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 // 2/02/00 Initial version
43 // 3/07/00 exp(inf) = inf but now does NOT call error support
44 // exp(-inf) = 0 but now does NOT call error support
45 // 4/04/00 Unwind support added
46 // 8/15/00 Bundle added after call to __libm_error_support to properly
47 // set [the previously overwritten] GR_Parameter_RESULT.
48 // 11/30/00 Reworked to shorten main path, widen main path to include all
49 // args in normal range, and add quick exit for 0, nan, inf.
50 // 12/05/00 Loaded constants earlier with setf to save 2 cycles.
51 // 02/05/02 Corrected uninitialize predicate in POSSIBLE_UNDERFLOW path
52 // 05/20/02 Cleaned up namespace and sf0 syntax
53 // 09/07/02 Force inexact flag
54 // 11/15/02 Split underflow path into zero/nonzero; eliminated fma in main path
55 // 05/30/03 Set inexact flag on unmasked overflow/underflow
56 // 03/31/05 Reformatted delimiters between data tables
59 //==============================================================
62 // Overview of operation
63 //==============================================================
64 // Take the input x. w is "how many log2/128 in x?"
67 // x = n log2/128 + r + delta
69 // n = 128M + index_1 + 2^4 index_2
70 // x = M log2 + (log2/128) index_1 + (log2/8) index_2 + r + delta
72 // exp(x) = 2^M 2^(index_1/128) 2^(index_2/8) exp(r) exp(delta)
74 // Get 2^(index_1/128) from table_1;
75 // Get 2^(index_2/8) from table_2;
76 // Calculate exp(r) by 5th order polynomial
77 // r = x - n (log2/128)_high
78 // delta = - n (log2/128)_low
79 // Calculate exp(delta) as 1 + delta
83 //==============================================================
95 // Overflow and Underflow
96 //=======================
97 // exp(x) = largest double normal when
98 // x = 709.7827 = 0x40862e42fefa39ef
100 // exp(x) = smallest double normal when
101 // x = -708.396 = 0xc086232bdd7abcd2
103 // exp(x) = largest round-to-nearest single zero when
104 // x = -745.1332 = 0xc0874910d52d3052
108 //==============================================================
109 // Floating Point registers used:
111 // f6 -> f15, f32 -> f49
113 // General registers used:
116 // Predicate registers used:
120 //==============================================================
150 GR_Parameter_RESULT = r39
151 GR_Parameter_TAG = r40
190 fMIN_DBL_OFLOW_ARG = f45
191 fMAX_DBL_ZERO_ARG = f46
192 fMAX_DBL_NORM_ARG = f47
193 fMIN_DBL_NORM_ARG = f48
199 //==============================================================
204 // ************* DO NOT CHANGE ORDER OF THESE TABLES ********************
206 // double-extended 1/ln(2)
207 // 3fff b8aa 3b29 5c17 f0bb be87fed0691d3e88
208 // 3fff b8aa 3b29 5c17 f0bc
209 // For speed the significand will be loaded directly with a movl and setf.sig
210 // and the exponent will be bias+63 instead of bias+0. Thus subsequent
211 // computations need to scale appropriately.
212 // The constant 128/ln(2) is needed for the computation of w. This is also
213 // obtained by scaling the computations.
215 // Two shifting constants are loaded directly with movl and setf.d.
216 // 1. fRSHF_2TO56 = 1.1000..00 * 2^(63-7)
217 // This constant is added to x*1/ln2 to shift the integer part of
218 // x*128/ln2 into the rightmost bits of the significand.
219 // The result of this fma is fW_2TO56_RSH.
220 // 2. fRSHF = 1.1000..00 * 2^(63)
221 // This constant is subtracted from fW_2TO56_RSH * 2^(-56) to give
222 // the integer part of w, n, as a floating-point number.
223 // The result of this fms is fNfloat.
