2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file lower_instructions.cpp
27 * Many GPUs lack native instructions for certain expression operations, and
28 * must replace them with some other expression tree. This pass lowers some
29 * of the most common cases, allowing the lowering code to be implemented once
30 * rather than in each driver backend.
32 * Currently supported transformations:
35 * - INT_DIV_TO_MUL_RCP
49 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
51 * This simplifies expression reassociation, and for many backends
52 * there is no subtract operation separate from adding the negation.
53 * For backends with native subtract operations, they will probably
54 * want to recognize add(op0, neg(op1)) or the other way around to
55 * produce a subtract anyway.
57 * DIV_TO_MUL_RCP and INT_DIV_TO_MUL_RCP:
58 * --------------------------------------
59 * Breaks an ir_binop_div expression down to op0 * (rcp(op1)).
61 * Many GPUs don't have a divide instruction (945 and 965 included),
62 * but they do have an RCP instruction to compute an approximate
63 * reciprocal. By breaking the operation down, constant reciprocals
64 * can get constant folded.
66 * DIV_TO_MUL_RCP only lowers floating point division; INT_DIV_TO_MUL_RCP
67 * handles the integer case, converting to and from floating point so that
70 * EXP_TO_EXP2 and LOG_TO_LOG2:
71 * ----------------------------
72 * Many GPUs don't have a base e log or exponent instruction, but they
73 * do have base 2 versions, so this pass converts exp and log to exp2
74 * and log2 operations.
78 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
79 * x**y to 2**(y * log2(x)).
83 * Breaks an ir_binop_mod expression down to (op0 - op1 * floor(op0 / op1))
85 * Many GPUs don't have a MOD instruction (945 and 965 included), and
86 * if we have to break it down like this anyway, it gives an
87 * opportunity to do things like constant fold the (1.0 / op1) easily.
89 * Note: before we used to implement this as op1 * fract(op / op1) but this
90 * implementation had significant precision errors.
94 * Converts ir_binop_ldexp to arithmetic and bit operations for float sources.
96 * DFREXP_DLDEXP_TO_ARITH:
98 * Converts ir_binop_ldexp, ir_unop_frexp_sig, and ir_unop_frexp_exp to
99 * arithmetic and bit ops for double arguments.
103 * Converts ir_carry into (x + y) < x.
107 * Converts ir_borrow into (x < y).
111 * Converts ir_unop_saturate into min(max(x, 0.0), 1.0)
115 * Converts double trunc, ceil, floor, round to fract
118 #include "c99_math.h"
119 #include "program/prog_instruction.h" /* for swizzle */
120 #include "glsl_types.h"
122 #include "ir_builder.h"
123 #include "ir_optimization.h"
125 using namespace ir_builder
;
129 class lower_instructions_visitor
: public ir_hierarchical_visitor
{
131 lower_instructions_visitor(unsigned lower
)
132 : progress(false), lower(lower
) { }
134 ir_visitor_status
visit_leave(ir_expression
*);
139 unsigned lower
; /** Bitfield of which operations to lower */
141 void sub_to_add_neg(ir_expression
*);
142 void div_to_mul_rcp(ir_expression
*);
143 void int_div_to_mul_rcp(ir_expression
*);
144 void mod_to_floor(ir_expression
*);
145 void exp_to_exp2(ir_expression
*);
146 void pow_to_exp2(ir_expression
*);
147 void log_to_log2(ir_expression
*);
148 void ldexp_to_arith(ir_expression
*);
149 void dldexp_to_arith(ir_expression
*);
150 void dfrexp_sig_to_arith(ir_expression
*);
151 void dfrexp_exp_to_arith(ir_expression
*);
152 void carry_to_arith(ir_expression
*);
153 void borrow_to_arith(ir_expression
*);
154 void sat_to_clamp(ir_expression
*);
155 void double_dot_to_fma(ir_expression
*);
156 void double_lrp(ir_expression
*);
157 void dceil_to_dfrac(ir_expression
*);
158 void dfloor_to_dfrac(ir_expression
*);
159 void dround_even_to_dfrac(ir_expression
*);
160 void dtrunc_to_dfrac(ir_expression
*);
161 void dsign_to_csel(ir_expression
*);
164 } /* anonymous namespace */
167 * Determine if a particular type of lowering should occur
169 #define lowering(x) (this->lower & x)
172 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
174 lower_instructions_visitor
v(what_to_lower
);
176 visit_list_elements(&v
, instructions
);
181 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
183 ir
->operation
= ir_binop_add
;
184 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
185 ir
->operands
[1], NULL
);
186 this->progress
= true;
190 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
192 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
194 /* New expression for the 1.0 / op1 */
196 expr
= new(ir
) ir_expression(ir_unop_rcp
,
197 ir
->operands
[1]->type
,
200 /* op0 / op1 -> op0 * (1.0 / op1) */
201 ir
->operation
= ir_binop_mul
;
202 ir
->operands
[1] = expr
;
204 this->progress
= true;
208 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
210 assert(ir
->operands
[1]->type
->is_integer());
212 /* Be careful with integer division -- we need to do it as a
213 * float and re-truncate, since rcp(n > 1) of an integer would
216 ir_rvalue
*op0
, *op1
;
217 const struct glsl_type
*vec_type
;
219 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
220 ir
->operands
[1]->type
->vector_elements
,
221 ir
