e_coshf.s

Glibc 2.3.2源代码(解压后有100多M)

      nop.i 999
}
;;

// Calculate coshf_FR_X2 = ax*ax and coshf_FR_X4 = ax*ax*ax*ax
{ .mmf
       nop.m 999
(p0)     ldfe       coshf_FR_P1 = [r34],16                 
(p0)     fma.s1     coshf_FR_X2 = coshf_FR_X, coshf_FR_X, f0 ;;           
}

{ .mmi
(p0)     ldfe       coshf_FR_P2 = [r34],16 ;;                 
(p0)     ldfe       coshf_FR_P3 = [r34],16                 
       nop.i 999 ;;
}

{ .mmi
(p0)     ldfe       coshf_FR_P4 = [r34],16 ;;                 
(p0)     ldfe       coshf_FR_P5 = [r34],16                 
       nop.i 999 ;;
}

{ .mfi
(p0)     ldfe       coshf_FR_P6 = [r34],16                 
(p0)     fma.s1     coshf_FR_X4 = coshf_FR_X2, coshf_FR_X2, f0         
       nop.i 999 ;;
}

// Calculate coshf_FR_podd = x4 *(x4 * P_5 + P_3) + P_1
{ .mfi
       nop.m 999
(p0)     fma.s1     coshf_FR_poly_podd_temp1 = coshf_FR_X4, coshf_FR_P5, coshf_FR_P3                
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fma.s1     coshf_FR_podd            = coshf_FR_X4, coshf_FR_poly_podd_temp1, coshf_FR_P1   
       nop.i 999
}

// Calculate coshf_FR_peven =  p_even = x4 *(x4 * (x4 * P_6 + P_4) + P_2)
{ .mfi
       nop.m 999
(p0)     fma.s1     coshf_FR_poly_peven_temp1 = coshf_FR_X4, coshf_FR_P6, coshf_FR_P4               
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fma.s1     coshf_FR_poly_peven_temp2 = coshf_FR_X4, coshf_FR_poly_peven_temp1, coshf_FR_P2 
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fma.s1     coshf_FR_peven       = coshf_FR_X4, coshf_FR_poly_peven_temp2, f0         
       nop.i 999 ;;
}

// Y_lo = x2*p_odd + p_even
// Calculate f8 = Y_hi + Y_lo 

{ .mfi
       nop.m 999
(p0)     fma.s1     coshf_FR_Y_lo         = coshf_FR_X2, coshf_FR_podd,  coshf_FR_peven    
       nop.i 999 ;;
}

{ .mfb
       nop.m 999
(p0)     fma.s.s0   f8                   = f1, f1, coshf_FR_Y_lo                        
(p0)     br.ret.sptk    b0 ;;                                                        
}


L(COSH_BY_TBL): 

// Now that we are at TBL; so far all we know is that |x| >= 0.25.
// The first two steps are the same for TBL and EXP, but if we are HUGE
// Double
// Go to HUGE if |x| >= 2^10, 10009 (register-biased) is e = 10 (true)
// Single
// Go to HUGE if |x| >= 2^7,  10006 (register-biased) is e =  7 (true)
// we want to leave now. Go to HUGE if |x| >= 2^14
// 1000d (register-biased) is e = 14 (true)

{ .mlx
       nop.m 999
(p0)     movl            r32 = 0x0000000000010006 ;;              
}

{ .mfi
(p0)     setf.exp        f9 = r32                              
       nop.f 999
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fcmp.ge.unc     p6,p7 = coshf_FR_X,f9                  
       nop.i 999 ;;
}

{ .mib
       nop.m 999
       nop.i 999
(p6)     br.cond.spnt    L(COSH_HUGE) ;;                             
}

// r32 = 1
// r34 = N-1 
// r35 = N
// r36 = j
// r37 = N+1

// TBL can never overflow
// coshf(x) = coshf(B+R)
//         = coshf(B) coshf(R) + sinh(B) sinh(R) 
// coshf(R) can be approximated by 1 + p_even
// sinh(R) can be approximated by p_odd

