.file "scalbf.s"// Copyright (c) 2000 - 2003, 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// 01/26/01 Scalb completely reworked and now standalone version// 05/20/02 Cleaned up namespace and sf0 syntax// 02/10/03 Reordered header: .section, .global, .proc, .align// 08/06/03 Improved performance//// API//==============================================================// float = scalbf  (float x, float n)// input  floating point f8 and floating point f9// output floating point f8//// int_type = 0 if int is 32 bits// int_type = 1 if int is 64 bits//// Returns x* 2**n using an fma and detects overflow// and underflow.////// Strategy://  Compute biased exponent of result exp_Result = N + exp_X//  Break into ranges://   exp_Result > 0x1007e                 -> Certain overflow//   exp_Result = 0x1007e                 -> Possible overflow//   0x0ff81 <= exp_Result < 0x1007e      -> No over/underflow (main path)//   0x0ff81 - 23 <= exp_Result < 0x0ff81 -> Possible underflow//   exp_Result < 0x0ff81 - 23            -> Certain underflowFR_Big         = f6FR_NBig        = f7FR_Floating_X  = f8FR_Result      = f8FR_Floating_N  = f9FR_Result2     = f9FR_Result3     = f10FR_Norm_X      = f11FR_Two_N       = f12FR_N_float_int = f13FR_Norm_N      = f14GR_neg_ov_limit= r14GR_big_exp     = r14GR_N_Biased    = r15GR_Big         = r16GR_exp_Result  = r18GR_pos_ov_limit= r19GR_exp_sure_ou = r19GR_Bias        = r20GR_N_as_int    = r21GR_signexp_X   = r22GR_exp_X       = r23GR_exp_mask    = r24GR_max_exp     = r25GR_min_exp     = r26GR_min_den_exp = r27GR_Scratch     = r28GR_signexp_N   = r29GR_exp_N       = r30GR_SAVE_B0          = r32GR_SAVE_GP          = r33GR_SAVE_PFS         = r34GR_Parameter_X      = r35GR_Parameter_Y      = r36GR_Parameter_RESULT = r37GR_Tag              = r38.section .textGLOBAL_IEEE754_ENTRY(scalbf)////   Is x NAN, INF, ZERO, +-?//   Build the exponent Bias//{    .mfi     getf.exp      GR_signexp_N = FR_Floating_N // Get signexp of n     fclass.m      p6,p0 = FR_Floating_X, 0xe7  // @snan | @qnan | @inf | @zero     mov           GR_Bias = 0x0ffff}{    .mfi     mov           GR_Big = 35000      // If N this big then certain overflow     fcvt.fx.trunc.s1   FR_N_float_int = FR_Floating_N // Get N in significand     nop.i         0};;{    .mfi     getf.exp      GR_signexp_X = FR_Floating_X // Get signexp of x     fclass.m      p7,p0 = FR_Floating_N, 0x0b  // Test for n=unorm     nop.i         0}////   Normalize n//{    .mfi     mov           GR_exp_mask = 0x1ffff     // Exponent mask     fnorm.s1      FR_Norm_N = FR_Floating_N     nop.i         0};;////   Is n NAN, INF, ZERO, +-?//{    .mfi     mov           GR_big_exp = 0x1003e      // Exponent at which n is integer     fclass.m      p9,p0 = FR_Floating_N, 0xe7  // @snan | @qnan | @inf | @zero     mov           GR_max_exp = 0x1007e      // Exponent of maximum float}////   Normalize x//{ .mfb     nop.m         0     fnorm.s1      FR_Norm_X = FR_Floating_X(p7) br.cond.spnt  SCALBF_N_UNORM             // Branch if n=unorm};;SCALBF_COMMON1:// Main path continues.  