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 // 02/02/00 Initial version
43 // 03/02/00 New Algorithm
44 // 04/04/00 Unwind support added
45 // 08/15/00 Bundle added after call to __libm_error_support to properly
46 // set [the previously overwritten] GR_Parameter_RESULT.
47 // 11/28/00 Set FR_Y to f9
48 // 03/11/02 Fixed flags for fmodl(qnan,zero)
49 // 05/20/02 Cleaned up namespace and sf0 syntax
50 // 02/10/03 Reordered header: .section, .global, .proc, .align
51 // 04/28/03 Fix: fmod(sNaN,0) no longer sets errno
54 //====================================================================
55 // long double fmodl(long double,long double);
57 // Overview of operation
58 //====================================================================
60 // where i is an integer such that, if b!=0,
61 // |i|<|a/b| and |a/b-i|<1
64 //====================================================================
65 // a). if |a|<|b|, return a
66 // b). get quotient and reciprocal overestimates accurate to
68 // c). if the exponent difference (exponent(a)-exponent(b))
69 // is less than 32, truncate quotient to integer and
70 // finish in one iteration
71 // d). if exponent(a)-exponent(b)>=32 (q2>=2^32)
72 // round quotient estimate to single precision (k=RN(q2)),
73 // calculate partial remainder (a'=a-k*b),
74 // get quotient estimate (a'*y2), and repeat from c).
77 //====================================================================
78 // Predicate registers: p6-p11
79 // General registers: r2,r29,r32 (ar.pfs), r33-r39
80 // Floating point registers: f6-f15
89 GR_Parameter_RESULT = r39
90 GR_Parameter_TAG = r40
98 GLOBAL_IEEE754_ENTRY(fmodl)
104 alloc r32=ar.pfs,1,4,4,0
119 frcpa.s1 f10,p6=f6,f7
123 // eliminate special cases
128 cmp.eq p7,p10=r29,r0;;
131 // Y +-NAN, +-inf, +-0? p7
134 (p10) fclass.m p7,p10 = f9, 0xe7
138 // qnan snan inf norm unorm 0 -+
141 // X +-NAN, +-inf, ? p9
145 fclass.m.unc p9,p11 = f8, 0xe3
149 // |x| < |y|? Return x p8
152 (p10) fcmp.lt.unc.s1 p8,p0 = f6,f7
159 (p6) fma.s1 f13=f6,f10,f0
163 // (3) e0 = 1 - b * y0
164 (p6) fnma.s1 f12=f7,f10,f1
168 // Y +-NAN, +-inf, +-0? p7
172 (p10) fclass.nm p7,p0 = f9, 0xff
176 // qnan snan inf norm unorm 0 -+
179 // X +-NAN, +-inf, ? p9
183 (p11) fclass.nm p9,p0 = f8, 0xff
189 // y denormal ? set D flag (if |x|<|y|)
190 (p8) fnma.s0 f10=f9,f1,f9
197 // normalize x (if |x|<|y|)
198 (p8) fma.s0 f8=f8,f1,f0
202 (p9) br.cond.spnt FMOD_X_NAN_INF
203 (p7) br.cond.spnt FMOD_Y_NAN_INF_ZERO
204 // if |x|<|y|, return
205 (p8) br.ret.spnt b0;;
210 // x denormal ? set D flag
216 // y denormal ? set D flag
225 (p6) fma.s1 f13=f12,f13,f13
230 // (5) e1 = e0 * e0 + 2^-34
231 (p6) fma.s1 f14=f12,f12,f11
240 // (6) y1 = y0 + e0 * y0
241 (p6) fma.s1 f10=f12,f10,f10
245 // set f12=1.25*2^{-24}
248 (p6) fma.s1 f13=f13,f14,f13
258 // (8) y2 = y1 + e1 * y1
259 (p6) fma.s1 f10=f14,f10,f10
261 cmp.ne.and p6,p10=r0,r0;;
270 fcmp.lt.unc.s1 p8,p7=f13,f15
275 // will truncate quotient to integer, if exponent<32 (in advance)
276 fcvt.fx.trunc.s1 f11=f13
281 // if exponent>32, round quotient to single precision (perform in advance)
282 fma.s.s1 f13=f13,f1,f0
290 (p8) fmerge.s f12=f8,f1
295 // normalize truncated quotient
301 // calculate remainder (assuming f13=RZ(Q))
302 (p7) fnma.s1 f14=f13,f7,f6
307 // also if exponent>32, round quotient to single precision
308 // and subtract 1 ulp: q=q-q*(1.25*2^{-24})
309 (p7) fnma.s.s1 f11=f13,f12,f13
315 // (p8) calculate remainder (82-bit format)
316 (p8) fnma.s1 f11=f13,f7,f6
321 // (p7) calculate remainder (assuming f11=RZ(Q))
322 (p7) fnma.s1 f6=f11,f7,f6
329 // Final iteration (p8): is f6 the correct remainder (quotient was not overestimated) ?
