.file "fmodl.s" // Copyright (c) 2000 - 2004, Intel Corporation // All rights reserved. // // Contributed 2000 by the Intel Numerics Group, Intel Corporation // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are // met: // // * Redistributions of source code must retain the above copyright // notice, this list of conditions and the following disclaimer. // // * Redistributions in binary form must reproduce the above copyright // notice, this list of conditions and the following disclaimer in the // documentation and/or other materials provided with the distribution. // // * The name of Intel Corporation may not be used to endorse or promote // products derived from this software without specific prior written // permission. // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. // // Intel Corporation is the author of this code, and requests that all // problem reports or change requests be submitted to it directly at // http://www.intel.com/software/products/opensource/libraries/num.htm. // // History //==================================================================== // 02/02/00 Initial version // 03/02/00 New Algorithm // 04/04/00 Unwind support added // 08/15/00 Bundle added after call to __libm_error_support to properly // set [ the previously overwritten ] GR_Parameter_RESULT. // 11/28/00 Set FR_Y to f9 // 03/11/02 Fixed flags for fmodl(qnan, zero) // 05/20/02 Cleaned up namespace and sf0 syntax // 02/10/03 Reordered header:.section,.global,.proc,.align // 04/28/03 Fix: fmod(sNaN, 0) no longer sets errno // 11/23/04 Reformatted routine and improved speed // // API //==================================================================== // long double fmodl(long double, long double); // // Overview of operation //==================================================================== // fmod(a, b)= a-i*b, // where i is an integer such that, if b!= 0, // |i|<|a/b| and |a/b-i|<1 // // Algorithm //==================================================================== // a). if |a|<|b|, return a // b). get quotient and reciprocal overestimates accurate to // 33 bits (q2, y2) // c). if the exponent difference (exponent(a)-exponent(b)) // is less than 32, truncate quotient to integer and // finish in one iteration // d). if exponent(a)-exponent(b)>= 32 (q2>= 2^32) // round quotient estimate to single precision (k= RN(q2)), // calculate partial remainder (a'= a-k*b), // get quotient estimate (a'*y2), and repeat from c). // // Registers used //==================================================================== GR_SMALLBIASEXP = r2 GR_2P32 = r3 GR_SMALLBIASEXP = r20 GR_ROUNDCONST = r21 GR_SIG_B = r22 GR_ARPFS = r23 GR_TMP1 = r24 GR_TMP2 = r25 GR_TMP3 = r26 GR_SAVE_B0 = r33 GR_SAVE_PFS = r34 GR_SAVE_GP = r35 GR_SAVE_SP = r36 GR_Parameter_X = r37 GR_Parameter_Y = r38 GR_Parameter_RESULT = r39 GR_Parameter_TAG = r40 FR_X = f10 FR_Y = f9 FR_RESULT = f8 FR_ABS_A = f6 FR_ABS_B = f7 FR_Y_INV = f10 FR_SMALLBIAS = f11 FR_E0 = f12 FR_Q = f13 FR_E1 = f14 FR_2P32 = f15 FR_TMPX = f32 FR_TMPY = f33 FR_ROUNDCONST = f34 FR_QINT = f35 FR_QRND24 = f36 FR_NORM_B = f37 FR_TMP = f38 FR_TMP2 = f39 FR_DFLAG = f40 FR_Y_INV0 = f41 FR_Y_INV1 = f42 FR_Q0 = f43 FR_Q1 = f44 FR_QINT_Z = f45 FR_QREM = f46 FR_B_SGN_A = f47 .section .text GLOBAL_IEEE754_ENTRY(fmodl) // inputs in f8, f9 // result in f8 { .mfi getf.sig GR_SIG_B = f9 // FR_ABS_A = |a| fmerge.s FR_ABS_A = f0, f8 mov GR_SMALLBIASEXP = 0x0ffdd } { .mfi nop.m 0 // FR_ABS_B = |b| fmerge.s FR_ABS_B = f0, f9 nop.i 0 } ;; { .mfi setf.exp FR_SMALLBIAS = GR_SMALLBIASEXP // (1) y0 frcpa.s1 FR_Y_INV0, p6 = FR_ABS_A, FR_ABS_B nop.i 0 } ;; { .mlx nop.m 0 movl GR_ROUNDCONST = 0x33a00000 } ;; // eliminate special cases { .mmi nop.m 0 nop.m 0 // y pseudo-zero ? cmp.eq p7, p10 = GR_SIG_B, r0 } ;; // set p7 if b +/-NAN, +/-inf, +/-0 { .mfi nop.m 0 (p10) fclass.m p7, p10 = f9, 0xe7 nop.i 0 } ;; { .mfi mov GR_2P32 = 0x1001f // (2) q0 = a*y0 (p6) fma.s1 FR_Q0 = FR_ABS_A, FR_Y_INV0, f0 nop.i 0 } { .mfi nop.m 0 // (3) e0 = 1 - b * y0 (p6) fnma.s1 FR_E0 = FR_ABS_B, FR_Y_INV0, f1 nop.i 0 } ;; // set p9 if a +/-NAN, +/-inf { .mfi nop.m 0 fclass.m.unc p9, p11 = f8, 0xe3 nop.i 0 } // |a| < |b|? Return a, p8=1 { .mfi nop.m 0 (p10) fcmp.lt.unc.s1 p8, p0 = FR_ABS_A, FR_ABS_B nop.i 0 } ;; // set p7 if b +/-NAN, +/-inf, +/-0 { .mfi nop.m 0 // pseudo-NaN ? (p10) fclass.nm p7, p0 = f9, 0xff nop.i 0 } ;; // set p9 if a is +/-NaN, +/-Inf { .mfi nop.m 0 (p11) fclass.nm p9, p0 = f8, 0xff nop.i 0 } { .mfi nop.m 0 // b denormal ? set D flag (if |a|<|b|) (p8) fnma.s0 FR_DFLAG = f9, f1, f9 nop.i 0 } ;; { .mfi // FR_2P32 = 2^32 setf.exp FR_2P32 = GR_2P32 // (4) q1 = q0+e0*q0 (p6) fma.s1 FR_Q1 = FR_E0, FR_Q0, FR_Q0 nop.i 0 } { .mfi nop.m 0 // (5) e1 = e0 * e0 + 2^-34 (p6) fma.s1 FR_E1 = FR_E0, FR_E0, FR_SMALLBIAS nop.i 0 } ;; { .mfi nop.m 0 // normalize a (if |a|<|b|) (p8) fma.s0 f8 = f8, f1, f0 nop.i 0 } { .bbb (p9) br.cond.spnt FMOD_A_NAN_INF (p7) br.cond.spnt FMOD_B_NAN_INF_ZERO // if |a|<|b|, return (p8) br.ret.spnt b0 } ;; { .mfi nop.m 0 // (6) y1 = y0 + e0 * y0 (p6) fma.s1 FR_Y_INV1 = FR_E0, FR_Y_INV0, FR_Y_INV0 nop.i 0 } ;; { .mfi nop.m 0 // a denormal ? set D flag // b denormal ? set D flag fcmp.eq.s0 p12,p0 = FR_ABS_A, FR_ABS_B nop.i 0 } { .mfi // set FR_ROUNDCONST = 1.25*2^{-24} setf.s FR_ROUNDCONST = GR_ROUNDCONST // (7) q2 = q1+e1*q1 (p6) fma.s1 FR_Q = FR_Q1, FR_E1, FR_Q1 nop.i 0 } ;; { .mfi nop.m 0 fmerge.s FR_B_SGN_A = f8, f9 nop.i 0 } { .mfi nop.m 0 // (8) y2 = y1 + e1 * y1 (p6) fma.s1 FR_Y_INV = FR_E1, FR_Y_INV1, FR_Y_INV1 // set p6 = 0, p10 = 0 cmp.ne.and p6, p10 = r0, r0 } ;; // will compute integer