e_acosh.s

glibc 2.9,最新版的C语言库函数

data8 0xc14cb69f7c5e55ab , 0x00003ffd //   log(1/frcpa(1+117/2^-8))
data8 0xc2c2abbb6e5fd56f , 0x00003ffd //   log(1/frcpa(1+118/2^-8))
data8 0xc439b2c193e6771e , 0x00003ffd //   log(1/frcpa(1+119/2^-8))
data8 0xc553acb9d5c67733 , 0x00003ffd //   log(1/frcpa(1+120/2^-8))
//
data8 0xc6cc96e441272441 , 0x00003ffd //   log(1/frcpa(1+121/2^-8))
data8 0xc8469753eca88c30 , 0x00003ffd //   log(1/frcpa(1+122/2^-8))
data8 0xc962cf3ce072b05c , 0x00003ffd //   log(1/frcpa(1+123/2^-8))
data8 0xcadeba8771f694aa , 0x00003ffd //   log(1/frcpa(1+124/2^-8))
data8 0xcc5bc08d1f72da94 , 0x00003ffd //   log(1/frcpa(1+125/2^-8))
//
data8 0xcd7a3f99ea035c29 , 0x00003ffd //   log(1/frcpa(1+126/2^-8))
data8 0xcef93860c8a53c35 , 0x00003ffd //   log(1/frcpa(1+127/2^-8))
data8 0xd0192f68a7ed23df , 0x00003ffd //   log(1/frcpa(1+128/2^-8))
data8 0xd19a201127d3c645 , 0x00003ffd //   log(1/frcpa(1+129/2^-8))
data8 0xd2bb92f4061c172c , 0x00003ffd //   log(1/frcpa(1+130/2^-8))
//
data8 0xd43e80b2ee8cc8fc , 0x00003ffd //   log(1/frcpa(1+131/2^-8))
data8 0xd56173601fc4ade4 , 0x00003ffd //   log(1/frcpa(1+132/2^-8))
data8 0xd6e6637efb54086f , 0x00003ffd //   log(1/frcpa(1+133/2^-8))
data8 0xd80ad9f58f3c8193 , 0x00003ffd //   log(1/frcpa(1+134/2^-8))
data8 0xd991d1d31aca41f8 , 0x00003ffd //   log(1/frcpa(1+135/2^-8))
//
data8 0xdab7d02231484a93 , 0x00003ffd //   log(1/frcpa(1+136/2^-8))
data8 0xdc40d532cde49a54 , 0x00003ffd //   log(1/frcpa(1+137/2^-8))
data8 0xdd685f79ed8b265e , 0x00003ffd //   log(1/frcpa(1+138/2^-8))
data8 0xde9094bbc0e17b1d , 0x00003ffd //   log(1/frcpa(1+139/2^-8))
data8 0xe01c91b78440c425 , 0x00003ffd //   log(1/frcpa(1+140/2^-8))
//
data8 0xe14658f26997e729 , 0x00003ffd //   log(1/frcpa(1+141/2^-8))
data8 0xe270cdc2391e0d23 , 0x00003ffd //   log(1/frcpa(1+142/2^-8))
data8 0xe3ffce3a2aa64922 , 0x00003ffd //   log(1/frcpa(1+143/2^-8))
data8 0xe52bdb274ed82887 , 0x00003ffd //   log(1/frcpa(1+144/2^-8))
data8 0xe6589852e75d7df6 , 0x00003ffd //   log(1/frcpa(1+145/2^-8))
//
data8 0xe786068c79937a7d , 0x00003ffd //   log(1/frcpa(1+146/2^-8))
data8 0xe91903adad100911 , 0x00003ffd //   log(1/frcpa(1+147/2^-8))
data8 0xea481236f7d35bb0 , 0x00003ffd //   log(1/frcpa(1+148/2^-8))
data8 0xeb77d48c692e6b14 , 0x00003ffd //   log(1/frcpa(1+149/2^-8))
data8 0xeca84b83d7297b87 , 0x00003ffd //   log(1/frcpa(1+150/2^-8))
//
