1.file "expl_m1.s"
2
3
4// Copyright (c) 2000 - 2003, Intel Corporation
5// All rights reserved.
6//
7//
8// Redistribution and use in source and binary forms, with or without
9// modification, are permitted provided that the following conditions are
10// met:
11//
12// * Redistributions of source code must retain the above copyright
13// notice, this list of conditions and the following disclaimer.
14//
15// * Redistributions in binary form must reproduce the above copyright
16// notice, this list of conditions and the following disclaimer in the
17// documentation and/or other materials provided with the distribution.
18//
19// * The name of Intel Corporation may not be used to endorse or promote
20// products derived from this software without specific prior written
21// permission.
22
23// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
26// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
27// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
28// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
29// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
30// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
31// OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
32// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
33// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34//
35// Intel Corporation is the author of this code, and requests that all
36// problem reports or change requests be submitted to it directly at
37// http://www.intel.com/software/products/opensource/libraries/num.htm.
38//
39// History
40//==============================================================
41// 02/02/00 Initial Version
42// 04/04/00 Unwind support added
43// 08/15/00 Bundle added after call to __libm_error_support to properly
44//          set [the previously overwritten] GR_Parameter_RESULT.
45// 07/07/01 Improved speed of all paths
46// 05/20/02 Cleaned up namespace and sf0 syntax
47// 02/10/03 Reordered header: .section, .global, .proc, .align;
48//          used data8 for long double table values
49// 03/11/03 Improved accuracy and performance, corrected missing inexact flags
50// 04/17/03 Eliminated misplaced and unused data label
51// 12/15/03 Eliminated call to error support on expm1l underflow
52//
53//*********************************************************************
54//
55// Function:   Combined expl(x) and expm1l(x), where
56//                        x
57//             expl(x) = e , for double-extended precision x values
58//                          x
59//             expm1l(x) = e  - 1  for double-extended precision x values
60//
61//*********************************************************************
62//
63// Resources Used:
64//
65//    Floating-Point Registers: f8  (Input and Return Value)
66//                              f9-f15,f32-f77
67//
68//    General Purpose Registers:
69//      r14-r38
70//      r35-r38 (Used to pass arguments to error handling routine)
71//
72//    Predicate Registers:      p6-p15
73//
74//*********************************************************************
75//
76// IEEE Special Conditions:
77//
78//    Denormal  fault raised on denormal inputs
79//    Overflow exceptions raised when appropriate for exp and expm1
80//    Underflow exceptions raised when appropriate for exp and expm1
81//    (Error Handling Routine called for overflow and Underflow)
82//    Inexact raised when appropriate by algorithm
83//
84//    exp(inf) = inf
85//    exp(-inf) = +0
86//    exp(SNaN) = QNaN
87//    exp(QNaN) = QNaN
88//    exp(0) = 1
89//    exp(EM_special Values) = QNaN
90//    exp(inf) = inf
91//    expm1(-inf) = -1
92//    expm1(SNaN) = QNaN
93//    expm1(QNaN) = QNaN
94//    expm1(0) = 0
95//    expm1(EM_special Values) = QNaN
96//
97//*********************************************************************
98//
99// Implementation and Algorithm Notes:
100//
101//  ker_exp_64( in_FR  : X,
102//            out_FR : Y_hi,
103//            out_FR : Y_lo,
104//            out_FR : scale,
105//            out_PR : Safe )
106//
107// On input, X is in register format
108// p6 for exp,
109// p7 for expm1,
110//
111// On output,
112//
113//   scale*(Y_hi + Y_lo)  approximates  exp(X)       if exp
114//   scale*(Y_hi + Y_lo)  approximates  exp(X)-1     if expm1
115//
116// The accuracy is sufficient for a highly accurate 64 sig.
117// bit implementation.  Safe is set if there is no danger of
118// overflow/underflow when the result is composed from scale,
119// Y_hi and Y_lo. Thus, we can have a fast return if Safe is set.
120// Otherwise, one must prepare to handle the possible exception
121// appropriately.  Note that SAFE not set (false) does not mean
122// that overflow/underflow will occur; only the setting of SAFE
123// guarantees the opposite.
124//
125// **** High Level Overview ****
126//
127// The method consists of three cases.
128//
129// If           |X| < Tiny	use case exp_tiny;
130// else if	|X| < 2^(-m)	use case exp_small; m=12 for exp, m=7 for expm1
131// else		use case exp_regular;
132//
133// Case exp_tiny:
134//
135//   1 + X     can be used to approximate exp(X)
136//   X + X^2/2 can be used to approximate exp(X) - 1
137//
138// Case exp_small:
139//
140//   Here, exp(X) and exp(X) - 1 can all be
141//   approximated by a relatively simple polynomial.
142//
143//   This polynomial resembles the truncated Taylor series
144//
145//	exp(w) = 1 + w + w^2/2! + w^3/3! + ... + w^n/n!
146//
147// Case exp_regular:
148//
149//   Here we use a table lookup method. The basic idea is that in
150//   order to compute exp(X), we accurately decompose X into
151//
152//   X = N * log(2)/(2^12)  + r,	|r| <= log(2)/2^13.
153//
154//   Hence
155//
156//   exp(X) = 2^( N / 2^12 ) * exp(r).
157//
158//   The value 2^( N / 2^12 ) is obtained by simple combinations
159//   of values calculated beforehand and stored in table; exp(r)
160//   is approximated by a short polynomial because |r| is small.
161//
162//   We elaborate this method in 4 steps.
163//
164//   Step 1: Reduction
165//
166//   The value 2^12/log(2) is stored as a double-extended number
167//   L_Inv.
168//
169//   N := round_to_nearest_integer( X * L_Inv )
170//
171//   The value log(2)/2^12 is stored as two numbers L_hi and L_lo so
172//   that r can be computed accurately via
173//
174//   r := (X - N*L_hi) - N*L_lo
175//
176//   We pick L_hi such that N*L_hi is representable in 64 sig. bits
177//   and thus the FMA   X - N*L_hi   is error free. So r is the
178//   1 rounding error from an exact reduction with respect to
179//
180//   L_hi + L_lo.
181//
182//   In particular, L_hi has 30 significant bit and can be stored
183//   as a double-precision number; L_lo has 64 significant bits and
184//   stored as a double-extended number.
185//
186//   Step 2: Approximation
187//
188//   exp(r) - 1 is approximated by a short polynomial of the form
189//
190//   r + A_1 r^2 + A_2 r^3 + A_3 r^4 .
191//
192//   Step 3: Composition from Table Values
193//
194//   The value 2^( N / 2^12 ) can be composed from a couple of tables
195//   of precalculated values. First, express N as three integers
196//   K, M_1, and M_2 as
197//
198//     N  =  K * 2^12  + M_1 * 2^6 + M_2
199//
200//   Where 0 <= M_1, M_2 < 2^6; and K can be positive or negative.
201//   When N is represented in 2's complement, M_2 is simply the 6
202//   lsb's, M_1 is the next 6, and K is simply N shifted right
203//   arithmetically (sign extended) by 12 bits.
204//
205//   Now, 2^( N / 2^12 ) is simply
206//
207//      2^K * 2^( M_1 / 2^6 ) * 2^( M_2 / 2^12 )
208//
209//   Clearly, 2^K needs no tabulation. The other two values are less
210//   trivial because if we store each accurately to more than working
211//   precision, than its product is too expensive to calculate. We
212//   use the following method.