quotient bits (24 bits per iteration) .align 32 loop64: { .mfi nop.m 0 // compare q2, 2^32 fcmp.lt.unc.s1 p8, p7 = FR_Q, FR_2P32 nop.i 0 } { .mfi nop.m 0 // will truncate quotient to integer, if exponent<32 (in advance) fcvt.fx.trunc.s1 FR_QINT = FR_Q nop.i 0 } ;; { .mfi nop.m 0 // if exponent>32 round quotient to single precision (perform in advance) fma.s.s1 FR_QRND24 = FR_Q, f1, f0 nop.i 0 } ;; { .mfi nop.m 0 // set FR_ROUNDCONST = sgn(a) (p8) fmerge.s FR_ROUNDCONST = f8, f1 nop.i 0 } { .mfi nop.m 0 // normalize truncated quotient (p8) fcvt.xf FR_QRND24 = FR_QINT nop.i 0 } ;; { .mfi nop.m 0 // calculate remainder (assuming FR_QRND24 = RZ(Q)) (p7) fnma.s1 FR_E1 = FR_QRND24, FR_ABS_B, FR_ABS_A nop.i 0 } { .mfi nop.m 0 // also if exponent>32, round quotient to single precision // and subtract 1 ulp: q = q-q*(1.25*2^{-24}) (p7) fnma.s.s1 FR_QINT_Z = FR_QRND24, FR_ROUNDCONST, FR_QRND24 nop.i 0 } ;; { .mfi nop.m 0 // (p8) calculate remainder (82-bit format) (p8) fnma.s1 FR_QREM = FR_QRND24, FR_ABS_B, FR_ABS_A nop.i 0 } { .mfi nop.m 0 // (p7) calculate remainder (assuming FR_QINT_Z = RZ(Q)) (p7) fnma.s1 FR_ABS_A = FR_QINT_Z, FR_ABS_B, FR_ABS_A nop.i 0 } ;; { .mfi nop.m 0 // Final iteration (p8): is FR_ABS_A the correct remainder // (quotient was not overestimated) ? (p8) fcmp.lt.unc.s1 p6, p10 = FR_QREM, f0 nop.i 0 } ;; { .mfi nop.m 0 // get new quotient estimation: a'*y2 (p7) fma.s1 FR_Q = FR_E1, FR_Y_INV, f0 nop.i 0 } { .mfb nop.m 0 // was FR_Q = RZ(Q) ? (then new remainder FR_E1> = 0) (p7) fcmp.lt.unc.s1 p7, p9 = FR_E1, f0 nop.b 0 } ;; .pred.rel "mutex", p6, p10 { .mfb nop.m 0 // add b to estimated remainder (to cover the case when the quotient was // overestimated) // also set correct sign by using // FR_B_SGN_A = |b|*sgn(a), FR_ROUNDCONST = sgn(a) (p6) fma.s0 f8 = FR_QREM, FR_ROUNDCONST, FR_B_SGN_A nop.b 0 } { .mfb nop.m 0 // set correct sign of result before returning: FR_ROUNDCONST = sgn(a) (p10) fma.s0 f8 = FR_QREM, FR_ROUNDCONST, f0 (p8) br.ret.sptk b0 } ;; { .mfi nop.m 0 // if f13! = RZ(Q), get alternative quotient estimation: a''*y2 (p7) fma.s1 FR_Q = FR_ABS_A, FR_Y_INV, f0 nop.i 0 } { .mfb nop.m 0 // if FR_E1 was RZ(Q), set remainder to FR_E1 (p9) fma.s1 FR_ABS_A = FR_E1, f1, f0 br.cond.sptk loop64 } ;; FMOD_A_NAN_INF: // b zero ? { .mfi nop.m 0 fclass.m p10, p0 = f8, 0xc3 // Test a = nan nop.i 0 } { .mfi nop.m 0 fma.s1 FR_NORM_B = f9, f1, f0 nop.i 0 } ;; { .mfi nop.m 0 fma.s0 f8 = f8, f1, f0 nop.i 0 } { .mfi nop.m 0 (p10) fclass.m p10, p0 = f9, 0x07 // Test x = nan, and y = zero nop.i 0 } ;; { .mfb nop.m 0 fcmp.eq.unc.s1 p11, p0 = FR_NORM_B, f0 (p10) br.ret.spnt b0 // Exit with result = a if a = nan and b = zero } ;; { .mib nop.m 0 nop.i 0 // if Y zero (p11) br.cond.spnt FMOD_B_ZERO } ;; // a= infinity? Return QNAN indefinite { .mfi // set