2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file lower_instructions.cpp
27 * Many GPUs lack native instructions for certain expression operations, and
28 * must replace them with some other expression tree. This pass lowers some
29 * of the most common cases, allowing the lowering code to be implemented once
30 * rather than in each driver backend.
32 * Currently supported transformations:
35 * - INT_DIV_TO_MUL_RCP
49 * Breaks an ir_binop_sub expression down to add(op0, neg(op1))
51 * This simplifies expression reassociation, and for many backends
52 * there is no subtract operation separate from adding the negation.
53 * For backends with native subtract operations, they will probably
54 * want to recognize add(op0, neg(op1)) or the other way around to
55 * produce a subtract anyway.
57 * DIV_TO_MUL_RCP and INT_DIV_TO_MUL_RCP:
58 * --------------------------------------
59 * Breaks an ir_binop_div expression down to op0 * (rcp(op1)).
61 * Many GPUs don't have a divide instruction (945 and 965 included),
62 * but they do have an RCP instruction to compute an approximate
63 * reciprocal. By breaking the operation down, constant reciprocals
64 * can get constant folded.
66 * DIV_TO_MUL_RCP only lowers floating point division; INT_DIV_TO_MUL_RCP
67 * handles the integer case, converting to and from floating point so that
70 * EXP_TO_EXP2 and LOG_TO_LOG2:
71 * ----------------------------
72 * Many GPUs don't have a base e log or exponent instruction, but they
73 * do have base 2 versions, so this pass converts exp and log to exp2
74 * and log2 operations.
78 * Many older GPUs don't have an x**y instruction. For these GPUs, convert
79 * x**y to 2**(y * log2(x)).
83 * Breaks an ir_binop_mod expression down to (op0 - op1 * floor(op0 / op1))
85 * Many GPUs don't have a MOD instruction (945 and 965 included), and
86 * if we have to break it down like this anyway, it gives an
87 * opportunity to do things like constant fold the (1.0 / op1) easily.
89 * Note: before we used to implement this as op1 * fract(op / op1) but this
90 * implementation had significant precision errors.
94 * Converts ir_binop_ldexp to arithmetic and bit operations for float sources.
96 * DFREXP_DLDEXP_TO_ARITH:
98 * Converts ir_binop_ldexp, ir_unop_frexp_sig, and ir_unop_frexp_exp to
99 * arithmetic and bit ops for double arguments.
103 * Converts ir_carry into (x + y) < x.
107 * Converts ir_borrow into (x < y).
111 * Converts ir_unop_saturate into min(max(x, 0.0), 1.0)
115 * Converts double trunc, ceil, floor, round to fract
118 #include "c99_math.h"
119 #include "program/prog_instruction.h" /* for swizzle */
120 #include "compiler/glsl_types.h"
122 #include "ir_builder.h"
123 #include "ir_optimization.h"
125 using namespace ir_builder
;
129 class lower_instructions_visitor
: public ir_hierarchical_visitor
{
131 lower_instructions_visitor(unsigned lower
)
132 : progress(false), lower(lower
) { }
134 ir_visitor_status
visit_leave(ir_expression
*);
139 unsigned lower
; /** Bitfield of which operations to lower */
141 void sub_to_add_neg(ir_expression
*);
142 void div_to_mul_rcp(ir_expression
*);
143 void int_div_to_mul_rcp(ir_expression
*);
144 void mod_to_floor(ir_expression
*);
145 void exp_to_exp2(ir_expression
*);
146 void pow_to_exp2(ir_expression
*);
147 void log_to_log2(ir_expression
*);
148 void ldexp_to_arith(ir_expression
*);
149 void dldexp_to_arith(ir_expression
*);
150 void dfrexp_sig_to_arith(ir_expression
*);
151 void dfrexp_exp_to_arith(ir_expression
*);
152 void carry_to_arith(ir_expression
*);
153 void borrow_to_arith(ir_expression
*);
154 void sat_to_clamp(ir_expression
*);
155 void double_dot_to_fma(ir_expression
*);
156 void double_lrp(ir_expression
*);
157 void dceil_to_dfrac(ir_expression
*);
158 void dfloor_to_dfrac(ir_expression
*);
159 void dround_even_to_dfrac(ir_expression
*);
160 void dtrunc_to_dfrac(ir_expression
*);
161 void dsign_to_csel(ir_expression
*);
162 void bit_count_to_math(ir_expression
*);
163 void extract_to_shifts(ir_expression
*);
164 void insert_to_shifts(ir_expression
*);
167 } /* anonymous namespace */
170 * Determine if a particular type of lowering should occur
172 #define lowering(x) (this->lower & x)
175 lower_instructions(exec_list
*instructions
, unsigned what_to_lower
)
177 lower_instructions_visitor
v(what_to_lower
);
179 visit_list_elements(&v
, instructions
);
184 lower_instructions_visitor::sub_to_add_neg(ir_expression
*ir
)
186 ir
->operation
= ir_binop_add
;
187 ir
->operands
[1] = new(ir
) ir_expression(ir_unop_neg
, ir
->operands
[1]->type
,
188 ir
->operands
[1], NULL
);
189 this->progress
= true;
193 lower_instructions_visitor::div_to_mul_rcp(ir_expression
*ir
)
195 assert(ir
->operands
[1]->type
->is_float() || ir
->operands
[1]->type
->is_double());
197 /* New expression for the 1.0 / op1 */
199 expr
= new(ir
) ir_expression(ir_unop_rcp
,
200 ir
->operands
[1]->type
,
203 /* op0 / op1 -> op0 * (1.0 / op1) */
204 ir
->operation
= ir_binop_mul
;
205 ir
->operands
[1] = expr
;
207 this->progress
= true;
211 lower_instructions_visitor::int_div_to_mul_rcp(ir_expression
*ir
)
213 assert(ir
->operands