330 (p8) fcmp.lt.unc.s1 p6,p10=f11,f0
335 // get new quotient estimation: a'*y2
336 (p7) fma.s1 f13=f14,f10,f0
341 // was f13=RZ(Q) ? (then new remainder f14>=0)
342 (p7) fcmp.lt.unc.s1 p7,p9=f14,f0
347 .pred.rel "mutex",p6,p10
350 // add b to estimated remainder (to cover the case when the quotient was overestimated)
351 // also set correct sign by using f9=|b|*sgn(a), f12=sgn(a)
352 (p6) fma.s0 f8=f11,f12,f9
357 // set correct sign of result before returning: f12=sgn(a)
358 (p10) fma.s0 f8=f11,f12,f0
359 (p8) br.ret.sptk b0;;
363 // if f13!=RZ(Q), get alternative quotient estimation: a''*y2
364 (p7) fma.s1 f13=f6,f10,f0
369 // if f14 was RZ(Q), set remainder to f14
371 br.cond.sptk loop64;;
381 fclass.m p10,p0=f8,0xc3 // Test x=nan
397 (p10) fclass.m p10,p0=f9,0x07 // Test x=nan, and y=zero
402 fcmp.eq.unc.s1 p11,p0=f10,f0
403 (p10) br.ret.spnt b0;; // Exit with result=x if x=nan and y=zero
409 (p11) br.cond.spnt FMOD_Y_ZERO;;
412 // X infinity? Return QNAN indefinite
416 fclass.m.unc p8,p9 = f8, 0x23
422 (p8) fclass.m p9,p8=f9,0xc3
425 // Y not pseudo-zero ? (r29 holds significand)
428 (p8) cmp.ne p7,p0=r29,r0
433 (p8) frcpa.s0 f8,p0 = f8,f8
438 // also set Denormal flag if necessary
439 (p7) fnma.s0 f9=f9,f1,f9
445 (p8) fma.s0 f8=f8,f1,f0
451 (p9) frcpa.s0 f8,p7=f8,f9
460 fclass.m.unc p7,p0 = f9, 0x23
466 (p7) fma.s0 f8=f8,f1,f0
467 (p7) br.ret.spnt b0 ;;
473 fclass.m.unc p9,p10 = f9, 0xc3
478 (p10) fclass.nm p9,p0 = f9, 0xff
484 (p9) fma.s0 f8=f9,f1,f0
485 (p9) br.ret.spnt b0 ;;
489 // Y zero? Must be zero at this point
490 // because it is the only choice left.
491 // Return QNAN indefinite
496 frcpa.s0 f12,p0=f0,f0
502 fclass.m.unc p9,p10 = f8, 0xc3
507 (p10) fclass.nm p9,p10 = f8, 0xff
513 (p9) frcpa.s0 f11,p7=f8,f0
520 (p10) frcpa.s0 f11,p7 = f9,f9
521 mov GR_Parameter_TAG = 120 ;;
526 fmerge.s f10 = f8, f8
533 br.sptk __libm_error_region;;
536 GLOBAL_IEEE754_END(fmodl)
539 LOCAL_LIBM_ENTRY(__libm_error_region)
542 add GR_Parameter_Y=-32,sp // Parameter 2 value
544 .save ar.pfs,GR_SAVE_PFS
545 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
549 add sp=-64,sp // Create new stack
551 mov GR_SAVE_GP=gp // Save gp
554 stfe [GR_Parameter_Y] = FR_Y,16 // Save Parameter 2 on stack
555 add GR_Parameter_X = 16,sp // Parameter 1 address
557 mov GR_SAVE_B0=b0 // Save b0
561 stfe [GR_Parameter_X] = FR_X // Store Parameter 1 on stack
562 add GR_Parameter_RESULT = 0,GR_Parameter_Y
563 nop.b 0 // Parameter 3 address
566 stfe [GR_Parameter_Y] = FR_RESULT // Store Parameter 3 on stack
567 add GR_Parameter_Y = -16,GR_Parameter_Y
568 br.call.sptk b0=__libm_error_support# // Call error handling function
573 add GR_Parameter_RESULT = 48,sp
576 ldfe f8 = [GR_Parameter_RESULT] // Get return result off stack
578 add sp = 64,sp // Restore stack pointer
579 mov b0 = GR_SAVE_B0 // Restore return address
582 mov gp = GR_SAVE_GP // Restore gp
583 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
584 br.ret.sptk b0 // Return
587 LOCAL_LIBM_END(__libm_error_region)
592 .type __libm_error_support#,@function
593 .global __libm_error_support#