226 LOCAL_OBJECT_START(exp_table_1)
227 data8 0x40862e42fefa39f0 // smallest dbl overflow arg, +709.7827
228 data8 0xc0874910d52d3052 // largest arg for rnd-to-nearest 0 result, -745.133
229 data8 0x40862e42fefa39ef // largest dbl arg to give normal dbl result, +709.7827
230 data8 0xc086232bdd7abcd2 // smallest dbl arg to give normal dbl result, -708.396
231 data8 0xb17217f7d1cf79ab , 0x00003ff7 // ln2/128 hi
232 data8 0xc9e3b39803f2f6af , 0x00003fb7 // ln2/128 lo
234 // Table 1 is 2^(index_1/128) where
235 // index_1 goes from 0 to 15
237 data8 0x8000000000000000 , 0x00003FFF
238 data8 0x80B1ED4FD999AB6C , 0x00003FFF
239 data8 0x8164D1F3BC030773 , 0x00003FFF
240 data8 0x8218AF4373FC25EC , 0x00003FFF
241 data8 0x82CD8698AC2BA1D7 , 0x00003FFF
242 data8 0x8383594EEFB6EE37 , 0x00003FFF
243 data8 0x843A28C3ACDE4046 , 0x00003FFF
244 data8 0x84F1F656379C1A29 , 0x00003FFF
245 data8 0x85AAC367CC487B15 , 0x00003FFF
246 data8 0x8664915B923FBA04 , 0x00003FFF
247 data8 0x871F61969E8D1010 , 0x00003FFF
248 data8 0x87DB357FF698D792 , 0x00003FFF
249 data8 0x88980E8092DA8527 , 0x00003FFF
250 data8 0x8955EE03618E5FDD , 0x00003FFF
251 data8 0x8A14D575496EFD9A , 0x00003FFF
252 data8 0x8AD4C6452C728924 , 0x00003FFF
253 LOCAL_OBJECT_END(exp_table_1)
255 // Table 2 is 2^(index_1/8) where
256 // index_2 goes from 0 to 7
257 LOCAL_OBJECT_START(exp_table_2)
258 data8 0x8000000000000000 , 0x00003FFF
259 data8 0x8B95C1E3EA8BD6E7 , 0x00003FFF
260 data8 0x9837F0518DB8A96F , 0x00003FFF
261 data8 0xA5FED6A9B15138EA , 0x00003FFF
262 data8 0xB504F333F9DE6484 , 0x00003FFF
263 data8 0xC5672A115506DADD , 0x00003FFF
264 data8 0xD744FCCAD69D6AF4 , 0x00003FFF
265 data8 0xEAC0C6E7DD24392F , 0x00003FFF
266 LOCAL_OBJECT_END(exp_table_2)
269 LOCAL_OBJECT_START(exp_p_table)
270 data8 0x3f8111116da21757 //P5
271 data8 0x3fa55555d787761c //P4
272 data8 0x3fc5555555555414 //P3
273 data8 0x3fdffffffffffd6a //P2
274 LOCAL_OBJECT_END(exp_p_table)
278 GLOBAL_IEEE754_ENTRY(exp)
282 movl rSig_inv_ln2 = 0xb8aa3b295c17f0bc // significand of 1/ln2
285 addl rAD_TB1 = @ltoff(exp_table_1), gp
286 movl rRshf_2to56 = 0x4768000000000000 // 1.10000 2^(63+56)
291 ld8 rAD_TB1 = [rAD_TB1]
292 fclass.m p8,p0 = f8,0x07 // Test for x=0
293 mov rExp_mask = 0x1ffff
296 mov rExp_bias = 0xffff
298 mov rExp_2tom56 = 0xffff-56
302 // Form two constants we need
303 // 1/ln2 * 2^63 to compute w = x * 1/ln2 * 128
304 // 1.1000..000 * 2^(63+63-7) to right shift int(w) into the significand