[1]->type
->is_integer());
215 /* Be careful with integer division -- we need to do it as a
216 * float and re-truncate, since rcp(n > 1) of an integer would
219 ir_rvalue
*op0
, *op1
;
220 const struct glsl_type
*vec_type
;
222 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
223 ir
->operands
[1]->type
->vector_elements
,
224 ir
->operands
[1]->type
->matrix_columns
);
226 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
)
227 op1
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[1], NULL
);
229 op1
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[1], NULL
);
231 op1
= new(ir
) ir_expression(ir_unop_rcp
, op1
->type
, op1
, NULL
);
233 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
234 ir
->operands
[0]->type
->vector_elements
,
235 ir
->operands
[0]->type
->matrix_columns
);
237 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
)
238 op0
= new(ir
) ir_expression(ir_unop_i2f
, vec_type
, ir
->operands
[0], NULL
);
240 op0
= new(ir
) ir_expression(ir_unop_u2f
, vec_type
, ir
->operands
[0], NULL
);
242 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
243 ir
->type
->vector_elements
,
244 ir
->type
->matrix_columns
);
246 op0
= new(ir
) ir_expression(ir_binop_mul
, vec_type
, op0
, op1
);
248 if (ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
) {
249 ir
->operation
= ir_unop_f2i
;
250 ir
->operands
[0] = op0
;
252 ir
->operation
= ir_unop_i2u
;
253 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_f2i
, op0
);
255 ir
->operands
[1] = NULL
;
257 this->progress
= true;
261 lower_instructions_visitor::exp_to_exp2(ir_expression
*ir
)
263 ir_constant
*log2_e
= new(ir
) ir_constant(float(M_LOG2E
));
265 ir
->operation
= ir_unop_exp2
;
266 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[0]->type
,
267 ir
->operands
[0], log2_e
);
268 this->progress
= true;
272 lower_instructions_visitor::pow_to_exp2(ir_expression
*ir
)
274 ir_expression
*const log2_x
=
275 new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
278 ir
->operation
= ir_unop_exp2
;
279 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_mul
, ir
->operands
[1]->type
,
280 ir
->operands
[1], log2_x
);
281 ir
->operands
[1] = NULL
;
282 this->progress
= true;
286 lower_instructions_visitor::log_to_log2(ir_expression
*ir
)
288 ir
->operation
= ir_binop_mul
;
289 ir
->operands
[0] = new(ir
) ir_expression(ir_unop_log2
, ir
->operands
[0]->type
,
290 ir
->operands
[0], NULL
);
291 ir
->operands
[1] = new(ir
) ir_constant(float(1.0 / M_LOG2E
));
292 this->progress
= true;
296 lower_instructions_visitor::mod_to_floor(ir_expression
*ir
)
298 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[0]->type
, "mod_x",
300 ir_variable
*y
= new(ir
) ir_variable(ir
->operands
[1]->type
, "mod_y",
302 this->base_ir
->insert_before(x
);
303 this->base_ir
->insert_before(y
);
305 ir_assignment
*const assign_x
=
306 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(x
),
307 ir
->operands
[0], NULL
);
308 ir_assignment
*const assign_y
=
309 new(ir
) ir_assignment(new(ir
) ir_dereference_variable(y
),
310 ir
->operands
[1], NULL
);
312 this->base_ir
->insert_before(assign_x
);
313 this->base_ir
->insert_before(assign_y
);
315 ir_expression
*const div_expr
=
316 new(ir
) ir_expression(ir_binop_div
, x
->type
,
317 new(ir
) ir_dereference_variable(x
),
318 new(ir
) ir_dereference_variable(y
));
320 /* Don't generate new IR that would need to be lowered in an additional
323 if (lowering(DIV_TO_MUL_RCP
) && (ir
->type
->is_float() || ir
->type
->is_double()))
324 div_to_mul_rcp(div_expr
);
326 ir_expression
*const floor_expr
=
327 new(ir
) ir_expression(ir_unop_floor
, x
->type
, div_expr
);
329 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
330 dfloor_to_dfrac(floor_expr
);
332 ir_expression
*const mul_expr
=
333 new(ir
) ir_expression(ir_binop_mul
,
334 new(ir
) ir_dereference_variable(y
),
337 ir
->operation
= ir_binop_sub
;
338 ir
->operands
[0] = new(ir
) ir_dereference_variable(x
);
339 ir
->operands
[1] = mul_expr
;
340 this->progress
= true;
344 lower_instructions_visitor::ldexp_to_arith(ir_expression
*ir
)
347 * ir_binop_ldexp x exp
350 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
351 * resulting_biased_exp = extracted_biased_exp + exp;
353 * if (resulting_biased_exp < 1 || x == 0.0f) {
354 * return copysign(0.0, x);
357 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
358 * lshift(i2u(resulting_biased_exp), exp_shift));
360 * which we can't actually implement as such, since the GLSL IR doesn't
361 * have vectorized if-statements. We actually implement it without branches
362 * using conditional-select:
364 * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
365 * resulting_biased_exp = extracted_biased_exp + exp;
367 * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
368 * gequal(resulting_biased_exp, 1);
369 * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
370 * resulting_biased_exp = csel(is_not_zero_or_underflow,
371 * resulting_biased_exp, 0);
373 * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
374 * lshift(i2u(resulting_biased_exp), exp_shift));
377 const unsigned vec_elem
= ir
->type
