s_log1pf.s

Glibc 2.3.2源代码(解压后有100多M)

//    Load Q2 
//    Load Q1 
//
(p8) fma.s1 FR_r = FR_G, FR_S_lo, FR_r 
	nop.i 999
}

{ .mfi
	nop.m 999
//
//    poly_lo = r * Q4 + Q3
//    rsq = r* r
//
(p0)  fma.s1 FR_h = FR_float_N, FR_log2_lo, FR_h 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
//
//    If (S_lo!=0) r = s_lo * G + r
//
(p0)  fma.s1 FR_poly_lo = FR_r, FR_Q4, FR_Q3 
	nop.i 999
}
//
//    Create a 0x00000....01
//    poly_lo = poly_lo * rsq + h
//

{ .mfi
(p0)  setf.sig FR_dummy = GR_Perturb 
(p0)  fmpy.s1 FR_rsq = FR_r, FR_r 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
//
//    h = N * log2_lo + h 
//    Y_hi = n * log2_hi + H 
//
(p0)  fma.s1 FR_poly_lo = FR_poly_lo, FR_r, FR_Q2 
	nop.i 999
}

{ .mfi
	nop.m 999
(p0)  fma.s1 FR_poly_hi = FR_Q1, FR_rsq, FR_r 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
//
//    poly_lo = r * poly_o + Q2 
//    poly_hi = Q1 * rsq + r 
//
(p0)  fmpy.s1 FR_poly_lo = FR_poly_lo, FR_r 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p0)  fma.s1 FR_poly_lo = FR_poly_lo, FR_rsq, FR_h 
	nop.i 999 ;;
}

{ .mfb
	nop.m 999
(p0)  fadd.s1 FR_Y_lo = FR_poly_hi, FR_poly_lo 
//
//    Create the FR for a binary "or"
//    Y_lo = poly_hi + poly_lo
//
// (p0)  for FR_dummy = FR_Y_lo,FR_dummy ;;
//
//    Turn the lsb of Y_lo ON
//
// (p0)  fmerge.se FR_Y_lo =  FR_Y_lo,FR_dummy ;;
//
//    Merge the new lsb into Y_lo, for alone doesn't
//
(p0)  br.cond.sptk L(LOG_main) ;; 
}


L(log1pf_near): 

{ .mmi
	nop.m 999
	nop.m 999
//    /*******************************************************/
//    /*********** Branch log1pf_near  ************************/
//    /*******************************************************/
(p0)  addl GR_Table_Base = @ltoff(Constants_P#),gp ;; 
}
//
//    Load base address of poly. coeff.
//
{.mmi
      nop.m 999
      ld8    GR_Table_Base = [GR_Table_Base]
      nop.i 999
};;

{ .mmb
(p0)  add GR_Table_ptr = 0x40,GR_Table_Base  
//
//    Address tables with separate pointers 
//
(p0)  ldfe FR_P8 = [GR_Table_Base],16 
	nop.b 999 ;;
}

{ .mmb
(p0)  ldfe FR_P4 = [GR_Table_ptr],16 
//
//    Load P4
//    Load P8
//
(p0)  ldfe FR_P7 = [GR_Table_Base],16 
	nop.b 999 ;;
}

{ .mmf
(p0)  ldfe FR_P3 = [GR_Table_ptr],16 
//
//    Load P3
//    Load P7
//
(p0)  ldfe FR_P6 = [GR_Table_Base],16 
(p0)  fmpy.s1 FR_wsq = FR_W, FR_W ;; 
}

{ .mfi
(p0)  ldfe FR_P2 = [GR_Table_ptr],16 
	nop.f 999
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p0)  fma.s1 FR_Y_hi = FR_W, FR_P4, FR_P3 
	nop.i 999
}
//
//    Load P2
//    Load P6
//    Wsq = w * w
//    Y_hi = p4 * w + p3
//

{ .mfi
(p0)  ldfe FR_P5 = [GR_Table_Base],16 
(p0)  fma.s1 FR_Y_lo = FR_W, FR_P8, FR_P7 
	nop.i 999 ;;
}

{ .mfi
(p0)  ldfe FR_P1 = [GR_Table_ptr],16 
//
//    Load P1
//    Load P5
//    Y_lo = p8 * w + P7
//
(p0)  fmpy.s1 FR_w4 = FR_wsq, FR_wsq 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p0)  fma.s1 FR_Y_hi = FR_W, FR_Y_hi, FR_P2 
	nop.i 999
}

