diff options
Diffstat (limited to 'usr/src/lib/libmvec/common/vis/__vpowf.S')
| -rw-r--r-- | usr/src/lib/libmvec/common/vis/__vpowf.S | 3139 |
1 files changed, 3139 insertions, 0 deletions
diff --git a/usr/src/lib/libmvec/common/vis/__vpowf.S b/usr/src/lib/libmvec/common/vis/__vpowf.S new file mode 100644 index 0000000000..cddb99ef99 --- /dev/null +++ b/usr/src/lib/libmvec/common/vis/__vpowf.S @@ -0,0 +1,3139 @@ +/* + * CDDL HEADER START + * + * The contents of this file are subject to the terms of the + * Common Development and Distribution License (the "License"). + * You may not use this file except in compliance with the License. + * + * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE + * or http://www.opensolaris.org/os/licensing. + * See the License for the specific language governing permissions + * and limitations under the License. + * + * When distributing Covered Code, include this CDDL HEADER in each + * file and include the License file at usr/src/OPENSOLARIS.LICENSE. + * If applicable, add the following below this CDDL HEADER, with the + * fields enclosed by brackets "[]" replaced with your own identifying + * information: Portions Copyright [yyyy] [name of copyright owner] + * + * CDDL HEADER END + */ +/* + * Copyright 2011 Nexenta Systems, Inc. All rights reserved. + */ +/* + * Copyright 2006 Sun Microsystems, Inc. All rights reserved. + * Use is subject to license terms. + */ + + .file "__vpowf.S" + +#include "libm.h" + + RO_DATA + .align 64 + +! __mt_constexp2fa: + .word 0x3ff00000, 0x00000000, 0x3ff00b1a, 0xfa5abcbf + .word 0x3ff0163d, 0xa9fb3335, 0x3ff02168, 0x143b0281 + .word 0x3ff02c9a, 0x3e778061, 0x3ff037d4, 0x2e11bbcc + .word 0x3ff04315, 0xe86e7f85, 0x3ff04e5f, 0x72f654b1 + .word 0x3ff059b0, 0xd3158574, 0x3ff0650a, 0x0e3c1f89 + .word 0x3ff0706b, 0x29ddf6de, 0x3ff07bd4, 0x2b72a836 + .word 0x3ff08745, 0x18759bc8, 0x3ff092bd, 0xf66607e0 + .word 0x3ff09e3e, 0xcac6f383, 0x3ff0a9c7, 0x9b1f3919 + .word 0x3ff0b558, 0x6cf9890f, 0x3ff0c0f1, 0x45e46c85 + .word 0x3ff0cc92, 0x2b7247f7, 0x3ff0d83b, 0x23395dec + .word 0x3ff0e3ec, 0x32d3d1a2, 0x3ff0efa5, 0x5fdfa9c5 + .word 0x3ff0fb66, 0xaffed31b, 0x3ff10730, 0x28d7233e + .word 0x3ff11301, 0xd0125b51, 0x3ff11edb, 0xab5e2ab6 + .word 0x3ff12abd, 0xc06c31cc, 0x3ff136a8, 0x14f204ab + .word 0x3ff1429a, 0xaea92de0, 0x3ff14e95, 0x934f312e + .word 0x3ff15a98, 0xc8a58e51, 0x3ff166a4, 0x5471c3c2 + .word 0x3ff172b8, 0x3c7d517b, 0x3ff17ed4, 0x8695bbc0 + .word 0x3ff18af9, 0x388c8dea, 0x3ff19726, 0x58375d2f + .word 0x3ff1a35b, 0xeb6fcb75, 0x3ff1af99, 0xf8138a1c + .word 0x3ff1bbe0, 0x84045cd4, 0x3ff1c82f, 0x95281c6b + .word 0x3ff1d487, 0x3168b9aa, 0x3ff1e0e7, 0x5eb44027 + .word 0x3ff1ed50, 0x22fcd91d, 0x3ff1f9c1, 0x8438ce4d + .word 0x3ff2063b, 0x88628cd6, 0x3ff212be, 0x3578a819 + .word 0x3ff21f49, 0x917ddc96, 0x3ff22bdd, 0xa27912d1 + .word 0x3ff2387a, 0x6e756238, 0x3ff2451f, 0xfb82140a + .word 0x3ff251ce, 0x4fb2a63f, 0x3ff25e85, 0x711ece75 + .word 0x3ff26b45, 0x65e27cdd, 0x3ff2780e, 0x341ddf29 + .word 0x3ff284df, 0xe1f56381, 0x3ff291ba, 0x7591bb70 + .word 0x3ff29e9d, 0xf51fdee1, 0x3ff2ab8a, 0x66d10f13 + .word 0x3ff2b87f, 0xd0dad990, 0x3ff2c57e, 0x39771b2f + .word 0x3ff2d285, 0xa6e4030b, 0x3ff2df96, 0x1f641589 + .word 0x3ff2ecaf, 0xa93e2f56, 0x3ff2f9d2, 0x4abd886b + .word 0x3ff306fe, 0x0a31b715, 0x3ff31432, 0xedeeb2fd + .word 0x3ff32170, 0xfc4cd831, 0x3ff32eb8, 0x3ba8ea32 + .word 0x3ff33c08, 0xb26416ff, 0x3ff34962, 0x66e3fa2d + .word 0x3ff356c5, 0x5f929ff1, 0x3ff36431, 0xa2de883b + .word 0x3ff371a7, 0x373aa9cb, 0x3ff37f26, 0x231e754a + .word 0x3ff38cae, 0x6d05d866, 0x3ff39a40, 0x1b7140ef + .word 0x3ff3a7db, 0x34e59ff7, 0x3ff3b57f, 0xbfec6cf4 + .word 0x3ff3c32d, 0xc313a8e5, 0x3ff3d0e5, 0x44ede173 + .word 0x3ff3dea6, 0x4c123422, 0x3ff3ec70, 0xdf1c5175 + .word 0x3ff3fa45, 0x04ac801c, 0x3ff40822, 0xc367a024 + .word 0x3ff4160a, 0x21f72e2a, 0x3ff423fb, 0x2709468a + .word 0x3ff431f5, 0xd950a897, 0x3ff43ffa, 0x3f84b9d4 + .word 0x3ff44e08, 0x6061892d, 0x3ff45c20, 0x42a7d232 + .word 0x3ff46a41, 0xed1d0057, 0x3ff4786d, 0x668b3237 + .word 0x3ff486a2, 0xb5c13cd0, 0x3ff494e1, 0xe192aed2 + .word 0x3ff4a32a, 0xf0d7d3de, 0x3ff4b17d, 0xea6db7d7 + .word 0x3ff4bfda, 0xd5362a27, 0x3ff4ce41, 0xb817c114 + .word 0x3ff4dcb2, 0x99fddd0d, 0x3ff4eb2d, 0x81d8abff + .word 0x3ff4f9b2, 0x769d2ca7, 0x3ff50841, 0x7f4531ee + .word 0x3ff516da, 0xa2cf6642, 0x3ff5257d, 0xe83f4eef + .word 0x3ff5342b, 0x569d4f82, 0x3ff542e2, 0xf4f6ad27 + .word 0x3ff551a4, 0xca5d920f, 0x3ff56070, 0xdde910d2 + .word 0x3ff56f47, 0x36b527da, 0x3ff57e27, 0xdbe2c4cf + .word 0x3ff58d12, 0xd497c7fd, 0x3ff59c08, 0x27ff07cc + .word 0x3ff5ab07, 0xdd485429, 0x3ff5ba11, 0xfba87a03 + .word 0x3ff5c926, 0x8a5946b7, 0x3ff5d845, 0x90998b93 + .word 0x3ff5e76f, 0x15ad2148, 0x3ff5f6a3, 0x20dceb71 + .word 0x3ff605e1, 0xb976dc09, 0x3ff6152a, 0xe6cdf6f4 + .word 0x3ff6247e, 0xb03a5585, 0x3ff633dd, 0x1d1929fd + .word 0x3ff64346, 0x34ccc320, 0x3ff652b9, 0xfebc8fb7 + .word 0x3ff66238, 0x82552225, 0x3ff671c1, 0xc70833f6 + .word 0x3ff68155, 0xd44ca973, 0x3ff690f4, 0xb19e9538 + .word 0x3ff6a09e, 0x667f3bcd, 0x3ff6b052, 0xfa75173e + .word 0x3ff6c012, 0x750bdabf, 0x3ff6cfdc, 0xddd47645 + .word 0x3ff6dfb2, 0x3c651a2f, 0x3ff6ef92, 0x98593ae5 + .word 0x3ff6ff7d, 