IA-64 implementation of tanl.
authordrepper <drepper>
Mon, 19 Feb 2001 09:04:50 +0000 (09:04 +0000)
committerdrepper <drepper>
Mon, 19 Feb 2001 09:04:50 +0000 (09:04 +0000)
sysdeps/ia64/fpu/s_tanl.S [new file with mode: 0644]

diff --git a/sysdeps/ia64/fpu/s_tanl.S b/sysdeps/ia64/fpu/s_tanl.S
new file mode 100644 (file)
index 0000000..d7cc3ee
--- /dev/null
@@ -0,0 +1,3057 @@
+.file "tanl.s"
+
+// Copyright (c) 2000, 2001, Intel Corporation
+// All rights reserved.
+// 
+// Contributed 2/2/2000 by John Harrison, Ted Kubaska, Bob Norin, Shane Story,
+// and Ping Tak Peter Tang of the Computational Software Lab, Intel Corporation.
+// 
+// WARRANTY DISCLAIMER
+// 
+// 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://developer.intel.com/opensource.
+//
+// *********************************************************************
+//
+// History: 
+//
+// 2/02/2000 (hand-optimized)
+// 4/04/00  Unwind support added
+// 12/28/00 Fixed false invalid flags
+//
+// *********************************************************************
+//
+// Function:   tanl(x) = tangent(x), for double-extended precision x values
+//
+// *********************************************************************
+//
+// Resources Used:
+//
+//    Floating-Point Registers: f8 (Input and Return Value)
+//                              f9-f15
+//                              f32-f112
+//
+//    General Purpose Registers:
+//      r32-r48
+//      r49-r50 (Used to pass arguments to pi_by_2 reduce routine)
+//
+//    Predicate Registers:      p6-p15
+//
+// *********************************************************************
+//
+// IEEE Special Conditions:
+//
+//    Denormal  fault raised on denormal inputs
+//    Overflow exceptions do not occur
+//    Underflow exceptions raised when appropriate for tan 
+//    (No specialized error handling for this routine)
+//    Inexact raised when appropriate by algorithm
+//
+//    tan(SNaN) = QNaN
+//    tan(QNaN) = QNaN
+//    tan(inf) = QNaN
+//    tan(+/-0) = +/-0
+//
+// *********************************************************************
+//
+// Mathematical Description
+//
+// We consider the computation of FPTANL of Arg. Now, given
+//
+//      Arg = N pi/2  + alpha,          |alpha| <= pi/4,
+//
+// basic mathematical relationship shows that
+//
+//      tan( Arg ) =  tan( alpha )     if N is even;
+//                 = -cot( alpha )      otherwise.
+//
+// The value of alpha is obtained by argument reduction and
+// represented by two working precision numbers r and c where
+//
+//      alpha =  r  +  c     accurately.
+//
+// The reduction method is described in a previous write up.
+// The argument reduction scheme identifies 4 cases. For Cases 2
+// and 4, because |alpha| is small, tan(r+c) and -cot(r+c) can be
+// computed very easily by 2 or 3 terms of the Taylor series
+// expansion as follows:
+//
+// Case 2:
+// -------
+//
+//      tan(r + c) = r + c + r^3/3          ...accurately
+//        -cot(r + c) = -1/(r+c) + r/3          ...accurately
+//
+// Case 4:
+// -------
+//
+//      tan(r + c) = r + c + r^3/3 + 2r^5/15     ...accurately
+//        -cot(r + c) = -1/(r+c) + r/3 + r^3/45     ...accurately
+//
+//
+// The only cases left are Cases 1 and 3 of the argument reduction
+// procedure. These two cases will be merged since after the
+// argument is reduced in either cases, we have the reduced argument
+// represented as r + c and that the magnitude |r + c| is not small
+// enough to allow the usage of a very short approximation.
+//
+// The greatest challenge of this task is that the second terms of
+// the Taylor series for tan(r) and -cot(r)
+//
+//      r + r^3/3 + 2 r^5/15 + ...
+//
+// and
+//
+//      -1/r + r/3 + r^3/45 + ...
+//
+// are not very small when |r| is close to pi/4 and the rounding
+// errors will be a concern if simple polynomial accumulation is
+// used. When |r| < 2^(-2), however, the second terms will be small
+// enough (5 bits or so of right shift) that a normal Horner
+// recurrence suffices. Hence there are two cases that we consider
+// in the accurate computation of tan(r) and cot(r), |r| <= pi/4.
+//
+// Case small_r: |r| < 2^(-2)
+// --------------------------
+//
+// Since Arg = N pi/4 + r + c accurately, we have
+//
+//      tan(Arg) =  tan(r+c)            for N even,
+//            = -cot(r+c)          otherwise.
+//
+// Here for this case, both tan(r) and -cot(r) can be approximated
+// by simple polynomials:
+//
+//      tan(r) =    r + P1_1 r^3 + P1_2 r^5 + ... + P1_9 r^19
+//        -cot(r) = -1/r + Q1_1 r   + Q1_2 r^3 + ... + Q1_7 r^13
+//
+// accurately. Since |r| is relatively small, tan(r+c) and
+// -cot(r+c) can be accurately approximated by replacing r with
+// r+c only in the first two terms of the corresponding polynomials.
+//
+// Note that P1_1 (and Q1_1 for that matter) approximates 1/3 to
+// almost 64 sig. bits, thus
+//
+//      P1_1 (r+c)^3 =  P1_1 r^3 + c * r^2     accurately.
+//
+// Hence,
+//
+//      tan(r+c) =    r + P1_1 r^3 + P1_2 r^5 + ... + P1_9 r^19
+//                     + c*(1 + r^2)
+//
+//        -cot(r+c) = -1/(r+c) + Q1_1 r   + Q1_2 r^3 + ... + Q1_7 r^13
+//               + Q1_1*c
+//
+//
+// Case normal_r: 2^(-2) <= |r| <= pi/4
+// ------------------------------------
+//
+// This case is more likely than the previous one if one considers
+// r to be uniformly distributed in [-pi/4 pi/4].
+//
+// The required calculation is either
+//
+//      tan(r + c)  =  tan(r)  +  correction,  or
+//        -cot(r + c)  = -cot(r)  +  correction.
+//
+// Specifically,
+//
+//      tan(r + c) =  tan(r) + c tan'(r)  + O(c^2)
+//              =  tan(r) + c sec^2(r) + O(c^2)
+//              =  tan(r) + c SEC_sq     ...accurately
+//                as long as SEC_sq approximates sec^2(r)
+//                to, say, 5 bits or so.
+//
+// Similarly,
+//
+//        -cot(r + c) = -cot(r) - c cot'(r)  + O(c^2)
+//              = -cot(r) + c csc^2(r) + O(c^2)
+//              = -cot(r) + c CSC_sq     ...accurately
+//                as long as CSC_sq approximates csc^2(r)
+//                to, say, 5 bits or so.
+//
+// We therefore concentrate on accurately calculating tan(r) and
+// cot(r) for a working-precision number r, |r| <= pi/4 to within
+// 0.1% or so.
+//
+// We will employ a table-driven approach. Let
+//
+//      r = sgn_r * 2^k * 1.b_1 b_2 ... b_5 ... b_63
+//        = sgn_r * ( B + x )
+//
+// where
+//
+//      B = 2^k * 1.b_1 b_2 ... b_5 1
+//         x = |r| - B
+//
+// Now,
+//                   tan(B)  +   tan(x)
+//      tan( B + x ) =  ------------------------
+//                   1 -  tan(B)*tan(x)
+//
+//               /                         \ 
+//               |   tan(B)  +   tan(x)          |
+
+//      = tan(B) +  | ------------------------ - tan(B) |
+//               |     1 -  tan(B)*tan(x)          |
+//               \                         /
+//
+//                 sec^2(B) * tan(x)
+//      = tan(B) + ------------------------
+//                 1 -  tan(B)*tan(x)
+//
+//                (1/[sin(B)*cos(B)]) * tan(x)
+//      = tan(B) + --------------------------------
+//                      cot(B)  -  tan(x)
+//
+//
+// Clearly, the values of tan(B), cot(B) and 1/(sin(B)*cos(B)) are
+// calculated beforehand and stored in a table. Since
+//
+//      |x| <= 2^k * 2^(-6)  <= 2^(-7)  (because k = -1, -2)
+//
+// a very short polynomial will be sufficient to approximate tan(x)
+// accurately. The details involved in computing the last expression
+// will be given in the next section on algorithm description.
+//
+//
+// Now, we turn to the case where cot( B + x ) is needed.
+//
+//
+//                   1 - tan(B)*tan(x)
+//      cot( B + x ) =  ------------------------
+//                   tan(B)  +  tan(x)
+//
+//               /                           \ 
+//               |   1 - tan(B)*tan(x)              |
+
+//      = cot(B) +  | ----------------------- - cot(B) |
+//               |     tan(B)  +  tan(x)            |
+//               \                           /
+//
+//               [tan(B) + cot(B)] * tan(x)
+//      = cot(B) - ----------------------------
+//                   tan(B)  +  tan(x)
+//
+//                (1/[sin(B)*cos(B)]) * tan(x)
+//      = cot(B) - --------------------------------
+//                      tan(B)  +  tan(x)
+//
+//
+// Note that the values of tan(B), cot(B) and 1/(sin(B)*cos(B)) that
+// are needed are the same set of values needed in the previous
+// case.
+//
+// Finally, we can put all the ingredients together as follows:
+//
+//      Arg = N * pi/2 +  r + c          ...accurately
+//
+//      tan(Arg) =  tan(r) + correction    if N is even;
+//            = -cot(r) + correction    otherwise.
+//
+// For Cases 2 and 4,
+//
+//     Case 2:
+//     tan(Arg) =  tan(r + c) = r + c + r^3/3           N even
+//              = -cot(r + c) = -1/(r+c) + r/3           N odd
+//     Case 4:
+//     tan(Arg) =  tan(r + c) = r + c + r^3/3 + 2r^5/15  N even
+//              = -cot(r + c) = -1/(r+c) + r/3 + r^3/45  N odd
+//
+//
+// For Cases 1 and 3,
+//
+//     Case small_r: |r| < 2^(-2)
+//
+//      tan(Arg) =  r + P1_1 r^3 + P1_2 r^5 + ... + P1_9 r^19
+//                     + c*(1 + r^2)               N even
+//
+//                  = -1/(r+c) + Q1_1 r   + Q1_2 r^3 + ... + Q1_7 r^13
+//               + Q1_1*c                    N odd
+//
+//     Case normal_r: 2^(-2) <= |r| <= pi/4
+//
+//      tan(Arg) =  tan(r) + c * sec^2(r)     N even
+//               = -cot(r) + c * csc^2(r)     otherwise
+//
+//     For N even,
+//
+//      tan(Arg) = tan(r) + c*sec^2(r)
+//               = tan( sgn_r * (B+x) ) + c * sec^2(|r|)
+//                  = sgn_r * ( tan(B+x)  + sgn_r*c*sec^2(|r|) )
+//                  = sgn_r * ( tan(B+x)  + sgn_r*c*sec^2(B) )
+//
+// since B approximates |r| to 2^(-6) in relative accuracy.
+//
+//                 /            (1/[sin(B)*cos(B)]) * tan(x)
+//    tan(Arg) = sgn_r * | tan(B) + --------------------------------
+//                 \                     cot(B)  -  tan(x)
+//                                        \ 
+//                       + CORR  |
+
+//                                     /
+// where
+//
+//    CORR = sgn_r*c*tan(B)*SC_inv(B);  SC_inv(B) = 1/(sin(B)*cos(B)).
+//
+// For N odd,
+//
+//      tan(Arg) = -cot(r) + c*csc^2(r)
+//               = -cot( sgn_r * (B+x) ) + c * csc^2(|r|)
+//                  = sgn_r * ( -cot(B+x)  + sgn_r*c*csc^2(|r|) )
+//                  = sgn_r * ( -cot(B+x)  + sgn_r*c*csc^2(B) )
+//
+// since B approximates |r| to 2^(-6) in relative accuracy.