data8 0xedd977f4962aa158 , 0x00003ffd //   log(1/frcpa(1+151/2^-8))
data8 0xef7179a22f257754 , 0x00003ffd //   log(1/frcpa(1+152/2^-8))
data8 0xf0a450d139366ca7 , 0x00003ffd //   log(1/frcpa(1+153/2^-8))
data8 0xf1d7e0524ff9ffdb , 0x00003ffd //   log(1/frcpa(1+154/2^-8))
data8 0xf30c29036a8b6cae , 0x00003ffd //   log(1/frcpa(1+155/2^-8))
//
data8 0xf4412bc411ea8d92 , 0x00003ffd //   log(1/frcpa(1+156/2^-8))
data8 0xf576e97564c8619d , 0x00003ffd //   log(1/frcpa(1+157/2^-8))
data8 0xf6ad62fa1b5f172f , 0x00003ffd //   log(1/frcpa(1+158/2^-8))
data8 0xf7e499368b55c542 , 0x00003ffd //   log(1/frcpa(1+159/2^-8))
data8 0xf91c8d10abaffe22 , 0x00003ffd //   log(1/frcpa(1+160/2^-8))
//
data8 0xfa553f7018c966f3 , 0x00003ffd //   log(1/frcpa(1+161/2^-8))
data8 0xfb8eb13e185d802c , 0x00003ffd //   log(1/frcpa(1+162/2^-8))
data8 0xfcc8e3659d9bcbed , 0x00003ffd //   log(1/frcpa(1+163/2^-8))
data8 0xfe03d6d34d487fd2 , 0x00003ffd //   log(1/frcpa(1+164/2^-8))
data8 0xff3f8c7581e9f0ae , 0x00003ffd //   log(1/frcpa(1+165/2^-8))
//
data8 0x803e029e280173ae , 0x00003ffe //   log(1/frcpa(1+166/2^-8))
data8 0x80dca10cc52d0757 , 0x00003ffe //   log(1/frcpa(1+167/2^-8))
data8 0x817ba200632755a1 , 0x00003ffe //   log(1/frcpa(1+168/2^-8))
data8 0x821b05f3b01d6774 , 0x00003ffe //   log(1/frcpa(1+169/2^-8))
data8 0x82bacd623ff19d06 , 0x00003ffe //   log(1/frcpa(1+170/2^-8))
//
data8 0x835af8c88e7a8f47 , 0x00003ffe //   log(1/frcpa(1+171/2^-8))
data8 0x83c5f8299e2b4091 , 0x00003ffe //   log(1/frcpa(1+172/2^-8))
data8 0x8466cb43f3d87300 , 0x00003ffe //   log(1/frcpa(1+173/2^-8))
data8 0x850803a67c80ca4b , 0x00003ffe //   log(1/frcpa(1+174/2^-8))
data8 0x85a9a1d11a23b461 , 0x00003ffe //   log(1/frcpa(1+175/2^-8))
//
data8 0x864ba644a18e6e05 , 0x00003ffe //   log(1/frcpa(1+176/2^-8))
data8 0x86ee1182dcc432f7 , 0x00003ffe //   log(1/frcpa(1+177/2^-8))
data8 0x875a925d7e48c316 , 0x00003ffe //   log(1/frcpa(1+178/2^-8))
data8 0x87fdaa109d23aef7 , 0x00003ffe //   log(1/frcpa(1+179/2^-8))
data8 0x88a129ed4becfaf2 , 0x00003ffe //   log(1/frcpa(1+180/2^-8))
//
data8 0x89451278ecd7f9cf , 0x00003ffe //   log(1/frcpa(1+181/2^-8))
data8 0x89b29295f8432617 , 0x00003ffe //   log(1/frcpa(1+182/2^-8))
data8 0x8a572ac5a5496882 , 0x00003ffe //   log(1/frcpa(1+183/2^-8))
data8 0x8afc2d0ce3b2dadf , 0x00003ffe //   log(1/frcpa(1+184/2^-8))
data8 0x8b6a69c608cfd3af , 0x00003ffe //   log(1/frcpa(1+185/2^-8))
//