0xf9519484, 0x3ff70f74, 0x66f42e87 + .word 0x3ff71f75, 0xe8ec5f74, 0x3ff72f82, 0x86ead08a + .word 0x3ff73f9a, 0x48a58174, 0x3ff74fbd, 0x35d7cbfd + .word 0x3ff75feb, 0x564267c9, 0x3ff77024, 0xb1ab6e09 + .word 0x3ff78069, 0x4fde5d3f, 0x3ff790b9, 0x38ac1cf6 + .word 0x3ff7a114, 0x73eb0187, 0x3ff7b17b, 0x0976cfdb + .word 0x3ff7c1ed, 0x0130c132, 0x3ff7d26a, 0x62ff86f0 + .word 0x3ff7e2f3, 0x36cf4e62, 0x3ff7f387, 0x8491c491 + .word 0x3ff80427, 0x543e1a12, 0x3ff814d2, 0xadd106d9 + .word 0x3ff82589, 0x994cce13, 0x3ff8364c, 0x1eb941f7 + .word 0x3ff8471a, 0x4623c7ad, 0x3ff857f4, 0x179f5b21 + .word 0x3ff868d9, 0x9b4492ed, 0x3ff879ca, 0xd931a436 + .word 0x3ff88ac7, 0xd98a6699, 0x3ff89bd0, 0xa478580f + .word 0x3ff8ace5, 0x422aa0db, 0x3ff8be05, 0xbad61778 + .word 0x3ff8cf32, 0x16b5448c, 0x3ff8e06a, 0x5e0866d9 + .word 0x3ff8f1ae, 0x99157736, 0x3ff902fe, 0xd0282c8a + .word 0x3ff9145b, 0x0b91ffc6, 0x3ff925c3, 0x53aa2fe2 + .word 0x3ff93737, 0xb0cdc5e5, 0x3ff948b8, 0x2b5f98e5 + .word 0x3ff95a44, 0xcbc8520f, 0x3ff96bdd, 0x9a7670b3 + .word 0x3ff97d82, 0x9fde4e50, 0x3ff98f33, 0xe47a22a2 + .word 0x3ff9a0f1, 0x70ca07ba, 0x3ff9b2bb, 0x4d53fe0d + .word 0x3ff9c491, 0x82a3f090, 0x3ff9d674, 0x194bb8d5 + .word 0x3ff9e863, 0x19e32323, 0x3ff9fa5e, 0x8d07f29e + .word 0x3ffa0c66, 0x7b5de565, 0x3ffa1e7a, 0xed8eb8bb + .word 0x3ffa309b, 0xec4a2d33, 0x3ffa42c9, 0x80460ad8 + .word 0x3ffa5503, 0xb23e255d, 0x3ffa674a, 0x8af46052 + .word 0x3ffa799e, 0x1330b358, 0x3ffa8bfe, 0x53c12e59 + .word 0x3ffa9e6b, 0x5579fdbf, 0x3ffab0e5, 0x21356eba + .word 0x3ffac36b, 0xbfd3f37a, 0x3ffad5ff, 0x3a3c2774 + .word 0x3ffae89f, 0x995ad3ad, 0x3ffafb4c, 0xe622f2ff + .word 0x3ffb0e07, 0x298db666, 0x3ffb20ce, 0x6c9a8952 + .word 0x3ffb33a2, 0xb84f15fb, 0x3ffb4684, 0x15b749b1 + .word 0x3ffb5972, 0x8de5593a, 0x3ffb6c6e, 0x29f1c52a + .word 0x3ffb7f76, 0xf2fb5e47, 0x3ffb928c, 0xf22749e4 + .word 0x3ffba5b0, 0x30a1064a, 0x3ffbb8e0, 0xb79a6f1f + .word 0x3ffbcc1e, 0x904bc1d2, 0x3ffbdf69, 0xc3f3a207 + .word 0x3ffbf2c2, 0x5bd71e09, 0x3ffc0628, 0x6141b33d + .word 0x3ffc199b, 0xdd85529c, 0x3ffc2d1c, 0xd9fa652c + .word 0x3ffc40ab, 0x5fffd07a, 0x3ffc5447, 0x78fafb22 + .word 0x3ffc67f1, 0x2e57d14b, 0x3ffc7ba8, 0x8988c933 + .word 0x3ffc8f6d, 0x9406e7b5, 0x3ffca340, 0x5751c4db + .word 0x3ffcb720, 0xdcef9069, 0x3ffccb0f, 0x2e6d1675 + .word 0x3ffcdf0b, 0x555dc3fa, 0x3ffcf315, 0x5b5bab74 + .word 0x3ffd072d, 0x4a07897c, 0x3ffd1b53, 0x2b08c968 + .word 0x3ffd2f87, 0x080d89f2, 0x3ffd43c8, 0xeacaa1d6 + .word 0x3ffd5818, 0xdcfba487, 0x3ffd6c76, 0xe862e6d3 + .word 0x3ffd80e3, 0x16c98398, 0x3ffd955d, 0x71ff6075 + .word 0x3ffda9e6, 0x03db3285, 0x3ffdbe7c, 0xd63a8315 + .word 0x3ffdd321, 0xf301b460, 0x3ffde7d5, 0x641c0658 + .word 0x3ffdfc97, 0x337b9b5f, 0x3ffe1167, 0x6b197d17 + .word 0x3ffe2646, 0x14f5a129, 0x3ffe3b33, 0x3b16ee12 + .word 0x3ffe502e, 0xe78b3ff6, 0x3ffe6539, 0x24676d76 + .word 0x3ffe7a51, 0xfbc74c83, 0x3ffe8f79, 0x77cdb740 + .word 0x3ffea4af, 0xa2a490da, 0x3ffeb9f4, 0x867cca6e + .word 0x3ffecf48, 0x2d8e67f1, 0x3ffee4aa, 0xa2188510 + .word 0x3ffefa1b, 0xee615a27, 0x3fff0f9c, 0x1cb6412a + .word 0x3fff252b, 0x376bba97, 0x3fff3ac9, 0x48dd7274 + .word 0x3fff5076, 0x5b6e4540, 0x3fff6632, 0x798844f8 + .word 0x3fff7bfd, 0xad9cbe14, 0x3fff91d8, 0x02243c89 + .word 0x3fffa7c1, 0x819e90d8, 0x3fffbdba, 0x3692d514 + .word 0x3fffd3c2, 0x2b8f71f1, 0x3fffe9d9, 0x6b2a23d9 + +! __mt_constexp2fb: + .word 0x36900000, 0x36a00000, 0x36b00000, 0x36c00000 + .word 0x36d00000, 0x36e00000, 0x36f00000, 0x37000000 + .word 0x37100000, 0x37200000, 0x37300000, 0x37400000 + .word 0x37500000, 0x37600000, 0x37700000, 0x37800000 + .word 0x37900000, 0x37a00000, 0x37b00000, 0x37c00000 + .word 0x37d00000, 0x37e00000, 0x37f00000, 0x38000000 + .word 0x38100000, 0x38200000, 0x38300000, 0x38400000 + .word 0x38500000, 0x38600000, 0x38700000, 0x38800000 + .word 0x38900000, 0x38a00000, 0x38b00000, 0x38c00000 + .word 0x38d00000, 0x38e00000, 0x38f00000, 0x39000000 + .word 0x39100000, 0x39200000, 0x39300000, 0x39400000 + .word 0x39500000, 0x39600000, 0x39700000, 0x39800000 + .word 0x39900000, 0x39a00000, 0x39b00000, 0x39c00000 + .word 0x39d00000, 0x39e00000, 0x39f00000, 0x3a000000 + .word 0x3a100000, 0x3a200000, 0x3a300000, 0x3a400000 + .word 0x3a500000, 0x3a600000, 0x3a700000, 0x3a800000 + .word 0x3a900000, 0x3aa00000, 0x3ab00000, 0x3ac00000 + .word 0x3ad00000, 0x3ae00000, 0x3af00000, 0x3b000000 + .word 0x3b100000, 0x3b200000, 0x3b300000, 0x3b400000 + .word 0x3b500000, 0x3b600000, 0x3b700000, 0x3b800000 + .word 0x3b900000, 0x3ba00000, 0x3bb00000, 0x3bc00000 + .word 0x3bd00000, 0x3be00000, 0x3bf00000, 0x3c000000 + .word 0x3c100000, 0x3c200000, 0x3c300000, 0x3c400000 + .word 0x3c500000, 0x3c600000, 0x3c700000, 0x3c800000 + .word 0x3c900000, 0x3ca00000, 0x3cb00000, 0x3cc00000 + .word 0x3cd00000, 0x3ce00000, 0x3cf00000, 0x3d000000 + .word 0x3d100000, 0x3d200000, 0x3d300000, 0x3d400000 + .word 0x3d500000, 0x3d600000, 0x3d700000, 0x3d800000 + .word 0x3d900000, 0x3da00000, 0x3db00000, 0x3dc00000 + .word 0x3dd00000, 0x3de00000, 0x3df00000, 0x3e000000 + .word 0x3e100000, 0x3e200000, 0x3e300000, 0x3e400000 + .word 0x3e500000, 0x3e600000, 0x3e700000, 0x3e800000 + .word 0x3e900000, 0x3ea00000, 0x3eb00000, 0x3ec00000 + .word 0x3ed00000, 0x3ee00000, 0x3ef00000, 0x3f000000 + .word 0x3f100000, 0x3f200000, 0x3f300000, 0x3f400000 + .word 0x3f500000, 0x3f600000, 0x3f700000, 0x3f800000 + .word 0x3f900000, 0x3fa00000, 0x3fb00000, 0x3fc00000 + .word 0x3fd00000, 0x3fe00000, 0x3ff00000, 0x40000000 + .word 0x40100000, 0x40200000, 0x40300000, 0x40400000 + .word 0x40500000, 0x40600000, 0x40700000, 0x40800000 + .word 0x40900000, 0x40a00000, 0x40b00000, 0x40c00000 + .word 0x40d00000, 0x40e00000, 0x40f00000, 0x41000000 + .word 0x41100000, 0x41200000, 0x41300000, 0x41400000 + .word 0x41500000, 0x41600000, 0x41700000, 0x41800000 + .word 0x41900000, 0x41a00000, 0x41b00000, 0x41c00000 + .word 0x41d00000, 0x41e00000, 0x41f00000, 0x42000000 + .word 0x42100000, 0x42200000, 0x42300000, 0x42400000 + .word 0x42500000, 0x42600000, 0x42700000, 0x42800000 + .word 0x42900000, 0x42a00000, 0x42b00000, 0x42c00000 + .word 0x42d00000, 0x42e00000, 0x42f00000, 0x43000000 + .word 0x43100000, 0x43200000, 0x43300000, 0x43400000 + .word 0x43500000, 0x43600000, 0x43700000, 0x43800000 + .word 0x43900000, 0x43a00000, 0x43b00000, 0x43c00000 + .word 0x43d00000, 0x43e00000, 0x43f00000, 0x44000000 + .word 0x44100000, 0x44200000, 0x44300000, 0x44400000 + .word 0x44500000, 0x44600000, 0x44700000, 0x44800000 + .word 0x44900000, 0x44a00000, 0x44b00000, 0x44c00000 + .word 0x44d00000, 0x44e00000, 0x44f00000, 0x45000000 + .word 0x45100000, 0x45200000, 0x45300000, 0x45400000 + .word 0x45500000, 0x45600000, 0x45700000, 0x45800000 + .word 0x45900000, 0x45a00000, 0x45b00000, 0x45c00000 + .word 0x45d00000, 0x45e00000, 0x45f00000, 0x46000000 + .word 0x46100000, 0x46200000, 0x46300000, 0x46400000 + .word 0x46500000, 0x46600000, 0x46700000, 0x46800000 + .word 0x46900000, 0x46a00000, 0x46b00000, 0x46c00000 + .word 0x46d00000, 0x46e00000, 0x46f00000, 0x47000000 + .word 0x47100000, 0x47200000, 0x47300000, 0x47400000 + .word 0x47500000, 0x47600000, 0x47700000, 0x47800000 + .word 0x47900000, 0x47a00000, 0x47b00000, 0x47c00000 + .word 0x47d00000, 0x47e00000, 0x47f00000, 0x00000000 + + .word 0,0,0,0 + .word 0,0,0,0 + +.CONST_TBL: +! __mt_constlog4f: + .word 0x00000000, 0x00000000, 0x3e800000, 0x00000000 + .word 0x4006fe50, 0xb6ef0851, 0x3e7fc07f, 0x01fc07f0 + .word 0x4016e796, 0x85c2d22a, 0x3e7f81f8, 0x1f81f820 + .word 0x40211cd1, 0xd5133413, 0x3e7f4465, 0x9e4a4271 + .word 0x4026bad3, 0x758efd87, 0x3e7f07c1, 0xf07c1f08 + .word 0x402c4dfa, 0xb90aab5f, 0x3e7ecc07, 0xb301ecc0 + .word 0x4030eb38, 0x9fa29f9b, 0x3e7e9131, 0xabf0b767 + .word 0x4033aa2f, 0xdd27f1c3, 0x3e7e573a, 0xc901e574 + .word 0x403663f6, 0xfac91316, 0x3e7e1e1e, 0x1e1e1e1e + .word 0x403918a1, 0x6e46335b, 0x3e7de5d6, 0xe3f8868a + .word 0x403bc842, 0x40adabba, 0x3e7dae60, 0x76b981db + .word 0x403e72ec, 0x117fa5b2, 0x3e7d77b6, 0x54b82c34 + .word 0x40408c58, 0x8cda79e4, 0x3e7d41d4, 0x1d41d41d + .word 0x4041dcd1, 0x97552b7b, 0x3e7d0cb5, 0x8f6ec074 + .word 0x40432ae9, 0xe278ae1a, 0x3e7cd856, 0x89039b0b + .word 0x404476a9, 0xf983f74d, 0x3e7ca4b3, 0x055ee191 + .word 0x4045c01a, 0x39fbd688, 0x3e7c71c7, 0x1c71c71c + .word 0x40470742, 0xd4ef027f, 0x3e7c3f8f, 0x01c3f8f0 + .word 0x40484c2b, 0xd02f03b3, 0x3e7c0e07, 0x0381c0e0 + .word 0x40498edd, 0x077e70df, 0x3e7bdd2b, 0x899406f7 + .word 0x404acf5e, 0x2db4ec94, 0x3e7bacf9, 0x14c1bad0 + .word 0x404c0db6, 0xcdd94dee, 0x3e7b7d6c, 0x3dda338b + .word 0x404d49ee, 0x4c325970, 0x3e7b4e81, 0xb4e81b4f + .word 0x404e840b, 0xe74e6a4d, 0x3e7b2036, 0x406c80d9 + .word 0x404fbc16, 0xb902680a, 0x3e7af286, 0xbca1af28 + .word 0x4050790a, 0xdbb03009, 0x3e7ac570, 0x1ac5701b + .word 0x40511307, 0xdad30b76, 0x3e7a98ef, 0x606a63be + .word 0x4051ac05, 0xb291f070, 0x3e7a6d01, 0xa6d01a6d + .word 0x40524407, 0xab0e073a, 0x3e7a41a4, 0x1a41a41a + .word 0x4052db10, 0xfc4d9aaf, 0x3e7a16d3, 0xf97a4b02 + .word 0x40537124, 0xcea4cded, 0x3e79ec8e, 0x951033d9 + .word 0x40540646, 0x3b1b0449, 0x3e79c2d1, 0x4ee4a102 + .word 0x40549a78, 0x4bcd1b8b, 0x3e799999, 0x9999999a + .word 0x40552dbd, 0xfc4c96b3, 0x3e7970e4, 0xf80cb872 + .word 0x4055c01a, 0x39fbd688, 0x3e7948b0, 0xfcd6e9e0 + .word 0x4056518f, 0xe4677ba7, 0x3e7920fb, 0x49d0e229 + .word 0x4056e221, 0xcd9d0cde, 0x3e78f9c1, 0x8f9c18fa + .word 0x405771d2, 0xba7efb3c, 0x3e78d301, 0x8d3018d3 + .word 0x405800a5, 0x63161c54, 0x3e78acb9, 0x0f6bf3aa + .word 0x40588e9c, 0x72e0b226, 0x3e7886e5, 0xf0abb04a + .word 0x40591bba, 0x891f1709, 0x3e786186, 0x18618618 + .word 0x4059a802, 0x391e232f, 0x3e783c97, 0x7ab2bedd + .word 0x405a3376, 0x0a7f6051, 0x3e781818, 0x18181818 + .word 0x405abe18, 0x797f1f49, 0x3e77f405, 0xfd017f40 + .word 0x405b47eb, 0xf73882a1, 0x3e77d05f, 0x417d05f4 + .word 0x405bd0f2, 0xe9e79031, 0x3e77ad22, 0x08e0ecc3 + .word 0x405c592f, 0xad295b56, 0x3e778a4c, 0x8178a4c8 + .word 0x405ce0a4, 0x923a587d, 0x3e7767dc, 0xe434a9b1 + .word 0x405d6753, 0xe032ea0f, 0x3e7745d1, 0x745d1746 + .word 0x405ded3f, 0xd442364c, 0x3e772428, 0x7f46debc + .word 0x405e726a, 0xa1e754d2, 0x3e7702e0, 0x5c0b8170 + .word 0x405ef6d6, 0x7328e220, 0x3e76e1f7, 0x6b4337c7 + .word 0x405f7a85, 0x68cb06cf, 0x3e76c16c, 0x16c16c17 + .word 0x405ffd79, 0x9a83ff9b, 0x3e76a13c, 0xd1537290 + .word 0x40603fda, 0x8b97997f, 0x3e768168, 0x16816817 + .word 0x4060809c, 0xf27f703d, 0x3e7661ec, 0x6a5122f9 + .word 0x4060c105, 0x00d63aa6, 0x3e7642c8, 0x590b2164 + .word 0x40610113, 0xb153c8ea, 