4 // Copyright (c) 2000 - 2003, 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 // 09/06/02 Improved performance and accuracy; no inexact flags on exact cases
45 // 01/29/03 Added missing } to bundle templates
48 //==============================================================
49 // float exp10f(float)
51 // Overview of operation
52 //==============================================================
57 // Let x= (K + fh + fl + r)/log2(10), where
58 // K is an integer, fh= 0.b1 b2 b3 b4 b5,
59 // fl= 2^{-5}* 0.b6 b7 b8 b8 b10 (fh, fl >= 0),
61 // Th is a table that stores 2^fh (32 entries) rounded to
62 // double extended precision (only mantissa is stored)
63 // Tl is a table that stores 2^fl (32 entries) rounded to
64 // double extended precision (only mantissa is stored)
66 // 10^x is approximated as
67 // 2^K * Th [ f ] * Tl [ f ] * (1+c1*r+c2*r^2)
69 // Note there are only 10 non-zero values that produce an exact result:
70 // 1.0, 2.0, ... 10.0.
71 // We test for these cases and use s1 to avoid setting the inexact flag.
74 //==============================================================
81 //==============================================================
118 GR_Parameter_RESULT = r39
119 GR_Parameter_TAG = r40
159 //==============================================================
165 LOCAL_OBJECT_START(poly_coeffs)
167 data8 0xd49a784bcd1b8afe, 0x00003fcb // log2(10)*2^(10-63)
168 data8 0xb17217f7d1cf79ab, 0x00004033 // C_1 * 2^53
169 data8 0xf5fdeffc162c7541, 0x00004066 // C_2 * 2^106
170 LOCAL_OBJECT_END(poly_coeffs)
173 LOCAL_OBJECT_START(T_table)
175 // 2^{0.00000 b6 b7 b8 b9 b10}
176 data8 0x8000000000000000, 0x8016302f17467628
177 data8 0x802c6436d0e04f50, 0x80429c17d77c18ed
178 data8 0x8058d7d2d5e5f6b0, 0x806f17687707a7af
179 data8 0x80855ad965e88b83, 0x809ba2264dada76a
180 data8 0x80b1ed4fd999ab6c, 0x80c83c56b50cf77f
181 data8 0x80de8f3b8b85a0af, 0x80f4e5ff089f763e
182 data8 0x810b40a1d81406d4, 0x81219f24a5baa59d
183 data8 0x813801881d886f7b, 0x814e67cceb90502c
184 data8 0x8164d1f3bc030773, 0x817b3ffd3b2f2e47
185 data8 0x8191b1ea15813bfd, 0x81a827baf7838b78
186 data8 0x81bea1708dde6055, 0x81d51f0b8557ec1c
187 data8 0x81eba08c8ad4536f, 0x820225f44b55b33b
188 data8 0x8218af4373fc25eb, 0x822f3c7ab205c89a
189 data8 0x8245cd9ab2cec048, 0x825c62a423d13f0c
190 data8 0x8272fb97b2a5894c, 0x828998760d01faf3
191 data8 0x82a0393fe0bb0ca8, 0x82b6ddf5dbc35906
194 // 2^{0.b1 b2 b3 b4 b5}
195 data8 0x8000000000000000, 0x82cd8698ac2ba1d7
196 data8 0x85aac367cc487b14, 0x88980e8092da8527
197 data8 0x8b95c1e3ea8bd6e6, 0x8ea4398b45cd53c0
198 data8 0x91c3d373ab11c336, 0x94f4efa8fef70961
199 data8 0x9837f0518db8a96f, 0x9b8d39b9d54e5538
200 data8 0x9ef5326091a111ad, 0xa27043030c496818
201 data8 0xa5fed6a9b15138ea, 0xa9a15ab4ea7c0ef8
202 data8 0xad583eea42a14ac6, 0xb123f581d2ac258f
203 data8 0xb504f333f9de6484, 0xb8fbaf4762fb9ee9
204 data8 0xbd08a39f580c36be, 0xc12c4cca66709456
205 data8 0xc5672a115506dadd, 0xc9b9bd866e2f27a2
206 data8 0xce248c151f8480e3, 0xd2a81d91f12ae45a
207 data8 0xd744fccad69d6af4, 0xdbfbb797daf23755
208 data8 0xe0ccdeec2a94e111, 0xe5b906e77c8348a8
209 data8 0xeac0c6e7dd24392e, 0xefe4b99bdcdaf5cb
210 data8 0xf5257d152486cc2c, 0xfa83b2db722a033a
211 LOCAL_OBJECT_END(T_table)
216 GLOBAL_IEEE754_ENTRY(exp10f)
220 alloc r32= ar.pfs, 1, 4, 4, 0
221 // will continue only for non-zero normal/denormal numbers
222 fclass.nm.unc p12, p7= f8, 0x1b
226 // GR_TBL_START= pointer to log2(10), C_1...C_4 followed by T_table
227 addl GR_TBL_START= @ltoff(poly_coeffs), gp
228 movl GR_ROUNDVAL= 0x3fc00000 // 1.5 (SP)
233 ld8 GR_COEFF_START= [ GR_TBL_START ] // Load pointer to coeff table
234 fcmp.lt.s1 p6, p8= f8, f0 // X<0 ?
