2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file lower_instructions.cpp
27 * Many GPUs lack native instructions for certain expression operations, and
28 * must replace them with some other expression tree. This pass lowers some
29 * of the most common cases, allowing the lowering code to be implemented once
30 * rather than in each driver backend.
32 * Currently supported transformations:
35 * - INT_DIV_TO_MUL_RCP
49 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
51 * This simplifies expression reassociation, and for many backends
52 * there is no subtract operation separate from adding the negation.
53 * For backends with native subtract operations, they will probably
54 * want to recognize add(op0, neg(op1)) or the other way around to
55 * produce a subtract anyway.
57 * DIV_TO_MUL_RCP and INT_DIV_TO_MUL_RCP:
58 * --------------------------------------
59 * Breaks an ir_binop_div expression down to op0 * (rcp(op1)).
61 * Many GPUs don't have a divide instruction (945 and 965 included),
62 * but they do have an RCP instruction to compute an approximate
63 * reciprocal. By breaking the operation down, constant reciprocals
64 * can get constant folded.
66 * DIV_TO_MUL_RCP only lowers floating point division; INT_DIV_TO_MUL_RCP
67 * handles the integer case, converting to and from floating point so that
70 * EXP_TO_EXP2 and LOG_TO_LOG2:
71 * ----------------------------
72 * Many GPUs don't have a base e log or exponent instruction, but they
73 * do have base 2 versions, so this pass converts exp and log to exp2
74 * and log2 operations.
78 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
79 * x**y to 2**(y * log2(x)).
83 * Breaks an ir_binop_mod expression down to (op0 - op1 * floor(op0 / op1))
85 * Many GPUs don't have a MOD instruction (945 and 965 included), and
86 * if we have to break it down like this anyway, it gives an
87 * opportunity to do things like constant fold the (1.0 / op1) easily.
89 * Note: before we used to implement this as op1 * fract(op / op1) but this
90 * implementation had significant precision errors.
94 * Converts ir_binop_ldexp to arithmetic and bit operations for float sources.
96 * DFREXP_DLDEXP_TO_ARITH:
98 * Converts ir_binop_ldexp, ir_unop_frexp_sig, and ir_unop_frexp_exp to
99 * arithmetic and bit ops for double arguments.
103 * Converts ir_carry into (x + y) < x.
107 * Converts ir_borrow into (x < y).
111 * Converts ir_unop_saturate into min(max(x, 0.0), 1.0)
115 * Converts double trunc, ceil, floor, round to fract
118 #include "c99_math.h"
119 #include "program/prog_instruction.h" /* for swizzle */
120 #include "compiler/glsl_types.h"
122 #include "ir_builder.h"
123 #include "ir_optimization.h"
125 using namespace ir_builder
;
129 class lower_instructions_visitor
: public ir_hierarchical_visitor
{
131 lower_instructions_visitor(unsigned lower
)
132 : progress(false), lower(lower
) { }
134 ir_visitor_status
visit_leave(ir_expression
*);
139 unsigned lower
; /** Bitfield of which operations to lower */
141 void sub_to_add_neg(ir_expression
*);
142 void div_to_mul_rcp(ir_expression
*);
143 void int_div_to_mul_rcp(ir_expression
*);
144 void mod_to_floor(ir_expression
*);
145 void exp_to_exp2(ir_expression
*);
146 void pow_to_exp2(ir_expression
*);
147 void log_to_log2(ir_expression
*);
148 void ldexp_to_arith(ir_expression
*);
149 void dldexp_to_arith(ir_expression
*);
150 void dfrexp_sig_to_arith(ir_expression
*);
151 void dfrexp_exp_to_arith(ir_expression
*);
152 void carry_to_arith(ir_expression
*);
153 void borrow_to_arith(ir_expression
*);
154 void sat_to_clamp(ir_expression
*);
155 void double_dot_to_fma(ir_expression
*);
156 void double_lrp(ir_expression
*);
157 void dceil_to_dfrac(ir_expression
*);
158 void dfloor_to_dfrac(ir_expression
*);
159 void dround_even_to_dfrac(ir_expression
*);
160 void dtrunc_to_dfrac(ir_expression
*);
161 void dsign_to_csel(ir_expression
*);
162 void bit_count_to_math(ir_expression
*);
163 void extract_to_shifts(ir_expression
*);
166 } /* anonymous namespace */
169 * Determine if a particular type of lowering should occur
171 #define lowering(x) (this->lower & x)
174 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
176 lower_instructions_visitor
v(what_to_lower
);
178 visit_list_elements(&v
, instructions
);
183 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
185 ir
->operation
= ir_binop_add
;
186 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
187 ir
->operands
[1], NULL
);
188 this->progress
= true;
192 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
194 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
196 /* New expression for the 1.0 / op1 */
198 expr
= new(ir
) ir_expression(ir_unop_rcp
,
199 ir
->operands
[1]->type
,
202 /* op0 / op1 -> op0 * (1.0 / op1) */
203 ir
->operation
= ir_binop_mul
;
204 ir
->operands
[1] = expr
;
206 this->progress
= true;
