s_cosf.s

来自「Glibc 2.3.2源代码(解压后有100多M)」· S 代码 · 共 700 行 · 第 1/2 页

}
{ .mfi
     addl           SIN_AD_PQ_1              =    @ltoff(sin_coeff_1_table),gp
     fnorm.s0 SIN_NORM_f8 = f8        // Sets denormal or invalid
     mov sin_GR_sincos_flag = 0x1
}
;;

{ .mfi
     ld8       SIN_AD_PQ_1    =    [SIN_AD_PQ_1]
     fclass.m.unc  p8,p0      =    f8, 0x07
     cmp.ne p9,p0 = r0,r0
}
{ .mfb
     nop.m 999
     nop.f 999
     br.sptk L(SINCOSF_COMMON)
}
;;


L(SINCOSF_COMMON):

// Here with p6 if sin, p7 if cos, p8 if sin(0), p9 if cos(0)


{ .mmf
     ldfpd      sin_coeff_Q3, sin_coeff_Q4     = [SIN_AD_PQ_1], 16
     nop.m 999
     fclass.m.unc  p11,p0      =    f8, 0x23	// Test for x=inf
}
;;

{ .mfb
     ldfpd      sin_coeff_Q5, sin_coeff_Q6     = [SIN_AD_PQ_1], 16
     fclass.m.unc  p10,p0      =    f8, 0xc3	// Test for x=nan
(p8) br.ret.spnt b0                   // Exit for sin(0)
}
{ .mfb
     nop.m 999
(p9) fma.s      f8 = f1,f1,f0
(p9) br.ret.spnt b0                   // Exit for cos(0)
}
;;

{ .mmf
     ldfpd      sin_coeff_P4, sin_coeff_P5     = [SIN_AD_PQ_1], 16
     addl gr_tmp = -1,r0
     fcvt.xf    sin_Mfloat                     =    sin_Mx
}
;;

{     .mfi
     getf.sig  sin_GR_Mint    =    sin_Mx
(p11) frcpa.s0      f8,p13      =    f0,f0  // qnan indef if x=inf
     nop.i 999
}
{     .mfb
     ldfpd      sin_coeff_P1, sin_coeff_P2     = [SIN_AD_PQ_1], 16
     nop.f 999
(p11) br.ret.spnt b0                   // Exit for x=inf
}
;;

{     .mfi
     ldfpd      sin_coeff_Q1, sin_coeff_Q2     = [SIN_AD_PQ_1], 16
     nop.f                      999
     cmp.ge    p8,p9          = -33,sin_GR_Mint
}
{     .mfb
     add       sin_GR_index   =    32,sin_GR_Mint
(p10) fma.s      f8 = f8,f1,f0         // Force qnan if x=nan
(p10) br.ret.spnt b0                   // Exit for x=nan
}
;;

{ .mmi
     ldfd      sin_coeff_P3   = [SIN_AD_PQ_1], 16
(p9) cmp.le    p8,p0        = 33, sin_GR_Mint 
     shl       sin_GR_index   =    sin_GR_index,4
}
;;


{     .mfi
     setf.sig fp_tmp = gr_tmp  // Create constant such that fmpy sets inexact
     fnma.s1   sin_r     =    f1,sin_Mfloat,SIN_NORM_f8
(p8) cmp.eq.unc p11,p12=sin_GR_sincos_flag,r0  // p11 if must call dbl cos
                                               // p12 if must call dbl sin
}
{    .mbb
     add       SIN_AD_PQ_2    =    sin_GR_index,SIN_AD_PQ_1
(p11) br.cond.spnt COS_DOUBLE
(p12) br.cond.spnt SIN_DOUBLE
}
;;

.pred.rel "mutex",p6,p7    //SIN_Sin_Flag, SIN_Cos_Flag
{     .mmi
(p6) ldfpd     sin_tbl_S,sin_tbl_C =    [SIN_AD_PQ_2]
(p7) ldfpd     sin_tbl_C,sin_tbl_S =    [SIN_AD_PQ_2]
               nop.i                           999
}
;;

{     .mfi
     nop.m                 999
(p6) fclass.m.unc p8,p0 = f8, 0x0b // If sin, note denormal input to set uflow
     nop.i                 999
}
{     .mfi
     nop.m                 999
     fma.s1    sin_t     =    sin_r,sin_r,f0
     nop.i                 999
}
;;

{     .mfi
     nop.m                 999
     fma.s1    sin_rcube =    sin_t,sin_r,f0
     nop.i                 999
}
{     .mfi
     nop.m                 999
     fma.s1    sin_tsq   =    sin_t,sin_t,f0
     nop.i                 999
}
;;

{     .mfi
     nop.m                      999
     fma.s1    sin_poly_q3    =    sin_t,sin_coeff_Q4,sin_coeff_Q3
     nop.i                      999
}
{     .mfi
     nop.m                      999
     fma.s1    sin_poly_q5    =    sin_t,sin_coeff_Q6,sin_coeff_Q5
     nop.i                      999
}
;;

{     .mfi
     nop.m                      999
     fma.s1    sin_poly_p1    =    sin_t,sin_coeff_P5,sin_coeff_P4
     nop.i                      999
}
{     .mfi
     nop.m                      999
     fma.s1    sin_poly_p2    =    sin_t,sin_coeff_P2,sin_coeff_P1
     nop.i                      999
}
;;

