1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
|
------------------------------------------------------------------------------
-- --
-- GNAT COMPILER COMPONENTS --
-- --
-- S Y S T E M . F A T _ G E N --
-- --
-- B o d y --
-- --
-- --
-- Copyright (C) 1992-2001 Free Software Foundation, Inc. --
-- --
-- GNAT is free software; you can redistribute it and/or modify it under --
-- terms of the GNU General Public License as published by the Free Soft- --
-- ware Foundation; either version 2, or (at your option) any later ver- --
-- sion. GNAT is distributed in the hope that it will be useful, but WITH- --
-- OUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY --
-- or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License --
-- for more details. You should have received a copy of the GNU General --
-- Public License distributed with GNAT; see file COPYING. If not, write --
-- to the Free Software Foundation, 59 Temple Place - Suite 330, Boston, --
-- MA 02111-1307, USA. --
-- --
-- As a special exception, if other files instantiate generics from this --
-- unit, or you link this unit with other files to produce an executable, --
-- this unit does not by itself cause the resulting executable to be --
-- covered by the GNU General Public License. This exception does not --
-- however invalidate any other reasons why the executable file might be --
-- covered by the GNU Public License. --
-- --
-- GNAT was originally developed by the GNAT team at New York University. --
-- Extensive contributions were provided by Ada Core Technologies Inc. --
-- --
------------------------------------------------------------------------------
-- The implementation here is portable to any IEEE implementation. It does
-- not handle non-binary radix, and also assumes that model numbers and
-- machine numbers are basically identical, which is not true of all possible
-- floating-point implementations. On a non-IEEE machine, this body must be
-- specialized appropriately, or better still, its generic instantiations
-- should be replaced by efficient machine-specific code.
with Ada.Unchecked_Conversion;
with System;
package body System.Fat_Gen is
Float_Radix : constant T := T (T'Machine_Radix);
Float_Radix_Inv : constant T := 1.0 / Float_Radix;
Radix_To_M_Minus_1 : constant T := Float_Radix ** (T'Machine_Mantissa - 1);
pragma Assert (T'Machine_Radix = 2);
-- This version does not handle radix 16
-- Constants for Decompose and Scaling
Rad : constant T := T (T'Machine_Radix);
Invrad : constant T := 1.0 / Rad;
subtype Expbits is Integer range 0 .. 6;
-- 2 ** (2 ** 7) might overflow. how big can radix-16 exponents get?
Log_Power : constant array (Expbits) of Integer := (1, 2, 4, 8, 16, 32, 64);
R_Power : constant array (Expbits) of T :=
(Rad ** 1,
Rad ** 2,
Rad ** 4,
Rad ** 8,
Rad ** 16,
Rad ** 32,
Rad ** 64);
R_Neg_Power : constant array (Expbits) of T :=
(Invrad ** 1,
Invrad ** 2,
Invrad ** 4,
Invrad ** 8,
Invrad ** 16,
Invrad ** 32,
Invrad ** 64);
-----------------------
-- Local Subprograms --
-----------------------
procedure Decompose (XX : T; Frac : out T; Expo : out UI);
-- Decomposes a floating-point number into fraction and exponent parts
function Gradual_Scaling (Adjustment : UI) return T;
-- Like Scaling with a first argument of 1.0, but returns the smallest
-- denormal rather than zero when the adjustment is smaller than
-- Machine_Emin. Used for Succ and Pred.
--------------
-- Adjacent --
--------------
function Adjacent (X, Towards : T) return T is
begin
if Towards = X then
return X;
elsif Towards > X then
return Succ (X);
else
return Pred (X);
end if;
end Adjacent;
-------------
-- Ceiling --
-------------
function Ceiling (X : T) return T is
XT : constant T := Truncation (X);
begin
if X <= 0.0 then
return XT;
elsif X = XT then
return X;
else
return XT + 1.0;
end if;
end Ceiling;
-------------
-- Compose --
-------------
function Compose (Fraction : T; Exponent : UI) return T is
Arg_Frac : T;
Arg_Exp : UI;
begin
Decompose (Fraction, Arg_Frac, Arg_Exp);
return Scaling (Arg_Frac, Exponent);
end Compose;
---------------
-- Copy_Sign --
---------------
function Copy_Sign (Value, Sign : T) return T is
Result : T;
function Is_Negative (V : T) return Boolean;
pragma Import (Intrinsic, Is_Negative);
begin
Result := abs Value;
if Is_Negative (Sign) then
return -Result;
else
return Result;
end if;
end Copy_Sign;
---------------
-- Decompose --
---------------
procedure Decompose (XX : T; Frac : out T; Expo : out UI) is
X : T := T'Machine (XX);
begin
if X = 0.0 then
Frac := X;
Expo := 0;
-- More useful would be defining Expo to be T'Machine_Emin - 1 or
-- T'Machine_Emin - T'Machine_Mantissa, which would preserve
-- monotonicity of the exponent function ???
