.file "__vsqrt.S"
#include "libm.h"
RO_DATA
.align 64
.CONST_TBL:
.word 0x3fe00000, 0x00000000 ! A1 = 5.00000000000000001789e-01
.word 0xbfbfffff, 0xfffd0bfd ! A2 = -1.24999999997314110667e-01
.word 0x3fafffff, 0xfffb5bfb ! A3 = 6.24999999978896565817e-02
.word 0xbfa4000f, 0xc00b4fc8 ! A4 = -3.90629693917215481458e-02
.word 0x3f9c0018, 0xc012da4e ! A5 = 2.73441188080261677282e-02
.word 0x000fffff, 0xffffffff ! DC0 = 0x000fffffffffffff
.word 0x00001000, 0x00000000 ! DC2 = 0x0000100000000000
.word 0x7fffe000, 0x00000000 ! DC3 = 0x7fffe00000000000
! i = [0,128]
! TBL[8*i+0] = 1.0 / (*(double*)&(0x3fe0000000000000LL + (i << 45)));
! TBL[8*i+1] = (double)(2.0 * sqrtl(*(double*)&(0x3fe0000000000000LL + (i << 45))));
! TBL[8*i+2] = (double)(2.0 * sqrtl(*(double*)&(0x3fe0000000000000LL + (i << 45))) - TBL[8*i+1]);
! TBL[8*i+3] = 0
! TBL[8*i+4] = 1.0 / (*(double*)&(0x3fe0000000000000LL + (i << 45)));
! TBL[8*i+5] = (double)(2.0 * sqrtl(2.0) * sqrtl(*(double*)&(0x3fe0000000000000LL + (i << 45))));
! TBL[8*i+6] = (double)(2.0 * sqrtl(2.0) * sqrtl(*(double*)&(0x3fe0000000000000LL + (i << 45))) - TBL[8*i+5]);
! TBL[8*i+7] = 0
.word 0x40000000, 0x00000000, 0x3ff6a09e, 0x667f3bcd
.word 0xbc9bdd34, 0x13b26456, 0x00000000, 0x00000000
.word 0x40000000, 0x00000000, 0x40000000, 0x00000000
.word 0xb8f00000, 0x00000000, 0x00000000, 0x00000000
.word 0x3fffc07f, 0x01fc07f0, 0x3ff6b733, 0xbfd8c648
.word 0x3c53b629, 0x05629048, 0x00000000, 0x00000000
.word 0x3fffc07f, 0x01fc07f0, 0x40000ff8, 0x07f60deb
.word 0x3c90655c, 0x648a53f1, 0x00000000, 0x00000000
.word 0x3fff81f8, 0x1f81f820, 0x3ff6cdb2, 0xbbb212eb
.word 0x3c960332, 0xcdbaba2d, 0x00000000, 0x00000000
.word 0x3fff81f8, 0x1f81f820, 0x40001fe0, 0x3f61bad0
.word 0x3ca2c41a, 0x15cbfaf2, 0x00000000, 0x00000000
.word 0x3fff4465, 0x9e4a4271, 0x3ff6e41b, 0x9bfb3b75
.word 0xbc925d8c, 0xfd6d5c87, 0x00000000, 0x00000000
.word 0x3fff4465, 0x9e4a4271, 0x40002fb8, 0xd4e30f48
.word 0xbca64203, 0xab1ba910, 0x00000000, 0x00000000
.word 0x3fff07c1, 0xf07c1f08, 0x3ff6fa6e, 0xa162d0f0
.word 0x3c691a24, 0x3d6297e9, 0x00000000, 0x00000000
.word 0x3fff07c1, 0xf07c1f08, 0x40003f81, 0xf636b80c
.word 0xbca0efc8, 0xba812a8c, 0x00000000, 0x00000000
.word 0x3ffecc07, 0xb301ecc0, 0x3ff710ac, 0x0b5e5e32
.word 0xbc991218, 0xb8d2850d, 0x00000000, 0x00000000
.word 0x3ffecc07, 0xb301ecc0, 0x40004f3b, 0xd03c0a64
.word 0x3c9ee2cf, 0x2d8ae22b, 0x00000000, 0x00000000
.word 0x3ffe9131, 0xabf0b767, 0x3ff726d4, 0x1832a0be
.word 0xbc2d9b1a, 0xa8ecb058, 0x00000000, 0x00000000
.word 0x3ffe9131, 0xabf0b767, 0x40005ee6, 0x8efad48b
.word 0xbc9c35f4, 0x8f4b89f7, 0x00000000, 0x00000000
.word 0x3ffe573a, 0xc901e574, 0x3ff73ce7, 0x04fb7b23
.word 0x3c91470b, 0x816b17a6, 0x00000000, 0x00000000
.word 0x3ffe573a, 0xc901e574, 0x40006e82, 0x5da8fc2b
.word 0x3c9a315a, 0x8bd8a03b, 0x00000000, 0x00000000
.word 0x3ffe1e1e, 0x1e1e1e1e, 0x3ff752e5, 0x0db3a3a2
.word 0xbc939331, 0x3eea4381, 0x00000000, 0x00000000
.word 0x3ffe1e1e, 0x1e1e1e1e, 0x40007e0f, 0x66afed07
.word 0xbc74a6e1, 0xdcd59eaf, 0x00000000, 0x00000000
.word 0x3ffde5d6, 0xe3f8868a, 0x3ff768ce, 0x6d3c11e0
.word 0xbc9478b8, 0xab33074d, 0x00000000, 0x00000000
.word 0x3ffde5d6, 0xe3f8868a, 0x40008d8d, 0xd3b1d9aa
.word 0x3c81d533, 0x85fe2b96, 0x00000000, 0x00000000
.word 0x3ffdae60, 0x76b981db, 0x3ff77ea3, 0x5d632e43
.word 0x3c92f714, 0x9a22fa4f, 0x00000000, 0x00000000
.word 0x3ffdae60, 0x76b981db, 0x40009cfd, 0xcd8ed009
.word 0xbc4862a9, 0xbcf7f372, 0x00000000, 0x00000000
.word 0x3ffd77b6, 0x54b82c34, 0x3ff79464, 0x16ebc56c
.word 0x3c9a7cd5, 0x224c7375, 0x00000000, 0x00000000
.word 0x3ffd77b6, 0x54b82c34, 0x4000ac5f, 0x7c69a3c8
.word 0x3ca94dff, 0x7bfa2757, 0x00000000, 0x00000000
.word 0x3ffd41d4, 0x1d41d41d, 0x3ff7aa10, 0xd193c22d
.word 0xbc790ed9, 0x403afe85, 0x00000000, 0x00000000
.word 0x3ffd41d4, 0x1d41d41d, 0x4000bbb3, 0x07acafdb
.word 0xbc852a97, 0x686f9d2e, 0x00000000, 0x00000000
.word 0x3ffd0cb5, 0x8f6ec074, 0x3ff7bfa9, 0xc41ab040
.word 0x3c8d6bc3, 0x02ae758f, 0x00000000, 0x00000000
.word 0x3ffd0cb5, 0x8f6ec074, 0x4000caf8, 0x960e710d
.word 0x3c9caa6b, 0xe2366171, 0x00000000, 0x00000000
.word 0x3ffcd856, 0x89039b0b, 0x3ff7d52f, 0x244809e9
.word 0x3c9081f6, 0xf3b99d5f, 0x00000000, 0x00000000
.word 0x3ffcd856, 0x89039b0b, 0x4000da30, 0x4d95fb06
.word 0xbc9e1269, 0x76855586, 0x00000000, 0x00000000
.word 0x3ffca4b3, 0x055ee191, 0x3ff7eaa1, 0x26f15284
.word 0xbc846ce4, 0x68c1882b, 0x00000000, 0x00000000
.word 0x3ffca4b3, 0x055ee191, 0x4000e95a, 0x539f492c
.word 0xbc80c73f, 0xc38a2184, 0x00000000, 0x00000000
.word 0x3ffc71c7, 0x1c71c71c, 0x3ff80000, 0x00000000
.word 0x00000000, 0x00000000, 0x00000000, 0x00000000
.word 0x3ffc71c7, 0x1c71c71c, 0x4000f876, 0xccdf6cd9
.word 0x3cab1a18, 0xf13a34c0, 0x00000000, 0x00000000
.word 0x3ffc3f8f, 0x01c3f8f0, 0x3ff8154b, 0xe2773526
.word 0xbc857147, 0xe067d0ee, 0x00000000, 0x00000000
.word 0x3ffc3f8f, 0x01c3f8f0, 0x40010785, 0xdd689a29
.word 0xbcaaabbe, 0x9e4d810a, 0x00000000, 0x00000000
.word 0x3ffc0e07, 0x0381c0e0, 0x3ff82a85, 0x00794e6c
.word 0xbc82edaa, 0x75e6ac5f, 0x00000000, 0x00000000
.word 0x3ffc0e07, 0x0381c0e0, 0x40011687, 0xa8ae14a3
.word 0x3cac9b43, 0xbcf06106, 0x00000000, 0x00000000
.word 0x3ffbdd2b, 0x899406f7, 0x3ff83fab, 0x8b4d4315
.word 0x3c829e06, 0x2d3e134d, 0x00000000, 0x00000000
.word 0x3ffbdd2b, 0x899406f7, 0x4001257c, 0x5187fd09
.word 0xbca4a750, 0xa83950a4, 0x00000000, 0x00000000
.word 0x3ffbacf9, 0x14c1bad0, 0x3ff854bf, 0xb363dc39
.word 0x3c99399f, 0xca38787e, 0x00000000, 0x00000000
.word 0x3ffbacf9, 0x14c1bad0, 0x40013463, 0xfa37014e
.word 0x3c7b295b, 0xaa698cd3, 0x00000000, 0x00000000
.word 0x3ffb7d6c, 0x3dda338b, 0x3ff869c1, 0xa85cc346
.word 0x3c9fcc99, 0xde11b1d1, 0x00000000, 0x00000000
.word 0x3ffb7d6c, 0x3dda338b, 0x4001433e, 0xc467effb
.word 0x3c92c031, 0x3b7278c8, 0x00000000, 0x00000000
.word 0x3ffb4e81, 0xb4e81b4f, 0x3ff87eb1, 0x990b697a
.word 0x3c7c43e9, 0xf593ea0f, 0x00000000, 0x00000000
.word 0x3ffb4e81, 0xb4e81b4f, 0x4001520c, 0xd1372feb
.word 0xbcadec22, 0x5d8e66d2, 0x00000000, 0x00000000
.word 0x3ffb2036, 0x406c80d9, 0x3ff8938f, 0xb37bc9c1
.word 0xbc7c115f, 0x9f5c8d6f, 0x00000000, 0x00000000
.word 0x3ffb2036, 0x406c80d9, 0x400160ce, 0x41341d74
.word 0x3c967036, 0x863a1bb2, 0x00000000, 0x00000000
.word 0x3ffaf286, 0xbca1af28, 0x3ff8a85c, 0x24f70659
.word 0x3c9f6e07, 0x6b588a50, 0x00000000, 0x00000000
.word 0x3ffaf286, 0xbca1af28, 0x40016f83, 0x34644df9
.word 0xbcae8679, 0x80a1c48e, 0x00000000, 0x00000000
.word 0x3ffac570, 0x1ac5701b, 0x3ff8bd17, 0x1a07e38a
.word 0x3c9c20b5, 0xa697f23f, 0x00000000, 0x00000000
.word 0x3ffac570, 0x1ac5701b, 0x40017e2b, 0xca46bab9
.word 0x3ca1519b, 0x10d04d5f, 0x00000000, 0x00000000
.word 0x3ffa98ef, 0x606a63be, 0x3ff8d1c0, 0xbe7f20ac
.word 0xbc8bdb8a, 0x6df021f3, 0x00000000, 0x00000000
.word 0x3ffa98ef, 0x606a63be, 0x40018cc8, 0x21d6d3e3
.word 0xbca30af1, 0xd725cc5b, 0x00000000, 0x00000000
.word 0x3ffa6d01, 0xa6d01a6d, 0x3ff8e659, 0x3d77b0b8
.word 0xbc7d99d7, 0x64769954, 0x00000000, 0x00000000
.word 0x3ffa6d01, 0xa6d01a6d, 0x40019b58, 0x598f7c9f
.word 0xbc72e0d8, 0x51c0e011, 0x00000000, 0x00000000
.word 0x3ffa41a4, 0x1a41a41a, 0x3ff8fae0, 0xc15ad38a
.word 0xbc7db7ad, 0xb6817f6d, 0x00000000, 0x00000000
.word 0x3ffa41a4, 0x1a41a41a, 0x4001a9dc, 0x8f6df104
.word 0xbcafc519, 0xc18dc1d5, 0x00000000, 0x00000000
.word 0x3ffa16d3, 0xf97a4b02, 0x3ff90f57, 0x73e410e4
.word 0x3c6fb605, 0xcee75482, 0x00000000, 0x00000000