307 setf.sig fINV_LN2_2TO63 = rSig_inv_ln2 // form 1/ln2 * 2^63
308 fclass.m p9,p0 = f8,0x22 // Test for x=-inf
312 setf.d fRSHF_2TO56 = rRshf_2to56 // Form const 1.100 * 2^(63+56)
313 movl rRshf = 0x43e8000000000000 // 1.10000 2^63 for right shift
318 ldfpd fMIN_DBL_OFLOW_ARG, fMAX_DBL_ZERO_ARG = [rAD_TB1],16
319 fclass.m p10,p0 = f8,0x1e1 // Test for x=+inf, nan, NaT
323 setf.exp f2TOM56 = rExp_2tom56 // form 2^-56 for scaling Nfloat
324 (p9) fma.d.s0 f8 = f0,f0,f0 // quick exit for x=-inf
330 ldfpd fMAX_DBL_NORM_ARG, fMIN_DBL_NORM_ARG = [rAD_TB1],16
335 setf.d fRSHF = rRshf // Form right shift const 1.100 * 2^63
336 (p8) fma.d.s0 f8 = f1,f1,f0 // quick exit for x=0
342 ldfe fLn2_by_128_hi = [rAD_TB1],16
343 (p10) fma.d.s0 f8 = f8,f8,f0 // Result if x=+inf, nan, NaT
344 (p10) br.ret.spnt b0 // quick exit for x=+inf, nan, NaT
349 ldfe fLn2_by_128_lo = [rAD_TB1],16
350 fcmp.eq.s0 p6,p0 = f8, f0 // Dummy to set D
355 // After that last load, rAD_TB1 points to the beginning of table 1
357 // W = X * Inv_log2_by_128
358 // By adding 1.10...0*2^63 we shift and get round_int(W) in significand.
359 // We actually add 1.10...0*2^56 to X * Inv_log2 to do the same thing.
363 fma.s1 fW_2TO56_RSH = fNormX, fINV_LN2_2TO63, fRSHF_2TO56
368 // Divide arguments into the following categories:
369 // Certain Underflow p11 - -inf < x <= MAX_DBL_ZERO_ARG
370 // Possible Underflow p13 - MAX_DBL_ZERO_ARG < x < MIN_DBL_NORM_ARG
371 // Certain Safe - MIN_DBL_NORM_ARG <= x <= MAX_DBL_NORM_ARG
372 // Possible Overflow p14 - MAX_DBL_NORM_ARG < x < MIN_DBL_OFLOW_ARG
373 // Certain Overflow p15 - MIN_DBL_OFLOW_ARG <= x < +inf
375 // If the input is really a double arg, then there will never be
376 // "Possible Overflow" arguments.
380 add rAD_TB2 = 0x100, rAD_TB1
381 fcmp.ge.s1 p15,p0 = fNormX,fMIN_DBL_OFLOW_ARG
387 add rAD_P = 0x80, rAD_TB2
388 fcmp.le.s1 p11,p0 = fNormX,fMAX_DBL_ZERO_ARG
394 ldfpd fP5, fP4 = [rAD_P] ,16
395 fcmp.gt.s1 p14,p0 = fNormX,fMAX_DBL_NORM_ARG
396 (p15) br.cond.spnt EXP_CERTAIN_OVERFLOW
400 // Nfloat = round_int(W)
401 // The signficand of fW_2TO56_RSH contains the rounded integer part of W,
402 // as a twos complement number in the lower bits (that is, it may be negative).
403 // That twos complement number (called N) is put into rN.
405 // Since fW_2TO56_RSH is scaled by 2^56, it must be multiplied by 2^-56
406 // before the shift constant 1.10000 * 2^63 is subtracted to yield fNfloat.