+//
+//                 /            (1/[sin(B)*cos(B)]) * tan(x)
+//    tan(Arg) = sgn_r * | -cot(B) + --------------------------------
+//                 \                     tan(B)  +  tan(x)
+//                                        \ 
+//                       + CORR  |
+
+//                                     /
+// where
+//
+//    CORR = sgn_r*c*cot(B)*SC_inv(B);  SC_inv(B) = 1/(sin(B)*cos(B)).
+//
+//
+// The actual algorithm prescribes how all the mathematical formulas
+// are calculated.
+//
+//
+// 2. Algorithmic Description
+// ==========================
+//
+// 2.1 Computation for Cases 2 and 4.
+// ----------------------------------
+//
+// For Case 2, we use two-term polynomials.
+//
+//    For N even,
+//
+//    rsq := r * r
+//    Result := c + r * rsq * P1_1
+//    Result := r + Result          ...in user-defined rounding
+//
+//    For N odd,
+//    S_hi  := -frcpa(r)               ...8 bits
+//    S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...16 bits
+//    S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...32 bits
+//    S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...64 bits
+//    S_lo  := S_hi*( (1 + S_hi*r) + S_hi*c )
+//    ...S_hi + S_lo is -1/(r+c) to extra precision
+//    S_lo  := S_lo + Q1_1*r
+//
+//    Result := S_hi + S_lo     ...in user-defined rounding
+//
+// For Case 4, we use three-term polynomials
+//
+//    For N even,
+//
+//    rsq := r * r
+//    Result := c + r * rsq * (P1_1 + rsq * P1_2)
+//    Result := r + Result          ...in user-defined rounding
+//
+//    For N odd,
+//    S_hi  := -frcpa(r)               ...8 bits
+//    S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...16 bits
+//    S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...32 bits
+//    S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...64 bits
+//    S_lo  := S_hi*( (1 + S_hi*r) + S_hi*c )
+//    ...S_hi + S_lo is -1/(r+c) to extra precision
+//    rsq   := r * r
+//    P      := Q1_1 + rsq*Q1_2
+//    S_lo  := S_lo + r*P
+//
+//    Result := S_hi + S_lo     ...in user-defined rounding
+//
+//
+// Note that the coefficients P1_1, P1_2, Q1_1, and Q1_2 are
+// the same as those used in the small_r case of Cases 1 and 3
+// below.
+//
+//
+// 2.2 Computation for Cases 1 and 3.
+// ----------------------------------
+// This is further divided into the case of small_r,
+// where |r| < 2^(-2), and the case of normal_r, where |r| lies between
+// 2^(-2) and pi/4.
+//
+// Algorithm for the case of small_r
+// ---------------------------------
+//
+// For N even,
+//      rsq   := r * r
+//      Poly1 := rsq*(P1_1 + rsq*(P1_2 + rsq*P1_3))
+//      r_to_the_8    := rsq * rsq
+//      r_to_the_8    := r_to_the_8 * r_to_the_8
+//      Poly2 := P1_4 + rsq*(P1_5 + rsq*(P1_6 + ... rsq*P1_9))
+//      CORR  := c * ( 1 + rsq )
+//      Poly  := Poly1 + r_to_the_8*Poly2
+//      Result := r*Poly + CORR
+//      Result := r + Result     ...in user-defined rounding
+//      ...note that Poly1 and r_to_the_8 can be computed in parallel
+//      ...with Poly2 (Poly1 is intentionally set to be much
+//      ...shorter than Poly2 so that r_to_the_8 and CORR can be hidden)
+//
+// For N odd,
+//      S_hi  := -frcpa(r)               ...8 bits
+//      S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...16 bits
+//      S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...32 bits
+//      S_hi  := S_hi + S_hi*(1 + S_hi*r)     ...64 bits
+//      S_lo  := S_hi*( (1 + S_hi*r) + S_hi*c )
+//      ...S_hi + S_lo is -1/(r+c) to extra precision
+//      S_lo  := S_lo + Q1_1*c
+//
+//      ...S_hi and S_lo are computed in parallel with
+//      ...the following
+//      rsq := r*r
+//      P   := Q1_1 + rsq*(Q1_2 + rsq*(Q1_3 + ... + rsq*Q1_7))
+//
+//      Result :=  r*P + S_lo
+//      Result :=  S_hi  +  Result      ...in user-defined rounding
+//
+//
+// Algorithm for the case of normal_r
+// ----------------------------------
+//
+// Here, we first consider the computation of tan( r + c ). As
+// presented in the previous section,
+//
+//      tan( r + c )  =  tan(r) + c * sec^2(r)
+//                 =  sgn_r * [ tan(B+x) + CORR ]
+//      CORR = sgn_r * c * tan(B) * 1/[sin(B)*cos(B)]
+//
+// because sec^2(r) = sec^(|r|), and B approximate |r| to 6.5 bits.
+//
+//      tan( r + c ) =
+//           /           (1/[sin(B)*cos(B)]) * tan(x)
+//      sgn_r * | tan(B) + --------------------------------  +
+//           \                     cot(B)  -  tan(x)
+//                                \ 
+//                          CORR  |
+
+//                                /
+//
+// The values of tan(B), cot(B) and 1/(sin(B)*cos(B)) are
+// calculated beforehand and stored in a table. Specifically,
+// the table values are
+//
+//      tan(B)                as  T_hi  +  T_lo;
+//      cot(B)             as  C_hi  +  C_lo;
+//      1/[sin(B)*cos(B)]  as  SC_inv
+//
+// T_hi, C_hi are in  double-precision  memory format;
+// T_lo, C_lo are in  single-precision  memory format;
+// SC_inv     is  in extended-precision memory format.
+//
+// The value of tan(x) will be approximated by a short polynomial of
+// the form
+//
+//      tan(x)  as  x  +  x * P, where
+//           P  =   x^2 * (P2_1 + x^2 * (P2_2 + x^2 * P2_3))
+//
+// Because |x| <= 2^(-7), cot(B) - x approximates cot(B) - tan(x)
+// to a relative accuracy better than 2^(-20). Thus, a good
+// initial guess of 1/( cot(B) - tan(x) ) to initiate the iterative
+// division is:
+//
+//      1/(cot(B) - tan(x))      is approximately
+//      1/(cot(B) -   x)         is
+//      tan(B)/(1 - x*tan(B))    is approximately
+//      T_hi / ( 1 - T_hi * x )  is approximately
+//
+//      T_hi * [ 1 + (Thi * x) + (T_hi * x)^2 ]
+//
+// The calculation of tan(r+c) therefore proceed as follows:
+//
+//      Tx     := T_hi * x
+//      xsq     := x * x
+//
+//      V_hi     := T_hi*(1 + Tx*(1 + Tx))
+//      P     := xsq * (P1_1 + xsq*(P1_2 + xsq*P1_3))
+//      ...V_hi serves as an initial guess of 1/(cot(B) - tan(x))
+//         ...good to about 20 bits of accuracy
+//
+//      tanx     := x + x*P
+//      D     := C_hi - tanx
+//      ...D is a double precision denominator: cot(B) - tan(x)
+//
+//      V_hi     := V_hi + V_hi*(1 - V_hi*D)
+//      ....V_hi approximates 1/(cot(B)-tan(x)) to 40 bits
+//
+//      V_lo     := V_hi * ( [ (1 - V_hi*C_hi) + V_hi*tanx ]
+//                           - V_hi*C_lo )   ...observe all order
+//         ...V_hi + V_lo approximates 1/(cot(B) - tan(x))
+//      ...to extra accuracy
+//
+//      ...               SC_inv(B) * (x + x*P)
+//      ...   tan(B) +      ------------------------- + CORR
+//         ...                cot(B) - (x + x*P)
+//      ...
+//      ... = tan(B) + SC_inv(B)*(x + x*P)*(V_hi + V_lo) + CORR
+//      ...
+//
+//      Sx     := SC_inv * x
+//      CORR     := sgn_r * c * SC_inv * T_hi
+//
+//      ...put the ingredients together to compute
+//      ...               SC_inv(B) * (x + x*P)
+//      ...   tan(B) +      ------------------------- + CORR
+//         ...                cot(B) - (x + x*P)
+//      ...
+//      ... = tan(B) + SC_inv(B)*(x + x*P)*(V_hi + V_lo) + CORR
+//      ...
+//      ... = T_hi + T_lo + CORR +
+//      ...    Sx * V_hi + Sx * V_lo + Sx * P *(V_hi + V_lo)
+//
+//      CORR := CORR + T_lo
+//      tail := V_lo + P*(V_hi + V_lo)
+//         tail := Sx * tail  +  CORR
+//      tail := Sx * V_hi  +  tail
+//         T_hi := sgn_r * T_hi
+//
+//         ...T_hi + sgn_r*tail  now approximate
+//      ...sgn_r*(tan(B+x) + CORR) accurately
+//
+//      Result :=  T_hi + sgn_r*tail  ...in user-defined
+//                           ...rounding control
+//      ...It is crucial that independent paths be fully
+//      ...exploited for performance's sake.
+//
+//
+// Next, we consider the computation of -cot( r + c ). As
+// presented in the previous section,
+//
+//        -cot( r + c )  =  -cot(r) + c * csc^2(r)
+//                 =  sgn_r * [ -cot(B+x) + CORR ]
+//      CORR = sgn_r * c * cot(B) * 1/[sin(B)*cos(B)]
+//
+// because csc^2(r) = csc^(|r|), and B approximate |r| to 6.5 bits.
+//
+//        -cot( r + c ) =
+//           /             (1/[sin(B)*cos(B)]) * tan(x)
+//      sgn_r * | -cot(B) + --------------------------------  +
+//           \                     tan(B)  +  tan(x)
+//                                \ 
+//                          CORR  |
+
+//                                /
+//
+// The values of tan(B), cot(B) and 1/(sin(B)*cos(B)) are
+// calculated beforehand and stored in a table. Specifically,
+// the table values are
+//
+//      tan(B)                as  T_hi  +  T_lo;
+//      cot(B)             as  C_hi  +  C_lo;
+//      1/[sin(B)*cos(B)]  as  SC_inv
+//
+// T_hi, C_hi are in  double-precision  memory format;
+// T_lo, C_lo are in  single-precision  memory format;
+// SC_inv     is  in extended-precision memory format.
+//
+// The value of tan(x) will be approximated by a short polynomial of
+// the form
+//
+//      tan(x)  as  x  +  x * P, where
+//           P  =   x^2 * (P2_1 + x^2 * (P2_2 + x^2 * P2_3))
+//
+// Because |x| <= 2^(-7), tan(B) + x approximates tan(B) + tan(x)
+// to a relative accuracy better than 2^(-18). Thus, a good
+// initial guess of 1/( tan(B) + tan(x) ) to initiate the iterative
+// division is:
+//
+//      1/(tan(B) + tan(x))      is approximately
+//      1/(tan(B) +   x)         is
+//      cot(B)/(1 + x*cot(B))    is approximately
+//      C_hi / ( 1 + C_hi * x )  is approximately
+//
+//      C_hi * [ 1 - (C_hi * x) + (C_hi * x)^2 ]
+//
+// The calculation of -cot(r+c) therefore proceed as follows:
+//
+//      Cx     := C_hi * x
+//      xsq     := x * x
+//
+//      V_hi     := C_hi*(1 - Cx*(1 - Cx))
+//      P     := xsq * (P1_1 + xsq*(P1_2 + xsq*P1_3))
+//      ...V_hi serves as an initial guess of 1/(tan(B) + tan(x))
+//         ...good to about 18 bits of accuracy
+//
+//      tanx     := x + x*P
+//      D     := T_hi + tanx
+//      ...D is a double precision denominator: tan(B) + tan(x)
+//
+//      V_hi     := V_hi + V_hi*(1 - V_hi*D)
+//      ....V_hi approximates 1/(tan(B)+tan(x)) to 40 bits
+//
+//      V_lo     := V_hi * ( [ (1 - V_hi*T_hi) - V_hi*tanx ]
+//                           - V_hi*T_lo )   ...observe all order
+//         ...V_hi + V_lo approximates 1/(tan(B) + tan(x))
+//      ...to extra accuracy
+//
+//      ...               SC_inv(B) * (x + x*P)
+//      ...  -cot(B) +      ------------------------- + CORR
+//         ...                tan(B) + (x + x*P)
+//      ...
