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"
126 #include "util/half_float.h"
128 using namespace ir_builder
;
132 class lower_instructions_visitor
: public ir_hierarchical_visitor
{
134 lower_instructions_visitor(unsigned lower
)
135 : progress(false), lower(lower
) { }
137 ir_visitor_status
visit_leave(ir_expression
*);
142 unsigned lower
; /** Bitfield of which operations to lower */
144 void sub_to_add_neg(ir_expression
*);
145 void div_to_mul_rcp(ir_expression
*);
146 void int_div_to_mul_rcp(ir_expression
*);
147 void mod_to_floor(ir_expression
*);
148 void exp_to_exp2(ir_expression
*);
149 void pow_to_exp2(ir_expression
*);
150 void log_to_log2(ir_expression
*);
151 void ldexp_to_arith(ir_expression
*);
152 void dldexp_to_arith(ir_expression
*);
153 void dfrexp_sig_to_arith(ir_expression
*);
154 void dfrexp_exp_to_arith(ir_expression
*);
155 void carry_to_arith(ir_expression
*);
156 void borrow_to_arith(ir_expression
*);
157 void sat_to_clamp(ir_expression
*);
158 void double_dot_to_fma(ir_expression
*);
159 void double_lrp(ir_expression
*);
160 void dceil_to_dfrac(ir_expression
*);
161 void dfloor_to_dfrac(ir_expression
*);
162 void dround_even_to_dfrac(ir_expression
*);
163 void dtrunc_to_dfrac(ir_expression
*);
164 void dsign_to_csel(ir_expression
*);
165 void bit_count_to_math(ir_expression
*);
166 void extract_to_shifts(ir_expression
*);
167 void insert_to_shifts(ir_expression
*);
168 void reverse_to_shifts(ir_expression
*ir
);
169 void find_lsb_to_float_cast(ir_expression
*ir
);
170 void find_msb_to_float_cast(ir_expression
*ir
);
171 void imul_high_to_mul(ir_expression
*ir
);
172 void sqrt_to_abs_sqrt(ir_expression
*ir
);
173 void mul64_to_mul_and_mul_high(ir_expression
*ir
);
175 ir_expression
*_carry(operand a
, operand b
);
177 static ir_constant
*_imm_fp(void *mem_ctx
,
178 const glsl_type
*type
,
180 unsigned vector_elements
=1);
183 } /* anonymous namespace */
186 * Determine if a particular type of lowering should occur
188 #define lowering(x) (this->lower & x)
191 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
193 lower_instructions_visitor
v(what_to_lower
);
195 visit_list_elements(&v
, instructions
);
200 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
202 ir
->operation
= ir_binop_add
;
203 ir
->init_num_operands();
204 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
205 ir
->operands
[1], NULL
);
206 this->progress
= true;
210 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
212 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
214 /* New expression for the 1.0 / op1 */
216 expr
= new(ir
) ir_expression(ir_unop_rcp
,
217 ir
->operands
[1]->type
,
220 /* op0 / op1 -> op0 * (1.0 / op1) */
221 ir
->operation
= ir_binop_mul
;
222 ir
->init_num_operands();
223 ir
->operands
[1] = expr
;
225 this->progress
= true;
229 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
231 assert(ir
->operands
[1]->type
->is_integer_32());
233 /* Be careful with integer division -- we need to do it as a
234 * float and re-truncate, since rcp(n > 1) of an integer would
237 ir_rvalue
*op0
, *op1
;
238 const struct glsl_type
*vec_type
;
240 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
241 ir
->operands
[1]->type
->vector_elements
,
242 ir
->operands
[1]->type
->matrix_columns
);
244 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
245 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
247 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
249 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
251 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
252 ir
->operands
[0]->type
->vector_elements
,
253 ir
->operands
[0]->type
->matrix_columns
);
255 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
256 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
258 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
260 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
261 ir
->type
->vector_elements
,
262 ir
->type
->matrix_columns
);
264 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
266 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
267 ir
->operation
= ir_unop_f2i
;
268 ir
->operands
[0] = op0
;
270 ir
->operation
= ir_unop_i2u
;
271 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
273 ir
->init_num_operands();
274 ir
->operands
[1] = NULL
;
276 this->progress
= true;
280 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
282 ir_constant
*log2_e
= _imm_fp(ir
, ir
->type
, M_LOG2E
);
284 ir
->operation
= ir_unop_exp2
;
285 ir
->init_num_operands();
286 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
287 ir
->operands
[0], log2_e
);
288 this->progress
= true;
292 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
294 ir_expression
*const log2_x
=
295 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
298 ir
->operation
= ir_unop_exp2
;
299 ir
->init_num_operands();
300 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
301 ir
->operands
[1], log2_x
);
302 ir
->operands
[1] = NULL
;
303 this->progress
= true;
307 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
309 ir
->operation
= ir_binop_mul
;
310 ir
->init_num_operands();
311 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
312 ir
->operands
[0], NULL
);
313 ir
->operands
[1] = _imm_fp(ir
, ir
->operands
[0]->type
, 1.0 / M_LOG2E
);
314 this->progress
= true;
318 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
320 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
322 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
324 this->base_ir
->insert_before(x
);
325 this->base_ir
->insert_before(y
);
327 ir_assignment
*const assign_x
=
328 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
330 ir_assignment
*const assign_y
=
331 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
334 this->base_ir
->insert_before(assign_x
);
335 this->base_ir
->insert_before(assign_y
);
337 ir_expression
*const div_expr
=
338 new(ir
) ir_expression(ir_binop_div
, x
->type
,
339 new(ir
) ir_dereference_variable(x
),
340 new(ir
) ir_dereference_variable(y
));
342 /* Don't generate new IR that would need to be lowered in an additional
345 if ((lowering(FDIV_TO_MUL_RCP
) && ir
->type
->is_float()) ||
346 (lowering(DDIV_TO_MUL_RCP
) && ir
->type
->is_double()))
347 div_to_mul_rcp(div_expr
);
349 ir_expression
*const floor_expr
=
350 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
352 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
353 dfloor_to_dfrac(floor_expr
);
355 ir_expression
*const mul_expr
=
356 new(ir
) ir_expression(ir_binop_mul
,
357 new(ir
) ir_dereference_variable(y
),
360 ir
->operation
= ir_binop_sub
;
361 ir
->init_num_operands();
362 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
363 ir
->operands
[1] = mul_expr
;
364 this->progress
= true;
368 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
371 * ir_binop_ldexp x exp
374 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
375 * resulting_biased_exp = min(extracted_biased_exp + exp, 255);
377 * if (extracted_biased_exp >= 255)
378 * return x; // +/-inf, NaN
380 * sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
382 * if (min(resulting_biased_exp, extracted_biased_exp) < 1)
383 * resulting_biased_exp = 0;
384 * if (resulting_biased_exp >= 255 ||
385 * min(resulting_biased_exp, extracted_biased_exp) < 1) {
386 * sign_mantissa &= sign_mask;
389 * return bitcast_u2f(sign_mantissa |
390 * lshift(i2u(resulting_biased_exp), exp_shift));
392 * which we can't actually implement as such, since the GLSL IR doesn't
393 * have vectorized if-statements. We actually implement it without branches
394 * using conditional-select:
396 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
397 * resulting_biased_exp = min(extracted_biased_exp + exp, 255);
399 * sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
401 * flush_to_zero = lequal(min(resulting_biased_exp, extracted_biased_exp), 0);
402 * resulting_biased_exp = csel(flush_to_zero, 0, resulting_biased_exp)
403 * zero_mantissa = logic_or(flush_to_zero,
404 * gequal(resulting_biased_exp, 255));
405 * sign_mantissa = csel(zero_mantissa, sign_mantissa & sign_mask, sign_mantissa);
407 * result = sign_mantissa |
408 * lshift(i2u(resulting_biased_exp), exp_shift));
410 * return csel(extracted_biased_exp >= 255, x, bitcast_u2f(result));
412 * The definition of ldexp in the GLSL spec says:
414 * "If this product is too large to be represented in the
415 * floating-point type, the result is undefined."
