aes/asm/aesni-x86[_64].pl update.
[openssl.git] / crypto / aes / asm / aes-ia64.S
index 1bad469e55094c6d19b0b785ce627dfac41b3050..7f6c4c366291c86c1214b726918757132b4eafa3 100644 (file)
 // much M-ports as there're I-ports on Itanium 2]. By sacrificing few
 // registers for small constants (255, 24 and 16) to be used with
 // 'shr' and 'and' instructions I can achieve better ILP, Intruction
-// Level Parallelism, and performance. This code outperforms gcc
-// generated code by almost factor of 2 (two). Improvement over HP C
-// is not that impressive, 20%...
+// Level Parallelism, and performance. This code outperforms GCC 3.3
+// generated code by over factor of 2 (two), GCC 3.4 - by 70% and
+// HP C - by 40%. Measured best-case scenario, i.e. aligned
+// big-endian input, ECB timing on Itanium 2 is (18 + 13*rounds)
+// ticks per block, or 9.25 CPU cycles per byte for 128 bit key.
 
-.ident "aes-ia64.S, version 1.0"
+// Version 1.2 mitigates the hazard of cache-timing attacks by
+// a) compressing S-boxes from 8KB to 2KB+256B, b) scheduling
+// references to S-boxes for L2 cache latency, c) prefetching T[ed]4
+// prior last round. As result performance dropped to (26 + 15*rounds)
+// ticks per block or 11 cycles per byte processed with 128-bit key.
+// This is ~16% deterioration. For reference Itanium 2 L1 cache has
+// 64 bytes line size and L2 - 128 bytes...
+
+.ident "aes-ia64.S, version 1.2"
 .ident "IA-64 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
 .explicit
 .text
 
 rk0=r8;     rk1=r9;
 
-prsave=r10;
+pfssave=r2;
+lcsave=r10;
+prsave=r3;
 maskff=r11;
 twenty4=r14;
 sixteen=r15;
@@ -42,132 +54,62 @@ te0=r40;    te1=r41;    te2=r42;    te3=r43;
 
 #if defined(_HPUX_SOURCE) && !defined(_LP64)
 # define ADDP  addp4
-# define KSZ   4
-# define LDKEY ld4
 #else
 # define ADDP  add
 #endif
 
+// Offsets from Te0
+#define TE0    0
+#define TE2    2
+#if defined(_HPUX_SOURCE) || defined(B_ENDIAN)
+#define TE1    3
+#define TE3    1
+#else
+#define TE1    1
+#define TE3    3
+#endif
+
+// This implies that AES_KEY comprises 32-bit key schedule elements
+// even on LP64 platforms.
 #ifndef        KSZ
-# define KSZ   8
-# define LDKEY ld8
+# define KSZ   4
+# define LDKEY ld4
 #endif
 
-// void AES_encrypt (const void *in,void *out,const AES_KEY *key);
-// measured timing om Itanium 2 is (48 + 14*rounds) cycles
-.global        AES_encrypt#
-.proc  AES_encrypt#
+.proc  _ia64_AES_encrypt#
+// Input:      rk0-rk1
+//             te0
+//             te3     as AES_KEY->rounds!!!
+//             s0-s3
+//             maskff,twenty4,sixteen
+// Output:     r16,r20,r24,r28 as s0-s3
+// Clobber:    r16-r31,rk0-rk1,r32-r43
 .align 32
-#if !defined(_HPUX_SOURCE)
-.skip  16
-#endif
-AES_encrypt:
+_ia64_AES_encrypt:
        .prologue
-       .fframe 0
-       .save   ar.pfs,r2
-       .save   ar.lc,r3
-{ .mii;        alloc   r2=ar.pfs,3,10,0,8
-       mov     r3=ar.lc        
-       mov     prsave=pr       };;
-
+       .altrp  b6
        .body
-{ .mmi;        and     r40=3,r32
-       ADDP    r32=0,r32
-       mov     pr.rot=7<<16    };;
-#if defined(_HPUX_SOURCE)      // HPUX is big-endian, cut 15 cycles...
-{ .mib; cmp.ne p6,p0=r40,r0
-       add     r41=4,r32               // 1st arg, borrow teN
-(p6)   br.dpnt.many    .Le_unaligned   };;
-
-{ .mmi;        ld4     r19=[r32],8
-       mov     r44=r33                 // save 2nd arg
-       mov     twenty4=24      }
-{ .mmi;        ld4     r23=[r41],8
-       addl    te0=@ltoff(Te0#),gp
-       ADDP    r35=KSZ*60,r34  };;     // &AES_KEY->rounds, borrow s1
-{ .mmi;        ld8     te0=[te0]
-       ld4     r35=[r35]               // AES_KEY->rounds
-       ADDP    rk0=0,r34       }//;;   // 3rd arg
-{ .mmi;        ld4     r27=[r32]
-       ld4     r31=[r41]               
-       ADDP    rk1=KSZ,r34     };;
-
-{ .mfi; LDKEY  t0=[rk0],2*KSZ
-       mov     sixteen=16      }
-{ .mfi;        LDKEY   t1=[rk1],2*KSZ
-       mov     maskff=0xff     };;
-{ .mfi;        LDKEY   t2=[rk0],2*KSZ
-       add     te1=1024,te0    }
+{ .mmi;        alloc   r16=ar.pfs,12,0,0,8
+       LDKEY   t0=[rk0],2*KSZ
+       mov     pr.rot=1<<16    }
+{ .mmi;        LDKEY   t1=[rk1],2*KSZ
+       add     te1=TE1,te0
+       add     te3=-3,te3      };;
+{ .mib;        LDKEY   t2=[rk0],2*KSZ
+       mov     ar.ec=2         }
 { .mib;        LDKEY   t3=[rk1],2*KSZ
-       add     te2=2048,te0
-       br.many .Le_common      };;
-#endif
-.Le_unaligned:
-{ .mfi;        ADDP    r40=0,r32               // 1st arg, borrow teN
-       ADDP    r41=1,r32       }
-{ .mfi;        ADDP    r42=2,r32
-       ADDP    r43=3,r32       };;
-{ .mmi;        ld1     r16=[r40],4
-       ld1     r17=[r41],4
-       mov     r44=r33         }//;;   // save 2nd arg
-{ .mmi;        ld1     r18=[r42],4
-       ld1     r19=[r43],4
-       ADDP    rk0=0,r34       };;     // 3rd arg
-{ .mmi;        ld1     r20=[r40],4
-       ld1     r21=[r41],4
-       ADDP    rk1=KSZ,r34     }//;;
-{ .mmi;        ld1     r22=[r42],4
-       ld1     r23=[r43],4
-       ADDP    r35=KSZ*60,r34  };;     // &AES_KEY->rounds, borrow s1
-{ .mmi;        ld1     r24=[r40],4
-       ld1     r25=[r41],4
-       mov     twenty4=24      }//;;
-{ .mmi;        ld1     r26=[r42],4
-       ld1     r27=[r43],4
-       mov     sixteen=16      };;
-{ .mmi;        ld1     r28=[r40]
-       ld1     r29=[r41]
-       mov     maskff=0xff     }//;;
-{ .mmi;        ld1     r30=[r42]
-       ld1     r31=[r43]
-       addl    te0=@ltoff(Te0#),gp     };;     // that was close...
-
-{ .mii;        ld8     te0=[te0]
-       dep     r19=r16,r19,24,8        //;;
-       dep     r23=r20,r23,24,8        }//;;
-{ .mii;        ld4     r35=[r35]               // AES_KEY->rounds
-       dep     r27=r24,r27,24,8        //;;
-       dep     r31=r28,r31,24,8        };;
-{ .mii;        LDKEY   t0=[rk0],2*KSZ
-       dep     r19=r17,r19,16,8        //;;
-       dep     r23=r21,r23,16,8        }//;;
-{ .mii;        LDKEY   t1=[rk1],2*KSZ
-       dep     r27=r25,r27,16,8        //;;
-       dep     r31=r29,r31,16,8        };;
-{ .mii;        LDKEY   t2=[rk0],2*KSZ
-       dep     r19=r18,r19,8,8         //;;
-       dep     r23=r22,r23,8,8         }//;;
-{ .mii;        LDKEY   t3=[rk1],2*KSZ
-       dep     r27=r26,r27,8,8         //;;
-       dep     r31=r30,r31,8,8         };;
-       
-{ .mib;        add     te1=1024,te0
-       add     te2=2048,te0    }
-.Le_common:
-{ .mib; add    te3=3072,te0
-       add     r35=-3,r35
-       brp.exit.imp    .Le_rounds_cexit,.Le_cexit_insn
-                               };;
-{ .mii;        mov     ar.lc=r35               // borrowed s1
-       mov     ar.ec=3         };;
-
-{ .mfi;        xor     s0=r19,t0
-       xor     s1=r23,t1       }
-{ .mfi;        xor     s2=r27,t2
-       xor     s3=r31,t3       };;
+       add     te2=TE2,te0
+       brp.loop.imp    .Le_top,.Le_end-16      };;
 
+{ .mmi;        xor     s0=s0,t0
+       xor     s1=s1,t1
+       mov     ar.lc=te3       }
+{ .mmi;        xor     s2=s2,t2
+       xor     s3=s3,t3
+       add     te3=TE3,te0     };;
+\f
 .align 32
-.Le_rounds:
+.Le_top:
 { .mmi;        (p0)    LDKEY   t0=[rk0],2*KSZ          // 0/0:rk[0]
        (p0)    and     te33=s3,maskff          // 0/0:s3&0xff
        (p0)    extr.u  te22=s2,8,8     }       // 0/0:s2>>8&0xff
@@ -175,144 +117,309 @@ AES_encrypt:
        (p0)    and     te30=s0,maskff          // 0/1:s0&0xff
        (p0)    shr.u   te00=s0,twenty4 };;     // 0/0:s0>>24
 { .mmi;        (p0)    LDKEY   t2=[rk0],2*KSZ          // 1/2:rk[2]
-       (p0)    shladd  te33=te33,2,te3         // 1/0:te0+s0>>24
+       (p0)    shladd  te33=te33,3,te3         // 1/0:te0+s0>>24
        (p0)    extr.u  te23=s3,8,8     }       // 1/1:s3>>8&0xff
 { .mmi;        (p0)    LDKEY   t3=[rk1],2*KSZ          // 1/3:rk[3]
-       (p0)    shladd  te30=te30,2,te3         // 1/1:te3+s0
+       (p0)    shladd  te30=te30,3,te3         // 1/1:te3+s0
        (p0)    shr.u   te01=s1,twenty4 };;     // 1/1:s1>>24
 { .mmi;        (p0)    ld4     te33=[te33]             // 2/0:te3[s3&0xff]
-       (p0)    shladd  te22=te22,2,te2         // 2/0:te2+s2>>8&0xff
+       (p0)    shladd  te22=te22,3,te2         // 2/0:te2+s2>>8&0xff
        (p0)    extr.u  te20=s0,8,8     }       // 2/2:s0>>8&0xff
 { .mmi;        (p0)    ld4     te30=[te30]             // 2/1:te3[s0]
-       (p0)    shladd  te23=te23,2,te2         // 2/1:te2+s3>>8
+       (p0)    shladd  te23=te23,3,te2         // 2/1:te2+s3>>8
        (p0)    shr.u   te02=s2,twenty4 };;     // 2/2:s2>>24
 { .mmi;        (p0)    ld4     te22=[te22]             // 3/0:te2[s2>>8]
-       (p0)    shladd  te20=te20,2,te2         // 3/2:te2+s0>>8
+       (p0)    shladd  te20=te20,3,te2         // 3/2:te2+s0>>8
        (p0)    extr.u  te21=s1,8,8     }       // 3/3:s1>>8&0xff
 { .mmi;        (p0)    ld4     te23=[te23]             // 3/1:te2[s3>>8]
-       (p0)    shladd  te00=te00,2,te0         // 3/0:te0+s0>>24
+       (p0)    shladd  te00=te00,3,te0         // 3/0:te0+s0>>24
        (p0)    shr.u   te03=s3,twenty4 };;     // 3/3:s3>>24
 { .mmi;        (p0)    ld4     te20=[te20]             // 4/2:te2[s0>>8]
-       (p0)    shladd  te21=te21,2,te2         // 4/3:te3+s2
+       (p0)    shladd  te21=te21,3,te2         // 4/3:te3+s2
        (p0)    extr.u  te11=s1,16,8    }       // 4/0:s1>>16&0xff
 { .mmi;        (p0)    ld4     te00=[te00]             // 4/0:te0[s0>>24]
-       (p0)    shladd  te01=te01,2,te0         // 4/1:te0+s1>>24
+       (p0)    shladd  te01=te01,3,te0         // 4/1:te0+s1>>24
        (p0)    shr.u   te13=s3,sixteen };;     // 4/2:s3>>16
 { .mmi;        (p0)    ld4     te21=[te21]             // 5/3:te2[s1>>8]
-       (p0)    shladd  te11=te11,2,te1         // 5/0:te1+s1>>16
+       (p0)    shladd  te11=te11,3,te1         // 5/0:te1+s1>>16
        (p0)    extr.u  te12=s2,16,8    }       // 5/1:s2>>16&0xff
 { .mmi;        (p0)    ld4     te01=[te01]             // 5/1:te0[s1>>24]
-       (p0)    shladd  te02=te02,2,te0         // 5/2:te0+s2>>24
+       (p0)    shladd  te02=te02,3,te0         // 5/2:te0+s2>>24
        (p0)    and     te31=s1,maskff  };;     // 5/2:s1&0xff
-
 { .mmi;        (p0)    ld4     te11=[te11]             // 6/0:te1[s1>>16]
-       (p0)    shladd  te12=te12,2,te1         // 6/1:te1+s2>>16
+       (p0)    shladd  te12=te12,3,te1         // 6/1:te1+s2>>16
        (p0)    extr.u  te10=s0,16,8    }       // 6/3:s0>>16&0xff
 { .mmi;        (p0)    ld4     te02=[te02]             // 6/2:te0[s2>>24]
-       (p0)    shladd  te03=te03,2,te0         // 6/3:te1+s0>>16
+       (p0)    shladd  te03=te03,3,te0         // 6/3:te1+s0>>16
        (p0)    and     te32=s2,maskff  };;     // 6/3:s2&0xff
+
 { .mmi;        (p0)    ld4     te12=[te12]             // 7/1:te1[s2>>16]
-       (p0)    shladd  te31=te31,2,te3         // 7/2:te3+s1&0xff
+       (p0)    shladd  te31=te31,3,te3         // 7/2:te3+s1&0xff
        (p0)    and     te13=te13,maskff}       // 7/2:s3>>16&0xff
 { .mmi;        (p0)    ld4     te03=[te03]             // 7/3:te0[s3>>24]
-       (p0)    shladd  te32=te32,2,te3         // 7/3:te3+s2
-       (p16)   cmp.eq  p0,p18=r0,r0    };;     // 7/clear (p18)
+       (p0)    shladd  te32=te32,3,te3         // 7/3:te3+s2
+       (p0)    xor     t0=t0,te33      };;     // 7/0:
 { .mmi;        (p0)    ld4     te31=[te31]             // 8/2:te3[s1]
-       (p0)    shladd  te13=te13,2,te1         // 8/2:te1+s3>>16
-       (p17)   xor     t0=t0,te33      }       // 8/0:
+       (p0)    shladd  te13=te13,3,te1         // 8/2:te1+s3>>16
+       (p0)    xor     t0=t0,te22      }       // 8/0:
 { .mmi;        (p0)    ld4     te32=[te32]             // 8/3:te3[s2]
-       (p0)    shladd  te10=te10,2,te1         // 8/3:te1+s0>>16
-       (p17)   xor     t1=t1,te30      };;     // 8/1:
+       (p0)    shladd  te10=te10,3,te1         // 8/3:te1+s0>>16
+       (p0)    xor     t1=t1,te30      };;     // 8/1:
 { .mmi;        (p0)    ld4     te13=[te13]             // 9/2:te1[s3>>16]
-       (p17)   xor     t0=t0,te22              // 9/0:
-       (p18)   add     te0=4096,te0    }       // 9/
-.Le_cexit_insn:
-{ .mmb;        (p0)    ld4     te10=[te10]             // 9/3:te1[s0>>16]
-       (p17)   xor     t1=t1,te23              // 9/1:
-       br.cexit.spnt.few       .Le_rounds_cexit
-                                       };;
-{ .mmi;        (p18)   xor     s2=s2,te20              // 10/2:
-       (p18)   xor     s0=s0,te00              // 10/0:
-       (p19)   add     te1=3072,te1    }       // 10/
-{ .mmi;        (p18)   xor     s3=s3,te21              // 10/3:
-       (p18)   xor     s1=s1,te01              // 10/1:
-       (p19)   add     te2=2048,te2    };;     // 10/
-{ .mfi;        (p18)   xor     s0=s0,te11              // 11/0:done!
-       (p18)   xor     s2=s2,te02      }       // 11/2:
-{ .mfi;        (p18)   xor     s1=s1,te12              // 11/1:done!
-       (p18)   xor     s3=s3,te03      };;     // 11/3:
-{ .mmi;        (p18)   xor     s2=s2,te31              // 12/2:
-       (p18)   xor     s3=s3,te32              // 12/3:
-       (p19)   add     te3=1024,te3    };;     // 12/
-{ .mib;        (p18)   xor     s2=s2,te13              // 13/2:done!
-       (p18)   xor     s3=s3,te10              // 13/3:done!
-       br      .Le_rounds      };;
+       (p0)    ld4     te10=[te10]             // 9/3:te1[s0>>16]
+       (p0)    xor     t0=t0,te00      };;     // 9/0:         !L2 scheduling
+{ .mmi;        (p0)    xor     t1=t1,te23              // 10[9]/1:     
+       (p0)    xor     t2=t2,te20              // 10[9]/2:
+       (p0)    xor     t3=t3,te21      };;     // 10[9]/3:
+{ .mmi;        (p0)    xor     t0=t0,te11              // 11[10]/0:done!
+       (p0)    xor     t1=t1,te01              // 11[10]/1:
+       (p0)    xor     t2=t2,te02      };;     // 11[10]/2:    !L2 scheduling
+{ .mmi;        (p0)    xor     t3=t3,te03              // 12[10]/3:
+       (p16)   cmp.eq  p0,p17=r0,r0    };;     // 12[10]/clear (p17)
+{ .mmi;        (p0)    xor     t1=t1,te12              // 13[11]/1:done!
+       (p0)    xor     t2=t2,te31              // 13[11]/2:
+       (p0)    xor     t3=t3,te32      }       // 13[11]/3:
+{ .mmi;        (p17)   add     te0=2048,te0            // 13[11]/
+       (p17)   add     te1=2048+64-TE1,te1};;  // 13[11]/
+{ .mib;        (p0)    xor     t2=t2,te13              // 14[12]/2:done!
+       (p17)   add     te2=2048+128-TE2,te2}   // 14[12]/
+{ .mib;        (p0)    xor     t3=t3,te10              // 14[12]/3:done!
+       (p17)   add     te3=2048+192-TE3,te3    // 14[12]/
+       br.ctop.sptk    .Le_top         };;
+.Le_end:
+\f
+
+{ .mmi;        ld8     te12=[te0]              // prefetch Te4
+       ld8     te31=[te1]      }
+{ .mmi;        ld8     te10=[te2]
+       ld8     te32=[te3]      }
+
+{ .mmi;        LDKEY   t0=[rk0],2*KSZ          // 0/0:rk[0]
+       and     te33=s3,maskff          // 0/0:s3&0xff
+       extr.u  te22=s2,8,8     }       // 0/0:s2>>8&0xff
+{ .mmi; LDKEY  t1=[rk1],2*KSZ          // 0/1:rk[1]
+       and     te30=s0,maskff          // 0/1:s0&0xff
+       shr.u   te00=s0,twenty4 };;     // 0/0:s0>>24
+{ .mmi;        LDKEY   t2=[rk0],2*KSZ          // 1/2:rk[2]
+       add     te33=te33,te0           // 1/0:te0+s0>>24
+       extr.u  te23=s3,8,8     }       // 1/1:s3>>8&0xff
+{ .mmi;        LDKEY   t3=[rk1],2*KSZ          // 1/3:rk[3]
+       add     te30=te30,te0           // 1/1:te0+s0
+       shr.u   te01=s1,twenty4 };;     // 1/1:s1>>24
+{ .mmi;        ld1     te33=[te33]             // 2/0:te0[s3&0xff]
+       add     te22=te22,te0           // 2/0:te0+s2>>8&0xff
+       extr.u  te20=s0,8,8     }       // 2/2:s0>>8&0xff
+{ .mmi;        ld1     te30=[te30]             // 2/1:te0[s0]
+       add     te23=te23,te0           // 2/1:te0+s3>>8
+       shr.u   te02=s2,twenty4 };;     // 2/2:s2>>24
+{ .mmi;        ld1     te22=[te22]             // 3/0:te0[s2>>8]
+       add     te20=te20,te0           // 3/2:te0+s0>>8
+       extr.u  te21=s1,8,8     }       // 3/3:s1>>8&0xff
+{ .mmi;        ld1     te23=[te23]             // 3/1:te0[s3>>8]
+       add     te00=te00,te0           // 3/0:te0+s0>>24
+       shr.u   te03=s3,twenty4 };;     // 3/3:s3>>24
+{ .mmi;        ld1     te20=[te20]             // 4/2:te0[s0>>8]
+       add     te21=te21,te0           // 4/3:te0+s2
+       extr.u  te11=s1,16,8    }       // 4/0:s1>>16&0xff
+{ .mmi;        ld1     te00=[te00]             // 4/0:te0[s0>>24]
+       add     te01=te01,te0           // 4/1:te0+s1>>24
+       shr.u   te13=s3,sixteen };;     // 4/2:s3>>16
+{ .mmi;        ld1     te21=[te21]             // 5/3:te0[s1>>8]
+       add     te11=te11,te0           // 5/0:te0+s1>>16
+       extr.u  te12=s2,16,8    }       // 5/1:s2>>16&0xff
+{ .mmi;        ld1     te01=[te01]             // 5/1:te0[s1>>24]
+       add     te02=te02,te0           // 5/2:te0+s2>>24
+       and     te31=s1,maskff  };;     // 5/2:s1&0xff
+{ .mmi;        ld1     te11=[te11]             // 6/0:te0[s1>>16]
+       add     te12=te12,te0           // 6/1:te0+s2>>16
+       extr.u  te10=s0,16,8    }       // 6/3:s0>>16&0xff
+{ .mmi;        ld1     te02=[te02]             // 6/2:te0[s2>>24]
+       add     te03=te03,te0           // 6/3:te0+s0>>16
+       and     te32=s2,maskff  };;     // 6/3:s2&0xff
 