210 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
212 assert(ir
->operands
[1]->type
->is_integer());
214 /* Be careful with integer division -- we need to do it as a
215 * float and re-truncate, since rcp(n > 1) of an integer would
218 ir_rvalue
*op0
, *op1
;
219 const struct glsl_type
*vec_type
;
221 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
222 ir
->operands
[1]->type
->vector_elements
,
223 ir
->operands
[1]->type
->matrix_columns
);
225 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
226 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
228 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
230 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
232 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
233 ir
->operands
[0]->type
->vector_elements
,
234 ir
->operands
[0]->type
->matrix_columns
);
236 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
237 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
239 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
241 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
242 ir
->type
->vector_elements
,
243 ir
->type
->matrix_columns
);
245 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
247 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
248 ir
->operation
= ir_unop_f2i
;
249 ir
->operands
[0] = op0
;
251 ir
->operation
= ir_unop_i2u
;
252 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
254 ir
->operands
[1] = NULL
;
256 this->progress
= true;
260 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
262 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
264 ir
->operation
= ir_unop_exp2
;
265 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
266 ir
->operands
[0], log2_e
);
267 this->progress
= true;
271 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
273 ir_expression
*const log2_x
=
274 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
277 ir
->operation
= ir_unop_exp2
;
278 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
279 ir
->operands
[1], log2_x
);
280 ir
->operands
[1] = NULL
;
281 this->progress
= true;
285 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
287 ir
->operation
= ir_binop_mul
;
288 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
289 ir
->operands
[0], NULL
);
290 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
291 this->progress
= true;
295 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
297 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
299 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
301 this->base_ir
->insert_before(x
);
302 this->base_ir
->insert_before(y
);
304 ir_assignment
*const assign_x
=
305 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
306 ir
->operands
[0], NULL
);
307 ir_assignment
*const assign_y
=
308 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
309 ir
->operands
[1], NULL
);
311 this->base_ir
->insert_before(assign_x
);
312 this->base_ir
->insert_before(assign_y
);
314 ir_expression
*const div_expr
=
315 new(ir
) ir_expression(ir_binop_div
, x
->type
,
316 new(ir
) ir_dereference_variable(x
),
317 new(ir
) ir_dereference_variable(y
));
319 /* Don't generate new IR that would need to be lowered in an additional
322 if (lowering(DIV_TO_MUL_RCP
) && (ir
->type
->is_float() || ir
->type
->is_double()))
323 div_to_mul_rcp(div_expr
);
325 ir_expression
*const floor_expr
=
326 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
328 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
329 dfloor_to_dfrac(floor_expr
);
331 ir_expression
*const mul_expr
=
332 new(ir
) ir_expression(ir_binop_mul
,
333 new(ir
) ir_dereference_variable(y
),
336 ir
->operation
= ir_binop_sub
;
337 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
338 ir
->operands
[1] = mul_expr
;
339 this->progress
= true;
343 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
346 * ir_binop_ldexp x exp
349 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
350 * resulting_biased_exp = extracted_biased_exp + exp;
352 * if (resulting_biased_exp < 1 || x == 0.0f) {
353 * return copysign(0.0, x);
356 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
357 * lshift(i2u(resulting_biased_exp), exp_shift));
359 * which we can't actually implement as such, since the GLSL IR doesn't
360 * have vectorized if-statements. We actually implement it without branches
361 * using conditional-select:
363 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
364 * resulting_biased_exp = extracted_biased_exp + exp;
366 * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
367 * gequal(resulting_biased_exp, 1);
368 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
369 * resulting_biased_exp = csel(is_not_zero_or_underflow,
370 * resulting_biased_exp, 0);
372 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
373 * lshift(i2u(resulting_biased_exp), exp_shift));
376 const unsigned vec_elem
= ir
->type
->vector_elements
;