.word 0x3ffa16d3, 0xf97a4b02, 0x4001b854, 0xe0f496a0
.word 0x3ca27006, 0x899b7c3a, 0x00000000, 0x00000000
.word 0x3ff9ec8e, 0x951033d9, 0x3ff923bd, 0x7e25164d
.word 0xbc9278d1, 0x901d3b40, 0x00000000, 0x00000000
.word 0x3ff9ec8e, 0x951033d9, 0x4001c6c1, 0x6b2db870
.word 0x3c887e1d, 0x8335fb28, 0x00000000, 0x00000000
.word 0x3ff9c2d1, 0x4ee4a102, 0x3ff93813, 0x088978c5
.word 0xbc54312c, 0x627e5c52, 0x00000000, 0x00000000
.word 0x3ff9c2d1, 0x4ee4a102, 0x4001d522, 0x4aae2ee1
.word 0x3ca91222, 0xf6aebdc9, 0x00000000, 0x00000000
.word 0x3ff99999, 0x9999999a, 0x3ff94c58, 0x3ada5b53
.word 0xbc9b7ed7, 0x50df3cca, 0x00000000, 0x00000000
.word 0x3ff99999, 0x9999999a, 0x4001e377, 0x9b97f4a8
.word 0xbc9f5063, 0x19fcfd19, 0x00000000, 0x00000000
.word 0x3ff970e4, 0xf80cb872, 0x3ff9608d, 0x3c41fb4b
.word 0x3c73df32, 0xeaa86b83, 0x00000000, 0x00000000
.word 0x3ff970e4, 0xf80cb872, 0x4001f1c1, 0x799ca8ff
.word 0xbca28b52, 0xeb725e0a, 0x00000000, 0x00000000
.word 0x3ff948b0, 0xfcd6e9e0, 0x3ff974b2, 0x334f2346
.word 0x3c814e4a, 0xd3ae9e3f, 0x00000000, 0x00000000
.word 0x3ff948b0, 0xfcd6e9e0, 0x40020000, 0x00000000
.word 0xb9000000, 0x00000000, 0x00000000, 0x00000000
.word 0x3ff920fb, 0x49d0e229, 0x3ff988c7, 0x45f88592
.word 0x3c95af70, 0x1a56047b, 0x00000000, 0x00000000
.word 0x3ff920fb, 0x49d0e229, 0x40020e33, 0x499a21a9
.word 0xbc924ba2, 0x74fea9a1, 0x00000000, 0x00000000
.word 0x3ff8f9c1, 0x8f9c18fa, 0x3ff99ccc, 0x999fff00
.word 0x3c866234, 0x063b88ee, 0x00000000, 0x00000000
.word 0x3ff8f9c1, 0x8f9c18fa, 0x40021c5b, 0x70d9f824
.word 0xbca844f9, 0x9eee6fc3, 0x00000000, 0x00000000
.word 0x3ff8d301, 0x8d3018d3, 0x3ff9b0c2, 0x5315c2ce
.word 0xbc87f64a, 0x65cc6887, 0x00000000, 0x00000000
.word 0x3ff8d301, 0x8d3018d3, 0x40022a78, 0x8fc76de5
.word 0x3c931e32, 0xd4e07a48, 0x00000000, 0x00000000
.word 0x3ff8acb9, 0x0f6bf3aa, 0x3ff9c4a8, 0x969b7077
.word 0xbc96ca9e, 0x5cd4517a, 0x00000000, 0x00000000
.word 0x3ff8acb9, 0x0f6bf3aa, 0x4002388a, 0xc0059c28
.word 0xbc96072f, 0xbe0e5da3, 0x00000000, 0x00000000
.word 0x3ff886e5, 0xf0abb04a, 0x3ff9d87f, 0x87e71422
.word 0xbc85fdd8, 0xb11b7b1d, 0x00000000, 0x00000000
.word 0x3ff886e5, 0xf0abb04a, 0x40024692, 0x1ad4ea49
.word 0xbcaa6d9b, 0x268ef62d, 0x00000000, 0x00000000
.word 0x3ff86186, 0x18618618, 0x3ff9ec47, 0x4a261264
.word 0xbc8540c4, 0x89ba5074, 0x00000000, 0x00000000
.word 0x3ff86186, 0x18618618, 0x4002548e, 0xb9151e85
.word 0x3c999820, 0x0a774879, 0x00000000, 0x00000000
.word 0x3ff83c97, 0x7ab2bedd, 0x3ffa0000, 0x00000000
.word 0x00000000, 0x00000000, 0x00000000, 0x00000000
.word 0x3ff83c97, 0x7ab2bedd, 0x40026280, 0xb3476096
.word 0x3c9ab88b, 0x5ffe1cf5, 0x00000000, 0x00000000
.word 0x3ff81818, 0x18181818, 0x3ffa13a9, 0xcb996651
.word 0xbc9f9ab9, 0x0e4e85c3, 0x00000000, 0x00000000
.word 0x3ff81818, 0x18181818, 0x40027068, 0x21902e9a
.word 0x3c90ff4c, 0x20f541f6, 0x00000000, 0x00000000
.word 0x3ff7f405, 0xfd017f40, 0x3ffa2744, 0xce9674f5
.word 0xbc8b936c, 0x81e54daa, 0x00000000, 0x00000000
.word 0x3ff7f405, 0xfd017f40, 0x40027e45, 0x1bb944c3
.word 0x3c8e4a16, 0x42099ef0, 0x00000000, 0x00000000
.word 0x3ff7d05f, 0x417d05f4, 0x3ffa3ad1, 0x2a1da160
.word 0x3c951168, 0xf4be5984, 0x00000000, 0x00000000
.word 0x3ff7d05f, 0x417d05f4, 0x40028c17, 0xb9337834
.word 0xbc8af150, 0xa0e88972, 0x00000000, 0x00000000
.word 0x3ff7ad22, 0x08e0ecc3, 0x3ffa4e4e, 0xfeda34de
.word 0x3c6afbb4, 0xdbdadd0d, 0x00000000, 0x00000000
.word 0x3ff7ad22, 0x08e0ecc3, 0x400299e0, 0x11188575
.word 0xbc9a6169, 0x3fb250e5, 0x00000000, 0x00000000
.word 0x3ff78a4c, 0x8178a4c8, 0x3ffa61be, 0x6cfec997
.word 0xbc8c37ea, 0xb2bb5ca0, 0x00000000, 0x00000000
.word 0x3ff78a4c, 0x8178a4c8, 0x4002a79e, 0x3a2cd2e6
.word 0xbca5ddd4, 0x9cc9ad59, 0x00000000, 0x00000000
.word 0x3ff767dc, 0xe434a9b1, 0x3ffa751f, 0x9447b724
.word 0x3c82b909, 0x477e9ed1, 0x00000000, 0x00000000
.word 0x3ff767dc, 0xe434a9b1, 0x4002b552, 0x4ae1278e
.word 0xbca2f2a9, 0x8841b934, 0x00000000, 0x00000000
.word 0x3ff745d1, 0x745d1746, 0x3ffa8872, 0x93fd6f34
.word 0x3c768ef2, 0x4f198721, 0x00000000, 0x00000000
.word 0x3ff745d1, 0x745d1746, 0x4002c2fc, 0x595456a7
.word 0xbc996f60, 0xb0fc7e96, 0x00000000, 0x00000000
.word 0x3ff72428, 0x7f46debc, 0x3ffa9bb7, 0x8af6cabc
.word 0x3c8ba60d, 0xc999aba7, 0x00000000, 0x00000000
.word 0x3ff72428, 0x7f46debc, 0x4002d09c, 0x7b54e03e
.word 0x3c98c747, 0xfdeda6de, 0x00000000, 0x00000000
.word 0x3ff702e0, 0x5c0b8170, 0x3ffaaeee, 0x979b4838
.word 0xbc91f08a, 0xef9ef6c0, 0x00000000, 0x00000000
.word 0x3ff702e0, 0x5c0b8170, 0x4002de32, 0xc6628741
.word 0x3ca78746, 0xc499a4f7, 0x00000000, 0x00000000
.word 0x3ff6e1f7, 0x6b4337c7, 0x3ffac217, 0xd7e53b66
.word 0xbc64282a, 0xaa967e4f, 0x00000000, 0x00000000
.word 0x3ff6e1f7, 0x6b4337c7, 0x4002ebbf, 0x4fafdd4b
.word 0xbca78a73, 0xb72d5c41, 0x00000000, 0x00000000
.word 0x3ff6c16c, 0x16c16c17, 0x3ffad533, 0x6963eefc
.word 0xbc977c4a, 0x537dbdd2, 0x00000000, 0x00000000
.word 0x3ff6c16c, 0x16c16c17, 0x4002f942, 0x2c23c47e
.word 0xbc827c85, 0xf29db65d, 0x00000000, 0x00000000
.word 0x3ff6a13c, 0xd1537290, 0x3ffae841, 0x693db8b4
.word 0x3c90f773, 0xcd7a0713, 0x00000000, 0x00000000
.word 0x3ff6a13c, 0xd1537290, 0x400306bb, 0x705ae7c3
.word 0x3caf4933, 0x907af47a, 0x00000000, 0x00000000
.word 0x3ff68168, 0x16816817, 0x3ffafb41, 0xf432002e
.word 0xbc7ac94a, 0xfdfe8c5b, 0x00000000, 0x00000000
.word 0x3ff68168, 0x16816817, 0x4003142b, 0x30a929ab
.word 0x3c98dc01, 0x081a6c5c, 0x00000000, 0x00000000
.word 0x3ff661ec, 0x6a5122f9, 0x3ffb0e35, 0x269b38f5
.word 0xbc4f69a8, 0x05c3271a, 0x00000000, 0x00000000
.word 0x3ff661ec, 0x6a5122f9, 0x40032191, 0x811b0a41
.word 0xbc9ce3f0, 0xb38c0bf7, 0x00000000, 0x00000000
.word 0x3ff642c8, 0x590b2164, 0x3ffb211b, 0x1c70d023
.word 0x3c2e4c5e, 0x66eae2f0, 0x00000000, 0x00000000
.word 0x3ff642c8, 0x590b2164, 0x40032eee, 0x75770416
.word 0x3caed8e7, 0x730eaff2, 0x00000000, 0x00000000
.word 0x3ff623fa, 0x77016240, 0x3ffb33f3, 0xf1490def
.word 0xbc95894b, 0xcb02373b, 0x00000000, 0x00000000
.word 0x3ff623fa, 0x77016240, 0x40033c42, 0x213ee0c9
.word 0x3ca84c24, 0x4ba98124, 0x00000000, 0x00000000
.word 0x3ff60581, 0x60581606, 0x3ffb46bf, 0xc05aeb89
.word 0x3c9b1c7c, 0xc39adc9f, 0x00000000, 0x00000000
.word 0x3ff60581, 0x60581606, 0x4003498c, 0x97b10540
.word 0x3c734193, 0xbc8543b4, 0x00000000, 0x00000000
.word 0x3ff5e75b, 0xb8d015e7, 0x3ffb597e, 0xa47fdda3
.word 0xbc923cc8, 0x9d1e4635, 0x00000000, 0x00000000
.word 0x3ff5e75b, 0xb8d015e7, 0x400356cd, 0xebc9b5e2
.word 0x3c96dee1, 0x46bb1571, 0x00000000, 0x00000000
.word 0x3ff5c988, 0x2b931057, 0x3ffb6c30, 0xb83593e6
.word 0x3c8f4e3f, 0xd28d84bc, 0x00000000, 0x00000000
.word 0x3ff5c988, 0x2b931057, 0x40036406, 0x30445306
.word 0xbca78d86, 0x2327430a, 0x00000000, 0x00000000
.word 0x3ff5ac05, 0x6b015ac0, 0x3ffb7ed6, 0x159fadc8
.word 0xbc899bcf, 0xf04d134b, 0x00000000, 0x00000000
.word 0x3ff5ac05, 0x6b015ac0, 0x40037135, 0x779c8dcb
.word 0xbc8fe126, 0xce9778ae, 0x00000000, 0x00000000
.word 0x3ff58ed2, 0x308158ed, 0x3ffb916e, 0xd68964ec
.word 0x3c826a5d, 0x5dbaae29, 0x00000000, 0x00000000
.word 0x3ff58ed2, 0x308158ed, 0x40037e5b, 0xd40f95a1
.word 0x3cac6ff5, 0xeca5d122, 0x00000000, 0x00000000
.word 0x3ff571ed, 0x3c506b3a, 0x3ffba3fb, 0x14672d7c
.word 0xbc8117d3, 0x97dcefc9, 0x00000000, 0x00000000
.word 0x3ff571ed, 0x3c506b3a, 0x40038b79, 0x579d3eab
.word 0xbcac254f, 0xc0db598e, 0x00000000, 0x00000000
.word 0x3ff55555, 0x55555555, 0x3ffbb67a, 0xe8584caa
.word 0x3c9cec95, 0xd0b5c1e3, 0x00000000, 0x00000000