+{ .mmi;        ld1     te12=[te12]             // 7/1:te0[s2>>16]
+       add     te31=te31,te0           // 7/2:te0+s1&0xff
+       dep     te33=te22,te33,8,8}     // 7/0:
+{ .mmi;        ld1     te03=[te03]             // 7/3:te0[s3>>24]
+       add     te32=te32,te0           // 7/3:te0+s2
+       and     te13=te13,maskff};;     // 7/2:s3>>16&0xff
+{ .mmi;        ld1     te31=[te31]             // 8/2:te0[s1]
+       add     te13=te13,te0           // 8/2:te0+s3>>16
+       dep     te30=te23,te30,8,8}     // 8/1:
+{ .mmi;        ld1     te32=[te32]             // 8/3:te0[s2]
+       add     te10=te10,te0           // 8/3:te0+s0>>16
+       shl     te00=te00,twenty4};;    // 8/0:
+{ .mii;        ld1     te13=[te13]             // 9/2:te0[s3>>16]
+       dep     te33=te11,te33,16,8     // 9/0:
+       shl     te01=te01,twenty4};;    // 9/1:
+{ .mii;        ld1     te10=[te10]             // 10/3:te0[s0>>16]
+       dep     te31=te20,te31,8,8      // 10/2:
+       shl     te02=te02,twenty4};;    // 10/2:
+{ .mii;        xor     t0=t0,te33              // 11/0:
+       dep     te32=te21,te32,8,8      // 11/3:
+       shl     te12=te12,sixteen};;    // 11/1:
+{ .mii;        xor     r16=t0,te00             // 12/0:done!
+       dep     te31=te13,te31,16,8     // 12/2:
+       shl     te03=te03,twenty4};;    // 12/3:
+{ .mmi;        xor     t1=t1,te01              // 13/1:
+       xor     t2=t2,te02              // 13/2:
+       dep     te32=te10,te32,16,8};;  // 13/3:
+{ .mmi;        xor     t1=t1,te30              // 14/1:
+       xor     r24=t2,te31             // 14/2:done!
+       xor     t3=t3,te32      };;     // 14/3:
+{ .mib;        xor     r20=t1,te12             // 15/1:done!
+       xor     r28=t3,te03             // 15/3:done!
+       br.ret.sptk     b6      };;
+.endp  _ia64_AES_encrypt#
+
+// void AES_encrypt (const void *in,void *out,const AES_KEY *key);
+.global        AES_encrypt#
+.proc  AES_encrypt#
 .align 32
-.Le_rounds_cexit:
-{ .mfi;        xor     te00=te00,s0            // "s0"
-       xor     te11=te11,s0    }
-{ .mfi;        xor     te22=te22,s0
-       xor     te33=te33,s0    }
-{ .mib;        xor     te01=te01,s1            // "s1"
-       xor     te12=te12,s1    }
-{ .mib;        xor     te23=te23,s1
-       xor     te30=te30,s1    }
-{ .mfi;        xor     te02=te02,s2            // "s2"
-       xor     te13=te13,s2    }
-{ .mfi;        xor     te20=te20,s2
-       xor     te31=te31,s2    }
-{ .mib;        xor     te03=te03,s3            // "s3"
-       xor     te10=te10,s3    }
-{ .mib;        xor     te21=te21,s3
-       xor     te32=te32,s3    };;
+AES_encrypt:
+       .prologue
+       .save   ar.pfs,pfssave
+{ .mmi;        alloc   pfssave=ar.pfs,3,1,12,0
+       and     out0=3,in0
+       mov     r3=ip                   }
+{ .mmi;        ADDP    in0=0,in0
+       mov     loc0=psr.um
+       ADDP    out11=KSZ*60,in2        };;     // &AES_KEY->rounds
+
+{ .mmi;        ld4     out11=[out11]                   // AES_KEY->rounds
+       add     out8=(AES_Te#-AES_encrypt#),r3  // Te0
+       .save   pr,prsave
+       mov     prsave=pr               }
+{ .mmi;        rum     1<<3                            // clear um.ac
+       .save   ar.lc,lcsave
+       mov     lcsave=ar.lc            };;
+
+       .body
+#if defined(_HPUX_SOURCE)      // HPUX is big-endian, cut 15+15 cycles...
+{ .mib; cmp.ne p6,p0=out0,r0
+       add     out0=4,in0
+(p6)   br.dpnt.many    .Le_i_unaligned };;
 
-{ .mii;        ADDP    r40=0,r44               // saved 2nd argument, snatch teN
-       extr.u  te22=te22,8,8
-       shr.u   te00=te00,twenty4       }//;;
-{ .mii;        ADDP    r41=1,r44
-       extr.u  te11=te11,16,8
-       shr.u   te01=te01,twenty4       }//;;
-{ .mii;        ADDP    r42=2,r44
-       extr.u  te23=te23,8,8
-       shr.u   te12=te12,sixteen       }//;;
-{ .mii;        ADDP    r43=3,r44
-       extr.u  te20=te20,8,8
-       shr.u   te02=te02,twenty4       };;
-{ .mii;        st1     [r43]=te33,4
-       extr.u  te13=te13,16,8
-       shr.u   te03=te03,twenty4       }//;;
-{ .mii;        st1     [r42]=te22,4
-       extr.u  te21=te21,8,8
-       shr.u   te10=te10,sixteen       }//;;
+{ .mmi;        ld4     out1=[in0],8            // s0
+       and     out9=3,in1
+       mov     twenty4=24              }
+{ .mmi;        ld4     out3=[out0],8           // s1
+       ADDP    rk0=0,in2
+       mov     sixteen=16              };;
+{ .mmi;        ld4     out5=[in0]              // s2
+       cmp.ne  p6,p0=out9,r0
+       mov     maskff=0xff             }
+{ .mmb;        ld4     out7=[out0]             // s3
+       ADDP    rk1=KSZ,in2
+       br.call.sptk.many       b6=_ia64_AES_encrypt    };;
 
-{ .mmi;        st1     [r41]=te11,4
-       st1     [r40]=te00,4            };;
-{ .mmi;        st1     [r43]=te30,4
-       st1     [r42]=te23,4            }//;;
-{ .mmi;        st1     [r41]=te12,4
-       st1     [r40]=te01,4            };;
-{ .mmi;        st1     [r43]=te31,4
-       st1     [r42]=te20,4            }//;;
-{ .mmi;        st1     [r41]=te13,4
-       st1     [r40]=te02,4
+{ .mib;        ADDP    in0=4,in1
+       ADDP    in1=0,in1
+(p6)   br.spnt .Le_o_unaligned         };;
+
+{ .mii;        mov     psr.um=loc0
+       mov     ar.pfs=pfssave
+       mov     ar.lc=lcsave            };;
+{ .mmi;        st4     [in1]=r16,8             // s0
+       st4     [in0]=r20,8             // s1
        mov     pr=prsave,0x1ffff       };;
-{ .mmi;        st1     [r43]=te32
-       st1     [r42]=te21
-       mov     ar.lc=r3                }//;;
-{ .mmb;        st1     [r41]=te10
-       st1     [r40]=te03
+{ .mmb;        st4     [in1]=r24               // s2
+       st4     [in0]=r28               // s3
+       br.ret.sptk.many        b0      };;
+#endif
+
+.align 32
+.Le_i_unaligned:
+{ .mmi;        add     out0=1,in0
+       add     out2=2,in0
+       add     out4=3,in0      };;
+{ .mmi;        ld1     r16=[in0],4
+       ld1     r17=[out0],4    }//;;
+{ .mmi;        ld1     r18=[out2],4
+       ld1     out1=[out4],4   };;     // s0
+{ .mmi;        ld1     r20=[in0],4
+       ld1     r21=[out0],4    }//;;
+{ .mmi;        ld1     r22=[out2],4
+       ld1     out3=[out4],4   };;     // s1
+{ .mmi;        ld1     r24=[in0],4
+       ld1     r25=[out0],4    }//;;
+{ .mmi;        ld1     r26=[out2],4
+       ld1     out5=[out4],4   };;     // s2
+{ .mmi;        ld1     r28=[in0]
+       ld1     r29=[out0]      }//;;
+{ .mmi;        ld1     r30=[out2]
+       ld1     out7=[out4]     };;     // s3
+
+{ .mii;
+       dep     out1=r16,out1,24,8      //;;
+       dep     out3=r20,out3,24,8      }//;;
+{ .mii;        ADDP    rk0=0,in2
+       dep     out5=r24,out5,24,8      //;;
+       dep     out7=r28,out7,24,8      };;
+{ .mii;        ADDP    rk1=KSZ,in2
+       dep     out1=r17,out1,16,8      //;;
+       dep     out3=r21,out3,16,8      }//;;
+{ .mii;        mov     twenty4=24
+       dep     out5=r25,out5,16,8      //;;
+       dep     out7=r29,out7,16,8      };;
+{ .mii;        mov     sixteen=16
+       dep     out1=r18,out1,8,8       //;;
+       dep     out3=r22,out3,8,8       }//;;
+{ .mii;        mov     maskff=0xff
+       dep     out5=r26,out5,8,8       //;;
+       dep     out7=r30,out7,8,8       };;
+
+{ .mib;        br.call.sptk.many       b6=_ia64_AES_encrypt    };;
+
+.Le_o_unaligned:
+{ .mii;        ADDP    out0=0,in1
+       extr.u  r17=r16,8,8                     // s0
+       shr.u   r19=r16,twenty4         }//;;
+{ .mii;        ADDP    out1=1,in1
+       extr.u  r18=r16,16,8
+       shr.u   r23=r20,twenty4         }//;;   // s1
+{ .mii;        ADDP    out2=2,in1
+       extr.u  r21=r20,8,8
+       shr.u   r22=r20,sixteen         }//;;
+{ .mii;        ADDP    out3=3,in1
+       extr.u  r25=r24,8,8                     // s2
+       shr.u   r27=r24,twenty4         };;
+{ .mii;        st1     [out3]=r16,4
+       extr.u  r26=r24,16,8
+       shr.u   r31=r28,twenty4         }//;;   // s3
+{ .mii;        st1     [out2]=r17,4
+       extr.u  r29=r28,8,8
+       shr.u   r30=r28,sixteen         }//;;
+
+{ .mmi;        st1     [out1]=r18,4
+       st1     [out0]=r19,4            };;
+{ .mmi;        st1     [out3]=r20,4
+       st1     [out2]=r21,4            }//;;
+{ .mmi;        st1     [out1]=r22,4
+       st1     [out0]=r23,4            };;
+{ .mmi;        st1     [out3]=r24,4
+       st1     [out2]=r25,4
+       mov     pr=prsave,0x1ffff       }//;;
+{ .mmi;        st1     [out1]=r26,4
+       st1     [out0]=r27,4
+       mov     ar.pfs=pfssave          };;
+{ .mmi;        st1     [out3]=r28
+       st1     [out2]=r29
+       mov     ar.lc=lcsave            }//;;
+{ .mmi;        st1     [out1]=r30
+       st1     [out0]=r31              }
+{ .mfb;        mov     psr.um=loc0                     // restore user mask
        br.ret.sptk.many        b0      };;
 .endp  AES_encrypt#
 
-// AES_decrypt is autogenerated by the following script:
+// *AES_decrypt are autogenerated by the following script:
 #if 0
 #!/usr/bin/env perl
-print "// AES_decrypt is autogenerated by the following script:\n#if 0\n";
+print "// *AES_decrypt are autogenerated by the following script:\n#if 0\n";
 open(PROG,'<'.$0); while(<PROG>) { print; } close(PROG);
 print "#endif\n";
 while(<>) {
-       $process=1      if (/\.global\s+AES_encrypt/);
+       $process=1      if (/\.proc\s+_ia64_AES_encrypt/);
        next            if (!$process);
 