0x3e7623fa, 0x77016240 + .word 0x406140c9, 0xfaa1e544, 0x3e760581, 0x60581606 + .word 0x40618028, 0xcf72976a, 0x3e75e75b, 0xb8d015e7 + .word 0x4061bf31, 0x1e95d00e, 0x3e75c988, 0x2b931057 + .word 0x4061fde3, 0xd30e8126, 0x3e75ac05, 0x6b015ac0 + .word 0x40623c41, 0xd42727c8, 0x3e758ed2, 0x308158ed + .word 0x40627a4c, 0x0585cbf8, 0x3e7571ed, 0x3c506b3a + .word 0x4062b803, 0x473f7ad1, 0x3e755555, 0x55555555 + .word 0x4062f568, 0x75eb3f26, 0x3e753909, 0x48f40feb + .word 0x4063327c, 0x6ab49ca7, 0x3e751d07, 0xeae2f815 + .word 0x40636f3f, 0xfb6d9162, 0x3e750150, 0x15015015 + .word 0x4063abb3, 0xfaa02167, 0x3e74e5e0, 0xa72f0539 + .word 0x4063e7d9, 0x379f7016, 0x3e74cab8, 0x8725af6e + .word 0x406423b0, 0x7e986aa9, 0x3e74afd6, 0xa052bf5b + .word 0x40645f3a, 0x98a20739, 0x3e749539, 0xe3b2d067 + .word 0x40649a78, 0x4bcd1b8b, 0x3e747ae1, 0x47ae147b + .word 0x4064d56a, 0x5b33cec4, 0x3e7460cb, 0xc7f5cf9a + .word 0x40651011, 0x8708a8f9, 0x3e7446f8, 0x6562d9fb + .word 0x40654a6e, 0x8ca5438e, 0x3e742d66, 0x25d51f87 + .word 0x40658482, 0x26989d34, 0x3e741414, 0x14141414 + .word 0x4065be4d, 0x0cb51435, 0x3e73fb01, 0x3fb013fb + .word 0x4065f7cf, 0xf41e09af, 0x3e73e22c, 0xbce4a902 + .word 0x4066310b, 0x8f553048, 0x3e73c995, 0xa47babe7 + .word 0x40666a00, 0x8e4788cc, 0x3e73b13b, 0x13b13b14 + .word 0x4066a2af, 0x9e5a0f0a, 0x3e73991c, 0x2c187f63 + .word 0x4066db19, 0x6a76194a, 0x3e738138, 0x13813814 + .word 0x4067133e, 0x9b156c7c, 0x3e73698d, 0xf3de0748 + .word 0x40674b1f, 0xd64e0754, 0x3e73521c, 0xfb2b78c1 + .word 0x406782bd, 0xbfdda657, 0x3e733ae4, 0x5b57bcb2 + .word 0x4067ba18, 0xf93502e4, 0x3e7323e3, 0x4a2b10bf + .word 0x4067f132, 0x2182cf16, 0x3e730d19, 0x0130d190 + .word 0x40682809, 0xd5be7073, 0x3e72f684, 0xbda12f68 + .word 0x40685ea0, 0xb0b27b26, 0x3e72e025, 0xc04b8097 + .word 0x406894f7, 0x4b06ef8b, 0x3e72c9fb, 0x4d812ca0 + .word 0x4068cb0e, 0x3b4b3bbe, 0x3e72b404, 0xad012b40 + .word 0x406900e6, 0x160002cd, 0x3e729e41, 0x29e4129e + .word 0x4069367f, 0x6da0ab2f, 0x3e7288b0, 0x1288b013 + .word 0x40696bda, 0xd2acb5f6, 0x3e727350, 0xb8812735 + .word 0x4069a0f8, 0xd3b0e050, 0x3e725e22, 0x708092f1 + .word 0x4069d5d9, 0xfd5010b3, 0x3e724924, 0x92492492 + .word 0x406a0a7e, 0xda4c112d, 0x3e723456, 0x789abcdf + .word 0x406a3ee7, 0xf38e181f, 0x3e721fb7, 0x8121fb78 + .word 0x406a7315, 0xd02f20c8, 0x3e720b47, 0x0c67c0d9 + .word 0x406aa708, 0xf58014d3, 0x3e71f704, 0x7dc11f70 + .word 0x406adac1, 0xe711c833, 0x3e71e2ef, 0x3b3fb874 + .word 0x406b0e41, 0x26bcc86c, 0x3e71cf06, 0xada2811d + .word 0x406b4187, 0x34a9008c, 0x3e71bb4a, 0x4046ed29 + .word 0x406b7494, 0x8f5532da, 0x3e71a7b9, 0x611a7b96 + .word 0x406ba769, 0xb39e4964, 0x3e719453, 0x808ca29c + .word 0x406bda07, 0x1cc67e6e, 0x3e718118, 0x11811812 + .word 0x406c0c6d, 0x447c5dd3, 0x3e716e06, 0x89427379 + .word 0x406c3e9c, 0xa2e1a055, 0x3e715b1e, 0x5f75270d + .word 0x406c7095, 0xae91e1c7, 0x3e71485f, 0x0e0acd3b + .word 0x406ca258, 0xdca93316, 0x3e7135c8, 0x1135c811 + .word 0x406cd3e6, 0xa0ca8907, 0x3e712358, 0xe75d3033 + .word 0x406d053f, 0x6d260896, 0x3e711111, 0x11111111 + .word 0x406d3663, 0xb27f31d5, 0x3e70fef0, 0x10fef011 + .word 0x406d6753, 0xe032ea0f, 0x3e70ecf5, 0x6be69c90 + .word 0x406d9810, 0x643d6615, 0x3e70db20, 0xa88f4696 + .word 0x406dc899, 0xab3ff56c, 0x3e70c971, 0x4fbcda3b + .word 0x406df8f0, 0x2086af2c, 0x3e70b7e6, 0xec259dc8 + .word 0x406e2914, 0x2e0e0140, 0x3e70a681, 0x0a6810a7 + .word 0x406e5906, 0x3c8822ce, 0x3e70953f, 0x39010954 + .word 0x406e88c6, 0xb3626a73, 0x3e708421, 0x08421084 + .word 0x406eb855, 0xf8ca88fb, 0x3e707326, 0x0a47f7c6 + .word 0x406ee7b4, 0x71b3a950, 0x3e70624d, 0xd2f1a9fc + .word 0x406f16e2, 0x81db7630, 0x3e705197, 0xf7d73404 + .word 0x406f45e0, 0x8bcf0655, 0x3e704104, 0x10410410 + .word 0x406f74ae, 0xf0efafae, 0x3e703091, 0xb51f5e1a + .word 0x406fa34e, 0x1177c233, 0x3e702040, 0x81020408 + .word 0x406fd1be, 0x4c7f2af9, 0x3e701010, 0x10101010 + .word 0x40700000, 0x00000000, 0x3e700000, 0x00000000 + +! __mt_constexp2f: + .word 0x3ff00000, 0x00000000, 0x3ff00b1a, 0xfa5abcbf + .word 0x3ff0163d, 0xa9fb3335, 0x3ff02168, 0x143b0281 + .word 0x3ff02c9a, 0x3e778061, 0x3ff037d4, 0x2e11bbcc + .word 0x3ff04315, 0xe86e7f85, 0x3ff04e5f, 0x72f654b1 + .word 0x3ff059b0, 0xd3158574, 0x3ff0650a, 0x0e3c1f89 + .word 0x3ff0706b, 0x29ddf6de, 0x3ff07bd4, 0x2b72a836 + .word 0x3ff08745, 0x18759bc8, 0x3ff092bd, 0xf66607e0 + .word 0x3ff09e3e, 0xcac6f383, 0x3ff0a9c7, 0x9b1f3919 + .word 0x3fefb558, 0x6cf9890f, 0x3fefc0f1, 0x45e46c85 + .word 0x3fefcc92, 0x2b7247f7, 0x3fefd83b, 0x23395dec + .word 0x3fefe3ec, 0x32d3d1a2, 0x3fefefa5, 0x5fdfa9c5 + .word 0x3feffb66, 0xaffed31b, 0x3ff00730, 0x28d7233e + .word 0x3ff01301, 0xd0125b51, 0x3ff01edb, 0xab5e2ab6 + .word 0x3ff02abd, 0xc06c31cc, 0x3ff036a8, 0x14f204ab + .word 0x3ff0429a, 0xaea92de0, 0x3ff04e95, 0x934f312e + .word 0x3ff05a98, 0xc8a58e51, 0x3ff066a4, 0x5471c3c2 + .word 0x3fef72b8, 0x3c7d517b, 0x3fef7ed4, 0x8695bbc0 + .word 0x3fef8af9, 0x388c8dea, 0x3fef9726, 0x58375d2f + .word 0x3fefa35b, 0xeb6fcb75, 0x3fefaf99, 0xf8138a1c + .word 0x3fefbbe0, 0x84045cd4, 0x3fefc82f, 0x95281c6b + .word 0x3fefd487, 0x3168b9aa, 0x3fefe0e7, 0x5eb44027 + .word 0x3fefed50, 0x22fcd91d, 0x3feff9c1, 0x8438ce4d + .word 0x3ff0063b, 0x88628cd6, 0x3ff012be, 0x3578a819 + .word 