417 * However, the definition of ldexp in the GLSL ES spec does not contain
418 * this sentence, so we do need to handle overflow correctly.
420 * There is additional language limiting the defined range of exp, but this
421 * is merely to allow implementations that store 2^exp in a temporary
425 const unsigned vec_elem
= ir
->type
->vector_elements
;
428 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
429 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
430 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
432 /* Temporary variables */
433 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
434 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
435 ir_variable
*result
= new(ir
) ir_variable(uvec
, "result", ir_var_temporary
);
437 ir_variable
*extracted_biased_exp
=
438 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
439 ir_variable
*resulting_biased_exp
=
440 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
442 ir_variable
*sign_mantissa
=
443 new(ir
) ir_variable(uvec
, "sign_mantissa", ir_var_temporary
);
445 ir_variable
*flush_to_zero
=
446 new(ir
) ir_variable(bvec
, "flush_to_zero", ir_var_temporary
);
447 ir_variable
*zero_mantissa
=
448 new(ir
) ir_variable(bvec
, "zero_mantissa", ir_var_temporary
);
450 ir_instruction
&i
= *base_ir
;
452 /* Copy <x> and <exp> arguments. */
454 i
.insert_before(assign(x
, ir
->operands
[0]));
455 i
.insert_before(exp
);
456 i
.insert_before(assign(exp
, ir
->operands
[1]));
458 /* Extract the biased exponent from <x>. */
459 i
.insert_before(extracted_biased_exp
);
460 i
.insert_before(assign(extracted_biased_exp
,
461 rshift(bitcast_f2i(abs(x
)),
462 new(ir
) ir_constant(23, vec_elem
))));
464 /* The definition of ldexp in the GLSL 4.60 spec says:
466 * "If exp is greater than +128 (single-precision) or +1024
467 * (double-precision), the value returned is undefined. If exp is less
468 * than -126 (single-precision) or -1022 (double-precision), the value
469 * returned may be flushed to zero."
471 * So we do not have to guard against the possibility of addition overflow,
472 * which could happen when exp is close to INT_MAX. Addition underflow
473 * cannot happen (the worst case is 0 + (-INT_MAX)).
475 i
.insert_before(resulting_biased_exp
);
476 i
.insert_before(assign(resulting_biased_exp
,
477 min2(add(extracted_biased_exp
, exp
),
478 new(ir
) ir_constant(255, vec_elem
))));
480 i
.insert_before(sign_mantissa
);
481 i
.insert_before(assign(sign_mantissa
,
482 bit_and(bitcast_f2u(x
),
483 new(ir
) ir_constant(0x807fffffu
, vec_elem
))));
485 /* We flush to zero if the original or resulting biased exponent is 0,
486 * indicating a +/-0.0 or subnormal input or output.
488 * The mantissa is set to 0 if the resulting biased exponent is 255, since
489 * an overflow should produce a +/-inf result.
491 * Note that NaN inputs are handled separately.
493 i
.insert_before(flush_to_zero
);
494 i
.insert_before(assign(flush_to_zero
,
495 lequal(min2(resulting_biased_exp
,
496 extracted_biased_exp
),
497 ir_constant::zero(ir
, ivec
))));
498 i
.insert_before(assign(resulting_biased_exp
,
500 ir_constant::zero(ir
, ivec
),
501 resulting_biased_exp
)));
503 i
.insert_before(zero_mantissa
);
504 i
.insert_before(assign(zero_mantissa
,
505 logic_or(flush_to_zero
,
506 equal(resulting_biased_exp
,
507 new(ir
) ir_constant(255, vec_elem
)))));
508 i
.insert_before(assign(sign_mantissa
,
510 bit_and(sign_mantissa
,
511 new(ir
) ir_constant(0x80000000u
, vec_elem
)),
514 /* Don't generate new IR that would need to be lowered in an additional
517 i
.insert_before(result
);
518 if (!lowering(INSERT_TO_SHIFTS
)) {
519 i
.insert_before(assign(result
,
520 bitfield_insert(sign_mantissa
,
521 i2u(resulting_biased_exp
),
522 new(ir
) ir_constant(23u, vec_elem
),
523 new(ir
) ir_constant(8u, vec_elem
))));
525 i
.insert_before(assign(result
,
526 bit_or(sign_mantissa
,
527 lshift(i2u(resulting_biased_exp
),
528 new(ir
) ir_constant(23, vec_elem
)))));
531 ir
->operation
= ir_triop_csel
;
532 ir
->init_num_operands();
533 ir
->operands
[0] = gequal(extracted_biased_exp
,
534 new(ir
) ir_constant(255, vec_elem
));
535 ir
->operands
[1] = new(ir
) ir_dereference_variable(x
);
536 ir
->operands
[2] = bitcast_u2f(result
);
538 this->progress
= true;
542 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
544 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
545 * from the significand.