->operands
[1]->type
->matrix_columns
);
223 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
224 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
226 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
228 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
230 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
231 ir
->operands
[0]->type
->vector_elements
,
232 ir
->operands
[0]->type
->matrix_columns
);
234 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
235 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
237 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
239 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
240 ir
->type
->vector_elements
,
241 ir
->type
->matrix_columns
);
243 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
245 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
246 ir
->operation
= ir_unop_f2i
;
247 ir
->operands
[0] = op0
;
249 ir
->operation
= ir_unop_i2u
;
250 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
252 ir
->operands
[1] = NULL
;
254 this->progress
= true;
258 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
260 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
262 ir
->operation
= ir_unop_exp2
;
263 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
264 ir
->operands
[0], log2_e
);
265 this->progress
= true;
269 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
271 ir_expression
*const log2_x
=
272 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
275 ir
->operation
= ir_unop_exp2
;
276 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
277 ir
->operands
[1], log2_x
);
278 ir
->operands
[1] = NULL
;
279 this->progress
= true;
283 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
285 ir
->operation
= ir_binop_mul
;
286 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
287 ir
->operands
[0], NULL
);
288 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
289 this->progress
= true;
293 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
295 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
297 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
299 this->base_ir
->insert_before(x
);
300 this->base_ir
->insert_before(y
);
302 ir_assignment
*const assign_x
=
303 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
304 ir
->operands
[0], NULL
);
305 ir_assignment
*const assign_y
=
306 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
307 ir
->operands
[1], NULL
);
309 this->base_ir
->insert_before(assign_x
);
310 this->base_ir
->insert_before(assign_y
);
312 ir_expression
*const div_expr
=
313 new(ir
) ir_expression(ir_binop_div
, x
->type
,
314 new(ir
) ir_dereference_variable(x
),
315 new(ir
) ir_dereference_variable(y
));
317 /* Don't generate new IR that would need to be lowered in an additional
320 if (lowering(DIV_TO_MUL_RCP
) && (ir
->type
->is_float() || ir
->type
->is_double()))
321 div_to_mul_rcp(div_expr
);
323 ir_expression
*const floor_expr
=
324 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
326 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
327 dfloor_to_dfrac(floor_expr
);
329 ir_expression
*const mul_expr
=
330 new(ir
) ir_expression(ir_binop_mul
,
331 new(ir
) ir_dereference_variable(y
),
334 ir
->operation
= ir_binop_sub
;
335 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
336 ir
->operands
[1] = mul_expr
;
337 this->progress
= true;
341 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
344 * ir_binop_ldexp x exp
347 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
348 * resulting_biased_exp = extracted_biased_exp + exp;
350 * if (resulting_biased_exp < 1) {
351 * return copysign(0.0, x);
354 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
355 * lshift(i2u(resulting_biased_exp), exp_shift));
357 * which we can't actually implement as such, since the GLSL IR doesn't
358 * have vectorized if-statements. We actually implement it without branches
359 * using conditional-select:
361 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
362 * resulting_biased_exp = extracted_biased_exp + exp;
364 * is_not_zero_or_underflow = gequal(resulting_biased_exp, 1);
365 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
366 * resulting_biased_exp = csel(is_not_zero_or_underflow,
367 * resulting_biased_exp, 0);
369 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
370 * lshift(i2u(resulting_biased_exp), exp_shift));
373 const unsigned vec_elem
= ir
->type
->vector_elements
;
376 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
377 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
380 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
382 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
384 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
385 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
387 /* Temporary variables */
388 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
389 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
391 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
394 ir_variable
*extracted_biased_exp
=
395 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
396 ir_variable
*resulting_biased_exp
=
397 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