+//      ... =-cot(B) + SC_inv(B)*(x + x*P)*(V_hi + V_lo) + CORR
+//      ...
+//
+//      Sx     := SC_inv * x
+//      CORR     := sgn_r * c * SC_inv * C_hi
+//
+//      ...put the ingredients together to compute
+//      ...               SC_inv(B) * (x + x*P)
+//      ...  -cot(B) +      ------------------------- + CORR
+//         ...                tan(B) + (x + x*P)
+//      ...
+//      ... =-cot(B) + SC_inv(B)*(x + x*P)*(V_hi + V_lo) + CORR
+//      ...
+//      ... =-C_hi - C_lo + CORR +
+//      ...    Sx * V_hi + Sx * V_lo + Sx * P *(V_hi + V_lo)
+//
+//      CORR := CORR - C_lo
+//      tail := V_lo + P*(V_hi + V_lo)
+//         tail := Sx * tail  +  CORR
+//      tail := Sx * V_hi  +  tail
+//         C_hi := -sgn_r * C_hi
+//
+//         ...C_hi + sgn_r*tail now approximates
+//      ...sgn_r*(-cot(B+x) + CORR) accurately
+//
+//      Result :=  C_hi + sgn_r*tail   in user-defined rounding control
+//      ...It is crucial that independent paths be fully
+//      ...exploited for performance's sake.
+//
+// 3. Implementation Notes
+// =======================
+//
+//   Table entries T_hi, T_lo; C_hi, C_lo; SC_inv
+//
+//   Recall that 2^(-2) <= |r| <= pi/4;
+//
+//      r = sgn_r * 2^k * 1.b_1 b_2 ... b_63
+//
+//   and
+//
+//        B = 2^k * 1.b_1 b_2 b_3 b_4 b_5 1
+//
+//   Thus, for k = -2, possible values of B are
+//
+//          B = 2^(-2) * ( 1 + index/32  +  1/64 ),
+//      index ranges from 0 to 31
+//
+//   For k = -1, however, since |r| <= pi/4 = 0.78...
+//   possible values of B are
+//
+//        B = 2^(-1) * ( 1 + index/32  +  1/64 )
+//      index ranges from 0 to 19.
+//
+//
+
+#include "libm_support.h"
+
+#ifdef _LIBC
+.rodata
+#else
+.data
+#endif
+.align 128
+
+TANL_BASE_CONSTANTS:
+ASM_TYPE_DIRECTIVE(TANL_BASE_CONSTANTS,@object)
+data4    0x4B800000, 0xCB800000, 0x38800000, 0xB8800000 // two**24, -two**24
+                                                        // two**-14, -two**-14
+data4    0x4E44152A, 0xA2F9836E, 0x00003FFE, 0x00000000 // two_by_pi
+data4    0xCE81B9F1, 0xC84D32B0, 0x00004016, 0x00000000 // P_0
+data4    0x2168C235, 0xC90FDAA2, 0x00003FFF, 0x00000000 // P_1
+data4    0xFC8F8CBB, 0xECE675D1, 0x0000BFBD, 0x00000000 // P_2
+data4    0xACC19C60, 0xB7ED8FBB, 0x0000BF7C, 0x00000000 // P_3
+data4    0x5F000000, 0xDF000000, 0x00000000, 0x00000000 // two_to_63, -two_to_63
+data4    0x6EC6B45A, 0xA397E504, 0x00003FE7, 0x00000000 // Inv_P_0
+data4    0xDBD171A1, 0x8D848E89, 0x0000BFBF, 0x00000000 // d_1
+data4    0x18A66F8E, 0xD5394C36, 0x0000BF7C, 0x00000000 // d_2
+data4    0x2168C234, 0xC90FDAA2, 0x00003FFE, 0x00000000 // PI_BY_4
+data4    0x2168C234, 0xC90FDAA2, 0x0000BFFE, 0x00000000 // MPI_BY_4
+data4    0x3E800000, 0xBE800000, 0x00000000, 0x00000000 // two**-2, -two**-2
+data4    0x2F000000, 0xAF000000, 0x00000000, 0x00000000 // two**-33, -two**-33
+data4    0xAAAAAABD, 0xAAAAAAAA, 0x00003FFD, 0x00000000 // P1_1
+data4    0x88882E6A, 0x88888888, 0x00003FFC, 0x00000000 // P1_2
+data4    0x0F0177B6, 0xDD0DD0DD, 0x00003FFA, 0x00000000 // P1_3
+data4    0x646B8C6D, 0xB327A440, 0x00003FF9, 0x00000000 // P1_4
+data4    0x1D5F7D20, 0x91371B25, 0x00003FF8, 0x00000000 // P1_5
+data4    0x61C67914, 0xEB69A5F1, 0x00003FF6, 0x00000000 // P1_6
+data4    0x019318D2, 0xBEDD37BE, 0x00003FF5, 0x00000000 // P1_7
+data4    0x3C794015, 0x9979B146, 0x00003FF4, 0x00000000 // P1_8
+data4    0x8C6EB58A, 0x8EBD21A3, 0x00003FF3, 0x00000000 // P1_9
+data4    0xAAAAAAB4, 0xAAAAAAAA, 0x00003FFD, 0x00000000 // Q1_1
+data4    0x0B5FC93E, 0xB60B60B6, 0x00003FF9, 0x00000000 // Q1_2
+data4    0x0C9BBFBF, 0x8AB355E0, 0x00003FF6, 0x00000000 // Q1_3
+data4    0xCBEE3D4C, 0xDDEBBC89, 0x00003FF2, 0x00000000 // Q1_4
+data4    0x5F80BBB6, 0xB3548A68, 0x00003FEF, 0x00000000 // Q1_5
+data4    0x4CED5BF1, 0x91362560, 0x00003FEC, 0x00000000 // Q1_6
+data4    0x8EE92A83, 0xF189D95A, 0x00003FE8, 0x00000000 // Q1_7
+data4    0xAAAB362F, 0xAAAAAAAA, 0x00003FFD, 0x00000000 // P2_1
+data4    0xE97A6097, 0x88888886, 0x00003FFC, 0x00000000 // P2_2
+data4    0x25E716A1, 0xDD108EE0, 0x00003FFA, 0x00000000 // P2_3
+//
+//  Entries T_hi   double-precision memory format
+//  Index = 0,1,...,31  B = 2^(-2)*(1+Index/32+1/64)
+//  Entries T_lo  single-precision memory format
+//  Index = 0,1,...,31  B = 2^(-2)*(1+Index/32+1/64)
+//
+data4    0x62400794, 0x3FD09BC3, 0x23A05C32, 0x00000000
+data4    0xDFFBC074, 0x3FD124A9, 0x240078B2, 0x00000000
+data4    0x5BD4920F, 0x3FD1AE23, 0x23826B8E, 0x00000000
+data4    0x15E2701D, 0x3FD23835, 0x22D31154, 0x00000000
+data4    0x63739C2D, 0x3FD2C2E4, 0x2265C9E2, 0x00000000
+data4    0xAFEEA48B, 0x3FD34E36, 0x245C05EB, 0x00000000
+data4    0x7DBB35D1, 0x3FD3DA31, 0x24749F2D, 0x00000000
+data4    0x67321619, 0x3FD466DA, 0x2462CECE, 0x00000000
+data4    0x1F94A4D5, 0x3FD4F437, 0x246D0DF1, 0x00000000
+data4    0x740C3E6D, 0x3FD5824D, 0x240A85B5, 0x00000000
+data4    0x4CB1E73D, 0x3FD61123, 0x23F96E33, 0x00000000
+data4    0xAD9EA64B, 0x3FD6A0BE, 0x247C5393, 0x00000000
+data4    0xB804FD01, 0x3FD73125, 0x241F3B29, 0x00000000
+data4    0xAB53EE83, 0x3FD7C25E, 0x2479989B, 0x00000000
+data4    0xE6640EED, 0x3FD8546F, 0x23B343BC, 0x00000000
+data4    0xE8AF1892, 0x3FD8E75F, 0x241454D1, 0x00000000
+data4    0x53928BDA, 0x3FD97B35, 0x238613D9, 0x00000000
+data4    0xEB9DE4DE, 0x3FDA0FF6, 0x22859FA7, 0x00000000
+data4    0x99ECF92D, 0x3FDAA5AB, 0x237A6D06, 0x00000000
+data4    0x6D8F1796, 0x3FDB3C5A, 0x23952F6C, 0x00000000
+data4    0x9CFB8BE4, 0x3FDBD40A, 0x2280FC95, 0x00000000
+data4    0x87943100, 0x3FDC6CC3, 0x245D2EC0, 0x00000000
+data4    0xB736C500, 0x3FDD068C, 0x23C4AD7D, 0x00000000
+data4    0xE1DDBC31, 0x3FDDA16D, 0x23D076E6, 0x00000000
+data4    0xEB515A93, 0x3FDE3D6E, 0x244809A6, 0x00000000
+data4    0xE6E9E5F1, 0x3FDEDA97, 0x220856C8, 0x00000000
+data4    0x1963CE69, 0x3FDF78F1, 0x244BE993, 0x00000000
+data4    0x7D635BCE, 0x3FE00C41, 0x23D21799, 0x00000000
+data4    0x1C302CD3, 0x3FE05CAB, 0x248A1B1D, 0x00000000
+data4    0xDB6A1FA0, 0x3FE0ADB9, 0x23D53E33, 0x00000000
+data4    0x4A20BA81, 0x3FE0FF72, 0x24DB9ED5, 0x00000000
+data4    0x153FA6F5, 0x3FE151D9, 0x24E9E451, 0x00000000
+//
+//  Entries T_hi   double-precision memory format
+//  Index = 0,1,...,19  B = 2^(-1)*(1+Index/32+1/64)
+//  Entries T_lo  single-precision memory format
+//  Index = 0,1,...,19  B = 2^(-1)*(1+Index/32+1/64)
+//
+data4    0xBA1BE39E, 0x3FE1CEC4, 0x24B60F9E, 0x00000000
+data4    0x5ABD9B2D, 0x3FE277E4, 0x248C2474, 0x00000000
+data4    0x0272B110, 0x3FE32418, 0x247B8311, 0x00000000
+data4    0x890E2DF0, 0x3FE3D38B, 0x24C55751, 0x00000000
+data4    0x46236871, 0x3FE4866D, 0x24E5BC34, 0x00000000
+data4    0x45E044B0, 0x3FE53CEE, 0x24001BA4, 0x00000000
+data4    0x82EC06E4, 0x3FE5F742, 0x24B973DC, 0x00000000
+data4    0x25DF43F9, 0x3FE6B5A1, 0x24895440, 0x00000000
+data4    0xCAFD348C, 0x3FE77844, 0x240021CA, 0x00000000
+data4    0xCEED6B92, 0x3FE83F6B, 0x24C45372, 0x00000000
+data4    0xA34F3665, 0x3FE90B58, 0x240DAD33, 0x00000000
+data4    0x2C1E56B4, 0x3FE9DC52, 0x24F846CE, 0x00000000
+data4    0x27041578, 0x3FEAB2A4, 0x2323FB6E, 0x00000000
+data4    0x9DD8C373, 0x3FEB8E9F, 0x24B3090B, 0x00000000
+data4    0x65C9AA7B, 0x3FEC709B, 0x2449F611, 0x00000000
+data4    0xACCF8435, 0x3FED58F4, 0x23616A7E, 0x00000000
+data4    0x97635082, 0x3FEE480F, 0x24C2FEAE, 0x00000000
+data4    0xF0ACC544, 0x3FEF3E57, 0x242CE964, 0x00000000
+data4    0xF7E06E4B, 0x3FF01E20, 0x2480D3EE, 0x00000000
+data4    0x8A798A69, 0x3FF0A125, 0x24DB8967, 0x00000000
+//
+//  Entries C_hi   double-precision memory format
+//  Index = 0,1,...,31  B = 2^(-2)*(1+Index/32+1/64)
+//  Entries C_lo  single-precision memory format
+//  Index = 0,1,...,31  B = 2^(-2)*(1+Index/32+1/64)
+//
+data4    0xE63EFBD0, 0x400ED3E2, 0x259D94D4, 0x00000000
+data4    0xC515DAB5, 0x400DDDB4, 0x245F0537, 0x00000000
+data4    0xBE19A79F, 0x400CF57A, 0x25D4EA9F, 0x00000000
+data4    0xD15298ED, 0x400C1A06, 0x24AE40A0, 0x00000000
+data4    0x164B2708, 0x400B4A4C, 0x25A5AAB6, 0x00000000
+data4    0x5285B068, 0x400A855A, 0x25524F18, 0x00000000
+data4    0x3FFA549F, 0x4009CA5A, 0x24C999C0, 0x00000000
+data4    0x646AF623, 0x4009188A, 0x254FD801, 0x00000000
+data4    0x6084D0E7, 0x40086F3C, 0x2560F5FD, 0x00000000
+data4    0xA29A76EE, 0x4007CDD2, 0x255B9D19, 0x00000000
+data4    0x6C8ECA95, 0x400733BE, 0x25CB021B, 0x00000000
+data4    0x1F8DDC52, 0x4006A07E, 0x24AB4722, 0x00000000
+data4    0xC298AD58, 0x4006139B, 0x252764E2, 0x00000000
+data4    0xBAD7164B, 0x40058CAB, 0x24DAF5DB, 0x00000000
+data4    0xAE31A5D3, 0x40050B4B, 0x25EA20F4, 0x00000000
+data4    0x89F85A8A, 0x40048F21, 0x2583A3E8, 0x00000000
+data4    0xA862380D, 0x400417DA, 0x25DCC4CC, 0x00000000
+data4    0x1088FCFE, 0x4003A52B, 0x2430A492, 0x00000000
+data4    0xCD3527D5, 0x400336CC, 0x255F77CF, 0x00000000
+data4    0x5760766D, 0x4002CC7F, 0x25DA0BDA, 0x00000000
+data4    0x11CE02E3, 0x40026607, 0x256FF4A2, 0x00000000