Also return here from u=unorm path.//   Handle special cases if x = Nan, Inf, Zero{ .mfb     nop.m         0     fcmp.lt.s1    p7,p0 = FR_Floating_N, f0  // Test N negative(p6) br.cond.spnt  SCALBF_NAN_INF_ZERO};;//   Handle special cases if n = Nan, Inf, Zero{    .mfi     getf.sig      GR_N_as_int = FR_N_float_int // Get n from significand     fclass.m      p8,p0 = FR_Floating_X, 0x0b // Test for x=unorm     mov           GR_exp_sure_ou = 0x1000e // Exp_N where x*2^N sure over/under}{    .mfb     mov           GR_min_exp = 0x0ff81      // Exponent of minimum float     fcvt.xf       FR_N_float_int = FR_N_float_int // Convert N to FP integer(p9) br.cond.spnt  SCALBF_NAN_INF_ZERO};;{    .mmi     and           GR_exp_N = GR_exp_mask, GR_signexp_N // Get exponent of N(p7) sub           GR_Big = r0, GR_Big          // Limit for N     nop.i         0};;{    .mib     cmp.lt        p9,p0 = GR_exp_N, GR_big_exp // N possible non-integer?     cmp.ge        p6,p0 = GR_exp_N, GR_exp_sure_ou // N certain over/under?(p8) br.cond.spnt  SCALBF_X_UNORM             // Branch if x=unorm};;SCALBF_COMMON2:// Main path continues.  Also return here from x=unorm path.//   Create biased exponent for 2**N{    .mmi(p6) mov           GR_N_as_int = GR_Big      // Limit N;;     add           GR_N_Biased = GR_Bias,GR_N_as_int     nop.i         0};;{    .mfi     setf.exp      FR_Two_N = GR_N_Biased               // Form 2**N(p9) fcmp.neq.unc.s1 p9,p0 = FR_Norm_N, FR_N_float_int  // Test if N an integer     and           GR_exp_X = GR_exp_mask, GR_signexp_X // Get exponent of X};;////   Compute biased result exponent//   Branch if N is not an integer//{    .mib     add           GR_exp_Result = GR_exp_X, GR_N_as_int     mov           GR_min_den_exp = 0x0ff81 - 23 // Exponent of min denorm float(p9) br.cond.spnt  SCALBF_N_NOT_INT};;////   Raise Denormal operand flag with compare//   Do final operation//{    .mfi     cmp.lt        p7,p6 = GR_exp_Result, GR_max_exp  // Test no overflow     fcmp.ge.s0    p0,p11 = FR_Floating_X,FR_Floating_N  // Dummy to set denorm     cmp.lt        p9,p0 = GR_exp_Result, GR_min_den_exp // Test sure underflow}{    .mfb     nop.m         0     fma.s.s0      FR_Result = FR_Two_N,FR_Norm_X,f0(p9) br.cond.spnt  SCALBF_UNDERFLOW           // Branch if certain underflow};;{    .mib(p6) cmp.gt.unc    p6,p8 = GR_exp_Result, GR_max_exp  // Test sure overflow(p7) cmp.ge.unc    p7,p9 = GR_exp_Result, GR_min_exp  // Test no over/underflow(p7) br.ret.sptk   b0                         // Return from main path};;{    .bbb(p6) br.cond.spnt  SCALBF_OVERFLOW            // Branch if certain overflow(p8) br.cond.spnt  SCALBF_POSSIBLE_OVERFLOW   // Branch if possible overflow(p9) br.cond.spnt  SCALBF_POSSIBLE_UNDERFLOW  // Branch if possible underflow};;// Here if possible underflow.// Resulting exponent: 0x0ff81-23 <= exp_Result < 0x0ff81SCALBF_POSSIBLE_UNDERFLOW://// Here if possible overflow.