        #s/te00=s0/td00=s0/;    s/te00/td00/g;
@@ -339,126 +446,47 @@ while(<>) {
 
        s/AES_encrypt/AES_decrypt/g;
        s/\.Le_/.Ld_/g;
-       s/Te0#/Td0#/g;
+       s/AES_Te#/AES_Td#/g;
 
        print;
 
        exit            if (/\.endp\s+AES_decrypt/);
 }
 #endif
-.global        AES_decrypt#
-.proc  AES_decrypt#
+.proc  _ia64_AES_decrypt#
+// Input:      rk0-rk1
+//             te0
+//             te3     as AES_KEY->rounds!!!
+//             s0-s3
+//             maskff,twenty4,sixteen
+// Output:     r16,r20,r24,r28 as s0-s3
+// Clobber:    r16-r31,rk0-rk1,r32-r43
 .align 32
-#if !defined(_HPUX_SOURCE)
-.skip  16
-#endif
-AES_decrypt:
+_ia64_AES_decrypt:
        .prologue
-       .fframe 0
-       .save   ar.pfs,r2
-       .save   ar.lc,r3
-{ .mii;        alloc   r2=ar.pfs,3,10,0,8
-       mov     r3=ar.lc        
-       mov     prsave=pr       };;
-
+       .altrp  b6
        .body
-{ .mmi;        and     r40=3,r32
-       ADDP    r32=0,r32
-       mov     pr.rot=7<<16    };;
-#if defined(_HPUX_SOURCE)      // HPUX is big-endian, cut 15 cycles...
-{ .mib; cmp.ne p6,p0=r40,r0
-       add     r41=4,r32               // 1st arg, borrow teN
-(p6)   br.dpnt.many    .Ld_unaligned   };;
-
-{ .mmi;        ld4     r19=[r32],8
-       mov     r44=r33                 // save 2nd arg
-       mov     twenty4=24      }
-{ .mmi;        ld4     r23=[r41],8
-       addl    te0=@ltoff(Td0#),gp
-       ADDP    r35=KSZ*60,r34  };;     // &AES_KEY->rounds, borrow s1
-{ .mmi;        ld8     te0=[te0]
-       ld4     r35=[r35]               // AES_KEY->rounds
-       ADDP    rk0=0,r34       }//;;   // 3rd arg
-{ .mmi;        ld4     r27=[r32]
-       ld4     r31=[r41]               
-       ADDP    rk1=KSZ,r34     };;
-
-{ .mfi; LDKEY  t0=[rk0],2*KSZ
-       mov     sixteen=16      }
-{ .mfi;        LDKEY   t1=[rk1],2*KSZ
-       mov     maskff=0xff     };;
-{ .mfi;        LDKEY   t2=[rk0],2*KSZ
-       add     te1=1024,te0    }
+{ .mmi;        alloc   r16=ar.pfs,12,0,0,8
+       LDKEY   t0=[rk0],2*KSZ
+       mov     pr.rot=1<<16    }
+{ .mmi;        LDKEY   t1=[rk1],2*KSZ
+       add     te1=TE1,te0
+       add     te3=-3,te3      };;
+{ .mib;        LDKEY   t2=[rk0],2*KSZ
+       mov     ar.ec=2         }
 { .mib;        LDKEY   t3=[rk1],2*KSZ
-       add     te2=2048,te0
-       br.many .Ld_common      };;
-#endif
-.Ld_unaligned:
-{ .mfi;        ADDP    r40=0,r32               // 1st arg, borrow teN
-       ADDP    r41=1,r32       }
-{ .mfi;        ADDP    r42=2,r32
-       ADDP    r43=3,r32       };;
-{ .mmi;        ld1     r16=[r40],4
-       ld1     r17=[r41],4
-       mov     r44=r33         }//;;   // save 2nd arg
-{ .mmi;        ld1     r18=[r42],4
-       ld1     r19=[r43],4
-       ADDP    rk0=0,r34       };;     // 3rd arg
-{ .mmi;        ld1     r20=[r40],4
-       ld1     r21=[r41],4
-       ADDP    rk1=KSZ,r34     }//;;
-{ .mmi;        ld1     r22=[r42],4
-       ld1     r23=[r43],4
-       ADDP    r35=KSZ*60,r34  };;     // &AES_KEY->rounds, borrow s1
-{ .mmi;        ld1     r24=[r40],4
-       ld1     r25=[r41],4
-       mov     twenty4=24      }//;;
-{ .mmi;        ld1     r26=[r42],4
-       ld1     r27=[r43],4
-       mov     sixteen=16      };;
-{ .mmi;        ld1     r28=[r40]
-       ld1     r29=[r41]
-       mov     maskff=0xff     }//;;
-{ .mmi;        ld1     r30=[r42]
-       ld1     r31=[r43]
-       addl    te0=@ltoff(Td0#),gp     };;     // that was close...
-
-{ .mii;        ld8     te0=[te0]
-       dep     r19=r16,r19,24,8        //;;
-       dep     r23=r20,r23,24,8        }//;;
-{ .mii;        ld4     r35=[r35]               // AES_KEY->rounds
-       dep     r27=r24,r27,24,8        //;;
-       dep     r31=r28,r31,24,8        };;
-{ .mii;        LDKEY   t0=[rk0],2*KSZ
-       dep     r19=r17,r19,16,8        //;;
-       dep     r23=r21,r23,16,8        }//;;
-{ .mii;        LDKEY   t1=[rk1],2*KSZ
-       dep     r27=r25,r27,16,8        //;;
-       dep     r31=r29,r31,16,8        };;
-{ .mii;        LDKEY   t2=[rk0],2*KSZ
-       dep     r19=r18,r19,8,8         //;;
-       dep     r23=r22,r23,8,8         }//;;
-{ .mii;        LDKEY   t3=[rk1],2*KSZ
-       dep     r27=r26,r27,8,8         //;;
-       dep     r31=r30,r31,8,8         };;
-       
-{ .mib;        add     te1=1024,te0
-       add     te2=2048,te0    }
-.Ld_common:
-{ .mib; add    te3=3072,te0
-       add     r35=-3,r35
-       brp.exit.imp    .Ld_rounds_cexit,.Ld_cexit_insn
-                               };;
-{ .mii;        mov     ar.lc=r35               // borrowed s1
-       mov     ar.ec=3         };;
-
-{ .mfi;        xor     s0=r19,t0
-       xor     s1=r23,t1       }
-{ .mfi;        xor     s2=r27,t2
-       xor     s3=r31,t3       };;
+       add     te2=TE2,te0
+       brp.loop.imp    .Ld_top,.Ld_end-16      };;
 
+{ .mmi;        xor     s0=s0,t0
+       xor     s1=s1,t1
+       mov     ar.lc=te3       }
+{ .mmi;        xor     s2=s2,t2
+       xor     s3=s3,t3
+       add     te3=TE3,te0     };;
+\f
 .align 32
-.Ld_rounds:
+.Ld_top:
 { .mmi;        (p0)    LDKEY   t0=[rk0],2*KSZ          // 0/0:rk[0]
        (p0)    and     te31=s1,maskff          // 0/0:s3&0xff
        (p0)    extr.u  te22=s2,8,8     }       // 0/0:s2>>8&0xff
@@ -466,789 +494,630 @@ AES_decrypt:
        (p0)    and     te32=s2,maskff          // 0/1:s0&0xff
        (p0)    shr.u   te00=s0,twenty4 };;     // 0/0:s0>>24
 { .mmi;        (p0)    LDKEY   t2=[rk0],2*KSZ          // 1/2:rk[2]
-       (p0)    shladd  te31=te31,2,te3         // 1/0:te0+s0>>24
+       (p0)    shladd  te31=te31,3,te3         // 1/0:te0+s0>>24
        (p0)    extr.u  te23=s3,8,8     }       // 1/1:s3>>8&0xff
 { .mmi;        (p0)    LDKEY   t3=[rk1],2*KSZ          // 1/3:rk[3]
-       (p0)    shladd  te32=te32,2,te3         // 1/1:te3+s0
+       (p0)    shladd  te32=te32,3,te3         // 1/1:te3+s0
        (p0)    shr.u   te01=s1,twenty4 };;     // 1/1:s1>>24
 { .mmi;        (p0)    ld4     te31=[te31]             // 2/0:te3[s3&0xff]
-       (p0)    shladd  te22=te22,2,te2         // 2/0:te2+s2>>8&0xff
+       (p0)    shladd  te22=te22,3,te2         // 2/0:te2+s2>>8&0xff
        (p0)    extr.u  te20=s0,8,8     }       // 2/2:s0>>8&0xff
 { .mmi;        (p0)    ld4     te32=[te32]             // 2/1:te3[s0]
-       (p0)    shladd  te23=te23,2,te2         // 2/1:te2+s3>>8
+       (p0)    shladd  te23=te23,3,te2         // 2/1:te2+s3>>8
        (p0)    shr.u   te02=s2,twenty4 };;     // 2/2:s2>>24
 { .mmi;        (p0)    ld4     te22=[te22]             // 3/0:te2[s2>>8]
-       (p0)    shladd  te20=te20,2,te2         // 3/2:te2+s0>>8
+       (p0)    shladd  te20=te20,3,te2         // 3/2:te2+s0>>8
        (p0)    extr.u  te21=s1,8,8     }       // 3/3:s1>>8&0xff
 { .mmi;        (p0)    ld4     te23=[te23]             // 3/1:te2[s3>>8]
-       (p0)    shladd  te00=te00,2,te0         // 3/0:te0+s0>>24
+       (p0)    shladd  te00=te00,3,te0         // 3/0:te0+s0>>24
        (p0)    shr.u   te03=s3,twenty4 };;     // 3/3:s3>>24
 { .mmi;        (p0)    ld4     te20=[te20]             // 4/2:te2[s0>>8]
-       (p0)    shladd  te21=te21,2,te2         // 4/3:te3+s2
+       (p0)    shladd  te21=te21,3,te2         // 4/3:te3+s2
        (p0)    extr.u  te13=s3,16,8    }       // 4/0:s1>>16&0xff
 { .mmi;        (p0)    ld4     te00=[te00]             // 4/0:te0[s0>>24]
-       (p0)    shladd  te01=te01,2,te0         // 4/1:te0+s1>>24
+       (p0)    shladd  te01=te01,3,te0         // 4/1:te0+s1>>24
        (p0)    shr.u   te11=s1,sixteen };;     // 4/2:s3>>16
 { .mmi;        (p0)    ld4     te21=[te21]             // 5/3:te2[s1>>8]
-       (p0)    shladd  te13=te13,2,te1         // 5/0:te1+s1>>16
+       (p0)    shladd  te13=te13,3,te1         // 5/0:te1+s1>>16
        (p0)    extr.u  te10=s0,16,8    }       // 5/1:s2>>16&0xff
 { .mmi;        (p0)    ld4     te01=[te01]             // 5/1:te0[s1>>24]
-       (p0)    shladd  te02=te02,2,te0         // 5/2:te0+s2>>24
+       (p0)    shladd  te02=te02,3,te0         // 5/2:te0+s2>>24
        (p0)    and     te33=s3,maskff  };;     // 5/2:s1&0xff
-
 { .mmi;        (p0)    ld4     te13=[te13]             // 6/0:te1[s1>>16]
-       (p0)    shladd  te10=te10,2,te1         // 6/1:te1+s2>>16
+       (p0)    shladd  te10=te10,3,te1         // 6/1:te1+s2>>16
        (p0)    extr.u  te12=s2,16,8    }       // 6/3:s0>>16&0xff
 { .mmi;        (p0)    ld4     te02=[te02]             // 6/2:te0[s2>>24]
-       (p0)    shladd  te03=te03,2,te0         // 6/3:te1+s0>>16
+       (p0)    shladd  te03=te03,3,te0         // 6/3:te1+s0>>16
        (p0)    and     te30=s0,maskff  };;     // 6/3:s2&0xff
+
 { .mmi;        (p0)    ld4     te10=[te10]             // 7/1:te1[s2>>16]
-       (p0)    shladd  te33=te33,2,te3         // 7/2:te3+s1&0xff
+       (p0)    shladd  te33=te33,3,te3         // 7/2:te3+s1&0xff
        (p0)    and     te11=te11,maskff}       // 7/2:s3>>16&0xff
 { .mmi;        (p0)    ld4     te03=[te03]             // 7/3:te0[s3>>24]
-       (p0)    shladd  te30=te30,2,te3         // 7/3:te3+s2
-       (p16)   cmp.eq  p0,p18=r0,r0    };;     // 7/clear (p18)
+       (p0)    shladd  te30=te30,3,te3         // 7/3:te3+s2
+       (p0)    xor     t0=t0,te31      };;     // 7/0:
 { .mmi;        (p0)    ld4     te33=[te33]             // 8/2:te3[s1]
-       (p0)    shladd  te11=te11,2,te1         // 8/2:te1+s3>>16
-       (p17)   xor     t0=t0,te31      }       // 8/0:
+       (p0)    shladd  te11=te11,3,te1         // 8/2:te1+s3>>16
+       (p0)    xor     t0=t0,te22      }       // 8/0:
 { .mmi;        (p0)    ld4     te30=[te30]             // 8/3:te3[s2]
-       (p0)    shladd  te12=te12,2,te1         // 8/3:te1+s0>>16
-       (p17)   xor     t1=t1,te32      };;     // 8/1:
+       (p0)    shladd  te12=te12,3,te1         // 8/3:te1+s0>>16
+       (p0)    xor     t1=t1,te32      };;     // 8/1:
 { .mmi;        (p0)    ld4     te11=[te11]             // 9/2:te1[s3>>16]
-       (p17)   xor     t0=t0,te22              // 9/0:
-       (p18)   add     te0=4096,te0    }       // 9/
-.Ld_cexit_insn:
-{ .mmb;        (p0)    ld4     te12=[te12]             // 9/3:te1[s0>>16]
-       (p17)   xor     t1=t1,te23              // 9/1:
-       br.cexit.spnt.few       .Ld_rounds_cexit
-                                       };;
-{ .mmi;        (p18)   xor     s2=s2,te20              // 10/2:
-       (p18)   xor     s0=s0,te00              // 10/0:
-       (p19)   add     te1=3072,te1    }       // 10/
-{ .mmi;        (p18)   xor     s3=s3,te21              // 10/3:
-       (p18)   xor     s1=s1,te01              // 10/1:
-       (p19)   add     te2=2048,te2    };;     // 10/
-{ .mfi;        (p18)   xor     s0=s0,te13              // 11/0:done!
-       (p18)   xor     s2=s2,te02      }       // 11/2:
-{ .mfi;        (p18)   xor     s1=s1,te10              // 11/1:done!
-       (p18)   xor     s3=s3,te03      };;     // 11/3:
-{ .mmi;        (p18)   xor     s2=s2,te33              // 12/2:
-       (p18)   xor     s3=s3,te30              // 12/3:
-       (p19)   add     te3=1024,te3    };;     // 12/
-{ .mib;        (p18)   xor     s2=s2,te11              // 13/2:done!
-       (p18)   xor     s3=s3,te12              // 13/3:done!
-       br      .Ld_rounds      };;
+       (p0)    ld4     te12=[te12]             // 9/3:te1[s0>>16]
+       (p0)    xor     t0=t0,te00      };;     // 9/0:         !L2 scheduling
+{ .mmi;        (p0)    xor     t1=t1,te23              // 10[9]/1:     
+       (p0)    xor     t2=t2,te20              // 10[9]/2:
+       (p0)    xor     t3=t3,te21      };;     // 10[9]/3:
+{ .mmi;        (p0)    xor     t0=t0,te13              // 11[10]/0:done!
+       (p0)    xor     t1=t1,te01              // 11[10]/1:
+       (p0)    xor     t2=t2,te02      };;     // 11[10]/2:    !L2 scheduling
+{ .mmi;        (p0)    xor     t3=t3,te03              // 12[10]/3:
+       (p16)   cmp.eq  p0,p17=r0,r0    };;     // 12[10]/clear (p17)
+{ .mmi;        (p0)    xor     t1=t1,te10              // 13[11]/1:done!
+       (p0)    xor     t2=t2,te33              // 13[11]/2:
+       (p0)    xor     t3=t3,te30      }       // 13[11]/3:
+{ .mmi;        (p17)   add     te0=2048,te0            // 13[11]/
+       (p17)   add     te1=2048+64-TE1,te1};;  // 13[11]/
+{ .mib;        (p0)    xor     t2=t2,te11              // 14[12]/2:done!
+       (p17)   add     te2=2048+128-TE2,te2}   // 14[12]/
+{ .mib;        (p0)    xor     t3=t3,te12              // 14[12]/3:done!
+       (p17)   add     te3=2048+192-TE3,te3    // 14[12]/
+       br.ctop.sptk    .Ld_top         };;
+.Ld_end:
+\f
+
+{ .mmi;        ld8     te10=[te0]              // prefetch Td4
+       ld8     te33=[te1]      }
+{ .mmi;        ld8     te12=[te2]
+       ld8     te30=[te3]      }
 