213//
214//   Define two mathematical values, delta_1 and delta_2, implicitly
215//   such that
216//
217//     T_1 = exp( [M_1 log(2)/2^6]  -  delta_1 )
218//     T_2 = exp( [M_2 log(2)/2^12] -  delta_2 )
219//
220//   are representable as 24 significant bits. To illustrate the idea,
221//   we show how we define delta_1:
222//
223//     T_1     := round_to_24_bits( exp( M_1 log(2)/2^6 ) )
224//     delta_1  = (M_1 log(2)/2^6) - log( T_1 )
225//
226//   The last equality means mathematical equality. We then tabulate
227//
228//     W_1 := exp(delta_1) - 1
229//     W_2 := exp(delta_2) - 1
230//
231//   Both in double precision.
232//
233//   From the tabulated values T_1, T_2, W_1, W_2, we compose the values
234//   T and W via
235//
236//     T := T_1 * T_2			...exactly
237//     W := W_1 + (1 + W_1)*W_2
238//
239//   W approximates exp( delta ) - 1  where delta = delta_1 + delta_2.
240//   The mathematical product of T and (W+1) is an accurate representation
241//   of 2^(M_1/2^6) * 2^(M_2/2^12).
242//
243//   Step 4. Reconstruction
244//
245//   Finally, we can reconstruct exp(X), exp(X) - 1.
246//   Because
247//
248//	X = K * log(2) + (M_1*log(2)/2^6  - delta_1)
249//		       + (M_2*log(2)/2^12 - delta_2)
250//		       + delta_1 + delta_2 + r 		...accurately
251//   We have
252//
253//	exp(X) ~=~ 2^K * ( T + T*[exp(delta_1+delta_2+r) - 1] )
254//	       ~=~ 2^K * ( T + T*[exp(delta + r) - 1]         )
255//	       ~=~ 2^K * ( T + T*[(exp(delta)-1)
256//				 + exp(delta)*(exp(r)-1)]   )
257//             ~=~ 2^K * ( T + T*( W + (1+W)*poly(r) ) )
258//             ~=~ 2^K * ( Y_hi  +  Y_lo )
259//
260//   where Y_hi = T  and Y_lo = T*(W + (1+W)*poly(r))
261//
262//   For exp(X)-1, we have
263//
264//	exp(X)-1 ~=~ 2^K * ( Y_hi + Y_lo ) - 1
265//		 ~=~ 2^K * ( Y_hi + Y_lo - 2^(-K) )
266//
267//   and we combine Y_hi + Y_lo - 2^(-N)  into the form of two
268//   numbers  Y_hi + Y_lo carefully.
269//
270//   **** Algorithm Details ****
271//
272//   A careful algorithm must be used to realize the mathematical ideas
273//   accurately. We describe each of the three cases. We assume SAFE
274//   is preset to be TRUE.
275//
276//   Case exp_tiny:
277//
278//   The important points are to ensure an accurate result under
279//   different rounding directions and a correct setting of the SAFE
280//   flag.
281//
282//   If expm1 is 1, then
283//      SAFE  := False	...possibility of underflow
284//      Scale := 1.0
285//      Y_hi  := X
286//      Y_lo  := 2^(-17000)
287//   Else
288//      Scale := 1.0
289//      Y_hi  := 1.0
290//      Y_lo  := X	...for different rounding modes
291//   Endif
292//
293//   Case exp_small:
294//
295//   Here we compute a simple polynomial. To exploit parallelism, we split
296//   the polynomial into several portions.
297//
298//   Let r = X
299//
300//   If exp 	...i.e. exp( argument )
301//
302//      rsq := r * r;
303//      r4  := rsq*rsq
304//      poly_lo := P_3 + r*(P_4 + r*(P_5 + r*P_6))
305//      poly_hi := r + rsq*(P_1 + r*P_2)
306//      Y_lo    := poly_hi + r4 * poly_lo
307//      Y_hi    := 1.0
308//      Scale   := 1.0
309//
310//   Else			...i.e. exp( argument ) - 1
311//
312//      rsq := r * r
313//      r4  := rsq * rsq
314//      poly_lo := Q_7 + r*(Q_8 + r*Q_9))
315//      poly_med:= Q_3 + r*Q_4 + rsq*(Q_5 + r*Q_6)
316//      poly_med:= poly_med + r4*poly_lo
317//      poly_hi := Q_1 + r*Q_2
318//      Y_lo    := rsq*(poly_hi +  rsq*poly_lo)
319//      Y_hi    := X
320//      Scale   := 1.0
321//
322//   Endif
323//
324//  Case exp_regular:
325//
326//  The previous description contain enough information except the
327//  computation of poly and the final Y_hi and Y_lo in the case for
328//  exp(X)-1.
329//
330//  The computation of poly for Step 2:
331//
332//   rsq := r*r
333//   poly := r + rsq*(A_1 + r*(A_2 + r*A_3))
334//
335//  For the case exp(X) - 1, we need to incorporate 2^(-K) into
336//  Y_hi and Y_lo at the end of Step 4.
337//
338//   If K > 10 then
339//      Y_lo := Y_lo - 2^(-K)
340//   Else
341//      If K < -10 then
342//	 Y_lo := Y_hi + Y_lo
343//	 Y_hi := -2^(-K)
344//      Else
345//	 Y_hi := Y_hi - 2^(-K)
346//      End If
347//   End If
348//
349//=======================================================
350// General Purpose Registers
351//
352GR_ad_Arg           = r14
353GR_ad_A             = r15
354GR_sig_inv_ln2      = r15
355GR_rshf_2to51       = r16
356GR_ad_PQ            = r16
357GR_ad_Q             = r16
358GR_signexp_x        = r17
359GR_exp_x            = r17
360GR_small_exp        = r18
361GR_rshf             = r18
362GR_exp_mask         = r19
363GR_ad_W1            = r20
364GR_exp_2tom51       = r20
365GR_ad_W2            = r21
366GR_exp_underflow    = r21
367GR_M2               = r22
368GR_huge_exp         = r22
369GR_M1               = r23
370GR_huge_signif      = r23
371GR_K                = r24
372GR_one              = r24
373GR_minus_one        = r24
374GR_exp_bias         = r25
375GR_ad_Limits        = r26
376GR_N_fix            = r26
377GR_exp_2_mk         = r26
378GR_ad_P             = r27
379GR_exp_2_k          = r27
380GR_big_expo_neg     = r28
381GR_very_small_exp   = r29
382GR_exp_half         = r29
383GR_ad_T1            = r30
384GR_ad_T2            = r31
385
386GR_SAVE_PFS         = r32
387GR_SAVE_B0          = r33
388GR_SAVE_GP          = r34
389GR_Parameter_X      = r35
390GR_Parameter_Y      = r36
391GR_Parameter_RESULT = r37
392GR_Parameter_TAG    = r38
393
394// Floating Point Registers
395//
396FR_norm_x           = f9
397FR_RSHF_2TO51       = f10
398FR_INV_LN2_2TO63    = f11
399FR_W_2TO51_RSH      = f12
400FR_2TOM51           = f13
401FR_RSHF             = f14
402FR_Y_hi             = f34
403FR_Y_lo             = f35
404FR_scale            = f36
405FR_tmp              = f37
406FR_float_N          = f38
407FR_N_signif         = f39
408FR_L_hi             = f40
409FR_L_lo             = f41
410FR_r                = f42
411FR_W1               = f43
412FR_T1               = f44
413FR_W2               = f45
414FR_T2               = f46
415FR_W1_p1            = f47
416FR_rsq              = f48
417FR_A2               = f49
418FR_r4               = f50
419FR_A3               = f51
420FR_poly             = f52
421FR_T                = f53
422FR_W                = f54
423FR_Wp1              = f55
424FR_p21              = f59
425FR_p210             = f59
426FR_p65              = f60
427FR_p654             = f60
428FR_p6543            = f60
429FR_2_mk             = f61
430FR_P4Q7             = f61
431FR_P4               = f61
432FR_Q7               = f61
433FR_P3Q6             = f62
434FR_P3               = f62
435FR_Q6               = f62
436FR_q65              = f62
437FR_q6543            = f62
438FR_P2Q5             = f63
439FR_P2               = f63
440FR_Q5               = f63
441FR_P1Q4             = f64
442FR_P1               = f64
443FR_Q4               = f64
444FR_q43              = f64
445FR_Q3               = f65
446FR_Q2               = f66
447FR_q21              = f66
448FR_Q1               = f67
449FR_A1               = f68
450FR_P6Q9             = f68
451FR_P6               = f68
452FR_Q9               = f68
453FR_P5Q8             = f69
454FR_P5               = f69
455FR_Q8               = f69
456FR_q987             = f69
457FR_q98              = f69
458FR_q9876543         = f69
459FR_min_oflow_x      = f70
460FR_huge_exp         = f70
461FR_zero_uflow_x     = f71
462FR_huge_signif      = f71
463FR_huge             = f72
464FR_small            = f72
465FR_half             = f73
466FR_T_scale          = f74
467FR_result_lo        = f75
468FR_W_T_scale        = f76
469FR_Wp1_T_scale      = f77
470FR_ftz              = f77
471FR_half_x           = f77
472//
473
474FR_X                = f9
475FR_Y                = f0
476FR_RESULT           = f15
477
478// ************* DO NOT CHANGE ORDER OF THESE TABLES ********************
479
480// double-extended 1/ln(2)
481// 3fff b8aa 3b29 5c17 f0bb be87fed0691d3e88
482// 3fff b8aa 3b29 5c17 f0bc
483// For speed the significand will be loaded directly with a movl and setf.sig
484//   and the exponent will be bias+63 instead of bias+0.  Thus subsequent
485//   computations need to scale appropriately.
