4
// Copyright (c) 2000 - 2005, Intel Corporation
5
// All rights reserved.
7
// Contributed 2000 by the Intel Numerics Group, Intel Corporation
9
// Redistribution and use in source and binary forms, with or without
10
// modification, are permitted provided that the following conditions are
13
// * Redistributions of source code must retain the above copyright
14
// notice, this list of conditions and the following disclaimer.
16
// * Redistributions in binary form must reproduce the above copyright
17
// notice, this list of conditions and the following disclaimer in the
18
// documentation and/or other materials provided with the distribution.
20
// * The name of Intel Corporation may not be used to endorse or promote
21
// products derived from this software without specific prior written
24
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
27
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
28
// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
29
// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
30
// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
31
// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
32
// OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
33
// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
34
// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36
// Intel Corporation is the author of this code, and requests that all
37
// problem reports or change requests be submitted to it directly at
38
// http://www.intel.com/software/products/opensource/libraries/num.htm.
41
//==============================================================
42
// 08/25/00 Initial version
43
// 05/20/02 Cleaned up namespace and sf0 syntax
44
// 09/05/02 Improved performance
45
// 01/17/03 Fixed to call error support when x=1024.0
46
// 03/31/05 Reformatted delimiters between data tables
49
//==============================================================
50
// double exp2(double)
52
// Overview of operation
53
//==============================================================
58
// Let x= (K + fh + fl + r), where
59
// K is an integer, fh= 0.b1 b2 b3 b4 b5,
60
// fl= 2^{-5}* 0.b6 b7 b8 b8 b10 (fh, fl >= 0),
62
// Th is a table that stores 2^fh (32 entries) rounded to
63
// double extended precision (only mantissa is stored)
64
// Tl is a table that stores 2^fl (32 entries) rounded to
65
// double extended precision (only mantissa is stored)
67
// 2^x is approximated as
68
// 2^K * Th [ f ] * Tl [ f ] * (1+c1*r+c2*r^2+c3*r^3+c4*r^4)
70
// Note: We use the following trick to speed up conversion from FP to integer:
72
// Let x = K + r, where K is an integer, and |r| <= 0.5
73
// Let N be the number of significand bits for the FP format used
74
// ( N=64 for double-extended, N=53 for double)
76
// Then let y = 1.5 * 2^(N-1) + x for RN mode
77
// K = y - 1.5 * 2^(N-1)
80
// If we want to obtain the integer part and the first m fractional bits of x,
81
// we can use the same trick, but with a constant of 1.5 * 2^(N-1-m):
84
// f = 0.b_1 b_2 ... b_m
87
// Then let y = 1.5 * 2^(N-1-m) + x for RN mode
88
// (K+f) = y - 1.5 * 2^(N-1-m)
93
//==============================================================
100
//==============================================================
137
GR_Parameter_RESULT = r39
138
GR_Parameter_TAG = r40
174
//==============================================================
180
LOCAL_OBJECT_START(poly_coeffs)
182
data8 0x3fac6b08d704a0c0, 0x3f83b2ab6fba4e77 // C_3 and C_4
183
data8 0xb17217f7d1cf79ab, 0x00003ffe // C_1
184
data8 0xf5fdeffc162c7541, 0x00003ffc // C_2
185
LOCAL_OBJECT_END(poly_coeffs)
188
LOCAL_OBJECT_START(T_table)
190
// 2^{0.00000 b6 b7 b8 b9 b10}
191
data8 0x8000000000000000, 0x8016302f17467628
192