+{ .mmi;        LDKEY   t0=[rk0],2*KSZ          // 0/0:rk[0]
+       and     te31=s1,maskff          // 0/0:s3&0xff
+       extr.u  te22=s2,8,8     }       // 0/0:s2>>8&0xff
+{ .mmi; LDKEY  t1=[rk1],2*KSZ          // 0/1:rk[1]
+       and     te32=s2,maskff          // 0/1:s0&0xff
+       shr.u   te00=s0,twenty4 };;     // 0/0:s0>>24
+{ .mmi;        LDKEY   t2=[rk0],2*KSZ          // 1/2:rk[2]
+       add     te31=te31,te0           // 1/0:te0+s0>>24
+       extr.u  te23=s3,8,8     }       // 1/1:s3>>8&0xff
+{ .mmi;        LDKEY   t3=[rk1],2*KSZ          // 1/3:rk[3]
+       add     te32=te32,te0           // 1/1:te0+s0
+       shr.u   te01=s1,twenty4 };;     // 1/1:s1>>24
+{ .mmi;        ld1     te31=[te31]             // 2/0:te0[s3&0xff]
+       add     te22=te22,te0           // 2/0:te0+s2>>8&0xff
+       extr.u  te20=s0,8,8     }       // 2/2:s0>>8&0xff
+{ .mmi;        ld1     te32=[te32]             // 2/1:te0[s0]
+       add     te23=te23,te0           // 2/1:te0+s3>>8
+       shr.u   te02=s2,twenty4 };;     // 2/2:s2>>24
+{ .mmi;        ld1     te22=[te22]             // 3/0:te0[s2>>8]
+       add     te20=te20,te0           // 3/2:te0+s0>>8
+       extr.u  te21=s1,8,8     }       // 3/3:s1>>8&0xff
+{ .mmi;        ld1     te23=[te23]             // 3/1:te0[s3>>8]
+       add     te00=te00,te0           // 3/0:te0+s0>>24
+       shr.u   te03=s3,twenty4 };;     // 3/3:s3>>24
+{ .mmi;        ld1     te20=[te20]             // 4/2:te0[s0>>8]
+       add     te21=te21,te0           // 4/3:te0+s2
+       extr.u  te13=s3,16,8    }       // 4/0:s1>>16&0xff
+{ .mmi;        ld1     te00=[te00]             // 4/0:te0[s0>>24]
+       add     te01=te01,te0           // 4/1:te0+s1>>24
+       shr.u   te11=s1,sixteen };;     // 4/2:s3>>16
+{ .mmi;        ld1     te21=[te21]             // 5/3:te0[s1>>8]
+       add     te13=te13,te0           // 5/0:te0+s1>>16
+       extr.u  te10=s0,16,8    }       // 5/1:s2>>16&0xff
+{ .mmi;        ld1     te01=[te01]             // 5/1:te0[s1>>24]
+       add     te02=te02,te0           // 5/2:te0+s2>>24
+       and     te33=s3,maskff  };;     // 5/2:s1&0xff
+{ .mmi;        ld1     te13=[te13]             // 6/0:te0[s1>>16]
+       add     te10=te10,te0           // 6/1:te0+s2>>16
+       extr.u  te12=s2,16,8    }       // 6/3:s0>>16&0xff
+{ .mmi;        ld1     te02=[te02]             // 6/2:te0[s2>>24]
+       add     te03=te03,te0           // 6/3:te0+s0>>16
+       and     te30=s0,maskff  };;     // 6/3:s2&0xff
+
+{ .mmi;        ld1     te10=[te10]             // 7/1:te0[s2>>16]
+       add     te33=te33,te0           // 7/2:te0+s1&0xff
+       dep     te31=te22,te31,8,8}     // 7/0:
+{ .mmi;        ld1     te03=[te03]             // 7/3:te0[s3>>24]
+       add     te30=te30,te0           // 7/3:te0+s2
+       and     te11=te11,maskff};;     // 7/2:s3>>16&0xff
+{ .mmi;        ld1     te33=[te33]             // 8/2:te0[s1]
+       add     te11=te11,te0           // 8/2:te0+s3>>16
+       dep     te32=te23,te32,8,8}     // 8/1:
+{ .mmi;        ld1     te30=[te30]             // 8/3:te0[s2]
+       add     te12=te12,te0           // 8/3:te0+s0>>16
+       shl     te00=te00,twenty4};;    // 8/0:
+{ .mii;        ld1     te11=[te11]             // 9/2:te0[s3>>16]
+       dep     te31=te13,te31,16,8     // 9/0:
+       shl     te01=te01,twenty4};;    // 9/1:
+{ .mii;        ld1     te12=[te12]             // 10/3:te0[s0>>16]
+       dep     te33=te20,te33,8,8      // 10/2:
+       shl     te02=te02,twenty4};;    // 10/2:
+{ .mii;        xor     t0=t0,te31              // 11/0:
+       dep     te30=te21,te30,8,8      // 11/3:
+       shl     te10=te10,sixteen};;    // 11/1:
+{ .mii;        xor     r16=t0,te00             // 12/0:done!
+       dep     te33=te11,te33,16,8     // 12/2:
+       shl     te03=te03,twenty4};;    // 12/3:
+{ .mmi;        xor     t1=t1,te01              // 13/1:
+       xor     t2=t2,te02              // 13/2:
+       dep     te30=te12,te30,16,8};;  // 13/3:
+{ .mmi;        xor     t1=t1,te32              // 14/1:
+       xor     r24=t2,te33             // 14/2:done!
+       xor     t3=t3,te30      };;     // 14/3:
+{ .mib;        xor     r20=t1,te10             // 15/1:done!
+       xor     r28=t3,te03             // 15/3:done!
+       br.ret.sptk     b6      };;
+.endp  _ia64_AES_decrypt#
+
+// void AES_decrypt (const void *in,void *out,const AES_KEY *key);
+.global        AES_decrypt#
+.proc  AES_decrypt#
 .align 32
-.Ld_rounds_cexit:
-{ .mfi;        xor     te00=te00,s0            // "s0"
-       xor     te13=te13,s0    }
-{ .mfi;        xor     te22=te22,s0
-       xor     te31=te31,s0    }
-{ .mib;        xor     te01=te01,s1            // "s1"
-       xor     te10=te10,s1    }
-{ .mib;        xor     te23=te23,s1
-       xor     te32=te32,s1    }
-{ .mfi;        xor     te02=te02,s2            // "s2"
-       xor     te11=te11,s2    }
-{ .mfi;        xor     te20=te20,s2
-       xor     te33=te33,s2    }
-{ .mib;        xor     te03=te03,s3            // "s3"
-       xor     te12=te12,s3    }
-{ .mib;        xor     te21=te21,s3
-       xor     te30=te30,s3    };;
+AES_decrypt:
+       .prologue
+       .save   ar.pfs,pfssave
+{ .mmi;        alloc   pfssave=ar.pfs,3,1,12,0
+       and     out0=3,in0
+       mov     r3=ip                   }
+{ .mmi;        ADDP    in0=0,in0
+       mov     loc0=psr.um
+       ADDP    out11=KSZ*60,in2        };;     // &AES_KEY->rounds
+
+{ .mmi;        ld4     out11=[out11]                   // AES_KEY->rounds
+       add     out8=(AES_Td#-AES_decrypt#),r3  // Te0
+       .save   pr,prsave
+       mov     prsave=pr               }
+{ .mmi;        rum     1<<3                            // clear um.ac
+       .save   ar.lc,lcsave
+       mov     lcsave=ar.lc            };;
+
+       .body
+#if defined(_HPUX_SOURCE)      // HPUX is big-endian, cut 15+15 cycles...
+{ .mib; cmp.ne p6,p0=out0,r0
+       add     out0=4,in0
+(p6)   br.dpnt.many    .Ld_i_unaligned };;
 
-{ .mii;        ADDP    r40=0,r44               // saved 2nd argument, snatch teN
-       extr.u  te22=te22,8,8
-       shr.u   te00=te00,twenty4       }//;;
-{ .mii;        ADDP    r41=1,r44
-       extr.u  te13=te13,16,8
-       shr.u   te01=te01,twenty4       }//;;
-{ .mii;        ADDP    r42=2,r44
-       extr.u  te23=te23,8,8
-       shr.u   te10=te10,sixteen       }//;;
-{ .mii;        ADDP    r43=3,r44
-       extr.u  te20=te20,8,8
-       shr.u   te02=te02,twenty4       };;
-{ .mii;        st1     [r43]=te31,4
-       extr.u  te11=te11,16,8
-       shr.u   te03=te03,twenty4       }//;;
-{ .mii;        st1     [r42]=te22,4
-       extr.u  te21=te21,8,8
-       shr.u   te12=te12,sixteen       }//;;
+{ .mmi;        ld4     out1=[in0],8            // s0
+       and     out9=3,in1
+       mov     twenty4=24              }
+{ .mmi;        ld4     out3=[out0],8           // s1
+       ADDP    rk0=0,in2
+       mov     sixteen=16              };;
+{ .mmi;        ld4     out5=[in0]              // s2
+       cmp.ne  p6,p0=out9,r0
+       mov     maskff=0xff             }
+{ .mmb;        ld4     out7=[out0]             // s3
+       ADDP    rk1=KSZ,in2
+       br.call.sptk.many       b6=_ia64_AES_decrypt    };;
 
-{ .mmi;        st1     [r41]=te13,4
-       st1     [r40]=te00,4            };;
-{ .mmi;        st1     [r43]=te32,4
-       st1     [r42]=te23,4            }//;;
-{ .mmi;        st1     [r41]=te10,4
-       st1     [r40]=te01,4            };;
-{ .mmi;        st1     [r43]=te33,4
-       st1     [r42]=te20,4            }//;;
-{ .mmi;        st1     [r41]=te11,4
-       st1     [r40]=te02,4
+{ .mib;        ADDP    in0=4,in1
+       ADDP    in1=0,in1
+(p6)   br.spnt .Ld_o_unaligned         };;
+
+{ .mii;        mov     psr.um=loc0
+       mov     ar.pfs=pfssave
+       mov     ar.lc=lcsave            };;
+{ .mmi;        st4     [in1]=r16,8             // s0
+       st4     [in0]=r20,8             // s1
        mov     pr=prsave,0x1ffff       };;
-{ .mmi;        st1     [r43]=te30
-       st1     [r42]=te21
-       mov     ar.lc=r3                }//;;
-{ .mmb;        st1     [r41]=te12
-       st1     [r40]=te03
+{ .mmb;        st4     [in1]=r24               // s2
+       st4     [in0]=r28               // s3
+       br.ret.sptk.many        b0      };;
+#endif
+
+.align 32
+.Ld_i_unaligned:
+{ .mmi;        add     out0=1,in0
+       add     out2=2,in0
+       add     out4=3,in0      };;
+{ .mmi;        ld1     r16=[in0],4
+       ld1     r17=[out0],4    }//;;
+{ .mmi;        ld1     r18=[out2],4
+       ld1     out1=[out4],4   };;     // s0
+{ .mmi;        ld1     r20=[in0],4
+       ld1     r21=[out0],4    }//;;
+{ .mmi;        ld1     r22=[out2],4
+       ld1     out3=[out4],4   };;     // s1
+{ .mmi;        ld1     r24=[in0],4
+       ld1     r25=[out0],4    }//;;
+{ .mmi;        ld1     r26=[out2],4
+       ld1     out5=[out4],4   };;     // s2
+{ .mmi;        ld1     r28=[in0]
+       ld1     r29=[out0]      }//;;
+{ .mmi;        ld1     r30=[out2]
+       ld1     out7=[out4]     };;     // s3
+
+{ .mii;
+       dep     out1=r16,out1,24,8      //;;
+       dep     out3=r20,out3,24,8      }//;;
+{ .mii;        ADDP    rk0=0,in2
+       dep     out5=r24,out5,24,8      //;;
+       dep     out7=r28,out7,24,8      };;
+{ .mii;        ADDP    rk1=KSZ,in2
+       dep     out1=r17,out1,16,8      //;;
+       dep     out3=r21,out3,16,8      }//;;
+{ .mii;        mov     twenty4=24
+       dep     out5=r25,out5,16,8      //;;
+       dep     out7=r29,out7,16,8      };;
+{ .mii;        mov     sixteen=16
+       dep     out1=r18,out1,8,8       //;;
+       dep     out3=r22,out3,8,8       }//;;
+{ .mii;        mov     maskff=0xff
+       dep     out5=r26,out5,8,8       //;;
+       dep     out7=r30,out7,8,8       };;
+
+{ .mib;        br.call.sptk.many       b6=_ia64_AES_decrypt    };;
+
+.Ld_o_unaligned:
+{ .mii;        ADDP    out0=0,in1
+       extr.u  r17=r16,8,8                     // s0
+       shr.u   r19=r16,twenty4         }//;;
+{ .mii;        ADDP    out1=1,in1
+       extr.u  r18=r16,16,8
+       shr.u   r23=r20,twenty4         }//;;   // s1
+{ .mii;        ADDP    out2=2,in1
+       extr.u  r21=r20,8,8
+       shr.u   r22=r20,sixteen         }//;;
+{ .mii;        ADDP    out3=3,in1
+       extr.u  r25=r24,8,8                     // s2
+       shr.u   r27=r24,twenty4         };;
+{ .mii;        st1     [out3]=r16,4
+       extr.u  r26=r24,16,8
+       shr.u   r31=r28,twenty4         }//;;   // s3
+{ .mii;        st1     [out2]=r17,4
+       extr.u  r29=r28,8,8
+       shr.u   r30=r28,sixteen         }//;;
+
+{ .mmi;        st1     [out1]=r18,4
+       st1     [out0]=r19,4            };;
+{ .mmi;        st1     [out3]=r20,4
+       st1     [out2]=r21,4            }//;;
+{ .mmi;        st1     [out1]=r22,4
+       st1     [out0]=r23,4            };;
+{ .mmi;        st1     [out3]=r24,4
+       st1     [out2]=r25,4
+       mov     pr=prsave,0x1ffff       }//;;
+{ .mmi;        st1     [out1]=r26,4
+       st1     [out0]=r27,4
+       mov     ar.pfs=pfssave          };;
+{ .mmi;        st1     [out3]=r28
+       st1     [out2]=r29
+       mov     ar.lc=lcsave            }//;;
+{ .mmi;        st1     [out1]=r30
+       st1     [out0]=r31              }
+{ .mfb;        mov     psr.um=loc0                     // restore user mask
        br.ret.sptk.many        b0      };;
 .endp  AES_decrypt#
 