486// The constant 2^12/ln(2) is needed for the computation of N.  This is also
487//   obtained by scaling the computations.
488//
489// Two shifting constants are loaded directly with movl and setf.d.
490//   1. RSHF_2TO51 = 1.1000..00 * 2^(63-12)
491//        This constant is added to x*1/ln2 to shift the integer part of
492//        x*2^12/ln2 into the rightmost bits of the significand.
493//        The result of this fma is N_signif.
494//   2. RSHF       = 1.1000..00 * 2^(63)
495//        This constant is subtracted from N_signif * 2^(-51) to give
496//        the integer part of N, N_fix, as a floating-point number.
497//        The result of this fms is float_N.
498
499RODATA
500.align 64
501LOCAL_OBJECT_START(Constants_exp_64_Arg)
502//data8 0xB8AA3B295C17F0BC,0x0000400B // Inv_L = 2^12/log(2)
503data8 0xB17217F400000000,0x00003FF2 // L_hi = hi part log(2)/2^12
504data8 0xF473DE6AF278ECE6,0x00003FD4 // L_lo = lo part log(2)/2^12
505LOCAL_OBJECT_END(Constants_exp_64_Arg)
506
507LOCAL_OBJECT_START(Constants_exp_64_Limits)
508data8 0xb17217f7d1cf79ac,0x0000400c // Smallest long dbl oflow x
509data8 0xb220000000000000,0x0000c00c // Small long dbl uflow zero x
510LOCAL_OBJECT_END(Constants_exp_64_Limits)
511
512LOCAL_OBJECT_START(Constants_exp_64_A)
513data8 0xAAAAAAABB1B736A0,0x00003FFA // A3
514data8 0xAAAAAAAB90CD6327,0x00003FFC // A2
515data8 0xFFFFFFFFFFFFFFFF,0x00003FFD // A1
516LOCAL_OBJECT_END(Constants_exp_64_A)
517
518LOCAL_OBJECT_START(Constants_exp_64_P)
519data8 0xD00D6C8143914A8A,0x00003FF2 // P6
520data8 0xB60BC4AC30304B30,0x00003FF5 // P5
521data8 0x888888887474C518,0x00003FF8 // P4
522data8 0xAAAAAAAA8DAE729D,0x00003FFA // P3
523data8 0xAAAAAAAAAAAAAF61,0x00003FFC // P2
524data8 0x80000000000004C7,0x00003FFE // P1
525LOCAL_OBJECT_END(Constants_exp_64_P)
526
527LOCAL_OBJECT_START(Constants_exp_64_Q)
528data8 0x93F2AC5F7471F32E, 0x00003FE9 // Q9
529data8 0xB8DA0F3550B3E764, 0x00003FEC // Q8
530data8 0xD00D00D0028E89C4, 0x00003FEF // Q7
531data8 0xD00D00DAEB8C4E91, 0x00003FF2 // Q6
532data8 0xB60B60B60B60B6F5, 0x00003FF5 // Q5
533data8 0x888888888886CC23, 0x00003FF8 // Q4
534data8 0xAAAAAAAAAAAAAAAB, 0x00003FFA // Q3
535data8 0xAAAAAAAAAAAAAAAB, 0x00003FFC // Q2
536data8 0x8000000000000000, 0x00003FFE // Q1
537LOCAL_OBJECT_END(Constants_exp_64_Q)
538
539LOCAL_OBJECT_START(Constants_exp_64_T1)
540data4 0x3F800000,0x3F8164D2,0x3F82CD87,0x3F843A29
541data4 0x3F85AAC3,0x3F871F62,0x3F88980F,0x3F8A14D5
542data4 0x3F8B95C2,0x3F8D1ADF,0x3F8EA43A,0x3F9031DC
543data4 0x3F91C3D3,0x3F935A2B,0x3F94F4F0,0x3F96942D
544data4 0x3F9837F0,0x3F99E046,0x3F9B8D3A,0x3F9D3EDA
545data4 0x3F9EF532,0x3FA0B051,0x3FA27043,0x3FA43516
546data4 0x3FA5FED7,0x3FA7CD94,0x3FA9A15B,0x3FAB7A3A
547data4 0x3FAD583F,0x3FAF3B79,0x3FB123F6,0x3FB311C4
548data4 0x3FB504F3,0x3FB6FD92,0x3FB8FBAF,0x3FBAFF5B
549data4 0x3FBD08A4,0x3FBF179A,0x3FC12C4D,0x3FC346CD
550data4 0x3FC5672A,0x3FC78D75,0x3FC9B9BE,0x3FCBEC15
551data4 0x3FCE248C,0x3FD06334,0x3FD2A81E,0x3FD4F35B
552data4 0x3FD744FD,0x3FD99D16,0x3FDBFBB8,0x3FDE60F5
553data4 0x3FE0CCDF,0x3FE33F89,0x3FE5B907,0x3FE8396A
554data4 0x3FEAC0C7,0x3FED4F30,0x3FEFE4BA,0x3FF28177
555data4 0x3FF5257D,0x3FF7D0DF,0x3FFA83B3,0x3FFD3E0C
556LOCAL_OBJECT_END(Constants_exp_64_T1)
557
558LOCAL_OBJECT_START(Constants_exp_64_T2)
559data4 0x3F800000,0x3F80058C,0x3F800B18,0x3F8010A4
560data4 0x3F801630,0x3F801BBD,0x3F80214A,0x3F8026D7
561data4 0x3F802C64,0x3F8031F2,0x3F803780,0x3F803D0E
562data4 0x3F80429C,0x3F80482B,0x3F804DB9,0x3F805349
563data4 0x3F8058D8,0x3F805E67,0x3F8063F7,0x3F806987
564data4 0x3F806F17,0x3F8074A8,0x3F807A39,0x3F807FCA
565data4 0x3F80855B,0x3F808AEC,0x3F80907E,0x3F809610
566data4 0x3F809BA2,0x3F80A135,0x3F80A6C7,0x3F80AC5A
567data4 0x3F80B1ED,0x3F80B781,0x3F80BD14,0x3F80C2A8
568data4 0x3F80C83C,0x3F80CDD1,0x3F80D365,0x3F80D8FA
569data4 0x3F80DE8F,0x3F80E425,0x3F80E9BA,0x3F80EF50
570data4 0x3F80F4E6,0x3F80FA7C,0x3F810013,0x3F8105AA
571data4 0x3F810B41,0x3F8110D8,0x3F81166F,0x3F811C07
572data4 0x3F81219F,0x3F812737,0x3F812CD0,0x3F813269
573data4 0x3F813802,0x3F813D9B,0x3F814334,0x3F8148CE
574data4 0x3F814E68,0x3F815402,0x3F81599C,0x3F815F37
575LOCAL_OBJECT_END(Constants_exp_64_T2)
576
577LOCAL_OBJECT_START(Constants_exp_64_W1)
578data8 0x0000000000000000, 0xBE384454171EC4B4
579data8 0xBE6947414AA72766, 0xBE5D32B6D42518F8
580data8 0x3E68D96D3A319149, 0xBE68F4DA62415F36
581data8 0xBE6DDA2FC9C86A3B, 0x3E6B2E50F49228FE
582data8 0xBE49C0C21188B886, 0x3E64BFC21A4C2F1F
583data8 0xBE6A2FBB2CB98B54, 0x3E5DC5DE9A55D329
584data8 0x3E69649039A7AACE, 0x3E54728B5C66DBA5
585data8 0xBE62B0DBBA1C7D7D, 0x3E576E0409F1AF5F
586data8 0x3E6125001A0DD6A1, 0xBE66A419795FBDEF