548 const unsigned vec_elem
= ir
->type
->vector_elements
;
551 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
552 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
555 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
557 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
559 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
560 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
561 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
563 /* Temporary variables */
564 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
565 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
567 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
570 ir_variable
*extracted_biased_exp
=
571 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
572 ir_variable
*resulting_biased_exp
=
573 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
575 ir_variable
*is_not_zero_or_underflow
=
576 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
578 ir_instruction
&i
= *base_ir
;
580 /* Copy <x> and <exp> arguments. */
582 i
.insert_before(assign(x
, ir
->operands
[0]));
583 i
.insert_before(exp
);
584 i
.insert_before(assign(exp
, ir
->operands
[1]));
586 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
587 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
588 dfrexp_exp_to_arith(frexp_exp
);
590 /* Extract the biased exponent from <x>. */
591 i
.insert_before(extracted_biased_exp
);
592 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
594 i
.insert_before(resulting_biased_exp
);
595 i
.insert_before(assign(resulting_biased_exp
,
596 add(extracted_biased_exp
, exp
)));
598 /* Test if result is ±0.0, subnormal, or underflow by checking if the
599 * resulting biased exponent would be less than 0x1. If so, the result is
600 * 0.0 with the sign of x. (Actually, invert the conditions so that
601 * immediate values are the second arguments, which is better for i965)
602 * TODO: Implement in a vector fashion.
604 i
.insert_before(zero_sign_x
);
605 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
606 ir_variable
*unpacked
=
607 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
608 i
.insert_before(unpacked
);
611 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
612 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
614 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
615 i
.insert_before(assign(zero_sign_x
,
616 expr(ir_unop_pack_double_2x32
, unpacked
),
619 i
.insert_before(is_not_zero_or_underflow
);
620 i
.insert_before(assign(is_not_zero_or_underflow
,
621 gequal(resulting_biased_exp
,
622 new(ir
) ir_constant(0x1, vec_elem
))));
623 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
625 i
.insert_before(assign(resulting_biased_exp
,
626 csel(is_not_zero_or_underflow
,
627 resulting_biased_exp
, zeroi
)));
629 /* We could test for overflows by checking if the resulting biased exponent
630 * would be greater than 0xFE. Turns out we don't need to because the GLSL
633 * "If this product is too large to be represented in the
634 * floating-point type, the result is undefined."
637 ir_rvalue
*results
[4] = {NULL
};
638 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
639 ir_variable
*unpacked
=
640 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
641 i
.insert_before(unpacked
);
644 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
646 ir_expression
*bfi
= bitfield_insert(
648 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
649 exp_shift
->clone(ir
, NULL
),
650 exp_width
->clone(ir
, NULL
));
652 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
654 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
657 ir
->operation
= ir_quadop_vector
;
658 ir
->init_num_operands();
659 ir
->operands
[0] = results
[0];
660 ir
->operands
[1] = results
[1];
661 ir
->operands
[2] = results
[2];
662 ir
->operands
[3] = results
[3];
664 /* Don't generate new IR that would need to be lowered in an additional
668 this->progress
= true;
672 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
674 const unsigned vec_elem
= ir
->type
->vector_elements
;
675 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
677 /* Double-precision floating-point values are stored as
682 * We're just extracting the significand here, so we only need to modify
683 * the upper 32-bit uint. Unfortunately we must extract each double
684 * independently as there is no vector version of unpackDouble.
687 ir_instruction
&i
= *base_ir
;
689 ir_variable
*is_not_zero
=
690 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
691 ir_rvalue
*results
[4] = {NULL
};
693 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
694 i
.insert_before(is_not_zero
);
697 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
699 /* TODO: Remake this as more vector-friendly when int64 support is
702 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
703 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
704 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
706 /* Exponent of double floating-point values in the range [0.5, 1.0). */
707 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
710 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
711 ir_variable
*unpacked
=
712 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
714 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
716 i
.insert_before(bits
);
717 i
.insert_before(unpacked
);
718 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
720 /* Manipulate the high uint to remove the exponent and replace it with
721 * either the default exponent or zero.
723 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
724 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
725 i
.insert_before(assign(bits
, bit_or(bits
,
726 csel(swizzle(is_not_zero
, elem
, 1),
729 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
730 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
733 /* Put the dvec back together */
734 ir
->operation
= ir_quadop_vector
;
735 ir
->init_num_operands();
736 ir
->operands
[0] = results
[0];
737 ir
->operands
[1] = results
[1];
738 ir
->operands
[2] = results
[2];
739 ir
->operands
[3] = results
[3];
741 this->progress
= true;
745 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
747 const unsigned vec_elem
= ir
->type
->vector_elements
;
748 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
749 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
751 /* Double-precision floating-point values are stored as
756 * We're just extracting the exponent here, so we only care about the upper
760 ir_instruction
&i
= *base_ir
;
762 ir_variable
*is_not_zero
=
763 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
764 ir_variable
*high_words
=
765 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
766 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
767 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
769 ir_rvalue
*absval
= abs(ir
->operands
[0]);
771 i
.insert_before(is_not_zero
);
772 i
.insert_before(high_words
);
773 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
775 /* Extract all of the upper uints. */
776 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
777 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
779 i
.insert_before(assign(high_words
,
780 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
784 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
785 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
787 /* For non-zero inputs, shift the exponent down and apply bias. */
788 ir
->operation
= ir_triop_csel
;
789 ir
->init_num_operands();
790 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
791 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
792 ir
->operands
[2] = izero
;
794 this->progress
= true;
798 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
803 * sum = ir_binop_add x y
804 * bcarry = ir_binop_less sum x
805 * carry = ir_unop_b2i bcarry
808 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
809 ir
->operation
= ir_unop_i2u
;
810 ir
->init_num_operands();
811 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
812 ir
->operands
[1] = NULL
;
814 this->progress
= true;
818 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
821 * ir_binop_borrow x y
823 * bcarry = ir_binop_less x y
824 * carry = ir_unop_b2i bcarry
827 ir
->operation
= ir_unop_i2u
;
828 ir
->init_num_operands();
829 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
830 ir
->operands
[1] = NULL
;
832 this->progress
= true;
836 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
841 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
844 ir
->operation
= ir_binop_min
;
845 ir
->init_num_operands();
847 ir_constant
*zero
= _imm_fp(ir
, ir
->operands
[0]->type
, 0.0);
848 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
849 ir
->operands
[0], zero
);
850 ir
->operands
[1] = _imm_fp(ir
, ir
->operands
[0]->type
, 1.0);
852 this->progress
= true;
856 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
858 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
860 this->base_ir
->insert_before(temp
);
862 int nc
= ir
->operands
[0]->type
->components();
863 for (int i
= nc
- 1; i
>= 1; i
--) {
864 ir_assignment
*assig
;
866 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
867 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
869 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
870 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
873 this->base_ir
->insert_before(assig
