c16
->clone(ir
, NULL
)))));
1625 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t2
, c16
->clone(ir
, NULL
))))));
1626 i
.insert_before(assign(lo
, add(lo
, lshift(t2
, c16
->clone(ir
, NULL
)))));
1628 if (different_signs
== NULL
) {
1629 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1631 ir
->operation
= ir_binop_add
;
1632 ir
->init_num_operands();
1633 ir
->operands
[0] = add(hi
, rshift(t1
, c16
->clone(ir
, NULL
)));
1634 ir
->operands
[1] = rshift(t2
, c16
->clone(ir
, NULL
));
1636 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1638 i
.insert_before(assign(hi
, add(add(hi
, rshift(t1
, c16
->clone(ir
, NULL
))),
1639 rshift(t2
, c16
->clone(ir
, NULL
)))));
1641 /* For channels where different_signs is set we have to perform a 64-bit
1642 * negation. This is *not* the same as just negating the high 32-bits.
1643 * Consider -3 * 2. The high 32-bits is 0, but the desired result is
1644 * -1, not -0! Recall -x == ~x + 1.
1646 ir_variable
*neg_hi
=
1647 new(ir
) ir_variable(glsl_type::ivec(elements
), "neg_hi", ir_var_temporary
);
1648 ir_constant
*c1
= new(ir
) ir_constant(1u, elements
);
1650 i
.insert_before(neg_hi
);
1651 i
.insert_before(assign(neg_hi
, add(bit_not(u2i(hi
)),
1652 u2i(_carry(bit_not(lo
), c1
)))));
1654 ir
->operation
= ir_triop_csel
;
1655 ir
->init_num_operands();
1656 ir
->operands
[0] = new(ir
) ir_dereference_variable(different_signs
);
1657 ir
->operands
[1] = new(ir
) ir_dereference_variable(neg_hi
);
1658 ir
->operands
[2] = u2i(hi
);
1663 lower_instructions_visitor::sqrt_to_abs_sqrt(ir_expression
*ir
)
1665 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_abs
, ir
->operands
[0]);
1666 this->progress
= true;
1670 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1672 switch (ir
->operation
) {
1674 if (ir
->operands
[0]->type
->is_double())
1675 double_dot_to_fma(ir
);
1678 if (ir
->operands
[0]->type
->is_double())
1682 if (lowering(SUB_TO_ADD_NEG
))
1687 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1688 int_div_to_mul_rcp(ir
);
1689 else if ((ir
->operands
[1]->type
->is_float() && lowering(FDIV_TO_MUL_RCP
)) ||
1690 (ir
->operands
[1]->type
->is_double() && lowering(DDIV_TO_MUL_RCP
)))
1695 if (lowering(EXP_TO_EXP2
))
1700 if (lowering(LOG_TO_LOG2
))
1705 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1710 if (lowering(POW_TO_EXP2
))
1714 case ir_binop_ldexp
:
1715 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1717 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1718 dldexp_to_arith(ir
);
1721 case ir_unop_frexp_exp
:
1722 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1723 dfrexp_exp_to_arith(ir
);
1726 case ir_unop_frexp_sig
:
1727 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1728 dfrexp_sig_to_arith(ir
);
1731 case ir_binop_carry
:
1732 if (lowering(CARRY_TO_ARITH
))
1736 case ir_binop_borrow
:
1737 if (lowering(BORROW_TO_ARITH
))
1738 borrow_to_arith(ir
);
1741 case ir_unop_saturate
:
1742 if (lowering(SAT_TO_CLAMP
))
1747 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1748 dtrunc_to_dfrac(ir
);
1752 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1757 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1758 dfloor_to_dfrac(ir
);
1761 case ir_unop_round_even
:
1762 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1763 dround_even_to_dfrac(ir
);
1767 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1771 case ir_unop_bit_count
:
1772 if (lowering(BIT_COUNT_TO_MATH
))
1773 bit_count_to_math(ir
);
1776 case ir_triop_bitfield_extract
:
1777 if (lowering(EXTRACT_TO_SHIFTS
))
1778 extract_to_shifts(ir
);
1781 case ir_quadop_bitfield_insert
:
1782 if (lowering(INSERT_TO_SHIFTS
))
1783 insert_to_shifts(ir
);
1786 case ir_unop_bitfield_reverse
:
1787 if (lowering(REVERSE_TO_SHIFTS
))
1788 reverse_to_shifts(ir
);
1791 case ir_unop_find_lsb
:
1792 if (lowering(FIND_LSB_TO_FLOAT_CAST
))
1793 find_lsb_to_float_cast(ir
);
1796 case ir_unop_find_msb
:
1797 if (lowering(FIND_MSB_TO_FLOAT_CAST
))
1798 find_msb_to_float_cast(ir
);
1801 case ir_binop_imul_high
:
1802 if (lowering(IMUL_HIGH_TO_MUL
))
1803 imul_high_to_mul(ir
);
1808 if (lowering(SQRT_TO_ABS_SQRT
))
1809 sqrt_to_abs_sqrt(ir
);
1813 return visit_continue
;
1816 return visit_continue
;