->vector_elements
;
380 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
381 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
384 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
386 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
, vec_elem
);
388 ir_constant
*exp_shift
= new(ir
) ir_constant(23, vec_elem
);
389 ir_constant
*exp_width
= new(ir
) ir_constant(8, vec_elem
);
391 /* Temporary variables */
392 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
393 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
395 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
398 ir_variable
*extracted_biased_exp
=
399 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
400 ir_variable
*resulting_biased_exp
=
401 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
403 ir_variable
*is_not_zero_or_underflow
=
404 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
406 ir_instruction
&i
= *base_ir
;
408 /* Copy <x> and <exp> arguments. */
410 i
.insert_before(assign(x
, ir
->operands
[0]));
411 i
.insert_before(exp
);
412 i
.insert_before(assign(exp
, ir
->operands
[1]));
414 /* Extract the biased exponent from <x>. */
415 i
.insert_before(extracted_biased_exp
);
416 i
.insert_before(assign(extracted_biased_exp
,
417 rshift(bitcast_f2i(abs(x
)), exp_shift
)));
419 i
.insert_before(resulting_biased_exp
);
420 i
.insert_before(assign(resulting_biased_exp
,
421 add(extracted_biased_exp
, exp
)));
423 /* Test if result is ±0.0, subnormal, or underflow by checking if the
424 * resulting biased exponent would be less than 0x1. If so, the result is
425 * 0.0 with the sign of x. (Actually, invert the conditions so that
426 * immediate values are the second arguments, which is better for i965)
428 i
.insert_before(zero_sign_x
);
429 i
.insert_before(assign(zero_sign_x
,
430 bitcast_u2f(bit_and(bitcast_f2u(x
), sign_mask
))));
432 i
.insert_before(is_not_zero_or_underflow
);
433 i
.insert_before(assign(is_not_zero_or_underflow
,
434 logic_and(nequal(x
, new(ir
) ir_constant(0.0f
, vec_elem
)),
435 gequal(resulting_biased_exp
,
436 new(ir
) ir_constant(0x1, vec_elem
)))));
437 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
439 i
.insert_before(assign(resulting_biased_exp
,
440 csel(is_not_zero_or_underflow
,
441 resulting_biased_exp
, zeroi
)));
443 /* We could test for overflows by checking if the resulting biased exponent
444 * would be greater than 0xFE. Turns out we don't need to because the GLSL
447 * "If this product is too large to be represented in the
448 * floating-point type, the result is undefined."
451 ir_constant
*exp_shift_clone
= exp_shift
->clone(ir
, NULL
);
452 ir
->operation
= ir_unop_bitcast_i2f
;
453 ir
->operands
[0] = bitfield_insert(bitcast_f2i(x
), resulting_biased_exp
,
454 exp_shift_clone
, exp_width
);
455 ir
->operands
[1] = NULL
;
457 this->progress
= true;
461 lower_instructions_visitor::dldexp_to_arith(ir_expression
*ir
)
463 /* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
464 * from the significand.
467 const unsigned vec_elem
= ir
->type
->vector_elements
;
470 const glsl_type
*ivec
= glsl_type::get_instance(GLSL_TYPE_INT
, vec_elem
, 1);
471 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
474 ir_constant
*zeroi
= ir_constant::zero(ir
, ivec
);
476 ir_constant
*sign_mask
= new(ir
) ir_constant(0x80000000u
);
478 ir_constant
*exp_shift
= new(ir
) ir_constant(20u);
479 ir_constant
*exp_width
= new(ir
) ir_constant(11u);
480 ir_constant
*exp_bias
= new(ir
) ir_constant(1022, vec_elem
);
482 /* Temporary variables */
483 ir_variable
*x
= new(ir
) ir_variable(ir
->type
, "x", ir_var_temporary
);
484 ir_variable
*exp
= new(ir
) ir_variable(ivec
, "exp", ir_var_temporary
);
486 ir_variable
*zero_sign_x
= new(ir
) ir_variable(ir
->type
, "zero_sign_x",
489 ir_variable
*extracted_biased_exp
=
490 new(ir
) ir_variable(ivec
, "extracted_biased_exp", ir_var_temporary
);
491 ir_variable
*resulting_biased_exp
=
492 new(ir
) ir_variable(ivec
, "resulting_biased_exp", ir_var_temporary
);
494 ir_variable
*is_not_zero_or_underflow
=
495 new(ir
) ir_variable(bvec
, "is_not_zero_or_underflow", ir_var_temporary
);
497 ir_instruction
&i
= *base_ir
;
499 /* Copy <x> and <exp> arguments. */
501 i
.insert_before(assign(x
, ir
->operands
[0]));
502 i
.insert_before(exp
);
503 i
.insert_before(assign(exp
, ir
->operands
[1]));
505 ir_expression
*frexp_exp
= expr(ir_unop_frexp_exp
, x
);
506 if (lowering(DFREXP_DLDEXP_TO_ARITH
))
507 dfrexp_exp_to_arith(frexp_exp
);
509 /* Extract the biased exponent from <x>. */
510 i
.insert_before(extracted_biased_exp
);
511 i
.insert_before(assign(extracted_biased_exp
, add(frexp_exp
, exp_bias
)));
513 i
.insert_before(resulting_biased_exp
);
514 i
.insert_before(assign(resulting_biased_exp
,
515 add(extracted_biased_exp
, exp
)));
517 /* Test if result is ±0.0, subnormal, or underflow by checking if the
518 * resulting biased exponent would be less than 0x1. If so, the result is
519 * 0.0 with the sign of x. (Actually, invert the conditions so that
520 * immediate values are the second arguments, which is better for i965)
521 * TODO: Implement in a vector fashion.