);
876 ir
->operation
= ir_triop_fma
;
877 ir
->init_num_operands();
878 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
879 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
880 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
882 this->progress
= true;
887 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
890 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
891 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
893 switch (op2
->type
->vector_elements
) {
895 swizval
= SWIZZLE_XXXX
;
898 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
899 swizval
= SWIZZLE_XYZW
;
903 ir
->operation
= ir_triop_fma
;
904 ir
->init_num_operands();
905 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
906 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
908 this->progress
= true;
912 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
916 * temp = sub(x, frtemp);
917 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
919 ir_instruction
&i
= *base_ir
;
920 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
921 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
922 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
925 i
.insert_before(frtemp
);
926 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
928 ir
->operation
= ir_binop_add
;
929 ir
->init_num_operands();
930 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
931 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
933 this->progress
= true;
937 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
941 * result = sub(x, frtemp);
943 ir
->operation
= ir_binop_sub
;
944 ir
->init_num_operands();
945 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
947 this->progress
= true;
950 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
955 * frtemp = frac(temp);
956 * t2 = sub(temp, frtemp);
957 * if (frac(x) == 0.5)
958 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
963 ir_instruction
&i
= *base_ir
;
964 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
966 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
968 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
970 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
971 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
972 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
974 i
.insert_before(temp
);
975 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
977 i
.insert_before(frtemp
);
978 i
.insert_before(assign(frtemp
, fract(temp
)));
981 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
983 ir
->operation
= ir_triop_csel
;
984 ir
->init_num_operands();
985 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
986 p5
->clone(ir
, NULL
));
987 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
991 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
993 this->progress
= true;
997 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
1001 * temp = sub(x, frtemp);
1002 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
1004 ir_rvalue
*arg
= ir
->operands
[0];
1005 ir_instruction
&i
= *base_ir
;
1007 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
1008 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
1009 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
1011 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
1014 i
.insert_before(frtemp
);
1015 i
.insert_before(assign(frtemp
, fract(arg
)));
1016 i
.insert_before(temp
);
1017 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
1019 ir
->operation
= ir_triop_csel
;
1020 ir
->init_num_operands();
1021 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
1022 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
1023 ir
->operands
[2] = add(temp
,
1024 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
1025 zero
->clone(ir
, NULL
),
1028 this->progress
= true;
1032 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
1035 * temp = x > 0.0 ? 1.0 : 0.0;
1036 * result = x < 0.0 ? -1.0 : temp;
1038 ir_rvalue
*arg
= ir
->operands
[0];
1039 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
1040 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
1041 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
1043 ir
->operation
= ir_triop_csel
;
1044 ir
->init_num_operands();
1045 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
1046 zero
->clone(ir
, NULL
));
1047 ir
->operands
[1] = neg_one
;
1048 ir
->operands
[2] = csel(greater(arg
, zero
),
1050 zero
->clone(ir
, NULL
));
1052 this->progress
= true;
1056 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
1058 /* For more details, see:
1060 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
1062 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1063 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
1065 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
1066 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
1067 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
1068 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
1069 ir_constant
*c1
= new(ir
) ir_constant(1u);
1070 ir_constant
*c2
= new(ir
) ir_constant(2u);
1071 ir_constant
*c4
= new(ir
) ir_constant(4u);
1072 ir_constant
*c24
= new(ir
) ir_constant(24u);
1074 base_ir
->insert_before(temp
);
1076 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1077 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
1079 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1080 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
1083 /* temp = temp - ((temp >> 1) & 0x55555555u); */
1084 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
1087 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
1088 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
1089 bit_and(rshift(temp
, c2
),
1090 c33333333
->clone(ir
, NULL
)))));
1092 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
1093 ir
->operation
= ir_unop_u2i
;
1094 ir
->init_num_operands();
1095 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1099 this->progress
= true;
1103 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1106 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1108 base_ir
->insert_before(bits
);
1109 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1111 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1113 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1115 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1116 ir_constant
*cFFFFFFFF
=
1117 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1119 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1120 * we'd get a mask of 0 when bits is 32. Special case it.
1122 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1124 ir_expression
*mask
= csel(equal(bits
, c32
),
1126 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1128 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1130 * If bits is zero, the result will be zero.
1132 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1133 * select as in the signed integer case.
1135 * (value >> offset) & mask;
1137 ir
->operation
= ir_binop_bit_and
;
1138 ir
->init_num_operands();
1139 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1140 ir
->operands
[1] = mask
;
1141 ir
->operands
[2] = NULL
;
1144 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1146 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1148 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1150 /* temp = 32 - bits; */
1151 base_ir
->insert_before(temp
);
1152 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1154 /* expr = value << (temp - offset)) >> temp; */
1155 ir_expression
*expr
=
1156 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1158 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1160 * If bits is zero, the result will be zero.
1162 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1163 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1166 * (bits == 0) ? 0 : e;
1168 ir
->operation
= ir_triop_csel
;
1169 ir
->init_num_operands();
1170 ir
->operands
[0] = equal(c0
, bits
);
1171 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1172 ir
->operands
[2] = expr
;
1175 this->progress
= true;
1179 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1183 ir_constant
*cFFFFFFFF
;
1184 ir_variable
*offset
=
1185 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1187 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1189 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1191 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1192 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1193 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1194 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1196 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1198 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1199 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1200 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1203 base_ir
->insert_before(offset
);
1204 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1206 base_ir
->insert_before(bits
);
1207 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1209 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1210 * we'd get a mask of 0 when bits is 32. Special case it.
1212 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1214 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1216 * The result will be undefined if offset or bits is negative, or if the
1217 * sum of offset and bits is greater than the number of bits used to
1218 * store the operand.