// ******************************************************
// STEP 1 (TBL and EXP)
// ******************************************************
// Get the following constants.
// f9  = Inv_log2by64
// f10 = log2by64_hi
// f11 = log2by64_lo

{ .mmi
(p0)     adds                 r32 = 0x1,r0      
(p0)     addl           r34   = @ltoff(single_coshf_arg_reduction), gp
         nop.i 999
}
;;


// We want 2^(N-1) and 2^(-N-1). So bias N-1 and -N-1 and
// put them in an exponent.
// coshf_FR_spos = 2^(N-1) and coshf_FR_sneg = 2^(-N-1)
// r39 = 0xffff + (N-1)  = 0xffff +N -1
// r40 = 0xffff - (N +1) = 0xffff -N -1

{ .mlx
         ld8 r34 = [r34]
(p0)     movl                r38 = 0x000000000000fffe ;; 
}

{ .mmi
(p0)     ldfe            coshf_FR_Inv_log2by64 = [r34],16 ;;            
(p0)     ldfe            coshf_FR_log2by64_hi  = [r34],16            
       nop.i 999 ;;
}

{ .mbb
(p0)     ldfe            coshf_FR_log2by64_lo  = [r34],16            
       nop.b 999
       nop.b 999 ;;
}

// Get the A coefficients
// f9  = A_1
// f10 = A_2
// f11 = A_3

{ .mmi
      nop.m 999
(p0)  addl           r34   = @ltoff(single_coshf_ab_table), gp
      nop.i 999
}
;;

{ .mmi
      ld8 r34 = [r34]
      nop.m 999
      nop.i 999
}
;;


// Calculate M and keep it as integer and floating point.
// M = round-to-integer(x*Inv_log2by64)
// coshf_FR_M = M = truncate(ax/(log2/64))
// Put the significand of M in r35
//    and the floating point representation of M in coshf_FR_M

{ .mfi
       nop.m 999
(p0)     fma.s1          coshf_FR_M      = coshf_FR_X, coshf_FR_Inv_log2by64, f0 
       nop.i 999
}

{ .mfi
(p0)     ldfe            coshf_FR_A1 = [r34],16            
       nop.f 999
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fcvt.fx.s1      coshf_FR_M_temp = coshf_FR_M                      
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fnorm.s1        coshf_FR_M      = coshf_FR_M_temp                 
       nop.i 999 ;;
}

{ .mfi
(p0)     getf.sig        r35       = coshf_FR_M_temp                 
       nop.f 999
       nop.i 999 ;;
}

// M is still in r35. Calculate j. j is the signed extension of the six lsb of M. It
// has a range of -32 thru 31.
// r35 = M
// r36 = j 

{ .mii
       nop.m 999
       nop.i 999 ;;
(p0)     and            r36 = 0x3f, r35 ;;   
}

// Calculate R
// f13 = f44 - f12*f10 = x - M*log2by64_hi
// f14 = f13 - f8*f11 = R = (x - M*log2by64_hi) - M*log2by64_lo

{ .mfi
       nop.m 999
(p0)     fnma.s1        coshf_FR_R_temp = coshf_FR_M, coshf_FR_log2by64_hi, coshf_FR_X      
       nop.i 999
}

{ .mfi
(p0)     ldfe            coshf_FR_A2 = [r34],16            
       nop.f 999
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fnma.s1        coshf_FR_R      = coshf_FR_M, coshf_FR_log2by64_lo, coshf_FR_R_temp 
       nop.i 999
}

// Get the B coefficients
// f15 = B_1
// f32 = B_2
// f33 = B_3

{ .mmi
(p0)     ldfe            coshf_FR_A3 = [r34],16 ;;            
(p0)     ldfe            coshf_FR_B1 = [r34],16            
       nop.i 999 ;;
}