523 i
.insert_before(zero_sign_x
);
524 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
525 ir_variable
*unpacked
=
526 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
527 i
.insert_before(unpacked
);
530 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
531 i
.insert_before(assign(unpacked
, bit_and(swizzle_y(unpacked
), sign_mask
->clone(ir
, NULL
)),
533 i
.insert_before(assign(unpacked
, ir_constant::zero(ir
, glsl_type::uint_type
), WRITEMASK_X
));
534 i
.insert_before(assign(zero_sign_x
,
535 expr(ir_unop_pack_double_2x32
, unpacked
),
538 i
.insert_before(is_not_zero_or_underflow
);
539 i
.insert_before(assign(is_not_zero_or_underflow
,
540 gequal(resulting_biased_exp
,
541 new(ir
) ir_constant(0x1, vec_elem
))));
542 i
.insert_before(assign(x
, csel(is_not_zero_or_underflow
,
544 i
.insert_before(assign(resulting_biased_exp
,
545 csel(is_not_zero_or_underflow
,
546 resulting_biased_exp
, zeroi
)));
548 /* We could test for overflows by checking if the resulting biased exponent
549 * would be greater than 0xFE. Turns out we don't need to because the GLSL
552 * "If this product is too large to be represented in the
553 * floating-point type, the result is undefined."
556 ir_rvalue
*results
[4] = {NULL
};
557 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
558 ir_variable
*unpacked
=
559 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
560 i
.insert_before(unpacked
);
563 expr(ir_unop_unpack_double_2x32
, swizzle(x
, elem
, 1))));
565 ir_expression
*bfi
= bitfield_insert(
567 i2u(swizzle(resulting_biased_exp
, elem
, 1)),
568 exp_shift
->clone(ir
, NULL
),
569 exp_width
->clone(ir
, NULL
));
571 i
.insert_before(assign(unpacked
, bfi
, WRITEMASK_Y
));
573 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
576 ir
->operation
= ir_quadop_vector
;
577 ir
->operands
[0] = results
[0];
578 ir
->operands
[1] = results
[1];
579 ir
->operands
[2] = results
[2];
580 ir
->operands
[3] = results
[3];
582 /* Don't generate new IR that would need to be lowered in an additional
586 this->progress
= true;
590 lower_instructions_visitor::dfrexp_sig_to_arith(ir_expression
*ir
)
592 const unsigned vec_elem
= ir
->type
->vector_elements
;
593 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
595 /* Double-precision floating-point values are stored as
600 * We're just extracting the significand here, so we only need to modify
601 * the upper 32-bit uint. Unfortunately we must extract each double
602 * independently as there is no vector version of unpackDouble.
605 ir_instruction
&i
= *base_ir
;
607 ir_variable
*is_not_zero
=
608 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
609 ir_rvalue
*results
[4] = {NULL
};
611 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
612 i
.insert_before(is_not_zero
);
615 nequal(abs(ir
->operands
[0]->clone(ir
, NULL
)), dzero
)));
617 /* TODO: Remake this as more vector-friendly when int64 support is
620 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
621 ir_constant
*zero
= new(ir
) ir_constant(0u, 1);
622 ir_constant
*sign_mantissa_mask
= new(ir
) ir_constant(0x800fffffu
, 1);
624 /* Exponent of double floating-point values in the range [0.5, 1.0). */
625 ir_constant
*exponent_value
= new(ir
) ir_constant(0x3fe00000u
, 1);
628 new(ir
) ir_variable(glsl_type::uint_type
, "bits", ir_var_temporary
);
629 ir_variable
*unpacked
=
630 new(ir
) ir_variable(glsl_type::uvec2_type
, "unpacked", ir_var_temporary
);
632 ir_rvalue
*x
= swizzle(ir
->operands
[0]->clone(ir
, NULL
), elem
, 1);
634 i
.insert_before(bits
);
635 i
.insert_before(unpacked
);
636 i
.insert_before(assign(unpacked
, expr(ir_unop_unpack_double_2x32
, x
)));
638 /* Manipulate the high uint to remove the exponent and replace it with
639 * either the default exponent or zero.