data8 0x8c101e106e899a83 , 0x00003ffe //   log(1/frcpa(1+186/2^-8))
data8 0x8cb63de258f9d626 , 0x00003ffe //   log(1/frcpa(1+187/2^-8))
data8 0x8d2539c5bd19e2b1 , 0x00003ffe //   log(1/frcpa(1+188/2^-8))
data8 0x8dcc0e064b29e6f1 , 0x00003ffe //   log(1/frcpa(1+189/2^-8))
data8 0x8e734f45d88357ae , 0x00003ffe //   log(1/frcpa(1+190/2^-8))
//
data8 0x8ee30cef034a20db , 0x00003ffe //   log(1/frcpa(1+191/2^-8))
data8 0x8f8b0515686d1d06 , 0x00003ffe //   log(1/frcpa(1+192/2^-8))
data8 0x90336bba039bf32f , 0x00003ffe //   log(1/frcpa(1+193/2^-8))
data8 0x90a3edd23d1c9d58 , 0x00003ffe //   log(1/frcpa(1+194/2^-8))
data8 0x914d0de2f5d61b32 , 0x00003ffe //   log(1/frcpa(1+195/2^-8))
//
data8 0x91be0c20d28173b5 , 0x00003ffe //   log(1/frcpa(1+196/2^-8))
data8 0x9267e737c06cd34a , 0x00003ffe //   log(1/frcpa(1+197/2^-8))
data8 0x92d962ae6abb1237 , 0x00003ffe //   log(1/frcpa(1+198/2^-8))
data8 0x9383fa6afbe2074c , 0x00003ffe //   log(1/frcpa(1+199/2^-8))
data8 0x942f0421651c1c4e , 0x00003ffe //   log(1/frcpa(1+200/2^-8))
//
data8 0x94a14a3845bb985e , 0x00003ffe //   log(1/frcpa(1+201/2^-8))
data8 0x954d133857f861e7 , 0x00003ffe //   log(1/frcpa(1+202/2^-8))
data8 0x95bfd96468e604c4 , 0x00003ffe //   log(1/frcpa(1+203/2^-8))
data8 0x9632d31cafafa858 , 0x00003ffe //   log(1/frcpa(1+204/2^-8))
data8 0x96dfaabd86fa1647 , 0x00003ffe //   log(1/frcpa(1+205/2^-8))
//
data8 0x9753261fcbb2a594 , 0x00003ffe //   log(1/frcpa(1+206/2^-8))
data8 0x9800c11b426b996d , 0x00003ffe //   log(1/frcpa(1+207/2^-8))
data8 0x9874bf4d45ae663c , 0x00003ffe //   log(1/frcpa(1+208/2^-8))
data8 0x99231f5ee9a74f79 , 0x00003ffe //   log(1/frcpa(1+209/2^-8))
data8 0x9997a18a56bcad28 , 0x00003ffe //   log(1/frcpa(1+210/2^-8))
//
data8 0x9a46c873a3267e79 , 0x00003ffe //   log(1/frcpa(1+211/2^-8))
data8 0x9abbcfc621eb6cb6 , 0x00003ffe //   log(1/frcpa(1+212/2^-8))
data8 0x9b310cb0d354c990 , 0x00003ffe //   log(1/frcpa(1+213/2^-8))
data8 0x9be14cf9e1b3515c , 0x00003ffe //   log(1/frcpa(1+214/2^-8))
data8 0x9c5710b8cbb73a43 , 0x00003ffe //   log(1/frcpa(1+215/2^-8))
//
data8 0x9ccd0abd301f399c , 0x00003ffe //   log(1/frcpa(1+216/2^-8))
data8 0x9d7e67f3bdce8888 , 0x00003ffe //   log(1/frcpa(1+217/2^-8))
data8 0x9df4ea81a99daa01 , 0x00003ffe //   log(1/frcpa(1+218/2^-8))
data8 0x9e6ba405a54514ba , 0x00003ffe //   log(1/frcpa(1+219/2^-8))
data8 0x9f1e21c8c7bb62b3 , 0x00003ffe //   log(1/frcpa(1+220/2^-8))
//
