2 // Copyright (C) 2000, 2001, Intel Corporation
3 // All rights reserved.
5 // Contributed 2/2/2000 by John Harrison, Cristina Iordache, Ted Kubaska,
6 // Bob Norin, Shane Story, and Ping Tak Peter Tang of the Computational
10 // Redistribution and use in source and binary forms, with or without
11 // modification, are permitted provided that the following conditions are
14 // * Redistributions of source code must retain the above copyright
15 // notice, this list of conditions and the following disclaimer.
17 // * Redistributions in binary form must reproduce the above copyright
18 // notice, this list of conditions and the following disclaimer in the
19 // documentation and/or other materials provided with the distribution.
21 // * The name of Intel Corporation may not be used to endorse or promote
22 // products derived from this software without specific prior written
25 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
28 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
29 // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
30 // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
31 // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
32 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
33 // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
34 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
35 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37 // Intel Corporation is the author of this code, and requests that all
38 // problem reports or change requests be submitted to it directly at
39 // http://developer.intel.com/opensource.
42 //====================================================================
43 // 2/02/00 Initial version
44 // 3/02/00 New algorithm
45 // 4/04/00 Unwind support added
46 // 7/21/00 Fixed quotient=2^{24*m+23} bug
47 // 8/15/00 Bundle added after call to __libm_error_support to properly
48 // set [the previously overwritten] GR_Parameter_RESULT.
49 //11/29/00 Set FR_Y to f9
52 //====================================================================
53 // float remainderf(float,float);
55 // Overview of operation
56 //====================================================================
57 // remainder(a,b)=a-i*b,
58 // where i is an integer such that, if b!=0 and a is finite,
59 // |a/b-i|<=1/2. If |a/b-i|=1/2, i is even.
62 //====================================================================
63 // a). eliminate special cases
64 // b). if |a/b|<0.25 (first quotient estimate), return a
65 // c). use single precision divide algorithm to get quotient q
66 // rounded to 24 bits of precision
67 // d). calculate partial remainders (using both q and q-ulp);
68 // select one and RZ(a/b) based on the sign of |a|-|b|*q
69 // e). if the exponent difference (exponent(a)-exponent(b))
70 // is less than 24 (quotient estimate<2^{24}-2), use RZ(a/b)
71 // and sticky bits to round to integer; exit loop and
72 // calculate final remainder
73 // f). if exponent(a)-exponent(b)>=24, select new value of a as
74 // the partial remainder calculated using RZ(a/b);
78 //====================================================================
79 // a=+/- Inf, or b=+/-0: return NaN, call libm_error_support
80 // a=NaN or b=NaN: return NaN
82 #include "libm_support.h"
86 //====================================================================
87 // Predicate registers: p6-p12
88 // General registers: r2,r3,r28,r29,r32 (ar.pfs), r33-r39
89 // Floating point registers: f6-f15
101 GR_Parameter_RESULT = r39
102 GR_Parameter_TAG = r40
117 .type __remainderf,@function
124 alloc r32=ar.pfs,1,4,4,0
141 // Y +-NAN, +-inf, +-0? p11
144 (p0) fclass.m.unc p11,p0 = f9, 0xe7
147 // qnan snan inf norm unorm 0 -+
150 // X +-NAN, +-inf, ? p9
153 (p0) fclass.m.unc p9,p0 = f8, 0xe3
164 cmp.eq.unc p7,p0=r0,r0
167 frcpa.s1 f10,p6=f13,f14
171 (p9) br.cond.spnt L(FREM_X_NAN_INF)
172 (p11) br.cond.spnt L(FREM_Y_NAN_INF_ZERO)
176 // set D flag if a (f8) is denormal
187 // q0 = a * y0 in f15
188 (p6) fma.s1 f15=f13,f10,f0
194 // e0 = 1 - b * y0 in f7
195 (p6) fnma.s1 f7=f14,f10,f1
206 // q1 = q0 + e0 * q0 in f6
207 (p6) fma.s1 f6=f7,f15,f15
213 // e1 = e0 * e0 in f7
214 (p6) fma.s1 f7=f7,f7,f0
218 (p7) getf.exp r29=f15
227 // q2 = q1 + e1 * q1 in f6
228 (p6) fma.s1 f6=f7,f6,f6
234 // e2 = e1 * e1 in f7
235 (p6) fma.s1 f7=f7,f7,f0
240 // q<1/4 ? (i.e. expon< -2)
241 (p7) cmp.gt.unc p7,p0=r28,r29
250 // if |a/b|<1/4, set D flag before returning
251 (p7) fma.s.s0 f9=f9,f0,f8
256 // can be combined with bundle above if sign of 0 or
257 // FTZ enabled are not important
258 (p7) fmerge.s f8=f8,f9
259 // return if |a|<4*|b| (estimated quotient < 1/4)
260 (p7) br.ret.spnt b0;;
264 // set f8 to current a value | sign
266 // r2=2^{-24}+2^{-48} (double prec.)