0x3ff01f49, 0x917ddc96, 0x3ff02bdd, 0xa27912d1 + .word 0x3fef387a, 0x6e756238, 0x3fef451f, 0xfb82140a + .word 0x3fef51ce, 0x4fb2a63f, 0x3fef5e85, 0x711ece75 + .word 0x3fef6b45, 0x65e27cdd, 0x3fef780e, 0x341ddf29 + .word 0x3fef84df, 0xe1f56381, 0x3fef91ba, 0x7591bb70 + .word 0x3fef9e9d, 0xf51fdee1, 0x3fefab8a, 0x66d10f13 + .word 0x3fefb87f, 0xd0dad990, 0x3fefc57e, 0x39771b2f + .word 0x3fefd285, 0xa6e4030b, 0x3fefdf96, 0x1f641589 + .word 0x3fefecaf, 0xa93e2f56, 0x3feff9d2, 0x4abd886b + .word 0x3fef06fe, 0x0a31b715, 0x3fef1432, 0xedeeb2fd + .word 0x3fef2170, 0xfc4cd831, 0x3fef2eb8, 0x3ba8ea32 + .word 0x3fef3c08, 0xb26416ff, 0x3fef4962, 0x66e3fa2d + .word 0x3fef56c5, 0x5f929ff1, 0x3fef6431, 0xa2de883b + .word 0x3fef71a7, 0x373aa9cb, 0x3fef7f26, 0x231e754a + .word 0x3fef8cae, 0x6d05d866, 0x3fef9a40, 0x1b7140ef + .word 0x3fefa7db, 0x34e59ff7, 0x3fefb57f, 0xbfec6cf4 + .word 0x3fefc32d, 0xc313a8e5, 0x3fefd0e5, 0x44ede173 + .word 0x3feedea6, 0x4c123422, 0x3feeec70, 0xdf1c5175 + .word 0x3feefa45, 0x04ac801c, 0x3fef0822, 0xc367a024 + .word 0x3fef160a, 0x21f72e2a, 0x3fef23fb, 0x2709468a + .word 0x3fef31f5, 0xd950a897, 0x3fef3ffa, 0x3f84b9d4 + .word 0x3fef4e08, 0x6061892d, 0x3fef5c20, 0x42a7d232 + .word 0x3fef6a41, 0xed1d0057, 0x3fef786d, 0x668b3237 + .word 0x3fef86a2, 0xb5c13cd0, 0x3fef94e1, 0xe192aed2 + .word 0x3fefa32a, 0xf0d7d3de, 0x3fefb17d, 0xea6db7d7 + .word 0x3feebfda, 0xd5362a27, 0x3feece41, 0xb817c114 + .word 0x3feedcb2, 0x99fddd0d, 0x3feeeb2d, 0x81d8abff + .word 0x3feef9b2, 0x769d2ca7, 0x3fef0841, 0x7f4531ee + .word 0x3fef16da, 0xa2cf6642, 0x3fef257d, 0xe83f4eef + .word 0x3fef342b, 0x569d4f82, 0x3fef42e2, 0xf4f6ad27 + .word 0x3fef51a4, 0xca5d920f, 0x3fef6070, 0xdde910d2 + .word 0x3fef6f47, 0x36b527da, 0x3fef7e27, 0xdbe2c4cf + .word 0x3fef8d12, 0xd497c7fd, 0x3fef9c08, 0x27ff07cc + .word 0x3feeab07, 0xdd485429, 0x3feeba11, 0xfba87a03 + .word 0x3feec926, 0x8a5946b7, 0x3feed845, 0x90998b93 + .word 0x3feee76f, 0x15ad2148, 0x3feef6a3, 0x20dceb71 + .word 0x3fef05e1, 0xb976dc09, 0x3fef152a, 0xe6cdf6f4 + .word 0x3fef247e, 0xb03a5585, 0x3fef33dd, 0x1d1929fd + .word 0x3fef4346, 0x34ccc320, 0x3fef52b9, 0xfebc8fb7 + .word 0x3fef6238, 0x82552225, 0x3fef71c1, 0xc70833f6 + .word 0x3fef8155, 0xd44ca973, 0x3fef90f4, 0xb19e9538 + .word 0x3feea09e, 0x667f3bcd, 0x3feeb052, 0xfa75173e + .word 0x3feec012, 0x750bdabf, 0x3feecfdc, 0xddd47645 + .word 0x3feedfb2, 0x3c651a2f, 0x3feeef92, 0x98593ae5 + .word 0x3feeff7d, 0xf9519484, 0x3fef0f74, 0x66f42e87 + .word 0x3fef1f75, 0xe8ec5f74, 0x3fef2f82, 0x86ead08a + .word 0x3fef3f9a, 0x48a58174, 0x3fef4fbd, 0x35d7cbfd + .word 0x3fef5feb, 0x564267c9, 0x3fef7024, 0xb1ab6e09 + .word 0x3fef8069, 0x4fde5d3f, 0x3fef90b9, 0x38ac1cf6 + .word 0x3feea114, 0x73eb0187, 0x3feeb17b, 0x0976cfdb + .word 0x3feec1ed, 0x0130c132, 0x3feed26a, 0x62ff86f0 + .word 0x3feee2f3, 0x36cf4e62, 0x3feef387, 0x8491c491 + .word 0x3fef0427, 0x543e1a12, 0x3fef14d2, 0xadd106d9 + .word 0x3fef2589, 0x994cce13, 0x3fef364c, 0x1eb941f7 + .word 0x3fef471a, 0x4623c7ad, 0x3fef57f4, 0x179f5b21 + .word 0x3fef68d9, 0x9b4492ed, 0x3fef79ca, 0xd931a436 + .word 0x3fef8ac7, 0xd98a6699, 0x3fef9bd0, 0xa478580f + .word 0x3feeace5, 0x422aa0db, 0x3feebe05, 0xbad61778 + .word 0x3feecf32, 0x16b5448c, 0x3feee06a, 0x5e0866d9 + .word 0x3feef1ae, 0x99157736, 0x3fef02fe, 0xd0282c8a + .word 0x3fef145b, 0x0b91ffc6, 0x3fef25c3, 0x53aa2fe2 + .word 0x3fef3737, 0xb0cdc5e5, 0x3fef48b8, 0x2b5f98e5 + .word 0x3fef5a44, 0xcbc8520f, 0x3fef6bdd, 0x9a7670b3 + .word 0x3fef7d82, 0x9fde4e50, 0x3fef8f33, 0xe47a22a2 + .word 0x3fefa0f1, 0x70ca07ba, 0x3fefb2bb, 0x4d53fe0d + .word 0x3feec491, 0x82a3f090, 0x3feed674, 0x194bb8d5 + .word 0x3feee863, 0x19e32323, 0x3feefa5e, 0x8d07f29e + .word 0x3fef0c66, 0x7b5de565, 0x3fef1e7a, 0xed8eb8bb + .word 0x3fef309b, 0xec4a2d33, 0x3fef42c9, 0x80460ad8 + .word 0x3fef5503, 0xb23e255d, 0x3fef674a, 0x8af46052 + .word 0x3fef799e, 0x1330b358, 0x3fef8bfe, 0x53c12e59 + .word 0x3fef9e6b, 0x5579fdbf, 0x3fefb0e5, 0x21356eba + .word 0x3fefc36b, 0xbfd3f37a, 0x3fefd5ff, 0x3a3c2774 + .word 0x3feee89f, 0x995ad3ad, 0x3feefb4c, 0xe622f2ff + .word 0x3fef0e07, 0x298db666, 0x3fef20ce, 0x6c9a8952 + .word 0x3fef33a2, 0xb84f15fb, 0x3fef4684, 0x15b749b1 + .word 0x3fef5972, 0x8de5593a, 0x3fef6c6e, 0x29f1c52a + .word 0x3fef7f76, 0xf2fb5e47, 0x3fef928c, 0xf22749e4 + .word 0x3fefa5b0, 0x30a1064a, 0x3fefb8e0, 0xb79a6f1f + .word 0x3fefcc1e, 0x904bc1d2, 0x3fefdf69, 0xc3f3a207 + .word 0x3feff2c2, 0x5bd71e09, 0x3ff00628, 0x6141b33d + .word 0x3fef199b, 0xdd85529c, 0x3fef2d1c, 0xd9fa652c + .word 0x3fef40ab, 0x5fffd07a, 0x3fef5447, 0x78fafb22 + .word 0x3fef67f1, 0x2e57d14b, 0x3fef7ba8, 0x8988c933 + .word 0x3fef8f6d, 0x9406e7b5, 0x3fefa340, 0x5751c4db + .word 0x3fefb720, 0xdcef9069, 0x3fefcb0f, 0x2e6d1675 + .word 0x3fefdf0b, 0x555dc3fa, 0x3feff315, 0x5b5bab74 + .word 0x3ff0072d, 0x4a07897c, 0x3ff01b53, 0x2b08c968 + .word 0x3ff02f87, 0x080d89f2, 0x3ff043c8, 0xeacaa1d6 + .word 0x3fef5818, 0xdcfba487, 0x3fef6c76, 0xe862e6d3 + .word 0x3fef80e3, 0x16c98398, 0x3fef955d, 0x71ff6075 + .word 0x3fefa9e6, 0x03db3285, 0x3fefbe7c, 0xd63a8315 + .word 0x3fefd321, 0xf301b460, 0x3fefe7d5, 0x641c0658 + .word 0x3feffc97, 0x337b9b5f, 0x3ff01167, 0x6b197d17 + .word 0x3ff02646, 0x14f5a129, 0x3ff03b33, 0x3b16ee12 + .word 0x3ff0502e, 0xe78b3ff6, 0x3ff06539, 0x24676d76 + .word 0x3ff07a51, 0xfbc74c83, 0x3ff08f79, 0x77cdb740 + .word 0x3fefa4af, 0xa2a490da, 0x3fefb9f4, 0x867cca6e + .word 0x3fefcf48, 0x2d8e67f1, 0x3fefe4aa, 0xa2188510 + .word 0x3feffa1b, 0xee615a27, 0x3ff00f9c, 0x1cb6412a + .word 0x3ff0252b, 0x376bba97, 0x3ff03ac9, 0x48dd7274 + .word 0x3ff05076, 0x5b6e4540, 0x3ff06632, 0x798844f8 + .word 0x3ff07bfd, 0xad9cbe14, 0x3ff091d8, 0x02243c89 + .word 0x3ff0a7c1, 0x819e90d8, 0x3ff0bdba, 0x3692d514 + .word 0x3ff0d3c2, 0x2b8f71f1, 0x3ff0e9d9, 0x6b2a23d9 + + .word 0xc057150d, 0x5f6e1c54 ! KA3 = -3.60659926599003171364e-01*256.0 + .word 0x405ec71c, 0x2e92efda ! KA2 = 4.80902715189356683026e-01*256.0 + .word 0xc0671547, 0x653cbec4 ! KA1 = -7.21347520569871841065e-01*256.0 + .word 0x40771547, 0x652af190 ! KA0 = 1.44269504088069658645e+00*256.0 + .word 0x3ecebfbe, 0x9d182250 ! KB2 = 3.66556671660783833261e-06 + .word 0x3f662e43, 0xe2528362 ! KB1 = 2.70760782821392980564e-03 + .word 0x40e00000, 0x00000000 ! HTHRESH = 32768.0 + .word 0xc0e2c000, 0x00000000 ! LTHRESH = -38400.0 ; 0.0f + .word 0x3f800000, 0x00000000 ! 1.0f ; free + +#define tmp_px STACK_BIAS-48 +#define tmp_py STACK_BIAS-40 +#define tmp_counter STACK_BIAS-32 +#define tmp0 STACK_BIAS-28 +#define tmp1 STACK_BIAS-24 +#define tmp2 STACK_BIAS-20 +#define tmp3 STACK_BIAS-16 +#define tmp4 STACK_BIAS-12 +#define tmp5 STACK_BIAS-8 +#define tmp6 STACK_BIAS-4 + + +#define KA3 %f34 +#define KA2 %f36 +#define KA1 %f38 +#define KA0 %f40 +#define KB2 %f42 +#define KB1 %f44 +#define HTHRESHOLD %f30 +#define LTHRESHOLD %f32 + +#define counter %o7 +#define stridex %i0 +#define stridey %i4 +#define stridez %l3 + +#define CONST_0x8000 %l1 +#define MASK_0x007fffff %l4 +#define MASK_0x7fffffff %l5 + +! sizeof temp storage - must be a multiple of 16 for V9 +#define tmps 0x30 + +!-------------------------------------------------------------------- +! !!!!! vpowf algorithm !!!!! +! uy = *(unsigned int*)py; +! ux = *(unsigned int*)px; +! ay = uy & 0x7fffffff; +! ax0 = ux & 0x7fffffff; +! sx = ux >> 31; +! yisint0 = 0; /* Y - non-integer */ +! if (ax0 >= 0x7f800000 || ay >= 0x7f800000) { /* |X| or |Y| = Inf,Nan */ +! if (ax0 > 0x7f800000 || ay > 0x7f800000) /* |X| or |Y| = Nan */ +! pz[0] = *px * *py; +! goto next; +! if (ay == 0x7f800000) { /* |Y| = Inf */ +! float fy; +! if (ax0 == 0x3f800000) fy = *py - *py; /* +-1 ** +-Inf = NaN */ +! else fy = ((ax0 < 0x3f800000) != (uy >> 31)) ? ZERO : *(float*) &ay; +! pz[0] = fy; +! goto next; +! } +! if (sx) { /* X = -Inf */ +! exp = ay >> 23; +! if (exp >= 0x97) /* |Y| >= 2^24 */ +! yisint0 = 2; /* Y - even */ +! else { +! if (exp >= 0x7f) { /* |Y| >= 1 */ +! i0 = ay >> ((0x7f + 23) - exp); +! if ((i0 << ((0x7f + 23) - exp)) == ay) yisint0 = 2 - (i0 & 1); +! } +! } +! } +! if (uy >> 31) ax0 = 0; +! ax0 += yisint0 << 31; +! pz[0] = *(float*)&ax0; +! goto next; +! } +! exp0 = (ax0 >> 23) - 127; +! if ((int)ux < 0x00800000) { /* X = denormal or negative */ +! if ((int)ax0 < 0x00800000) { /* X = denormal */ +! *((float*) &ax0) = (float) (int)ax0; +! exp0 = (ax0 >> 23) - (127 + 149); +! } +! if ((int)ux <= 0) { /* X <= 0 */ +! exp = ay >> 23; +! if (exp >= 0x97) /* |Y| >= 2^24 */ +! yisint0 = 2; /* Y - even */ +! else { +! if (exp >= 0x7f) { /* |Y| >= 1 */ +! i0 = ay >> ((0x7f + 23) - exp); +! if ((i0 << ((0x7f + 23) - exp)) == ay) yisint0 = 2 - (i0 & 1); +! } +! } +! if (ax0 == 0) { /* pow(0,Y) */ +! float fy; +! fy = (uy >> 31) ? ONE / ZERO : ZERO; +! if (sx & yisint0) fy = -fy; +! pz[0] = fy; +! goto next; +! } +! if (yisint0 == 0) { /* pow(neg,non-integer) */ +! pz[0] = ZERO / ZERO; /* NaN */ +! goto next; +! } +! } +! } +! +! ax0 = *px; +! exp0 = ax0 & 0x7fffffff; +! exp0 >>= 23; +! exp0 -= 127; +! exp0 <<= 8; +! ax0 &= 0x007fffff; +! i0 = ax0 + 0x8000; +! i0 &= 0xffff0000; +! ind0 = i0 >> 12; +! ind0 &= -8; +! i0 = ax0 - i0; +! dtmp0 = (double) i0; +! dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); +! y0 = dtmp0 * dtmp1; +! dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); +! dtmp1 = (double) exp0; +! yy0 = dtmp0 + dtmp1; +! dtmp0 = KA3 * y0; +! dtmp0 += KA2; +! dtmp0 *= y0; +! dtmp0 += KA1; +! dtmp0 *= y0; +! dtmp0 += KA0; +! dtmp0 *= y0; +! yy0 += dtmp0; +! ftmp0 = *py0; +! dtmp0 = (double)ftmp0; +! yy0 *= dtmp0; +! if (yy0 >= HTHRESH) +! yy0 = HTHRESH; +! if (yy0 <= LTHRESH) +! yy0 = LTHRESH; +! ind0 = (int) yy0; +! ((int*)&dtmp1)[0] = ind0; +! ((int*)&dtmp1)[1] = 0; +! dtmp1 = vis_fpackfix(dtmp1); +! dtmp0 = (double)ind0; +! y0 = yy0 - dtmp0; +! dtmp0 = KB2 * y0; +! dtmp0 += KB1; +! yy0 = dtmp0 * y0; +! ind0 &= 255; +! ind0 <<= 3; +! di0 = *(double*)((char*)__mt_constexp2f + ind0); +! di0 = vis_fpadd32(di0,dtmp1); +! yy0 *= di0; +! yy0 += di0; +! ftmp0 = (float)yy0; +! *pz0 = ftmp0; +!