399 ir_variable
*is_not_zero_or_underflow
=
400 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
402 ir_instruction
&i
= *base_ir
;
404 /* Copy <x> and <exp> arguments. */
406 i
.insert_before(assign(x
, ir
->operands
[0]));
407 i
.insert_before(exp
);
408 i
.insert_before(assign(exp
, ir
->operands
[1]));
410 /* Extract the biased exponent from <x>. */
411 i
.insert_before(extracted_biased_exp
);
412 i
.insert_before(assign(extracted_biased_exp
,
413 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
415 i
.insert_before(resulting_biased_exp
);
416 i
.insert_before(assign(resulting_biased_exp
,
417 add(extracted_biased_exp
, exp
)));
419 /* Test if result is ±0.0, subnormal, or underflow by checking if the
420 * resulting biased exponent would be less than 0x1. If so, the result is
421 * 0.0 with the sign of x. (Actually, invert the conditions so that
422 * immediate values are the second arguments, which is better for i965)
424 i
.insert_before(zero_sign_x
);
425 i
.insert_before(assign(zero_sign_x
,
426 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
428 i
.insert_before(is_not_zero_or_underflow
);
429 i
.insert_before(assign(is_not_zero_or_underflow
,
430 gequal(resulting_biased_exp
,
431 new(ir
) ir_constant(0x1, vec_elem
))));
432 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
434 i
.insert_before(assign(resulting_biased_exp
,
435 csel(is_not_zero_or_underflow
,
436 resulting_biased_exp
, zeroi
)));
438 /* We could test for overflows by checking if the resulting biased exponent
439 * would be greater than 0xFE. Turns out we don't need to because the GLSL
442 * "If this product is too large to be represented in the
443 * floating-point type, the result is undefined."
446 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
447 ir
->operation
= ir_unop_bitcast_i2f
;
448 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
449 exp_shift_clone
, exp_width
);
450 ir
->operands
[1] = NULL
;
452 this->progress
= true;
456 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
458 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
459 * from the significand.
462 const unsigned vec_elem
= ir
->type
->vector_elements
;
465 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
466 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
469 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
471 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
473 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
474 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
475 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
477 /* Temporary variables */
478 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
479 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
481 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
484 ir_variable
*extracted_biased_exp
=
485 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
486 ir_variable
*resulting_biased_exp
=
487 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
489 ir_variable
*is_not_zero_or_underflow
=
490 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
492 ir_instruction
&i
= *base_ir
;
494 /* Copy <x> and <exp> arguments. */
496 i
.insert_before(assign(x
, ir
->operands
[0]));
497 i
.insert_before(exp
);
498 i
.insert_before(assign(exp
, ir
->operands
[1]));
500 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
501 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
502 dfrexp_exp_to_arith(frexp_exp
);
504 /* Extract the biased exponent from <x>. */
505 i
.insert_before(extracted_biased_exp
);
506 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
508 i
.insert_before(resulting_biased_exp
);
509 i
.insert_before(assign(resulting_biased_exp
,
510 add(extracted_biased_exp
, exp
)));
512 /* Test if result is ±0.0, subnormal, or underflow by checking if the
513 * resulting biased exponent would be less than 0x1. If so, the result is
514 * 0.0 with the sign of x. (Actually, invert the conditions so that
515 * immediate values are the second arguments, which is better for i965)
516 * TODO: Implement in a vector fashion.
518 i
.insert_before(zero_sign_x
);
519 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
520 ir_variable
*unpacked
=
521 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
522 i
.insert_before(unpacked
);
525 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
526 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
528 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
529 i
.insert_before(assign(zero_sign_x
,
530 expr(ir_unop_pack_double_2x32
, unpacked
),
533 i
.insert_before(is_not_zero_or_underflow
);
534 i
.insert_before(assign(is_not_zero_or_underflow
,
535 gequal(resulting_biased_exp
,
536 new(ir
) ir_constant(0x1, vec_elem
))));
537 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
539 i
.insert_before(assign(resulting_biased_exp
,
540 csel(is_not_zero_or_underflow
,
541 resulting_biased_exp
, zeroi
)));
543 /* We could test for overflows by checking if the resulting biased exponent
544 * would be greater than 0xFE. Turns out we don't need to because the GLSL
547 * "If this product is too large to be represented in the
548 * floating-point type, the result is undefined."