// Resulting exponent: 0x1007e = exp_ResultSCALBF_POSSIBLE_OVERFLOW://   Set up necessary status fields////   S0 user supplied status//   S2 user supplied status + WRE + TD  (Overflows)//   S3 user supplied status + FZ + TD   (Underflows)//{    .mfi     mov           GR_pos_ov_limit = 0x1007f // Exponent for positive overflow     fsetc.s3      0x7F,0x41     nop.i         0}{    .mfi     mov           GR_neg_ov_limit = 0x3007f // Exponent for negative overflow     fsetc.s2      0x7F,0x42     nop.i         0};;////   Do final operation with s2 and s3//{    .mfi     setf.exp      FR_NBig = GR_neg_ov_limit     fma.s.s3      FR_Result3 = FR_Two_N,FR_Norm_X,f0     nop.i         0}{    .mfi     setf.exp      FR_Big = GR_pos_ov_limit     fma.s.s2      FR_Result2 = FR_Two_N,FR_Norm_X,f0     nop.i         0};;//   Check for overflow or underflow.//   Restore s3//   Restore s2//{    .mfi     nop.m         0     fsetc.s3      0x7F,0x40     nop.i         0}{    .mfi     nop.m         0     fsetc.s2      0x7F,0x40     nop.i         0};;////   Is the result zero?//{    .mfi     nop.m         0     fclass.m      p6, p0 =  FR_Result3, 0x007     nop.i         0}{    .mfi     nop.m         0     fcmp.ge.s1    p7, p8 = FR_Result2 , FR_Big     nop.i         0};;////   Detect masked underflow - Tiny + Inexact Only//{    .mfi     nop.m         0(p6) fcmp.neq.unc.s1 p6, p0 = FR_Result , FR_Result2     nop.i         0};;////   Is result bigger the allowed range?//   Branch out for underflow//{    .mfb     nop.m          0(p8) fcmp.le.unc.s1 p9, p10 = FR_Result2 , FR_NBig(p6) br.cond.spnt   SCALBF_UNDERFLOW};;////   Branch out for overflow//{ .bbb(p7) br.cond.spnt   SCALBF_OVERFLOW(p9) br.cond.spnt   SCALBF_OVERFLOW     br.ret.sptk    b0             //   Return from main path.};;// Here if result overflowsSCALBF_OVERFLOW:{ .mib     alloc         r32=ar.pfs,3,0,4,0     addl          GR_Tag = 55, r0     // Set error tag for overflow     br.cond.sptk  __libm_error_region // Call error support for overflow};;// Here if result underflowsSCALBF_UNDERFLOW:{ .mib     alloc         r32=ar.pfs,3,0,4,0     addl          GR_Tag = 56, r0     // Set error tag for underflow     br.cond.sptk  __libm_error_region // Call error support for underflow};;SCALBF_NAN_INF_ZERO:////   Before entry, N has been converted to a fp integer in significand of //     FR_N_float_int////   Convert  N_float_int to floating point value//{    .mfi     getf.sig     GR_N_as_int = FR_N_float_int     fclass.m     p6,p0 = FR_Floating_N, 0xc3 //@snan | @qnan     nop.i        0}{    .mfi     addl         GR_Scratch = 1,r0     fcvt.xf      FR_N_float_int = FR_N_float_int     nop.i        0};;{    .mfi     nop.m        0     fclass.m     p7,p0 = FR_Floating_X, 0xc3 //@snan | @qnan     shl          GR_Scratch = GR_Scratch,63};;{    .mfi     nop.m        0     fclass.m     p8,p0 = FR_Floating_N, 0x21 // @inf     nop.i        0}{    .mfi     nop.m        0     fclass.m     p9,p0 = FR_Floating_N, 0x22 // @-inf     nop.i        0};;////   Either X or N is a Nan, return result and possible raise invalid.