407 // Thus, fNfloat contains the floating point version of N
410 ldfpd fP3, fP2 = [rAD_P]
411 fms.s1 fNfloat = fW_2TO56_RSH, f2TOM56, fRSHF
412 (p11) br.cond.spnt EXP_CERTAIN_UNDERFLOW
417 getf.sig rN = fW_2TO56_RSH
423 // rIndex_1 has index_1
424 // rIndex_2_16 has index_2 * 16
426 // rIndex_1_16 has index_1 * 16
429 // r = x - Nfloat * ln2_by_128_hi
430 // f = 1 - Nfloat * ln2_by_128_lo
432 and rIndex_1 = 0x0f, rN
433 fnma.s1 fR = fNfloat, fLn2_by_128_hi, fNormX
437 and rIndex_2_16 = 0x70, rN
438 fnma.s1 fF = fNfloat, fLn2_by_128_lo, f1
443 // rAD_T1 has address of T1
444 // rAD_T2 has address if T2
447 add rBiased_M = rExp_bias, rM
448 add rAD_T2 = rAD_TB2, rIndex_2_16
449 shladd rAD_T1 = rIndex_1, 4, rAD_TB1
453 // Create Scale = 2^M
455 setf.exp f2M = rBiased_M
464 fmpy.s0 fTmp = fLn2_by_128_lo, fLn2_by_128_lo // Force inexact
471 fma.s1 fRsq = fR, fR, f0
476 fma.s1 fP54 = fR, fP5, fP4
483 fcmp.lt.s1 p13,p0 = fNormX,fMIN_DBL_NORM_ARG
488 fma.s1 fP32 = fR, fP3, fP2
495 fma.s1 fP5432 = fRsq, fP54, fP32
502 fma.s1 fS1 = f2M,fT1,f0
507 fma.s1 fS2 = fF,fT2,f0
514 fma.s1 fP = fRsq, fP5432, fR
519 fma.s1 fS = fS1,fS2,f0
526 (p13) br.cond.spnt EXP_POSSIBLE_UNDERFLOW
527 (p14) br.cond.spnt EXP_POSSIBLE_OVERFLOW
533 fma.d.s0 f8 = fS, fP, fS
534 br.ret.sptk b0 // Normal path exit
539 EXP_POSSIBLE_OVERFLOW:
541 // Here if fMAX_DBL_NORM_ARG < x < fMIN_DBL_OFLOW_ARG
542 // This cannot happen if input is a double, only if input higher precision.
543 // Overflow is a possibility, not a certainty.
545 // Recompute result using status field 2 with user's rounding mode,
546 // and wre set. If result is larger than largest double, then we have
550 mov rGt_ln = 0x103ff // Exponent for largest dbl + 1 ulp
551 fsetc.s2 0x7F,0x42 // Get user's round mode, set wre
557 setf.exp fGt_pln = rGt_ln // Create largest double + 1 ulp
558 fma.d.s2 fWre_urm_f8 = fS, fP, fS // Result with wre set
565 fsetc.s2 0x7F,0x40 // Turn off wre in sf2
572 fcmp.ge.s1 p6, p0 = fWre_urm_f8, fGt_pln // Test for overflow
580 (p6) br.cond.spnt EXP_CERTAIN_OVERFLOW // Branch if overflow
586 fma.d.s0 f8 = fS, fP, fS
587 br.ret.sptk b0 // Exit if really no overflow
591 EXP_CERTAIN_OVERFLOW:
593 sub rTmp = rExp_mask, r0, 1
601 alloc r32=ar.pfs,1,4,4,0
602 fmerge.s FR_X = f8,f8
606 mov GR_Parameter_TAG = 14
607 fma.d.s0 FR_RESULT = fTmp, fTmp, fTmp // Set I,O and +INF result
608 br.cond.sptk __libm_error_region
612 EXP_POSSIBLE_UNDERFLOW:
614 // Here if fMAX_DBL_ZERO_ARG < x < fMIN_DBL_NORM_ARG
615 // Underflow is a possibility, not a certainty
617 // We define an underflow when the answer with
619 // is zero (tiny numbers become zero)
621 // Notice (from below) that if we have an unlimited exponent range,
622 // then there is an extra machine number E between the largest denormal and
623 // the smallest normal.
625 // So if with unbounded exponent we round to E or below, then we are
626 // tiny and underflow has occurred.