551 ir_rvalue
*results
[4] = {NULL
};
552 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
553 ir_variable
*unpacked
=
554 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
555 i
.insert_before(unpacked
);
558 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
560 ir_expression
*bfi
= bitfield_insert(
562 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
563 exp_shift
->clone(ir
, NULL
),
564 exp_width
->clone(ir
, NULL
));
566 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
568 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
571 ir
->operation
= ir_quadop_vector
;
572 ir
->operands
[0] = results
[0];
573 ir
->operands
[1] = results
[1];
574 ir
->operands
[2] = results
[2];
575 ir
->operands
[3] = results
[3];
577 /* Don't generate new IR that would need to be lowered in an additional
581 this->progress
= true;
585 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
587 const unsigned vec_elem
= ir
->type
->vector_elements
;
588 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
590 /* Double-precision floating-point values are stored as
595 * We're just extracting the significand here, so we only need to modify
596 * the upper 32-bit uint. Unfortunately we must extract each double
597 * independently as there is no vector version of unpackDouble.
600 ir_instruction
&i
= *base_ir
;
602 ir_variable
*is_not_zero
=
603 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
604 ir_rvalue
*results
[4] = {NULL
};
606 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
607 i
.insert_before(is_not_zero
);
610 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
612 /* TODO: Remake this as more vector-friendly when int64 support is
615 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
616 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
617 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
619 /* Exponent of double floating-point values in the range [0.5, 1.0). */
620 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
623 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
624 ir_variable
*unpacked
=
625 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
627 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
629 i
.insert_before(bits
);
630 i
.insert_before(unpacked
);
631 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
633 /* Manipulate the high uint to remove the exponent and replace it with
634 * either the default exponent or zero.
636 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
637 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
638 i
.insert_before(assign(bits
, bit_or(bits
,
639 csel(swizzle(is_not_zero
, elem
, 1),
642 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
643 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
646 /* Put the dvec back together */
647 ir
->operation
= ir_quadop_vector
;
648 ir
->operands
[0] = results
[0];
649 ir
->operands
[1] = results
[1];
650 ir
->operands
[2] = results
[2];
651 ir
->operands
[3] = results
[3];
653 this->progress
= true;
657 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
659 const unsigned vec_elem
= ir
->type
->vector_elements
;
660 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
661 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
663 /* Double-precision floating-point values are stored as
668 * We're just extracting the exponent here, so we only care about the upper
672 ir_instruction
&i
= *base_ir
;
674 ir_variable
*is_not_zero
=
675 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
676 ir_variable
*high_words
=
677 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
678 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
679 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
681 ir_rvalue
*absval
= abs(ir
->operands
[0]);
683 i
.insert_before(is_not_zero
);
684 i
.insert_before(high_words
);
685 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
687 /* Extract all of the upper uints. */
688 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
689 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
691 i
.insert_before(assign(high_words
,
692 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
696 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
697 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
699 /* For non-zero inputs, shift the exponent down and apply bias. */
700 ir
->operation
= ir_triop_csel
;
701 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
702 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
703 ir
->operands
[2] = izero
;
705 this->progress
= true;
709 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
714 * sum = ir_binop_add x y
715 * bcarry = ir_binop_less sum x
716 * carry = ir_unop_b2i bcarry
719 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
720 ir
->operation
= ir_unop_i2u
;
721 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
722 ir
->operands
[1] = NULL
;
724 this->progress
= true;
728 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
731 * ir_binop_borrow x y
733 * bcarry = ir_binop_less x y
734 * carry = ir_unop_b2i bcarry
737 ir
->operation
= ir_unop_i2u
;
738 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
739 ir
->operands
[1] = NULL
;
741 this->progress
= true;
745 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
750 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
753 ir
->operation
= ir_binop_min
;
754 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
756 new(ir
) ir_constant(0.0f
));
757 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
759 this->progress
= true;
763 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
765 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
767 this->base_ir
->insert_before(temp
);
769 int nc
= ir
->operands
[0]->type
->components();
770 for (int i
= nc
- 1; i
>= 1; i
--) {
771 ir_assignment
*assig
;
773 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
774 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
776 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