.word 0x3ff55555, 0x55555555, 0x4003988e, 0x1409212e
.word 0x3caf40c8, 0x6450c869, 0x00000000, 0x00000000
.word 0x3ff53909, 0x48f40feb, 0x3ffbc8ee, 0x6b2865b9
.word 0x3c9394eb, 0x90f645c8, 0x00000000, 0x00000000
.word 0x3ff53909, 0x48f40feb, 0x4003a59a, 0x1adbb257
.word 0x3ca6adce, 0x020a308d, 0x00000000, 0x00000000
.word 0x3ff51d07, 0xeae2f815, 0x3ffbdb55, 0xb550fdbc
.word 0x3c7365e9, 0x6aa5fae3, 0x00000000, 0x00000000
.word 0x3ff51d07, 0xeae2f815, 0x4003b29d, 0x7d635662
.word 0x3cac99b0, 0x5e282129, 0x00000000, 0x00000000
.word 0x3ff50150, 0x15015015, 0x3ffbedb0, 0xdefaf661
.word 0x3c91a627, 0xb279170d, 0x00000000, 0x00000000
.word 0x3ff50150, 0x15015015, 0x4003bf98, 0x4cb56c77
.word 0x3ca8f653, 0xbcc0c4a1, 0x00000000, 0x00000000
.word 0x3ff4e5e0, 0xa72f0539, 0x3ffc0000, 0x00000000
.word 0x00000000, 0x00000000, 0x00000000, 0x00000000
.word 0x3ff4e5e0, 0xa72f0539, 0x4003cc8a, 0x99af5453
.word 0xbc486364, 0x4f05f2be, 0x00000000, 0x00000000
.word 0x3ff4cab8, 0x8725af6e, 0x3ffc1243, 0x2fec0329
.word 0x3c96e0d7, 0x8dd23a7d, 0x00000000, 0x00000000
.word 0x3ff4cab8, 0x8725af6e, 0x4003d974, 0x74f76df2
.word 0x3c82e3c9, 0xfdbbbdc2, 0x00000000, 0x00000000
.word 0x3ff4afd6, 0xa052bf5b, 0x3ffc247a, 0x85fe81fa
.word 0x3c89d8ee, 0xf6854220, 0x00000000, 0x00000000
.word 0x3ff4afd6, 0xa052bf5b, 0x4003e655, 0xeefe1367
.word 0x3c80eb35, 0xbb532559, 0x00000000, 0x00000000
.word 0x3ff49539, 0xe3b2d067, 0x3ffc36a6, 0x192bf168
.word 0xbc9083d8, 0x1a423b11, 0x00000000, 0x00000000
.word 0x3ff49539, 0xe3b2d067, 0x4003f32f, 0x17fe8d04
.word 0xbc905d6c, 0x1c437de0, 0x00000000, 0x00000000
.word 0x3ff47ae1, 0x47ae147b, 0x3ffc48c6, 0x001f0ac0
.word 0xbc92d481, 0x189efd6b, 0x00000000, 0x00000000
.word 0x3ff47ae1, 0x47ae147b, 0x40040000, 0x00000000
.word 0x00000000, 0x00000000, 0x00000000, 0x00000000
.word 0x3ff460cb, 0xc7f5cf9a, 0x3ffc5ada, 0x513a1593
.word 0xbc7aaedd, 0x014f5f03, 0x00000000, 0x00000000
.word 0x3ff460cb, 0xc7f5cf9a, 0x40040cc8, 0xb6d657c2
.word 0xbc9c05ab, 0xf480ce19, 0x00000000, 0x00000000
.word 0x3ff446f8, 0x6562d9fb, 0x3ffc6ce3, 0x22982a3f
.word 0x3c891b2d, 0xf3e15f29, 0x00000000, 0x00000000
.word 0x3ff446f8, 0x6562d9fb, 0x40041989, 0x4c2329f0
.word 0x3c976037, 0x46da0ea6, 0x00000000, 0x00000000
.word 0x3ff42d66, 0x25d51f87, 0x3ffc7ee0, 0x8a0e6d4c
.word 0x3c991c54, 0xc53e75c8, 0x00000000, 0x00000000
.word 0x3ff42d66, 0x25d51f87, 0x40042641, 0xcf569572
.word 0xbcadf80b, 0x1442c029, 0x00000000, 0x00000000
.word 0x3ff41414, 0x14141414, 0x3ffc90d2, 0x9d2d43ce
.word 0xbc9edadb, 0x07f1137a, 0x00000000, 0x00000000
.word 0x3ff41414, 0x14141414, 0x400432f2, 0x4fb01c7a
.word 0x3ca38bfe, 0x0e012c1c, 0x00000000, 0x00000000
.word 0x3ff3fb01, 0x3fb013fb, 0x3ffca2b9, 0x714180f7
.word 0xbc81a63d, 0x6750c57c, 0x00000000, 0x00000000
.word 0x3ff3fb01, 0x3fb013fb, 0x40043f9a, 0xdc3f79ce
.word 0x3c66d2b1, 0x767ae30a, 0x00000000, 0x00000000
.word 0x3ff3e22c, 0xbce4a902, 0x3ffcb495, 0x1b558d17
.word 0x3c8fcbcb, 0x357f2308, 0x00000000, 0x00000000
.word 0x3ff3e22c, 0xbce4a902, 0x40044c3b, 0x83e57153
.word 0x3c98c853, 0xc6be5ee1, 0x00000000, 0x00000000
.word 0x3ff3c995, 0xa47babe7, 0x3ffcc665, 0xb0328622
.word 0xbc91baa4, 0xd369f814, 0x00000000, 0x00000000
.word 0x3ff3c995, 0xa47babe7, 0x400458d4, 0x55549c1a
.word 0x3ca02d72, 0x8d9a6054, 0x00000000, 0x00000000
.word 0x3ff3b13b, 0x13b13b14, 0x3ffcd82b, 0x446159f3
.word 0x3c983fb7, 0xb33cdfe8, 0x00000000, 0x00000000
.word 0x3ff3b13b, 0x13b13b14, 0x40046565, 0x5f122ff6
.word 0x3ca862c5, 0xd2f0ca4c, 0x00000000, 0x00000000
.word 0x3ff3991c, 0x2c187f63, 0x3ffce9e5, 0xec2bda80
.word 0xbc94ccf3, 0xd8e249ab, 0x00000000, 0x00000000
.word 0x3ff3991c, 0x2c187f63, 0x400471ee, 0xaf76c2c6
.word 0x3c975c62, 0xeff26e8e, 0x00000000, 0x00000000
.word 0x3ff38138, 0x13813814, 0x3ffcfb95, 0xbb9dcc0c
.word 0x3c92cea2, 0x0857ae03, 0x00000000, 0x00000000
.word 0x3ff38138, 0x13813814, 0x40047e70, 0x54af0989
.word 0x3c9d8c33, 0xc0054830, 0x00000000, 0x00000000
.word 0x3ff3698d, 0xf3de0748, 0x3ffd0d3a, 0xc685eda4
.word 0x3c94115a, 0x0ff4cf9e, 0x00000000, 0x00000000
.word 0x3ff3698d, 0xf3de0748, 0x40048aea, 0x5cbc935f
.word 0xbca8cb00, 0x12d14ff5, 0x00000000, 0x00000000
.word 0x3ff3521c, 0xfb2b78c1, 0x3ffd1ed5, 0x2076fbe9
.word 0x3c8f48a8, 0x6b72875f, 0x00000000, 0x00000000
.word 0x3ff3521c, 0xfb2b78c1, 0x4004975c, 0xd5768088
.word 0xbca1731e, 0xbc02f748, 0x00000000, 0x00000000
.word 0x3ff33ae4, 0x5b57bcb2, 0x3ffd3064, 0xdcc8ae67
.word 0x3c93480e, 0x805158ba, 0x00000000, 0x00000000
.word 0x3ff33ae4, 0x5b57bcb2, 0x4004a3c7, 0xcc8a358a
.word 0xbc9d8f7f, 0xd2726ffa, 0x00000000, 0x00000000
.word 0x3ff323e3, 0x4a2b10bf, 0x3ffd41ea, 0x0e98af91
.word 0x3c824640, 0x0309962f, 0x00000000, 0x00000000
.word 0x3ff323e3, 0x4a2b10bf, 0x4004b02b, 0x4f7c0a88
.word 0xbcaf71e1, 0xf6cafde2, 0x00000000, 0x00000000
.word 0x3ff30d19, 0x0130d190, 0x3ffd5364, 0xc8cb8f86
.word 0x3c8ad003, 0xc00630e1, 0x00000000, 0x00000000
.word 0x3ff30d19, 0x0130d190, 0x4004bc87, 0x6ba7f6ec
.word 0x3c9c1edb, 0x2be943b8, 0x00000000, 0x00000000
.word 0x3ff2f684, 0xbda12f68, 0x3ffd64d5, 0x1e0db1c6
.word 0xbc911ed3, 0x6986d362, 0x00000000, 0x00000000
.word 0x3ff2f684, 0xbda12f68, 0x4004c8dc, 0x2e423980
.word 0xbc949d1f, 0x46ef5d2c, 0x00000000, 0x00000000
.word 0x3ff2e025, 0xc04b8097, 0x3ffd763b, 0x20d435ef
.word 0x3c9d6780, 0xf76cb258, 0x00000000, 0x00000000
.word 0x3ff2e025, 0xc04b8097, 0x4004d529, 0xa457fcfc
.word 0xbca1404a, 0x46484e3d, 0x00000000, 0x00000000
.word 0x3ff2c9fb, 0x4d812ca0, 0x3ffd8796, 0xe35ddbb2
.word 0x3c83fdd9, 0x1aeb637a, 0x00000000, 0x00000000
.word 0x3ff2c9fb, 0x4d812ca0, 0x4004e16f, 0xdacff937
.word 0xbca1deb9, 0xd3815ad2, 0x00000000, 0x00000000
.word 0x3ff2b404, 0xad012b40, 0x3ffd98e8, 0x77b3e207
.word 0xbc48c301, 0xee02dee8, 0x00000000, 0x00000000
.word 0x3ff2b404, 0xad012b40, 0x4004edae, 0xde6b10fe
.word 0x3ca99709, 0x4a91a780, 0x00000000, 0x00000000
.word 0x3ff29e41, 0x29e4129e, 0x3ffdaa2f, 0xefaae1d8
.word 0xbc63fe0e, 0x03f44594, 0x00000000, 0x00000000
.word 0x3ff29e41, 0x29e4129e, 0x4004f9e6, 0xbbc4ecb3
.word 0x3c6ce5a6, 0x018493f1, 0x00000000, 0x00000000
.word 0x3ff288b0, 0x1288b013, 0x3ffdbb6d, 0x5ce3a42f
.word 0xbc922c27, 0xf71c8337, 0x00000000, 0x00000000
.word 0x3ff288b0, 0x1288b013, 0x40050617, 0x7f5491bb
.word 0xbc9e591e, 0x7b2a6d1a, 0x00000000, 0x00000000
.word 0x3ff27350, 0xb8812735, 0x3ffdcca0, 0xd0cbf408
.word 0x3c7a6d16, 0x2310db57, 0x00000000, 0x00000000
.word 0x3ff27350, 0xb8812735, 0x40051241, 0x356cf6e0
.word 0x3ca37dc2, 0x60e8bc2d, 0x00000000, 0x00000000
.word 0x3ff25e22, 0x708092f1, 0x3ffdddca, 0x5c9f6be8
.word 0x3c818520, 0xf0a3f809, 0x00000000, 0x00000000
.word 0x3ff25e22, 0x708092f1, 0x40051e63, 0xea3d95b0
.word 0x3caecf78, 0x2e88d5ce, 0x00000000, 0x00000000
.word 0x3ff24924, 0x92492492, 0x3ffdeeea, 0x11683f49
.word 0x3c802aae, 0x4bfa7c27, 0x00000000, 0x00000000
.word 0x3ff24924, 0x92492492, 0x40052a7f, 0xa9d2f8ea
.word 0xbca21c62, 0xb033c079, 0x00000000, 0x00000000
.word 0x3ff23456, 0x789abcdf, 0x3ffe0000, 0x00000000
.word 0x00000000, 0x00000000, 0x00000000, 0x00000000
.word 0x3ff23456, 0x789abcdf, 0x40053694, 0x80174810
.word 0xbc9c3ec1, 0xa4ee7c21, 0x00000000, 0x00000000
.word 0x3ff21fb7, 0x8121fb78, 0x3ffe110c, 0x39105faf
.word 0x3c776161, 0x4c513964, 0x00000000, 0x00000000
.word 0x3ff21fb7, 0x8121fb78, 0x400542a2, 0x78d2d036
.word 0xbca495c2, 0x45254df4, 0x00000000, 0x00000000
.word 0x3ff20b47, 0x0c67c0d9, 0x3ffe220e, 0xcd13ed60