587data8 0xBE5CDE8CE1BD41FC, 0xBE621376EA54964F
588data8 0x3E6370BE476E76EE, 0x3E390D1A3427EB92
589data8 0x3E1336DE2BF82BF8, 0xBE5FF1CBD0F7BD9E
590data8 0xBE60A3550CEB09DD, 0xBE5CA37E0980F30D
591data8 0xBE5C541B4C082D25, 0xBE5BBECA3B467D29
592data8 0xBE400D8AB9D946C5, 0xBE5E2A0807ED374A
593data8 0xBE66CB28365C8B0A, 0x3E3AAD5BD3403BCA
594data8 0x3E526055C7EA21E0, 0xBE442C75E72880D6
595data8 0x3E58B2BB85222A43, 0xBE5AAB79522C42BF
596data8 0xBE605CB4469DC2BC, 0xBE589FA7A48C40DC
597data8 0xBE51C2141AA42614, 0xBE48D087C37293F4
598data8 0x3E367A1CA2D673E0, 0xBE51BEBB114F7A38
599data8 0xBE6348E5661A4B48, 0xBDF526431D3B9962
600data8 0x3E3A3B5E35A78A53, 0xBE46C46C1CECD788
601data8 0xBE60B7EC7857D689, 0xBE594D3DD14F1AD7
602data8 0xBE4F9C304C9A8F60, 0xBE52187302DFF9D2
603data8 0xBE5E4C8855E6D68F, 0xBE62140F667F3DC4
604data8 0xBE36961B3BF88747, 0x3E602861C96EC6AA
605data8 0xBE3B5151D57FD718, 0x3E561CD0FC4A627B
606data8 0xBE3A5217CA913FEA, 0x3E40A3CC9A5D193A
607data8 0xBE5AB71310A9C312, 0x3E4FDADBC5F57719
608data8 0x3E361428DBDF59D5, 0x3E5DB5DB61B4180D
609data8 0xBE42AD5F7408D856, 0x3E2A314831B2B707
610LOCAL_OBJECT_END(Constants_exp_64_W1)
611
612LOCAL_OBJECT_START(Constants_exp_64_W2)
613data8 0x0000000000000000, 0xBE641F2537A3D7A2
614data8 0xBE68DD57AD028C40, 0xBE5C77D8F212B1B6
615data8 0x3E57878F1BA5B070, 0xBE55A36A2ECAE6FE
616data8 0xBE620608569DFA3B, 0xBE53B50EA6D300A3
617data8 0x3E5B5EF2223F8F2C, 0xBE56A0D9D6DE0DF4
618data8 0xBE64EEF3EAE28F51, 0xBE5E5AE2367EA80B
619data8 0x3E47CB1A5FCBC02D, 0xBE656BA09BDAFEB7
620data8 0x3E6E70C6805AFEE7, 0xBE6E0509A3415EBA
621data8 0xBE56856B49BFF529, 0x3E66DD3300508651
622data8 0x3E51165FC114BC13, 0x3E53333DC453290F
623data8 0x3E6A072B05539FDA, 0xBE47CD877C0A7696
624data8 0xBE668BF4EB05C6D9, 0xBE67C3E36AE86C93
625data8 0xBE533904D0B3E84B, 0x3E63E8D9556B53CE
626data8 0x3E212C8963A98DC8, 0xBE33138F032A7A22
627data8 0x3E530FA9BC584008, 0xBE6ADF82CCB93C97
628data8 0x3E5F91138370EA39, 0x3E5443A4FB6A05D8
629data8 0x3E63DACD181FEE7A, 0xBE62B29DF0F67DEC
630data8 0x3E65C4833DDE6307, 0x3E5BF030D40A24C1
631data8 0x3E658B8F14E437BE, 0xBE631C29ED98B6C7
632data8 0x3E6335D204CF7C71, 0x3E529EEDE954A79D
633data8 0x3E5D9257F64A2FB8, 0xBE6BED1B854ED06C
634data8 0x3E5096F6D71405CB, 0xBE3D4893ACB9FDF5
635data8 0xBDFEB15801B68349, 0x3E628D35C6A463B9
636data8 0xBE559725ADE45917, 0xBE68C29C042FC476
637data8 0xBE67593B01E511FA, 0xBE4A4313398801ED
638data8 0x3E699571DA7C3300, 0x3E5349BE08062A9E
639data8 0x3E5229C4755BB28E, 0x3E67E42677A1F80D
640data8 0xBE52B33F6B69C352, 0xBE6B3550084DA57F
641data8 0xBE6DB03FD1D09A20, 0xBE60CBC42161B2C1
642data8 0x3E56ED9C78A2B771, 0xBE508E319D0FA795
643data8 0xBE59482AFD1A54E9, 0xBE2A17CEB07FD23E
644data8 0x3E68BF5C17365712, 0x3E3956F9B3785569
645LOCAL_OBJECT_END(Constants_exp_64_W2)
646
647
648.section .text
649
650GLOBAL_IEEE754_ENTRY(expm1l)
651
652//
653//    Set p7 true for expm1, p6 false
654//
655
656{ .mlx
657      getf.exp GR_signexp_x = f8  // Get sign and exponent of x, redo if unorm
658      movl GR_sig_inv_ln2 = 0xb8aa3b295c17f0bc  // significand of 1/ln2
659}
660{ .mlx
661      addl GR_ad_Arg = @ltoff(Constants_exp_64_Arg#),gp
662      movl GR_rshf_2to51 = 0x4718000000000000 // 1.10000 2^(63+51)
663}
664;;
665
666{ .mfi
667      ld8  GR_ad_Arg = [GR_ad_Arg]       // Point to Arg table
668      fclass.m p8, p0 =  f8, 0x1E7       // Test x for natval, nan, inf, zero
669      cmp.eq  p7, p6 =  r0, r0
670}
671{ .mfb
672      mov GR_exp_half = 0x0FFFE          // Exponent of 0.5, for very small path
673      fnorm.s1 FR_norm_x = f8            // Normalize x
674      br.cond.sptk exp_continue
675}
676;;
677
678GLOBAL_IEEE754_END(expm1l)
679libm_alias_ldouble_other (__expm1, expm1)
680
681
682GLOBAL_IEEE754_ENTRY(expl)
683//
684//    Set p7 false for exp, p6 true
685//
686{ .mlx
687      getf.exp GR_signexp_x = f8  // Get sign and exponent of x, redo if unorm
688      movl GR_sig_inv_ln2 = 0xb8aa3b295c17f0bc  // significand of 1/ln2
689}
690{ .mlx
691      addl GR_ad_Arg = @ltoff(Constants_exp_64_Arg#),gp
692      movl GR_rshf_2to51 = 0x4718000000000000 // 1.10000 2^(63+51)
693}
694;;
695
696{ .mfi
697      ld8  GR_ad_Arg = [GR_ad_Arg]       // Point to Arg table
698      fclass.m p8, p0 =  f8, 0x1E7       // Test x for natval, nan, inf, zero
699      cmp.eq  p6, p7 =  r0, r0
700}
701{ .mfi
702      mov GR_exp_half = 0x0FFFE          // Exponent of 0.5, for very small path
703      fnorm.s1 FR_norm_x = f8            // Normalize x
704      nop.i 999
705}
706;;
707
708exp_continue:
709// Form two constants we need
710//  1/ln2 * 2^63  to compute  w = x * 1/ln2 * 128
711//  1.1000..000 * 2^(63+63-12) to right shift int(N) into the significand