777 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
780 this->base_ir
->insert_before(assig
);
783 ir
->operation
= ir_triop_fma
;
784 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
785 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
786 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
788 this->progress
= true;
793 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
796 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
797 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
799 switch (op2
->type
->vector_elements
) {
801 swizval
= SWIZZLE_XXXX
;
804 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
805 swizval
= SWIZZLE_XYZW
;
809 ir
->operation
= ir_triop_fma
;
810 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
811 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
813 this->progress
= true;
817 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
821 * temp = sub(x, frtemp);
822 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
824 ir_instruction
&i
= *base_ir
;
825 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
826 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
827 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
830 i
.insert_before(frtemp
);
831 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
833 ir
->operation
= ir_binop_add
;
834 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
835 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
837 this->progress
= true;
841 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
845 * result = sub(x, frtemp);
847 ir
->operation
= ir_binop_sub
;
848 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
850 this->progress
= true;
853 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
858 * frtemp = frac(temp);
859 * t2 = sub(temp, frtemp);
860 * if (frac(x) == 0.5)
861 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
866 ir_instruction
&i
= *base_ir
;
867 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
869 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
871 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
873 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
874 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
875 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
877 i
.insert_before(temp
);
878 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
880 i
.insert_before(frtemp
);
881 i
.insert_before(assign(frtemp
, fract(temp
)));
884 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
886 ir
->operation
= ir_triop_csel
;
887 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
888 p5
->clone(ir
, NULL
));
889 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
893 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
895 this->progress
= true;
899 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
903 * temp = sub(x, frtemp);
904 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
906 ir_rvalue
*arg
= ir
->operands
[0];
907 ir_instruction
&i
= *base_ir
;
909 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
910 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
911 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
913 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
916 i
.insert_before(frtemp
);
917 i
.insert_before(assign(frtemp
, fract(arg
)));
918 i
.insert_before(temp
);
919 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
921 ir
->operation
= ir_triop_csel
;
922 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
923 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
924 ir
->operands
[2] = add(temp
,
925 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
926 zero
->clone(ir
, NULL
),
929 this->progress
= true;
933 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
936 * temp = x > 0.0 ? 1.0 : 0.0;
937 * result = x < 0.0 ? -1.0 : temp;
939 ir_rvalue
*arg
= ir
->operands
[0];
940 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
941 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
942 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
944 ir
->operation
= ir_triop_csel
;
945 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
946 zero
->clone(ir
, NULL
));
947 ir
->operands
[1] = neg_one
;
948 ir
->operands
[2] = csel(greater(arg
, zero
),
950 zero
->clone(ir
, NULL
));
952 this->progress
= true;
956 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
958 switch (ir
->operation
) {
960 if (ir
->operands
[0]->type
->is_double())
961 double_dot_to_fma(ir
);
964 if (ir
->operands
[0]->type
->is_double())
968 if (lowering(SUB_TO_ADD_NEG
))
973 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
974 int_div_to_mul_rcp(ir
);
975 else if ((ir
->operands
[1]->type
->is_float() ||
976 ir
->operands
[1]->type
->is_double()) && lowering(DIV_TO_MUL_RCP
))
981 if (lowering(EXP_TO_EXP2
))
986 if (lowering(LOG_TO_LOG2
))
991 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
996 if (lowering(POW_TO_EXP2
))
1000 case ir_binop_ldexp
:
1001 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1003 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1004 dldexp_to_arith(ir
);
1007 case ir_unop_frexp_exp
:
1008 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1009 dfrexp_exp_to_arith(ir
);
1012 case ir_unop_frexp_sig
:
1013 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1014 dfrexp_sig_to_arith(ir
);
1017 case ir_binop_carry
:
1018 if (lowering(CARRY_TO_ARITH
))
1022 case ir_binop_borrow
:
1023 if (lowering(BORROW_TO_ARITH
))
1024 borrow_to_arith(ir
);
1027 case ir_unop_saturate
:
1028 if (lowering(SAT_TO_CLAMP
))
1033 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1034 dtrunc_to_dfrac(ir
);
1038 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1043 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1044 dfloor_to_dfrac(ir
);
1047 case ir_unop_round_even
:
1048 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1049 dround_even_to_dfrac(ir
);
1053 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1057 return visit_continue
;
1060 return visit_continue
;