.word 0xbc729f01, 0xf18c9dc9, 0x00000000, 0x00000000
.word 0x3ff20b47, 0x0c67c0d9, 0x40054ea9, 0x9fac8a0f
.word 0x3c80cfbb, 0x19353b3d, 0x00000000, 0x00000000
.word 0x3ff1f704, 0x7dc11f70, 0x3ffe3307, 0xcc56cf5c
.word 0xbc81f04e, 0xc3189131, 0x00000000, 0x00000000
.word 0x3ff1f704, 0x7dc11f70, 0x40055aaa, 0x002a9d5a
.word 0xbc4bf504, 0x76241f94, 0x00000000, 0x00000000
.word 0x3ff1e2ef, 0x3b3fb874, 0x3ffe43f7, 0x46f7795b
.word 0xbc931e7f, 0x8af68f8c, 0x00000000, 0x00000000
.word 0x3ff1e2ef, 0x3b3fb874, 0x400566a3, 0xa5b2e1b1
.word 0x3caa1fd2, 0x8cc92e33, 0x00000000, 0x00000000
.word 0x3ff1cf06, 0xada2811d, 0x3ffe54dd, 0x4ce75f1e
.word 0xbc811b19, 0x5dfc62e5, 0x00000000, 0x00000000
.word 0x3ff1cf06, 0xada2811d, 0x40057296, 0x9b8b5cd8
.word 0x3ca30cbf, 0x1c53312e, 0x00000000, 0x00000000
.word 0x3ff1bb4a, 0x4046ed29, 0x3ffe65b9, 0xedeba38e
.word 0xbc7bb732, 0x51e8c364, 0x00000000, 0x00000000
.word 0x3ff1bb4a, 0x4046ed29, 0x40057e82, 0xecdabe8d
.word 0xbc7c2aed, 0xf3c4c4bd, 0x00000000, 0x00000000
.word 0x3ff1a7b9, 0x611a7b96, 0x3ffe768d, 0x399dc470
.word 0xbc9a8c81, 0x3405c01c, 0x00000000, 0x00000000
.word 0x3ff1a7b9, 0x611a7b96, 0x40058a68, 0xa4a8d9f3
.word 0x3ca50798, 0xe67012d9, 0x00000000, 0x00000000
.word 0x3ff19453, 0x808ca29c, 0x3ffe8757, 0x3f6c42c5
.word 0x3c9dbf9c, 0xf7bbcda3, 0x00000000, 0x00000000
.word 0x3ff19453, 0x808ca29c, 0x40059647, 0xcddf1ca5
.word 0x3ca14a95, 0xf35dea0b, 0x00000000, 0x00000000
.word 0x3ff18118, 0x11811812, 0x3ffe9818, 0x0e9b47f2
.word 0xbc9b6bd7, 0x4396d08e, 0x00000000, 0x00000000
.word 0x3ff18118, 0x11811812, 0x4005a220, 0x73490377
.word 0xbcadd036, 0x39925812, 0x00000000, 0x00000000
.word 0x3ff16e06, 0x89427379, 0x3ffea8cf, 0xb64547ab
.word 0x3c8721b2, 0x6374e19f, 0x00000000, 0x00000000
.word 0x3ff16e06, 0x89427379, 0x4005adf2, 0x9f948cfb
.word 0xbca42520, 0xf7716fa6, 0x00000000, 0x00000000
.word 0x3ff15b1e, 0x5f75270d, 0x3ffeb97e, 0x455b9edb
.word 0x3c999b45, 0x40857883, 0x00000000, 0x00000000
.word 0x3ff15b1e, 0x5f75270d, 0x4005b9be, 0x5d52a9da
.word 0x3c9098cd, 0x1b3af777, 0x00000000, 0x00000000
.word 0x3ff1485f, 0x0e0acd3b, 0x3ffeca23, 0xcaa72f73
.word 0x3c7e3ed5, 0x29679959, 0x00000000, 0x00000000
.word 0x3ff1485f, 0x0e0acd3b, 0x4005c583, 0xb6f7ab03
.word 0x3ca963bc, 0x9d795b51, 0x00000000, 0x00000000
.word 0x3ff135c8, 0x1135c811, 0x3ffedac0, 0x54c8f94c
.word 0x3c90b5c1, 0x15a56207, 0x00000000, 0x00000000
.word 0x3ff135c8, 0x1135c811, 0x4005d142, 0xb6dbadc5
.word 0x3ca6f1f5, 0x5323d116, 0x00000000, 0x00000000
.word 0x3ff12358, 0xe75d3033, 0x3ffeeb53, 0xf23ab028
.word 0xbc8617e4, 0xb5384f5d, 0x00000000, 0x00000000
.word 0x3ff12358, 0xe75d3033, 0x4005dcfb, 0x673b05df
.word 0xbca099df, 0xc321634f, 0x00000000, 0x00000000
.word 0x3ff11111, 0x11111111, 0x3ffefbde, 0xb14f4eda
.word 0xbc93a145, 0xfe1be078, 0x00000000, 0x00000000
.word 0x3ff11111, 0x11111111, 0x4005e8ad, 0xd236a58f
.word 0xbc7ef8c7, 0xc0d1fec6, 0x00000000, 0x00000000
.word 0x3ff0fef0, 0x10fef011, 0x3fff0c60, 0xa033a7b3
.word 0xbc91b0fc, 0x15cd89c6, 0x00000000, 0x00000000
.word 0x3ff0fef0, 0x10fef011, 0x4005f45a, 0x01d483b4
.word 0xbc94a237, 0xdc0fa105, 0x00000000, 0x00000000
.word 0x3ff0ecf5, 0x6be69c90, 0x3fff1cd9, 0xcceef239
.word 0x3c91afd8, 0x64eab60a, 0x00000000, 0x00000000
.word 0x3ff0ecf5, 0x6be69c90, 0x40060000, 0x00000000
.word 0x00000000, 0x00000000, 0x00000000, 0x00000000
.word 0x3ff0db20, 0xa88f4696, 0x3fff2d4a, 0x45635640
.word 0xbc8eebae, 0xea670bc2, 0x00000000, 0x00000000
.word 0x3ff0db20, 0xa88f4696, 0x40060b9f, 0xd68a4554
.word 0x3ca328e1, 0x70dae176, 0x00000000, 0x00000000
.word 0x3ff0c971, 0x4fbcda3b, 0x3fff3db2, 0x174e7468
.word 0x3c9e1513, 0x2d6ac52a, 0x00000000, 0x00000000
.word 0x3ff0c971, 0x4fbcda3b, 0x40061739, 0x8f2aaa48
.word 0xbc9b672b, 0xba260735, 0x00000000, 0x00000000
.word 0x3ff0b7e6, 0xec259dc8, 0x3fff4e11, 0x5049ec26
.word 0xbc9b6656, 0xb6bd5d76, 0x00000000, 0x00000000
.word 0x3ff0b7e6, 0xec259dc8, 0x400622cd, 0x337f0fe8
.word 0x3c9fe207, 0x3279559f, 0x00000000, 0x00000000
.word 0x3ff0a681, 0x0a6810a7, 0x3fff5e67, 0xfdcbdf44
.word 0xbc98af06, 0x1849d6fc, 0x00000000, 0x00000000
.word 0x3ff0a681, 0x0a6810a7, 0x40062e5a, 0xcd0c3ebe
.word 0xbca2c50e, 0x2092203a, 0x00000000, 0x00000000
.word 0x3ff0953f, 0x39010954, 0x3fff6eb6, 0x2d27730d
.word 0xbc9401d9, 0x5ca1ce34, 0x00000000, 0x00000000
.word 0x3ff0953f, 0x39010954, 0x400639e2, 0x653e421b
.word 0xbc9f75e0, 0x5835e4b9, 0x00000000, 0x00000000
.word 0x3ff08421, 0x08421084, 0x3fff7efb, 0xeb8d4f12
.word 0xbc7e84e8, 0xa6ff3256, 0x00000000, 0x00000000
.word 0x3ff08421, 0x08421084, 0x40064564, 0x0568c1c3
.word 0x3cad1778, 0x7e4c8970, 0x00000000, 0x00000000
.word 0x3ff07326, 0x0a47f7c6, 0x3fff8f39, 0x460c19a8
.word 0x3c989b4e, 0x16ee9aaf, 0x00000000, 0x00000000
.word 0x3ff07326, 0x0a47f7c6, 0x400650df, 0xb6c759f4
.word 0x3c99063c, 0x91db4c77, 0x00000000, 0x00000000
.word 0x3ff0624d, 0xd2f1a9fc, 0x3fff9f6e, 0x4990f227
.word 0x3c8b42e5, 0xb5d1e808, 0x00000000, 0x00000000
.word 0x3ff0624d, 0xd2f1a9fc, 0x40065c55, 0x827df1d2
.word 0xbca3923d, 0xf03e1e2f, 0x00000000, 0x00000000
.word 0x3ff05197, 0xf7d73404, 0x3fffaf9b, 0x02e7e8f2
.word 0x3c897a76, 0x8f34e1c2, 0x00000000, 0x00000000
.word 0x3ff05197, 0xf7d73404, 0x400667c5, 0x7199104b
.word 0x3c875b89, 0x6f332e70, 0x00000000, 0x00000000
.word 0x3ff04104, 0x10410410, 0x3fffbfbf, 0x7ebc755f
.word 0xbc9b2a94, 0x084da0b6, 0x00000000, 0x00000000
.word 0x3ff04104, 0x10410410, 0x4006732f, 0x8d0e2f77
.word 0xbc93dffd, 0x470422e3, 0x00000000, 0x00000000
.word 0x3ff03091, 0xb51f5e1a, 0x3fffcfdb, 0xc999e97d
.word 0x3c82be17, 0xecdd3bbc, 0x00000000, 0x00000000
.word 0x3ff03091, 0xb51f5e1a, 0x40067e93, 0xddbc0e73
.word 0xbc86eb9f, 0x32ac1a5c, 0x00000000, 0x00000000
.word 0x3ff02040, 0x81020408, 0x3fffdfef, 0xefebe3d6
.word 0xbc909afc, 0xfc7c1f3b, 0x00000000, 0x00000000
.word 0x3ff02040, 0x81020408, 0x400689f2, 0x6c6b01d0
.word 0x3cae816f, 0x9d2a1032, 0x00000000, 0x00000000
.word 0x3ff01010, 0x10101010, 0x3fffeffb, 0xfdfebf1f
.word 0x3c95dee5, 0x1994f18b, 0x00000000, 0x00000000
.word 0x3ff01010, 0x10101010, 0x4006954b, 0x41cd4293
.word 0x3ca3d5bc, 0xcc443076, 0x00000000, 0x00000000
.word 0x3ff00000, 0x00000000, 0x40000000, 0x00000000
.word 0x00000000, 0x00000000, 0x00000000, 0x00000000
.word 0x3ff00000, 0x00000000, 0x4006a09e, 0x667f3bcd
.word 0xbcabdd34, 0x13b26456, 0x00000000, 0x00000000
#define A5 %f32
#define A4 %f30
#define A3 %f28
#define A2 %f26
#define A1 %f56
#define DC0 %f8
#define DC2 %f6
#define DC3 %f4
#define counter %l3
#define TBL %l5
#define stridex %l6
#define stridey %l7
#define _0x00001ff8 %i0
#define _0x7ff00000 %o0
#define _0x00100000 %o2
#define tmp_counter STACK_BIAS-0x40
#define tmp_px STACK_BIAS-0x38
#define tmp0 STACK_BIAS-0x30
#define tmp1 STACK_BIAS-0x28
#define tmp2 STACK_BIAS-0x20
#define tmp3 STACK_BIAS-0x18
#define tmp4 STACK_BIAS-0x10
#define tmp5 STACK_BIAS-0x08
! sizeof temp storage - must be a multiple of 16 for V9
#define tmps 0x40
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
! !!!!! algorithm !!!!!
! ((float*)&res)[0] = ((float*)px)[0];
! ((float*)&res)[1] = ((float*)px)[1];
! hx = *(int*)px;
! px += stridex;
!
! if ( hx >= 0x7ff00000 )
! {
! res = sqrt(res);
! ((float*)py)[0] = ((float*)&res)[0];
! ((float*)py)[1] = ((float*)&res)[1];
! py += stridey;
! goto next;
! }
! if ( hx < 0x00100000 )
! {
! res = sqrt(res);
! ((float*)py)[0] = ((float*)&res)[0];
! ((float*)py)[1] = ((float*)&res)[1];
! py += stridey;
! goto next;
! }
!