712
713{ .mfi
714      setf.sig  FR_INV_LN2_2TO63 = GR_sig_inv_ln2 // form 1/ln2 * 2^63
715      fclass.nm.unc p9, p0 =  f8, 0x1FF  // Test x for unsupported
716      mov GR_exp_2tom51 = 0xffff-51
717}
718{ .mlx
719      setf.d  FR_RSHF_2TO51 = GR_rshf_2to51 // Form const 1.1000 * 2^(63+51)
720      movl GR_rshf = 0x43e8000000000000  // 1.10000 2^63 for right shift
721}
722;;
723
724{ .mfi
725      setf.exp FR_half = GR_exp_half     // Form 0.5 for very small path
726      fma.s1 FR_scale = f1,f1,f0         // Scale = 1.0
727      mov GR_exp_bias = 0x0FFFF          // Set exponent bias
728}
729{ .mib
730      add GR_ad_Limits = 0x20, GR_ad_Arg // Point to Limits table
731      mov GR_exp_mask = 0x1FFFF          // Form exponent mask
732(p8)  br.cond.spnt EXP_64_SPECIAL        // Branch if natval, nan, inf, zero
733}
734;;
735
736{ .mfi
737      setf.exp FR_2TOM51 = GR_exp_2tom51 // Form 2^-51 for scaling float_N
738      nop.f 999
739      add GR_ad_A = 0x40, GR_ad_Arg      // Point to A table
740}
741{ .mib
742      setf.d  FR_RSHF = GR_rshf          // Form right shift const 1.1000 * 2^63
743      add GR_ad_T1 = 0x160, GR_ad_Arg    // Point to T1 table
744(p9)  br.cond.spnt EXP_64_UNSUPPORTED    // Branch if unsupported
745}
746;;
747
748.pred.rel "mutex",p6,p7
749{ .mfi
750      ldfe FR_L_hi = [GR_ad_Arg],16      // Get L_hi
751      fcmp.eq.s0 p9,p0 =  f8, f0         // Dummy op to flag denormals
752(p6)  add GR_ad_PQ = 0x30, GR_ad_A       // Point to P table for exp
753}
754{ .mfi
755      ldfe FR_min_oflow_x = [GR_ad_Limits],16 // Get min x to cause overflow
756      fmpy.s1 FR_rsq = f8, f8            // rsq = x * x for small path
757(p7)  add GR_ad_PQ = 0x90, GR_ad_A       // Point to Q table for expm1
758};;
759
760{ .mmi
761      ldfe FR_L_lo = [GR_ad_Arg],16      // Get L_lo
762      ldfe FR_zero_uflow_x = [GR_ad_Limits],16 // Get x for zero uflow result
763      add GR_ad_W1 = 0x200, GR_ad_T1     // Point to W1 table
764}
765;;
766
767{ .mfi
768      ldfe FR_P6Q9 = [GR_ad_PQ],16       // P6(exp) or Q9(expm1) for small path
769      mov FR_r = FR_norm_x               // r = X for small path
770      mov GR_very_small_exp = -60        // Exponent of x for very small path
771}
772{ .mfi
773      add GR_ad_W2 = 0x400, GR_ad_T1     // Point to W2 table
774      nop.f 999
775(p7)  mov GR_small_exp = -7              // Exponent of x for small path expm1
776}
777;;
778
779{ .mmi
780      ldfe FR_P5Q8 = [GR_ad_PQ],16       // P5(exp) or Q8(expm1) for small path
781      and  GR_exp_x = GR_signexp_x, GR_exp_mask
782(p6)  mov GR_small_exp = -12             // Exponent of x for small path exp
783}
784;;
785
786// N_signif = X * Inv_log2_by_2^12
787// By adding 1.10...0*2^63 we shift and get round_int(N_signif) in significand.
788// We actually add 1.10...0*2^51 to X * Inv_log2 to do the same thing.
789{ .mfi
790      ldfe FR_P4Q7 = [GR_ad_PQ],16       // P4(exp) or Q7(expm1) for small path
791      fma.s1 FR_N_signif = FR_norm_x, FR_INV_LN2_2TO63, FR_RSHF_2TO51
792      nop.i 999
793}
794{ .mfi
795      sub GR_exp_x = GR_exp_x, GR_exp_bias // Get exponent
796      fmpy.s1 FR_r4 = FR_rsq, FR_rsq     // Form r4 for small path
797      cmp.eq.unc  p15, p0 =  r0, r0      // Set Safe as default
798}
799;;
800
801{ .mmi
802      ldfe FR_P3Q6 = [GR_ad_PQ],16       // P3(exp) or Q6(expm1) for small path
803      cmp.lt  p14, p0 =  GR_exp_x, GR_very_small_exp // Is |x| < 2^-60?
804      nop.i 999
805}
806;;
807
808{ .mfi
809      ldfe FR_P2Q5 = [GR_ad_PQ],16       // P2(exp) or Q5(expm1) for small path
810      fmpy.s1 FR_half_x = FR_half, FR_norm_x // 0.5 * x for very small path
811      cmp.lt  p13, p0 =  GR_exp_x, GR_small_exp // Is |x| < 2^-m?
812}
813{ .mib
814      nop.m 999
815      nop.i 999
816(p14) br.cond.spnt EXP_VERY_SMALL        // Branch if |x| < 2^-60
817}
818;;
819
820{ .mfi
821      ldfe FR_A3 = [GR_ad_A],16          // Get A3 for normal path
822      fcmp.ge.s1 p10,p0 = FR_norm_x, FR_min_oflow_x // Will result overflow?
823      mov GR_big_expo_neg = -16381       // -0x3ffd
824}
825{ .mfb
826      ldfe FR_P1Q4 = [GR_ad_PQ],16       // P1(exp) or Q4(expm1) for small path
827      nop.f 999
828(p13) br.cond.spnt EXP_SMALL             // Branch if |x| < 2^-m
829                                         // m=12 for exp, m=7 for expm1
830}
831;;
832
833// Now we are on the main path for |x| >= 2^-m, m=12 for exp, m=7 for expm1
834//
835// float_N = round_int(N_signif)
836// The signficand of N_signif contains the rounded integer part of X * 2^12/ln2,
837// as a twos complement number in the lower bits (that is, it may be negative).