{     .mfi
     nop.m                      999
     fma.s1    sin_poly_q1    =    sin_t,sin_coeff_Q2,sin_coeff_Q1
     nop.i                      999
}
{     .mfi
     nop.m                      999
     fma.s1    sin_S_t   =    sin_t,sin_tbl_S,f0
     nop.i                      999
}
;;

{     .mfi
     nop.m                 999
(p8) fmpy.s.s0 fp_tmp2 = f8,f8  // Dummy mult to set underflow if sin(denormal)
     nop.i                 999
}
{     .mfi
     nop.m                 999
     fma.s1    sin_r7    =    sin_rcube,sin_tsq,f0
     nop.i                 999
}
;;

{     .mfi
     nop.m                      999
     fma.s1    sin_poly_q3456 =    sin_tsq,sin_poly_q5,sin_poly_q3
     nop.i                      999
}
;;

{     .mfi
     nop.m                      999
     fma.s1    sin_poly_p3    =    sin_t,sin_poly_p1,sin_coeff_P3
     nop.i                      999
}
{     .mfi
     nop.m                      999
     fma.s1    sin_poly_p4    =    sin_rcube,sin_poly_p2,sin_r
     nop.i                      999
}
;;

{     .mfi
     nop.m                           999
     fma.s1    sin_tbl_S_tcube     =    sin_S_t,sin_tsq,f0
     nop.i                           999
}
{     .mfi
     nop.m                      999
     fma.s1    sin_poly_q12   =    sin_S_t,sin_poly_q1,sin_tbl_S
     nop.i                      999
}
;;

{     .mfi
     nop.m                 999
     fma.d.s1  sin_of_r  =    sin_r7,sin_poly_p3,sin_poly_p4
     nop.i                 999
}
;;

{     .mfi
     nop.m                           999
     fma.d.s1  sin_tbl_S_cos_of_r  =    sin_tbl_S_tcube,sin_poly_q3456,sin_poly_q12
     nop.i                           999
}
{     .mfi
     nop.m                           999
     fmpy.s0   fp_tmp = fp_tmp, fp_tmp  // Dummy mult to set inexact
     nop.i                           999
}
;;


.pred.rel "mutex",p6,p7    //SIN_Sin_Flag, SIN_Cos_Flag
{     .mfi
               nop.m            999
//(SIN_Sin_Flag) fma.s     f8   =    sin_tbl_C,sin_of_r,sin_tbl_S_cos_of_r
(p6) fma.s     f8   =    sin_tbl_C,sin_of_r,sin_tbl_S_cos_of_r
               nop.i            999
}
{     .mfb
               nop.m            999
//(SIN_Cos_Flag) fnma.s    f8   =    sin_tbl_C,sin_of_r,sin_tbl_S_cos_of_r
(p7) fnma.s    f8   =    sin_tbl_C,sin_of_r,sin_tbl_S_cos_of_r
               br.ret.sptk     b0
}

.endp sinf
ASM_SIZE_DIRECTIVE(sinf)


.proc SIN_DOUBLE 
SIN_DOUBLE:
.prologue
{ .mfi
        nop.m 0
        nop.f 0
.save   ar.pfs,GR_SAVE_PFS
        mov  GR_SAVE_PFS=ar.pfs
}
;;

{ .mfi
        mov GR_SAVE_GP=gp
        nop.f 0
.save   b0, GR_SAVE_B0
        mov GR_SAVE_B0=b0
}

.body
{ .mmb
       nop.m 999
       nop.m 999
       br.call.sptk.many   b0=sin 
}
;;

{ .mfi
       mov gp        = GR_SAVE_GP
       nop.f 999
       mov b0        = GR_SAVE_B0
}
;;

{ .mfi
      nop.m 999
      fma.s f8 = f8,f1,f0
(p0)  mov ar.pfs    = GR_SAVE_PFS
}
{ .mib
      nop.m 999
      nop.i 999
(p0)  br.ret.sptk     b0 
}
;;

.endp  SIN_DOUBLE
ASM_SIZE_DIRECTIVE(SIN_DOUBLE)


.proc COS_DOUBLE 
COS_DOUBLE:
.prologue
{ .mfi
        nop.m 0
        nop.f 0
.save   ar.pfs,GR_SAVE_PFS
        mov  GR_SAVE_PFS=ar.pfs
}
;;

{ .mfi
        mov GR_SAVE_GP=gp
        nop.f 0
.save   b0, GR_SAVE_B0
        mov GR_SAVE_B0=b0
}

.body
{ .mmb
       nop.m 999
       nop.m 999
       br.call.sptk.many   b0=cos 
}
;;

{ .mfi
       mov gp        = GR_SAVE_GP
       nop.f 999
       mov b0        = GR_SAVE_B0
}
;;

{ .mfi
      nop.m 999
      fma.s f8 = f8,f1,f0
(p0)  mov ar.pfs    = GR_SAVE_PFS
}
{ .mib
      nop.m 999
      nop.i 999
(p0)  br.ret.sptk     b0 
}
;;

.endp  COS_DOUBLE
ASM_SIZE_DIRECTIVE(COS_DOUBLE)



.type sin,@function
.global sin 
.type cos,@function
.global cos