379 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
380 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
383 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
385 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
387 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
388 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
390 /* Temporary variables */
391 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
392 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
394 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
397 ir_variable
*extracted_biased_exp
=
398 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
399 ir_variable
*resulting_biased_exp
=
400 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
402 ir_variable
*is_not_zero_or_underflow
=
403 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
405 ir_instruction
&i
= *base_ir
;
407 /* Copy <x> and <exp> arguments. */
409 i
.insert_before(assign(x
, ir
->operands
[0]));
410 i
.insert_before(exp
);
411 i
.insert_before(assign(exp
, ir
->operands
[1]));
413 /* Extract the biased exponent from <x>. */
414 i
.insert_before(extracted_biased_exp
);
415 i
.insert_before(assign(extracted_biased_exp
,
416 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
418 i
.insert_before(resulting_biased_exp
);
419 i
.insert_before(assign(resulting_biased_exp
,
420 add(extracted_biased_exp
, exp
)));
422 /* Test if result is ±0.0, subnormal, or underflow by checking if the
423 * resulting biased exponent would be less than 0x1. If so, the result is
424 * 0.0 with the sign of x. (Actually, invert the conditions so that
425 * immediate values are the second arguments, which is better for i965)
427 i
.insert_before(zero_sign_x
);
428 i
.insert_before(assign(zero_sign_x
,
429 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
431 i
.insert_before(is_not_zero_or_underflow
);
432 i
.insert_before(assign(is_not_zero_or_underflow
,
433 logic_and(nequal(x
, new(ir
) ir_constant(0.0f
, vec_elem
)),
434 gequal(resulting_biased_exp
,
435 new(ir
) ir_constant(0x1, vec_elem
)))));
436 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
438 i
.insert_before(assign(resulting_biased_exp
,
439 csel(is_not_zero_or_underflow
,
440 resulting_biased_exp
, zeroi
)));
442 /* We could test for overflows by checking if the resulting biased exponent
443 * would be greater than 0xFE. Turns out we don't need to because the GLSL
446 * "If this product is too large to be represented in the
447 * floating-point type, the result is undefined."
450 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
451 ir
->operation
= ir_unop_bitcast_i2f
;
452 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
453 exp_shift_clone
, exp_width
);
454 ir
->operands
[1] = NULL
;
456 this->progress
= true;
460 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
462 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
463 * from the significand.
466 const unsigned vec_elem
= ir
->type
->vector_elements
;
469 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
470 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
473 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
475 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
477 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
478 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
479 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
481 /* Temporary variables */
482 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
483 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
485 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
488 ir_variable
*extracted_biased_exp
=
489 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
490 ir_variable
*resulting_biased_exp
=
491 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
493 ir_variable
*is_not_zero_or_underflow
=
494 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
496 ir_instruction
&i
= *base_ir
;
498 /* Copy <x> and <exp> arguments. */
500 i
.insert_before(assign(x
, ir
->operands
[0]));
501 i
.insert_before(exp
);
502 i
.insert_before(assign(exp
, ir
->operands
[1]));
504 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
505 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
506 dfrexp_exp_to_arith(frexp_exp
);
508 /* Extract the biased exponent from <x>. */
509 i
.insert_before(extracted_biased_exp
);
510 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
512 i
.insert_before(resulting_biased_exp
);
513 i
.insert_before(assign(resulting_biased_exp
,
514 add(extracted_biased_exp
, exp
)));
516 /* Test if result is ±0.0, subnormal, or underflow by checking if the
517 * resulting biased exponent would be less than 0x1. If so, the result is
518 * 0.0 with the sign of x. (Actually, invert the conditions so that
519 * immediate values are the second arguments, which is better for i965)
520 * TODO: Implement in a vector fashion.
