e_exp.s

Glibc 2.3.2源代码(解压后有100多M)

//

{ .mfi
      add        EXP_AD_TB2 = exp_TB1_size, EXP_AD_TB1
      fcmp.ge.s1  p15,p14 = EXP_NORM_f8,EXP_MIN_DBL_OFLOW_ARG
      nop.i 999
;;                        
}

{ .mfi
      add        EXP_AD_P = exp_TB2_size, EXP_AD_TB2
      fcmp.le.s1  p11,p12 = EXP_NORM_f8,EXP_MAX_DBL_ZERO_ARG
      nop.i 999
;;
}

{ .mfb
      ldfpd      exp_P4, exp_P3  = [EXP_AD_P] ,16
(p14) fcmp.gt.unc.s1  p14,p0 = EXP_NORM_f8,EXP_MAX_DBL_NORM_ARG
(p15) br.cond.spnt L(EXP_CERTAIN_OVERFLOW)
;;
}


// Nfloat = round_int(W) 
// The signficand of EXP_W_2TO56_RSH contains the rounded integer part of W,
// as a twos complement number in the lower bits (that is, it may be negative).
// That twos complement number (called N) is put into exp_GR_N.

// Since EXP_W_2TO56_RSH is scaled by 2^56, it must be multiplied by 2^-56
// before the shift constant 1.10000 * 2^63 is subtracted to yield EXP_Nfloat.
// Thus, EXP_Nfloat contains the floating point version of N


{ .mfi
      nop.m 999
(p12) fcmp.le.unc  p12,p0 = EXP_NORM_f8,EXP_MAX_DBL_UFLOW_ARG
      nop.i 999
}
{ .mfb
      ldfpd      exp_P2, exp_P1  = [EXP_AD_P]                                  
      fms.s1          EXP_Nfloat = EXP_W_2TO56_RSH, EXP_2TOM56, EXP_RSHF 
(p11) br.cond.spnt L(EXP_CERTAIN_UNDERFLOW_ZERO)
;;
}

{ .mfi
      getf.sig        exp_GR_N        = EXP_W_2TO56_RSH
(p13) fcmp.lt.unc  p13,p0 = EXP_NORM_f8,EXP_MIN_DBL_NORM_ARG
      nop.i 999
;;
}


// exp_GR_index_1 has index_1
// exp_GR_index_2_16 has index_2 * 16
// exp_GR_biased_M has M
// exp_GR_index_1_16 has index_1 * 16

// r2 has true M
{ .mfi
      and            exp_GR_index_1 = 0x0f, exp_GR_N
      fnma.s1    exp_r   = EXP_Nfloat, exp_ln2_by_128_hi, EXP_NORM_f8 
      shr            r2 = exp_GR_N,  0x7
}
{ .mfi
      and            exp_GR_index_2_16 = 0x70, exp_GR_N
      fnma.s1    exp_f   = EXP_Nfloat, exp_ln2_by_128_lo, f1 
      nop.i 999
;;                            
}


// EXP_AD_T1 has address of T1                           
// EXP_AD_T2 has address if T2                            

{ .mmi
      addl           exp_GR_biased_M = 0xffff, r2 
      add            EXP_AD_T2 = EXP_AD_TB2, exp_GR_index_2_16 
      shladd         EXP_AD_T1 = exp_GR_index_1, 4, EXP_AD_TB1
;;                            
}


// Create Scale = 2^M
// r = x - Nfloat * ln2_by_128_hi 
// f = 1 - Nfloat * ln2_by_128_lo 

{ .mmi
      setf.exp        EXP_2M = exp_GR_biased_M                              
      ldfe       exp_T2  = [EXP_AD_T2]                                
      nop.i 999
;;
}

// Load T1 and T2
{ .mfi
      ldfe       exp_T1  = [EXP_AD_T1]                                
      nop.f 999
      nop.i 999
;;
}


{ .mfi
        nop.m 999
        fma.s1           exp_rsq = exp_r, exp_r, f0 
        nop.i 999
}
{ .mfi
        nop.m 999
        fma.s1        exp_rP4pP3 = exp_r, exp_P4, exp_P3               
        nop.i 999
;;
}



{ .mfi
        nop.m 999
        fma.s1           exp_rcube = exp_r, exp_rsq, f0 
        nop.i 999 
}
{ .mfi
        nop.m 999
        fma.s1        exp_P_lo  = exp_r, exp_rP4pP3, exp_P2            
        nop.i 999
;;
}


{ .mfi
        nop.m 999
        fma.s1        exp_P_hi  = exp_rsq, exp_P1, exp_r              
        nop.i 999
}
{ .mfi
        nop.m 999
        fma.s1        exp_S2  = exp_f,exp_T2,f0                       
        nop.i 999
;;
}

{ .mfi
        nop.m 999
        fma.s1        exp_S1  = EXP_2M,exp_T1,f0                      
        nop.i 999
;;
}


{ .mfi
        nop.m 999
        fma.s1        exp_P     = exp_rcube, exp_P_lo, exp_P_hi       
        nop.i 999
;;
}

{ .mfi
        nop.m 999
        fma.s1        exp_S   = exp_S1,exp_S2,f0                      
        nop.i 999
;;
}

{ .bbb
(p12)   br.cond.spnt  L(EXP_CERTAIN_UNDERFLOW)
(p13)   br.cond.spnt  L(EXP_POSSIBLE_UNDERFLOW)
(p14)   br.cond.spnt  L(EXP_POSSIBLE_OVERFLOW)
;;
}


{ .mfb
        nop.m 999
        fma.d      f8 = exp_S, exp_P, exp_S 
        br.ret.sptk     b0 ;;               // Normal path exit 
}


L(EXP_POSSIBLE_OVERFLOW): 

// We got an answer. EXP_MAX_DBL_NORM_ARG < x < EXP_MIN_DBL_OFLOW_ARG
// overflow is a possibility, not a certainty

{ .mfi
	nop.m 999
        fsetc.s2 0x7F,0x42                                          
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
        fma.d.s2      exp_wre_urm_f8 = exp_S, exp_P, exp_S          
	nop.i 999 ;;
}

// We define an overflow when the answer with
//    WRE set
//    user-defined rounding mode
// is ldn +1

// Is the exponent 1 more than the largest double?
// If so, go to ERROR RETURN, else get the answer and 
// leave.