-- Check for infinities, transfinites, whatnot.
elsif X > T'Safe_Last then
Frac := Invrad;
Expo := T'Machine_Emax + 1;
elsif X < T'Safe_First then
Frac := -Invrad;
Expo := T'Machine_Emax + 2; -- how many extra negative values?
else
-- Case of nonzero finite x. Essentially, we just multiply
-- by Rad ** (+-2**N) to reduce the range.
declare
Ax : T := abs X;
Ex : UI := 0;
-- Ax * Rad ** Ex is invariant.
begin
if Ax >= 1.0 then
while Ax >= R_Power (Expbits'Last) loop
Ax := Ax * R_Neg_Power (Expbits'Last);
Ex := Ex + Log_Power (Expbits'Last);
end loop;
-- Ax < Rad ** 64
for N in reverse Expbits'First .. Expbits'Last - 1 loop
if Ax >= R_Power (N) then
Ax := Ax * R_Neg_Power (N);
Ex := Ex + Log_Power (N);
end if;
-- Ax < R_Power (N)
end loop;
-- 1 <= Ax < Rad
Ax := Ax * Invrad;
Ex := Ex + 1;
else
-- 0 < ax < 1
while Ax < R_Neg_Power (Expbits'Last) loop
Ax := Ax * R_Power (Expbits'Last);
Ex := Ex - Log_Power (Expbits'Last);
end loop;
-- Rad ** -64 <= Ax < 1
for N in reverse Expbits'First .. Expbits'Last - 1 loop
if Ax < R_Neg_Power (N) then
Ax := Ax * R_Power (N);
Ex := Ex - Log_Power (N);
end if;
-- R_Neg_Power (N) <= Ax < 1
end loop;
end if;
if X > 0.0 then
Frac := Ax;
else
Frac := -Ax;
end if;
Expo := Ex;
end;
end if;
end Decompose;
--------------
-- Exponent --
--------------
function Exponent (X : T) return UI is
X_Frac : T;
X_Exp : UI;
begin
Decompose (X, X_Frac, X_Exp);
return X_Exp;
end Exponent;
-----------
-- Floor --
-----------
function Floor (X : T) return T is
XT : constant T := Truncation (X);
begin
if X >= 0.0 then
return XT;
elsif XT = X then
return X;
else
return XT - 1.0;
end if;
end Floor;
--------------
-- Fraction --
--------------
function Fraction (X : T) return T is
X_Frac : T;
X_Exp : UI;
begin
Decompose (X, X_Frac, X_Exp);
return X_Frac;
end Fraction;
---------------------
-- Gradual_Scaling --
---------------------
function Gradual_Scaling (Adjustment : UI) return T is
Y : T;
Y1 : T;
Ex : UI := Adjustment;
begin
if Adjustment < T'Machine_Emin then
Y := 2.0 ** T'Machine_Emin;
Y1 := Y;
Ex := Ex - T'Machine_Emin;
while Ex <= 0 loop
Y := T'Machine (Y / 2.0);
if Y = 0.0 then
return Y1;
end if;
Ex := Ex + 1;
Y1 := Y;
end loop;
return Y1;
else
return Scaling (1.0, Adjustment);
end if;
end Gradual_Scaling;
------------------
-- Leading_Part --
------------------
function Leading_Part (X : T; Radix_Digits : UI) return T is
L : UI;
Y, Z : T;
begin
if Radix_Digits >= T'Machine_Mantissa then
return X;
else
L := Exponent (X) - Radix_Digits;
Y := Truncation (Scaling (X, -L));
Z := Scaling (Y, L);
return Z;
end if;
end Leading_Part;
-------------
-- Machine --
-------------
-- The trick with Machine is to force the compiler to store the result
-- in memory so that we do not have extra precision used. The compiler
-- is clever, so we have to outwit its possible optimizations! We do
-- this by using an intermediate pragma Volatile location.