data8 0x9f956593f6b6355c , 0x00003ffe //   log(1/frcpa(1+221/2^-8))
data8 0xa00ce1092e5498c3 , 0x00003ffe //   log(1/frcpa(1+222/2^-8))
data8 0xa0c08309c4b912c1 , 0x00003ffe //   log(1/frcpa(1+223/2^-8))
data8 0xa1388a8c6faa2afa , 0x00003ffe //   log(1/frcpa(1+224/2^-8))
data8 0xa1b0ca7095b5f985 , 0x00003ffe //   log(1/frcpa(1+225/2^-8))
//
data8 0xa22942eb47534a00 , 0x00003ffe //   log(1/frcpa(1+226/2^-8))
data8 0xa2de62326449d0a3 , 0x00003ffe //   log(1/frcpa(1+227/2^-8))
data8 0xa357690f88bfe345 , 0x00003ffe //   log(1/frcpa(1+228/2^-8))
data8 0xa3d0a93f45169a4b , 0x00003ffe //   log(1/frcpa(1+229/2^-8))
data8 0xa44a22f7ffe65f30 , 0x00003ffe //   log(1/frcpa(1+230/2^-8))
//
data8 0xa500c5e5b4c1aa36 , 0x00003ffe //   log(1/frcpa(1+231/2^-8))
data8 0xa57ad064eb2ebbc2 , 0x00003ffe //   log(1/frcpa(1+232/2^-8))
data8 0xa5f5152dedf4384e , 0x00003ffe //   log(1/frcpa(1+233/2^-8))
data8 0xa66f9478856233ec , 0x00003ffe //   log(1/frcpa(1+234/2^-8))
data8 0xa6ea4e7cca02c32e , 0x00003ffe //   log(1/frcpa(1+235/2^-8))
//
data8 0xa765437325341ccf , 0x00003ffe //   log(1/frcpa(1+236/2^-8))
data8 0xa81e21e6c75b4020 , 0x00003ffe //   log(1/frcpa(1+237/2^-8))
data8 0xa899ab333fe2b9ca , 0x00003ffe //   log(1/frcpa(1+238/2^-8))
data8 0xa9157039c51ebe71 , 0x00003ffe //   log(1/frcpa(1+239/2^-8))
data8 0xa991713433c2b999 , 0x00003ffe //   log(1/frcpa(1+240/2^-8))
//
data8 0xaa0dae5cbcc048b3 , 0x00003ffe //   log(1/frcpa(1+241/2^-8))
data8 0xaa8a27ede5eb13ad , 0x00003ffe //   log(1/frcpa(1+242/2^-8))
data8 0xab06de228a9e3499 , 0x00003ffe //   log(1/frcpa(1+243/2^-8))
data8 0xab83d135dc633301 , 0x00003ffe //   log(1/frcpa(1+244/2^-8))
data8 0xac3fb076adc7fe7a , 0x00003ffe //   log(1/frcpa(1+245/2^-8))
//
data8 0xacbd3cbbe47988f1 , 0x00003ffe //   log(1/frcpa(1+246/2^-8))
data8 0xad3b06b1a5dc57c3 , 0x00003ffe //   log(1/frcpa(1+247/2^-8))
data8 0xadb90e94af887717 , 0x00003ffe //   log(1/frcpa(1+248/2^-8))
data8 0xae3754a218f7c816 , 0x00003ffe //   log(1/frcpa(1+249/2^-8))
data8 0xaeb5d9175437afa2 , 0x00003ffe //   log(1/frcpa(1+250/2^-8))
//
data8 0xaf349c322e9c7cee , 0x00003ffe //   log(1/frcpa(1+251/2^-8))
data8 0xafb39e30d1768d1c , 0x00003ffe //   log(1/frcpa(1+252/2^-8))
data8 0xb032df51c2c93116 , 0x00003ffe //   log(1/frcpa(1+253/2^-8))
data8 0xb0b25fd3e6035ad9 , 0x00003ffe //   log(1/frcpa(1+254/2^-8))
data8 0xb1321ff67cba178c , 0x00003ffe //   log(1/frcpa(1+255/2^-8))
LOCAL_OBJECT_END(log_table_3)