{ .mfi
	nop.m 999
(p0)  fma.s1 FR_Y_lo = FR_W, FR_Y_lo, FR_P6 
(p0)  add GR_Perturb = 0x1, r0 ;; 
}

{ .mfi
	nop.m 999
//
//    w4 = w2 * w2 
//    Y_hi = y_hi * w + p2 
//    Y_lo = y_lo * w + p6 
//    Create perturbation bit
//
(p0)  fmpy.s1 FR_w6 = FR_w4, FR_wsq 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p0)  fma.s1 FR_Y_hi = FR_W, FR_Y_hi, FR_P1 
	nop.i 999
}
//
//    Y_hi = y_hi * w + p1 
//    w6 = w4 * w2 
//

{ .mfi
(p0)  setf.sig FR_Q4 = GR_Perturb 
(p0)  fma.s1 FR_Y_lo = FR_W, FR_Y_lo, FR_P5 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p0)  fma.s1 FR_Y_hi = FR_wsq,FR_Y_hi, FR_W 
	nop.i 999
}

{ .mfb
	nop.m 999
//
//    Y_hi = y_hi * wsq + w 
//    Y_lo = y_lo * w + p5 
//
(p0)  fmpy.s1 FR_Y_lo = FR_w6, FR_Y_lo 
//
//    Y_lo = y_lo * w6  
//
// (p0)  for FR_dummy = FR_Y_lo,FR_dummy ;;
//
//    Set lsb on: Taken out to improve performance 
//
// (p0)  fmerge.se FR_Y_lo =  FR_Y_lo,FR_dummy ;;
//
//    Make sure it's on in Y_lo also.  Taken out to improve
//    performance
//
(p0)  br.cond.sptk L(LOG_main) ;; 
}


L(log1pf_small): 

{ .mmi
	nop.m 999
	nop.m 999
//  /*******************************************************/
//  /*********** Branch log1pf_small  ***********************/
//  /*******************************************************/
(p0)  addl GR_Table_Base = @ltoff(Constants_Threshold#),gp 
}

{ .mfi
	nop.m 999
(p0)  mov FR_Em1 = FR_W 
(p0)  cmp.eq.unc  p7, p0 = r0, r0 ;; 
}

{ .mlx
      ld8    GR_Table_Base = [GR_Table_Base]
(p0)  movl GR_Expo_Range = 0x0000000000000002 ;; 
}
//
//    Set Safe to true
//    Set Expo_Range = 0 for single
//    Set Expo_Range = 2 for double 
//    Set Expo_Range = 4 for double-extended 
//

{ .mmi
(p0)  shladd GR_Table_Base = GR_Expo_Range,4,GR_Table_Base ;; 
(p0)  ldfe FR_Threshold = [GR_Table_Base],16 
	nop.i 999
}

{ .mlx
	nop.m 999
(p0)  movl GR_Bias = 0x000000000000FF9B ;; 
}

{ .mfi
(p0)  ldfe FR_Tiny = [GR_Table_Base],0 
	nop.f 999
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p0)  fcmp.gt.unc.s1 p13, p12 =  FR_abs_W, FR_Threshold 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p13) fnmpy.s1 FR_Y_lo = FR_W, FR_W 
	nop.i 999
}

{ .mfi
	nop.m 999
(p13) fadd FR_SCALE = f0, f1 
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p12) fsub.s1 FR_Y_lo = f0, FR_Tiny 
(p12) cmp.ne.unc  p7, p0 = r0, r0 
}

{ .mfi
(p12) setf.exp FR_SCALE = GR_Bias 
	nop.f 999
	nop.i 999 ;;
}

//
//    Set p7 to SAFE = FALSE
//    Set Scale = 2^-100 
//
{ .mfb
	nop.m 999
(p0)  fma.s.s0 FR_Input_X = FR_Y_lo,FR_SCALE,FR_Y_hi
(p0)  br.ret.sptk   b0
}
;;

L(LOG_64_one): 

{ .mfb
	nop.m 999
(p0)  fmpy.s.s0 FR_Input_X = FR_Input_X, f0 
(p0)  br.ret.sptk   b0
}
;;
//    
//    Raise divide by zero for +/-0 input.
//    

L(LOG_64_zero): 

