e_exp2l.s

glibc 2.9,最新版的C语言库函数

       // GR_CONST1 = bias+15
       mov GR_CONST1 = 0xffff+15
       nop.f 0
       mov GR_CONST3 = 0x1ffff ;;
}

{.mfi
       // load start address for C_1...C_6 followed by T_table
       ld8 GR_ADDR = [ GR_ADDR0 ]
       nop.f 0
       // get sign of argument
       andcm GR_SGN = GR_EBIAS, GR_CONST3
}
{.mfi
       // GR_D_ADDR = pointer to D_table
       ld8 GR_D_ADDR = [ GR_D_ADDR0 ]
       nop.f 0
       // get argument exponent
       and GR_ARGEXP = GR_CONST3, GR_EBIAS ;;
}

{.mfi
       alloc GR_SREG = ar.pfs, 1, 4, 4, 0
       nop.f 0
       // p6 = 1 if sign = 1
       cmp.ne p6, p8 = GR_SGN, r0
}
{.mfi
       // p7 = 1 if exponent> = 15 (argument out of range)
       cmp.ge p7, p0 = GR_ARGEXP, GR_CONST1
       nop.f 0
       sub GR_EXPON = GR_CONST2, GR_ARGEXP ;;
}

{.mib
       // load C_3, C_4
       ldfpd FR_COEFF3, FR_COEFF4 = [ GR_ADDR ], 16
       // get first exponent+8 bits
       shr.u GR_LEADBITS = GR_SIGNIF, GR_EXPON
 (p12) br.cond.spnt SPECIAL_exp2l
}
{.mib
       mov GR_256 = 256
       // exponent- = 63
       sub GR_EM63 = GR_EBIAS, GR_63
  (p7) br.cond.spnt OUT_RANGE_exp2l ;;
}

{.mlx
       // load C_5, C_6
       ldfpd FR_COEFF5, FR_COEFF6 = [ GR_ADDR ], 16
       // GR_2P14 = 2^14
       movl GR_2P14 = 0x46800000 ;;
}

{.mfi
       // load C_1
       ldfe FR_COEFF1 = [ GR_ADDR ], 16
       fma.s0 f8 = f8, f1, f0
       // GR_BM63 = bias-63
       mov GR_BM63 = 0xffff-63 ;;
}

{.mlx
       setf.s FR_2P14 = GR_2P14
       // GR_UF_TEST = -2^14-62
       movl GR_UF_TEST = 0xc6807c00
}
{.mfi
       // load C_2
       ldfe FR_COEFF2 = [ GR_ADDR ], 16
       nop.f 0
       mov GR_255 = 255 ;;
}

{.mib
       // get 8-bit index
       and GR_INDEX = GR_255, GR_LEADBITS
       // get K = integer part
       shr.u GR_K = GR_LEADBITS, 8
       nop.b 0 ;;
}

{.mmi
       // if sign = 1 && f>0, set p7 = 1
  (p6) cmp.gt.unc p7, p0 = GR_INDEX, r0
       setf.s FR_UF_TEST = GR_UF_TEST
       shl GR_KF = GR_LEADBITS, GR_EXPON ;;
}

{.mfi
       // if sign = 1 && f>0, set f = 1-f
  (p7) sub GR_INDEX = GR_256, GR_INDEX
       nop.f 0
       // if sign = 1 && f>0, set K = K+1
  (p7) add GR_K = GR_K, r0, 1 ;;
}

{.mfi
       // FR_EXP63 = 2^{expon-63}
       setf.exp FR_EXP63 = GR_EM63
       nop.f 0
       nop.i 0 ;;
}

.pred.rel "mutex", p6, p8
{.mfi
       // if sign = 0, set scale factor exponent S = K+bias-63
  (p8) add GR_K = GR_K, GR_BM63
       nop.f 0
       // if sign = 1, set scale factor exponent S = -K+bias-63
  (p6) sub GR_K = GR_BM63, GR_K ;;
}