.section .text
GLOBAL_LIBM_ENTRY(acosh)

{ .mfi
      getf.exp   acosh_GR_f8 = f8
      fclass.m   p6,p0 = f8, 0xc3                    // Test for x = NaN
      mov        log_GR_comp2 = 0x1003e
}
{ .mfi
      addl       NR_table_address = @ltoff(log_table_1), gp
      fms.s1     log_y = f8, f8, f1                  // y = x^2-1
      nop.i      0
}
;;

{ .mfi
      getf.sig   acosh_GR_f8_sig = f8
      fclass.m   p11,p0 = f8, 0x21                   // Test for x=+inf
      mov        log_GR_exp_17_ones = 0x1ffff
}
{ .mfi
      ld8        NR_table_address = [NR_table_address]
      fms.s1     log_w = f8,f1,f1                    // w = x - 1
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fcmp.lt.s1 p7,p8 = f8, f1            // Test for x<1.0
      addl       log_GR_comp = 0x10020C,r0 // Upper 21 bits of signif of 1.0005
}
{ .mfb
      mov        log_GR_exp_16_ones = 0xffff         //BIAS
(p6)  fma.d.s0   f8 = f8,f1,f0      // quietize nan result if x=nan
(p6)  br.ret.spnt b0                // Exit for x=nan
}
;;

{ .mfb
      //get second table address
      adds       log_table_address2 = 0x40, NR_table_address
      fcmp.eq.s1 p10,p0 = f8, f1      // Test for x=+1.0
(p11) br.ret.spnt b0                  // Exit for x=+inf
}
;;

{ .mfi
      ldfpd      NR1,NR2 = [log_table_address2],16
      frsqrta.s1 log_y_rs,p0 = log_y  // z=1/sqrt(y)
      nop.i      0
}
{ .mfb
      nop.m      0
      fma.s1     log_arg = f8,f1,f8
(p7)  br.cond.spnt ACOSH_LESS_ONE     // Branch if path 7, x < 1.0
}
;;

{ .mfi
      ldfe       log_C4 = [log_table_address2],16
(p8)  fcmp.eq.s0 p6,p0 = f8, f0       // Dummy op sets denorm flag if unorm>=1.0
      nop.i      0
}
{ .mfb
(p8)  cmp.le.unc p13,p0 = log_GR_comp2,acosh_GR_f8
      nop.f      0
(p13) br.cond.spnt LOG_COMMON1        // Branch if path 4, x >= 2^63
}
;;

{ .mfi
      ldfe       log_C3 = [log_table_address2],16
(p10) fmerge.s   f8 = f0, f0          // Return 0 if x=1.0
      shr.u      acosh_GR_f8_sig = acosh_GR_f8_sig,43
}
{ .mib
      cmp.eq     p14,p0 = log_GR_exp_16_ones,acosh_GR_f8
      nop.i      0
(p10) br.ret.spnt b0                  // Exit for x=1.0
}
;;