p7 t0 0 cmp.ne p7, p0 = r0, r0 fclass.m.unc p8, p9 = f8, 0x23 nop.i 0 } ;; // b NaN ? { .mfi nop.m 0 (p8) fclass.m p9, p8 = f9, 0xc3 nop.i 0 } ;; // b not pseudo-zero ? (GR_SIG_B holds significand) { .mii nop.m 0 (p8) cmp.ne p7, p0 = GR_SIG_B, r0 nop.i 0 } ;; { .mfi nop.m 0 (p8) frcpa.s0 f8, p0 = f8, f8 nop.i 0 } { .mfi nop.m 0 // also set Denormal flag if necessary (p7) fnma.s0 f9 = f9, f1, f9 nop.i 0 } ;; { .mfb nop.m 0 (p8) fma.s0 f8 = f8, f1, f0 nop.b 0 } ;; { .mfb nop.m 0 (p9) frcpa.s0 f8, p7 = f8, f9 br.ret.sptk b0 } ;; FMOD_B_NAN_INF_ZERO: // b INF { .mfi nop.m 0 fclass.m.unc p7, p0 = f9, 0x23 nop.i 0 } ;; { .mfb nop.m 0 (p7) fma.s0 f8 = f8, f1, f0 (p7) br.ret.spnt b0 } ;; // b NAN? { .mfi nop.m 0 fclass.m.unc p9, p10 = f9, 0xc3 nop.i 0 } ;; { .mfi nop.m 0 (p10) fclass.nm p9, p0 = f9, 0xff nop.i 0 } ;; { .mfb nop.m 0 (p9) fma.s0 f8 = f9, f1, f0 (p9) br.ret.spnt b0 } ;; FMOD_B_ZERO: // Y zero? Must be zero at this point // because it is the only choice left. // Return QNAN indefinite { .mfi nop.m 0 // set Invalid frcpa.s0 FR_TMP, p0 = f0, f0 nop.i 0 } ;; // a NAN? { .mfi nop.m 0 fclass.m.unc p9, p10 = f8, 0xc3 nop.i 0 } ;; { .mfi alloc GR_ARPFS = ar.pfs, 1, 4, 4, 0 (p10) fclass.nm p9, p10 = f8, 0xff nop.i 0 } ;; { .mfi nop.m 0 (p9) frcpa.s0 FR_TMP2, p7 = f8, f0 nop.i 0 } ;; { .mfi nop.m 0 (p10) frcpa.s0 FR_TMP2, p7 = f9, f9 mov GR_Parameter_TAG = 120 } ;; { .mfi nop.m 0 fmerge.s FR_X = f8, f8 nop.i 0 } { .mfb nop.m 0 fma.s0 f8 = FR_TMP2, f1, f0 br.sptk __libm_error_region } ;; GLOBAL_IEEE754_END(fmodl) LOCAL_LIBM_ENTRY(__libm_error_region) .prologue { .mfi add GR_Parameter_Y = -32, sp // Parameter 2 value nop.f 0 .save ar.pfs, GR_SAVE_PFS mov GR_SAVE_PFS = ar.pfs // Save ar.pfs } { .mfi .fframe 64 add sp = -64, sp // Create new stack nop.f 0 mov GR_SAVE_GP = gp // Save gp } ;; { .mmi stfe [ GR_Parameter_Y ] = FR_Y, 16 // Save Parameter 2 on stack add GR_Parameter_X = 16, sp // Parameter 1 address .save b0, GR_SAVE_B0 mov GR_SAVE_B0 = b0 // Save b0 } ;; .body { .mib stfe [ GR_Parameter_X ] = FR_X // Store Parameter 1 on stack add GR_Parameter_RESULT = 0, GR_Parameter_Y nop.b 0 // Parameter 3 address } { .mib stfe [ GR_Parameter_Y ] = FR_RESULT // Store Parameter 3 on stack add GR_Parameter_Y = -16, GR_Parameter_Y br.call.sptk b0 = __libm_error_support# // Call error handling function } ;; { .mmi nop.m 0 nop.m 0 add GR_Parameter_RESULT = 48, sp } ;; { .mmi ldfe f8 = [ GR_Parameter_RESULT ] // Get return result off stack .restore sp add sp = 64, sp // Restore stack pointer mov b0 = GR_SAVE_B0 // Restore return address } ;; { .mib mov gp = GR_SAVE_GP // Restore gp mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs br.ret.sptk b0 // Return } ;; LOCAL_LIBM_END(__libm_error_region) .type __libm_error_support#, @function .global __libm_error_support#