function Machine (X : T) return T is
Temp : T;
pragma Volatile (Temp);
begin
Temp := X;
return Temp;
end Machine;
-----------
-- Model --
-----------
-- We treat Model as identical to Machine. This is true of IEEE and other
-- nice floating-point systems, but not necessarily true of all systems.
function Model (X : T) return T is
begin
return Machine (X);
end Model;
----------
-- Pred --
----------
-- Subtract from the given number a number equivalent to the value of its
-- least significant bit. Given that the most significant bit represents
-- a value of 1.0 * radix ** (exp - 1), the value we want is obtained by
-- shifting this by (mantissa-1) bits to the right, i.e. decreasing the
-- exponent by that amount.
-- Zero has to be treated specially, since its exponent is zero
function Pred (X : T) return T is
X_Frac : T;
X_Exp : UI;
begin
if X = 0.0 then
return -Succ (X);
else
Decompose (X, X_Frac, X_Exp);
-- A special case, if the number we had was a positive power of
-- two, then we want to subtract half of what we would otherwise
-- subtract, since the exponent is going to be reduced.
if X_Frac = 0.5 and then X > 0.0 then
return X - Gradual_Scaling (X_Exp - T'Machine_Mantissa - 1);
-- Otherwise the exponent stays the same
else
return X - Gradual_Scaling (X_Exp - T'Machine_Mantissa);
end if;
end if;
end Pred;
---------------
-- Remainder --
---------------
function Remainder (X, Y : T) return T is
A : T;
B : T;
Arg : T;
P : T;
Arg_Frac : T;
P_Frac : T;
Sign_X : T;
IEEE_Rem : T;
Arg_Exp : UI;
P_Exp : UI;
K : UI;
P_Even : Boolean;
begin
if X > 0.0 then
Sign_X := 1.0;
Arg := X;
else
Sign_X := -1.0;
Arg := -X;
end if;
P := abs Y;
if Arg < P then
P_Even := True;
IEEE_Rem := Arg;
P_Exp := Exponent (P);
else
Decompose (Arg, Arg_Frac, Arg_Exp);
Decompose (P, P_Frac, P_Exp);
P := Compose (P_Frac, Arg_Exp);
K := Arg_Exp - P_Exp;
P_Even := True;
IEEE_Rem := Arg;
for Cnt in reverse 0 .. K loop
if IEEE_Rem >= P then
P_Even := False;
IEEE_Rem := IEEE_Rem - P;
else
P_Even := True;
end if;
P := P * 0.5;
end loop;
end if;
-- That completes the calculation of modulus remainder. The final
-- step is get the IEEE remainder. Here we need to compare Rem with
-- (abs Y) / 2. We must be careful of unrepresentable Y/2 value
-- caused by subnormal numbers
if P_Exp >= 0 then
A := IEEE_Rem;
B := abs Y * 0.5;
else
A := IEEE_Rem * 2.0;
B := abs Y;
end if;
if A > B or else (A = B and then not P_Even) then
IEEE_Rem := IEEE_Rem - abs Y;
end if;
return Sign_X * IEEE_Rem;
end Remainder;
--------------
-- Rounding --
--------------
function Rounding (X : T) return T is
Result : T;
Tail : T;
begin
Result := Truncation (abs X);
Tail := abs X - Result;
if Tail >= 0.5 then
Result := Result + 1.0;
end if;
if X > 0.0 then
return Result;
elsif X < 0.0 then
return -Result;
-- For zero case, make sure sign of zero is preserved
else
return X;
end if;
end Rounding;
-------------
-- Scaling --
-------------
-- Return x * rad ** adjustment quickly,
-- or quietly underflow to zero, or overflow naturally.