{.mmi
       // FR_KF0 = 2^{63-expon}*(K+f)
       setf.sig FR_KF0 = GR_KF
       nop.m 0
       // GR_EMIN = EMIN = 2-2^14
       mov GR_EMIN = 0x18cfff ;;
}

{.mfi
       // get T_table index
       shladd GR_IT = GR_INDEX, 3, GR_ADDR
       // p7 = 1 if x> = 2^10
       fcmp.ge.s1 p7, p12 = f8, FR_2P14
       // get D_table index
       shladd GR_ID = GR_INDEX, 2, GR_D_ADDR ;;
}

{.mfi
       // load T_table value
       ldf8 FR_T = [ GR_IT ]
       // p7 = 1 if x<-2^10-50
 (p12) fcmp.lt.s1 p7, p0 = f8, FR_UF_TEST
       // GR_EMIN1 = EMIN = 2-2^14
       shl GR_EMIN1 = GR_EMIN, 11 ;;
}

{.mmb
       // f50 = scale factor = 2^{K-63}
       setf.exp FR_2EXP = GR_K
       // load D_table value
       ldfs FR_D = [ GR_ID ]
  (p7) br.cond.spnt OUT_RANGE_exp2l ;;
}

{.mfi
       nop.m 0
       // get r = x-(K+f)
       fnma.s1 FR_R = FR_KF0, FR_EXP63, f8
       nop.i 0 ;;
}

{.mfi
       // FR_EMIN = EMIN
       setf.s FR_EMIN = GR_EMIN1
       // P34 = C_4*r+C_3
       fma.s1 FR_P34 = FR_COEFF4, FR_R, FR_COEFF3
       nop.i 0
}
{.mfi
       nop.m 0
       // P56 = C_6*r+C_5
       fma.s1 FR_P56 = FR_COEFF6, FR_R, FR_COEFF5
       nop.i 0 ;;
}

{.mfi
       nop.m 0
       //  r*r
       fma.s1 FR_R2 = FR_R, FR_R, f0
       nop.i 0
}
{.mfi
       nop.m 0
       // P12 = C_2*r+C_1
       fma.s1 FR_P12 = FR_COEFF2, FR_R, FR_COEFF1
       nop.i 0 ;;
}

{.mfi
       nop.m 0
       // T* = scaling factor
       fma.s1 FR_TS = FR_T, FR_2EXP, f0
       nop.i 0
}
{.mfi
       nop.m 0
       // P36 = P34+r2*P56
       fma.s1 FR_P36 = FR_P56, FR_R2, FR_P34
       nop.i 0 ;;
}

{.mfi
       nop.m 0
       // P02 = D+r*P12
       fma.s1 FR_P02 = FR_P12, FR_R, FR_D
       nop.i 0
}
{.mfi
       nop.m 0
       // GR_ID = r*r2
       fma.s1 FR_R3 = FR_R2, FR_R, f0
       nop.i 0 ;;
}

{.mfi
       nop.m 0
       // P06 = P02+r3*P36
       fma.s1 FR_P06 = FR_P36, FR_R3, FR_P02
       nop.i 0 ;;
}

{.mfi
       nop.m 0
       // underflow (x<EMIN) ?
       fcmp.lt.s0 p6, p8 = f8, FR_EMIN
       nop.i 0 ;;
}

{.mfb
       nop.m 0
       // result = T+T*P06
       fma.s0 f8 = FR_TS, FR_P06, FR_TS
       // return
  (p8) br.ret.sptk b0
}
{.mfb
  (p6) mov GR_Parameter_TAG = 160
       nop.f 0
  (p6) br.cond.sptk __libm_error_region ;;
}


SPECIAL_exp2l:

{.mfi
       nop.m 0
       // x = -Infinity ?
       fclass.m p6, p0 = f8, 0x22
       nop.i 0 ;;
}

{.mfi
       nop.m 0
       // x = +Infinity ?
       fclass.m p7, p0 = f8, 0x21
       nop.i 0 ;;
}