1220 * Since it's undefined, there are a couple other ways this could be
1221 * implemented. The other way that was considered was to put the csel
1222 * around the whole thing:
1224 * final_result = bits == 32 ? insert : ... ;
1226 base_ir
->insert_before(mask
);
1228 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1230 lshift(sub(lshift(c1
, bits
),
1231 c1
->clone(ir
, NULL
)),
1234 /* (base & ~mask) | ((insert << offset) & mask) */
1235 ir
->operation
= ir_binop_bit_or
;
1236 ir
->init_num_operands();
1237 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1238 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1239 ir
->operands
[2] = NULL
;
1240 ir
->operands
[3] = NULL
;
1242 this->progress
= true;
1246 lower_instructions_visitor::reverse_to_shifts(ir_expression
*ir
)
1248 /* For more details, see:
1250 * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
1253 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1255 new(ir
) ir_constant(2u, ir
->operands
[0]->type
->vector_elements
);
1257 new(ir
) ir_constant(4u, ir
->operands
[0]->type
->vector_elements
);
1259 new(ir
) ir_constant(8u, ir
->operands
[0]->type
->vector_elements
);
1261 new(ir
) ir_constant(16u, ir
->operands
[0]->type
->vector_elements
);
1262 ir_constant
*c33333333
=
1263 new(ir
) ir_constant(0x33333333u
, ir
->operands
[0]->type
->vector_elements
);
1264 ir_constant
*c55555555
=
1265 new(ir
) ir_constant(0x55555555u
, ir
->operands
[0]->type
->vector_elements
);
1266 ir_constant
*c0F0F0F0F
=
1267 new(ir
) ir_constant(0x0F0F0F0Fu
, ir
->operands
[0]->type
->vector_elements
);
1268 ir_constant
*c00FF00FF
=
1269 new(ir
) ir_constant(0x00FF00FFu
, ir
->operands
[0]->type
->vector_elements
);
1271 new(ir
) ir_variable(glsl_type::uvec(ir
->operands
[0]->type
->vector_elements
),
1272 "temp", ir_var_temporary
);
1273 ir_instruction
&i
= *base_ir
;
1275 i
.insert_before(temp
);
1277 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1278 i
.insert_before(assign(temp
, ir
->operands
[0]));
1280 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1281 i
.insert_before(assign(temp
, i2u(ir
->operands
[0])));
1284 /* Swap odd and even bits.
1286 * temp = ((temp >> 1) & 0x55555555u) | ((temp & 0x55555555u) << 1);
1288 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c1
), c55555555
),
1289 lshift(bit_and(temp
, c55555555
->clone(ir
, NULL
)),
1290 c1
->clone(ir
, NULL
)))));
1291 /* Swap consecutive pairs.
1293 * temp = ((temp >> 2) & 0x33333333u) | ((temp & 0x33333333u) << 2);
1295 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c2
), c33333333
),
1296 lshift(bit_and(temp
, c33333333
->clone(ir
, NULL
)),
1297 c2
->clone(ir
, NULL
)))));
1301 * temp = ((temp >> 4) & 0x0F0F0F0Fu) | ((temp & 0x0F0F0F0Fu) << 4);
1303 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c4
), c0F0F0F0F
),
1304 lshift(bit_and(temp
, c0F0F0F0F
->clone(ir
, NULL
)),
1305 c4
->clone(ir
, NULL
)))));
1307 /* The last step is, basically, bswap. Swap the bytes, then swap the
1308 * words. When this code is run through GCC on x86, it does generate a
1309 * bswap instruction.
1311 * temp = ((temp >> 8) & 0x00FF00FFu) | ((temp & 0x00FF00FFu) << 8);
1312 * temp = ( temp >> 16 ) | ( temp << 16);
1314 i
.insert_before(assign(temp
, bit_or(bit_and(rshift(temp
, c8
), c00FF00FF
),
1315 lshift(bit_and(temp
, c00FF00FF
->clone(ir
, NULL
)),
1316 c8
->clone(ir
, NULL
)))));
1318 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1319 ir
->operation
= ir_binop_bit_or
;
1320 ir
->init_num_operands();
1321 ir
->operands
[0] = rshift(temp
, c16
);
1322 ir
->operands
[1] = lshift(temp
, c16
->clone(ir
, NULL
));
1324 ir
->operation
= ir_unop_u2i
;
1325 ir
->init_num_operands();
1326 ir
->operands
[0] = bit_or(rshift(temp
, c16
),
1327 lshift(temp
, c16
->clone(ir
, NULL
)));
1330 this->progress
= true;
1334 lower_instructions_visitor::find_lsb_to_float_cast(ir_expression
*ir
)
1336 /* For more details, see:
1338 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1340 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1341 ir_constant
*c0
= new(ir
) ir_constant(unsigned(0), elements
);
1342 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1343 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1344 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1346 new(ir
) ir_variable(glsl_type::ivec(elements
), "temp", ir_var_temporary
);
1347 ir_variable
*lsb_only
=
1348 new(ir
) ir_variable(glsl_type::uvec(elements
), "lsb_only", ir_var_temporary
);
1349 ir_variable
*as_float
=
1350 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1352 new(ir
) ir_variable(glsl_type::ivec(elements
), "lsb", ir_var_temporary
);
1354 ir_instruction
&i
= *base_ir
;
1356 i
.insert_before(temp
);
1358 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1359 i
.insert_before(assign(temp
, ir
->operands
[0]));
1361 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1362 i
.insert_before(assign(temp
, u2i(ir
->operands
[0])));
1365 /* The int-to-float conversion is lossless because (value & -value) is
1366 * either a power of two or zero. We don't use the result in the zero
1367 * case. The uint() cast is necessary so that 0x80000000 does not
1368 * generate a negative value.
1370 * uint lsb_only = uint(value & -value);
1371 * float as_float = float(lsb_only);
1373 i
.insert_before(lsb_only
);
1374 i
.insert_before(assign(lsb_only
, i2u(bit_and(temp
, neg(temp
)))));
1376 i
.insert_before(as_float
);
1377 i
.insert_before(assign(as_float
, u2f(lsb_only
)));
1379 /* This is basically an open-coded frexp. Implementations that have a
1380 * native frexp instruction would be better served by that. This is
1381 * optimized versus a full-featured open-coded implementation in two ways:
1383 * - We don't care about a correct result from subnormal numbers (including
1384 * 0.0), so the raw exponent can always be safely unbiased.