! sqrt_exp = hx >> 21;
! sqrt_exp -= 512;
! sqrt_exp <<= 52;
! dsqrt_exp = *(double*)&sqrt_exp;
! bit = hx >> 15;
! bit &= 32;
! ind0 = hx >> 7;
! ind0 &= 0x1ff8;
! ind0 += 32;
! ind0 &= -64;
! ind1 = ind0;
! ind1 += bit;
!
! res = vis_fand(res,DC0);
! res = vis_for(res,A1);
! res_c = vis_fpadd32(res,DC2);
! res_c = vis_fand(res_c,DC3);
!
! pind = (char*)TBL + ind1;
! dexp_hi = ((double*)pind)[1];
! dexp_lo = ((double*)pind)[2];
!
! dtmp0 = ((double*)pind)[0];
! xx = (res - res_c);
! xx *= dtmp0;
!
! res = A5 * xx;
! res += A4;
! res *= xx;
! res += A3;
! res *= xx;
! res += A2;
! res *= xx;
! res += A1;
! res *= xx;
!
! res = dexp_hi * res;
! res += dexp_lo;
! res += dexp_hi;
!
! dtmp0 = vis_fpadd32(dsqrt_exp,res);
! ((float*)py)[0] = ((float*)&dtmp0)[0];
! ((float*)py)[1] = ((float*)&dtmp0)[1];
! py += stridey;
!
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
ENTRY(__vsqrt)
save %sp,-SA(MINFRAME)-tmps,%sp
PIC_SETUP(l7)
PIC_SET(l7,.CONST_TBL,l5)
wr %g0,0x82,%asi
ldd [TBL],A1
sll %i2,3,stridex
or %g0,%i3,%o4
ldd [TBL+8],A2
sll %i4,3,stridey
or %g0,0x7ff,%o0
ldd [TBL+16],A3
sll %o0,20,_0x7ff00000
or %g0,0x001,%o2
ldd [TBL+24],A4
sll %o2,20,_0x00100000
ldd [TBL+32],A5
ldd [TBL+40],DC0
ldd [TBL+48],DC2
ldd [TBL+56],DC3
add TBL,64,TBL
add %g0,1023,%o5
st %i0,[%fp+tmp_counter]
sll %o5,3,_0x00001ff8
stx %i1,[%fp+tmp_px]
.begin:
ld [%fp+tmp_counter],counter
ldx [%fp+tmp_px],%l2
st %g0,[%fp+tmp_counter]
.begin1:
cmp counter,0
ble,pn %icc,.exit
lda [%l2]%asi,%o5 ! (5_1) hx = *(int*)px;
lda [%l2]%asi,%f10 ! (5_0) ((float*)&res)[0] = ((float*)px)[0];
lda [%l2+4]%asi,%f11 ! (5_0) ((float*)&res)[1] = ((float*)px)[1];
cmp %o5,_0x7ff00000 ! (5_1) hx ? 0x7ff00000
bge,pn %icc,.spec ! (5_1) if ( hx >= 0x7ff00000 )
nop
cmp %o5,_0x00100000 ! (5_1) hx ? 0x00100000
bl,pn %icc,.spec ! (5_1) if ( hx < 0x00100000 )
nop
add %l2,stridex,%l2 ! px += stridex
fand %f10,DC0,%f50 ! (5_1) res = vis_fand(res,DC0);
for %f50,A1,%f40 ! (5_1) res = vis_for(res,A1);
sra %o5,21,%l1 ! (5_1) sqrt_exp = hx >> 21;
sra %o5,15,%i1 ! (5_1) bit = hx >> 15;
sra %o5,7,%o1 ! (5_1) ind0 = hx >> 7;
sub %l1,512,%o3 ! (5_1) sqrt_exp -= 512;
and %o1,_0x00001ff8,%o1 ! (5_1) ind0 &= 0x1ff8;
lda [%l2]%asi,%f10 ! (0_0) ((float*)&res)[0] = ((float*)px)[0];
add %o1,32,%o1 ! (5_1) ind0 += 32;
lda [%l2+4]%asi,%f11 ! (0_0) ((float*)&res)[1] = ((float*)px)[1];
and %i1,32,%i4 ! (5_1) bit &= 32;
and %o1,-64,%o1 ! (5_1) ind0 &= -8;
sll %o1,0,%o7 ! (5_1) ind1 = ind0;
sllx %o3,52,%o3 ! (5_1) sqrt_exp <<= 52;
add %o7,%i4,%l0 ! (5_1) ind1 += bit;
lda [%l2]%asi,%o5 ! (0_0) hx = *(int*)px;
stx %o3,[%fp+tmp0] ! (5_1) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (0_0) res = vis_fand(res,DC0);
add %l2,stridex,%l2 ! px += stridex
fpadd32 %f40,DC2,%f54 ! (5_1) res_c = vis_fpadd32(res,DC2);
add %l0,TBL,%o1 ! (5_1) pind = (char*)TBL + ind1
cmp %o5,_0x7ff00000 ! (0_0) hx ? 0x7ff00000
bge,pn %icc,.update0 ! (0_0) if ( hx >= 0x7ff00000 )
for %f50,A1,%f42 ! (0_0) res = vis_for(res,A1);
.cont0:
sra %o5,21,%l1 ! (0_0) sqrt_exp = hx >> 21;
sra %o5,15,%i2 ! (0_0) bit = hx >> 15;
ldd [%o1],%f50 ! (5_1) dtmp0 = ((double*)pind)[0];
sra %o5,7,%o1 ! (0_0) ind0 = hx >> 7;
sub %l1,512,%o3 ! (0_0) sqrt_exp -= 512;
fand %f54,DC3,%f54 ! (5_1) res_c = vis_fand(res_c,DC3);
and %o1,_0x00001ff8,%o1 ! (0_0) ind0 &= 0x1ff8;
lda [%l2]%asi,%f10 ! (1_0) ((float*)&res)[0] = ((float*)px)[0];
add %o1,32,%o1 ! (0_0) ind0 += 32;
lda [%l2+4]%asi,%f11 ! (1_0) ((float*)&res)[1] = ((float*)px)[1];
and %i2,32,%i4 ! (0_0) bit &= 32;
and %o1,-64,%o1 ! (0_0) ind0 &= -8;
fsubd %f40,%f54,%f40 ! (5_1) xx = (res - res_c);
sll %o1,0,%o7 ! (0_0) ind1 = ind0;
cmp %o5,_0x00100000 ! (0_0) hx ? 0x00100000
bl,pn %icc,.update1 ! (0_0) if ( hx < 0x00100000 )
lda [%l2]%asi,%o5 ! (1_0) hx = *(int*)px;
.cont1:
sllx %o3,52,%o3 ! (0_0) sqrt_exp <<= 52;
add %o7,%i4,%i1 ! (0_0) ind1 += bit;
fmuld %f40,%f50,%f40 ! (5_1) xx *= dtmp0;
stx %o3,[%fp+tmp1] ! (0_0) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (1_0) res = vis_fand(res,DC0);
add %l2,stridex,%l2 ! px += stridex
fpadd32 %f42,DC2,%f54 ! (0_0) res_c = vis_fpadd32(res,DC2);
add %i1,TBL,%o1 ! (0_0) pind = (char*)TBL + ind1
cmp %o5,_0x7ff00000 ! (1_0) hx ? 0x7ff00000
bge,pn %icc,.update2 ! (1_0) if ( hx >= 0x7ff00000 )
for %f50,A1,%f14 ! (1_0) res = vis_for(res,A1);
.cont2:
sra %o5,21,%l1 ! (1_0) sqrt_exp = hx >> 21;
sra %o5,15,%g5 ! (1_0) bit = hx >> 15;
ldd [%o1],%f50 ! (0_0) dtmp0 = ((double*)pind)[0];
fmuld A5,%f40,%f52 ! (5_1) res = A5 * xx;
sra %o5,7,%o1 ! (1_0) ind0 = hx >> 7;
sub %l1,512,%o3 ! (1_0) sqrt_exp -= 512;
fand %f54,DC3,%f54 ! (0_0) res_c = vis_fand(res_c,DC3);
and %o1,_0x00001ff8,%o1 ! (1_0) ind0 &= 0x1ff8;
lda [%l2]%asi,%f10 ! (2_0) ((float*)&res)[0] = ((float*)px)[0];
add %o1,32,%o1 ! (1_0) ind0 += 32;
lda [%l2+4]%asi,%f11 ! (2_0) ((float*)&res)[1] = ((float*)px)[1];
and %g5,32,%i4 ! (1_0) bit &= 32;
and %o1,-64,%o1 ! (1_0) ind0 &= -8;
fsubd %f42,%f54,%f42 ! (0_0) xx = (res - res_c);
sll %o1,0,%o7 ! (1_0) ind1 = ind0;
faddd %f52,A4,%f54 ! (5_1) res += A4;
cmp %o5,_0x00100000 ! (1_0) hx ? 0x00100000
bl,pn %icc,.update3 ! (1_0) if ( hx < 0x00100000 )
lda [%l2]%asi,%o5 ! (2_0) hx = *(int*)px;
.cont3:
sllx %o3,52,%o3 ! (1_0) sqrt_exp <<= 52;
add %o7,%i4,%i2 ! (1_0) ind1 += bit;
fmuld %f42,%f50,%f42 ! (0_0) xx *= dtmp0;
stx %o3,[%fp+tmp2] ! (1_0) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (2_0) res = vis_fand(res,DC0);
fmuld %f54,%f40,%f34 ! (5_1) res *= xx;
fpadd32 %f14,DC2,%f54 ! (1_0) res_c = vis_fpadd32(res,DC2);
add %l2,stridex,%l2 ! px += stridex
add %i2,TBL,%o1 ! (1_0) pind = (char*)TBL + ind1
cmp %o5,_0x7ff00000 ! (2_0) hx ? 0x7ff00000
bge,pn %icc,.update4 ! (2_0) if ( hx >= 0x7ff00000 )
for %f50,A1,%f18 ! (2_0) res = vis_for(res,A1);
.cont4:
sra %o5,21,%l1 ! (2_0) sqrt_exp = hx >> 21;
sra %o5,15,%g1 ! (2_0) bit = hx >> 15;
ldd [%o1],%f50 ! (1_0) dtmp0 = ((double*)pind)[0];
fmuld A5,%f42,%f52 ! (0_0) res = A5 * xx;
sra %o5,7,%o1 ! (2_0) ind0 = hx >> 7;
sub %l1,512,%o3 ! (2_0) sqrt_exp -= 512;
fand %f54,DC3,%f54 ! (1_0) res_c = vis_fand(res_c,DC3);
and %o1,_0x00001ff8,%o1 ! (2_0) ind0 &= 0x1ff8;
lda [%l2]%asi,%f10 ! (3_0) ((float*)&res)[0] = ((float*)px)[0];
faddd %f34,A3,%f62 ! (5_1) res += A3;
add %o1,32,%o1 ! (2_0) ind0 += 32;
lda [%l2+4]%asi,%f11 ! (3_0) ((float*)&res)[1] = ((float*)px)[1];
and %g1,32,%i4 ! (2_0) bit &= 32;
and %o1,-64,%o1 ! (2_0) ind0 &= -8;
fsubd %f14,%f54,%f14 ! (1_0) xx = (res - res_c);
sll %o1,0,%o7 ! (2_0) ind1 = ind0;
faddd %f52,A4,%f54 ! (0_0) res += A4;
fmuld %f62,%f40,%f52 ! (5_1) res *= xx;
cmp %o5,_0x00100000 ! (2_0) hx ? 0x00100000
bl,pn %icc,.update5 ! (2_0) if ( hx < 0x00100000 )
lda [%l2]%asi,%o5 ! (3_0) hx = *(int*)px;
.cont5:
sllx %o3,52,%o3 ! (2_0) sqrt_exp <<= 52;
add %o7,%i4,%g5 ! (2_0) ind1 += bit;
fmuld %f14,%f50,%f14 ! (1_0) xx *= dtmp0;
stx %o3,[%fp+tmp3] ! (2_0) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (3_0) res = vis_fand(res,DC0);
fmuld %f54,%f42,%f34 ! (0_0) res *= xx;
fpadd32 %f18,DC2,%f54 ! (2_0) res_c = vis_fpadd32(res,DC2);
add %l2,stridex,%l2 ! px += stridex
add %g5,TBL,%o1 ! (2_0) pind = (char*)TBL + ind1
faddd %f52,A2,%f20 ! (5_1) res += A2;
cmp %o5,_0x7ff00000 ! (3_0) hx ? 0x7ff00000
bge,pn %icc,.update6 ! (3_0) if ( hx >= 0x7ff00000 )
for %f50,A1,%f44 ! (3_0) res = vis_for(res,A1);
.cont6:
sra %o5,21,%l1 ! (3_0) sqrt_exp = hx >> 21;
sra %o5,15,%i3 ! (3_0) bit = hx >> 15;
ldd [%o1],%f50 ! (2_0) dtmp0 = ((double*)pind)[0];
fmuld A5,%f14,%f52 ! (1_0) res = A5 * xx;
sra %o5,7,%o1 ! (3_0) ind0 = hx >> 7;
sub %l1,512,%o3 ! (3_0) sqrt_exp -= 512;
fand %f54,DC3,%f54 ! (2_0) res_c = vis_fand(res_c,DC3);
fmuld %f20,%f40,%f20 ! (5_1) res *= xx;
and %o1,_0x00001ff8,%o1 ! (3_0) ind0 &= 0x1ff8;
lda [%l2]%asi,%f10 ! (4_0) ((float*)&res)[0] = ((float*)px)[0];
faddd %f34,A3,%f62 ! (0_0) res += A3;
add %o1,32,%o1 ! (3_0) ind0 += 32;
lda [%l2+4]%asi,%f11 ! (4_0) ((float*)&res)[1] = ((float*)px)[1];
and %i3,32,%i4 ! (3_0) bit &= 32;
and %o1,-64,%o1 ! (3_0) ind0 &= -8;
fsubd %f18,%f54,%f18 ! (2_0) xx = (res - res_c);