 // leave it in .text segment...
 .align 64
-.type  Te0#,@object
-Te0:   data4   0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
-       data4   0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
-       data4   0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
-       data4   0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
-       data4   0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
-       data4   0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
-       data4   0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
-       data4   0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
-       data4   0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
-       data4   0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
-       data4   0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
-       data4   0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
-       data4   0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
-       data4   0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
-       data4   0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
-       data4   0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
-       data4   0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
-       data4   0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
-       data4   0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
-       data4   0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
-       data4   0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
-       data4   0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
-       data4   0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
-       data4   0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
-       data4   0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
-       data4   0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
-       data4   0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
-       data4   0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
-       data4   0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
-       data4   0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
-       data4   0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
-       data4   0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
-       data4   0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
-       data4   0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
-       data4   0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
-       data4   0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
-       data4   0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
-       data4   0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
-       data4   0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
-       data4   0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
-       data4   0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
-       data4   0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
-       data4   0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
-       data4   0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
-       data4   0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
-       data4   0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
-       data4   0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
-       data4   0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
-       data4   0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
-       data4   0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
-       data4   0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
-       data4   0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
-       data4   0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
-       data4   0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
-       data4   0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
-       data4   0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
-       data4   0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
-       data4   0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
-       data4   0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
-       data4   0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
-       data4   0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
-       data4   0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
-       data4   0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
-       data4   0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
-// Te1:
-       data4   0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b
-       data4   0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5
-       data4   0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b
-       data4   0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676
-       data4   0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d
-       data4   0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0
-       data4   0xec41adad, 0x67b3d4d4, 0xfd5fa2a2, 0xea45afaf
-       data4   0xbf239c9c, 0xf753a4a4, 0x96e47272, 0x5b9bc0c0
-       data4   0xc275b7b7, 0x1ce1fdfd, 0xae3d9393, 0x6a4c2626
-       data4   0x5a6c3636, 0x417e3f3f, 0x02f5f7f7, 0x4f83cccc
-       data4   0x5c683434, 0xf451a5a5, 0x34d1e5e5, 0x08f9f1f1
-       data4   0x93e27171, 0x73abd8d8, 0x53623131, 0x3f2a1515
-       data4   0x0c080404, 0x5295c7c7, 0x65462323, 0x5e9dc3c3
-       data4   0x28301818, 0xa1379696, 0x0f0a0505, 0xb52f9a9a
-       data4   0x090e0707, 0x36241212, 0x9b1b8080, 0x3ddfe2e2
-       data4   0x26cdebeb, 0x694e2727, 0xcd7fb2b2, 0x9fea7575
-       data4   0x1b120909, 0x9e1d8383, 0x74582c2c, 0x2e341a1a
-       data4   0x2d361b1b, 0xb2dc6e6e, 0xeeb45a5a, 0xfb5ba0a0
-       data4   0xf6a45252, 0x4d763b3b, 0x61b7d6d6, 0xce7db3b3
-       data4   0x7b522929, 0x3edde3e3, 0x715e2f2f, 0x97138484
-       data4   0xf5a65353, 0x68b9d1d1, 0x00000000, 0x2cc1eded
-       data4   0x60402020, 0x1fe3fcfc, 0xc879b1b1, 0xedb65b5b
-       data4   0xbed46a6a, 0x468dcbcb, 0xd967bebe, 0x4b723939
-       data4   0xde944a4a, 0xd4984c4c, 0xe8b05858, 0x4a85cfcf
-       data4   0x6bbbd0d0, 0x2ac5efef, 0xe54faaaa, 0x16edfbfb
-       data4   0xc5864343, 0xd79a4d4d, 0x55663333, 0x94118585
-       data4   0xcf8a4545, 0x10e9f9f9, 0x06040202, 0x81fe7f7f
-       data4   0xf0a05050, 0x44783c3c, 0xba259f9f, 0xe34ba8a8
-       data4   0xf3a25151, 0xfe5da3a3, 0xc0804040, 0x8a058f8f
-       data4   0xad3f9292, 0xbc219d9d, 0x48703838, 0x04f1f5f5
-       data4   0xdf63bcbc, 0xc177b6b6, 0x75afdada, 0x63422121
-       data4   0x30201010, 0x1ae5ffff, 0x0efdf3f3, 0x6dbfd2d2
-       data4   0x4c81cdcd, 0x14180c0c, 0x35261313, 0x2fc3ecec
-       data4   0xe1be5f5f, 0xa2359797, 0xcc884444, 0x392e1717
-       data4   0x5793c4c4, 0xf255a7a7, 0x82fc7e7e, 0x477a3d3d
-       data4   0xacc86464, 0xe7ba5d5d, 0x2b321919, 0x95e67373
-       data4   0xa0c06060, 0x98198181, 0xd19e4f4f, 0x7fa3dcdc
-       data4   0x66442222, 0x7e542a2a, 0xab3b9090, 0x830b8888
-       data4   0xca8c4646, 0x29c7eeee, 0xd36bb8b8, 0x3c281414
-       data4   0x79a7dede, 0xe2bc5e5e, 0x1d160b0b, 0x76addbdb
-       data4   0x3bdbe0e0, 0x56643232, 0x4e743a3a, 0x1e140a0a
-       data4   0xdb924949, 0x0a0c0606, 0x6c482424, 0xe4b85c5c
-       data4   0x5d9fc2c2, 0x6ebdd3d3, 0xef43acac, 0xa6c46262
-       data4   0xa8399191, 0xa4319595, 0x37d3e4e4, 0x8bf27979
-       data4   0x32d5e7e7, 0x438bc8c8, 0x596e3737, 0xb7da6d6d
-       data4   0x8c018d8d, 0x64b1d5d5, 0xd29c4e4e, 0xe049a9a9
-       data4   0xb4d86c6c, 0xfaac5656, 0x07f3f4f4, 0x25cfeaea
-       data4   0xafca6565, 0x8ef47a7a, 0xe947aeae, 0x18100808
-       data4   0xd56fbaba, 0x88f07878, 0x6f4a2525, 0x725c2e2e
-       data4   0x24381c1c, 0xf157a6a6, 0xc773b4b4, 0x5197c6c6
-       data4   0x23cbe8e8, 0x7ca1dddd, 0x9ce87474, 0x213e1f1f
-       data4   0xdd964b4b, 0xdc61bdbd, 0x860d8b8b, 0x850f8a8a
-       data4   0x90e07070, 0x427c3e3e, 0xc471b5b5, 0xaacc6666
-       data4   0xd8904848, 0x05060303, 0x01f7f6f6, 0x121c0e0e
-       data4   0xa3c26161, 0x5f6a3535, 0xf9ae5757, 0xd069b9b9
-       data4   0x91178686, 0x5899c1c1, 0x273a1d1d, 0xb9279e9e
-       data4   0x38d9e1e1, 0x13ebf8f8, 0xb32b9898, 0x33221111
-       data4   0xbbd26969, 0x70a9d9d9, 0x89078e8e, 0xa7339494
-       data4   0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9
-       data4   0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf
-       data4   0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d
-       data4   0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868
-       data4   0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f
-       data4   0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616
-// Te2:
-       data4   0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b
-       data4   0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5
-       data4   0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b
-       data4   0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76
-       data4   0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d
-       data4   0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0
-       data4   0xadec41ad, 0xd467b3d4, 0xa2fd5fa2, 0xafea45af
-       data4   0x9cbf239c, 0xa4f753a4, 0x7296e472, 0xc05b9bc0
-       data4   0xb7c275b7, 0xfd1ce1fd, 0x93ae3d93, 0x266a4c26
-       data4   0x365a6c36, 0x3f417e3f, 0xf702f5f7, 0xcc4f83cc
-       data4   0x345c6834, 0xa5f451a5, 0xe534d1e5, 0xf108f9f1
-       data4   0x7193e271, 0xd873abd8, 0x31536231, 0x153f2a15
-       data4   0x040c0804, 0xc75295c7, 0x23654623, 0xc35e9dc3
-       data4   0x18283018, 0x96a13796, 0x050f0a05, 0x9ab52f9a
-       data4   0x07090e07, 0x12362412, 0x809b1b80, 0xe23ddfe2
-       data4   0xeb26cdeb, 0x27694e27, 0xb2cd7fb2, 0x759fea75
-       data4   0x091b1209, 0x839e1d83, 0x2c74582c, 0x1a2e341a
-       data4   0x1b2d361b, 0x6eb2dc6e, 0x5aeeb45a, 0xa0fb5ba0
-       data4   0x52f6a452, 0x3b4d763b, 0xd661b7d6, 0xb3ce7db3
-       data4   0x297b5229, 0xe33edde3, 0x2f715e2f, 0x84971384
-       data4   0x53f5a653, 0xd168b9d1, 0x00000000, 0xed2cc1ed
-       data4   0x20604020, 0xfc1fe3fc, 0xb1c879b1, 0x5bedb65b
-       data4   0x6abed46a, 0xcb468dcb, 0xbed967be, 0x394b7239
-       data4   0x4ade944a, 0x4cd4984c, 0x58e8b058, 0xcf4a85cf
-       data4   0xd06bbbd0, 0xef2ac5ef, 0xaae54faa, 0xfb16edfb
-       data4   0x43c58643, 0x4dd79a4d, 0x33556633, 0x85941185
-       data4   0x45cf8a45, 0xf910e9f9, 0x02060402, 0x7f81fe7f
-       data4   0x50f0a050, 0x3c44783c, 0x9fba259f, 0xa8e34ba8
-       data4   0x51f3a251, 0xa3fe5da3, 0x40c08040, 0x8f8a058f
-       data4   0x92ad3f92, 0x9dbc219d, 0x38487038, 0xf504f1f5
-       data4   0xbcdf63bc, 0xb6c177b6, 0xda75afda, 0x21634221
-       data4   0x10302010, 0xff1ae5ff, 0xf30efdf3, 0xd26dbfd2
-       data4   0xcd4c81cd, 0x0c14180c, 0x13352613, 0xec2fc3ec
-       data4   0x5fe1be5f, 0x97a23597, 0x44cc8844, 0x17392e17
-       data4   0xc45793c4, 0xa7f255a7, 0x7e82fc7e, 0x3d477a3d
-       data4   0x64acc864, 0x5de7ba5d, 0x192b3219, 0x7395e673
-       data4   0x60a0c060, 0x81981981, 0x4fd19e4f, 0xdc7fa3dc
-       data4   0x22664422, 0x2a7e542a, 0x90ab3b90, 0x88830b88
-       data4   0x46ca8c46, 0xee29c7ee, 0xb8d36bb8, 0x143c2814
-       data4   0xde79a7de, 0x5ee2bc5e, 0x0b1d160b, 0xdb76addb
-       data4   0xe03bdbe0, 0x32566432, 0x3a4e743a, 0x0a1e140a
-       data4   0x49db9249, 0x060a0c06, 0x246c4824, 0x5ce4b85c
-       data4   0xc25d9fc2, 0xd36ebdd3, 0xacef43ac, 0x62a6c462
-       data4   0x91a83991, 0x95a43195, 0xe437d3e4, 0x798bf279
-       data4   0xe732d5e7, 0xc8438bc8, 0x37596e37, 0x6db7da6d
-       data4   0x8d8c018d, 0xd564b1d5, 0x4ed29c4e, 0xa9e049a9
-       data4   0x6cb4d86c, 0x56faac56, 0xf407f3f4, 0xea25cfea
-       data4   0x65afca65, 0x7a8ef47a, 0xaee947ae, 0x08181008
-       data4   0xbad56fba, 0x7888f078, 0x256f4a25, 0x2e725c2e
-       data4   0x1c24381c, 0xa6f157a6, 0xb4c773b4, 0xc65197c6
-       data4   0xe823cbe8, 0xdd7ca1dd, 0x749ce874, 0x1f213e1f
-       data4   0x4bdd964b, 0xbddc61bd, 0x8b860d8b, 0x8a850f8a
-       data4   0x7090e070, 0x3e427c3e, 0xb5c471b5, 0x66aacc66
-       data4   0x48d89048, 0x03050603, 0xf601f7f6, 0x0e121c0e
-       data4   0x61a3c261, 0x355f6a35, 0x57f9ae57, 0xb9d069b9
-       data4   0x86911786, 0xc15899c1, 0x1d273a1d, 0x9eb9279e
-       data4   0xe138d9e1, 0xf813ebf8, 0x98b32b98, 0x11332211
-       data4   0x69bbd269, 0xd970a9d9, 0x8e89078e, 0x94a73394
-       data4   0x9bb62d9b, 0x1e223c1e, 0x87921587, 0xe920c9e9
-       data4   0xce4987ce, 0x55ffaa55, 0x28785028, 0xdf7aa5df
-       data4   0x8c8f038c, 0xa1f859a1, 0x89800989, 0x0d171a0d
-       data4   0xbfda65bf, 0xe631d7e6, 0x42c68442, 0x68b8d068
-       data4   0x41c38241, 0x99b02999, 0x2d775a2d, 0x0f111e0f
-       data4   0xb0cb7bb0, 0x54fca854, 0xbbd66dbb, 0x163a2c16
-// Te3:
-       data4   0x6363a5c6, 0x7c7c84f8, 0x777799ee, 0x7b7b8df6
-       data4   0xf2f20dff, 0x6b6bbdd6, 0x6f6fb1de, 0xc5c55491
-       data4   0x30305060, 0x01010302, 0x6767a9ce, 0x2b2b7d56
-       data4   0xfefe19e7, 0xd7d762b5, 0xababe64d, 0x76769aec
-       data4   0xcaca458f, 0x82829d1f, 0xc9c94089, 0x7d7d87fa
-       data4   0xfafa15ef, 0x5959ebb2, 0x4747c98e, 0xf0f00bfb
-       data4   0xadadec41, 0xd4d467b3, 0xa2a2fd5f, 0xafafea45
-       data4   0x9c9cbf23, 0xa4a4f753, 0x727296e4, 0xc0c05b9b
-       data4   0xb7b7c275, 0xfdfd1ce1, 0x9393ae3d, 0x26266a4c
-       data4   0x36365a6c, 0x3f3f417e, 0xf7f702f5, 0xcccc4f83
-       data4   0x34345c68, 0xa5a5f451, 0xe5e534d1, 0xf1f108f9
-       data4   0x717193e2, 0xd8d873ab, 0x31315362, 0x15153f2a
-       data4   0x04040c08, 0xc7c75295, 0x23236546, 0xc3c35e9d
-       data4   0x18182830, 0x9696a137, 0x05050f0a, 0x9a9ab52f
-       data4   0x0707090e, 0x12123624, 0x80809b1b, 0xe2e23ddf
-       data4   0xebeb26cd, 0x2727694e, 0xb2b2cd7f, 0x75759fea
-       data4   0x09091b12, 0x83839e1d, 0x2c2c7458, 0x1a1a2e34
-       data4   0x1b1b2d36, 0x6e6eb2dc, 0x5a5aeeb4, 0xa0a0fb5b
-       data4   0x5252f6a4, 0x3b3b4d76, 0xd6d661b7, 0xb3b3ce7d
-       data4   0x29297b52, 0xe3e33edd, 0x2f2f715e, 0x84849713
-       data4   0x5353f5a6, 0xd1d168b9, 0x00000000, 0xeded2cc1
-       data4   0x20206040, 0xfcfc1fe3, 0xb1b1c879, 0x5b5bedb6
-       data4   0x6a6abed4, 0xcbcb468d, 0xbebed967, 0x39394b72
-       data4   0x4a4ade94, 0x4c4cd498, 0x5858e8b0, 0xcfcf4a85
-       data4   0xd0d06bbb, 0xefef2ac5, 0xaaaae54f, 0xfbfb16ed
-       data4   0x4343c586, 0x4d4dd79a, 0x33335566, 0x85859411
-       data4   0x4545cf8a, 0xf9f910e9, 0x02020604, 0x7f7f81fe
-       data4   0x5050f0a0, 0x3c3c4478, 0x9f9fba25, 0xa8a8e34b
-       data4   0x5151f3a2, 0xa3a3fe5d, 0x4040c080, 0x8f8f8a05
-       data4   0x9292ad3f, 0x9d9dbc21, 0x38384870, 0xf5f504f1
-       data4   0xbcbcdf63, 0xb6b6c177, 0xdada75af, 0x21216342
-       data4   0x10103020, 0xffff1ae5, 0xf3f30efd, 0xd2d26dbf
-       data4   0xcdcd4c81, 0x0c0c1418, 0x13133526, 0xecec2fc3
-       data4   0x5f5fe1be, 0x9797a235, 0x4444cc88, 0x1717392e
-       data4   0xc4c45793, 0xa7a7f255, 0x7e7e82fc, 0x3d3d477a
-       data4   0x6464acc8, 0x5d5de7ba, 0x19192b32, 0x737395e6
-       data4   0x6060a0c0, 0x81819819, 0x4f4fd19e, 0xdcdc7fa3
-       data4   0x22226644, 0x2a2a7e54, 0x9090ab3b, 0x8888830b
-       data4   0x4646ca8c, 0xeeee29c7, 0xb8b8d36b, 0x14143c28
-       data4   0xdede79a7, 0x5e5ee2bc, 0x0b0b1d16, 0xdbdb76ad
-       data4   0xe0e03bdb, 0x32325664, 0x3a3a4e74, 0x0a0a1e14
-       data4   0x4949db92, 0x06060a0c, 0x24246c48, 0x5c5ce4b8
-       data4   0xc2c25d9f, 0xd3d36ebd, 0xacacef43, 0x6262a6c4