838// That twos complement number (called N) is put into GR_N.
839
840// Since N_signif is scaled by 2^51, it must be multiplied by 2^-51
841// before the shift constant 1.10000 * 2^63 is subtracted to yield float_N.
842// Thus, float_N contains the floating point version of N
843
844
845{ .mfi
846      ldfe FR_A2 = [GR_ad_A],16          // Get A2 for main path
847      fcmp.lt.s1 p11,p0 = FR_norm_x, FR_zero_uflow_x // Certain zero, uflow?
848      add GR_ad_T2 = 0x100, GR_ad_T1     // Point to T2 table
849}
850{ .mfi
851      nop.m 999
852      fms.s1 FR_float_N = FR_N_signif, FR_2TOM51, FR_RSHF // Form float_N
853      nop.i 999
854}
855;;
856
857{ .mbb
858      getf.sig GR_N_fix = FR_N_signif    // Get N from significand
859(p10) br.cond.spnt  EXP_OVERFLOW         // Branch if result will overflow
860(p11) br.cond.spnt  EXP_CERTAIN_UNDERFLOW_ZERO // Branch if certain zero, uflow
861}
862;;
863
864{ .mfi
865      ldfe FR_A1 = [GR_ad_A],16          // Get A1 for main path
866      fnma.s1 FR_r = FR_L_hi, FR_float_N, FR_norm_x  // r = -L_hi * float_N + x
867      extr.u GR_M1 = GR_N_fix, 6, 6      // Extract index M_1
868}
869{ .mfi
870      and GR_M2 = 0x3f, GR_N_fix         // Extract index M_2
871      nop.f 999
872      nop.i 999
873}
874;;
875
876// N_fix is only correct up to 50 bits because of our right shift technique.
877// Actually in the normal path we will have restricted K to about 14 bits.
878// Somewhat arbitrarily we extract 32 bits.
879{ .mfi
880      shladd GR_ad_W1 = GR_M1,3,GR_ad_W1 // Point to W1
881      nop.f 999
882      extr GR_K = GR_N_fix, 12, 32       // Extract limited range K
883}
884{ .mfi
885      shladd GR_ad_T1 = GR_M1,2,GR_ad_T1 // Point to T1
886      nop.f 999
887      shladd GR_ad_T2 = GR_M2,2,GR_ad_T2 // Point to T2
888}
889;;
890
891{ .mmi
892      ldfs  FR_T1 = [GR_ad_T1],0         // Get T1
893      ldfd  FR_W1 = [GR_ad_W1],0         // Get W1
894      add GR_exp_2_k = GR_exp_bias, GR_K // Form exponent of 2^k
895}
896;;
897
898{ .mmi
899      ldfs  FR_T2 = [GR_ad_T2],0         // Get T2
900      shladd GR_ad_W2 = GR_M2,3,GR_ad_W2 // Point to W2
901      sub GR_exp_2_mk = GR_exp_bias, GR_K // Form exponent of 2^-k
902}
903;;
904
905{ .mmf
906      ldfd  FR_W2 = [GR_ad_W2],0         // Get W2
907      setf.exp FR_scale = GR_exp_2_k     // Set scale = 2^k
908      fnma.s1 FR_r = FR_L_lo, FR_float_N, FR_r // r = -L_lo * float_N + r
909}
910;;
911
912{ .mfi
913      setf.exp FR_2_mk = GR_exp_2_mk     // Form 2^-k
914      fma.s1 FR_poly = FR_r, FR_A3, FR_A2 // poly = r * A3 + A2
915      cmp.lt p8,p15 = GR_K,GR_big_expo_neg // Set Safe if K > big_expo_neg
916}
917{ .mfi
918      nop.m 999
919      fmpy.s1 FR_rsq = FR_r, FR_r         // rsq = r * r
920      nop.i 999
921}
922;;
923
924{ .mfi
925      nop.m 999
926      fmpy.s1 FR_T = FR_T1, FR_T2         // T = T1 * T2
927      nop.i 999
928}
929{ .mfi
930      nop.m 999
931      fadd.s1 FR_W1_p1 = FR_W1, f1        // W1_p1 = W1 + 1.0
932      nop.i 999
933}
934;;
935
936{ .mfi
937(p7)  cmp.lt.unc  p8, p9 =  10, GR_K       // If expm1, set p8 if K > 10
938      fma.s1 FR_poly = FR_r, FR_poly, FR_A1 // poly = r * poly + A1
939      nop.i 999
940}
941;;
942
943{ .mfi
944(p7)  cmp.eq  p15, p0 =  r0, r0            // If expm1, set Safe flag
945      fma.s1 FR_T_scale = FR_T, FR_scale, f0 // T_scale = T * scale
946(p9)  cmp.gt.unc  p9, p10 =  -10, GR_K     // If expm1, set p9 if K < -10
947                                           // If expm1, set p10 if -10<=K<=10
948}
949{ .mfi
950      nop.m 999
951      fma.s1 FR_W = FR_W2, FR_W1_p1, FR_W1 // W = W2 * (W1+1.0) + W1
952      nop.i 999
953}
954;;
955
956{ .mfi
957      nop.m 999
958      mov FR_Y_hi = FR_T                   // Assume Y_hi = T
959      nop.i 999
960}
961;;
962
963{ .mfi
964      nop.m 999
965      fma.s1 FR_poly = FR_rsq, FR_poly, FR_r // poly = rsq * poly + r
966      nop.i 999
967}
968;;
969
970{ .mfi
971      nop.m 999
972      fma.s1 FR_Wp1_T_scale = FR_W, FR_T_scale, FR_T_scale // (W+1)*T*scale
973      nop.i 999
974}
975{ .mfi
976      nop.m 999
977      fma.s1 FR_W_T_scale = FR_W, FR_T_scale, f0 // W*T*scale
978      nop.i 999
979}
980;;
981
982{ .mfi
983      nop.m 999
984(p9)  fsub.s1 FR_Y_hi = f0, FR_2_mk      // If expm1, if K < -10 set Y_hi
985      nop.i 999
986}
987{ .mfi
988      nop.m 999
989(p10) fsub.s1 FR_Y_hi = FR_T, FR_2_mk    // If expm1, if |K|<=10 set Y_hi
990      nop.i 999
991}
992;;
993
994{ .mfi
995      nop.m 999
996      fma.s1 FR_result_lo = FR_Wp1_T_scale, FR_poly, FR_W_T_scale
997      nop.i 999
998}
999;;
1000
1001.pred.rel "mutex",p8,p9
1002// If K > 10 adjust result_lo = result_lo - scale * 2^-k
1003// If |K| <= 10 adjust result_lo = result_lo + scale * T
1004{ .mfi
1005      nop.m 999
1006(p8)  fnma.s1 FR_result_lo = FR_scale, FR_2_mk, FR_result_lo // If K > 10
1007      nop.i 999
1008}
1009{ .mfi
1010      nop.m 999
1011(p9)  fma.s1 FR_result_lo = FR_T_scale, f1, FR_result_lo // If |K| <= 10
1012      nop.i 999
1013}
1014;;
1015
1016{ .mfi
1017      nop.m 999
1018      fmpy.s0 FR_tmp = FR_A1, FR_A1         // Dummy op to set inexact
1019      nop.i 999
1020}
1021{ .mfb
1022      nop.m 999
1023(p15) fma.s0 f8 = FR_Y_hi, FR_scale, FR_result_lo  // Safe result
1024(p15) br.ret.sptk b0                        // Safe exit for normal path
1025}
1026;;
1027
1028// Here if unsafe, will only be here for exp with K < big_expo_neg
1029{ .mfb
1030      nop.m 999
1031      fma.s0 FR_RESULT = FR_Y_hi, FR_scale, FR_result_lo  // Prelim result
1032      br.cond.sptk EXP_POSSIBLE_UNDERFLOW  // Branch to unsafe code