241 movl GR_UF_LIMIT= 0xc2349e35 // (-2^7-22) / log2(10)
244 setf.s FR_ROUNDVAL= GR_ROUNDVAL
245 movl GR_OF_LIMIT= 0x421a209a // Overflow threshold
250 ldfe FR_LOG2_10= [ GR_COEFF_START ], 16 // load log2(10)*2^(10-63)
252 (p12) br.cond.spnt SPECIAL_exp10 // Branch if nan, inf, zero
257 setf.s FR_OF_LIMIT= GR_OF_LIMIT // Set overflow limit
258 fma.s0 f8= f8, f1, f0 // normalize x
265 (p8) fcvt.fx.s1 FR_int_x = f8 // Convert x to integer
269 setf.s FR_UF_LIMIT= GR_UF_LIMIT // Set underflow limit
270 fma.s1 FR_KF0= f8, FR_LOG2_10, FR_ROUNDVAL // y= (x*log2(10)*2^10 +
271 // 1.5*2^63) * 2^(-63)
272 mov GR_EXP_CORR= 0xffff-126
277 ldfe FR_COEFF1= [ GR_COEFF_START ], 16 // load C_1
278 fms.s1 FR_KF= FR_KF0, f1, FR_ROUNDVAL // (K+f)*2^(10-63)
284 ldfe FR_COEFF2= [ GR_COEFF_START ], 16 // load C_2
291 getf.sig GR_KF0= FR_KF0 // (K+f)*2^10= round_to_int(y)
292 (p8) movl GR_exact_limit= 0x41200000 // Largest x for exact result,
298 add GR_LOG_TBL= 256, GR_COEFF_START // Pointer to high T_table
299 fcmp.gt.s1 p12, p7= f8, FR_OF_LIMIT // x>overflow threshold ?
305 (p8) setf.s FR_exact_limit = GR_exact_limit // Largest x for exact result
306 (p8) fcvt.xf FR_int_x = FR_int_x // Integral part of x
307 shr GR_K= GR_KF0, 10 // K
310 and GR_F_high= GR_MASK, GR_KF0 // f_high*32
311 fms.s1 FR_R= f8, FR_LOG2_10, FR_KF // r*2^(-53)= [ x*log2(10)-
312 // (K+f) ] *2^{10-63}
313 and GR_F_low= GR_KF0, GR_MASK_low // f_low
318 shladd GR_Flow_ADDR= GR_F_low, 3, GR_COEFF_START // address of 2^{f_low}
319 add GR_BIAS= GR_K, GR_EXP_CORR // K= bias-2*63
320 shr GR_Fh= GR_F_high, 5 // f_high
325 setf.exp FR_2_TO_K= GR_BIAS // 2^{K-126}
326 (p7) fcmp.lt.s1 p12, p7= f8, FR_UF_LIMIT // x<underflow threshold ?