641 i
.insert_before(assign(bits
, swizzle_y(unpacked
)));
642 i
.insert_before(assign(bits
, bit_and(bits
, sign_mantissa_mask
)));
643 i
.insert_before(assign(bits
, bit_or(bits
,
644 csel(swizzle(is_not_zero
, elem
, 1),
647 i
.insert_before(assign(unpacked
, bits
, WRITEMASK_Y
));
648 results
[elem
] = expr(ir_unop_pack_double_2x32
, unpacked
);
651 /* Put the dvec back together */
652 ir
->operation
= ir_quadop_vector
;
653 ir
->operands
[0] = results
[0];
654 ir
->operands
[1] = results
[1];
655 ir
->operands
[2] = results
[2];
656 ir
->operands
[3] = results
[3];
658 this->progress
= true;
662 lower_instructions_visitor::dfrexp_exp_to_arith(ir_expression
*ir
)
664 const unsigned vec_elem
= ir
->type
->vector_elements
;
665 const glsl_type
*bvec
= glsl_type::get_instance(GLSL_TYPE_BOOL
, vec_elem
, 1);
666 const glsl_type
*uvec
= glsl_type::get_instance(GLSL_TYPE_UINT
, vec_elem
, 1);
668 /* Double-precision floating-point values are stored as
673 * We're just extracting the exponent here, so we only care about the upper
677 ir_instruction
&i
= *base_ir
;
679 ir_variable
*is_not_zero
=
680 new(ir
) ir_variable(bvec
, "is_not_zero", ir_var_temporary
);
681 ir_variable
*high_words
=
682 new(ir
) ir_variable(uvec
, "high_words", ir_var_temporary
);
683 ir_constant
*dzero
= new(ir
) ir_constant(0.0, vec_elem
);
684 ir_constant
*izero
= new(ir
) ir_constant(0, vec_elem
);
686 ir_rvalue
*absval
= abs(ir
->operands
[0]);
688 i
.insert_before(is_not_zero
);
689 i
.insert_before(high_words
);
690 i
.insert_before(assign(is_not_zero
, nequal(absval
->clone(ir
, NULL
), dzero
)));
692 /* Extract all of the upper uints. */
693 for (unsigned elem
= 0; elem
< vec_elem
; elem
++) {
694 ir_rvalue
*x
= swizzle(absval
->clone(ir
, NULL
), elem
, 1);
696 i
.insert_before(assign(high_words
,
697 swizzle_y(expr(ir_unop_unpack_double_2x32
, x
)),
701 ir_constant
*exponent_shift
= new(ir
) ir_constant(20, vec_elem
);
702 ir_constant
*exponent_bias
= new(ir
) ir_constant(-1022, vec_elem
);
704 /* For non-zero inputs, shift the exponent down and apply bias. */
705 ir
->operation
= ir_triop_csel
;
706 ir
->operands
[0] = new(ir
) ir_dereference_variable(is_not_zero
);
707 ir
->operands
[1] = add(exponent_bias
, u2i(rshift(high_words
, exponent_shift
)));
708 ir
->operands
[2] = izero
;
710 this->progress
= true;
714 lower_instructions_visitor::carry_to_arith(ir_expression
*ir
)
719 * sum = ir_binop_add x y
720 * bcarry = ir_binop_less sum x
721 * carry = ir_unop_b2i bcarry
724 ir_rvalue
*x_clone
= ir
->operands
[0]->clone(ir
, NULL
);
725 ir
->operation
= ir_unop_i2u
;
726 ir
->operands
[0] = b2i(less(add(ir
->operands
[0], ir
->operands
[1]), x_clone
));
727 ir
->operands
[1] = NULL
;
729 this->progress
= true;
733 lower_instructions_visitor::borrow_to_arith(ir_expression
*ir
)
736 * ir_binop_borrow x y
738 * bcarry = ir_binop_less x y
739 * carry = ir_unop_b2i bcarry
742 ir
->operation
= ir_unop_i2u
;
743 ir
->operands
[0] = b2i(less(ir
->operands
[0], ir
->operands
[1]));
744 ir
->operands
[1] = NULL
;
746 this->progress
= true;
750 lower_instructions_visitor::sat_to_clamp(ir_expression
*ir
)
755 * ir_binop_min (ir_binop_max(x, 0.0), 1.0)
758 ir
->operation
= ir_binop_min
;
759 ir
->operands
[0] = new(ir
) ir_expression(ir_binop_max
, ir
->operands
[0]->type
,
761 new(ir
) ir_constant(0.0f
));
762 ir
->operands
[1] = new(ir
) ir_constant(1.0f
);
764 this->progress
= true;
768 lower_instructions_visitor::double_dot_to_fma(ir_expression
*ir
)
770 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
->get_base_type(), "dot_res",
772 this->base_ir
->insert_before(temp
);
774 int nc
= ir
->operands
[0]->type
->components();
775 for (int i
= nc
- 1; i
>= 1; i
--) {
776 ir_assignment
*assig
;
778 assig
= assign(temp
, mul(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
779 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1)));
781 assig
= assign(temp
, fma(swizzle(ir
->operands
[0]->clone(ir
, NULL
), i
, 1),
782 swizzle(ir
->operands
[1]->clone(ir
, NULL
), i
, 1),
785 this->base_ir
->insert_before(assig
);
788 ir
->operation