sll %o1,0,%o7 ! (3_0) ind1 = ind0;
faddd %f52,A4,%f54 ! (1_0) res += A4;
fmuld %f62,%f42,%f52 ! (0_0) res *= xx;
cmp %o5,_0x00100000 ! (3_0) hx ? 0x00100000
bl,pn %icc,.update7 ! (3_0) if ( hx < 0x00100000 )
faddd %f20,A1,%f12 ! (5_1) res += A1;
.cont7:
lda [%l2]%asi,%o5 ! (4_0) hx = *(int*)px;
sllx %o3,52,%o3 ! (3_0) sqrt_exp <<= 52;
add %o7,%i4,%g1 ! (3_0) ind1 += bit;
fmuld %f18,%f50,%f18 ! (2_0) xx *= dtmp0;
add %l0,TBL,%l0 ! (5_1) pind = (char*)TBL + ind1;
stx %o3,[%fp+tmp4] ! (3_0) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (4_0) res = vis_fand(res,DC0);
fmuld %f54,%f14,%f34 ! (1_0) res *= xx;
add %l2,stridex,%l2 ! px += stridex
ldd [%l0+16],%f36 ! (5_1) dexp_lo = ((double*)pind)[2];
fpadd32 %f44,DC2,%f54 ! (3_0) res_c = vis_fpadd32(res,DC2);
fmuld %f12,%f40,%f12 ! (5_1) res *= xx;
add %g1,TBL,%o1 ! (3_0) (char*)div_arr+ind0
ldd [%l0+8],%f40 ! (5_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (0_0) res += A2;
cmp %o5,_0x7ff00000 ! (4_0) hx ? 0x7ff00000
bge,pn %icc,.update8 ! (4_0) if ( hx >= 0x7ff00000 )
for %f50,A1,%f24 ! (4_0) res = vis_for(res,A1);
.cont8:
sra %o5,21,%l1 ! (4_0) sqrt_exp = hx >> 21;
sra %o5,15,%l0 ! (4_0) bit = hx >> 15;
ldd [%o1],%f22 ! (3_0) dtmp0 = ((double*)pind)[0];
fmuld A5,%f18,%f52 ! (2_0) res = A5 * xx;
sra %o5,7,%o1 ! (4_0) ind0 = hx >> 7;
sub %l1,512,%o3 ! (4_0) sqrt_exp -= 512;
fand %f54,DC3,%f54 ! (3_0) res_c = vis_fand(res_c,DC3);
fmuld %f20,%f42,%f20 ! (0_0) res *= xx;
and %o1,_0x00001ff8,%o1 ! (4_0) ind0 &= 0x1ff8;
lda [%l2]%asi,%f10 ! (5_0) ((float*)&res)[0] = ((float*)px)[0];
faddd %f34,A3,%f62 ! (1_0) res += A3;
fmuld %f40,%f12,%f34 ! (5_1) res = dexp_hi * res;
add %o1,32,%o1 ! (4_0) ind0 += 32;
lda [%l2+4]%asi,%f11 ! (5_0) ((float*)&res)[1] = ((float*)px)[1];
and %l0,32,%i4 ! (4_0) bit &= 32;
cmp %o5,_0x00100000 ! (4_0) hx ? 0x00100000
bl,pn %icc,.update9 ! (4_0) if ( hx < 0x00100000 )
fsubd %f44,%f54,%f44 ! (3_0) xx = (res - res_c);
.cont9:
and %o1,-64,%o1 ! (4_0) ind0 &= -8;
faddd %f52,A4,%f54 ! (2_0) res += A4;
cmp counter,6
bl,pn %icc,.tail
or %g0,%o4,%l0
ba .main_loop
nop
.align 16
.main_loop:
fmuld %f62,%f14,%f52 ! (1_1) res *= xx;
sll %o1,0,%i3 ! (4_1) ind1 = ind0;
add %i1,TBL,%i1 ! (0_1) pind = (char*)TBL + ind1;
faddd %f20,A1,%f12 ! (0_1) res += A1;
lda [%l2]%asi,%o5 ! (5_1) hx = *(int*)px;
sllx %o3,52,%o3 ! (4_1) sqrt_exp <<= 52;
add %i3,%i4,%i3 ! (4_1) ind1 += bit;
faddd %f34,%f36,%f60 ! (5_2) res += dexp_lo;
fmuld %f44,%f22,%f44 ! (3_1) xx *= dtmp0;
add %l2,stridex,%l2 ! px += stridex
stx %o3,[%fp+tmp5] ! (4_1) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (5_1) res = vis_fand(res,DC0);
fmuld %f54,%f18,%f34 ! (2_1) res *= xx;
nop
ldd [%i1+16],%f36 ! (0_1) dexp_lo = ((double*)pind)[2];
fpadd32 %f24,DC2,%f54 ! (4_1) res_c = vis_fpadd32(res,DC2);
fmuld %f12,%f42,%f16 ! (0_1) res *= xx;
sra %o5,21,%l1 ! (5_1) sqrt_exp = hx >> 21;
ldd [%i1+8],%f42 ! (0_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (1_1) res += A2;
ldd [%fp+tmp0],%f48 ! (5_2) dsqrt_exp = *(double*)&sqrt_exp;
cmp %o5,_0x7ff00000 ! (5_1) hx ? 0x7ff00000
bge,pn %icc,.update10 ! (5_1) if ( hx >= 0x7ff00000 )
faddd %f60,%f40,%f60 ! (5_2) res += dexp_hi;
.cont10:
lda [%l2]%asi,%f10 ! (0_0) ((float*)&res)[0] = ((float*)px)[0];
sra %o5,15,%i1 ! (5_1) bit = hx >> 15;
add %i3,TBL,%o7 ! (4_1) pind = (char*)TBL + ind1
for %f50,A1,%f40 ! (5_1) res = vis_for(res,A1);
fmuld A5,%f44,%f52 ! (3_1) res = A5 * xx;
sra %o5,7,%o1 ! (5_1) ind0 = hx >> 7;
ldd [%o7],%f22 ! (4_1) dtmp0 = ((double*)pind)[0];
fand %f54,DC3,%f54 ! (4_1) res_c = vis_fand(res_c,DC3);
fmuld %f20,%f14,%f20 ! (1_1) res *= xx;
and %o1,_0x00001ff8,%o1 ! (5_1) ind0 &= 0x1ff8;
sub %l1,512,%o3 ! (5_1) sqrt_exp -= 512;
faddd %f34,A3,%f62 ! (2_1) res += A3;
fpadd32 %f48,%f60,%f12 ! (5_2) dtmp0 = vis_fpadd32(dsqrt_exp,res);
add %o1,32,%o1 ! (5_1) ind0 += 32;
st %f12,[%l0] ! (5_2) ((float*)py)[0] = ((float*)&dtmp0)[0];
fmuld %f42,%f16,%f34 ! (0_1) res = dexp_hi * res;
lda [%l2+4]%asi,%f11 ! (0_0) ((float*)&res)[1] = ((float*)px)[1];
and %i1,32,%i4 ! (5_1) bit &= 32;
and %o1,-64,%o1 ! (5_1) ind0 &= -8;
fsubd %f24,%f54,%f24 ! (4_1) xx = (res - res_c);
sll %o1,0,%o7 ! (5_1) ind1 = ind0;
add %l0,stridey,%i1 ! py += stridey
st %f13,[%l0+4] ! (5_2) ((float*)py)[1] = ((float*)&dtmp0)[1];
faddd %f52,A4,%f54 ! (3_1) res += A4;
fmuld %f62,%f18,%f52 ! (2_1) res *= xx;
cmp %o5,_0x00100000 ! (5_1) hx ? 0x00100000
bl,pn %icc,.update11 ! (5_1) if ( hx < 0x00100000 )
faddd %f20,A1,%f12 ! (1_1) res += A1;
.cont11:
sllx %o3,52,%o3 ! (5_1) sqrt_exp <<= 52;
add %o7,%i4,%l0 ! (5_1) ind1 += bit;
lda [%l2]%asi,%o5 ! (0_0) hx = *(int*)px;
faddd %f34,%f36,%f60 ! (0_1) res += dexp_lo;
fmuld %f24,%f22,%f24 ! (4_1) xx *= dtmp0;
add %i2,TBL,%i2 ! (1_1) pind = (char*)TBL + ind1;
stx %o3,[%fp+tmp0] ! (5_1) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (0_0) res = vis_fand(res,DC0);
fmuld %f54,%f44,%f34 ! (3_1) res *= xx;
add %l2,stridex,%l2 ! px += stridex
ldd [%i2+16],%f36 ! (1_1) dexp_lo = ((double*)pind)[2];
fpadd32 %f40,DC2,%f54 ! (5_1) res_c = vis_fpadd32(res,DC2);
fmuld %f12,%f14,%f16 ! (1_1) res *= xx;
sra %o5,21,%l1 ! (0_0) sqrt_exp = hx >> 21;
ldd [%i2+8],%f14 ! (1_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (2_1) res += A2;
ldd [%fp+tmp1],%f48 ! (0_1) dsqrt_exp = *(double*)&sqrt_exp;
cmp %o5,_0x7ff00000 ! (0_0) hx ? 0x7ff00000
bge,pn %icc,.update12 ! (0_0) if ( hx >= 0x7ff00000 )
faddd %f60,%f42,%f60 ! (0_1) res += dexp_hi;
.cont12:
lda [%l2]%asi,%f10 ! (1_0) ((float*)&res)[0] = ((float*)px)[0];
sra %o5,15,%i2 ! (0_0) bit = hx >> 15;
add %l0,TBL,%o7 ! (5_1) pind = (char*)TBL + ind1
for %f50,A1,%f42 ! (0_0) res = vis_for(res,A1);
fmuld A5,%f24,%f52 ! (4_1) res = A5 * xx;
sra %o5,7,%o1 ! (0_0) ind0 = hx >> 7;
ldd [%o7],%f22 ! (5_1) dtmp0 = ((double*)pind)[0];
fand %f54,DC3,%f54 ! (5_1) res_c = vis_fand(res_c,DC3);
fmuld %f20,%f18,%f20 ! (2_1) res *= xx;
and %o1,_0x00001ff8,%o1 ! (0_0) ind0 &= 0x1ff8;
sub %l1,512,%o3 ! (0_0) sqrt_exp -= 512;
faddd %f34,A3,%f62 ! (3_1) res += A3;
fpadd32 %f48,%f60,%f12 ! (0_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
add %o1,32,%o1 ! (0_0) ind0 += 32;
st %f12,[%i1] ! (0_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
fmuld %f14,%f16,%f34 ! (1_1) res = dexp_hi * res;
lda [%l2+4]%asi,%f11 ! (1_0) ((float*)&res)[1] = ((float*)px)[1];
and %i2,32,%i4 ! (0_0) bit &= 32;
and %o1,-64,%o1 ! (0_0) ind0 &= -8;
fsubd %f40,%f54,%f40 ! (5_1) xx = (res - res_c);
sll %o1,0,%o7 ! (0_0) ind1 = ind0;
add %i1,stridey,%i2 ! py += stridey
st %f13,[%i1+4] ! (0_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
faddd %f52,A4,%f54 ! (4_1) res += A4;
fmuld %f62,%f44,%f52 ! (3_1) res *= xx;
cmp %o5,_0x00100000 ! (0_0) hx ? 0x00100000
bl,pn %icc,.update13 ! (0_0) if ( hx < 0x00100000 )
faddd %f20,A1,%f12 ! (2_1) res += A1;
.cont13:
lda [%l2]%asi,%o5 ! (1_0) hx = *(int*)px;
sllx %o3,52,%o3 ! (0_0) sqrt_exp <<= 52;
add %o7,%i4,%i1 ! (0_0) ind1 += bit;
faddd %f34,%f36,%f60 ! (1_1) res += dexp_lo;
fmuld %f40,%f22,%f40 ! (5_1) xx *= dtmp0;
add %g5,TBL,%g5 ! (2_1) pind = (char*)TBL + ind1;
stx %o3,[%fp+tmp1] ! (0_0) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (1_0) res = vis_fand(res,DC0);
fmuld %f54,%f24,%f34 ! (4_1) res *= xx;
add %l2,stridex,%l2 ! px += stridex
ldd [%g5+16],%f36 ! (2_1) dexp_lo = ((double*)pind)[2];
fpadd32 %f42,DC2,%f54 ! (0_0) res_c = vis_fpadd32(res,DC2);
fmuld %f12,%f18,%f16 ! (2_1) res *= xx;
sra %o5,21,%l1 ! (1_0) sqrt_exp = hx >> 21;
ldd [%g5+8],%f18 ! (2_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (3_1) res += A2;
ldd [%fp+tmp2],%f48 ! (1_1) dsqrt_exp = *(double*)&sqrt_exp;
cmp %o5,_0x7ff00000 ! (1_0) hx ? 0x7ff00000
bge,pn %icc,.update14 ! (1_0) if ( hx >= 0x7ff00000 )
faddd %f60,%f14,%f60 ! (1_1) res += dexp_hi;
.cont14:
lda [%l2]%asi,%f10 ! (2_0) ((float*)&res)[0] = ((float*)px)[0];
sra %o5,15,%g5 ! (1_0) bit = hx >> 15;
add %i1,TBL,%o7 ! (0_0) pind = (char*)TBL + ind1
for %f50,A1,%f14 ! (1_0) res = vis_for(res,A1);
fmuld A5,%f40,%f52 ! (5_1) res = A5 * xx;
sra %o5,7,%o1 ! (1_0) ind0 = hx >> 7;