function Scaling (X : T; Adjustment : UI) return T is
begin
if X = 0.0 or else Adjustment = 0 then
return X;
end if;
-- Nonzero x. essentially, just multiply repeatedly by Rad ** (+-2**n).
declare
Y : T := X;
Ex : UI := Adjustment;
-- Y * Rad ** Ex is invariant
begin
if Ex < 0 then
while Ex <= -Log_Power (Expbits'Last) loop
Y := Y * R_Neg_Power (Expbits'Last);
Ex := Ex + Log_Power (Expbits'Last);
end loop;
-- -64 < Ex <= 0
for N in reverse Expbits'First .. Expbits'Last - 1 loop
if Ex <= -Log_Power (N) then
Y := Y * R_Neg_Power (N);
Ex := Ex + Log_Power (N);
end if;
-- -Log_Power (N) < Ex <= 0
end loop;
-- Ex = 0
else
-- Ex >= 0
while Ex >= Log_Power (Expbits'Last) loop
Y := Y * R_Power (Expbits'Last);
Ex := Ex - Log_Power (Expbits'Last);
end loop;
-- 0 <= Ex < 64
for N in reverse Expbits'First .. Expbits'Last - 1 loop
if Ex >= Log_Power (N) then
Y := Y * R_Power (N);
Ex := Ex - Log_Power (N);
end if;
-- 0 <= Ex < Log_Power (N)
end loop;
-- Ex = 0
end if;
return Y;
end;
end Scaling;
----------
-- Succ --
----------
-- Similar computation to that of Pred: find value of least significant
-- bit of given number, and add. Zero has to be treated specially since
-- the exponent can be zero, and also we want the smallest denormal if
-- denormals are supported.
function Succ (X : T) return T is
X_Frac : T;
X_Exp : UI;
X1, X2 : T;
begin
if X = 0.0 then
X1 := 2.0 ** T'Machine_Emin;
-- Following loop generates smallest denormal
loop
X2 := T'Machine (X1 / 2.0);
exit when X2 = 0.0;
X1 := X2;
end loop;
return X1;
else
Decompose (X, X_Frac, X_Exp);
-- A special case, if the number we had was a negative power of
-- two, then we want to add half of what we would otherwise add,
-- since the exponent is going to be reduced.
if X_Frac = 0.5 and then X < 0.0 then
return X + Gradual_Scaling (X_Exp - T'Machine_Mantissa - 1);
-- Otherwise the exponent stays the same
else
return X + Gradual_Scaling (X_Exp - T'Machine_Mantissa);
end if;
end if;
end Succ;
----------------
-- Truncation --
----------------
-- The basic approach is to compute
-- T'Machine (RM1 + N) - RM1.
-- where N >= 0.0 and RM1 = radix ** (mantissa - 1)
-- This works provided that the intermediate result (RM1 + N) does not
-- have extra precision (which is why we call Machine). When we compute
-- RM1 + N, the exponent of N will be normalized and the mantissa shifted
-- shifted appropriately so the lower order bits, which cannot contribute
-- to the integer part of N, fall off on the right. When we subtract RM1
-- again, the significant bits of N are shifted to the left, and what we
-- have is an integer, because only the first e bits are different from
-- zero (assuming binary radix here).
function Truncation (X : T) return T is
Result : T;
begin
Result := abs X;
if Result >= Radix_To_M_Minus_1 then
return Machine (X);
else
Result := Machine (Radix_To_M_Minus_1 + Result) - Radix_To_M_Minus_1;
if Result > abs X then
Result := Result - 1.0;
end if;
if X > 0.0 then
return Result;
elsif X < 0.0 then
return -Result;
-- For zero case, make sure sign of zero is preserved
else
return X;
end if;
end if;
end Truncation;
-----------------------
-- Unbiased_Rounding --
-----------------------
function Unbiased_Rounding (X : T) return T is
Abs_X : constant T := abs X;
Result : T;
Tail : T;
begin
Result := Truncation (Abs_X);
Tail := Abs_X - Result;
if Tail > 0.5 then
Result := Result + 1.0;
elsif Tail = 0.5 then
Result := 2.0 * Truncation ((Result / 2.0) + 0.5);
end if;
if X > 0.0 then
return Result;
elsif X < 0.0 then
return -Result;
-- For zero case, make sure sign of zero is preserved
else
return X;
end if;
end Unbiased_Rounding;
-----------
-- Valid --
-----------
function Valid (X : access T) return Boolean is
IEEE_Emin : constant Integer := T'Machine_Emin - 1;
IEEE_Emax : constant Integer := T'Machine_Emax - 1;
IEEE_Bias : constant Integer := -(IEEE_Emin - 1);
subtype IEEE_Exponent_Range is
Integer range IEEE_Emin - 1 .. IEEE_Emax + 1;
-- The implementation of this floating point attribute uses
-- a representation type Float_Rep that allows direct access to
-- the exponent and mantissa parts of a floating point number.