-       data4   0x9191a839, 0x9595a431, 0xe4e437d3, 0x79798bf2
-       data4   0xe7e732d5, 0xc8c8438b, 0x3737596e, 0x6d6db7da
-       data4   0x8d8d8c01, 0xd5d564b1, 0x4e4ed29c, 0xa9a9e049
-       data4   0x6c6cb4d8, 0x5656faac, 0xf4f407f3, 0xeaea25cf
-       data4   0x6565afca, 0x7a7a8ef4, 0xaeaee947, 0x08081810
-       data4   0xbabad56f, 0x787888f0, 0x25256f4a, 0x2e2e725c
-       data4   0x1c1c2438, 0xa6a6f157, 0xb4b4c773, 0xc6c65197
-       data4   0xe8e823cb, 0xdddd7ca1, 0x74749ce8, 0x1f1f213e
-       data4   0x4b4bdd96, 0xbdbddc61, 0x8b8b860d, 0x8a8a850f
-       data4   0x707090e0, 0x3e3e427c, 0xb5b5c471, 0x6666aacc
-       data4   0x4848d890, 0x03030506, 0xf6f601f7, 0x0e0e121c
-       data4   0x6161a3c2, 0x35355f6a, 0x5757f9ae, 0xb9b9d069
-       data4   0x86869117, 0xc1c15899, 0x1d1d273a, 0x9e9eb927
-       data4   0xe1e138d9, 0xf8f813eb, 0x9898b32b, 0x11113322
-       data4   0x6969bbd2, 0xd9d970a9, 0x8e8e8907, 0x9494a733
-       data4   0x9b9bb62d, 0x1e1e223c, 0x87879215, 0xe9e920c9
-       data4   0xcece4987, 0x5555ffaa, 0x28287850, 0xdfdf7aa5
-       data4   0x8c8c8f03, 0xa1a1f859, 0x89898009, 0x0d0d171a
-       data4   0xbfbfda65, 0xe6e631d7, 0x4242c684, 0x6868b8d0
-       data4   0x4141c382, 0x9999b029, 0x2d2d775a, 0x0f0f111e
-       data4   0xb0b0cb7b, 0x5454fca8, 0xbbbbd66d, 0x16163a2c
+.global        AES_Te#
+.type  AES_Te#,@object
+AES_Te:        data4   0xc66363a5,0xc66363a5, 0xf87c7c84,0xf87c7c84
+       data4   0xee777799,0xee777799, 0xf67b7b8d,0xf67b7b8d
+       data4   0xfff2f20d,0xfff2f20d, 0xd66b6bbd,0xd66b6bbd
+       data4   0xde6f6fb1,0xde6f6fb1, 0x91c5c554,0x91c5c554
+       data4   0x60303050,0x60303050, 0x02010103,0x02010103
+       data4   0xce6767a9,0xce6767a9, 0x562b2b7d,0x562b2b7d
+       data4   0xe7fefe19,0xe7fefe19, 0xb5d7d762,0xb5d7d762
+       data4   0x4dababe6,0x4dababe6, 0xec76769a,0xec76769a
+       data4   0x8fcaca45,0x8fcaca45, 0x1f82829d,0x1f82829d
+       data4   0x89c9c940,0x89c9c940, 0xfa7d7d87,0xfa7d7d87
+       data4   0xeffafa15,0xeffafa15, 0xb25959eb,0xb25959eb
+       data4   0x8e4747c9,0x8e4747c9, 0xfbf0f00b,0xfbf0f00b
+       data4   0x41adadec,0x41adadec, 0xb3d4d467,0xb3d4d467
+       data4   0x5fa2a2fd,0x5fa2a2fd, 0x45afafea,0x45afafea
+       data4   0x239c9cbf,0x239c9cbf, 0x53a4a4f7,0x53a4a4f7
+       data4   0xe4727296,0xe4727296, 0x9bc0c05b,0x9bc0c05b
+       data4   0x75b7b7c2,0x75b7b7c2, 0xe1fdfd1c,0xe1fdfd1c
+       data4   0x3d9393ae,0x3d9393ae, 0x4c26266a,0x4c26266a
+       data4   0x6c36365a,0x6c36365a, 0x7e3f3f41,0x7e3f3f41
+       data4   0xf5f7f702,0xf5f7f702, 0x83cccc4f,0x83cccc4f
+       data4   0x6834345c,0x6834345c, 0x51a5a5f4,0x51a5a5f4
+       data4   0xd1e5e534,0xd1e5e534, 0xf9f1f108,0xf9f1f108
+       data4   0xe2717193,0xe2717193, 0xabd8d873,0xabd8d873
+       data4   0x62313153,0x62313153, 0x2a15153f,0x2a15153f
+       data4   0x0804040c,0x0804040c, 0x95c7c752,0x95c7c752
+       data4   0x46232365,0x46232365, 0x9dc3c35e,0x9dc3c35e
+       data4   0x30181828,0x30181828, 0x379696a1,0x379696a1
+       data4   0x0a05050f,0x0a05050f, 0x2f9a9ab5,0x2f9a9ab5
+       data4   0x0e070709,0x0e070709, 0x24121236,0x24121236
+       data4   0x1b80809b,0x1b80809b, 0xdfe2e23d,0xdfe2e23d
+       data4   0xcdebeb26,0xcdebeb26, 0x4e272769,0x4e272769
+       data4   0x7fb2b2cd,0x7fb2b2cd, 0xea75759f,0xea75759f
+       data4   0x1209091b,0x1209091b, 0x1d83839e,0x1d83839e
+       data4   0x582c2c74,0x582c2c74, 0x341a1a2e,0x341a1a2e
+       data4   0x361b1b2d,0x361b1b2d, 0xdc6e6eb2,0xdc6e6eb2
+       data4   0xb45a5aee,0xb45a5aee, 0x5ba0a0fb,0x5ba0a0fb
+       data4   0xa45252f6,0xa45252f6, 0x763b3b4d,0x763b3b4d
+       data4   0xb7d6d661,0xb7d6d661, 0x7db3b3ce,0x7db3b3ce
+       data4   0x5229297b,0x5229297b, 0xdde3e33e,0xdde3e33e
+       data4   0x5e2f2f71,0x5e2f2f71, 0x13848497,0x13848497
+       data4   0xa65353f5,0xa65353f5, 0xb9d1d168,0xb9d1d168
+       data4   0x00000000,0x00000000, 0xc1eded2c,0xc1eded2c
+       data4   0x40202060,0x40202060, 0xe3fcfc1f,0xe3fcfc1f
+       data4   0x79b1b1c8,0x79b1b1c8, 0xb65b5bed,0xb65b5bed
+       data4   0xd46a6abe,0xd46a6abe, 0x8dcbcb46,0x8dcbcb46
+       data4   0x67bebed9,0x67bebed9, 0x7239394b,0x7239394b
+       data4   0x944a4ade,0x944a4ade, 0x984c4cd4,0x984c4cd4
+       data4   0xb05858e8,0xb05858e8, 0x85cfcf4a,0x85cfcf4a
+       data4   0xbbd0d06b,0xbbd0d06b, 0xc5efef2a,0xc5efef2a
+       data4   0x4faaaae5,0x4faaaae5, 0xedfbfb16,0xedfbfb16
+       data4   0x864343c5,0x864343c5, 0x9a4d4dd7,0x9a4d4dd7
+       data4   0x66333355,0x66333355, 0x11858594,0x11858594
+       data4   0x8a4545cf,0x8a4545cf, 0xe9f9f910,0xe9f9f910
+       data4   0x04020206,0x04020206, 0xfe7f7f81,0xfe7f7f81
+       data4   0xa05050f0,0xa05050f0, 0x783c3c44,0x783c3c44
+       data4   0x259f9fba,0x259f9fba, 0x4ba8a8e3,0x4ba8a8e3
+       data4   0xa25151f3,0xa25151f3, 0x5da3a3fe,0x5da3a3fe
+       data4   0x804040c0,0x804040c0, 0x058f8f8a,0x058f8f8a
+       data4   0x3f9292ad,0x3f9292ad, 0x219d9dbc,0x219d9dbc
+       data4   0x70383848,0x70383848, 0xf1f5f504,0xf1f5f504
+       data4   0x63bcbcdf,0x63bcbcdf, 0x77b6b6c1,0x77b6b6c1
+       data4   0xafdada75,0xafdada75, 0x42212163,0x42212163
+       data4   0x20101030,0x20101030, 0xe5ffff1a,0xe5ffff1a
+       data4   0xfdf3f30e,0xfdf3f30e, 0xbfd2d26d,0xbfd2d26d
+       data4   0x81cdcd4c,0x81cdcd4c, 0x180c0c14,0x180c0c14
+       data4   0x26131335,0x26131335, 0xc3ecec2f,0xc3ecec2f
+       data4   0xbe5f5fe1,0xbe5f5fe1, 0x359797a2,0x359797a2
+       data4   0x884444cc,0x884444cc, 0x2e171739,0x2e171739
+       data4   0x93c4c457,0x93c4c457, 0x55a7a7f2,0x55a7a7f2
+       data4   0xfc7e7e82,0xfc7e7e82, 0x7a3d3d47,0x7a3d3d47
+       data4   0xc86464ac,0xc86464ac, 0xba5d5de7,0xba5d5de7
+       data4   0x3219192b,0x3219192b, 0xe6737395,0xe6737395
+       data4   0xc06060a0,0xc06060a0, 0x19818198,0x19818198
+       data4   0x9e4f4fd1,0x9e4f4fd1, 0xa3dcdc7f,0xa3dcdc7f
+       data4   0x44222266,0x44222266, 0x542a2a7e,0x542a2a7e
+       data4   0x3b9090ab,0x3b9090ab, 0x0b888883,0x0b888883
+       data4   0x8c4646ca,0x8c4646ca, 0xc7eeee29,0xc7eeee29
+       data4   0x6bb8b8d3,0x6bb8b8d3, 0x2814143c,0x2814143c
+       data4   0xa7dede79,0xa7dede79, 0xbc5e5ee2,0xbc5e5ee2
+       data4   0x160b0b1d,0x160b0b1d, 0xaddbdb76,0xaddbdb76
+       data4   0xdbe0e03b,0xdbe0e03b, 0x64323256,0x64323256
+       data4   0x743a3a4e,0x743a3a4e, 0x140a0a1e,0x140a0a1e
+       data4   0x924949db,0x924949db, 0x0c06060a,0x0c06060a
+       data4   0x4824246c,0x4824246c, 0xb85c5ce4,0xb85c5ce4
+       data4   0x9fc2c25d,0x9fc2c25d, 0xbdd3d36e,0xbdd3d36e
+       data4   0x43acacef,0x43acacef, 0xc46262a6,0xc46262a6
+       data4   0x399191a8,0x399191a8, 0x319595a4,0x319595a4
+       data4   0xd3e4e437,0xd3e4e437, 0xf279798b,0xf279798b
+       data4   0xd5e7e732,0xd5e7e732, 0x8bc8c843,0x8bc8c843
+       data4   0x6e373759,0x6e373759, 0xda6d6db7,0xda6d6db7
+       data4   0x018d8d8c,0x018d8d8c, 0xb1d5d564,0xb1d5d564
+       data4   0x9c4e4ed2,0x9c4e4ed2, 0x49a9a9e0,0x49a9a9e0
+       data4   0xd86c6cb4,0xd86c6cb4, 0xac5656fa,0xac5656fa
+       data4   0xf3f4f407,0xf3f4f407, 0xcfeaea25,0xcfeaea25
+       data4   0xca6565af,0xca6565af, 0xf47a7a8e,0xf47a7a8e
+       data4   0x47aeaee9,0x47aeaee9, 0x10080818,0x10080818
+       data4   0x6fbabad5,0x6fbabad5, 0xf0787888,0xf0787888
+       data4   0x4a25256f,0x4a25256f, 0x5c2e2e72,0x5c2e2e72
+       data4   0x381c1c24,0x381c1c24, 0x57a6a6f1,0x57a6a6f1
+       data4   0x73b4b4c7,0x73b4b4c7, 0x97c6c651,0x97c6c651
+       data4   0xcbe8e823,0xcbe8e823, 0xa1dddd7c,0xa1dddd7c
+       data4   0xe874749c,0xe874749c, 0x3e1f1f21,0x3e1f1f21
+       data4   0x964b4bdd,0x964b4bdd, 0x61bdbddc,0x61bdbddc
+       data4   0x0d8b8b86,0x0d8b8b86, 0x0f8a8a85,0x0f8a8a85
+       data4   0xe0707090,0xe0707090, 0x7c3e3e42,0x7c3e3e42
+       data4   0x71b5b5c4,0x71b5b5c4, 0xcc6666aa,0xcc6666aa
+       data4   0x904848d8,0x904848d8, 0x06030305,0x06030305
+       data4   0xf7f6f601,0xf7f6f601, 0x1c0e0e12,0x1c0e0e12
+       data4   0xc26161a3,0xc26161a3, 0x6a35355f,0x6a35355f
+       data4   0xae5757f9,0xae5757f9, 0x69b9b9d0,0x69b9b9d0
+       data4   0x17868691,0x17868691, 0x99c1c158,0x99c1c158
+       data4   0x3a1d1d27,0x3a1d1d27, 0x279e9eb9,0x279e9eb9
+       data4   0xd9e1e138,0xd9e1e138, 0xebf8f813,0xebf8f813
+       data4   0x2b9898b3,0x2b9898b3, 0x22111133,0x22111133
+       data4   0xd26969bb,0xd26969bb, 0xa9d9d970,0xa9d9d970
+       data4   0x078e8e89,0x078e8e89, 0x339494a7,0x339494a7
+       data4   0x2d9b9bb6,0x2d9b9bb6, 0x3c1e1e22,0x3c1e1e22
+       data4   0x15878792,0x15878792, 0xc9e9e920,0xc9e9e920
+       data4   0x87cece49,0x87cece49, 0xaa5555ff,0xaa5555ff
+       data4   0x50282878,0x50282878, 0xa5dfdf7a,0xa5dfdf7a
+       data4   0x038c8c8f,0x038c8c8f, 0x59a1a1f8,0x59a1a1f8
+       data4   0x09898980,0x09898980, 0x1a0d0d17,0x1a0d0d17
+       data4   0x65bfbfda,0x65bfbfda, 0xd7e6e631,0xd7e6e631
+       data4   0x844242c6,0x844242c6, 0xd06868b8,0xd06868b8
+       data4   0x824141c3,0x824141c3, 0x299999b0,0x299999b0
+       data4   0x5a2d2d77,0x5a2d2d77, 0x1e0f0f11,0x1e0f0f11
+       data4   0x7bb0b0cb,0x7bb0b0cb, 0xa85454fc,0xa85454fc
+       data4   0x6dbbbbd6,0x6dbbbbd6, 0x2c16163a,0x2c16163a
 // Te4:
-       data4   0x63636363, 0x7c7c7c7c, 0x77777777, 0x7b7b7b7b
-       data4   0xf2f2f2f2, 0x6b6b6b6b, 0x6f6f6f6f, 0xc5c5c5c5
-       data4   0x30303030, 0x01010101, 0x67676767, 0x2b2b2b2b
-       data4   0xfefefefe, 0xd7d7d7d7, 0xabababab, 0x76767676
-       data4   0xcacacaca, 0x82828282, 0xc9c9c9c9, 0x7d7d7d7d
-       data4   0xfafafafa, 0x59595959, 0x47474747, 0xf0f0f0f0
-       data4   0xadadadad, 0xd4d4d4d4, 0xa2a2a2a2, 0xafafafaf
-       data4   0x9c9c9c9c, 0xa4a4a4a4, 0x72727272, 0xc0c0c0c0
-       data4   0xb7b7b7b7, 0xfdfdfdfd, 0x93939393, 0x26262626
-       data4   0x36363636, 0x3f3f3f3f, 0xf7f7f7f7, 0xcccccccc
-       data4   0x34343434, 0xa5a5a5a5, 0xe5e5e5e5, 0xf1f1f1f1
-       data4   0x71717171, 0xd8d8d8d8, 0x31313131, 0x15151515
-       data4   0x04040404, 0xc7c7c7c7, 0x23232323, 0xc3c3c3c3
-       data4   0x18181818, 0x96969696, 0x05050505, 0x9a9a9a9a
-       data4   0x07070707, 0x12121212, 0x80808080, 0xe2e2e2e2
-       data4   0xebebebeb, 0x27272727, 0xb2b2b2b2, 0x75757575
-       data4   0x09090909, 0x83838383, 0x2c2c2c2c, 0x1a1a1a1a
-       data4   0x1b1b1b1b, 0x6e6e6e6e, 0x5a5a5a5a, 0xa0a0a0a0
-       data4   0x52525252, 0x3b3b3b3b, 0xd6d6d6d6, 0xb3b3b3b3
-       data4   0x29292929, 0xe3e3e3e3, 0x2f2f2f2f, 0x84848484
-       data4   0x53535353, 0xd1d1d1d1, 0x00000000, 0xedededed
-       data4   0x20202020, 0xfcfcfcfc, 0xb1b1b1b1, 0x5b5b5b5b
-       data4   0x6a6a6a6a, 0xcbcbcbcb, 0xbebebebe, 0x39393939
-       data4   0x4a4a4a4a, 0x4c4c4c4c, 0x58585858, 0xcfcfcfcf
-       data4   0xd0d0d0d0, 0xefefefef, 0xaaaaaaaa, 0xfbfbfbfb
-       data4   0x43434343, 0x4d4d4d4d, 0x33333333, 0x85858585
-       data4   0x45454545, 0xf9f9f9f9, 0x02020202, 0x7f7f7f7f
-       data4   0x50505050, 0x3c3c3c3c, 0x9f9f9f9f, 0xa8a8a8a8
-       data4   0x51515151, 0xa3a3a3a3, 0x40404040, 0x8f8f8f8f
-       data4   0x92929292, 0x9d9d9d9d, 0x38383838, 0xf5f5f5f5
-       data4   0xbcbcbcbc, 0xb6b6b6b6, 0xdadadada, 0x21212121
-       data4   0x10101010, 0xffffffff, 0xf3f3f3f3, 0xd2d2d2d2
-       data4   0xcdcdcdcd, 0x0c0c0c0c, 0x13131313, 0xecececec
-       data4   0x5f5f5f5f, 0x97979797, 0x44444444, 0x17171717
-       data4   0xc4c4c4c4, 0xa7a7a7a7, 0x7e7e7e7e, 0x3d3d3d3d
-       data4   0x64646464, 0x5d5d5d5d, 0x19191919, 0x73737373
-       data4   0x60606060, 0x81818181, 0x4f4f4f4f, 0xdcdcdcdc
-       data4   0x22222222, 0x2a2a2a2a, 0x90909090, 0x88888888
-       data4   0x46464646, 0xeeeeeeee, 0xb8b8b8b8, 0x14141414
-       data4   0xdededede, 0x5e5e5e5e, 0x0b0b0b0b, 0xdbdbdbdb
-       data4   0xe0e0e0e0, 0x32323232, 0x3a3a3a3a, 0x0a0a0a0a
-       data4   0x49494949, 0x06060606, 0x24242424, 0x5c5c5c5c
-       data4   0xc2c2c2c2, 0xd3d3d3d3, 0xacacacac, 0x62626262
-       data4   0x91919191, 0x95959595, 0xe4e4e4e4, 0x79797979
-       data4   0xe7e7e7e7, 0xc8c8c8c8, 0x37373737, 0x6d6d6d6d
-       data4   0x8d8d8d8d, 0xd5d5d5d5, 0x4e4e4e4e, 0xa9a9a9a9
-       data4   0x6c6c6c6c, 0x56565656, 0xf4f4f4f4, 0xeaeaeaea
-       data4   0x65656565, 0x7a7a7a7a, 0xaeaeaeae, 0x08080808
-       data4   0xbabababa, 0x78787878, 0x25252525, 0x2e2e2e2e
-       data4   0x1c1c1c1c, 0xa6a6a6a6, 0xb4b4b4b4, 0xc6c6c6c6
-       data4   0xe8e8e8e8, 0xdddddddd, 0x74747474, 0x1f1f1f1f
-       data4   0x4b4b4b4b, 0xbdbdbdbd, 0x8b8b8b8b, 0x8a8a8a8a
-       data4   0x70707070, 0x3e3e3e3e, 0xb5b5b5b5, 0x66666666
-       data4   0x48484848, 0x03030303, 0xf6f6f6f6, 0x0e0e0e0e
-       data4   0x61616161, 0x35353535, 0x57575757, 0xb9b9b9b9
-       data4   0x86868686, 0xc1c1c1c1, 0x1d1d1d1d, 0x9e9e9e9e
-       data4   0xe1e1e1e1, 0xf8f8f8f8, 0x98989898, 0x11111111
-       data4   0x69696969, 0xd9d9d9d9, 0x8e8e8e8e, 0x94949494
-       data4   0x9b9b9b9b, 0x1e1e1e1e, 0x87878787, 0xe9e9e9e9
-       data4   0xcececece, 0x55555555, 0x28282828, 0xdfdfdfdf
-       data4   0x8c8c8c8c, 0xa1a1a1a1, 0x89898989, 0x0d0d0d0d
-       data4   0xbfbfbfbf, 0xe6e6e6e6, 0x42424242, 0x68686868
-       data4   0x41414141, 0x99999999, 0x2d2d2d2d, 0x0f0f0f0f
-       data4   0xb0b0b0b0, 0x54545454, 0xbbbbbbbb, 0x16161616
-.size  Te0#,5*256*4    // HP-UX assembler fails to ".-Te0#"
+       data1   0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
+       data1   0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
+       data1   0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
+       data1   0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
+       data1   0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
+       data1   0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
+       data1   0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
+       data1   0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
+       data1   0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
+       data1   0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
+       data1   0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
+       data1   0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
+       data1   0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
+       data1   0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
+       data1   0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
+       data1   0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
+       data1   0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
+       data1   0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
+       data1   0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
+       data1   0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
+       data1   0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
+       data1   0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
+       data1   0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
+       data1   0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
+       data1   0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
+       data1   0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
+       data1   0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
+       data1   0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
+       data1   0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
+       data1   0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
+       data1   0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
+       data1   0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
+.size  AES_Te#,2048+256        // HP-UX assembler fails to ".-AES_Te#"
 