{ .mmi
(p0)     ldfe            coshf_FR_B2 = [r34],16 ;;            
(p0)     ldfe            coshf_FR_B3 = [r34],16            
       nop.i 999 ;;
}

{ .mii
       nop.m 999
(p0)     shl            r34 = r36,  0x2 ;;   
(p0)     sxt1           r37 = r34 ;;         
}

// ******************************************************
// STEP 2 (TBL and EXP)
// ******************************************************
// Calculate Rsquared and Rcubed in preparation for p_even and p_odd
// f12 = R*R*R
// f13 = R*R
// f14 = R <== from above

{ .mfi
       nop.m 999
(p0)     fma.s1          coshf_FR_Rsq  = coshf_FR_R,   coshf_FR_R, f0  
(p0)     shr            r36 = r37,  0x2 ;;   
}

// r34 = M-j = r35 - r36
// r35 = N = (M-j)/64

{ .mii
(p0)     sub                  r34 = r35, r36    
       nop.i 999 ;;
(p0)     shr                  r35 = r34, 0x6 ;;    
}

{ .mii
(p0)     sub                 r40 = r38, r35           
(p0)     adds                 r37 = 0x1, r35    
(p0)     add                 r39 = r38, r35 ;;           
}

// Get the address of the J table, add the offset,
// addresses are sinh_AD_mJ and sinh_AD_J, get the T value
// f32 = T(j)_hi
// f33 = T(j)_lo
// f34 = T(-j)_hi
// f35 = T(-j)_lo

{ .mmi
(p0)     sub                  r34 = r35, r32    
(p0)     addl    r37   = @ltoff(single_coshf_j_table), gp
         nop.i 999
}
;;

{ .mfi
      ld8 r37 = [r37]
(p0)     fma.s1          coshf_FR_Rcub = coshf_FR_Rsq, coshf_FR_R, f0  
       nop.i 999
}

// ******************************************************
// STEP 3 Now decide if we need to branch to EXP
// ******************************************************
// Put 32 in f9; p6 true if x < 32

{ .mlx
       nop.m 999
(p0)     movl                r32 = 0x0000000000010004 ;;               
}

// Calculate p_even
// f34 = B_2 + Rsq *B_3
// f35 = B_1 + Rsq*f34      = B_1 + Rsq * (B_2 + Rsq *B_3)
// f36 = peven = Rsq * f35 = Rsq * (B_1 + Rsq * (B_2 + Rsq *B_3))

{ .mfi
       nop.m 999
(p0)     fma.s1          coshf_FR_peven_temp1 = coshf_FR_Rsq, coshf_FR_B3,          coshf_FR_B2  
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fma.s1          coshf_FR_peven_temp2 = coshf_FR_Rsq, coshf_FR_peven_temp1, coshf_FR_B1  
       nop.i 999
}

// Calculate p_odd
// f34 = A_2 + Rsq *A_3
// f35 = A_1 + Rsq * (A_2 + Rsq *A_3)
// f37 = podd = R + Rcub * (A_1 + Rsq * (A_2 + Rsq *A_3))

{ .mfi
       nop.m 999
(p0)     fma.s1          coshf_FR_podd_temp1 = coshf_FR_Rsq,        coshf_FR_A3,         coshf_FR_A2  
       nop.i 999 ;;
}

{ .mfi
(p0)     setf.exp            coshf_FR_N_temp1 = r39            
       nop.f 999
       nop.i 999 ;;
}

{ .mfi
       nop.m 999
(p0)     fma.s1          coshf_FR_peven       = coshf_FR_Rsq, coshf_FR_peven_temp2, f0     
       nop.i 999
}

{ .mfi
       nop.m 999
(p0)     fma.s1          coshf_FR_podd_temp2 = coshf_FR_Rsq,        coshf_FR_podd_temp1, coshf_FR_A1  
       nop.i 999 ;;
}

{ .mfi
(p0)     setf.exp            f9  = r32                              
       nop.f 999
       nop.i 999 ;;