522 i
.insert_before(zero_sign_x
);
523 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
524 ir_variable
*unpacked
=
525 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
526 i
.insert_before(unpacked
);
529 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
530 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
532 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
533 i
.insert_before(assign(zero_sign_x
,
534 expr(ir_unop_pack_double_2x32
, unpacked
),
537 i
.insert_before(is_not_zero_or_underflow
);
538 i
.insert_before(assign(is_not_zero_or_underflow
,
539 gequal(resulting_biased_exp
,
540 new(ir
) ir_constant(0x1, vec_elem
))));
541 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
543 i
.insert_before(assign(resulting_biased_exp
,
544 csel(is_not_zero_or_underflow
,
545 resulting_biased_exp
, zeroi
)));
547 /* We could test for overflows by checking if the resulting biased exponent
548 * would be greater than 0xFE. Turns out we don't need to because the GLSL
551 * "If this product is too large to be represented in the
552 * floating-point type, the result is undefined."
555 ir_rvalue
*results
[4] = {NULL
};
556 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
557 ir_variable
*unpacked
=
558 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
559 i
.insert_before(unpacked
);
562 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
564 ir_expression
*bfi
= bitfield_insert(
566 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
567 exp_shift
->clone(ir
, NULL
),
568 exp_width
->clone(ir
, NULL
));
570 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
572 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
575 ir
->operation
= ir_quadop_vector
;
576 ir
->operands
[0] = results
[0];
577 ir
->operands
[1] = results
[1];
578 ir
->operands
[2] = results
[2];
579 ir
->operands
[3] = results
[3];
581 /* Don't generate new IR that would need to be lowered in an additional
585 this->progress
= true;
589 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
591 const unsigned vec_elem
= ir
->type
->vector_elements
;
592 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
594 /* Double-precision floating-point values are stored as
599 * We're just extracting the significand here, so we only need to modify
600 * the upper 32-bit uint. Unfortunately we must extract each double
601 * independently as there is no vector version of unpackDouble.
604 ir_instruction
&i
= *base_ir
;
606 ir_variable
*is_not_zero
=
607 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
608 ir_rvalue
*results
[4] = {NULL
};
610 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
611 i
.insert_before(is_not_zero
);
614 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
616 /* TODO: Remake this as more vector-friendly when int64 support is
619 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
620 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
621 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
623 /* Exponent of double floating-point values in the range [0.5, 1.0). */
624 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
627 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
628 ir_variable
*unpacked
=
629 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
631 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
633 i
.insert_before(bits
);
634 i
.insert_before(unpacked
);
635 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
637 /* Manipulate the high uint to remove the exponent and replace it with
638 * either the default exponent or zero.