+data4    0xD37BBE04, 0x4002032C, 0x25208AED, 0x00000000
+data4    0x7F050775, 0x4001A3BD, 0x24B72DD6, 0x00000000
+data4    0xA554848A, 0x40014789, 0x24AB4DAA, 0x00000000
+data4    0x323E81B7, 0x4000EE65, 0x2584C440, 0x00000000
+data4    0x21CF1293, 0x40009827, 0x25C9428D, 0x00000000
+data4    0x3D415EEB, 0x400044A9, 0x25DC8482, 0x00000000
+data4    0xBD72C577, 0x3FFFE78F, 0x257F5070, 0x00000000
+data4    0x75EFD28E, 0x3FFF4AC3, 0x23EBBF7A, 0x00000000
+data4    0x60B52DDE, 0x3FFEB2AF, 0x22EECA07, 0x00000000
+data4    0x35204180, 0x3FFE1F19, 0x24191079, 0x00000000
+data4    0x54F7E60A, 0x3FFD8FCA, 0x248D3058, 0x00000000
+//
+//  Entries C_hi   double-precision memory format
+//  Index = 0,1,...,19  B = 2^(-1)*(1+Index/32+1/64)
+//  Entries C_lo  single-precision memory format
+//  Index = 0,1,...,19  B = 2^(-1)*(1+Index/32+1/64)
+//
+data4    0x79F6FADE, 0x3FFCC06A, 0x239C7886, 0x00000000
+data4    0x891662A6, 0x3FFBB91F, 0x250BD191, 0x00000000
+data4    0x529F155D, 0x3FFABFB6, 0x256CC3E6, 0x00000000
+data4    0x2E964AE9, 0x3FF9D300, 0x250843E3, 0x00000000
+data4    0x89DCB383, 0x3FF8F1EF, 0x2277C87E, 0x00000000
+data4    0x7C87DBD6, 0x3FF81B93, 0x256DA6CF, 0x00000000
+data4    0x1042EDE4, 0x3FF74F14, 0x2573D28A, 0x00000000
+data4    0x1784B360, 0x3FF68BAF, 0x242E489A, 0x00000000
+data4    0x7C923C4C, 0x3FF5D0B5, 0x2532D940, 0x00000000
+data4    0xF418EF20, 0x3FF51D88, 0x253C7DD6, 0x00000000
+data4    0x02F88DAE, 0x3FF4719A, 0x23DB59BF, 0x00000000
+data4    0x49DA0788, 0x3FF3CC66, 0x252B4756, 0x00000000
+data4    0x0B980DB8, 0x3FF32D77, 0x23FE585F, 0x00000000
+data4    0xE56C987A, 0x3FF2945F, 0x25378A63, 0x00000000
+data4    0xB16523F6, 0x3FF200BD, 0x247BB2E0, 0x00000000
+data4    0x8CE27778, 0x3FF17235, 0x24446538, 0x00000000
+data4    0xFDEFE692, 0x3FF0E873, 0x2514638F, 0x00000000
+data4    0x33154062, 0x3FF0632C, 0x24A7FC27, 0x00000000
+data4    0xB3EF115F, 0x3FEFC42E, 0x248FD0FE, 0x00000000
+data4    0x135D26F6, 0x3FEEC9E8, 0x2385C719, 0x00000000
+//
+//  Entries SC_inv in Swapped IEEE format (extended)
+//  Index = 0,1,...,31  B = 2^(-2)*(1+Index/32+1/64)
+//
+data4    0x1BF30C9E, 0x839D6D4A, 0x00004001, 0x00000000
+data4    0x554B0EB0, 0x80092804, 0x00004001, 0x00000000
+data4    0xA1CF0DE9, 0xF959F94C, 0x00004000, 0x00000000
+data4    0x77378677, 0xF3086BA0, 0x00004000, 0x00000000
+data4    0xCCD4723C, 0xED154515, 0x00004000, 0x00000000
+data4    0x1C27CF25, 0xE7790944, 0x00004000, 0x00000000
+data4    0x8DDACB88, 0xE22D037D, 0x00004000, 0x00000000
+data4    0x89C73522, 0xDD2B2D8A, 0x00004000, 0x00000000
+data4    0xBB2C1171, 0xD86E1A23, 0x00004000, 0x00000000
+data4    0xDFF5E0F9, 0xD3F0E288, 0x00004000, 0x00000000
+data4    0x283BEBD5, 0xCFAF16B1, 0x00004000, 0x00000000
+data4    0x0D88DD53, 0xCBA4AFAA, 0x00004000, 0x00000000
+data4    0xCA67C43D, 0xC7CE03CC, 0x00004000, 0x00000000
+data4    0x0CA0DDB0, 0xC427BC82, 0x00004000, 0x00000000
+data4    0xF13D8CAB, 0xC0AECD57, 0x00004000, 0x00000000
+data4    0x71ECE6B1, 0xBD606C38, 0x00004000, 0x00000000
+data4    0xA44C4929, 0xBA3A0A96, 0x00004000, 0x00000000
+data4    0xE5CCCEC1, 0xB7394F6F, 0x00004000, 0x00000000
+data4    0x9637D8BC, 0xB45C1203, 0x00004000, 0x00000000
+data4    0x92CB051B, 0xB1A05528, 0x00004000, 0x00000000
+data4    0x6BA2FFD0, 0xAF04432B, 0x00004000, 0x00000000
+data4    0x7221235F, 0xAC862A23, 0x00004000, 0x00000000
+data4    0x5F00A9D1, 0xAA2478AF, 0x00004000, 0x00000000
+data4    0x81E082BF, 0xA7DDBB0C, 0x00004000, 0x00000000
+data4    0x45684FEE, 0xA5B0987D, 0x00004000, 0x00000000
+data4    0x627A8F53, 0xA39BD0F5, 0x00004000, 0x00000000
+data4    0x6EC5C8B0, 0xA19E3B03, 0x00004000, 0x00000000
+data4    0x91CD7C66, 0x9FB6C1F0, 0x00004000, 0x00000000
+data4    0x1FA3DF8A, 0x9DE46410, 0x00004000, 0x00000000
+data4    0xA8F6B888, 0x9C263139, 0x00004000, 0x00000000
+data4    0xC27B0450, 0x9A7B4968, 0x00004000, 0x00000000
+data4    0x5EE614EE, 0x98E2DB7E, 0x00004000, 0x00000000
+//
+//  Entries SC_inv in Swapped IEEE format (extended)
+//  Index = 0,1,...,19  B = 2^(-1)*(1+Index/32+1/64)
+//
+data4    0x13B2B5BA, 0x969F335C, 0x00004000, 0x00000000
+data4    0xD4C0F548, 0x93D446D9, 0x00004000, 0x00000000
+data4    0x61B798AF, 0x9147094F, 0x00004000, 0x00000000
+data4    0x758787AC, 0x8EF317CC, 0x00004000, 0x00000000
+data4    0xB99EEFDB, 0x8CD498B3, 0x00004000, 0x00000000
+data4    0xDFF8BC37, 0x8AE82A7D, 0x00004000, 0x00000000
+data4    0xE3C55D42, 0x892AD546, 0x00004000, 0x00000000
+data4    0xD15573C1, 0x8799FEA9, 0x00004000, 0x00000000
+data4    0x435A4B4C, 0x86335F88, 0x00004000, 0x00000000
+data4    0x3E93A87B, 0x84F4FB6E, 0x00004000, 0x00000000
+data4    0x80A382FB, 0x83DD1952, 0x00004000, 0x00000000
+data4    0xA4CB8C9E, 0x82EA3D7F, 0x00004000, 0x00000000
+data4    0x6861D0A8, 0x821B247C, 0x00004000, 0x00000000
+data4    0x63E8D244, 0x816EBED1, 0x00004000, 0x00000000
+data4    0x27E4CFC6, 0x80E42D91, 0x00004000, 0x00000000
+data4    0x28E64AFD, 0x807ABF8D, 0x00004000, 0x00000000
+data4    0x863B4FD8, 0x8031EF26, 0x00004000, 0x00000000
+data4    0xAE8C11FD, 0x800960AD, 0x00004000, 0x00000000
+data4    0x5FDBEC21, 0x8000E147, 0x00004000, 0x00000000
+data4    0xA07791FA, 0x80186650, 0x00004000, 0x00000000
+ASM_SIZE_DIRECTIVE(TANL_BASE_CONSTANTS)
+
+Arg                 = f8   
+Result              = f8
+fp_tmp              = f9
+U_2                 = f10
+rsq                =  f11
+C_hi                = f12
+C_lo                = f13
+T_hi                = f14
+T_lo                = f15
+
+N_0                 = f32
+d_1                 = f33
+MPI_BY_4            = f34
+tail                = f35
+tanx                = f36
+Cx                  = f37
+Sx                  = f38
+sgn_r               = f39
+CORR                = f40
+P                   = f41
+D                   = f42
+ArgPrime            = f43
+P_0                 = f44
+
+P2_1                = f45
+P2_2                = f46
+P2_3                = f47
+
+P1_1                = f45
+P1_2                = f46
+P1_3                = f47
+
+P1_4                = f48
+P1_5                = f49
+P1_6                = f50
+P1_7                = f51
+P1_8                = f52
+P1_9                = f53
+
+TWO_TO_63           = f54
+NEGTWO_TO_63        = f55
+x                   = f56
+xsq                 = f57
+Tx                  = f58
+Tx1                 = f59
+Set                 = f60
+poly1               = f61
+poly2               = f62
+Poly                = f63
+Poly1               = f64
+Poly2               = f65
+r_to_the_8          = f66
+B                   = f67
+SC_inv              = f68
+Pos_r               = f69
+N_0_fix             = f70
+PI_BY_4             = f71
+NEGTWO_TO_NEG2      = f72
+TWO_TO_24           = f73
+TWO_TO_NEG14        = f74
+TWO_TO_NEG33        = f75
+NEGTWO_TO_24        = f76
+NEGTWO_TO_NEG14     = f76
+NEGTWO_TO_NEG33     = f77
+two_by_PI           = f78
+N                   = f79
+N_fix               = f80
+P_1                 = f81
+P_2                 = f82
+P_3                 = f83
+s_val               = f84
+w                   = f85
+c                   = f86
+r                   = f87
+A                   = f89
+a                   = f90
+t                   = f91
+U_1                 = f92
+d_2                 = f93
+TWO_TO_NEG2         = f94
+Q1_1                = f95
+Q1_2                = f96
+Q1_3                = f97
+Q1_4                = f98
+Q1_5                = f99
+Q1_6                = f100
+Q1_7                = f101
+Q1_8                = f102
+S_hi                = f103
+S_lo                = f104
+V_hi                = f105
+V_lo                = f106
+U_hi                = f107
+U_lo                = f108
+U_hiabs             = f109
+V_hiabs             = f110
+V                   = f111
+Inv_P_0             = f112
+
+GR_SAVE_B0     = r33
+GR_SAVE_GP     = r34
+GR_SAVE_PFS    = r35
+delta1         = r36
+table_ptr1     = r37
+table_ptr2     = r38
+i_0            = r39
+i_1            = r40 
+N_fix_gr       = r41 
+N_inc          = r42 
+exp_Arg        = r43 
+exp_r          = r44 
+sig_r          = r45 
+lookup         = r46   
+table_offset   = r47 
+Create_B       = r48 
+gr_tmp         = r49
+
+.section .text
+.global tanl
+.proc tanl
+tanl:
+#ifdef _LIBC
+.global __tanl
+.proc __tanl
+__tanl:
+#endif
+{ .mfi
+alloc r32 = ar.pfs, 0,17,2,0
+(p0)   fclass.m.unc  p6,p0 = Arg, 0x1E7
+      addl gr_tmp = -1,r0             
+}
+{ .mfi
+  nop.m 0
+(p0)   fclass.nm.unc  p7,p0 = Arg, 0x1FF
+  nop.i 0
+};;
+
+{ .mfi
+(p0)  addl           table_ptr1   = @ltoff(TANL_BASE_CONSTANTS), gp
+       nop.f 999
+      nop.i 0
+}
+;;
+{ .mmi
+(p0)  ld8 table_ptr1 = [table_ptr1]
+      setf.sig fp_tmp = gr_tmp   // Make a constant so fmpy produces inexact
+      nop.i 999
+}
+;;
+
+//
+//     Check for NatVals, Infs , NaNs, and Zeros 
+//     Check for everything - if false, then must be pseudo-zero
+//     or pseudo-nan.