ldd [%o7],%f22 ! (0_0) dtmp0 = ((double*)pind)[0];
fand %f54,DC3,%f54 ! (0_0) res_c = vis_fand(res_c,DC3);
fmuld %f20,%f44,%f20 ! (3_1) res *= xx;
and %o1,_0x00001ff8,%o1 ! (1_0) ind0 &= 0x1ff8;
sub %l1,512,%o3 ! (1_0) sqrt_exp -= 512;
faddd %f34,A3,%f62 ! (4_1) res += A3;
fpadd32 %f48,%f60,%f12 ! (1_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
add %o1,32,%o1 ! (1_0) ind0 += 32;
st %f12,[%i2] ! (1_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
fmuld %f18,%f16,%f34 ! (2_1) res = dexp_hi * res;
lda [%l2+4]%asi,%f11 ! (2_0) ((float*)&res)[1] = ((float*)px)[1];
and %g5,32,%i4 ! (1_0) bit &= 32;
and %o1,-64,%o1 ! (1_0) ind0 &= -8;
fsubd %f42,%f54,%f42 ! (0_0) xx = (res - res_c);
sll %o1,0,%o7 ! (1_0) ind1 = ind0;
add %i2,stridey,%g5 ! py += stridey
st %f13,[%i2+4] ! (1_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
faddd %f52,A4,%f54 ! (5_1) res += A4;
fmuld %f62,%f24,%f52 ! (4_1) res *= xx;
cmp %o5,_0x00100000 ! (1_0) hx ? 0x00100000
bl,pn %icc,.update15 ! (1_0) if ( hx < 0x00100000 )
faddd %f20,A1,%f12 ! (3_1) res += A1;
.cont15:
lda [%l2]%asi,%o5 ! (2_0) hx = *(int*)px;
sllx %o3,52,%o3 ! (1_0) sqrt_exp <<= 52;
add %o7,%i4,%i2 ! (1_0) ind1 += bit;
faddd %f34,%f36,%f60 ! (2_1) res += dexp_lo;
fmuld %f42,%f22,%f42 ! (0_0) xx *= dtmp0;
add %g1,TBL,%g1 ! (3_1) pind = (char*)TBL + ind1;
stx %o3,[%fp+tmp2] ! (1_0) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (2_0) res = vis_fand(res,DC0);
fmuld %f54,%f40,%f34 ! (5_1) res *= xx;
fpadd32 %f14,DC2,%f54 ! (1_0) res_c = vis_fpadd32(res,DC2);
add %l2,stridex,%l2 ! px += stridex
ldd [%g1+16],%f36 ! (3_1) dexp_lo = ((double*)pind)[2];
fmuld %f12,%f44,%f16 ! (3_1) res *= xx;
sra %o5,21,%l1 ! (2_0) sqrt_exp = hx >> 21;
ldd [%g1+8],%f44 ! (3_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (4_1) res += A2;
ldd [%fp+tmp3],%f48 ! (2_1) dsqrt_exp = *(double*)&sqrt_exp;
cmp %o5,_0x7ff00000 ! (2_0) hx ? 0x7ff00000
bge,pn %icc,.update16 ! (2_0) if ( hx >= 0x7ff00000 )
faddd %f60,%f18,%f60 ! (2_1) res += dexp_hi;
.cont16:
lda [%l2]%asi,%f10 ! (3_0) ((float*)&res)[0] = ((float*)px)[0];
sra %o5,15,%g1 ! (2_0) bit = hx >> 15;
add %i2,TBL,%o7 ! (1_0) pind = (char*)TBL + ind1
for %f50,A1,%f18 ! (2_0) res = vis_for(res,A1);
fmuld A5,%f42,%f52 ! (0_0) res = A5 * xx;
sra %o5,7,%o1 ! (2_0) ind0 = hx >> 7;
ldd [%o7],%f22 ! (1_0) dtmp0 = ((double*)pind)[0];
fand %f54,DC3,%f54 ! (1_0) res_c = vis_fand(res_c,DC3);
fmuld %f20,%f24,%f20 ! (4_1) res *= xx;
and %o1,_0x00001ff8,%o1 ! (2_0) ind0 &= 0x1ff8;
sub %l1,512,%o3 ! (2_0) sqrt_exp -= 512;
faddd %f34,A3,%f62 ! (5_1) res += A3;
fpadd32 %f48,%f60,%f12 ! (2_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
add %o1,32,%o1 ! (2_0) ind0 += 32;
st %f12,[%g5] ! (2_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
fmuld %f44,%f16,%f34 ! (3_1) res = dexp_hi * res;
lda [%l2+4]%asi,%f11 ! (3_0) ((float*)&res)[1] = ((float*)px)[1];
and %g1,32,%i4 ! (2_0) bit &= 32;
and %o1,-64,%o1 ! (2_0) ind0 &= -8;
fsubd %f14,%f54,%f14 ! (1_0) xx = (res - res_c);
sll %o1,0,%o7 ! (2_0) ind1 = ind0;
add %g5,stridey,%g1 ! py += stridey
st %f13,[%g5+4] ! (2_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
faddd %f52,A4,%f54 ! (0_0) res += A4;
fmuld %f62,%f40,%f52 ! (5_1) res *= xx;
cmp %o5,_0x00100000 ! (2_0) hx ? 0x00100000
bl,pn %icc,.update17 ! (2_0) if ( hx < 0x00100000 )
faddd %f20,A1,%f12 ! (4_1) res += A1;
.cont17:
lda [%l2]%asi,%o5 ! (3_0) hx = *(int*)px;
sllx %o3,52,%o3 ! (2_0) sqrt_exp <<= 52;
add %o7,%i4,%g5 ! (2_0) ind1 += bit;
faddd %f34,%f36,%f60 ! (3_1) res += dexp_lo;
fmuld %f14,%f22,%f14 ! (1_0) xx *= dtmp0;
add %i3,TBL,%i3 ! (4_1) pind = (char*)TBL + ind1;
stx %o3,[%fp+tmp3] ! (2_0) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (3_0) res = vis_fand(res,DC0);
fmuld %f54,%f42,%f34 ! (0_0) res *= xx;
fpadd32 %f18,DC2,%f54 ! (2_0) res_c = vis_fpadd32(res,DC2);
add %l2,stridex,%l2 ! px += stridex
ldd [%i3+16],%f36 ! (4_1) dexp_lo = ((double*)pind)[2];
fmuld %f12,%f24,%f16 ! (4_1) res *= xx;
sra %o5,21,%l1 ! (3_0) sqrt_exp = hx >> 21;
ldd [%i3+8],%f24 ! (4_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (5_1) res += A2;
ldd [%fp+tmp4],%f48 ! (3_1) dsqrt_exp = *(double*)&sqrt_exp;
cmp %o5,_0x7ff00000 ! (3_0) hx ? 0x7ff00000
bge,pn %icc,.update18 ! (3_0) if ( hx >= 0x7ff00000 )
faddd %f60,%f44,%f60 ! (3_1) res += dexp_hi;
.cont18:
lda [%l2]%asi,%f10 ! (4_0) ((float*)&res)[0] = ((float*)px)[0];
sra %o5,15,%i3 ! (3_0) bit = hx >> 15;
add %g5,TBL,%o7 ! (2_0) pind = (char*)TBL + ind1
for %f50,A1,%f44 ! (3_0) res = vis_for(res,A1);
fmuld A5,%f14,%f52 ! (1_0) res = A5 * xx;
sra %o5,7,%o1 ! (3_0) ind0 = hx >> 7;
ldd [%o7],%f22 ! (2_0) dtmp0 = ((double*)pind)[0];
fand %f54,DC3,%f54 ! (2_0) res_c = vis_fand(res_c,DC3);
fmuld %f20,%f40,%f20 ! (5_1) res *= xx;
and %o1,_0x00001ff8,%o1 ! (3_0) ind0 &= 0x1ff8;
sub %l1,512,%o3 ! (3_0) sqrt_exp -= 512;
faddd %f34,A3,%f62 ! (0_0) res += A3;
fpadd32 %f48,%f60,%f12 ! (3_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
add %o1,32,%o1 ! (3_0) ind0 += 32;
st %f12,[%g1] ! (3_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
fmuld %f24,%f16,%f34 ! (4_1) res = dexp_hi * res;
lda [%l2+4]%asi,%f11 ! (4_0) ((float*)&res)[1] = ((float*)px)[1];
and %i3,32,%i4 ! (3_0) bit &= 32;
and %o1,-64,%o1 ! (3_0) ind0 &= -8;
fsubd %f18,%f54,%f18 ! (2_0) xx = (res - res_c);
or %g0,%o1,%o7 ! (3_0) ind1 = ind0;
add %g1,stridey,%i3 ! py += stridey
st %f13,[%g1+4] ! (3_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
faddd %f52,A4,%f54 ! (1_0) res += A4;
fmuld %f62,%f42,%f52 ! (0_0) res *= xx;
cmp %o5,_0x00100000 ! (3_0) hx ? 0x00100000
bl,pn %icc,.update19 ! (3_0) if ( hx < 0x00100000 )
faddd %f20,A1,%f12 ! (5_1) res += A1;
.cont19:
lda [%l2]%asi,%o5 ! (4_0) hx = *(int*)px;
sllx %o3,52,%o3 ! (3_0) sqrt_exp <<= 52;
add %o7,%i4,%g1 ! (3_0) ind1 += bit;
faddd %f34,%f36,%f60 ! (4_1) res += dexp_lo;
fmuld %f18,%f22,%f18 ! (2_0) xx *= dtmp0;
add %l0,TBL,%l0 ! (5_1) pind = (char*)TBL + ind1;
stx %o3,[%fp+tmp4] ! (3_0) dsqrt_exp = *(double*)&sqrt_exp;
fand %f10,DC0,%f50 ! (4_0) res = vis_fand(res,DC0);
fmuld %f54,%f14,%f34 ! (1_0) res *= xx;
add %l2,stridex,%l2 ! px += stridex
ldd [%l0+16],%f36 ! (5_1) dexp_lo = ((double*)pind)[2];
fpadd32 %f44,DC2,%f54 ! (3_0) res_c = vis_fpadd32(res,DC2);
fmuld %f12,%f40,%f16 ! (5_1) res *= xx;
sra %o5,21,%l1 ! (4_0) sqrt_exp = hx >> 21;
ldd [%l0+8],%f40 ! (5_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (0_0) res += A2;
ldd [%fp+tmp5],%f48 ! (4_1) dsqrt_exp = *(double*)&sqrt_exp;
cmp %o5,_0x7ff00000 ! (4_0) hx ? 0x7ff00000
bge,pn %icc,.update20 ! (4_0) if ( hx >= 0x7ff00000 )
faddd %f60,%f24,%f60 ! (4_1) res += dexp_hi;
.cont20:
lda [%l2]%asi,%f10 ! (5_0) ((float*)&res)[0] = ((float*)px)[0];
sra %o5,15,%l0 ! (4_0) bit = hx >> 15;
add %g1,TBL,%o7 ! (3_0) (char*)div_arr+ind0
for %f50,A1,%f24 ! (4_0) res = vis_for(res,A1);
fmuld A5,%f18,%f52 ! (2_0) res = A5 * xx;
sra %o5,7,%o1 ! (4_0) ind0 = hx >> 7;
ldd [%o7],%f22 ! (3_0) dtmp0 = ((double*)pind)[0];
fand %f54,DC3,%f54 ! (3_0) res_c = vis_fand(res_c,DC3);
fmuld %f20,%f42,%f20 ! (0_0) res *= xx;
and %o1,_0x00001ff8,%o1 ! (4_0) ind0 &= 0x1ff8;
sub %l1,512,%o3 ! (4_0) sqrt_exp -= 512;
faddd %f34,A3,%f62 ! (1_0) res += A3;
lda [%l2+4]%asi,%f11 ! (5_0) ((float*)&res)[1] = ((float*)px)[1];
add %o1,32,%o1 ! (4_0) ind0 += 32;
fpadd32 %f48,%f60,%f12 ! (4_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
fmuld %f40,%f16,%f34 ! (5_1) res = dexp_hi * res;
and %l0,32,%i4 ! (4_0) bit &= 32;
cmp %o5,_0x00100000 ! (4_0) hx ? 0x00100000
bl,pn %icc,.update21 ! (4_0) if ( hx < 0x00100000 )
fsubd %f44,%f54,%f44 ! (3_0) xx = (res - res_c);
.cont21:
and %o1,-64,%o1 ! (4_0) ind0 &= -8;
sub counter,6,counter ! counter
st %f12,[%i3] ! (4_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
faddd %f52,A4,%f54 ! (2_0) res += A4;
st %f13,[%i3+4] ! (4_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
cmp counter,6
bge,pt %icc,.main_loop
add %i3,stridey,%l0 ! py += stridey
.tail:
subcc counter,1,counter
bneg .begin
or %g0,%l0,%o4
fmuld %f62,%f14,%f52 ! (1_1) res *= xx;