{.mfi
       nop.m 0
       // x = +/-Zero ?
       fclass.m p8, p0 = f8, 0x7
       nop.i 0
}
{.mfb
       nop.m 0
       // exp2l(-Infinity) = 0
  (p6) mov f8 = f0
  (p6) br.ret.spnt b0 ;;
}

{.mfb
       nop.m 0
       // exp2l(+Infinity) = +Infinity
       nop.f 0
  (p7) br.ret.spnt b0 ;;
}

{.mfb
       nop.m 0
       // exp2l(+/-0) = 1
  (p8) mov f8 = f1
  (p8) br.ret.spnt b0 ;;
}

{.mfb
       nop.m 0
       // Remaining cases: NaNs
       fma.s0 f8 = f8, f1, f0
       br.ret.sptk b0 ;;
}


OUT_RANGE_exp2l:


{.mfi
       // overflow: p8 = 1
  (p8) mov GR_EM63 = 0x1fffe
       // normalize input, to detect pseudo-zeroes
       fma.s0 f8 = f8, f1, f0
       nop.i 0 ;;
}

{.mfi
       nop.m 0
       // f8 = 0?
       fcmp.eq.s1 p7, p0 = f8, f0
       nop.i 0 ;;
}

{.mmb
  (p8) mov GR_Parameter_TAG = 159
  (p8) setf.exp FR_TS = GR_EM63
       nop.b 999 ;;
}

{.mfb
       nop.m 0
       // pseudo-zero
  (p7) mov f8 = f1
  (p7) br.ret.sptk b0 ;;
}

{.mfi
       nop.m 999
  (p8) fma.s0 f8 = FR_TS, FR_TS, f0
       nop.i 999
}
{.mii
       nop.m 0
       // underflow: p6 = 1
  (p6) mov GR_EM63 = 1
       nop.i 0 ;;
}

{.mmb
  (p6) mov GR_Parameter_TAG = 160
  (p6) setf.exp FR_TS = GR_EM63
       nop.b 999 ;;
}

{.mfb
       nop.m 999
  (p6) fma.s0 f8 = FR_TS, FR_TS, f0
       nop.b 0 ;;
}


GLOBAL_LIBM_END(exp2l)


LOCAL_LIBM_ENTRY(__libm_error_region)
.prologue
{.mfi
       add GR_Parameter_Y = -32, sp // Parameter 2 value
       nop.f 0
.save ar.pfs, GR_SAVE_PFS
       mov GR_SAVE_PFS = ar.pfs // Save ar.pfs
}
{.mfi
.fframe 64
       add sp = -64, sp // Create new stack
       nop.f 0
       mov GR_SAVE_GP = gp ;; // Save gp
}

{.mmi
       stfe [ GR_Parameter_Y ] = FR_Y, 16 // STORE Parameter 2 on stack
       add GR_Parameter_X = 16, sp // Parameter 1 address
.save b0, GR_SAVE_B0
       mov GR_SAVE_B0 = b0 ;; // Save b0
}

.body
{.mib
       stfe [ GR_Parameter_X ] = FR_X // STORE Parameter 1 on stack
       add GR_Parameter_RESULT = 0, GR_Parameter_Y // Parameter 3 address
       nop.b 0
}
{.mib
       stfe [ GR_Parameter_Y ] = FR_RESULT // STORE Parameter 3 on stack
       add GR_Parameter_Y = -16, GR_Parameter_Y
       br.call.sptk b0 = __libm_error_support# ;; // Call error handling function
}

{.mmi
       add GR_Parameter_RESULT = 48, sp
       nop.m 0
       nop.i 0 ;;
}

{.mmi
       ldfe f8 = [ GR_Parameter_RESULT ] // Get return result off stack
.restore sp
       add sp = 64, sp // Restore stack pointer
       mov b0 = GR_SAVE_B0 ;; // Restore return address
}

{.mib
       mov gp = GR_SAVE_GP // Restore gp
       mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
       br.ret.sptk b0 ;; // Return
}


LOCAL_LIBM_END(__libm_error_region)
.type __libm_error_support#, @function
.global __libm_error_support#