+//     Local table pointer
+//
+{ .mbb
+(p0)   add table_ptr2 = 96, table_ptr1
+(p6)   br.cond.spnt L(TANL_SPECIAL) 
+(p7)   br.cond.spnt L(TANL_SPECIAL) ;;
+}
+//
+//     Point to Inv_P_0
+//     Branch out to deal with unsupporteds and special values. 
+//
+{ .mmf
+(p0)   ldfs TWO_TO_24 = [table_ptr1],4
+(p0)   ldfs TWO_TO_63 = [table_ptr2],4
+//
+//     Load -2**24, load -2**63.
+//
+(p0)   fcmp.eq.s0 p0, p6 = Arg, f1 ;;
+}
+{ .mfi
+(p0)   ldfs NEGTWO_TO_63 = [table_ptr2],12
+(p0)   fnorm.s1     Arg = Arg
+       nop.i 999
+}
+//
+//     Load 2**24, Load 2**63.
+//
+{ .mmi
+(p0)   ldfs NEGTWO_TO_24 = [table_ptr1],12 ;;
+//
+//     Do fcmp to generate Denormal exception 
+//     - can't do FNORM (will generate Underflow when U is unmasked!)
+//     Normalize input argument.
+//
+(p0)   ldfe two_by_PI = [table_ptr1],16
+       nop.i 999
+}
+{ .mmi
+(p0)   ldfe Inv_P_0 = [table_ptr2],16 ;;
+(p0)   ldfe d_1 = [table_ptr2],16
+       nop.i 999
+}
+//
+//     Decide about the paths to take:
+//     PR_1 and PR_3 set if -2**24 < Arg < 2**24 - CASE 1 OR 2
+//     OTHERWISE - CASE 3 OR 4
+//     Load inverse of P_0 .
+//     Set PR_6 if Arg <= -2**63
+//     Are there any Infs, NaNs, or zeros?
+//
+{ .mmi
+(p0)   ldfe P_0 = [table_ptr1],16 ;;
+(p0)   ldfe d_2 = [table_ptr2],16
+       nop.i 999
+}
+//
+//     Set PR_8 if Arg <= -2**24
+//     Set PR_6 if Arg >=  2**63
+//
+{ .mmi
+(p0)   ldfe P_1 = [table_ptr1],16 ;;
+(p0)   ldfe PI_BY_4 = [table_ptr2],16
+       nop.i 999
+}
+//
+//     Set PR_8 if Arg >= 2**24
+//
+{ .mmi
+(p0)   ldfe P_2 = [table_ptr1],16 ;;
+(p0)   ldfe   MPI_BY_4 = [table_ptr2],16
+       nop.i 999
+}
+//
+//     Load  P_2 and PI_BY_4
+//
+{ .mfi
+(p0)   ldfe   P_3 = [table_ptr1],16
+       nop.f 999
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p0)   fcmp.le.unc.s1 p6,p7 = Arg,NEGTWO_TO_63
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p0)   fcmp.le.unc.s1 p8,p9 = Arg,NEGTWO_TO_24
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p7)   fcmp.ge.s1 p6,p0 = Arg,TWO_TO_63
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p9)   fcmp.ge.s1 p8,p0 = Arg,TWO_TO_24
+       nop.i 999 ;;
+}
+{ .mib
+       nop.m 999
+       nop.i 999
+//
+//     Load  P_3 and -PI_BY_4
+//
+(p6)   br.cond.spnt L(TANL_ARG_TOO_LARGE) ;;
+}
+{ .mib
+       nop.m 999
+       nop.i 999
+//
+//     Load 2**(-2).
+//     Load -2**(-2).
+//     Branch out if we have a special argument.
+//     Branch out if the magnitude of the input argument is too large
+//     - do this branch before the next.
+//
+(p8)   br.cond.spnt L(TANL_LARGER_ARG) ;;
+}
+//
+//     Branch to Cases 3 or 4 if Arg <= -2**24 or Arg >= 2**24
+//
+{ .mfi
+(p0)   ldfs TWO_TO_NEG2 = [table_ptr2],4
+//     ARGUMENT REDUCTION CODE - CASE 1 and 2
+//     Load 2**(-2).
+//     Load -2**(-2).
+(p0)   fmpy.s1 N = Arg,two_by_PI
+       nop.i 999 ;;
+}
+{ .mfi
+(p0)   ldfs NEGTWO_TO_NEG2 = [table_ptr2],12
+//
+//     N = Arg * 2/pi
+//
+(p0)   fcmp.lt.unc.s1 p8,p9= Arg,PI_BY_4
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     if Arg < pi/4,  set PR_8.
+//
+(p8)   fcmp.gt.s1 p8,p9= Arg,MPI_BY_4
+       nop.i 999 ;;
+}
+//
+//     Case 1: Is |r| < 2**(-2).
+//     Arg is the same as r in this case.
+//     r = Arg
+//     c = 0
+//
+{ .mfi
+(p8)   mov N_fix_gr = r0
+//
+//     if Arg > -pi/4, reset PR_8.
+//     Select the case when |Arg| < pi/4 - set PR[8] = true.
+//     Else Select the case when |Arg| >= pi/4 - set PR[9] = true.
+//
+(p0)   fcvt.fx.s1 N_fix = N
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Grab the integer part of N .
+//
+(p8)   mov r = Arg
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p8)   mov c = f0
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p8)   fcmp.lt.unc.s1 p10, p11 = Arg, TWO_TO_NEG2
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p10)  fcmp.gt.s1 p10,p0 = Arg, NEGTWO_TO_NEG2
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 2: Place integer part of N in GP register.
+//
+(p9)   fcvt.xf N = N_fix
+       nop.i 999 ;;
+}
+{ .mib
+(p9)   getf.sig N_fix_gr = N_fix
+       nop.i 999
+//
+//     Case 2: Convert integer N_fix back to normalized floating-point value.
+//
+(p10)  br.cond.spnt L(TANL_SMALL_R) ;;
+}
+{ .mib
+       nop.m 999
+       nop.i 999
+(p8)   br.cond.sptk L(TANL_NORMAL_R) ;;
+}
+//
+//     Case 1: PR_3 is only affected  when PR_1 is set.
+//
+{ .mmi
+(p9)   ldfs TWO_TO_NEG33 = [table_ptr2], 4 ;;
+//
+//     Case 2: Load 2**(-33).
+//
+(p9)   ldfs NEGTWO_TO_NEG33 = [table_ptr2], 4
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 2: Load -2**(-33).
+//
+(p9)   fnma.s1 s_val = N, P_1, Arg
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p9)   fmpy.s1 w = N, P_2
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 2: w = N * P_2
+//     Case 2: s_val = -N * P_1  + Arg
+//
+(p0)   fcmp.lt.unc.s1 p9,p8 = s_val, TWO_TO_NEG33
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Decide between case_1 and case_2 reduce:
+//
+(p9)   fcmp.gt.s1 p9, p8 = s_val, NEGTWO_TO_NEG33
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 1_reduce:  s <= -2**(-33) or s >= 2**(-33)
+//     Case 2_reduce: -2**(-33) < s < 2**(-33)
+//
+(p8)   fsub.s1 r = s_val, w
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p9)   fmpy.s1 w = N, P_3
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p9)   fma.s1  U_1 = N, P_2, w
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 1_reduce: Is |r| < 2**(-2), if so set PR_10
+//     else set PR_11.
+//
+(p8)   fsub.s1 c = s_val, r
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 1_reduce: r = s + w (change sign)
+//     Case 2_reduce: w = N * P_3 (change sign)
+//
+(p8)   fcmp.lt.unc.s1 p10, p11 = r, TWO_TO_NEG2
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p10)  fcmp.gt.s1 p10, p11 = r, NEGTWO_TO_NEG2
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p9)   fsub.s1 r = s_val, U_1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 1_reduce: c is complete here.
+//     c = c + w (w has not been negated.)
+//     Case 2_reduce: r is complete here - continue to calculate c .
+//     r = s - U_1
+//
+(p9)   fms.s1 U_2 = N, P_2, U_1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 1_reduce: c = s - r
+//     Case 2_reduce: U_1 = N * P_2 + w
+//
+(p8)   fsub.s1 c = c, w
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p9)   fsub.s1 s_val = s_val, r
+       nop.i 999
+}
+{ .mfb
+       nop.m 999
+//
+//     Case 2_reduce:
+//     U_2 = N * P_2 - U_1
+//     Not needed until later.
+//
+(p9)   fadd.s1 U_2 = U_2, w
+//
+//     Case 2_reduce:
+//     s = s - r
+//     U_2 = U_2 + w
+//
+(p10)  br.cond.spnt L(TANL_SMALL_R) ;;
+}
+{ .mib
+       nop.m 999
+       nop.i 999
+(p11)  br.cond.sptk L(TANL_NORMAL_R) ;;
+}
+{ .mii
+       nop.m 999
+//
+//     Case 2_reduce:
+//     c = c - U_2
+//     c is complete here
+//     Argument reduction ends here.
+//
+(p9)   extr.u i_1 = N_fix_gr, 0, 1 ;;
+(p9)   cmp.eq.unc p11, p12 = 0x0000,i_1 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     Is i_1  even or odd?
+//     if i_1 == 0, set p11, else set p12.
+//
+(p11)  fmpy.s1 rsq = r, r
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12)  frcpa.s1 S_hi,p0 = f1, r
+       nop.i 999
+}
+
+
+
+//
+//     Case 1: Branch to SMALL_R or NORMAL_R.
+//     Case 1 is done now.
+//
+
+{ .mfi
+(p9)   addl           table_ptr1   = @ltoff(TANL_BASE_CONSTANTS), gp
+(p9)   fsub.s1 c = s_val, U_1
+       nop.i 999 ;;
+}
+;;
+
+{ .mmi
+(p9)  ld8 table_ptr1 = [table_ptr1]
+      nop.m 999
+      nop.i 999
+}
+;;
+
+
+{ .mmi
+(p9)   add table_ptr1 = 224, table_ptr1 ;;
+(p9)   ldfe P1_1 = [table_ptr1],144
+       nop.i 999 ;;
+}
+//
+//     Get [i_1] -  lsb of N_fix_gr .