640 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
641 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
642 i
.insert_before(assign(bits
, bit_or(bits
,
643 csel(swizzle(is_not_zero
, elem
, 1),
646 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
647 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
650 /* Put the dvec back together */
651 ir
->operation
= ir_quadop_vector
;
652 ir
->operands
[0] = results
[0];
653 ir
->operands
[1] = results
[1];
654 ir
->operands
[2] = results
[2];
655 ir
->operands
[3] = results
[3];
657 this->progress
= true;
661 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
663 const unsigned vec_elem
= ir
->type
->vector_elements
;
664 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
665 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
667 /* Double-precision floating-point values are stored as
672 * We're just extracting the exponent here, so we only care about the upper
676 ir_instruction
&i
= *base_ir
;
678 ir_variable
*is_not_zero
=
679 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
680 ir_variable
*high_words
=
681 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
682 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
683 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
685 ir_rvalue
*absval
= abs(ir
->operands
[0]);
687 i
.insert_before(is_not_zero
);
688 i
.insert_before(high_words
);
689 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
691 /* Extract all of the upper uints. */
692 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
693 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
695 i
.insert_before(assign(high_words
,
696 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
700 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
701 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
703 /* For non-zero inputs, shift the exponent down and apply bias. */
704 ir
->operation
= ir_triop_csel
;
705 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
706 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
707 ir
->operands
[2] = izero
;
709 this->progress
= true;
713 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
718 * sum = ir_binop_add x y
719 * bcarry = ir_binop_less sum x
720 * carry = ir_unop_b2i bcarry
723 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
724 ir
->operation
= ir_unop_i2u
;
725 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
726 ir
->operands
[1] = NULL
;
728 this->progress
= true;
732 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
735 * ir_binop_borrow x y
737 * bcarry = ir_binop_less x y
738 * carry = ir_unop_b2i bcarry
741 ir
->operation
= ir_unop_i2u
;
742 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
743 ir
->operands
[1] = NULL
;
745 this->progress
= true;
749 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
754 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
757 ir
->operation
= ir_binop_min
;
758 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
760 new(ir
) ir_constant(0.0f
));
761 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
763 this->progress
= true;
767 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
769 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
771 this->base_ir
->insert_before(temp
);
773 int nc
= ir
->operands
[0]->type
->components();
774 for (int i
= nc
- 1; i
>= 1; i
--) {
775 ir_assignment
*assig
;
777 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
778 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
780 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
781 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
784 this->base_ir
->insert_before(assig
);
787 ir
->operation
= ir_triop_fma
;
788 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
789 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
790 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
792 this->progress
= true;
797 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
800 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
801 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
803 switch (op2
->type
->vector_elements
) {
805 swizval
= SWIZZLE_XXXX
;
808 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
809 swizval
= SWIZZLE_XYZW
;
813 ir
->operation
= ir_triop_fma
;
814 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
815 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
817 this->progress
= true;
821 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
825 * temp = sub(x, frtemp);
826 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
828 ir_instruction
&i
= *base_ir
;
829 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
830 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
831 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
834 i
.insert_before(frtemp
);
835 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
837 ir
->operation
= ir_binop_add
;
838 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
839 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
841 this->progress
= true;
845 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
849 * result = sub(x, frtemp);
851 ir
->operation
= ir_binop_sub
;
852 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
854 this->progress
= true;
857 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
862 * frtemp = frac(temp);
863 * t2 = sub(temp, frtemp);
864 * if (frac(x) == 0.5)
865 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
870 ir_instruction
&i
= *base_ir
;
871 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
873 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
875 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
877 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
878 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
879 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
881 i
.insert_before(temp
);
882 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
884 i
.insert_before(frtemp
);
885 i
.insert_before(assign(frtemp
, fract(temp
)));
888 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
890 ir
->operation
= ir_triop_csel
;
891 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
892 p5
->clone(ir
, NULL
));
893 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
897 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
899 this->progress
= true;
903 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
907 * temp = sub(x, frtemp);
908 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
910 ir_rvalue
*arg
= ir
->operands
[0];
911 ir_instruction
&i
= *base_ir
;
913 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
914 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
915 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
917 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
920 i
.insert_before(frtemp
);
921 i
.insert_before(assign(frtemp
, fract(arg
)));
922 i
.insert_before(temp
);
923 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
925 ir
->operation
= ir_triop_csel
;
926 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
927 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
928 ir
->operands
[2] = add(temp
,
929 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
930 zero
->clone(ir
, NULL
),
933 this->progress
= true;
937 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
940 * temp = x > 0.0 ? 1.0 : 0.0;
941 * result = x < 0.0 ? -1.0 : temp;
943 ir_rvalue
*arg
= ir
->operands
[0];
944 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
945 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
946 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
948 ir
->operation
= ir_triop_csel
;
949 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
950 zero
->clone(ir
, NULL
));
951 ir
->operands
[1] = neg_one
;
952 ir
->operands
[2] = csel(greater(arg
, zero
),
954 zero
->clone(ir
, NULL
));
956 this->progress
= true;
960 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
962 /* For more details, see:
964 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
966 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
967 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
969 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
970 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
971 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
972 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
973 ir_constant
*c1
= new(ir
) ir_constant(1u);
974 ir_constant
*c2
= new(ir
) ir_constant(2u);
975 ir_constant
*c4
= new(ir
) ir_constant(4u);
976 ir_constant
*c24
= new(ir
) ir_constant(24u);
978 base_ir
->insert_before(temp
);
980 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
981 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
983 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
984 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
987 /* temp = temp - ((temp >> 1) & 0x55555555u); */
988 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
991 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
992 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
993 bit_and(rshift(temp
, c2
),
994 c33333333
->clone(ir
, NULL
)))));
996 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
997 ir
->operation
= ir_unop_u2i
;
998 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1002 this->progress
= true;
1006 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1009 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1011 base_ir
->insert_before(bits
);
1012 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1014 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1016 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1018 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1019 ir_constant
*cFFFFFFFF
=
1020 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1022 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1023 * we'd get a mask of 0 when bits is 32. Special case it.