1386 * - The value cannot be negative, so it does not need to be masked off to
1387 * extract the exponent.
1389 * int lsb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1391 i
.insert_before(lsb
);
1392 i
.insert_before(assign(lsb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1394 /* Use lsb_only in the comparison instead of temp so that the & (far above)
1395 * can possibly generate the result without an explicit comparison.
1397 * (lsb_only == 0) ? -1 : lsb;
1399 * Since our input values are all integers, the unbiased exponent must not
1400 * be negative. It will only be negative (-0x7f, in fact) if lsb_only is
1401 * 0. Instead of using (lsb_only == 0), we could use (lsb >= 0). Which is
1402 * better is likely GPU dependent. Either way, the difference should be
1405 ir
->operation
= ir_triop_csel
;
1406 ir
->init_num_operands();
1407 ir
->operands
[0] = equal(lsb_only
, c0
);
1408 ir
->operands
[1] = cminus1
;
1409 ir
->operands
[2] = new(ir
) ir_dereference_variable(lsb
);
1411 this->progress
= true;
1415 lower_instructions_visitor::find_msb_to_float_cast(ir_expression
*ir
)
1417 /* For more details, see:
1419 * http://graphics.stanford.edu/~seander/bithacks.html#ZerosOnRightFloatCast
1421 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1422 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1423 ir_constant
*cminus1
= new(ir
) ir_constant(int(-1), elements
);
1424 ir_constant
*c23
= new(ir
) ir_constant(int(23), elements
);
1425 ir_constant
*c7F
= new(ir
) ir_constant(int(0x7F), elements
);
1426 ir_constant
*c000000FF
= new(ir
) ir_constant(0x000000FFu
, elements
);
1427 ir_constant
*cFFFFFF00
= new(ir
) ir_constant(0xFFFFFF00u
, elements
);
1429 new(ir
) ir_variable(glsl_type::uvec(elements
), "temp", ir_var_temporary
);
1430 ir_variable
*as_float
=
1431 new(ir
) ir_variable(glsl_type::vec(elements
), "as_float", ir_var_temporary
);
1433 new(ir
) ir_variable(glsl_type::ivec(elements
), "msb", ir_var_temporary
);
1435 ir_instruction
&i
= *base_ir
;
1437 i
.insert_before(temp
);
1439 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1440 i
.insert_before(assign(temp
, ir
->operands
[0]));
1442 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1444 /* findMSB(uint(abs(some_int))) almost always does the right thing.
1445 * There are two problem values:
1447 * * 0x80000000. Since abs(0x80000000) == 0x80000000, findMSB returns
1448 * 31. However, findMSB(int(0x80000000)) == 30.
1450 * * 0xffffffff. Since abs(0xffffffff) == 1, findMSB returns
1451 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1453 * For a value of zero or negative one, -1 will be returned.
1455 * For all negative number cases, including 0x80000000 and 0xffffffff,
1456 * the correct value is obtained from findMSB if instead of negating the
1457 * (already negative) value the logical-not is used. A conditonal
1458 * logical-not can be achieved in two instructions.
1460 ir_variable
*as_int
=
1461 new(ir
) ir_variable(glsl_type::ivec(elements
), "as_int", ir_var_temporary
);
1462 ir_constant
*c31
= new(ir
) ir_constant(int(31), elements
);
1464 i
.insert_before(as_int
);
1465 i
.insert_before(assign(as_int
, ir
->operands
[0]));
1466 i
.insert_before(assign(temp
, i2u(expr(ir_binop_bit_xor
,
1468 rshift(as_int
, c31
)))));
1471 /* The int-to-float conversion is lossless because bits are conditionally
1472 * masked off the bottom of temp to ensure the value has at most 24 bits of
1473 * data or is zero. We don't use the result in the zero case. The uint()
1474 * cast is necessary so that 0x80000000 does not generate a negative value.
1476 * float as_float = float(temp > 255 ? temp & ~255 : temp);
1478 i
.insert_before(as_float
);
1479 i
.insert_before(assign(as_float
, u2f(csel(greater(temp
, c000000FF
),
1480 bit_and(temp
, cFFFFFF00
),
1483 /* This is basically an open-coded frexp. Implementations that have a
1484 * native frexp instruction would be better served by that. This is
1485 * optimized versus a full-featured open-coded implementation in two ways:
1487 * - We don't care about a correct result from subnormal numbers (including
1488 * 0.0), so the raw exponent can always be safely unbiased.
1490 * - The value cannot be negative, so it does not need to be masked off to
1491 * extract the exponent.
1493 * int msb = (floatBitsToInt(as_float) >> 23) - 0x7f;
1495 i
.insert_before(msb
);
1496 i
.insert_before(assign(msb
, sub(rshift(bitcast_f2i(as_float
), c23
), c7F
)));
1498 /* Use msb in the comparison instead of temp so that the subtract can
1499 * possibly generate the result without an explicit comparison.
1501 * (msb < 0) ? -1 : msb;
1503 * Since our input values are all integers, the unbiased exponent must not
1504 * be negative. It will only be negative (-0x7f, in fact) if temp is 0.