+//     Load P1_1 and point to Q1_1 .
+//
+{ .mfi
+(p9)   ldfe Q1_1 = [table_ptr1] , 0
+//
+//     N even: rsq = r * Z
+//     N odd:  S_hi = frcpa(r)
+//
+(p12)  fmerge.ns S_hi = S_hi, S_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//     Case 2_reduce:
+//     c = s - U_1
+//
+(p9)   fsub.s1 c = c, U_2
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12)  fma.s1  poly1 = S_hi, r, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N odd:  Change sign of S_hi
+//
+(p11)  fmpy.s1 rsq = rsq, P1_1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12)  fma.s1 S_hi = S_hi, poly1, S_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N even: rsq = rsq * P1_1
+//     N odd:  poly1 =  1.0 +  S_hi * r    16 bits partial  account for necessary
+//
+(p11)  fma.s1 Result = r, rsq, c
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N even: Result = c  + r * rsq
+//     N odd:  S_hi  = S_hi + S_hi*poly1  16 bits account for necessary
+//
+(p12)  fma.s1 poly1 = S_hi, r, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N even: Result = Result + r
+//     N odd:  poly1  = 1.0 + S_hi * r        32 bits partial
+//
+(p11)  fadd.s0 Result = r, Result
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12)  fma.s1  S_hi = S_hi, poly1, S_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N even: Result1 = Result + r
+//     N odd:   S_hi  = S_hi * poly1 + S_hi   32 bits
+//
+(p12)  fma.s1 poly1 = S_hi, r, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N odd:  poly1  =  S_hi * r + 1.0       64 bits partial
+//
+(p12)  fma.s1 S_hi = S_hi, poly1, S_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N odd:  poly1  =  S_hi * poly + 1.0    64 bits
+//
+(p12)  fma.s1 poly1 = S_hi, r, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N odd:  poly1  =  S_hi * r + 1.0
+//
+(p12)  fma.s1 poly1 = S_hi, c, poly1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N odd:  poly1  =  S_hi * c + poly1
+//
+(p12)  fmpy.s1 S_lo = S_hi, poly1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//     N odd:  S_lo  =  S_hi *  poly1
+//
+(p12)  fma.s1 S_lo = Q1_1, r, S_lo
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//     N odd:  Result =  S_hi + S_lo
+//
+(p0)   fmpy.s0 fp_tmp = fp_tmp, fp_tmp  // Dummy mult to set inexact
+       nop.i 999 ;;
+}
+{ .mfb
+       nop.m 999
+//
+//     N odd:  S_lo  =  S_lo + Q1_1 * r
+//
+(p12)  fadd.s0 Result = S_hi, S_lo
+(p0)   br.ret.sptk b0 ;;
+}
+
+
+L(TANL_LARGER_ARG): 
+
+//
+// ARGUMENT REDUCTION CODE - CASE 3 and 4
+//
+
+{ .mfi
+(p0)  addl           table_ptr1   = @ltoff(TANL_BASE_CONSTANTS), gp
+(p0)  fmpy.s1 N_0 = Arg, Inv_P_0 
+       nop.i 999
+}
+;;
+
+{ .mmi
+(p0)  ld8 table_ptr1 = [table_ptr1]
+      nop.m 999
+      nop.i 999
+}
+;;
+
+
+//
+//    Adjust table_ptr1 to beginning of table.
+//    N_0 = Arg * Inv_P_0
+//
+{ .mmi
+(p0)  add table_ptr1 = 8, table_ptr1 ;;
+//
+//    Point to  2*-14
+//
+(p0)  ldfs TWO_TO_NEG14 = [table_ptr1], 4
+       nop.i 999 ;;
+}
+//
+//    Load 2**(-14).
+//
+{ .mmi
+(p0)  ldfs NEGTWO_TO_NEG14 = [table_ptr1], 180 ;;
+//
+//    N_0_fix  = integer part of N_0 .
+//    Adjust table_ptr1 to beginning of table.
+//
+(p0)  ldfs TWO_TO_NEG2 = [table_ptr1], 4
+       nop.i 999 ;;
+}
+//
+//    Make N_0 the integer part.
+//
+{ .mfi
+(p0)  ldfs NEGTWO_TO_NEG2 = [table_ptr1]
+//
+//    Load -2**(-14).
+//
+(p0)  fcvt.fx.s1 N_0_fix = N_0
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p0)  fcvt.xf N_0 = N_0_fix
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p0)  fnma.s1 ArgPrime = N_0, P_0, Arg
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p0)  fmpy.s1 w = N_0, d_1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    ArgPrime = -N_0 * P_0 + Arg
+//    w  = N_0 * d_1
+//
+(p0)  fmpy.s1 N = ArgPrime, two_by_PI
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N = ArgPrime * 2/pi
+//
+(p0)  fcvt.fx.s1 N_fix = N
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N_fix is the integer part.
+//
+(p0)  fcvt.xf N = N_fix
+       nop.i 999 ;;
+}
+{ .mfi
+(p0)  getf.sig N_fix_gr = N_fix
+       nop.f 999
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N is the integer part of the reduced-reduced argument.
+//    Put the integer in a GP register.
+//
+(p0)  fnma.s1 s_val = N, P_1, ArgPrime
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p0)  fnma.s1 w = N, P_2, w
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    s_val = -N*P_1 + ArgPrime
+//    w = -N*P_2 + w
+//
+(p0)  fcmp.lt.unc.s1 p11, p10 = s_val, TWO_TO_NEG14
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p11) fcmp.gt.s1 p11, p10 = s_val, NEGTWO_TO_NEG14
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 3: r = s_val + w (Z complete)
+//    Case 4: U_hi = N_0 * d_1
+//
+(p10) fmpy.s1 V_hi = N, P_2
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p11) fmpy.s1 U_hi = N_0, d_1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 3: r = s_val + w (Z complete)
+//    Case 4: U_hi = N_0 * d_1
+//
+(p11) fmpy.s1 V_hi = N, P_2
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p11) fmpy.s1 U_hi = N_0, d_1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Decide between case 3 and 4:
+//    Case 3:  s <= -2**(-14) or s >= 2**(-14)
+//    Case 4: -2**(-14) < s < 2**(-14)
+//
+(p10) fadd.s1 r = s_val, w
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p11) fmpy.s1 w = N, P_3
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: We need abs of both U_hi and V_hi - dont
+//    worry about switched sign of V_hi .
+//
+(p11) fsub.s1 A = U_hi, V_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: A =  U_hi + V_hi
+//    Note: Worry about switched sign of V_hi, so subtract instead of add.
+//
+(p11) fnma.s1 V_lo = N, P_2, V_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p11) fms.s1 U_lo = N_0, d_1, U_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p11) fabs V_hiabs = V_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: V_hi = N * P_2
+//            w = N * P_3
+//    Note the product does not include the (-) as in the writeup
+//    so (-) missing for V_hi and w .
+(p10) fadd.s1 r = s_val, w
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 3: c = s_val - r
+//    Case 4: U_lo = N_0 * d_1 - U_hi
+//
+(p11) fabs U_hiabs = U_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p11) fmpy.s1 w = N, P_3
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: Set P_12 if U_hiabs >= V_hiabs
+//
+(p11) fadd.s1 C_hi = s_val, A
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: C_hi = s_val + A
+//
+(p11) fadd.s1 t = U_lo, V_lo
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 3: Is |r| < 2**(-2), if so set PR_7
+//    else set PR_8.
+//    Case 3: If PR_7 is set, prepare to branch to Small_R.
+//    Case 3: If PR_8 is set, prepare to branch to Normal_R.
+//
+(p10) fsub.s1 c = s_val, r
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 3: c = (s - r) + w (c complete)
+//
+(p11) fcmp.ge.unc.s1 p12, p13 = U_hiabs, V_hiabs
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p11) fms.s1 w = N_0, d_2, w
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: V_hi = N * P_2
+//            w = N * P_3
+//    Note the product does not include the (-) as in the writeup
+//    so (-) missing for V_hi and w .
+//
+(p10) fcmp.lt.unc.s1 p14, p15 = r, TWO_TO_NEG2
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p14) fcmp.gt.s1 p14, p15 = r, NEGTWO_TO_NEG2
+       nop.i 999 ;;
+}
+{ .mfb
+       nop.m 999
+//
+//    Case 4: V_lo = -N * P_2 - V_hi (U_hi is in place of V_hi in writeup)
+//    Note: the (-) is still missing for V_hi .
+//    Case 4: w = w + N_0 * d_2
+//    Note: the (-) is now incorporated in w .
+//
+(p10) fadd.s1 c = c, w
+//
+//    Case 4: t = U_lo + V_lo
+//    Note: remember V_lo should be (-), subtract instead of add. NO
+//
+(p14) br.cond.spnt L(TANL_SMALL_R) ;;
+}
+{ .mib
+       nop.m 999
+       nop.i 999
+(p15) br.cond.spnt L(TANL_NORMAL_R) ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 3: Vector off when |r| < 2**(-2).  Recall that PR_3 will be true.
+//    The remaining stuff is for Case 4.
+//
+(p12) fsub.s1 a = U_hi, A
+(p11) extr.u i_1 = N_fix_gr, 0, 1 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: C_lo = s_val - C_hi
+//
+(p11) fadd.s1 t = t, w
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p13) fadd.s1 a = V_hi, A
+       nop.i 999 ;;
+}
+
+
+
+//
+//    Case 4: a = U_hi - A
+//            a = V_hi - A (do an add to account for missing (-) on V_hi
+//
+
+{ .mfi
+(p11)  addl           table_ptr1   = @ltoff(TANL_BASE_CONSTANTS), gp
+(p11) fsub.s1 C_lo = s_val, C_hi
+       nop.i 999
+}
+;;
+
+
+
+//
+//    Case 4: a = (U_hi - A)  + V_hi
+//            a = (V_hi - A)  + U_hi
+//    In each case account for negative missing form V_hi .
+//
+
+
+{ .mmi
+(p11)  ld8 table_ptr1 = [table_ptr1]
+      nop.m 999
+      nop.i 999
+}
+;;
+
+
+//
+//    Case 4: C_lo = (s_val - C_hi) + A
+//
+{ .mmi
+(p11) add table_ptr1 = 224, table_ptr1 ;;
+(p11) ldfe P1_1 = [table_ptr1], 16
+       nop.i 999 ;;
+}
+{ .mfi
+(p11) ldfe P1_2 = [table_ptr1], 128
+//
+//    Case 4: w = U_lo + V_lo  + w
+//
+(p12) fsub.s1 a = a, V_hi
+       nop.i 999 ;;
+}
+//
+//    Case 4: r = C_hi + C_lo
+//
+{ .mfi
+(p11) ldfe Q1_1 = [table_ptr1], 16
+(p11) fadd.s1 C_lo = C_lo, A
+       nop.i 999 ;;
+}
+//
+//    Case 4: c = C_hi - r
+//    Get [i_1] - lsb of N_fix_gr.
+//
+{ .mfi
+(p11) ldfe Q1_2 = [table_ptr1], 16
+       nop.f 999
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p13) fsub.s1 a = U_hi, a
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p11) fadd.s1 t = t, a
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: t = t + a
+//
+(p11) fadd.s1 C_lo = C_lo, t
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: C_lo = C_lo + t
+//
+(p11) fadd.s1 r = C_hi, C_lo
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p11) fsub.s1 c = C_hi, r
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    Case 4: c = c + C_lo  finished.
+//    Is i_1  even or odd?
+//    if i_1 == 0, set PR_4, else set PR_5.
+//
+// r and c have been computed.
+// We known whether this is the sine or cosine routine.
