e_log2f.s

glibc 2.9,最新版的C语言库函数

  shl r24=r24,60
}
{.mfi
  mov r26=0x804
  nop.f 0
  // r23=bias-1
  mov r23=0xfffe;;
}

{.mmf
  getf.exp r29=f8
  // load start address for C_1...C_6 followed by T_table
  ld8 r2=[r2]
  // will continue only for positive normal/denormal numbers          
  fclass.nm.unc p12,p7 = f8, 0x19 ;;
}

.pred.rel "mutex",p8,p10
{.mfi
  // denormal input, repeat get significand (after normalization)
  (p8) getf.sig r25=f7
  // x=1 ?
  fcmp.eq.s0 p6,p0=f8,f1
  // get T_index
  (p10) shr.u r28=r25,63-8
}
{.mfi
  // f12=0.5
  setf.exp f12=r23
  nop.f 0
  // r27=bias
  mov r27=0xffff;;
}

{.mfb
  // denormal input, repeat get exponent (after normalization)
  (p8) getf.exp r29=f7
  nop.f 0
  (p12) br.cond.spnt SPECIAL_log2f
}
{.mfi
  cmp.geu p12,p0=r25,r24
  nop.f 0
  mov r23=0xff;;
}

{.mfi
  add r3=32,r2
  // r=1-x*y
  fms.s1 f6=f6,f8,f1
  // r26=0x80400...0 (threshold for using polynomial approximation)
  shl r26=r26,64-12
}
{.mfi
  // load C_3, C_4
  ldfpd f10,f11=[r2],16
  nop.f 0
  // r27=bias-1 (if index >=128, will add exponent+1) 
  (p12) mov r27=0xfffe;;
}

{.mfi
  // load C_1
  ldfe f14=[r2],32
  // x=1, return 0
  (p6) fma.s.s0 f8=f0,f0,f0
  (p8) shr.u r28=r25,63-8
}
{.mib
  // load C_2
  ldfe f13=[r3]
  // r29=exponent-bias
  sub r29=r29,r27
  // x=1, return
  (p6) br.ret.spnt b0;;
}


{.mfi
  // get T_index
  and r28=r28,r23
  fmerge.se f7=f1,f7
  // if first 9 bits after leading 1 are all zero, then p8=1
  cmp.ltu p8,p12=r25,r26;;
}
{.mfi
  // f8=expon - bias 
  setf.sig f8=r29
  nop.f 0
  // get T address
  shladd r2=r28,3,r2
}
{.mfi
  // first 8 bits after leading 1 are all ones ?
  cmp.eq p10,p0=r23,r28
  // if first 8 bits after leading bit are 0, use polynomial approx. only
  (p8) fms.s1 f6=f7,f1,f1
  nop.i 0;;
}
{.mfi
  //r26=1
  mov r26=1
  // if first 8 bits after leading 1 are all ones, use polynomial approx. only
  (p10) fms.s1 f6=f7,f12,f1
  nop.i 0;;
}

.pred.rel "mutex",p8,p12
{.mmf
  // load T (unless first 9 bits after leading 1 are 0)
  (p12) ldfd f12=[r2]
  nop.m 0
  // set T=0 (if first 9 bits after leading 1 are 0)
  (p8) fma.s1 f12=f0,f0,f0;;
}

{.mfi
  nop.m 0
  // P34=C_3+C_4*r
  fma.s1 f10=f11,f6,f10
  // r26=2^{63}
  shl r26=r26,63
}
{.mfi
  nop.m 0
  // r2=r*r
  fma.s1 f11=f6,f6,f0
  nop.i 0;;
}
{.mfi
  // significand of x is 1 ?
  cmp.eq p0,p6=r25,r26
  // P12=C_1+C_2*r
  fma.s1 f14=f13,f6,f14
  nop.i 0;;
}
{.mfi
  nop.m 0
  // normalize additive term (l=exponent of x)
  fcvt.xf f8=f8
  // if significand(x)=1, return exponent (l)
  nop.i 0;;
}
{.mfi
  nop.m 0
  // add T+l
  (p6) fma.s1 f8=f8,f1,f12
  nop.i 0
}
{.mfi
  nop.m 0
  // P14=P12+r2*P34
  (p6) fma.s1 f13=f10,f11,f14
  nop.i 0;;
}

{.mfb
  nop.m 0
  // result=T+l+r*P14
  (p6) fma.s.s0 f8=f13,f6,f8
  // return
  br.ret.sptk b0;;
}


SPECIAL_log2f:
{.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;;
}
{.mfi
  nop.m 0
  // x=-Infinity, -normal, -denormal ?
  fclass.m p6,p0=f8,0x3a
  nop.i 0;;
}
{.mfb
  nop.m 0
  // log2f(+Infinity)=+Infinity
  nop.f 0
  (p7) br.ret.spnt b0;;
}
{.mfi
  (p8) mov GR_Parameter_TAG = 172                          
  // log2f(+/-0)=-infinity, raises Divide by Zero
  // set f8=-0
  (p8) fmerge.ns f8=f0,f8
  nop.i 0;;
}
{.mfb
  nop.m 0
  (p8) frcpa.s0 f8,p0=f1,f8
  (p8) br.cond.sptk __libm_error_region;;
}
{.mfb
  (p6) mov GR_Parameter_TAG = 173 
  // x<0: return NaN, raise Invalid
  (p6) frcpa.s0 f8,p0=f0,f0
  (p6) br.cond.sptk __libm_error_region;;
}                          
  

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

GLOBAL_LIBM_END(log2f)


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
        stfs [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
        stfs [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
        stfs [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
        ldfs  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#