-------------------------------------------------------------------- +! !!!!! vpowf algorithm,stridex=0 !!!!! +! +! ax = ax0 = *px; +! exp0 = ax0 & 0x7fffffff; +! exp0 >>= 23; +! exp0 -= 127; +! exp0 <<= 8; +! ax0 &= 0x007fffff; +! i0 = ax0 + 0x8000; +! i0 &= 0xffff0000; +! ind0 = i0 >> 12; +! ind0 &= -8; +! i0 = ax0 - i0; +! dtmp0 = (double) i0; +! dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); +! y0 = dtmp0 * dtmp1; +! dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); +! dtmp1 = (double) exp0; +! yy0 = dtmp0 + dtmp1; +! dtmp0 = KA3 * y0; +! dtmp0 += KA2; +! dtmp0 *= y0; +! dtmp0 += KA1; +! dtmp0 *= y0; +! dtmp0 += KA0; +! dtmp0 *= y0; +! yy = yy0 + dtmp0; +! +! uy = ((int*)py)[0]; +! ay = uy & 0x7fffffff; +! if (ay >= 0x7f800000) { /* |Y| = Inf or Nan */ +! float fy; +! if (ay > 0x7f800000) fy = *py + *py; /* |Y| = Nan */ +! else fy = ((ax < 0x3f800000) != (uy >> 31)) ? ZERO : *(float*)&ay; +! pz[0] = fy; +! goto next; +! } +! +! +! ftmp0 = py[0]; +! dtmp0 = (double)ftmp0; +! yy0 = dtmp0 * yy; +! if (yy0 >= HTHRESH) +! if (yy0 <= LTHRESH) +! yy0 = HTHRESH; +! yy0 = LTHRESH; +! ii0 = (int) yy0; +! dtmp0 = (double)ii0; +! i0 = ii0 >> 5; +! i0 &= -8; +! di0 = ((double*)((char*)(__mt_constexp2fb + 150) + i0))[0]; +! y0 = yy0 - dtmp0; +! dtmp0 = KB2 * y0; +! dtmp0 += KB1; +! yy0 = dtmp0 * y0; +! ii0 &= 255; +! ii0 <<= 3; +! dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; +! di0 *= dtmp0; +! dtmp0 = yy0 * di0; +! dtmp0 += di0; +! ftmp0 = (float)dtmp0; +! pz[0] = ftmp0; +!-------------------------------------------------------------------- + ENTRY(__vpowf) + save %sp,-SA(MINFRAME)-tmps,%sp + PIC_SETUP(l7) + PIC_SET(l7,.CONST_TBL,l2) + wr %g0,0x60,%gsr + +#ifdef __sparcv9 + ldx [%fp+STACK_BIAS+176],stridez +#else + ld [%fp+STACK_BIAS+92],stridez +#endif + + ld [%i1],%o3 + add %l2,2064,%l0 + st %i0,[%fp+tmp_counter] + add %l0,2048,%l6 + ldd [%l6],KA3 + ldd [%l6+8],KA2 + sll stridey,2,stridey + ldd [%l6+16],KA1 + sll stridez,2,stridez + ldd [%l6+24],KA0 + sll %i2,2,stridex + ldd [%l6+32],KB2 + sethi %hi(0x7ffffc00),MASK_0x7fffffff + fzero %f2 + ldd [%l6+40],KB1 + add MASK_0x7fffffff,1023,MASK_0x7fffffff + fzero %f10 + ldd [%l6+48],HTHRESHOLD + sethi %hi(0x7ffc00),MASK_0x007fffff + fzero %f20 + ldd [%l6+56],LTHRESHOLD + sethi %hi(0x8000),CONST_0x8000 + add MASK_0x007fffff,1023,MASK_0x007fffff + + cmp stridex,0 + bne,pt %icc,.common_case + sethi %hi(0x00800000),%l6 + + cmp %o3,%l6 + bl,pn %icc,.common_case + sethi %hi(0x7f800000),%o1 + + cmp %o3,%o1 + bge,pn %icc,.common_case + sethi %hi(0x3f800000),%l6 + + cmp %o3,%l6 + bne,pt %icc,.stridex_zero + nop + +.common_case: + stx %i1,[%fp+tmp_px] + stx %i3,[%fp+tmp_py] +.begin: + ld [%fp+tmp_counter],counter + ldx [%fp+tmp_px],%o2 + ldx [%fp+tmp_py],%i2 + st %g0,[%fp+tmp_counter] +.begin1: + cmp counter,0 + ble,pn %icc,.exit + lda [%o2]0x82,%i1 ! (Y0_2) ax0 = *px; + + lda [%i2]0x82,%l7 + sethi %hi(0xffff0000),%l6 + sethi %hi(0x7f800000),%o5 + + and %i1,MASK_0x7fffffff,%i3 ! (Y0_2) exp0 = ax0 & 0x7fffffff; + and %i1,MASK_0x007fffff,%g5 ! (Y0_2) ax0 &= 0x007fffff; + + cmp %i3,%o5 ! (Y0_2) ax0 ? 0x7f800000 + bge,pn %icc,.spec1 ! (Y0_2) if( ax0 >= 0x7f800000 ) + and %l7,MASK_0x7fffffff,%o4 + + cmp %o4,%o5 ! (Y0_2) ay0 ? 0x7f800000 + bge,pn %icc,.spec1 ! (Y0_2) if( ay0 >= 0x7f800000 ) + nop + + cmp %i1,MASK_0x007fffff ! (Y0_2) ux0 ? 0x800000 + ble,pn %icc,.spec2 ! (Y0_2) if(ux0 < 0x800000) + srl %i3,23,%o3 ! (Y0_2) exp0 >>= 23; + + sub %o3,127,%o3 ! (Y0_2) exp0 -= 127; + + add %g5,CONST_0x8000,%i3 ! (Y0_2) i0 = ax0 + 0x8000; + + sll %o3,8,%o4 ! (Y0_2) exp0 <<= 8; + and %i3,%l6,%i3 ! (Y0_2) i0 &= 0xffff0000; + st %o4,[%fp+tmp3] ! (Y0_2) STORE exp0 + + sub %g5,%i3,%o4 ! (Y0_2) i0 = ax0 - i0; + st %o4,[%fp+tmp2] ! (Y0_2) STORE i0 + add %o2,stridex,%o2 ! px += stridex + + sra %i3,12,%o0 ! (Y0_2) ind0 = i0 >> 12; + lda [%o2]0x82,%o3 ! (Y1_2) ax0 = *px; + + and %o0,-8,%g5 ! (Y0_2) ind0 &= -8; + ld [%fp+tmp2],%f14 ! (Y0_2) dtmp0 = (double) i0; + + and %o3,MASK_0x7fffffff,%i3 ! (Y1_2) exp0 = ax0 & 0x7fffffff; + and %o3,MASK_0x007fffff,%o0 ! (Y1_2) ax0 &= 0x007fffff; + + cmp %i3,%o5 ! (Y1_2) ax0 ? 0x7f800000 + add %l2,%g5,%g1 ! (Y0_2) (char*)__mt_constlog4f + ind0 + + srl %i3,23,%i3 ! (Y1_2) exp0 >>= 23; + add %o0,CONST_0x8000,%i1 ! (Y1_2) i0 = ax0 + 0x8000; + + ldd [%g1+8],%f48 ! (Y0_2) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + sub %i3,127,%i3 ! (Y1_2) exp0 -= 127; + fitod %f14,%f60 ! (Y0_2) dtmp0 = (double) i0; + + sll %i3,8,%i3 ! (Y1_2) exp0 <<= 8; + and %i1,%l6,%i1 ! (Y1_2) i0 &= 0xffff0000; + st %i3,[%fp+tmp4] ! (Y1_2) STORE exp0 + + sub %o0,%i1,%o0 ! (Y1_2) i0 = ax0 - i0; + st %o0,[%fp+tmp5] ! (Y1_2) STORE i0 + bge,pn %icc,.update0 ! (Y1_2) if(ax0 >= 0x7f800000) + nop +.cont0: + cmp %o3,MASK_0x007fffff ! (Y1_2) ux0 ? 0x800000 + + fmuld %f60,%f48,%f48 ! (Y0_2) y0 = dtmp0 * dtmp1; + ble,pn %icc,.update1 ! (Y1_2) if(ux0 < 0x800000) + nop +.cont1: + fmuld KA3,%f48,%f62 ! (Y0_2) dtmp0 = KA3 * y0; + + faddd %f62,KA2,%f22 ! (Y0_2) dtmp0 += KA2; + + sra %i1,12,%o1 ! (Y1_2) ind0 = i0 >> 12; + add %o2,stridex,%i3 ! px += stridex + lda [stridex+%o2]0x82,%g1 ! (Y2_2) ax0 = *px; + + and %o1,-8,%o0 ! (Y1_2) ind0 &= -8; + ld [%fp+tmp5],%f12 ! (Y1_2) LOAD i0 + + and %g1,MASK_0x7fffffff,%i1 ! (Y2_2) exp0 = ax0 & 0x7fffffff; + and %g1,MASK_0x007fffff,%o2 ! (Y2_2) ax0 &= 0x007fffff; + lda [%i2]0x82,%f0 ! (Y0_2) ftmp0 = *py0; + + srl %i1,23,%o3 ! (Y2_2) exp0 >>= 23; + cmp %i1,%o5 ! (Y2_2) ax0 ? 