-- The Float_Rep type is an array of Float_Word elements. This
-- representation is chosen to make it possible to size the
-- type based on a generic parameter.
-- The following conditions must be met for all possible
-- instantiations of the attributes package:
-- - T'Size is an integral multiple of Float_Word'Size
-- - The exponent and sign are completely contained in a single
-- component of Float_Rep, named Most_Significant_Word (MSW).
-- - The sign occupies the most significant bit of the MSW
-- and the exponent is in the following bits.
-- Unused bits (if any) are in the least significant part.
type Float_Word is mod 2**32;
type Rep_Index is range 0 .. 7;
Rep_Last : constant Rep_Index := (T'Size - 1) / Float_Word'Size;
type Float_Rep is array (Rep_Index range 0 .. Rep_Last) of Float_Word;
Most_Significant_Word : constant Rep_Index :=
Rep_Last * Standard'Default_Bit_Order;
-- Finding the location of the Exponent_Word is a bit tricky.
-- In general we assume Word_Order = Bit_Order.
-- This expression needs to be refined for VMS.
Exponent_Factor : constant Float_Word :=
2**(Float_Word'Size - 1) /
Float_Word (IEEE_Emax - IEEE_Emin + 3) *
Boolean'Pos (T'Size /= 96) +
Boolean'Pos (T'Size = 96);
-- Factor that the extracted exponent needs to be divided by
-- to be in range 0 .. IEEE_Emax - IEEE_Emin + 2.
-- Special kludge: Exponent_Factor is 0 for x86 double extended
-- as GCC adds 16 unused bits to the type.
Exponent_Mask : constant Float_Word :=
Float_Word (IEEE_Emax - IEEE_Emin + 2) *
Exponent_Factor;
-- Value needed to mask out the exponent field.
-- This assumes that the range IEEE_Emin - 1 .. IEEE_Emax + 1
-- contains 2**N values, for some N in Natural.
function To_Float is new Ada.Unchecked_Conversion (Float_Rep, T);
type Float_Access is access all T;
function To_Address is
new Ada.Unchecked_Conversion (Float_Access, System.Address);
XA : constant System.Address := To_Address (Float_Access (X));
R : Float_Rep;
pragma Import (Ada, R);
for R'Address use XA;
-- R is a view of the input floating-point parameter. Note that we
-- must avoid copying the actual bits of this parameter in float
-- form (since it may be a signalling NaN.
E : constant IEEE_Exponent_Range :=
Integer ((R (Most_Significant_Word) and Exponent_Mask) /
Exponent_Factor)
- IEEE_Bias;
-- Mask/Shift T to only get bits from the exponent
-- Then convert biased value to integer value.
SR : Float_Rep;
-- Float_Rep representation of significant of X.all
begin
if T'Denorm then
-- All denormalized numbers are valid, so only invalid numbers
-- are overflows and NaN's, both with exponent = Emax + 1.
return E /= IEEE_Emax + 1;
end if;
-- All denormalized numbers except 0.0 are invalid
-- Set exponent of X to zero, so we end up with the significand, which
-- definitely is a valid number and can be converted back to a float.
SR := R;
SR (Most_Significant_Word) :=
(SR (Most_Significant_Word)
and not Exponent_Mask) + Float_Word (IEEE_Bias) * Exponent_Factor;
return (E in IEEE_Emin .. IEEE_Emax) or else
((E = IEEE_Emin - 1) and then abs To_Float (SR) = 1.0);
end Valid;
end System.Fat_Gen;
|