= ir_triop_fma
;
789 ir
->operands
[0] = swizzle(ir
->operands
[0], 0, 1);
790 ir
->operands
[1] = swizzle(ir
->operands
[1], 0, 1);
791 ir
->operands
[2] = new(ir
) ir_dereference_variable(temp
);
793 this->progress
= true;
798 lower_instructions_visitor::double_lrp(ir_expression
*ir
)
801 ir_rvalue
*op0
= ir
->operands
[0], *op2
= ir
->operands
[2];
802 ir_constant
*one
= new(ir
) ir_constant(1.0, op2
->type
->vector_elements
);
804 switch (op2
->type
->vector_elements
) {
806 swizval
= SWIZZLE_XXXX
;
809 assert(op0
->type
->vector_elements
== op2
->type
->vector_elements
);
810 swizval
= SWIZZLE_XYZW
;
814 ir
->operation
= ir_triop_fma
;
815 ir
->operands
[0] = swizzle(op2
, swizval
, op0
->type
->vector_elements
);
816 ir
->operands
[2] = mul(sub(one
, op2
->clone(ir
, NULL
)), op0
);
818 this->progress
= true;
822 lower_instructions_visitor::dceil_to_dfrac(ir_expression
*ir
)
826 * temp = sub(x, frtemp);
827 * result = temp + ((frtemp != 0.0) ? 1.0 : 0.0);
829 ir_instruction
&i
= *base_ir
;
830 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
831 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
832 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
835 i
.insert_before(frtemp
);
836 i
.insert_before(assign(frtemp
, fract(ir
->operands
[0])));
838 ir
->operation
= ir_binop_add
;
839 ir
->operands
[0] = sub(ir
->operands
[0]->clone(ir
, NULL
), frtemp
);
840 ir
->operands
[1] = csel(nequal(frtemp
, zero
), one
, zero
->clone(ir
, NULL
));
842 this->progress
= true;
846 lower_instructions_visitor::dfloor_to_dfrac(ir_expression
*ir
)
850 * result = sub(x, frtemp);
852 ir
->operation
= ir_binop_sub
;
853 ir
->operands
[1] = fract(ir
->operands
[0]->clone(ir
, NULL
));
855 this->progress
= true;
858 lower_instructions_visitor::dround_even_to_dfrac(ir_expression
*ir
)
863 * frtemp = frac(temp);
864 * t2 = sub(temp, frtemp);
865 * if (frac(x) == 0.5)
866 * result = frac(t2 * 0.5) == 0 ? t2 : t2 - 1;
871 ir_instruction
&i
= *base_ir
;
872 ir_variable
*frtemp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "frtemp",
874 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
876 ir_variable
*t2
= new(ir
) ir_variable(ir
->operands
[0]->type
, "t2",
878 ir_constant
*p5
= new(ir
) ir_constant(0.5, ir
->operands
[0]->type
->vector_elements
);
879 ir_constant
*one
= new(ir
) ir_constant(1.0, ir
->operands
[0]->type
->vector_elements
);
880 ir_constant
*zero
= new(ir
) ir_constant(0.0, ir
->operands
[0]->type
->vector_elements
);
882 i
.insert_before(temp
);
883 i
.insert_before(assign(temp
, add(ir
->operands
[0], p5
)));
885 i
.insert_before(frtemp
);
886 i
.insert_before(assign(frtemp
, fract(temp
)));
889 i
.insert_before(assign(t2
, sub(temp
, frtemp
)));
891 ir
->operation
= ir_triop_csel
;
892 ir
->operands
[0] = equal(fract(ir
->operands
[0]->clone(ir
, NULL
)),
893 p5
->clone(ir
, NULL
));
894 ir
->operands
[1] = csel(equal(fract(mul(t2
, p5
->clone(ir
, NULL
))),
898 ir
->operands
[2] = new(ir
) ir_dereference_variable(t2
);
900 this->progress
= true;
904 lower_instructions_visitor::dtrunc_to_dfrac(ir_expression
*ir
)
908 * temp = sub(x, frtemp);
909 * result = x >= 0 ? temp : temp + (frtemp == 0.0) ? 0 : 1;
911 ir_rvalue
*arg
= ir
->operands
[0];
912 ir_instruction
&i
= *base_ir
;
914 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
915 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
916 ir_variable
*frtemp
= new(ir
) ir_variable(arg
->type
, "frtemp",
918 ir_variable
*temp
= new(ir
) ir_variable(ir
->operands
[0]->type
, "temp",
921 i
.insert_before(frtemp
);
922 i
.insert_before(assign(frtemp
, fract(arg
)));
923 i
.insert_before(temp
);
924 i
.insert_before(assign(temp
, sub(arg
->clone(ir
, NULL
), frtemp
)));
926 ir
->operation
= ir_triop_csel
;
927 ir
->operands
[0] = gequal(arg
->clone(ir
, NULL
), zero
);
928 ir
->operands
[1] = new (ir
) ir_dereference_variable(temp
);
929 ir
->operands
[2] = add(temp
,
930 csel(equal(frtemp
, zero
->clone(ir
, NULL
)),
931 zero
->clone(ir
, NULL
),
934 this->progress
= true;
938 lower_instructions_visitor::dsign_to_csel(ir_expression
*ir
)