add %i1,TBL,%i1 ! (0_1) pind = (char*)TBL + ind1;
faddd %f20,A1,%f12 ! (0_1) res += A1;
faddd %f34,%f36,%f60 ! (5_2) res += dexp_lo;
fmuld %f44,%f22,%f44 ! (3_1) xx *= dtmp0;
add %l2,stridex,%l2 ! px += stridex
fmuld %f54,%f18,%f34 ! (2_1) res *= xx;
ldd [%i1+16],%f36 ! (0_1) dexp_lo = ((double*)pind)[2];
fmuld %f12,%f42,%f12 ! (0_1) res *= xx;
ldd [%i1+8],%f42 ! (0_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (1_1) res += A2;
ldd [%fp+tmp0],%f48 ! (5_2) dsqrt_exp = *(double*)&sqrt_exp;
faddd %f60,%f40,%f60 ! (5_2) res += dexp_hi;
fmuld A5,%f44,%f52 ! (3_1) res = A5 * xx;
fmuld %f20,%f14,%f20 ! (1_1) res *= xx;
faddd %f34,A3,%f62 ! (2_1) res += A3;
fmuld %f42,%f12,%f34 ! (0_1) res = dexp_hi * res;
fpadd32 %f48,%f60,%f12 ! (5_2) dtmp0 = vis_fpadd32(dsqrt_exp,res);
st %f12,[%l0] ! (5_2) ((float*)py)[0] = ((float*)&dtmp0)[0];
add %l0,stridey,%i1 ! py += stridey
st %f13,[%l0+4] ! (5_2) ((float*)py)[1] = ((float*)&dtmp0)[1];
faddd %f52,A4,%f54 ! (3_1) res += A4;
subcc counter,1,counter
bneg .begin
or %g0,%i1,%o4
fmuld %f62,%f18,%f52 ! (2_1) res *= xx;
faddd %f20,A1,%f12 ! (1_1) res += A1;
faddd %f34,%f36,%f60 ! (0_1) res += dexp_lo;
add %i2,TBL,%i2 ! (1_1) pind = (char*)TBL + ind1;
fmuld %f54,%f44,%f34 ! (3_1) res *= xx;
add %l2,stridex,%l2 ! px += stridex
ldd [%i2+16],%f36 ! (1_1) dexp_lo = ((double*)pind)[2];
fmuld %f12,%f14,%f12 ! (1_1) res *= xx;
ldd [%i2+8],%f14 ! (1_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (2_1) res += A2;
ldd [%fp+tmp1],%f48 ! (0_1) dsqrt_exp = *(double*)&sqrt_exp;
faddd %f60,%f42,%f60 ! (0_1) res += dexp_hi;
fmuld %f20,%f18,%f20 ! (2_1) res *= xx;
faddd %f34,A3,%f62 ! (3_1) res += A3;
fmuld %f14,%f12,%f34 ! (1_1) res = dexp_hi * res;
fpadd32 %f48,%f60,%f12 ! (0_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
st %f12,[%i1] ! (0_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
add %i1,stridey,%i2 ! py += stridey
st %f13,[%i1+4] ! (0_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
subcc counter,1,counter
bneg .begin
or %g0,%i2,%o4
fmuld %f62,%f44,%f52 ! (3_1) res *= xx;
faddd %f20,A1,%f12 ! (2_1) res += A1;
faddd %f34,%f36,%f60 ! (1_1) res += dexp_lo;
add %g5,TBL,%g5 ! (2_1) pind = (char*)TBL + ind1;
add %l2,stridex,%l2 ! px += stridex
ldd [%g5+16],%f36 ! (2_1) dexp_lo = ((double*)pind)[2];
fmuld %f12,%f18,%f12 ! (2_1) res *= xx;
ldd [%g5+8],%f18 ! (2_1) dexp_hi = ((double*)pind)[1];
faddd %f52,A2,%f20 ! (3_1) res += A2;
ldd [%fp+tmp2],%f48 ! (1_1) dsqrt_exp = *(double*)&sqrt_exp;
faddd %f60,%f14,%f60 ! (1_1) res += dexp_hi;
fmuld %f20,%f44,%f20 ! (3_1) res *= xx;
fmuld %f18,%f12,%f34 ! (2_1) res = dexp_hi * res;
fpadd32 %f48,%f60,%f12 ! (1_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
st %f12,[%i2] ! (1_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
add %i2,stridey,%g5 ! py += stridey
st %f13,[%i2+4] ! (1_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
subcc counter,1,counter
bneg .begin
or %g0,%g5,%o4
faddd %f20,A1,%f12 ! (3_1) res += A1;
faddd %f34,%f36,%f60 ! (2_1) res += dexp_lo;
add %g1,TBL,%g1 ! (3_1) pind = (char*)TBL + ind1;
add %l2,stridex,%l2 ! px += stridex
ldd [%g1+16],%f36 ! (3_1) dexp_lo = ((double*)pind)[2];
fmuld %f12,%f44,%f12 ! (3_1) res *= xx;
ldd [%g1+8],%f44 ! (3_1) dexp_hi = ((double*)pind)[1];
ldd [%fp+tmp3],%f48 ! (2_1) dsqrt_exp = *(double*)&sqrt_exp;
faddd %f60,%f18,%f60 ! (2_1) res += dexp_hi;
fmuld %f44,%f12,%f34 ! (3_1) res = dexp_hi * res;
fpadd32 %f48,%f60,%f12 ! (2_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
st %f12,[%g5] ! (2_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
add %g5,stridey,%g1 ! py += stridey
st %f13,[%g5+4] ! (2_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
subcc counter,1,counter
bneg .begin
or %g0,%g1,%o4
faddd %f34,%f36,%f60 ! (3_1) res += dexp_lo;
add %l2,stridex,%l2 ! px += stridex
ldd [%fp+tmp4],%f48 ! (3_1) dsqrt_exp = *(double*)&sqrt_exp;
faddd %f60,%f44,%f60 ! (3_1) res += dexp_hi;
fpadd32 %f48,%f60,%f12 ! (3_1) dtmp0 = vis_fpadd32(dsqrt_exp,res);
st %f12,[%g1] ! (3_1) ((float*)py)[0] = ((float*)&dtmp0)[0];
add %g1,stridey,%i3 ! py += stridey
st %f13,[%g1+4] ! (3_1) ((float*)py)[1] = ((float*)&dtmp0)[1];
ba .begin
or %g0,%i3,%o4
.align 16
.spec:
fsqrtd %f10,%f10
add %l2,stridex,%l2
st %f10,[%o4]
st %f11,[%o4+4]
add %o4,stridey,%o4
ba .begin1
sub counter,1,counter
.align 16
.update0:
cmp counter,1
ble .cont0
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,1,counter
st counter,[%fp+tmp_counter]
ba .cont0
or %g0,1,counter
.align 16
.update1:
cmp counter,1
ble .cont1
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,1,counter
st counter,[%fp+tmp_counter]
ba .cont1
or %g0,1,counter
.align 16
.update2:
cmp counter,2
ble .cont2
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,2,counter
st counter,[%fp+tmp_counter]
ba .cont2
or %g0,2,counter
.align 16
.update3:
cmp counter,2
ble .cont3
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,2,counter
st counter,[%fp+tmp_counter]
ba .cont3
or %g0,2,counter
.align 16
.update4:
cmp counter,3
ble .cont4
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,3,counter
st counter,[%fp+tmp_counter]
ba .cont4
or %g0,3,counter
.align 16
.update5:
cmp counter,3
ble .cont5
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,3,counter
st counter,[%fp+tmp_counter]
ba .cont5
or %g0,3,counter
.align 16
.update6:
cmp counter,4
ble .cont6
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,4,counter
st counter,[%fp+tmp_counter]
ba .cont6
or %g0,4,counter
.align 16
.update7:
cmp counter,4
ble .cont7
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,4,counter
st counter,[%fp+tmp_counter]
ba .cont7
or %g0,4,counter
.align 16
.update8:
cmp counter,5
ble .cont8
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,5,counter
st counter,[%fp+tmp_counter]
ba .cont8
or %g0,5,counter
.align 16
.update9:
cmp counter,5
ble .cont9
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,5,counter
st counter,[%fp+tmp_counter]
ba .cont9
or %g0,5,counter
.align 16
.update10:
cmp counter,6
ble .cont10
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,6,counter
st counter,[%fp+tmp_counter]
ba .cont10
or %g0,6,counter
.align 16
.update11:
cmp counter,6
ble .cont11
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,6,counter
st counter,[%fp+tmp_counter]
ba .cont11
or %g0,6,counter
.align 16
.update12:
cmp counter,7
ble .cont12
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,7,counter
st counter,[%fp+tmp_counter]
ba .cont12
or %g0,7,counter
.align 16
.update13:
cmp counter,7
ble .cont13
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,7,counter
st counter,[%fp+tmp_counter]
ba .cont13
or %g0,7,counter
.align 16
.update14:
cmp counter,8
ble .cont14
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,8,counter
st counter,[%fp+tmp_counter]
ba .cont14
or %g0,8,counter
.align 16
.update15:
cmp counter,8
ble .cont15
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,8,counter
st counter,[%fp+tmp_counter]
ba .cont15
or %g0,8,counter
.align 16
.update16:
cmp counter,9
ble .cont16
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,9,counter
st counter,[%fp+tmp_counter]
ba .cont16
or %g0,9,counter
.align 16
.update17:
cmp counter,9
ble .cont17
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,9,counter
st counter,[%fp+tmp_counter]
ba .cont17
or %g0,9,counter
.align 16
.update18:
cmp counter,10
ble .cont18
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,10,counter
st counter,[%fp+tmp_counter]
ba .cont18
or %g0,10,counter
.align 16
.update19:
cmp counter,10
ble .cont19
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,10,counter
st counter,[%fp+tmp_counter]
ba .cont19
or %g0,10,counter
.align 16
.update20:
cmp counter,11
ble .cont20
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,11,counter
st counter,[%fp+tmp_counter]
ba .cont20
or %g0,11,counter
.align 16
.update21:
cmp counter,11
ble .cont21
nop
sub %l2,stridex,%i5
stx %i5,[%fp+tmp_px]
sub counter,11,counter
st counter,[%fp+tmp_counter]
ba .cont21
or %g0,11,counter
.exit:
ret
restore
SET_SIZE(__vsqrt)