// Largest double is 7FE (biased double)
//                   7FE - 3FF + FFFF = 103FE
// Create + largest_double_plus_ulp
// Create - largest_double_plus_ulp
// Calculate answer with WRE set.

// Cases when answer is ldn+1  are as follows:
//  ldn                   ldn+1
// --+----------|----------+------------
//              | 
//    +inf          +inf      -inf
//                  RN         RN
//                             RZ 

{ .mfi
	nop.m 999
        fsetc.s2 0x7F,0x40                                          
        mov           exp_GR_gt_ln  = 0x103ff ;;                      
}

{ .mfi
        setf.exp      exp_gt_pln    = exp_GR_gt_ln                 
	nop.f 999
	nop.i 999 ;;
}

{ .mfi
	nop.m 999
       fcmp.ge.unc.s1 p6, p0 =  exp_wre_urm_f8, exp_gt_pln 	  
	nop.i 999 ;;
}

{ .mfb
	nop.m 999
	nop.f 999
(p6)   br.cond.spnt L(EXP_CERTAIN_OVERFLOW) ;; // Branch if really overflow
}

{ .mfb
	nop.m 999
       fma.d        f8 = exp_S, exp_P, exp_S                      
       br.ret.sptk     b0 ;;             // Exit if really no overflow
}

L(EXP_CERTAIN_OVERFLOW):
{ .mmi
      sub   exp_GR_17ones_m1 = exp_GR_17ones, r0, 1 ;;
      setf.exp     f9 = exp_GR_17ones_m1
      nop.i 999 ;;
}

{ .mfi
      nop.m 999
      fmerge.s FR_X = f8,f8
      nop.i 999
}
{ .mfb
      mov        GR_Parameter_TAG = 14
      fma.d       FR_RESULT = f9, f9, f0    // Set I,O and +INF result
      br.cond.sptk  __libm_error_region ;;                             
}

L(EXP_POSSIBLE_UNDERFLOW): 

// We got an answer. EXP_MAX_DBL_UFLOW_ARG < x < EXP_MIN_DBL_NORM_ARG
// underflow is a possibility, not a certainty

// We define an underflow when the answer with
//    ftz set
// is zero (tiny numbers become zero)

// Notice (from below) that if we have an unlimited exponent range,
// then there is an extra machine number E between the largest denormal and
// the smallest normal.

// So if with unbounded exponent we round to E or below, then we are
// tiny and underflow has occurred.

// But notice that you can be in a situation where we are tiny, namely
// rounded to E, but when the exponent is bounded we round to smallest
// normal. So the answer can be the smallest normal with underflow.

//                           E
// -----+--------------------+--------------------+-----
//      |                    |                    |
//   1.1...10 2^-3fff    1.1...11 2^-3fff    1.0...00 2^-3ffe
//   0.1...11 2^-3ffe                                   (biased, 1)
//    largest dn                               smallest normal

{ .mfi
	nop.m 999
       fsetc.s2 0x7F,0x41                                          
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
       fma.d.s2      exp_ftz_urm_f8 = exp_S, exp_P, exp_S          
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
       fsetc.s2 0x7F,0x40                                          
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
       fcmp.eq.unc.s1 p6, p0 =  exp_ftz_urm_f8, f0 	          
	nop.i 999 ;;
}
{ .mfb
	nop.m 999
	nop.f 999
(p6)   br.cond.spnt L(EXP_CERTAIN_UNDERFLOW) ;; // Branch if really underflow
}
{ .mfb
	nop.m 999
       fma.d        f8 = exp_S, exp_P, exp_S                      
       br.ret.sptk     b0 ;;                // Exit if really no underflow
}

L(EXP_CERTAIN_UNDERFLOW):
{ .mfi
      nop.m 999
      fmerge.s FR_X = f8,f8
      nop.i 999
}
{ .mfb
      mov        GR_Parameter_TAG = 15
      fma.d       FR_RESULT  = exp_S, exp_P, exp_S // Set I,U and tiny result
      br.cond.sptk  __libm_error_region ;;                             
}

L(EXP_CERTAIN_UNDERFLOW_ZERO):
{ .mmi
      mov   exp_GR_one = 1 ;;
      setf.exp     f9 = exp_GR_one
      nop.i 999 ;;
}

{ .mfi
      nop.m 999
      fmerge.s FR_X = f8,f8
      nop.i 999
}
{ .mfb
      mov        GR_Parameter_TAG = 15
      fma.d       FR_RESULT = f9, f9, f0    // Set I,U and tiny (+0.0) result
      br.cond.sptk  __libm_error_region ;;                             
}

.endp exp
ASM_SIZE_DIRECTIVE(exp)


.proc __libm_error_region
__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
        stfd [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
        stfd [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
        stfd [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
        nop.m 0
        nop.m 0
        add   GR_Parameter_RESULT = 48,sp
};;
{ .mmi
        ldfd  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
};; 

.endp __libm_error_region
ASM_SIZE_DIRECTIVE(__libm_error_region)
.type   __libm_error_support#,@function
.global __libm_error_support#