//{    .mfb     nop.m        0(p6) fma.s.s0     FR_Result = FR_Floating_N,FR_Floating_X,f0(p6) br.ret.spnt  b0};;{    .mfb     nop.m        0(p7) fma.s.s0     FR_Result = FR_Floating_N,FR_Floating_X,f0(p7) br.ret.spnt  b0};;////   If N + Inf do something special//   For N = -Inf, create Int//{    .mfb     nop.m        0(p8) fma.s.s0     FR_Result = FR_Floating_X, FR_Floating_N,f0(p8) br.ret.spnt  b0}{    .mfi     nop.m        0(p9) fnma.s.s0    FR_Floating_N = FR_Floating_N, f1, f0     nop.i        0};;////   If N==-Inf,return x/(-N)//{    .mfb     cmp.ne       p7,p0 = GR_N_as_int,GR_Scratch(p9) frcpa.s0     FR_Result,p0 = FR_Floating_X,FR_Floating_N(p9) br.ret.spnt  b0};;////   Is N an integer.//{    .mfi     nop.m        0(p7) fcmp.neq.unc.s1 p7,p0 = FR_Norm_N, FR_N_float_int     nop.i        0};;////   If N not an int, return NaN and raise invalid.//{    .mfb     nop.m        0(p7) frcpa.s0     FR_Result,p0 = f0,f0(p7) br.ret.spnt  b0};;////   Always return x in other path.//{    .mfb     nop.m        0     fma.s.s0     FR_Result = FR_Floating_X,f1,f0     br.ret.sptk  b0};;// Here if n not int// Return NaN and raise invalid.SCALBF_N_NOT_INT:{    .mfb     nop.m        0     frcpa.s0     FR_Result,p0 = f0,f0     br.ret.sptk  b0};;// Here if n=unormSCALBF_N_UNORM:{ .mfb     getf.exp      GR_signexp_N = FR_Norm_N // Get signexp of normalized n     fcvt.fx.trunc.s1   FR_N_float_int = FR_Norm_N // Get N in significand     br.cond.sptk  SCALBF_COMMON1            // Return to main path};;// Here if x=unormSCALBF_X_UNORM:{ .mib     getf.exp      GR_signexp_X = FR_Norm_X // Get signexp of normalized x     nop.i         0     br.cond.sptk  SCALBF_COMMON2            // Return to main path};;GLOBAL_IEEE754_END(scalbf)LOCAL_LIBM_ENTRY(__libm_error_region)//// Get stack address of N//.prologue{ .mfi    add   GR_Parameter_Y=-32,sp    nop.f 0.save   ar.pfs,GR_SAVE_PFS    mov  GR_SAVE_PFS=ar.pfs}//// Adjust sp//{ .mfi.fframe 64   add sp=-64,sp   nop.f 0   mov GR_SAVE_GP=gp};;////  Store N on stack in correct position//  Locate the address of x on stack//{ .mmi   stfs [GR_Parameter_Y] = FR_Norm_N,16   add GR_Parameter_X = 16,sp.save   b0, GR_SAVE_B0   mov GR_SAVE_B0=b0};;//// Store x on the stack.// Get address for result on stack.//.body{ .mib   stfs [GR_Parameter_X] = FR_Norm_X   add   GR_Parameter_RESULT = 0,GR_Parameter_Y   nop.b 0}{ .mib   stfs [GR_Parameter_Y] = FR_Result   add   GR_Parameter_Y = -16,GR_Parameter_Y   br.call.sptk b0=__libm_error_support#};;////  Get location of result on stack//{ .mmi   add   GR_Parameter_RESULT = 48,sp   nop.m 0   nop.i 0};;////  Get the new result//{ .mmi   ldfs  FR_Result = [GR_Parameter_RESULT].restore sp   add   sp = 64,sp   mov   b0 = GR_SAVE_B0};;////  Restore gp, ar.pfs and return//{ .mib   mov   gp = GR_SAVE_GP   mov   ar.pfs = GR_SAVE_PFS   br.ret.sptk     b0};;LOCAL_LIBM_END(__libm_error_region).type   __libm_error_support#,@function.global __libm_error_support#