data8 0x802c6436d0e04f50, 0x80429c17d77c18ed
193
data8 0x8058d7d2d5e5f6b0, 0x806f17687707a7af
194
data8 0x80855ad965e88b83, 0x809ba2264dada76a
195
data8 0x80b1ed4fd999ab6c, 0x80c83c56b50cf77f
196
data8 0x80de8f3b8b85a0af, 0x80f4e5ff089f763e
197
data8 0x810b40a1d81406d4, 0x81219f24a5baa59d
198
data8 0x813801881d886f7b, 0x814e67cceb90502c
199
data8 0x8164d1f3bc030773, 0x817b3ffd3b2f2e47
200
data8 0x8191b1ea15813bfd, 0x81a827baf7838b78
201
data8 0x81bea1708dde6055, 0x81d51f0b8557ec1c
202
data8 0x81eba08c8ad4536f, 0x820225f44b55b33b
203
data8 0x8218af4373fc25eb, 0x822f3c7ab205c89a
204
data8 0x8245cd9ab2cec048, 0x825c62a423d13f0c
205
data8 0x8272fb97b2a5894c, 0x828998760d01faf3
206
data8 0x82a0393fe0bb0ca8, 0x82b6ddf5dbc35906
208
// 2^{0.b1 b2 b3 b4 b5}
209
data8 0x8000000000000000, 0x82cd8698ac2ba1d7
210
data8 0x85aac367cc487b14, 0x88980e8092da8527
211
data8 0x8b95c1e3ea8bd6e6, 0x8ea4398b45cd53c0
212
data8 0x91c3d373ab11c336, 0x94f4efa8fef70961
213
data8 0x9837f0518db8a96f, 0x9b8d39b9d54e5538
214
data8 0x9ef5326091a111ad, 0xa27043030c496818
215
data8 0xa5fed6a9b15138ea, 0xa9a15ab4ea7c0ef8
216
data8 0xad583eea42a14ac6, 0xb123f581d2ac258f
217
data8 0xb504f333f9de6484, 0xb8fbaf4762fb9ee9
218
data8 0xbd08a39f580c36be, 0xc12c4cca66709456
219
data8 0xc5672a115506dadd, 0xc9b9bd866e2f27a2
220
data8 0xce248c151f8480e3, 0xd2a81d91f12ae45a
221
data8 0xd744fccad69d6af4, 0xdbfbb797daf23755
222
data8 0xe0ccdeec2a94e111, 0xe5b906e77c8348a8
223
data8 0xeac0c6e7dd24392e, 0xefe4b99bdcdaf5cb
224
data8 0xf5257d152486cc2c, 0xfa83b2db722a033a
225
LOCAL_OBJECT_END(T_table)
230
GLOBAL_LIBM_ENTRY(exp2)
234
alloc r32= ar.pfs, 1, 4, 4, 0
235
// will continue only for non-zero normal/denormal numbers
236
fclass.nm p12, p0= f8, 0x1b
237
// GR_TBL_START= pointer to C_1...C_4 followed by T_table
238
addl GR_TBL_START= @ltoff(poly_coeffs), gp
241
mov GR_OF_LIMIT= 0xffff + 10 // Exponent of overflow limit
242
movl GR_ROUNDVAL= 0x5a400000 // 1.5*2^(63-10) (SP)
246
// Form special constant 1.5*2^(63-10) to give integer part and first 10
247
// fractional bits of x
249
setf.s FR_ROUNDVAL= GR_ROUNDVAL // Form special constant
250
fcmp.lt.s1 p6, p8= f8, f0 // X<0 ?
254
ld8 GR_COEFF_START= [ GR_TBL_START ] // Load pointer to coeff table
256
(p12) br.cond.spnt SPECIAL_exp2 // Branch if nan, inf, zero
261
setf.exp FR_OF_LIMIT= GR_OF_LIMIT // Set overflow limit
262
movl GR_UF_LIMIT= 0xc4866000 // (-2^10-51) = -1075
267
ldfpd FR_COEFF3, FR_COEFF4= [ GR_COEFF_START ], 16 // load C_3, C_4
268
fma.s0 f8= f8, f1, f0 // normalize x
274
setf.s FR_UF_LIMIT= GR_UF_LIMIT // Set underflow limit
275
ldfe FR_COEFF1= [ GR_COEFF_START ], 16 // load C_1
276
mov GR_EXP_CORR= 0xffff-126
281
ldfe FR_COEFF2= [ GR_COEFF_START ], 16 // load C_2
282
fma.s1 FR_KF0= f8, f1, FR_ROUNDVAL // y= x + 1.5*2^(63-10)
289
fms.s1 FR_KF= FR_KF0, f1, FR_ROUNDVAL // (K+f)
295
getf.sig GR_KF0= FR_KF0 // (K+f)*2^10= round_to_int(y)
296
fcmp.ge.s1 p12, p7= f8, FR_OF_LIMIT // x >= overflow threshold ?
297
add GR_LOG_TBL= 256, GR_COEFF_START // Pointer to high T_table
302
and GR_F_low= GR_KF0, GR_MASK_low // f_low
303
and GR_F_high= GR_MASK, GR_KF0 // f_high*32
304
shr GR_K= GR_KF0, 10 // K
309
shladd GR_Flow_ADDR= GR_F_low, 3, GR_COEFF_START // address of 2^{f_low}
310
add GR_BIAS= GR_K, GR_EXP_CORR // K= bias-2*63
311
shr GR_Fh= GR_F_high, 5 // f_high
316
setf.exp FR_2_TO_K= GR_BIAS // 2^{K-126}
317
fnma.s1 FR_R= FR_KF, f1, f8 // r= x - (K+f)
318
shladd GR_Fh_ADDR= GR_Fh, 3, GR_LOG_TBL // address of 2^{f_high}
321
ldf8 FR_T_low= [ GR_Flow_ADDR ] // load T_low= 2^{f_low}
322
movl GR_EMIN= 0xc47f8000 // EMIN= -1022
327
ldf8 FR_T_high= [ GR_Fh_ADDR ] // load T_high= 2^{f_high}
328
(p7) fcmp.lt.s1 p12, p7= f8, FR_UF_LIMIT // x<underflow threshold ?