{ .mfi
      ldfe       log_C2 = [log_table_address2],16
      frsqrta.s1 acosh_w_rs,p0 = log_w // t=1/sqrt(w)
      nop.i      0
}
{ .mfb
(p14) cmp.lt.unc p15,p0 = acosh_GR_f8_sig,log_GR_comp
      nop.f      0
(p15) br.cond.spnt ACOSH_NEAR_ONE     // Branch if path 2, 1.0 < x < 1.0005
}
;;

// Here is main path, 1.0005 <= x < 2^63
/////////////// The first iteration //////////////////////////////////
{ .mfi
      ldfpd      acosh_comp,log_P5 = [NR_table_address],16
      fma.s1     log_y_rs_iter = log_y_rs,log_y,f0              // y*z
      nop.i      0
}
;;

{ .mfi
      ldfpd      log_P4,log_P3 = [NR_table_address],16
      fnma.s1    log_y_rs_iter = log_y_rs_iter,log_y_rs,NR2     // 3-(y*z)*z
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     log_y_rs_iter1 = log_y_rs,NR1,f0               // 0.5*z
      nop.i      0
}
;;

{ .mfi
      ldfpd      log_P2,log_P1 = [NR_table_address],16
      //(0.5*z)*(3-(y*z)*z)
      fma.s1     log_y_rs_iter = log_y_rs_iter1,log_y_rs_iter,f0
      nop.i      0
}
;;

/////////////////////////// The second iteration /////////////////////////////
{ .mfi
      nop.m      0
      fma.s1     log_y_rs = log_y_rs_iter,log_y,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fnma.s1    log_y_rs = log_y_rs,log_y_rs_iter,NR2
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     log_y_rs_iter1 = log_y_rs_iter,NR1,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      //(0.5*z)*(3-(y*z)*z)
      fma.s1     log_y_rs_iter = log_y_rs_iter1,log_y_rs,f0
      nop.i      0
}
{ .mfi
      nop.m      0
      //(0.5*z)*(3-(y*z)*z)
      fma.s1     log_arg_early = log_y_rs_iter1,log_y_rs,f0
      nop.i      0
}
;;

//////////////////////////////////////// The third iteration /////////////////
{ .mfi
      nop.m      0
      fma.s1     log_y_rs = log_y_rs_iter,log_y,f0
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     log_y_rs_iter1 = log_y_rs_iter,NR1,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     log_arg_early = log_arg_early,log_y,f8
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fnma.s1    log_y_rs = log_y_rs,log_y_rs_iter,NR2
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     log_y_rs_iter1 = log_y_rs_iter1,log_y,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      frcpa.s1   log_C,p0 = f1,log_arg_early
      nop.i      0
}
;;

{ .mfi
      getf.exp   log_GR_signexp_f8 = log_arg_early
      nop.f      0
      nop.i      0
}
;;

{ .mfi
      getf.sig   log_GR_significand_f8 = log_arg_early
      fma.s1     log_arg = log_y_rs_iter1,log_y_rs,f8 // (0.5*z)*(3-(y*z)*z)
      adds       log_table_address3 = 0x70, NR_table_address
}
;;

///////////////////////////////// The end NR iterations /////////////////////
{ .mfi
      ldfe       log2 = [NR_table_address],16
      nop.f      0
      nop.i      0
}
;;

{ .mmi
      //significant bit destruction
      and        log_GR_exp_f8 = log_GR_signexp_f8, log_GR_exp_17_ones
;;
      //BIAS subtraction
      sub        log_GR_true_exp_f8 = log_GR_exp_f8, log_GR_exp_16_ones
      nop.i      0
}
;;

{ .mfi
      setf.sig   log_int_Nfloat = log_GR_true_exp_f8
      fms.s1     log_r = log_C,log_arg,f1  // C = frcpa(x); r = C * x - 1