1506 ir
->operation
= ir_triop_csel
;
1507 ir
->init_num_operands();
1508 ir
->operands
[0] = less(msb
, c0
);
1509 ir
->operands
[1] = cminus1
;
1510 ir
->operands
[2] = new(ir
) ir_dereference_variable(msb
);
1512 this->progress
= true;
1516 lower_instructions_visitor::_carry(operand a
, operand b
)
1518 if (lowering(CARRY_TO_ARITH
))
1519 return i2u(b2i(less(add(a
, b
),
1520 a
.val
->clone(ralloc_parent(a
.val
), NULL
))));
1526 lower_instructions_visitor::_imm_fp(void *mem_ctx
,
1527 const glsl_type
*type
,
1529 unsigned vector_elements
)
1531 switch (type
->base_type
) {
1532 case GLSL_TYPE_FLOAT
:
1533 return new(mem_ctx
) ir_constant((float) f
, vector_elements
);
1534 case GLSL_TYPE_DOUBLE
:
1535 return new(mem_ctx
) ir_constant((double) f
, vector_elements
);
1536 case GLSL_TYPE_FLOAT16
:
1537 return new(mem_ctx
) ir_constant(float16_t(f
), vector_elements
);
1539 assert(!"unknown float type for immediate");
1545 lower_instructions_visitor::imul_high_to_mul(ir_expression
*ir
)
1550 * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
1552 * In GLSL, (a * b) becomes
1554 * uint m1 = (a & 0x0000ffffu) * (b & 0x0000ffffu);
1555 * uint m2 = (a & 0x0000ffffu) * (b >> 16);
1556 * uint m3 = (a >> 16) * (b & 0x0000ffffu);
1557 * uint m4 = (a >> 16) * (b >> 16);
1564 * lo_result = uaddCarry(m1, m2 << 16, c1);
1565 * hi_result = m4 + c1;
1566 * lo_result = uaddCarry(lo_result, m3 << 16, c2);
1567 * hi_result = hi_result + c2;
1568 * hi_result = hi_result + (m2 >> 16) + (m3 >> 16);
1570 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1572 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1", ir_var_temporary
);
1573 ir_variable
*src1h
=
1574 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1h", ir_var_temporary
);
1575 ir_variable
*src1l
=
1576 new(ir
) ir_variable(glsl_type::uvec(elements
), "src1l", ir_var_temporary
);
1578 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2", ir_var_temporary
);
1579 ir_variable
*src2h
=
1580 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2h", ir_var_temporary
);
1581 ir_variable
*src2l
=
1582 new(ir
) ir_variable(glsl_type::uvec(elements
), "src2l", ir_var_temporary
);
1584 new(ir
) ir_variable(glsl_type::uvec(elements
), "t1", ir_var_temporary
);
1586 new(ir
) ir_variable(glsl_type::uvec(elements
), "t2", ir_var_temporary
);
1588 new(ir
) ir_variable(glsl_type::uvec(elements
), "lo", ir_var_temporary
);
1590 new(ir
) ir_variable(glsl_type::uvec(elements
), "hi", ir_var_temporary
);
1591 ir_variable
*different_signs
= NULL
;
1592 ir_constant
*c0000FFFF
= new(ir
) ir_constant(0x0000FFFFu
, elements
);
1593 ir_constant
*c16
= new(ir
) ir_constant(16u, elements
);
1595 ir_instruction
&i
= *base_ir
;
1597 i
.insert_before(src1
);
1598 i
.insert_before(src2
);
1599 i
.insert_before(src1h
);
1600 i
.insert_before(src2h
);
1601 i
.insert_before(src1l
);
1602 i
.insert_before(src2l
);
1604 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1605 i
.insert_before(assign(src1
, ir
->operands
[0]));
1606 i
.insert_before(assign(src2
, ir
->operands
[1]));
1608 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1610 ir_variable
*itmp1
=
1611 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp1", ir_var_temporary
);
1612 ir_variable
*itmp2
=
1613 new(ir
) ir_variable(glsl_type::ivec(elements
), "itmp2", ir_var_temporary
);
1614 ir_constant
*c0
= new(ir
) ir_constant(int(0), elements
);
1616 i
.insert_before(itmp1
);
1617 i
.insert_before(itmp2
);
1618 i
.insert_before(assign(itmp1
, ir
->operands
[0]));
1619 i
.insert_before(assign(itmp2
, ir
->operands
[1]));
1622 new(ir
) ir_variable(glsl_type::bvec(elements
), "different_signs",
1625 i
.insert_before(different_signs
);
1626 i
.insert_before(assign(different_signs
, expr(ir_binop_logic_xor
,
1628 less(itmp2
, c0
->clone(ir
, NULL
)))));
1630 i
.insert_before(assign(src1
, i2u(abs(itmp1
))));
1631 i
.insert_before(assign(src2
, i2u(abs(itmp2
))));
1634 i
.insert_before(assign(src1l
, bit_and(src1
, c0000FFFF
)));
1635 i
.insert_before(assign(src2l
, bit_and(src2
, c0000FFFF
->clone(ir
, NULL
))));
1636 i
.insert_before(assign(src1h
, rshift(src1
, c16
)));
1637 i
.insert_before(assign(src2h
, rshift(src2
, c16
->clone(ir
, NULL
))));
1639 i
.insert_before(lo
);
1640 i
.insert_before(hi
);
1641 i
.insert_before(t1
);
1642 i
.insert_before(t2
);
1644 i
.insert_before(assign(lo
, mul(src1l
, src2l
)));
1645 i
.insert_before(assign(t1
, mul(src1l
, src2h
)));
1646 i
.insert_before(assign(t2
, mul(src1h
, src2l
)));
1647 i
.insert_before(assign(hi
, mul(src1h
, src2h
)));
1649 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t1
, c16
->clone(ir
, NULL
))))));
1650 i
.insert_before(assign(lo
, add(lo
, lshift(t1
, c16
->clone(ir
, NULL
)))));
1652 i
.insert_before(assign(hi
, add(hi
, _carry(lo
, lshift(t2
, c16
->clone(ir
, NULL
))))));
1653 i
.insert_before(assign(lo
, add(lo
, lshift(t2
, c16
->clone(ir
, NULL
)))));
1655 if (different_signs
== NULL
) {
1656 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1658 ir
->operation
= ir_binop_add
;
1659 ir
->init_num_operands();
1660 ir
->operands
[0] = add(hi
, rshift(t1
, c16
->clone(ir
, NULL
)));
1661 ir
->operands
[1] = rshift(t2
, c16
->clone(ir
, NULL
));
1663 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
1665 i
.insert_before(assign(hi
, add(add(hi
, rshift(t1
, c16
->clone(ir
, NULL
))),
1666 rshift(t2
, c16
->clone(ir
, NULL
)))));
1668 /* For channels where different_signs is set we have to perform a 64-bit
1669 * negation. This is *not* the same as just negating the high 32-bits.
1670 * Consider -3 * 2. The high 32-bits is 0, but the desired result is
1671 * -1, not -0! Recall -x == ~x + 1.