275 // q3 = q2 + e2 * q2 in f6
276 (p6) fma.d.s1 f6=f7,f6,f6
283 (p6) fma.s.s1 f11=f6,f1,f0
289 // last step ? (q3<2^{24}-2 --> q<2^{24})
290 fcmp.lt.unc.s1 p0,p12=f6,f12
301 fnma.s1 f6=f14,f11,f13
306 // q'=q-q*(1.25*2^{-24}) (q'=q-ulp)
307 fnma.s.s1 f15=f11,f15,f11
314 fnma.s1 f13=f14,f15,f13
319 // r>0 iff q=RZ(a/b) and inexact
320 fcmp.gt.unc.s1 p8,p0=f6,f0
325 // r<0 iff q'=RZ(a/b) and inexact
326 fcmp.lt.unc.s1 p9,p10=f6,f0
329 .pred.rel "mutex",p8,p9
332 // (p8) Q=q+(last iteration ? sticky bits:0)
333 // i.e. Q=q+q*x (x=2^{-32} or 0)
334 (p8) fma.s1 f11=f11,f12,f11
339 // (p9) Q=q'+(last iteration ? sticky bits:0)
340 // i.e. Q=q'+q'*x (x=2^{-24} or 0: if expon. difference=23, want to round back to q)
341 (p9) fma.s1 f11=f15,f7,f15
347 // (p9) set r=r2 (new a, if not last iteration)
350 (p12) br.cond.sptk L(remloop24);;
380 // This can be removed if sign of 0 is not important
381 // get remainder using sf1
382 fnma.s.s1 f12=f9,f11,f8
388 fnma.s.s0 f8=f9,f11,f8
397 // This can be removed if sign of 0 is not important
398 fcmp.eq.unc.s1 p8,p0=f12,f0
403 // if f8=0, set sign correctly
404 // This can be removed if sign of 0 is not important
405 (p8) fmerge.s f8=f7,f8
421 fcmp.eq.unc.s1 p11,p0=f10,f0
428 (p11) br.cond.spnt L(FREM_Y_ZERO);;
431 // X infinity? Return QNAN indefinite
434 (p0) fclass.m.unc p8,p0 = f8, 0x23
437 // X infinity? Return QNAN indefinite
440 (p0) fclass.m.unc p11,p0 = f8, 0x23
446 (p8) fclass.m.unc p0,p8=f9,0xc3
451 // also set Denormal flag if necessary
452 (p8) fma.s0 f9=f9,f1,f0
457 (p8) frcpa.s0 f8,p7 = f8,f8
468 (p8) fma.s f8=f8,f1,f0
475 (p11) br.cond.spnt L(EXP_ERROR_RETURN);;
484 L(FREM_Y_NAN_INF_ZERO):
489 (p0) fclass.m.unc p7,p0 = f9, 0x23
495 (p7) fma.s f8=f8,f1,f0
496 (p7) br.ret.spnt b0 ;;
502 (p0) fclass.m.unc p9,p0 = f9, 0xc3
508 (p9) fma.s f8=f9,f1,f0
509 (p9) br.ret.spnt b0 ;;
513 // Y zero? Must be zero at this point
514 // because it is the only choice left.
515 // Return QNAN indefinite
520 (p0) fclass.m.unc p9,p10 = f8, 0xc3
525 (p10) fclass.nm p9,p10 = f8, 0xff
531 (p9) frcpa f11,p7=f8,f0
537 (p10) frcpa f11,p7 = f0,f0
543 (p0) fmerge.s f10 = f8, f8
549 (p0) fma.s f8=f11,f1,f0
557 (p0) mov GR_Parameter_TAG = 125
559 (p0) br.sptk __libm_error_region;;
563 ASM_SIZE_DIRECTIVE(remainderf)
565 ASM_SIZE_DIRECTIVE(__remainderf)
570 .proc __libm_error_region
574 add GR_Parameter_Y=-32,sp // Parameter 2 value
576 .save ar.pfs,GR_SAVE_PFS
577 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
581 add sp=-64,sp // Create new stack
583 mov GR_SAVE_GP=gp // Save gp
586 stfs [GR_Parameter_Y] = FR_Y,16 // Save Parameter 2 on stack
587 add GR_Parameter_X = 16,sp // Parameter 1 address
589 mov GR_SAVE_B0=b0 // Save b0
593 stfs [GR_Parameter_X] = FR_X // Store Parameter 1 on stack
594 add GR_Parameter_RESULT = 0,GR_Parameter_Y
595 nop.b 0 // Parameter 3 address
598 stfs [GR_Parameter_Y] = FR_RESULT // Store Parameter 3 on stack
599 add GR_Parameter_Y = -16,GR_Parameter_Y
600 br.call.sptk b0=__libm_error_support#;; // Call error handling function
605 add GR_Parameter_RESULT = 48,sp
608 ldfs f8 = [GR_Parameter_RESULT] // Get return result off stack
610 add sp = 64,sp // Restore stack pointer
611 mov b0 = GR_SAVE_B0 // Restore return address
614 mov gp = GR_SAVE_GP // Restore gp
615 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
616 br.ret.sptk b0 // Return
619 .endp __libm_error_region
620 ASM_SIZE_DIRECTIVE(__libm_error_region)
623 .type __libm_error_support#,@function
624 .global __libm_error_support#