941 * temp = x > 0.0 ? 1.0 : 0.0;
942 * result = x < 0.0 ? -1.0 : temp;
944 ir_rvalue
*arg
= ir
->operands
[0];
945 ir_constant
*zero
= new(ir
) ir_constant(0.0, arg
->type
->vector_elements
);
946 ir_constant
*one
= new(ir
) ir_constant(1.0, arg
->type
->vector_elements
);
947 ir_constant
*neg_one
= new(ir
) ir_constant(-1.0, arg
->type
->vector_elements
);
949 ir
->operation
= ir_triop_csel
;
950 ir
->operands
[0] = less(arg
->clone(ir
, NULL
),
951 zero
->clone(ir
, NULL
));
952 ir
->operands
[1] = neg_one
;
953 ir
->operands
[2] = csel(greater(arg
, zero
),
955 zero
->clone(ir
, NULL
));
957 this->progress
= true;
961 lower_instructions_visitor::bit_count_to_math(ir_expression
*ir
)
963 /* For more details, see:
965 * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetPaallel
967 const unsigned elements
= ir
->operands
[0]->type
->vector_elements
;
968 ir_variable
*temp
= new(ir
) ir_variable(glsl_type::uvec(elements
), "temp",
970 ir_constant
*c55555555
= new(ir
) ir_constant(0x55555555u
);
971 ir_constant
*c33333333
= new(ir
) ir_constant(0x33333333u
);
972 ir_constant
*c0F0F0F0F
= new(ir
) ir_constant(0x0F0F0F0Fu
);
973 ir_constant
*c01010101
= new(ir
) ir_constant(0x01010101u
);
974 ir_constant
*c1
= new(ir
) ir_constant(1u);
975 ir_constant
*c2
= new(ir
) ir_constant(2u);
976 ir_constant
*c4
= new(ir
) ir_constant(4u);
977 ir_constant
*c24
= new(ir
) ir_constant(24u);
979 base_ir
->insert_before(temp
);
981 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
982 base_ir
->insert_before(assign(temp
, ir
->operands
[0]));
984 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
);
985 base_ir
->insert_before(assign(temp
, i2u(ir
->operands
[0])));
988 /* temp = temp - ((temp >> 1) & 0x55555555u); */
989 base_ir
->insert_before(assign(temp
, sub(temp
, bit_and(rshift(temp
, c1
),
992 /* temp = (temp & 0x33333333u) + ((temp >> 2) & 0x33333333u); */
993 base_ir
->insert_before(assign(temp
, add(bit_and(temp
, c33333333
),
994 bit_and(rshift(temp
, c2
),
995 c33333333
->clone(ir
, NULL
)))));
997 /* int(((temp + (temp >> 4) & 0xF0F0F0Fu) * 0x1010101u) >> 24); */
998 ir
->operation
= ir_unop_u2i
;
999 ir
->operands
[0] = rshift(mul(bit_and(add(temp
, rshift(temp
, c4
)), c0F0F0F0F
),
1003 this->progress
= true;
1007 lower_instructions_visitor::extract_to_shifts(ir_expression
*ir
)
1010 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1012 base_ir
->insert_before(bits
);
1013 base_ir
->insert_before(assign(bits
, ir
->operands
[2]));
1015 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
) {
1017 new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1019 new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1020 ir_constant
*cFFFFFFFF
=
1021 new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1023 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1024 * we'd get a mask of 0 when bits is 32. Special case it.
1026 * mask = bits == 32 ? 0xffffffff : (1u << bits) - 1u;
1028 ir_expression
*mask
= csel(equal(bits
, c32
),
1030 sub(lshift(c1
, bits
), c1
->clone(ir
, NULL
)));
1032 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1034 * If bits is zero, the result will be zero.
1036 * Since (1 << 0) - 1 == 0, we don't need to bother with the conditional
1037 * select as in the signed integer case.
1039 * (value >> offset) & mask;
1041 ir
->operation
= ir_binop_bit_and
;
1042 ir
->operands
[0] = rshift(ir
->operands
[0], ir
->operands
[1]);
1043 ir
->operands
[1] = mask
;
1044 ir
->operands
[2] = NULL
;
1047 new(ir
) ir_constant(int(0), ir
->operands
[0]->type
->vector_elements
);
1049 new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1051 new(ir
) ir_variable(ir
->operands
[0]->type
, "temp", ir_var_temporary
);
1053 /* temp = 32 - bits; */
1054 base_ir
->insert_before(temp
);
1055 base_ir
->insert_before(assign(temp
, sub(c32
, bits
)));
1057 /* expr = value << (temp - offset)) >> temp; */
1058 ir_expression
*expr
=
1059 rshift(lshift(ir
->operands
[0], sub(temp
, ir
->operands
[1])), temp
);
1061 /* Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1063 * If bits is zero, the result will be zero.