 .align 64
-.type  Td0#,@object
-Td0:   data4   0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
-       data4   0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
-       data4   0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
-       data4   0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
-       data4   0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
-       data4   0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
-       data4   0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
-       data4   0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
-       data4   0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
-       data4   0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
-       data4   0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
-       data4   0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
-       data4   0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
-       data4   0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
-       data4   0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
-       data4   0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
-       data4   0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
-       data4   0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
-       data4   0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
-       data4   0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
-       data4   0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
-       data4   0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
-       data4   0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
-       data4   0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
-       data4   0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
-       data4   0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
-       data4   0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
-       data4   0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
-       data4   0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
-       data4   0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
-       data4   0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
-       data4   0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
-       data4   0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
-       data4   0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
-       data4   0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
-       data4   0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
-       data4   0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
-       data4   0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
-       data4   0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
-       data4   0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
-       data4   0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
-       data4   0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
-       data4   0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
-       data4   0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
-       data4   0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
-       data4   0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
-       data4   0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
-       data4   0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
-       data4   0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
-       data4   0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
-       data4   0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
-       data4   0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
-       data4   0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
-       data4   0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
-       data4   0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
-       data4   0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
-       data4   0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
-       data4   0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
-       data4   0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
-       data4   0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
-       data4   0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
-       data4   0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
-       data4   0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
-       data4   0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
-// Td1:
-       data4   0x5051f4a7, 0x537e4165, 0xc31a17a4, 0x963a275e
-       data4   0xcb3bab6b, 0xf11f9d45, 0xabacfa58, 0x934be303
-       data4   0x552030fa, 0xf6ad766d, 0x9188cc76, 0x25f5024c
-       data4   0xfc4fe5d7, 0xd7c52acb, 0x80263544, 0x8fb562a3
-       data4   0x49deb15a, 0x6725ba1b, 0x9845ea0e, 0xe15dfec0
-       data4   0x02c32f75, 0x12814cf0, 0xa38d4697, 0xc66bd3f9
-       data4   0xe7038f5f, 0x9515929c, 0xebbf6d7a, 0xda955259
-       data4   0x2dd4be83, 0xd3587421, 0x2949e069, 0x448ec9c8
-       data4   0x6a75c289, 0x78f48e79, 0x6b99583e, 0xdd27b971
-       data4   0xb6bee14f, 0x17f088ad, 0x66c920ac, 0xb47dce3a
-       data4   0x1863df4a, 0x82e51a31, 0x60975133, 0x4562537f
-       data4   0xe0b16477, 0x84bb6bae, 0x1cfe81a0, 0x94f9082b
-       data4   0x58704868, 0x198f45fd, 0x8794de6c, 0xb7527bf8
-       data4   0x23ab73d3, 0xe2724b02, 0x57e31f8f, 0x2a6655ab
-       data4   0x07b2eb28, 0x032fb5c2, 0x9a86c57b, 0xa5d33708
-       data4   0xf2302887, 0xb223bfa5, 0xba02036a, 0x5ced1682
-       data4   0x2b8acf1c, 0x92a779b4, 0xf0f307f2, 0xa14e69e2
-       data4   0xcd65daf4, 0xd50605be, 0x1fd13462, 0x8ac4a6fe
-       data4   0x9d342e53, 0xa0a2f355, 0x32058ae1, 0x75a4f6eb
-       data4   0x390b83ec, 0xaa4060ef, 0x065e719f, 0x51bd6e10
-       data4   0xf93e218a, 0x3d96dd06, 0xaedd3e05, 0x464de6bd
-       data4   0xb591548d, 0x0571c45d, 0x6f0406d4, 0xff605015
-       data4   0x241998fb, 0x97d6bde9, 0xcc894043, 0x7767d99e
-       data4   0xbdb0e842, 0x8807898b, 0x38e7195b, 0xdb79c8ee
-       data4   0x47a17c0a, 0xe97c420f, 0xc9f8841e, 0x00000000
-       data4   0x83098086, 0x48322bed, 0xac1e1170, 0x4e6c5a72
-       data4   0xfbfd0eff, 0x560f8538, 0x1e3daed5, 0x27362d39
-       data4   0x640a0fd9, 0x21685ca6, 0xd19b5b54, 0x3a24362e
-       data4   0xb10c0a67, 0x0f9357e7, 0xd2b4ee96, 0x9e1b9b91
-       data4   0x4f80c0c5, 0xa261dc20, 0x695a774b, 0x161c121a
-       data4   0x0ae293ba, 0xe5c0a02a, 0x433c22e0, 0x1d121b17
-       data4   0x0b0e090d, 0xadf28bc7, 0xb92db6a8, 0xc8141ea9
-       data4   0x8557f119, 0x4caf7507, 0xbbee99dd, 0xfda37f60
-       data4   0x9ff70126, 0xbc5c72f5, 0xc544663b, 0x345bfb7e
-       data4   0x768b4329, 0xdccb23c6, 0x68b6edfc, 0x63b8e4f1
-       data4   0xcad731dc, 0x10426385, 0x40139722, 0x2084c611
-       data4   0x7d854a24, 0xf8d2bb3d, 0x11aef932, 0x6dc729a1
-       data4   0x4b1d9e2f, 0xf3dcb230, 0xec0d8652, 0xd077c1e3
-       data4   0x6c2bb316, 0x99a970b9, 0xfa119448, 0x2247e964
-       data4   0xc4a8fc8c, 0x1aa0f03f, 0xd8567d2c, 0xef223390
-       data4   0xc787494e, 0xc1d938d1, 0xfe8ccaa2, 0x3698d40b
-       data4   0xcfa6f581, 0x28a57ade, 0x26dab78e, 0xa43fadbf
-       data4   0xe42c3a9d, 0x0d507892, 0x9b6a5fcc, 0x62547e46
-       data4   0xc2f68d13, 0xe890d8b8, 0x5e2e39f7, 0xf582c3af
-       data4   0xbe9f5d80, 0x7c69d093, 0xa96fd52d, 0xb3cf2512
-       data4   0x3bc8ac99, 0xa710187d, 0x6ee89c63, 0x7bdb3bbb
-       data4   0x09cd2678, 0xf46e5918, 0x01ec9ab7, 0xa8834f9a
-       data4   0x65e6956e, 0x7eaaffe6, 0x0821bccf, 0xe6ef15e8
-       data4   0xd9bae79b, 0xce4a6f36, 0xd4ea9f09, 0xd629b07c
-       data4   0xaf31a4b2, 0x312a3f23, 0x30c6a594, 0xc035a266
-       data4   0x37744ebc, 0xa6fc82ca, 0xb0e090d0, 0x1533a7d8
-       data4   0x4af10498, 0xf741ecda, 0x0e7fcd50, 0x2f1791f6
-       data4   0x8d764dd6, 0x4d43efb0, 0x54ccaa4d, 0xdfe49604
-       data4   0xe39ed1b5, 0x1b4c6a88, 0xb8c12c1f, 0x7f466551
-       data4   0x049d5eea, 0x5d018c35, 0x73fa8774, 0x2efb0b41
-       data4   0x5ab3671d, 0x5292dbd2, 0x33e91056, 0x136dd647
-       data4   0x8c9ad761, 0x7a37a10c, 0x8e59f814, 0x89eb133c
-       data4   0xeecea927, 0x35b761c9, 0xede11ce5, 0x3c7a47b1
-       data4   0x599cd2df, 0x3f55f273, 0x791814ce, 0xbf73c737
-       data4   0xea53f7cd, 0x5b5ffdaa, 0x14df3d6f, 0x867844db
-       data4   0x81caaff3, 0x3eb968c4, 0x2c382434, 0x5fc2a340
-       data4   0x72161dc3, 0x0cbce225, 0x8b283c49, 0x41ff0d95
-       data4   0x7139a801, 0xde080cb3, 0x9cd8b4e4, 0x906456c1
-       data4   0x617bcb84, 0x70d532b6, 0x74486c5c, 0x42d0b857
-// Td2:
-       data4   0xa75051f4, 0x65537e41, 0xa4c31a17, 0x5e963a27
-       data4   0x6bcb3bab, 0x45f11f9d, 0x58abacfa, 0x03934be3
-       data4   0xfa552030, 0x6df6ad76, 0x769188cc, 0x4c25f502
-       data4   0xd7fc4fe5, 0xcbd7c52a, 0x44802635, 0xa38fb562
-       data4   0x5a49deb1, 0x1b6725ba, 0x0e9845ea, 0xc0e15dfe
-       data4   0x7502c32f, 0xf012814c, 0x97a38d46, 0xf9c66bd3
-       data4   0x5fe7038f, 0x9c951592, 0x7aebbf6d, 0x59da9552
-       data4   0x832dd4be, 0x21d35874, 0x692949e0, 0xc8448ec9
-       data4   0x896a75c2, 0x7978f48e, 0x3e6b9958, 0x71dd27b9
-       data4   0x4fb6bee1, 0xad17f088, 0xac66c920, 0x3ab47dce
-       data4   0x4a1863df, 0x3182e51a, 0x33609751, 0x7f456253
-       data4   0x77e0b164, 0xae84bb6b, 0xa01cfe81, 0x2b94f908
-       data4   0x68587048, 0xfd198f45, 0x6c8794de, 0xf8b7527b
-       data4   0xd323ab73, 0x02e2724b, 0x8f57e31f, 0xab2a6655
-       data4   0x2807b2eb, 0xc2032fb5, 0x7b9a86c5, 0x08a5d337
-       data4   0x87f23028, 0xa5b223bf, 0x6aba0203, 0x825ced16
-       data4   0x1c2b8acf, 0xb492a779, 0xf2f0f307, 0xe2a14e69
-       data4   0xf4cd65da, 0xbed50605, 0x621fd134, 0xfe8ac4a6
-       data4   0x539d342e, 0x55a0a2f3, 0xe132058a, 0xeb75a4f6
-       data4   0xec390b83, 0xefaa4060, 0x9f065e71, 0x1051bd6e
-       data4   0x8af93e21, 0x063d96dd, 0x05aedd3e, 0xbd464de6
-       data4   0x8db59154, 0x5d0571c4, 0xd46f0406, 0x15ff6050
-       data4   0xfb241998, 0xe997d6bd, 0x43cc8940, 0x9e7767d9
-       data4   0x42bdb0e8, 0x8b880789, 0x5b38e719, 0xeedb79c8
-       data4   0x0a47a17c, 0x0fe97c42, 0x1ec9f884, 0x00000000
-       data4   0x86830980, 0xed48322b, 0x70ac1e11, 0x724e6c5a
-       data4   0xfffbfd0e, 0x38560f85, 0xd51e3dae, 0x3927362d
-       data4   0xd9640a0f, 0xa621685c, 0x54d19b5b, 0x2e3a2436
-       data4   0x67b10c0a, 0xe70f9357, 0x96d2b4ee, 0x919e1b9b
-       data4   0xc54f80c0, 0x20a261dc, 0x4b695a77, 0x1a161c12
-       data4   0xba0ae293, 0x2ae5c0a0, 0xe0433c22, 0x171d121b
-       data4   0x0d0b0e09, 0xc7adf28b, 0xa8b92db6, 0xa9c8141e
-       data4   0x198557f1, 0x074caf75, 0xddbbee99, 0x60fda37f
-       data4   0x269ff701, 0xf5bc5c72, 0x3bc54466, 0x7e345bfb
-       data4   0x29768b43, 0xc6dccb23, 0xfc68b6ed, 0xf163b8e4
-       data4   0xdccad731, 0x85104263, 0x22401397, 0x112084c6
-       data4   0x247d854a, 0x3df8d2bb, 0x3211aef9, 0xa16dc729
-       data4   0x2f4b1d9e, 0x30f3dcb2, 0x52ec0d86, 0xe3d077c1
-       data4   0x166c2bb3, 0xb999a970, 0x48fa1194, 0x642247e9
-       data4   0x8cc4a8fc, 0x3f1aa0f0, 0x2cd8567d, 0x90ef2233
-       data4   0x4ec78749, 0xd1c1d938, 0xa2fe8cca, 0x0b3698d4
-       data4   0x81cfa6f5, 0xde28a57a, 0x8e26dab7, 0xbfa43fad
-       data4   0x9de42c3a, 0x920d5078, 0xcc9b6a5f, 0x4662547e
-       data4   0x13c2f68d, 0xb8e890d8, 0xf75e2e39, 0xaff582c3
-       data4   0x80be9f5d, 0x937c69d0, 0x2da96fd5, 0x12b3cf25
-       data4   0x993bc8ac, 0x7da71018, 0x636ee89c, 0xbb7bdb3b
-       data4   0x7809cd26, 0x18f46e59, 0xb701ec9a, 0x9aa8834f
-       data4   0x6e65e695, 0xe67eaaff, 0xcf0821bc, 0xe8e6ef15
-       data4   0x9bd9bae7, 0x36ce4a6f, 0x09d4ea9f, 0x7cd629b0
-       data4   0xb2af31a4, 0x23312a3f, 0x9430c6a5, 0x66c035a2
-       data4   0xbc37744e, 0xcaa6fc82, 0xd0b0e090, 0xd81533a7
-       data4   0x984af104, 0xdaf741ec, 0x500e7fcd, 0xf62f1791
-       data4   0xd68d764d, 0xb04d43ef, 0x4d54ccaa, 0x04dfe496
-       data4   0xb5e39ed1, 0x881b4c6a, 0x1fb8c12c, 0x517f4665
-       data4   0xea049d5e, 0x355d018c, 0x7473fa87, 0x412efb0b
-       data4   0x1d5ab367, 0xd25292db, 0x5633e910, 0x47136dd6
-       data4   0x618c9ad7, 0x0c7a37a1, 0x148e59f8, 0x3c89eb13
-       data4   0x27eecea9, 0xc935b761, 0xe5ede11c, 0xb13c7a47
-       data4   0xdf599cd2, 0x733f55f2, 0xce791814, 0x37bf73c7
-       data4   0xcdea53f7, 0xaa5b5ffd, 0x6f14df3d, 0xdb867844
-       data4   0xf381caaf, 0xc43eb968, 0x342c3824, 0x405fc2a3
-       data4   0xc372161d, 0x250cbce2, 0x498b283c, 0x9541ff0d
-       data4   0x017139a8, 0xb3de080c, 0xe49cd8b4, 0xc1906456
-       data4   0x84617bcb, 0xb670d532, 0x5c74486c, 0x5742d0b8
-// Td3:
-       data4   0xf4a75051, 0x4165537e, 0x17a4c31a, 0x275e963a
-       data4   0xab6bcb3b, 0x9d45f11f, 0xfa58abac, 0xe303934b
-       data4   0x30fa5520, 0x766df6ad, 0xcc769188, 0x024c25f5
-       data4   0xe5d7fc4f, 0x2acbd7c5, 0x35448026, 0x62a38fb5
-       data4   0xb15a49de, 0xba1b6725, 0xea0e9845, 0xfec0e15d
-       data4   0x2f7502c3, 0x4cf01281, 0x4697a38d, 0xd3f9c66b
-       data4   0x8f5fe703, 0x929c9515, 0x6d7aebbf, 0x5259da95
-       data4   0xbe832dd4, 0x7421d358, 0xe0692949, 0xc9c8448e
-       data4   0xc2896a75, 0x8e7978f4, 0x583e6b99, 0xb971dd27
-       data4   0xe14fb6be, 0x88ad17f0, 0x20ac66c9, 0xce3ab47d
-       data4   0xdf4a1863, 0x1a3182e5, 0x51336097, 0x537f4562
-       data4   0x6477e0b1, 0x6bae84bb, 0x81a01cfe, 0x082b94f9
-       data4   0x48685870, 0x45fd198f, 0xde6c8794, 0x7bf8b752
-       data4   0x73d323ab, 0x4b02e272, 0x1f8f57e3, 0x55ab2a66
-       data4   0xeb2807b2, 0xb5c2032f, 0xc57b9a86, 0x3708a5d3
-       data4   0x2887f230, 0xbfa5b223, 0x036aba02, 0x16825ced
-       data4   0xcf1c2b8a, 0x79b492a7, 0x07f2f0f3, 0x69e2a14e
-       data4   0xdaf4cd65, 0x05bed506, 0x34621fd1, 0xa6fe8ac4
-       data4   0x2e539d34, 0xf355a0a2, 0x8ae13205, 0xf6eb75a4
-       data4   0x83ec390b, 0x60efaa40, 0x719f065e, 0x6e1051bd
-       data4   0x218af93e, 0xdd063d96, 0x3e05aedd, 0xe6bd464d
-       data4   0x548db591, 0xc45d0571, 0x06d46f04, 0x5015ff60
-       data4   0x98fb2419, 0xbde997d6, 0x4043cc89, 0xd99e7767
-       data4   0xe842bdb0, 0x898b8807, 0x195b38e7, 0xc8eedb79
-       data4   0x7c0a47a1, 0x420fe97c, 0x841ec9f8, 0x00000000
-       data4   0x80868309, 0x2bed4832, 0x1170ac1e, 0x5a724e6c
-       data4   0x0efffbfd, 0x8538560f, 0xaed51e3d, 0x2d392736
-       data4   0x0fd9640a, 0x5ca62168, 0x5b54d19b, 0x362e3a24
-       data4   0x0a67b10c, 0x57e70f93, 0xee96d2b4, 0x9b919e1b
-       data4   0xc0c54f80, 0xdc20a261, 0x774b695a, 0x121a161c
-       data4   0x93ba0ae2, 0xa02ae5c0, 0x22e0433c, 0x1b171d12
-       data4   0x090d0b0e, 0x8bc7adf2, 0xb6a8b92d, 0x1ea9c814
-       data4   0xf1198557, 0x75074caf, 0x99ddbbee, 0x7f60fda3
-       data4   0x01269ff7, 0x72f5bc5c, 0x663bc544, 0xfb7e345b
-       data4   0x4329768b, 0x23c6dccb, 0xedfc68b6, 0xe4f163b8
-       data4   0x31dccad7, 0x63851042, 0x97224013, 0xc6112084
-       data4   0x4a247d85, 0xbb3df8d2, 0xf93211ae, 0x29a16dc7
-       data4   0x9e2f4b1d, 0xb230f3dc, 0x8652ec0d, 0xc1e3d077
-       data4   0xb3166c2b, 0x70b999a9, 0x9448fa11, 0xe9642247
-       data4   0xfc8cc4a8, 0xf03f1aa0, 0x7d2cd856, 0x3390ef22
-       data4   0x494ec787, 0x38d1c1d9, 0xcaa2fe8c, 0xd40b3698
-       data4   0xf581cfa6, 0x7ade28a5, 0xb78e26da, 0xadbfa43f
-       data4   0x3a9de42c, 0x78920d50, 0x5fcc9b6a, 0x7e466254
-       data4   0x8d13c2f6, 0xd8b8e890, 0x39f75e2e, 0xc3aff582
-       data4   0x5d80be9f, 0xd0937c69, 0xd52da96f, 0x2512b3cf