{ .mfi
(p0)  mov   GR_Parameter_TAG = 142 
//
//    If we have log1pf(0), return -Inf.
//  
(p0)  fsub.s0 FR_Output_X_tmp = f0, f1 
      nop.i 999 ;;
}
{ .mfb
      nop.m 999
(p0)  frcpa.s0 FR_Output_X_tmp, p8 =  FR_Output_X_tmp, f0 
(p0)  br.cond.sptk L(LOG_ERROR_Support) ;; 
}

L(LOG_64_special): 

{ .mfi
      nop.m 999
//    
//    Return -Inf or value from handler.
//    
(p0)  fclass.m.unc p7, p0 =  FR_Input_X, 0x1E1 
      nop.i 999 ;;
}

{ .mfb
      nop.m 999
//     
//    Check for Natval, QNan, SNaN, +Inf   
//    
(p7)  fmpy.s.s0  f8 =  FR_Input_X, f1 
//     
//    For SNaN raise invalid and return QNaN.
//    For QNaN raise invalid and return QNaN.
//    For +Inf return +Inf.
//    
(p7)  br.ret.sptk   b0
}
;;

//    
//    For -Inf raise invalid and return QNaN.
//    

{ .mfb
(p0)  mov   GR_Parameter_TAG = 143 
(p0)  fmpy.s.s0  FR_Output_X_tmp =  FR_Input_X, f0 
(p0)  br.cond.sptk L(LOG_ERROR_Support) ;; 
}

//
//    Report that log1pf(-Inf) computed
//     

L(LOG_64_unsupported): 

//    
//    Return generated NaN or other value .
//    

{ .mfb
      nop.m 999
(p0)  fmpy.s.s0 FR_Input_X = FR_Input_X, f0 
(p0)  br.ret.sptk   b0 ;;
}

L(LOG_64_negative): 

{ .mfi
      nop.m 999
//     
//    Deal with x < 0 in a special way 
//    
(p0)  frcpa.s0 FR_Output_X_tmp, p8 =  f0, f0 
//     
//    Deal with x < 0 in a special way - raise
//    invalid and produce QNaN indefinite.
//    
(p0)  mov   GR_Parameter_TAG = 143;;
}

.endp log1pf#
ASM_SIZE_DIRECTIVE(log1pf)

.proc __libm_error_region
__libm_error_region:
L(LOG_ERROR_Support): 
.prologue

// (1)
{ .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
};;


// (2)
{ .mmi
        stfs [GR_Parameter_Y] = f0,16         // 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
// (3)
{ .mib
        stfs [GR_Parameter_X] =FR_Input_X               // STORE Parameter 1 on stack
        add   GR_Parameter_RESULT = 0,GR_Parameter_Y    // Parameter 3 address
        nop.b 0                                      
}
{ .mib
        stfs [GR_Parameter_Y] = FR_Output_X_tmp         // 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
        nop.m 0
        nop.m 0
        add   GR_Parameter_RESULT = 48,sp
};;

// (4)
{ .mmi
        ldfs  FR_Input_X = [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 
};;

.endp __libm_error_region
ASM_SIZE_DIRECTIVE(__libm_error_region)


.proc __libm_LOG_main 
__libm_LOG_main:
L(LOG_main): 

//
//    kernel_log_64 computes ln(X + E)
//

{ .mfi
	nop.m 999
(p7)  fadd.s.s0 FR_Input_X = FR_Y_lo,FR_Y_hi
        nop.i 999
}

{ .mmi
	nop.m 999
	nop.m 999
(p14) addl GR_Table_Base = @ltoff(Constants_1_by_LN10#),gp ;; 
}

{ .mmi
      nop.m 999
(p14) ld8    GR_Table_Base = [GR_Table_Base]
      nop.i 999
};;

{ .mmi
(p14) ldfe FR_1LN10_hi = [GR_Table_Base],16 ;; 
(p14) ldfe FR_1LN10_lo = [GR_Table_Base]
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p14) fmpy.s1 FR_Output_X_tmp = FR_Y_lo,FR_1LN10_hi
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
(p14) fma.s1  FR_Output_X_tmp = FR_Y_hi,FR_1LN10_lo,FR_Output_X_tmp
	nop.i 999 ;;
}

{ .mfb
	nop.m 999
(p14) fma.s.s0 FR_Input_X = FR_Y_hi,FR_1LN10_hi,FR_Output_X_tmp
(p0)  br.ret.sptk   b0 ;; 
}
.endp __libm_LOG_main
ASM_SIZE_DIRECTIVE(__libm_LOG_main)


.type   __libm_error_support#,@function
.global __libm_error_support#