1065 * Due to the (x << (y%32)) behavior mentioned before, the (value <<
1066 * (32-0)) doesn't "erase" all of the data as we would like, so finish
1069 * (bits == 0) ? 0 : e;
1071 ir
->operation
= ir_triop_csel
;
1072 ir
->operands
[0] = equal(c0
, bits
);
1073 ir
->operands
[1] = c0
->clone(ir
, NULL
);
1074 ir
->operands
[2] = expr
;
1077 this->progress
= true;
1081 lower_instructions_visitor::insert_to_shifts(ir_expression
*ir
)
1085 ir_constant
*cFFFFFFFF
;
1086 ir_variable
*offset
=
1087 new(ir
) ir_variable(ir
->operands
[0]->type
, "offset", ir_var_temporary
);
1089 new(ir
) ir_variable(ir
->operands
[0]->type
, "bits", ir_var_temporary
);
1091 new(ir
) ir_variable(ir
->operands
[0]->type
, "mask", ir_var_temporary
);
1093 if (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
) {
1094 c1
= new(ir
) ir_constant(int(1), ir
->operands
[0]->type
->vector_elements
);
1095 c32
= new(ir
) ir_constant(int(32), ir
->operands
[0]->type
->vector_elements
);
1096 cFFFFFFFF
= new(ir
) ir_constant(int(0xFFFFFFFF), ir
->operands
[0]->type
->vector_elements
);
1098 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
);
1100 c1
= new(ir
) ir_constant(1u, ir
->operands
[0]->type
->vector_elements
);
1101 c32
= new(ir
) ir_constant(32u, ir
->operands
[0]->type
->vector_elements
);
1102 cFFFFFFFF
= new(ir
) ir_constant(0xFFFFFFFFu
, ir
->operands
[0]->type
->vector_elements
);
1105 base_ir
->insert_before(offset
);
1106 base_ir
->insert_before(assign(offset
, ir
->operands
[2]));
1108 base_ir
->insert_before(bits
);
1109 base_ir
->insert_before(assign(bits
, ir
->operands
[3]));
1111 /* At least some hardware treats (x << y) as (x << (y%32)). This means
1112 * we'd get a mask of 0 when bits is 32. Special case it.
1114 * mask = (bits == 32 ? 0xffffffff : (1u << bits) - 1u) << offset;
1116 * Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
1118 * The result will be undefined if offset or bits is negative, or if the
1119 * sum of offset and bits is greater than the number of bits used to
1120 * store the operand.
1122 * Since it's undefined, there are a couple other ways this could be
1123 * implemented. The other way that was considered was to put the csel
1124 * around the whole thing:
1126 * final_result = bits == 32 ? insert : ... ;
1128 base_ir
->insert_before(mask
);
1130 base_ir
->insert_before(assign(mask
, csel(equal(bits
, c32
),
1132 lshift(sub(lshift(c1
, bits
),
1133 c1
->clone(ir
, NULL
)),
1136 /* (base & ~mask) | ((insert << offset) & mask) */
1137 ir
->operation
= ir_binop_bit_or
;
1138 ir
->operands
[0] = bit_and(ir
->operands
[0], bit_not(mask
));
1139 ir
->operands
[1] = bit_and(lshift(ir
->operands
[1], offset
), mask
);
1140 ir
->operands
[2] = NULL
;
1141 ir
->operands
[3] = NULL
;
1143 this->progress
= true;
1147 lower_instructions_visitor::visit_leave(ir_expression
*ir
)
1149 switch (ir
->operation
) {
1151 if (ir
->operands
[0]->type
->is_double())
1152 double_dot_to_fma(ir
);
1155 if (ir
->operands
[0]->type
->is_double())
1159 if (lowering(SUB_TO_ADD_NEG
))
1164 if (ir
->operands
[1]->type
->is_integer() && lowering(INT_DIV_TO_MUL_RCP
))
1165 int_div_to_mul_rcp(ir
);
1166 else if ((ir
->operands
[1]->type
->is_float() ||
1167 ir
->operands
[1]->type
->is_double()) && lowering(DIV_TO_MUL_RCP
))
1172 if (lowering(EXP_TO_EXP2
))
1177 if (lowering(LOG_TO_LOG2
))
1182 if (lowering(MOD_TO_FLOOR
) && (ir
->type
->is_float() || ir
->type
->is_double()))
1187 if (lowering(POW_TO_EXP2
))
1191 case ir_binop_ldexp
:
1192 if (lowering(LDEXP_TO_ARITH
) && ir
->type
->is_float())
1194 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->type
->is_double())
1195 dldexp_to_arith(ir
);
1198 case ir_unop_frexp_exp
:
1199 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1200 dfrexp_exp_to_arith(ir
);
1203 case ir_unop_frexp_sig
:
1204 if (lowering(DFREXP_DLDEXP_TO_ARITH
) && ir
->operands
[0]->type
->is_double())
1205 dfrexp_sig_to_arith(ir
);
1208 case ir_binop_carry
:
1209 if (lowering(CARRY_TO_ARITH
))
1213 case ir_binop_borrow
:
1214 if (lowering(BORROW_TO_ARITH
))
1215 borrow_to_arith(ir
);
1218 case ir_unop_saturate
:
1219 if (lowering(SAT_TO_CLAMP
))
1224 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1225 dtrunc_to_dfrac(ir
);
1229 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1234 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1235 dfloor_to_dfrac(ir
);
1238 case ir_unop_round_even
:
1239 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1240 dround_even_to_dfrac(ir
);
1244 if (lowering(DOPS_TO_DFRAC
) && ir
->type
->is_double())
1248 case ir_unop_bit_count
:
1249 if (lowering(BIT_COUNT_TO_MATH
))
1250 bit_count_to_math(ir
);
1253 case ir_triop_bitfield_extract
:
1254 if (lowering(EXTRACT_TO_SHIFTS
))
1255 extract_to_shifts(ir
);
1258 case ir_quadop_bitfield_insert
:
1259 if (lowering(INSERT_TO_SHIFTS
))
1260 insert_to_shifts(ir
);
1265 return visit_continue
;
1268 return visit_continue
;