1033}
1034;;
1035
1036
1037EXP_SMALL:
1038// Here if 2^-60 < |x| < 2^-m, m=12 for exp, m=7 for expm1
1039{ .mfi
1040(p7)  ldfe FR_Q3 = [GR_ad_Q],16          // Get Q3 for small path, if expm1
1041(p6)  fma.s1 FR_p65 = FR_P6, FR_r, FR_P5  // If exp, p65 = P6 * r + P5
1042      nop.i 999
1043}
1044{ .mfi
1045      mov GR_minus_one = -1
1046(p7)  fma.s1 FR_q98 = FR_Q9, FR_r, FR_Q8  // If expm1, q98 = Q9 * r + Q8
1047      nop.i 999
1048}
1049;;
1050
1051{ .mfi
1052(p7)  ldfe FR_Q2 = [GR_ad_Q],16           // Get Q2 for small path, if expm1
1053(p7)  fma.s1 FR_q65 = FR_Q6, FR_r, FR_Q5  // If expm1, q65 = Q6 * r + Q5
1054      nop.i 999
1055}
1056;;
1057
1058{ .mfi
1059      setf.sig FR_tmp = GR_minus_one      // Create value to force inexact
1060(p6)  fma.s1 FR_p21 = FR_P2, FR_r, FR_P1  // If exp, p21 = P2 * r + P1
1061      nop.i 999
1062}
1063{ .mfi
1064(p7)  ldfe FR_Q1 = [GR_ad_Q],16           // Get Q1 for small path, if expm1
1065(p7)  fma.s1 FR_q43 = FR_Q4, FR_r, FR_Q3  // If expm1, q43 = Q4 * r + Q3
1066      nop.i 999
1067}
1068;;
1069
1070{ .mfi
1071      nop.m 999
1072(p6)  fma.s1 FR_p654 = FR_p65, FR_r, FR_P4 // If exp, p654 = p65 * r + P4
1073      nop.i 999
1074}
1075{ .mfi
1076      nop.m 999
1077(p7)  fma.s1 FR_q987 = FR_q98, FR_r, FR_Q7 // If expm1, q987 = q98 * r + Q7
1078      nop.i 999
1079}
1080;;
1081
1082{ .mfi
1083      nop.m 999
1084(p7)  fma.s1 FR_q21 = FR_Q2, FR_r, FR_Q1  // If expm1, q21 = Q2 * r + Q1
1085      nop.i 999
1086}
1087;;
1088
1089{ .mfi
1090      nop.m 999
1091(p6)  fma.s1 FR_p210 = FR_p21, FR_rsq, FR_r // If exp, p210 = p21 * r + P0
1092      nop.i 999
1093}
1094{ .mfi
1095      nop.m 999
1096(p7)  fma.s1 FR_q6543 = FR_q65, FR_rsq, FR_q43 // If expm1, q6543 = q65*r2+q43
1097      nop.i 999
1098}
1099;;
1100
1101{ .mfi
1102      nop.m 999
1103(p6)  fma.s1 FR_p6543 = FR_p654, FR_r, FR_P3 // If exp, p6543 = p654 * r + P3
1104      nop.i 999
1105}
1106{ .mfi
1107      nop.m 999
1108(p7)  fma.s1 FR_q9876543 = FR_q987, FR_r4, FR_q6543 // If expm1, q9876543 = ...
1109      nop.i 999
1110}
1111;;
1112
1113{ .mfi
1114      nop.m 999
1115(p6)  fma.s1 FR_Y_lo = FR_p6543, FR_r4, FR_p210 // If exp, form Y_lo
1116      nop.i 999
1117}
1118;;
1119
1120{ .mfi
1121      nop.m 999
1122(p7)  fma.s1 FR_Y_lo = FR_q9876543, FR_rsq, FR_q21 // If expm1, form Y_lo
1123      nop.i 999
1124}
1125;;
1126
1127{ .mfi
1128      nop.m 999
1129      fmpy.s0  FR_tmp = FR_tmp, FR_tmp   // Dummy op to set inexact
1130      nop.i 999
1131}
1132;;
1133
1134.pred.rel "mutex",p6,p7
1135{ .mfi
1136      nop.m 999
1137(p6)  fma.s0 f8 = FR_Y_lo, f1, f1          // If exp, result = 1 + Y_lo
1138      nop.i 999
1139}
1140{ .mfb
1141      nop.m 999
1142(p7)  fma.s0 f8 = FR_Y_lo, FR_rsq, FR_norm_x // If expm1, result = Y_lo*r2+x
1143      br.ret.sptk  b0                      // Exit for 2^-60 <= |x| < 2^-m
1144                                           // m=12 for exp, m=7 for expm1
1145}
1146;;
1147
1148
1149EXP_VERY_SMALL:
1150//
1151// Here if 0 < |x| < 2^-60
1152// If exp, result = 1.0 + x
1153// If expm1, result = x +x*x/2, but have to check for possible underflow
1154//
1155
1156{ .mfi
1157(p7)  mov GR_exp_underflow = -16381        // Exponent for possible underflow
1158(p6)  fadd.s0 f8 = f1, FR_norm_x           // If exp, result = 1+x
1159      nop.i 999
1160}
1161{ .mfi
1162      nop.m 999
1163(p7)  fmpy.s1 FR_result_lo = FR_half_x, FR_norm_x  // If expm1 result_lo = x*x/2
1164      nop.i 999
1165}
1166;;
1167
1168{ .mfi
1169(p7)  cmp.lt.unc p0, p8 = GR_exp_x, GR_exp_underflow // Unsafe if expm1 x small
1170(p7)  mov FR_Y_hi = FR_norm_x              // If expm1, Y_hi = x
1171(p7)  cmp.lt p0, p15 = GR_exp_x, GR_exp_underflow // Unsafe if expm1 x small
1172}
1173;;
1174
1175{ .mfb
1176      nop.m 999
1177(p8)  fma.s0 f8 = FR_norm_x, f1, FR_result_lo // If expm1, result=x+x*x/2
1178(p15) br.ret.sptk b0                       // If Safe, exit
1179}
1180;;
1181
1182// Here if expm1 and 0 < |x| < 2^-16381;  may be possible underflow
1183{ .mfb
1184      nop.m 999
1185      fma.s0 FR_RESULT = FR_Y_hi, FR_scale, FR_result_lo // Prelim result
1186      br.cond.sptk EXP_POSSIBLE_UNDERFLOW  // Branch to unsafe code
1187}
1188;;
1189
1190EXP_CERTAIN_UNDERFLOW_ZERO:
1191// Here if x < zero_uflow_x
1192// For exp, set result to tiny+0.0 and set I, U, and branch to error handling
1193// For expm1, set result to tiny-1.0 and set I, and exit
1194{ .mmi
1195      alloc GR_SAVE_PFS = ar.pfs,0,3,4,0
1196      nop.m 999
1197      mov GR_one = 1
1198}
1199;;
1200
1201{ .mmi
1202      setf.exp FR_small = GR_one               // Form small value
1203      nop.m 999
1204(p6)  mov GR_Parameter_TAG = 13                // Error tag for exp underflow
1205}
1206;;
1207
1208{ .mfi
1209      nop.m 999
1210      fmerge.s FR_X = f8,f8                    // Save x for error call
1211      nop.i 999
1212}
1213;;
1214
1215.pred.rel "mutex",p6,p7
1216{ .mfb
1217      nop.m 999
1218(p6)  fma.s0 FR_RESULT = FR_small, FR_small, f0 // If exp, set I,U, tiny result
1219(p6)  br.cond.sptk __libm_error_region          // If exp, go to error handling
1220}
1221{ .mfb
1222      nop.m 999
1223(p7)  fms.s0 f8 = FR_small, FR_small, f1        // If expm1, set I, result -1.0
1224(p7)  br.ret.sptk  b0                           // If expm1, exit
1225}
1226;;
1227
1228
1229EXP_OVERFLOW:
1230// Here if x >= min_oflow_x
1231{ .mmi
1232      alloc GR_SAVE_PFS = ar.pfs,0,3,4,0
1233      mov GR_huge_exp = 0x1fffe
1234      nop.i 999
1235}
1236{ .mfi
1237      mov GR_huge_signif = -0x1