+// Make sure ftz mode is set - should be automatic when using wre
+(p0)  fmpy.s1 rsq = r, r
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p11) fadd.s1 c = c , C_lo
+(p11) cmp.eq.unc p11, p12 =  0x0000, i_1 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) frcpa.s1 S_hi, p0 = f1, r
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd: Change sign of S_hi
+//
+(p11) fma.s1 Result = rsq, P1_2, P1_1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 P = rsq, Q1_2, Q1_1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:  Result  =  S_hi + S_lo      (User supplied rounding mode for C1)
+//
+(p0)   fmpy.s0 fp_tmp = fp_tmp, fp_tmp  // Dummy mult to set inexact
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: rsq = r * r
+//    N odd:  S_hi = frcpa(r)
+//
+(p12) fmerge.ns S_hi = S_hi, S_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: rsq = rsq * P1_2 + P1_1
+//    N odd:  poly1 =  1.0 +  S_hi * r    16 bits partial  account for necessary
+//
+(p11) fmpy.s1 Result = rsq, Result
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly1 = S_hi, r,f1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: Result =  Result * rsq
+//    N odd:  S_hi  = S_hi + S_hi*poly1  16 bits account for necessary
+//
+(p11) fma.s1 Result = r, Result, c
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 S_hi = S_hi, poly1, S_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:   S_hi  = S_hi * poly1 + S_hi   32 bits
+//
+(p11) fadd.s0 Result= r, Result
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly1 =  S_hi, r, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: Result = Result * r + c
+//    N odd:  poly1  = 1.0 + S_hi * r        32 bits partial
+//
+(p12) fma.s1 S_hi = S_hi, poly1, S_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly1 = S_hi, r, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: Result1 = Result + r  (Rounding mode S0)
+//    N odd:  poly1  =  S_hi * r + 1.0       64 bits partial
+//
+(p12) fma.s1 S_hi = S_hi, poly1, S_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:  poly1  =  S_hi * poly + S_hi    64 bits
+//
+(p12) fma.s1 poly1 = S_hi, r, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:  poly1  =  S_hi * r + 1.0
+//
+(p12) fma.s1 poly1 = S_hi, c, poly1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:  poly1  =  S_hi * c + poly1
+//
+(p12) fmpy.s1 S_lo = S_hi, poly1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:  S_lo  =  S_hi *  poly1
+//
+(p12) fma.s1 S_lo = P, r, S_lo
+       nop.i 999 ;;
+}
+{ .mfb
+       nop.m 999
+//
+//    N odd:  S_lo  =  S_lo + r * P
+//
+(p12) fadd.s0 Result = S_hi, S_lo
+(p0)   br.ret.sptk b0 ;;
+}
+
+
+L(TANL_SMALL_R): 
+{ .mii
+       nop.m 999
+(p0)  extr.u i_1 = N_fix_gr, 0, 1 ;;
+(p0)  cmp.eq.unc p11, p12 = 0x0000, i_1
+}
+{ .mfi
+       nop.m 999
+(p0)  fmpy.s1 rsq = r, r
+       nop.i 999 ;;
+}
+{ .mfi
+(p0)  addl           table_ptr1   = @ltoff(TANL_BASE_CONSTANTS), gp
+(p12) frcpa.s1 S_hi, p0 = f1, r
+       nop.i 999
+}
+;;
+
+
+{ .mmi
+(p0)  ld8 table_ptr1 = [table_ptr1]
+      nop.m 999
+      nop.i 999
+}
+;;
+
+// *****************************************************************
+// *****************************************************************
+// *****************************************************************
+
+
+{ .mmi
+(p0)  add table_ptr1 = 224, table_ptr1 ;;
+(p0)  ldfe P1_1 = [table_ptr1], 16
+       nop.i 999 ;;
+}
+//    r and c have been computed.
+//    We known whether this is the sine or cosine routine.
+//    Make sure ftz mode is set - should be automatic when using wre
+//    |r| < 2**(-2)
+{ .mfi
+(p0)  ldfe P1_2 = [table_ptr1], 16
+(p11) fmpy.s1 r_to_the_8 = rsq, rsq
+       nop.i 999 ;;
+}
+//
+//    Set table_ptr1 to beginning of constant table.
+//    Get [i_1] - lsb of N_fix_gr.
+//
+{ .mfi
+(p0)  ldfe P1_3 = [table_ptr1], 96
+//
+//    N even: rsq = r * r
+//    N odd:  S_hi = frcpa(r)
+//
+(p12) fmerge.ns S_hi = S_hi, S_hi
+       nop.i 999 ;;
+}
+//
+//    Is i_1  even or odd?
+//    if i_1 == 0, set PR_11.
+//    if i_1 != 0, set PR_12.
+//
+{ .mfi
+(p11) ldfe P1_9 = [table_ptr1], -16
+//
+//    N even: Poly2 = P1_7 + Poly2 * rsq
+//    N odd:  poly2 = Q1_5 + poly2 * rsq
+//
+(p11) fadd.s1 CORR = rsq, f1
+       nop.i 999 ;;
+}
+{ .mmi
+(p11) ldfe P1_8 = [table_ptr1], -16 ;;
+//
+//    N even: Poly1 = P1_2 + P1_3 * rsq
+//    N odd:  poly1 =  1.0 +  S_hi * r     
+//    16 bits partial  account for necessary (-1)
+//
+(p11) ldfe P1_7 = [table_ptr1], -16
+       nop.i 999 ;;
+}
+//
+//    N even: Poly1 = P1_1 + Poly1 * rsq
+//    N odd:  S_hi  =  S_hi + S_hi * poly1)     16 bits account for necessary
+//
+{ .mfi
+(p11) ldfe P1_6 = [table_ptr1], -16
+//
+//    N even: Poly2 = P1_5 + Poly2 * rsq
+//    N odd:  poly2 = Q1_3 + poly2 * rsq
+//
+(p11) fmpy.s1 r_to_the_8 = r_to_the_8, r_to_the_8
+       nop.i 999 ;;
+}
+//
+//    N even: Poly1 =  Poly1 * rsq
+//    N odd:  poly1  = 1.0 + S_hi * r         32 bits partial
+//
+{ .mfi
+(p11) ldfe P1_5 = [table_ptr1], -16
+(p12) fma.s1 poly1 =  S_hi, r, f1
+       nop.i 999 ;;
+}
+
+//
+//    N even: CORR =  CORR * c
+//    N odd:  S_hi  =  S_hi * poly1 + S_hi    32 bits
+//
+
+//
+//    N even: Poly2 = P1_6 + Poly2 * rsq
+//    N odd:  poly2 = Q1_4 + poly2 * rsq
+//
+
+{ .mmf
+(p11) ldfe P1_4 = [table_ptr1], -16
+(p0)  addl           table_ptr2   = @ltoff(TANL_BASE_CONSTANTS), gp
+(p11) fmpy.s1 CORR =  CORR, c
+}
+;;
+
+
+{ .mmi
+(p0)  ld8 table_ptr2 = [table_ptr2]
+      nop.m 999
+      nop.i 999
+}
+;;
+
+
+{ .mii
+(p0)  add table_ptr2 = 464, table_ptr2
+       nop.i 999 ;;
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p11) fma.s1 Poly1 = P1_3, rsq, P1_2
+       nop.i 999 ;;
+}
+{ .mfi
+(p0)  ldfe Q1_7 = [table_ptr2], -16
+(p12) fma.s1 S_hi = S_hi, poly1, S_hi
+       nop.i 999 ;;
+}
+{ .mfi
+(p0)  ldfe Q1_6 = [table_ptr2], -16
+(p11) fma.s1 Poly2 = P1_9, rsq, P1_8
+       nop.i 999 ;;
+}
+{ .mmi
+(p0)  ldfe Q1_5 = [table_ptr2], -16 ;;
+(p12) ldfe Q1_4 = [table_ptr2], -16
+       nop.i 999 ;;
+}
+{ .mfi
+(p12) ldfe Q1_3 = [table_ptr2], -16
+//
+//    N even: Poly2 = P1_8 + P1_9 * rsq
+//    N odd:  poly2 = Q1_6 + Q1_7 * rsq
+//
+(p11) fma.s1 Poly1 = Poly1, rsq, P1_1
+       nop.i 999 ;;
+}
+{ .mfi
+(p12) ldfe Q1_2 = [table_ptr2], -16
+(p12) fma.s1 poly1 = S_hi, r, f1
+       nop.i 999 ;;
+}
+{ .mfi
+(p12) ldfe Q1_1 = [table_ptr2], -16
+(p11) fma.s1 Poly2 = Poly2, rsq, P1_7
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: CORR =  rsq + 1
+//    N even: r_to_the_8 =  rsq * rsq
+//
+(p11) fmpy.s1 Poly1 = Poly1, rsq
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 S_hi = S_hi, poly1, S_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly2 = Q1_7, rsq, Q1_6
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p11) fma.s1 Poly2 = Poly2, rsq, P1_6
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly1 = S_hi, r, f1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly2 = poly2, rsq, Q1_5
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p11) fma.s1 Poly2= Poly2, rsq, P1_5
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 S_hi =  S_hi, poly1, S_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly2 = poly2, rsq, Q1_4
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: r_to_the_8 = r_to_the_8 * r_to_the_8
+//    N odd:  poly1  =  S_hi * r + 1.0       64 bits partial
+//
+(p11) fma.s1 Poly2 = Poly2, rsq, P1_4
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: Result = CORR + Poly * r
+//    N odd:  P = Q1_1 + poly2 * rsq
+//
+(p12) fma.s1 poly1 = S_hi, r, f1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly2 = poly2, rsq, Q1_3
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: Poly2 = P1_4 + Poly2 * rsq
+//    N odd:  poly2 = Q1_2 + poly2 * rsq
+//
+(p11) fma.s1 Poly = Poly2, r_to_the_8, Poly1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly1 = S_hi, c, poly1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 poly2 = poly2, rsq, Q1_2
+       nop.i 999 ;;
+}
+
+{ .mfi
+       nop.m 999
+//
+//    N even: Poly = Poly1 + Poly2 * r_to_the_8
+//    N odd:  S_hi =  S_hi * poly1 + S_hi    64 bits
+//
+(p11) fma.s1 Result = Poly, r, CORR
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: Result =  r + Result  (User supplied rounding mode)
+//    N odd:  poly1  =  S_hi * c + poly1
+//
+(p12) fmpy.s1 S_lo = S_hi, poly1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fma.s1 P = poly2, rsq, Q1_1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:  poly1  =  S_hi * r + 1.0
+//
+//
+//    N odd:  S_lo  =  S_hi *  poly1
+//
+(p11) fadd.s0 Result = Result, r
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:  S_lo  =  Q1_1 * c + S_lo
+//
+(p12) fma.s1 S_lo = Q1_1, c, S_lo
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p0)   fmpy.s0 fp_tmp = fp_tmp, fp_tmp  // Dummy mult to set inexact
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd:  Result =  S_lo + r * P
+//
+(p12) fma.s1 Result = P, r, S_lo
+       nop.i 999 ;;
+}
+{ .mfb
+       nop.m 999
+//
+//    N odd:  Result = Result + S_hi  (user supplied rounding mode)
+//
+(p12) fadd.s0 Result = Result, S_hi
+(p0)   br.ret.sptk b0 ;;
+}
+
+
+L(TANL_NORMAL_R): 
+{ .mfi
+(p0)  getf.sig sig_r = r
+// *******************************************************************
+// *******************************************************************
+// *******************************************************************
+//
+//    r and c have been computed.
+//    Make sure ftz mode is set - should be automatic when using wre
+//
+//
+//    Get [i_1] -  lsb of N_fix_gr alone.
+//
+(p0)  fmerge.s  Pos_r = f1, r
+(p0)  extr.u i_1 = N_fix_gr, 0, 1 ;;
+}
+{ .mfi
+       nop.m 999
+(p0)  fmerge.s  sgn_r =  r, f1
+(p0)  cmp.eq.unc p11, p12 = 0x0000, i_1 ;;
+}
+{ .mfi
+       nop.m 999
+       nop.f 999
+(p0)  extr.u lookup = sig_r, 58, 5
+}
+{ .mlx
+       nop.m 999
+(p0)  movl Create_B = 0x8200000000000000 ;;
+}
+{ .mfi
+(p0)  addl           table_ptr1   = @ltoff(TANL_BASE_CONSTANTS), gp
+       nop.f 999
+(p0)  dep Create_B = lookup, Create_B, 58, 5
+}
+;;
+
+
+//
+//    Get [i_1] -  lsb of N_fix_gr alone.