334
setf.s FR_EXPMIN= GR_EMIN // FR_EXPMIN= EMIN
335
fma.s1 FR_P34= FR_COEFF4, FR_R, FR_COEFF3 // P34= C_3+C_4*r
340
fma.s1 FR_R2= FR_R, FR_R, f0 // r*r
341
(p12) br.cond.spnt OUT_RANGE_exp2
347
fma.s1 FR_P12= FR_COEFF2, FR_R, FR_COEFF1 // P12= C_1+C_2*r
354
fma.s1 FR_T_low_K= FR_T_low, FR_2_TO_K, f0 // T= 2^{K-126}*T_low
361
fma.s1 FR_P14= FR_R2, FR_P34, FR_P12 // P14= P12+r2*P34
368
fma.s1 FR_T= FR_T_low_K, FR_T_high, f0 // T= T*T_high
375
fcmp.lt.s0 p6, p8= f8, FR_EXPMIN // underflow (x<EMIN) ?
382
fma.s1 FR_P= FR_P14, FR_R, f0 // P= P14*r
389
fma.d.s0 f8= FR_P, FR_T, FR_T // result= T+T*P
390
(p8) br.ret.sptk b0 // return
395
(p6) mov GR_Parameter_TAG= 162
397
(p6) br.cond.sptk __libm_error_region
405
fclass.m p6, p0= f8, 0x22 // x= -Infinity ?
412
fclass.m p7, p0= f8, 0x21 // x= +Infinity ?
419
fclass.m p8, p0= f8, 0x7 // x= +/-Zero ?
424
(p6) mov f8= f0 // exp2(-Infinity)= 0
432
(p7) br.ret.spnt b0 // exp2(+Infinity)= +Infinity
438
(p8) mov f8= f1 // exp2(+/-0)= 1
445
fma.d.s0 f8= f8, f1, f0 // Remaining cases: NaNs
456
(p8) mov GR_EXPMAX= 0x1fffe
463
(p8) mov GR_Parameter_TAG= 161
464
(p8) setf.exp FR_R= GR_EXPMAX
471
(p8) fma.d.s0 f8= FR_R, FR_R, f0 // Create overflow
476
(p6) mov GR_Parameter_TAG= 162
477
(p6) mov GR_EXPMAX= 1
484
(p6) setf.exp FR_R= GR_EXPMAX
491
(p6) fma.d.s0 f8= FR_R, FR_R, f0 // Create underflow
496
GLOBAL_LIBM_END(exp2)
499
LOCAL_LIBM_ENTRY(__libm_error_region)
503
add GR_Parameter_Y= -32, sp // Parameter 2 value
505
.save ar.pfs, GR_SAVE_PFS
506
mov GR_SAVE_PFS= ar.pfs // Save ar.pfs
511
add sp= -64, sp // Create new stack
513
mov GR_SAVE_GP= gp // Save gp
518
stfd [ GR_Parameter_Y ]= FR_Y, 16 // STORE Parameter 2 on stack
519
add GR_Parameter_X= 16, sp // Parameter 1 address
521
mov GR_SAVE_B0= b0 // Save b0
527
stfd [ GR_Parameter_X ]= FR_X // STORE Parameter 1 on stack
528
add GR_Parameter_RESULT= 0, GR_Parameter_Y // Parameter 3 address
532
stfd [ GR_Parameter_Y ]= FR_RESULT // STORE Parameter 3 on stack
533
add GR_Parameter_Y= -16, GR_Parameter_Y
534
br.call.sptk b0= __libm_error_support# // Call error handling function
539
add GR_Parameter_RESULT= 48, sp
546
ldfd f8= [ GR_Parameter_RESULT ] // Get return result off stack
548
add sp= 64, sp // Restore stack pointer
549
mov b0= GR_SAVE_B0 // Restore return address
554
mov gp= GR_SAVE_GP // Restore gp
555
mov ar.pfs= GR_SAVE_PFS // Restore ar.pfs
556
br.ret.sptk b0 // Return
561
LOCAL_LIBM_END(__libm_error_region)
563
.type __libm_error_support#, @function
564
.global __libm_error_support#