628 // But notice that you can be in a situation where we are tiny, namely
629 // rounded to E, but when the exponent is bounded we round to smallest
630 // normal. So the answer can be the smallest normal with underflow.
633 // -----+--------------------+--------------------+-----
635 // 1.1...10 2^-3fff 1.1...11 2^-3fff 1.0...00 2^-3ffe
636 // 0.1...11 2^-3ffe (biased, 1)
637 // largest dn smallest normal
641 fsetc.s2 0x7F,0x41 // Get user's round mode, set ftz
648 fma.d.s2 fFtz_urm_f8 = fS, fP, fS // Result with ftz set
655 fsetc.s2 0x7F,0x40 // Turn off ftz in sf2
662 fcmp.eq.s1 p6, p7 = fFtz_urm_f8, f0 // Test for underflow
667 fma.d.s0 f8 = fS, fP, fS // Compute result, set I, maybe U
674 (p6) br.cond.spnt EXP_UNDERFLOW_COMMON // Branch if really underflow
675 (p7) br.ret.sptk b0 // Exit if really no underflow
679 EXP_CERTAIN_UNDERFLOW:
680 // Here if x < fMAX_DBL_ZERO_ARG
681 // Result will be zero (or smallest denorm if round to +inf) with I, U set
685 setf.exp fTmp = rTmp // Form small normal
692 fmerge.se fTmp = fTmp, fLn2_by_128_lo // Small with signif lsb 1
699 fma.d.s0 f8 = fTmp, fTmp, f0 // Set I,U, tiny (+0.0) result
700 br.cond.sptk EXP_UNDERFLOW_COMMON
704 EXP_UNDERFLOW_COMMON:
705 // Determine if underflow result is zero or nonzero
707 alloc r32=ar.pfs,1,4,4,0
708 fcmp.eq.s1 p6, p0 = f8, f0
715 fmerge.s FR_X = fNormX,fNormX
716 (p6) br.cond.spnt EXP_UNDERFLOW_ZERO
720 EXP_UNDERFLOW_NONZERO:
721 // Here if x < fMIN_DBL_NORM_ARG and result nonzero;
724 mov GR_Parameter_TAG = 15
725 nop.f 0 // FR_RESULT already set
726 br.cond.sptk __libm_error_region
731 // Here if x < fMIN_DBL_NORM_ARG and result zero;
734 mov GR_Parameter_TAG = 15
735 nop.f 0 // FR_RESULT already set
736 br.cond.sptk __libm_error_region
740 GLOBAL_IEEE754_END(exp)
743 LOCAL_LIBM_ENTRY(__libm_error_region)
746 add GR_Parameter_Y=-32,sp // Parameter 2 value
748 .save ar.pfs,GR_SAVE_PFS
749 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
753 add sp=-64,sp // Create new stack
755 mov GR_SAVE_GP=gp // Save gp
758 stfd [GR_Parameter_Y] = FR_Y,16 // STORE Parameter 2 on stack
759 add GR_Parameter_X = 16,sp // Parameter 1 address
761 mov GR_SAVE_B0=b0 // Save b0
765 stfd [GR_Parameter_X] = FR_X // STORE Parameter 1 on stack
766 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
770 stfd [GR_Parameter_Y] = FR_RESULT // STORE Parameter 3 on stack
771 add GR_Parameter_Y = -16,GR_Parameter_Y
772 br.call.sptk b0=__libm_error_support# // Call error handling function
775 add GR_Parameter_RESULT = 48,sp
780 ldfd f8 = [GR_Parameter_RESULT] // Get return result off stack
782 add sp = 64,sp // Restore stack pointer
783 mov b0 = GR_SAVE_B0 // Restore return address
786 mov gp = GR_SAVE_GP // Restore gp
787 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
788 br.ret.sptk b0 // Return
791 LOCAL_LIBM_END(__libm_error_region)
792 .type __libm_error_support#,@function
793 .global __libm_error_support#