327 shladd GR_Fh_ADDR= GR_Fh, 3, GR_LOG_TBL // address of 2^{f_high}
330 ldf8 FR_T_low= [ GR_Flow_ADDR ] // load T_low= 2^{f_low}
337 ldf8 FR_T_high= [ GR_Fh_ADDR ] // load T_high= 2^{f_high}
339 (p12) br.cond.spnt OUT_RANGE_exp10
345 fma.s1 FR_P12= FR_COEFF2, FR_R, FR_COEFF1 // P12= C_1+C_2*r
346 cmp.eq p7,p9= r0,r0 // Assume inexact result
352 (p8) fcmp.eq.s1 p9,p7= FR_int_x, f8 // Test x positive integer
357 fma.s1 FR_T_low_K= FR_T_low, FR_2_TO_K, f0 // T= 2^{K-126}*T_low
364 fma.s1 FR_P= FR_P12, FR_R, f0 // P= P12*r
369 // If x a positive integer, will it produce an exact result?
370 // p7 result will be inexact
371 // p9 result will be exact
374 (p9) fcmp.le.s1 p9,p7= f8, FR_exact_limit // Test x gives exact result
379 fma.s1 FR_T= FR_T_low_K, FR_T_high, f0 // T= T*T_high
384 .pred.rel "mutex",p7,p9
387 (p7) fma.s.s0 f8= FR_P, FR_T, FR_T // result= T+T*P, inexact set
392 (p9) fma.s.s1 f8= FR_P, FR_T, FR_T // result= T+T*P, exact use s1
393 br.ret.sptk b0 // return
401 fclass.m p6, p0= f8, 0x22 // x= -Infinity ?
408 fclass.m p7, p0= f8, 0x21 // x= +Infinity ?
415 fclass.m p8, p0= f8, 0x7 // x= +/-Zero ?
420 (p6) mov f8= f0 // exp10(-Infinity)= 0
428 (p7) br.ret.spnt b0 // exp10(+Infinity)= +Infinity
434 (p8) mov f8= f1 // exp10(+/-0)= 1
441 fma.s.s0 f8= f8, f1, f0 // Remaining cases: NaNs
452 (p8) mov GR_EXPMAX= 0x1fffe
460 (p8) mov GR_Parameter_TAG= 167
461 (p8) setf.exp FR_R= GR_EXPMAX
468 (p8) fma.s.s0 f8= FR_R, FR_R, f0 // Create overflow
474 (p6) mov GR_EXPMAX= 1
481 (p6) setf.exp FR_R= GR_EXPMAX
488 (p6) fma.s.s0 f8= FR_R, FR_R, f0 // Create underflow
489 (p6) br.ret.sptk b0 // will not call libm_error for underflow
493 GLOBAL_IEEE754_END(exp10f)
494 weak_alias (exp10f, pow10f)
496 LOCAL_LIBM_ENTRY(__libm_error_region)
500 add GR_Parameter_Y= -32, sp // Parameter 2 value
502 .save ar.pfs, GR_SAVE_PFS
503 mov GR_SAVE_PFS= ar.pfs // Save ar.pfs
508 add sp= -64, sp // Create new stack
510 mov GR_SAVE_GP= gp // Save gp
515 stfs [ GR_Parameter_Y ]= FR_Y, 16 // STORE Parameter 2 on stack
516 add GR_Parameter_X= 16, sp // Parameter 1 address
518 mov GR_SAVE_B0= b0 // Save b0
524 stfs [ GR_Parameter_X ]= FR_X // STORE Parameter 1 on stack
525 add GR_Parameter_RESULT= 0, GR_Parameter_Y // Parameter 3 address
529 stfs [ GR_Parameter_Y ]= FR_RESULT // STORE Parameter 3 on stack
530 add GR_Parameter_Y= -16, GR_Parameter_Y
531 br.call.sptk b0= __libm_error_support# // Call error handling function
536 add GR_Parameter_RESULT= 48, sp
543 ldfs f8= [ GR_Parameter_RESULT ] // Get return result off stack
545 add sp= 64, sp // Restore stack pointer
546 mov b0= GR_SAVE_B0 // Restore return address
551 mov gp= GR_SAVE_GP // Restore gp
552 mov ar.pfs= GR_SAVE_PFS // Restore ar.pfs
553 br.ret.sptk b0 // Return
558 LOCAL_LIBM_END(__libm_error_region)
560 .type __libm_error_support#, @function
561 .global __libm_error_support#