-       data4   0xac993bc8, 0x187da710, 0x9c636ee8, 0x3bbb7bdb
-       data4   0x267809cd, 0x5918f46e, 0x9ab701ec, 0x4f9aa883
-       data4   0x956e65e6, 0xffe67eaa, 0xbccf0821, 0x15e8e6ef
-       data4   0xe79bd9ba, 0x6f36ce4a, 0x9f09d4ea, 0xb07cd629
-       data4   0xa4b2af31, 0x3f23312a, 0xa59430c6, 0xa266c035
-       data4   0x4ebc3774, 0x82caa6fc, 0x90d0b0e0, 0xa7d81533
-       data4   0x04984af1, 0xecdaf741, 0xcd500e7f, 0x91f62f17
-       data4   0x4dd68d76, 0xefb04d43, 0xaa4d54cc, 0x9604dfe4
-       data4   0xd1b5e39e, 0x6a881b4c, 0x2c1fb8c1, 0x65517f46
-       data4   0x5eea049d, 0x8c355d01, 0x877473fa, 0x0b412efb
-       data4   0x671d5ab3, 0xdbd25292, 0x105633e9, 0xd647136d
-       data4   0xd7618c9a, 0xa10c7a37, 0xf8148e59, 0x133c89eb
-       data4   0xa927eece, 0x61c935b7, 0x1ce5ede1, 0x47b13c7a
-       data4   0xd2df599c, 0xf2733f55, 0x14ce7918, 0xc737bf73
-       data4   0xf7cdea53, 0xfdaa5b5f, 0x3d6f14df, 0x44db8678
-       data4   0xaff381ca, 0x68c43eb9, 0x24342c38, 0xa3405fc2
-       data4   0x1dc37216, 0xe2250cbc, 0x3c498b28, 0x0d9541ff
-       data4   0xa8017139, 0x0cb3de08, 0xb4e49cd8, 0x56c19064
-       data4   0xcb84617b, 0x32b670d5, 0x6c5c7448, 0xb85742d0
+.global        AES_Td#
+.type  AES_Td#,@object
+AES_Td:        data4   0x51f4a750,0x51f4a750, 0x7e416553,0x7e416553
+       data4   0x1a17a4c3,0x1a17a4c3, 0x3a275e96,0x3a275e96
+       data4   0x3bab6bcb,0x3bab6bcb, 0x1f9d45f1,0x1f9d45f1
+       data4   0xacfa58ab,0xacfa58ab, 0x4be30393,0x4be30393
+       data4   0x2030fa55,0x2030fa55, 0xad766df6,0xad766df6
+       data4   0x88cc7691,0x88cc7691, 0xf5024c25,0xf5024c25
+       data4   0x4fe5d7fc,0x4fe5d7fc, 0xc52acbd7,0xc52acbd7
+       data4   0x26354480,0x26354480, 0xb562a38f,0xb562a38f
+       data4   0xdeb15a49,0xdeb15a49, 0x25ba1b67,0x25ba1b67
+       data4   0x45ea0e98,0x45ea0e98, 0x5dfec0e1,0x5dfec0e1
+       data4   0xc32f7502,0xc32f7502, 0x814cf012,0x814cf012
+       data4   0x8d4697a3,0x8d4697a3, 0x6bd3f9c6,0x6bd3f9c6
+       data4   0x038f5fe7,0x038f5fe7, 0x15929c95,0x15929c95
+       data4   0xbf6d7aeb,0xbf6d7aeb, 0x955259da,0x955259da
+       data4   0xd4be832d,0xd4be832d, 0x587421d3,0x587421d3
+       data4   0x49e06929,0x49e06929, 0x8ec9c844,0x8ec9c844
+       data4   0x75c2896a,0x75c2896a, 0xf48e7978,0xf48e7978
+       data4   0x99583e6b,0x99583e6b, 0x27b971dd,0x27b971dd
+       data4   0xbee14fb6,0xbee14fb6, 0xf088ad17,0xf088ad17
+       data4   0xc920ac66,0xc920ac66, 0x7dce3ab4,0x7dce3ab4
+       data4   0x63df4a18,0x63df4a18, 0xe51a3182,0xe51a3182
+       data4   0x97513360,0x97513360, 0x62537f45,0x62537f45
+       data4   0xb16477e0,0xb16477e0, 0xbb6bae84,0xbb6bae84
+       data4   0xfe81a01c,0xfe81a01c, 0xf9082b94,0xf9082b94
+       data4   0x70486858,0x70486858, 0x8f45fd19,0x8f45fd19
+       data4   0x94de6c87,0x94de6c87, 0x527bf8b7,0x527bf8b7
+       data4   0xab73d323,0xab73d323, 0x724b02e2,0x724b02e2
+       data4   0xe31f8f57,0xe31f8f57, 0x6655ab2a,0x6655ab2a
+       data4   0xb2eb2807,0xb2eb2807, 0x2fb5c203,0x2fb5c203
+       data4   0x86c57b9a,0x86c57b9a, 0xd33708a5,0xd33708a5
+       data4   0x302887f2,0x302887f2, 0x23bfa5b2,0x23bfa5b2
+       data4   0x02036aba,0x02036aba, 0xed16825c,0xed16825c
+       data4   0x8acf1c2b,0x8acf1c2b, 0xa779b492,0xa779b492
+       data4   0xf307f2f0,0xf307f2f0, 0x4e69e2a1,0x4e69e2a1
+       data4   0x65daf4cd,0x65daf4cd, 0x0605bed5,0x0605bed5
+       data4   0xd134621f,0xd134621f, 0xc4a6fe8a,0xc4a6fe8a
+       data4   0x342e539d,0x342e539d, 0xa2f355a0,0xa2f355a0
+       data4   0x058ae132,0x058ae132, 0xa4f6eb75,0xa4f6eb75
+       data4   0x0b83ec39,0x0b83ec39, 0x4060efaa,0x4060efaa
+       data4   0x5e719f06,0x5e719f06, 0xbd6e1051,0xbd6e1051
+       data4   0x3e218af9,0x3e218af9, 0x96dd063d,0x96dd063d
+       data4   0xdd3e05ae,0xdd3e05ae, 0x4de6bd46,0x4de6bd46
+       data4   0x91548db5,0x91548db5, 0x71c45d05,0x71c45d05
+       data4   0x0406d46f,0x0406d46f, 0x605015ff,0x605015ff
+       data4   0x1998fb24,0x1998fb24, 0xd6bde997,0xd6bde997
+       data4   0x894043cc,0x894043cc, 0x67d99e77,0x67d99e77
+       data4   0xb0e842bd,0xb0e842bd, 0x07898b88,0x07898b88
+       data4   0xe7195b38,0xe7195b38, 0x79c8eedb,0x79c8eedb
+       data4   0xa17c0a47,0xa17c0a47, 0x7c420fe9,0x7c420fe9
+       data4   0xf8841ec9,0xf8841ec9, 0x00000000,0x00000000
+       data4   0x09808683,0x09808683, 0x322bed48,0x322bed48
+       data4   0x1e1170ac,0x1e1170ac, 0x6c5a724e,0x6c5a724e
+       data4   0xfd0efffb,0xfd0efffb, 0x0f853856,0x0f853856
+       data4   0x3daed51e,0x3daed51e, 0x362d3927,0x362d3927
+       data4   0x0a0fd964,0x0a0fd964, 0x685ca621,0x685ca621
+       data4   0x9b5b54d1,0x9b5b54d1, 0x24362e3a,0x24362e3a
+       data4   0x0c0a67b1,0x0c0a67b1, 0x9357e70f,0x9357e70f
+       data4   0xb4ee96d2,0xb4ee96d2, 0x1b9b919e,0x1b9b919e
+       data4   0x80c0c54f,0x80c0c54f, 0x61dc20a2,0x61dc20a2
+       data4   0x5a774b69,0x5a774b69, 0x1c121a16,0x1c121a16
+       data4   0xe293ba0a,0xe293ba0a, 0xc0a02ae5,0xc0a02ae5
+       data4   0x3c22e043,0x3c22e043, 0x121b171d,0x121b171d
+       data4   0x0e090d0b,0x0e090d0b, 0xf28bc7ad,0xf28bc7ad
+       data4   0x2db6a8b9,0x2db6a8b9, 0x141ea9c8,0x141ea9c8
+       data4   0x57f11985,0x57f11985, 0xaf75074c,0xaf75074c
+       data4   0xee99ddbb,0xee99ddbb, 0xa37f60fd,0xa37f60fd
+       data4   0xf701269f,0xf701269f, 0x5c72f5bc,0x5c72f5bc
+       data4   0x44663bc5,0x44663bc5, 0x5bfb7e34,0x5bfb7e34
+       data4   0x8b432976,0x8b432976, 0xcb23c6dc,0xcb23c6dc
+       data4   0xb6edfc68,0xb6edfc68, 0xb8e4f163,0xb8e4f163
+       data4   0xd731dcca,0xd731dcca, 0x42638510,0x42638510
+       data4   0x13972240,0x13972240, 0x84c61120,0x84c61120
+       data4   0x854a247d,0x854a247d, 0xd2bb3df8,0xd2bb3df8
+       data4   0xaef93211,0xaef93211, 0xc729a16d,0xc729a16d
+       data4   0x1d9e2f4b,0x1d9e2f4b, 0xdcb230f3,0xdcb230f3
+       data4   0x0d8652ec,0x0d8652ec, 0x77c1e3d0,0x77c1e3d0
+       data4   0x2bb3166c,0x2bb3166c, 0xa970b999,0xa970b999
+       data4   0x119448fa,0x119448fa, 0x47e96422,0x47e96422
+       data4   0xa8fc8cc4,0xa8fc8cc4, 0xa0f03f1a,0xa0f03f1a
+       data4   0x567d2cd8,0x567d2cd8, 0x223390ef,0x223390ef
+       data4   0x87494ec7,0x87494ec7, 0xd938d1c1,0xd938d1c1
+       data4   0x8ccaa2fe,0x8ccaa2fe, 0x98d40b36,0x98d40b36
+       data4   0xa6f581cf,0xa6f581cf, 0xa57ade28,0xa57ade28
+       data4   0xdab78e26,0xdab78e26, 0x3fadbfa4,0x3fadbfa4
+       data4   0x2c3a9de4,0x2c3a9de4, 0x5078920d,0x5078920d
+       data4   0x6a5fcc9b,0x6a5fcc9b, 0x547e4662,0x547e4662
+       data4   0xf68d13c2,0xf68d13c2, 0x90d8b8e8,0x90d8b8e8
+       data4   0x2e39f75e,0x2e39f75e, 0x82c3aff5,0x82c3aff5
+       data4   0x9f5d80be,0x9f5d80be, 0x69d0937c,0x69d0937c
+       data4   0x6fd52da9,0x6fd52da9, 0xcf2512b3,0xcf2512b3
+       data4   0xc8ac993b,0xc8ac993b, 0x10187da7,0x10187da7
+       data4   0xe89c636e,0xe89c636e, 0xdb3bbb7b,0xdb3bbb7b
+       data4   0xcd267809,0xcd267809, 0x6e5918f4,0x6e5918f4
+       data4   0xec9ab701,0xec9ab701, 0x834f9aa8,0x834f9aa8
+       data4   0xe6956e65,0xe6956e65, 0xaaffe67e,0xaaffe67e
+       data4   0x21bccf08,0x21bccf08, 0xef15e8e6,0xef15e8e6
+       data4   0xbae79bd9,0xbae79bd9, 0x4a6f36ce,0x4a6f36ce
+       data4   0xea9f09d4,0xea9f09d4, 0x29b07cd6,0x29b07cd6
+       data4   0x31a4b2af,0x31a4b2af, 0x2a3f2331,0x2a3f2331
+       data4   0xc6a59430,0xc6a59430, 0x35a266c0,0x35a266c0
+       data4   0x744ebc37,0x744ebc37, 0xfc82caa6,0xfc82caa6
+       data4   0xe090d0b0,0xe090d0b0, 0x33a7d815,0x33a7d815
+       data4   0xf104984a,0xf104984a, 0x41ecdaf7,0x41ecdaf7
+       data4   0x7fcd500e,0x7fcd500e, 0x1791f62f,0x1791f62f
+       data4   0x764dd68d,0x764dd68d, 0x43efb04d,0x43efb04d
+       data4   0xccaa4d54,0xccaa4d54, 0xe49604df,0xe49604df
+       data4   0x9ed1b5e3,0x9ed1b5e3, 0x4c6a881b,0x4c6a881b
+       data4   0xc12c1fb8,0xc12c1fb8, 0x4665517f,0x4665517f
+       data4   0x9d5eea04,0x9d5eea04, 0x018c355d,0x018c355d
+       data4   0xfa877473,0xfa877473, 0xfb0b412e,0xfb0b412e
+       data4   0xb3671d5a,0xb3671d5a, 0x92dbd252,0x92dbd252
+       data4   0xe9105633,0xe9105633, 0x6dd64713,0x6dd64713
+       data4   0x9ad7618c,0x9ad7618c, 0x37a10c7a,0x37a10c7a
+       data4   0x59f8148e,0x59f8148e, 0xeb133c89,0xeb133c89
+       data4   0xcea927ee,0xcea927ee, 0xb761c935,0xb761c935
+       data4   0xe11ce5ed,0xe11ce5ed, 0x7a47b13c,0x7a47b13c
+       data4   0x9cd2df59,0x9cd2df59, 0x55f2733f,0x55f2733f
+       data4   0x1814ce79,0x1814ce79, 0x73c737bf,0x73c737bf
+       data4   0x53f7cdea,0x53f7cdea, 0x5ffdaa5b,0x5ffdaa5b
+       data4   0xdf3d6f14,0xdf3d6f14, 0x7844db86,0x7844db86
+       data4   0xcaaff381,0xcaaff381, 0xb968c43e,0xb968c43e
+       data4   0x3824342c,0x3824342c, 0xc2a3405f,0xc2a3405f
+       data4   0x161dc372,0x161dc372, 0xbce2250c,0xbce2250c
+       data4   0x283c498b,0x283c498b, 0xff0d9541,0xff0d9541
+       data4   0x39a80171,0x39a80171, 0x080cb3de,0x080cb3de
+       data4   0xd8b4e49c,0xd8b4e49c, 0x6456c190,0x6456c190
+       data4   0x7bcb8461,0x7bcb8461, 0xd532b670,0xd532b670
+       data4   0x486c5c74,0x486c5c74, 0xd0b85742,0xd0b85742
 // Td4:
-       data4   0x52525252, 0x09090909, 0x6a6a6a6a, 0xd5d5d5d5
-       data4   0x30303030, 0x36363636, 0xa5a5a5a5, 0x38383838
-       data4   0xbfbfbfbf, 0x40404040, 0xa3a3a3a3, 0x9e9e9e9e
-       data4   0x81818181, 0xf3f3f3f3, 0xd7d7d7d7, 0xfbfbfbfb
-       data4   0x7c7c7c7c, 0xe3e3e3e3, 0x39393939, 0x82828282
-       data4   0x9b9b9b9b, 0x2f2f2f2f, 0xffffffff, 0x87878787
-       data4   0x34343434, 0x8e8e8e8e, 0x43434343, 0x44444444
-       data4   0xc4c4c4c4, 0xdededede, 0xe9e9e9e9, 0xcbcbcbcb
-       data4   0x54545454, 0x7b7b7b7b, 0x94949494, 0x32323232
-       data4   0xa6a6a6a6, 0xc2c2c2c2, 0x23232323, 0x3d3d3d3d
-       data4   0xeeeeeeee, 0x4c4c4c4c, 0x95959595, 0x0b0b0b0b
-       data4   0x42424242, 0xfafafafa, 0xc3c3c3c3, 0x4e4e4e4e
-       data4   0x08080808, 0x2e2e2e2e, 0xa1a1a1a1, 0x66666666
-       data4   0x28282828, 0xd9d9d9d9, 0x24242424, 0xb2b2b2b2
-       data4   0x76767676, 0x5b5b5b5b, 0xa2a2a2a2, 0x49494949
-       data4   0x6d6d6d6d, 0x8b8b8b8b, 0xd1d1d1d1, 0x25252525
-       data4   0x72727272, 0xf8f8f8f8, 0xf6f6f6f6, 0x64646464
-       data4   0x86868686, 0x68686868, 0x98989898, 0x16161616
-       data4   0xd4d4d4d4, 0xa4a4a4a4, 0x5c5c5c5c, 0xcccccccc
-       data4   0x5d5d5d5d, 0x65656565, 0xb6b6b6b6, 0x92929292
-       data4   0x6c6c6c6c, 0x70707070, 0x48484848, 0x50505050
-       data4   0xfdfdfdfd, 0xedededed, 0xb9b9b9b9, 0xdadadada
-       data4   0x5e5e5e5e, 0x15151515, 0x46464646, 0x57575757
-       data4   0xa7a7a7a7, 0x8d8d8d8d, 0x9d9d9d9d, 0x84848484
-       data4   0x90909090, 0xd8d8d8d8, 0xabababab, 0x00000000
-       data4   0x8c8c8c8c, 0xbcbcbcbc, 0xd3d3d3d3, 0x0a0a0a0a
-       data4   0xf7f7f7f7, 0xe4e4e4e4, 0x58585858, 0x05050505
-       data4   0xb8b8b8b8, 0xb3b3b3b3, 0x45454545, 0x06060606
-       data4   0xd0d0d0d0, 0x2c2c2c2c, 0x1e1e1e1e, 0x8f8f8f8f
-       data4   0xcacacaca, 0x3f3f3f3f, 0x0f0f0f0f, 0x02020202
-       data4   0xc1c1c1c1, 0xafafafaf, 0xbdbdbdbd, 0x03030303
-       data4   0x01010101, 0x13131313, 0x8a8a8a8a, 0x6b6b6b6b
-       data4   0x3a3a3a3a, 0x91919191, 0x11111111, 0x41414141
-       data4   0x4f4f4f4f, 0x67676767, 0xdcdcdcdc, 0xeaeaeaea
-       data4   0x97979797, 0xf2f2f2f2, 0xcfcfcfcf, 0xcececece
-       data4   0xf0f0f0f0, 0xb4b4b4b4, 0xe6e6e6e6, 0x73737373
-       data4   0x96969696, 0xacacacac, 0x74747474, 0x22222222
-       data4   0xe7e7e7e7, 0xadadadad, 0x35353535, 0x85858585
-       data4   0xe2e2e2e2, 0xf9f9f9f9, 0x37373737, 0xe8e8e8e8
-       data4   0x1c1c1c1c, 0x75757575, 0xdfdfdfdf, 0x6e6e6e6e
-       data4   0x47474747, 0xf1f1f1f1, 0x1a1a1a1a, 0x71717171
-       data4   0x1d1d1d1d, 0x29292929, 0xc5c5c5c5, 0x89898989
-       data4   0x6f6f6f6f, 0xb7b7b7b7, 0x62626262, 0x0e0e0e0e
-       data4   0xaaaaaaaa, 0x18181818, 0xbebebebe, 0x1b1b1b1b
-       data4   0xfcfcfcfc, 0x56565656, 0x3e3e3e3e, 0x4b4b4b4b
-       data4   0xc6c6c6c6, 0xd2d2d2d2, 0x79797979, 0x20202020
-       data4   0x9a9a9a9a, 0xdbdbdbdb, 0xc0c0c0c0, 0xfefefefe
-       data4   0x78787878, 0xcdcdcdcd, 0x5a5a5a5a, 0xf4f4f4f4
-       data4   0x1f1f1f1f, 0xdddddddd, 0xa8a8a8a8, 0x33333333
-       data4   0x88888888, 0x07070707, 0xc7c7c7c7, 0x31313131
-       data4   0xb1b1b1b1, 0x12121212, 0x10101010, 0x59595959
-       data4   0x27272727, 0x80808080, 0xecececec, 0x5f5f5f5f
-       data4   0x60606060, 0x51515151, 0x7f7f7f7f, 0xa9a9a9a9
-       data4   0x19191919, 0xb5b5b5b5, 0x4a4a4a4a, 0x0d0d0d0d
-       data4   0x2d2d2d2d, 0xe5e5e5e5, 0x7a7a7a7a, 0x9f9f9f9f
-       data4   0x93939393, 0xc9c9c9c9, 0x9c9c9c9c, 0xefefefef
-       data4   0xa0a0a0a0, 0xe0e0e0e0, 0x3b3b3b3b, 0x4d4d4d4d
-       data4   0xaeaeaeae, 0x2a2a2a2a, 0xf5f5f5f5, 0xb0b0b0b0
-       data4   0xc8c8c8c8, 0xebebebeb, 0xbbbbbbbb, 0x3c3c3c3c
-       data4   0x83838383, 0x53535353, 0x99999999, 0x61616161
-       data4   0x17171717, 0x2b2b2b2b, 0x04040404, 0x7e7e7e7e
-       data4   0xbabababa, 0x77777777, 0xd6d6d6d6, 0x26262626
-       data4   0xe1e1e1e1, 0x69696969, 0x14141414, 0x63636363
-       data4   0x55555555, 0x21212121, 0x0c0c0c0c, 0x7d7d7d7d
-.size  Td0#,5*256*4    // HP-UX assembler fails to ".-Td0#"
+       data1   0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
+       data1   0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
+       data1   0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
+       data1   0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
+       data1   0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
+       data1   0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
+       data1   0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
+       data1   0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
+       data1   0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
+       data1   0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
+       data1   0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
+       data1   0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
+       data1   0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
+       data1   0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
+       data1   0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
+       data1   0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
+       data1   0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
+       data1   0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
+       data1   0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
+       data1   0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
+       data1   0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
+       data1   0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
+       data1   0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
+       data1   0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
+       data1   0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
+       data1   0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
+       data1   0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
+       data1   0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
+       data1   0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
+       data1   0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
+       data1   0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
+       data1   0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
+.size  AES_Td#,2048+256        // HP-UX assembler fails to ".-AES_Td#"