1238      nop.f 999
1239(p6)  mov GR_Parameter_TAG = 12                // Error tag for exp overflow
1240}
1241;;
1242
1243{ .mmf
1244      setf.exp FR_huge_exp = GR_huge_exp       // Create huge value
1245      setf.sig FR_huge_signif = GR_huge_signif // Create huge value
1246      fmerge.s FR_X = f8,f8                    // Save x for error call
1247}
1248;;
1249
1250{ .mfi
1251      nop.m 999
1252      fmerge.se FR_huge = FR_huge_exp, FR_huge_signif
1253(p7)  mov GR_Parameter_TAG = 39                // Error tag for expm1 overflow
1254}
1255;;
1256
1257{ .mfb
1258      nop.m 999
1259      fma.s0 FR_RESULT = FR_huge, FR_huge, FR_huge // Force I, O, and Inf
1260      br.cond.sptk __libm_error_region         // Branch to error handling
1261}
1262;;
1263
1264
1265
1266EXP_POSSIBLE_UNDERFLOW:
1267// Here if exp and zero_uflow_x < x < about -11356 [where k < -16381]
1268// Here if expm1 and |x| < 2^-16381
1269{ .mfi
1270      alloc GR_SAVE_PFS = ar.pfs,0,3,4,0
1271      fsetc.s2 0x7F,0x41                   // Set FTZ and disable traps
1272      nop.i 999
1273}
1274;;
1275
1276{ .mfi
1277      nop.m 999
1278      fma.s2 FR_ftz = FR_Y_hi, FR_scale, FR_result_lo   // Result with FTZ
1279      nop.i 999
1280}
1281;;
1282
1283{ .mfi
1284      nop.m 999
1285      fsetc.s2 0x7F,0x40                   // Disable traps (set s2 default)
1286      nop.i 999
1287}
1288;;
1289
1290{ .mfi
1291      nop.m 999
1292(p6)  fclass.m.unc p11, p0 = FR_ftz, 0x00F // If exp, FTZ result denorm or zero?
1293      nop.i 999
1294}
1295;;
1296
1297{ .mfb
1298(p11) mov   GR_Parameter_TAG = 13             // exp underflow
1299      fmerge.s FR_X = f8,f8                   // Save x for error call
1300(p11) br.cond.spnt __libm_error_region        // Branch on exp underflow
1301}
1302;;
1303
1304{ .mfb
1305      nop.m 999
1306      mov   f8     = FR_RESULT                // Was safe after all
1307      br.ret.sptk   b0
1308}
1309;;
1310
1311
1312EXP_64_SPECIAL:
1313// Here if x natval, nan, inf, zero
1314// If x natval, +inf, or if expm1 and x zero, just return x.
1315// The other cases must be tested for, and results set.
1316// These cases do not generate exceptions.
1317{ .mfi
1318      nop.m 999
1319      fclass.m p8, p0 =  f8, 0x0c3            // Is x nan?
1320      nop.i 999
1321}
1322;;
1323
1324{ .mfi
1325      nop.m 999
1326(p6)  fclass.m.unc p13, p0 =  f8, 0x007       // If exp, is x zero?
1327      nop.i 999
1328}
1329;;
1330
1331{ .mfi
1332      nop.m 999
1333(p6)  fclass.m.unc p11, p0 =  f8, 0x022       // If exp, is x -inf?
1334      nop.i 999
1335}
1336{ .mfi
1337      nop.m 999
1338(p8)  fadd.s0 f8 = f8, f1                     // If x nan, result quietized x
1339      nop.i 999
1340}
1341;;
1342
1343{ .mfi
1344      nop.m 999
1345(p7)  fclass.m.unc p10, p0 =  f8, 0x022       // If expm1, is x -inf?
1346      nop.i 999
1347}
1348{ .mfi
1349      nop.m 999
1350(p13) fadd.s0 f8 = f0, f1                     // If exp and x zero, result 1.0
1351      nop.i 999
1352}
1353;;
1354
1355{ .mfi
1356      nop.m 999
1357(p11) mov f8 = f0                             // If exp and x -inf, result 0
1358      nop.i 999
1359}
1360;;
1361
1362{ .mfb
1363      nop.m 999
1364(p10) fsub.s1 f8 = f0, f1                     // If expm1, x -inf, result -1.0
1365      br.ret.sptk b0                          // Exit special cases
1366}
1367;;
1368
1369
1370EXP_64_UNSUPPORTED:
1371// Here if x unsupported type
1372{ .mfb
1373      nop.m 999
1374      fmpy.s0 f8 = f8, f0                     // Return nan
1375      br.ret.sptk   b0
1376}
1377;;
1378
1379GLOBAL_IEEE754_END(expl)
1380libm_alias_ldouble_other (__exp, exp)
1381
1382LOCAL_LIBM_ENTRY(__libm_error_region)
1383.prologue
1384{ .mfi
1385        add   GR_Parameter_Y=-32,sp             // Parameter 2 value
1386        nop.f 0
1387.save   ar.pfs,GR_SAVE_PFS
1388        mov  GR_SAVE_PFS=ar.pfs                 // Save ar.pfs
1389}
1390{ .mfi
1391.fframe 64
1392        add sp=-64,sp                           // Create new stack
1393        nop.f 0
1394        mov GR_SAVE_GP=gp                       // Save gp
1395};;
1396{ .mmi
1397        stfe [GR_Parameter_Y] = FR_Y,16         // Save Parameter 2 on stack
1398        add GR_Parameter_X = 16,sp              // Parameter 1 address
1399.save   b0, GR_SAVE_B0
1400        mov GR_SAVE_B0=b0                       // Save b0
1401};;
1402.body
1403{ .mib
1404        stfe [GR_Parameter_X] = FR_X            // Store Parameter 1 on stack
1405        add   GR_Parameter_RESULT = 0,GR_Parameter_Y
1406        nop.b 0                                 // Parameter 3 address
1407}
1408{ .mib
1409        stfe [GR_Parameter_Y] = FR_RESULT      // Store Parameter 3 on stack
1410        add   GR_Parameter_Y = -16,GR_Parameter_Y
1411        br.call.sptk b0=__libm_error_support#  // Call error handling function
1412};;
1413{ .mmi
1414        add   GR_Parameter_RESULT = 48,sp
1415        nop.m 0
1416        nop.i 0
1417};;
1418{ .mmi
1419        ldfe  f8 = [GR_Parameter_RESULT]       // Get return result off stack
1420.restore sp
1421        add   sp = 64,sp                       // Restore stack pointer
1422        mov   b0 = GR_SAVE_B0                  // Restore return address
1423};;
1424{ .mib
1425        mov   gp = GR_SAVE_GP                  // Restore gp
1426        mov   ar.pfs = GR_SAVE_PFS             // Restore ar.pfs
1427        br.ret.sptk     b0                     // Return
1428};;
1429LOCAL_LIBM_END(__libm_error_region#)
1430
1431.type   __libm_error_support#,@function
1432.global __libm_error_support#
1433