1673 ir_variable
*neg_hi
=
1674 new(ir
) ir_variable(glsl_type::ivec(elements
), "neg_hi", ir_var_temporary
);
1675 ir_constant
*c1
= new(ir
) ir_constant(1u, elements
);
1677 i
.insert_before(neg_hi
);
1678 i
.insert_before(assign(neg_hi
, add(bit_not(u2i(hi
)),
1679 u2i(_carry(bit_not(lo
), c1
)))));
1681 ir
->operation
= ir_triop_csel
;
1682 ir
->init_num_operands();
1683 ir
->operands
[0] = new(ir
) ir_dereference_variable(different_signs
);
1684 ir
->operands
[1] = new(ir
) ir_dereference_variable(neg_hi
);
1685 ir
->operands
[2] = u2i(hi
);
1690 lower_instructions_visitor::sqrt_to_abs_sqrt(ir_expression
*ir
)
1692 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_abs
, ir
->operands
[0]);
1693 this->progress
= true;
1697 lower_instructions_visitor::mul64_to_mul_and_mul_high(ir_expression
*ir
)
1699 /* Lower 32x32-> 64 to
1700 * msb = imul_high(x_lo, y_lo)
1701 * lsb = mul(x_lo, y_lo)
1703 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
1705 const ir_expression_operation operation
=
1706 ir
->type
->base_type
== GLSL_TYPE_UINT64
? ir_unop_pack_uint_2x32
1707 : ir_unop_pack_int_2x32
;
1709 const glsl_type
*var_type
= ir
->type
->base_type
== GLSL_TYPE_UINT64
1710 ? glsl_type::uvec(elements
)
1711 : glsl_type::ivec(elements
);
1713 const glsl_type
*ret_type
= ir
->type
->base_type
== GLSL_TYPE_UINT64
1714 ? glsl_type::uvec2_type
1715 : glsl_type::ivec2_type
;
1717 ir_instruction
&i
= *base_ir
;
1720 new(ir
) ir_variable(var_type
, "msb", ir_var_temporary
);
1722 new(ir
) ir_variable(var_type
, "lsb", ir_var_temporary
);
1724 new(ir
) ir_variable(var_type
, "x", ir_var_temporary
);
1726 new(ir
) ir_variable(var_type
, "y", ir_var_temporary
);
1729 i
.insert_before(assign(x
, ir
->operands
[0]));
1731 i
.insert_before(assign(y
, ir
->operands
[1]));
1732 i
.insert_before(msb
);
1733 i
.insert_before(lsb
);
1735 i
.insert_before(assign(msb
, imul_high(x
, y
)));
1736 i
.insert_before(assign(lsb
, mul(x
, y
)));
1738 ir_rvalue
*result
[4] = {NULL
};
1739 for (unsigned elem
= 0; elem
< elements
; elem
++) {
1740 ir_rvalue
*val
= new(ir
) ir_expression(ir_quadop_vector
, ret_type
,
1741 swizzle(lsb
, elem
, 1),
1742 swizzle(msb
, elem
, 1), NULL
, NULL
);
1743 result
[elem
] = expr(operation
, val
);
1746 ir
->operation
= ir_quadop_vector
;
1747 ir
->init_num_operands();
1748 ir
->operands
[0] = result
[0];
1749 ir
->operands
[1] = result
[1];
1750 ir
->operands
[2] = result
[2];
1751 ir
->operands
[3] = result
[3];
1753 this->progress
= true;
1757 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1759 switch (ir
->operation
) {
1761 if (ir
->operands
[0]->type
->is_double())
1762 double_dot_to_fma(ir
);
1765 if (ir
->operands
[0]->type
->is_double())
1769 if (lowering(SUB_TO_ADD_NEG
))
1774 if (ir
->operands
[1]->type
->is_integer_32() && lowering(INT_DIV_TO_MUL_RCP
))
1775 int_div_to_mul_rcp(ir
);
1776 else if ((ir
->operands
[1]->type
->is_float() && lowering(FDIV_TO_MUL_RCP
)) ||
1777 (ir
->operands
[1]->type
->is_double() && lowering(DDIV_TO_MUL_RCP
)))
1782 if (lowering(EXP_TO_EXP2
))
1787 if (lowering(LOG_TO_LOG2
))
1792 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1797 if (lowering(POW_TO_EXP2
))
1801 case ir_binop_ldexp
:
1802 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1804 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1805 dldexp_to_arith(ir
);
1808 case ir_unop_frexp_exp
:
1809 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1810 dfrexp_exp_to_arith(ir
);
1813 case ir_unop_frexp_sig
:
1814 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1815 dfrexp_sig_to_arith(ir
);
1818 case ir_binop_carry
:
1819 if (lowering(CARRY_TO_ARITH
))
1823 case ir_binop_borrow
:
1824 if (lowering(BORROW_TO_ARITH
))
1825 borrow_to_arith(ir
);
1828 case ir_unop_saturate
:
1829 if (lowering(SAT_TO_CLAMP
))
1834 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1835 dtrunc_to_dfrac(ir
);
1839 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1844 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1845 dfloor_to_dfrac(ir
);
1848 case ir_unop_round_even
:
1849 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1850 dround_even_to_dfrac(ir
);
1854 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1858 case ir_unop_bit_count
:
1859 if (lowering(BIT_COUNT_TO_MATH
))
1860 bit_count_to_math(ir
);
1863 case ir_triop_bitfield_extract
:
1864 if (lowering(EXTRACT_TO_SHIFTS
))
1865 extract_to_shifts(ir
);
1868 case ir_quadop_bitfield_insert
:
1869 if (lowering(INSERT_TO_SHIFTS
))
1870 insert_to_shifts(ir
);
1873 case ir_unop_bitfield_reverse
:
1874 if (lowering(REVERSE_TO_SHIFTS
))
1875 reverse_to_shifts(ir
);
1878 case ir_unop_find_lsb
:
1879 if (lowering(FIND_LSB_TO_FLOAT_CAST
))
1880 find_lsb_to_float_cast(ir
);
1883 case ir_unop_find_msb
:
1884 if (lowering(FIND_MSB_TO_FLOAT_CAST
))
1885 find_msb_to_float_cast(ir
);
1888 case ir_binop_imul_high
:
1889 if (lowering(IMUL_HIGH_TO_MUL
))
1890 imul_high_to_mul(ir
);
1894 if (lowering(MUL64_TO_MUL_AND_MUL_HIGH
) &&
1895 (ir
->type
->base_type
== GLSL_TYPE_INT64
||
1896 ir
->type
->base_type
== GLSL_TYPE_UINT64
) &&
1897 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
||
1898 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_UINT
))
1899 mul64_to_mul_and_mul_high(ir
);
1904 if (lowering(SQRT_TO_ABS_SQRT
))
1905 sqrt_to_abs_sqrt(ir
);
1909 return visit_continue
;
1912 return visit_continue
;