+//    Pos_r = abs (r)
+//
+
+
+{ .mmi
+(p0)  ld8 table_ptr1 = [table_ptr1]
+      nop.m 999
+      nop.i 999
+}
+;;
+
+
+{ .mmi
+       nop.m 999
+(p0)  setf.sig B = Create_B
+//
+//    Set table_ptr1 and table_ptr2 to base address of
+//    constant table.
+//
+(p0)  add table_ptr1 = 480, table_ptr1 ;;
+}
+{ .mmb
+       nop.m 999
+//
+//    Is i_1 or i_0  == 0 ?
+//    Create the constant  1 00000 1000000000000000000000...
+//
+(p0)  ldfe P2_1 = [table_ptr1], 16
+       nop.b 999
+}
+{ .mmi
+       nop.m 999 ;;
+(p0)  getf.exp exp_r = Pos_r
+       nop.i 999
+}
+//
+//    Get r's exponent
+//    Get r's significand
+//
+{ .mmi
+(p0)  ldfe P2_2 = [table_ptr1], 16 ;;
+//
+//    Get the 5 bits or r for the lookup.   1.xxxxx ....
+//    from sig_r.
+//    Grab  lsb of exp of B
+//
+(p0)  ldfe P2_3 = [table_ptr1], 16
+       nop.i 999 ;;
+}
+{ .mii
+       nop.m 999
+(p0)  andcm table_offset = 0x0001, exp_r ;;
+(p0)  shl table_offset = table_offset, 9 ;;
+}
+{ .mii
+       nop.m 999
+//
+//    Deposit   0 00000 1000000000000000000000... on
+//              1 xxxxx yyyyyyyyyyyyyyyyyyyyyy...,
+//    getting rid of the ys.
+//    Is  B = 2** -2 or  B= 2** -1? If 2**-1, then
+//    we want an offset of 512 for table addressing.
+//
+(p0)  shladd table_offset = lookup, 4, table_offset ;;
+//
+//    B =  ........ 1xxxxx 1000000000000000000...
+//
+(p0)  add table_ptr1 = table_ptr1, table_offset ;;
+}
+{ .mmb
+       nop.m 999
+//
+//   B =  ........ 1xxxxx 1000000000000000000...
+//   Convert B so it has the same exponent as Pos_r
+//
+(p0)  ldfd T_hi = [table_ptr1], 8
+       nop.b 999 ;;
+}
+
+
+
+//
+//    x = |r| - B
+//    Load T_hi.
+//    Load C_hi.
+//
+
+{ .mmf
+(p0)  addl           table_ptr2   = @ltoff(TANL_BASE_CONSTANTS), gp
+(p0)  ldfs T_lo = [table_ptr1]
+(p0)  fmerge.se B = Pos_r, B
+}
+;;
+
+
+{ .mmi
+(p0)  ld8 table_ptr2 = [table_ptr2]
+      nop.m 999
+      nop.i 999
+}
+;;
+
+
+{ .mii
+(p0)  add table_ptr2 = 1360, table_ptr2
+       nop.i 999 ;;
+(p0)  add table_ptr2 = table_ptr2, table_offset ;;
+}
+{ .mfi
+(p0)  ldfd C_hi = [table_ptr2], 8
+(p0)  fsub.s1 x = Pos_r, B
+       nop.i 999 ;;
+}
+{ .mii
+(p0)  ldfs C_lo = [table_ptr2],255
+       nop.i 999 ;;
+//
+//    xsq = x * x
+//    N even: Tx = T_hi * x
+//    Load T_lo.
+//    Load C_lo - increment pointer to get SC_inv 
+//    - cant get all the way, do an add later.
+//
+(p0)  add table_ptr2 = 569, table_ptr2 ;;
+}
+//
+//    N even: Tx1 = Tx + 1
+//    N odd:  Cx1 = 1 - Cx
+//
+{ .mfi
+(p0)  ldfe SC_inv = [table_ptr2], 0
+       nop.f 999
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p0)  fmpy.s1 xsq = x, x
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p11) fmpy.s1 Tx = T_hi, x
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fmpy.s1 Cx = C_hi, x
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd: Cx = C_hi * x
+//
+(p0)  fma.s1 P = P2_3, xsq, P2_2
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: P = P2_3 + P2_2 * xsq
+//
+(p11) fadd.s1 Tx1 = Tx, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: D = C_hi - tanx
+//    N odd: D = T_hi + tanx
+//
+(p11) fmpy.s1 CORR = SC_inv, T_hi
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p0)  fmpy.s1 Sx = SC_inv, x
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fmpy.s1 CORR = SC_inv, C_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fsub.s1 V_hi = f1, Cx
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p0)  fma.s1 P = P, xsq, P2_1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: P = P2_1 + P * xsq
+//
+(p11) fma.s1 V_hi = Tx, Tx1, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: Result  = sgn_r * tail + T_hi (user rounding mode for C1)
+//    N odd:  Result  = sgn_r * tail + C_hi (user rounding mode for C1)
+//
+(p0)   fmpy.s0 fp_tmp = fp_tmp, fp_tmp  // Dummy mult to set inexact
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p0)  fmpy.s1 CORR = CORR, c
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fnma.s1 V_hi = Cx,V_hi,f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: V_hi = Tx * Tx1 + 1
+//    N odd: Cx1 = 1 - Cx * Cx1
+//
+(p0)  fmpy.s1 P = P, xsq
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: P = P * xsq
+//
+(p11) fmpy.s1 V_hi = V_hi, T_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: tail = P * tail + V_lo
+//
+(p11) fmpy.s1 T_hi = sgn_r, T_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p0)  fmpy.s1 CORR = CORR, sgn_r
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+(p12) fmpy.s1 V_hi = V_hi,C_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: V_hi = T_hi * V_hi
+//    N odd: V_hi  = C_hi * V_hi
+//
+(p0)  fma.s1 tanx = P, x, x
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fnmpy.s1 C_hi = sgn_r, C_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: V_lo = 1 - V_hi + C_hi
+//    N odd: V_lo = 1 - V_hi + T_hi
+//
+(p11) fadd.s1 CORR = CORR, T_lo
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fsub.s1 CORR = CORR, C_lo
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: tanx = x + x * P
+//    N even and odd: Sx = SC_inv * x
+//
+(p11) fsub.s1 D = C_hi, tanx
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fadd.s1 D = T_hi, tanx
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N odd: CORR = SC_inv * C_hi
+//    N even: CORR = SC_inv * T_hi
+//
+(p0)  fnma.s1 D = V_hi, D, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: D = 1 - V_hi * D
+//    N even and odd: CORR = CORR * c
+//
+(p0)  fma.s1 V_hi = V_hi, D, V_hi
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: V_hi = V_hi + V_hi * D
+//    N even and odd: CORR = sgn_r * CORR
+//
+(p11) fnma.s1 V_lo = V_hi, C_hi, f1
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fnma.s1 V_lo = V_hi, T_hi, f1
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: CORR = COOR + T_lo
+//    N odd: CORR = CORR - C_lo
+//
+(p11) fma.s1 V_lo = tanx, V_hi, V_lo
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fnma.s1 V_lo = tanx, V_hi, V_lo
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: V_lo = V_lo + V_hi * tanx
+//    N odd: V_lo = V_lo - V_hi * tanx
+//
+(p11) fnma.s1 V_lo = C_lo, V_hi, V_lo
+       nop.i 999
+}
+{ .mfi
+       nop.m 999
+(p12) fnma.s1 V_lo = T_lo, V_hi, V_lo
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N  even: V_lo = V_lo - V_hi * C_lo
+//    N  odd: V_lo = V_lo - V_hi * T_lo
+//
+(p0)  fmpy.s1 V_lo = V_hi, V_lo
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: V_lo = V_lo * V_hi
+//
+(p0)  fadd.s1 tail = V_hi, V_lo
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: tail = V_hi + V_lo
+//
+(p0)  fma.s1 tail = tail, P, V_lo
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even: T_hi = sgn_r * T_hi
+//    N odd : C_hi = -sgn_r * C_hi
+//
+(p0)  fma.s1 tail = tail, Sx, CORR
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even and odd: tail = Sx * tail + CORR
+//
+(p0)  fma.s1 tail = V_hi, Sx, tail
+       nop.i 999 ;;
+}
+{ .mfi
+       nop.m 999
+//
+//    N even an odd: tail = Sx * V_hi + tail
+//
+(p11) fma.s0 Result = sgn_r, tail, T_hi
+       nop.i 999
+}
+{ .mfb
+       nop.m 999
+(p12) fma.s0 Result = sgn_r, tail, C_hi
+(p0)   br.ret.sptk b0 ;;
+}
+
+L(TANL_SPECIAL):
+{ .mfb
+        nop.m 999
+(p0)   fmpy.s0 Arg = Arg, f0
+(p0)   br.ret.sptk b0 ;;
+}
+//
+//     Code for NaNs, Unsupporteds, Infs, or +/- zero ?
+//     Invalid raised for Infs and SNaNs.
+//
+
+.endp  tanl
+ASM_SIZE_DIRECTIVE(tanl)
+
+// *******************************************************************
+// *******************************************************************
+// *******************************************************************
+//
+//     Special Code to handle very large argument case.
+//     Call int pi_by_2_reduce(&x,&r,&c)
+//     for |arguments| >= 2**63
+//     (Arg or x) is in f8
+//     Address to save r and c as double
+// *******************************************************************
+// *******************************************************************
+// *******************************************************************
+
+.proc __libm_callout
+__libm_callout:
+L(TANL_ARG_TOO_LARGE): 
+.prologue
+{ .mfi
+        add   r50=-32,sp                        // Parameter: r address
+        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 [r50] = f0,16                      // Clear Parameter r on stack
+        add  r49 = 16,sp                        // Parameter x address
+.save   b0, GR_SAVE_B0
+        mov GR_SAVE_B0=b0                       // Save b0
+};;
+.body
+{ .mib
+        stfe [r50] = f0,-16                     // Clear Parameter c on stack
+        nop.i 0
+        nop.b 0
+}
+{ .mib
+        stfe [r49] = Arg                        // Store Parameter x on stack
+        nop.i 0
+(p0)    br.call.sptk b0=__libm_pi_by_2_reduce# ;;
+};;
+//
+//     Load 2^-2
+//
+{ .mmi
+(p0)   ldfe  Arg =[r49],16   
+//
+//     Call argument reduction
+//
+(p0)   ldfs  TWO_TO_NEG2 = [table_ptr2],4
+//     Get Arg off stack
+//     Get r off stack - hi order part
+//     Get c off stack - lo order part
+(p0)   mov   N_fix_gr = r8 ;;
+}
+{ .mmb
+(p0)   ldfe  r =[r50],16  
+(p0)   ldfs  NEGTWO_TO_NEG2 = [table_ptr2],4
+       nop.b 999 ;;
+}
+{ .mfi
+(p0)   ldfe  c =[r50],-32  
+       nop.f 999
+       nop.i 999 ;;
+}
+{ .mfi
+.restore sp
+       add   sp = 64,sp                       // Restore stack pointer
+//
+//     Is |r| < 2**(-2)
+//
+(p0)   fcmp.lt.unc.s1  p6, p0 = r, TWO_TO_NEG2 
+mov    b0 = GR_SAVE_B0                        // Restore return address
+};;
+{ .mfi
+       mov   gp = GR_SAVE_GP                  // Restore gp
+(p6)   fcmp.gt.unc.s1  p6, p0 = r, NEGTWO_TO_NEG2 
+       mov   ar.pfs = GR_SAVE_PFS             // Restore gp
+};;
+{ .mbb
+       nop.m 999
+(p6)   br.cond.spnt L(TANL_SMALL_R)
+(p0)   br.cond.sptk L(TANL_NORMAL_R) ;;
+}
+
+.endp __libm_callout
+ASM_SIZE_DIRECTIVE(__libm_callout)
+
+.type __libm_pi_by_2_reduce#,@function
+.global __libm_pi_by_2_reduce#