1025 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1027 ir_expression
*mask
= csel(equal(bits
, c32
),
1029 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1031 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1033 * If bits is zero, the result will be zero.
1035 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1036 * select as in the signed integer case.
1038 * (value >> offset) & mask;
1040 ir
->operation
= ir_binop_bit_and
;
1041 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1042 ir
->operands
[1] = mask
;
1043 ir
->operands
[2] = NULL
;
1046 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1048 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1050 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1052 /* temp = 32 - bits; */
1053 base_ir
->insert_before(temp
);
1054 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1056 /* expr = value << (temp - offset)) >> temp; */
1057 ir_expression
*expr
=
1058 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1060 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1062 * If bits is zero, the result will be zero.
1064 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1065 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1068 * (bits == 0) ? 0 : e;
1070 ir
->operation
= ir_triop_csel
;
1071 ir
->operands
[0] = equal(c0
, bits
);
1072 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1073 ir
->operands
[2] = expr
;
1076 this->progress
= true;
1080 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1082 switch (ir
->operation
) {
1084 if (ir
->operands
[0]->type
->is_double())
1085 double_dot_to_fma(ir
);
1088 if (ir
->operands
[0]->type
->is_double())
1092 if (lowering(SUB_TO_ADD_NEG
))
1097 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1098 int_div_to_mul_rcp(ir
);
1099 else if ((ir
->operands
[1]->type
->is_float() ||
1100 ir
->operands
[1]->type
->is_double()) && lowering(DIV_TO_MUL_RCP
))
1105 if (lowering(EXP_TO_EXP2
))
1110 if (lowering(LOG_TO_LOG2
))
1115 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1120 if (lowering(POW_TO_EXP2
))
1124 case ir_binop_ldexp
:
1125 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1127 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1128 dldexp_to_arith(ir
);
1131 case ir_unop_frexp_exp
:
1132 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1133 dfrexp_exp_to_arith(ir
);
1136 case ir_unop_frexp_sig
:
1137 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1138 dfrexp_sig_to_arith(ir
);
1141 case ir_binop_carry
:
1142 if (lowering(CARRY_TO_ARITH
))
1146 case ir_binop_borrow
:
1147 if (lowering(BORROW_TO_ARITH
))
1148 borrow_to_arith(ir
);
1151 case ir_unop_saturate
:
1152 if (lowering(SAT_TO_CLAMP
))
1157 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1158 dtrunc_to_dfrac(ir
);
1162 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1167 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1168 dfloor_to_dfrac(ir
);
1171 case ir_unop_round_even
:
1172 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1173 dround_even_to_dfrac(ir
);
1177 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1181 case ir_unop_bit_count
:
1182 if (lowering(BIT_COUNT_TO_MATH
))
1183 bit_count_to_math(ir
);
1186 case ir_triop_bitfield_extract
:
1187 if (lowering(EXTRACT_TO_SHIFTS
))
1188 extract_to_shifts(ir
);
1192 return visit_continue
;
1195 return visit_continue
;