0x7f800000 + + fmuld %f22,%f48,%f26 ! (Y0_2) dtmp0 *= y0; + add %l2,%o0,%i1 ! (Y1_2) (char*)__mt_constlog4f + ind0 + sub %o3,127,%l7 ! (Y2_2) exp0 -= 127; + + add %o2,CONST_0x8000,%o1 ! (Y2_2) i0 = ax0 + 0x8000; + ldd [%i1+8],%f50 ! (Y1_2) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + fitod %f12,%f28 ! (Y1_2) dtmp0 = (double) i0; + + sll %l7,8,%l7 ! (Y2_2) exp0 <<= 8; + and %o1,%l6,%o1 ! (Y2_2) i0 &= 0xffff0000; + st %l7,[%fp+tmp6] ! (Y2_2) STORE exp0 + + sub %o2,%o1,%i1 ! (Y2_2) i0 = ax0 - i0; + st %i1,[%fp+tmp2] ! (Y2_2) STORE i0 + bge,pn %icc,.update2 ! (Y2_2) if(ax0 >= 0x7f800000) + nop +.cont2: + cmp %g1,MASK_0x007fffff ! (Y2_2) ux0 ? 0x800000 + + fmuld %f28,%f50,%f46 ! (Y1_2) y0 = dtmp0 * dtmp1; + ble,pn %icc,.update3 ! (Y2_2) if(ux0 < 0x800000) + faddd %f26,KA1,%f50 ! (Y0_2) dtmp0 += KA1; +.cont3: + ld [%fp+tmp3],%f4 ! (Y0_2) dtmp1 = (double) exp0; + + fstod %f0,%f24 ! (Y0_2) dtmp0 = (double)ftmp0; + + fmuld KA3,%f46,%f28 ! (Y1_1) dtmp0 = KA3 * y0; + + fitod %f4,%f26 ! (Y0_1) dtmp1 = (double) exp0; + + fmuld %f50,%f48,%f50 ! (Y0_1) dtmp0 *= y0; + + faddd %f28,KA2,%f28 ! (Y1_1) dtmp0 += KA2; + + ldd [%l2+%g5],%f60 ! (Y0_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + add %i3,stridex,%o2 ! px += stridex + + lda [%o2]0x82,%i1 ! (Y0_2) ax0 = *px; + sra %o1,12,%g5 ! (Y2_1) ind0 = i0 >> 12; + + faddd %f50,KA0,%f58 ! (Y0_1) dtmp0 += KA0; + and %g5,-8,%o1 ! (Y2_1) ind0 &= -8; + ld [%fp+tmp2],%f6 ! (Y2_1) dtmp0 = (double) i0; + + and %i1,MASK_0x7fffffff,%i3 ! (Y0_2) exp0 = ax0 & 0x7fffffff; + and %i1,MASK_0x007fffff,%g5 ! (Y0_2) ax0 &= 0x007fffff; + + srl %i3,23,%o3 ! (Y0_2) exp0 >>= 23; + add %l2,%o1,%g1 ! (Y2_1) (char*)__mt_constlog4f + ind0 + faddd %f60,%f26,%f26 ! (Y0_1) yy0 = dtmp0 + dtmp1; + + fmuld %f28,%f46,%f50 ! (Y1_1) dtmp0 *= y0; + sub %o3,127,%o3 ! (Y0_2) exp0 -= 127; + cmp %i3,%o5 ! (Y0_2) ax0 ? 0x7f800000 + + fmuld %f58,%f48,%f48 ! (Y0_1) dtmp0 *= y0; + add %g5,CONST_0x8000,%i3 ! (Y0_2) i0 = ax0 + 0x8000; + ldd [%g1+8],%f58 ! (Y2_1) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + fitod %f6,%f54 ! (Y2_1) dtmp0 = (double) i0; + + sll %o3,8,%o4 ! (Y0_2) exp0 <<= 8; + and %i3,%l6,%i3 ! (Y0_2) i0 &= 0xffff0000; + st %o4,[%fp+tmp3] ! (Y0_2) STORE exp0 + + sub %g5,%i3,%o4 ! (Y0_2) i0 = ax0 - i0; + st %o4,[%fp+tmp2] ! (Y0_2) STORE i0 + bge,pn %icc,.update4 ! (Y0_2) if( ax0 >= 0x7f800000 ) + nop +.cont4: + lda [stridey+%i2]0x82,%g1 ! (Y1_1) ay0 = *(unsigned*)py0 + add %i2,stridey,%o4 ! py += stridey + cmp %i1,MASK_0x007fffff ! (Y0_2) ux0 ? 0x800000 + + fmuld %f54,%f58,%f28 ! (Y2_1) y0 = dtmp0 * dtmp1; + lda [stridey+%i2]0x82,%f2 ! (Y1_1) ftmp0 = *py0; + ble,pn %icc,.update5 ! (Y0_2) if(ux0 < 0x800000) + faddd %f50,KA1,%f54 ! (Y1_1) dtmp0 += KA1; +.cont5: + and %g1,MASK_0x7fffffff,%g1 ! (Y1_1) ay0 &= 0x7fffffff; + ld [%fp+tmp4],%f1 ! (Y1_1) LOAD exp0 + faddd %f26,%f48,%f58 ! (Y0_1) yy0 += dtmp0; + + cmp %g1,%o5 ! (Y1_1) ay0 ? 0x7f800000 + bge,pn %icc,.update6 ! (Y1_1) if(ay0 >= 0x7f800000) + nop +.cont6: + fmuld KA3,%f28,%f62 ! (Y2_1) dtmp0 = KA3 * y0; + fstod %f2,%f22 ! (Y1_1) dtmp0 = (double)ftmp0; + + fmuld %f24,%f58,%f58 ! (Y0_1) yy0 *= dtmp0; + + fitod %f1,%f48 ! (Y1_1) dtmp1 = (double) exp0; + + fmuld %f54,%f46,%f54 ! (Y1_1) dtmp0 *= y0; + + faddd %f62,KA2,%f26 ! (Y2_1) dtmp0 += KA2; + + add %o2,stridex,%o2 ! px += stridex + ldd [%l2+%o0],%f60 ! (Y1_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + fcmped %fcc0,HTHRESHOLD,%f58 ! (Y0_1) if (yy0 >= HTHRESH) + + sra %i3,12,%o0 ! (Y0_2) ind0 = i0 >> 12; + lda [%o2]0x82,%o3 ! (Y1_2) ax0 = *px; + + faddd %f54,KA0,%f56 ! (Y1_1) dtmp0 += KA0; + and %o0,-8,%g5 ! (Y0_2) ind0 &= -8; + ld [%fp+tmp2],%f14 ! (Y0_2) dtmp0 = (double) i0; + + and %o3,MASK_0x7fffffff,%i3 ! (Y1_2) exp0 = ax0 & 0x7fffffff; + and %o3,MASK_0x007fffff,%o0 ! (Y1_2) ax0 &= 0x007fffff; + + cmp %i3,%o5 ! (Y1_2) ax0 ? 0x7f800000 + add %l2,%g5,%g1 ! (Y0_2) (char*)__mt_constlog4f + ind0 + faddd %f60,%f48,%f12 ! (Y1_1) yy0 = dtmp0 + dtmp1; + + fmuld %f26,%f28,%f50 ! (Y2_1) dtmp0 *= y0; + srl %i3,23,%i3 ! (Y1_2) exp0 >>= 23; + add %o0,CONST_0x8000,%i1 ! (Y1_2) i0 = ax0 + 0x8000; + fcmped %fcc1,LTHRESHOLD,%f58 ! (Y0_1) if (yy0 <= LTHRESH) + + fmuld %f56,%f46,%f46 ! (Y1_1) dtmp0 *= y0; + ldd [%g1+8],%f48 ! (Y0_2) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + sub %i3,127,%i3 ! (Y1_2) exp0 -= 127; + fitod %f14,%f60 ! (Y0_2) dtmp0 = (double) i0; + + sll %i3,8,%i2 ! (Y1_2) exp0 <<= 8; + and %i1,%l6,%i1 ! (Y1_2) i0 &= 0xffff0000; + st %i2,[%fp+tmp4] ! (Y1_2) STORE exp0 + + sub %o0,%i1,%o0 ! (Y1_2) i0 = ax0 - i0; + st %o0,[%fp+tmp5] ! (Y1_2) STORE i0 + bge,pn %icc,.update7 ! (Y1_2) if(ax0 >= 0x7f800000) + nop +.cont7: + lda [stridey+%o4]0x82,%i3 ! Y(2_1) ay0 = *py0 + cmp %o3,MASK_0x007fffff ! (Y1_2) ux0 ? 0x800000 + add %o4,stridey,%i2 ! py += stridey; + fmovdl %fcc0,HTHRESHOLD,%f58 ! (Y0_1) yy0 = HTHRESH; + + fmuld %f60,%f48,%f48 ! (Y0_2) y0 = dtmp0 * dtmp1; + lda [stridey+%o4]0x82,%f16 ! (Y2_1) ftmp0 = *py0; + ble,pn %icc,.update8 ! (Y1_2) if(ux0 < 0x800000) + faddd %f50,KA1,%f52 ! (Y2_1) dtmp0 += KA1; +.cont8: + and %i3,MASK_0x7fffffff,%i3 ! (Y2_1) ay0 &= 0x7fffffff + ld [%fp+tmp6],%f17 ! (Y2_1) dtmp1 = (double) exp0; + faddd %f12,%f46,%f60 ! (Y1_1) yy0 += dtmp0; + + cmp %i3,%o5 ! (Y2_1) ay0 ? 0x7f800000 + bge,pn %icc,.update9 ! (Y2_1) if(ay0 >= 0x7f800000) + nop + +.cont9: + fmovdg %fcc1,LTHRESHOLD,%f58 ! (Y0_1) yy0 = LTHRESH; + + fmuld KA3,%f48,%f62 ! (Y0_2) dtmp0 = KA3 * y0; + fstod %f16,%f54 ! (Y2_1) dtmp0 = (double)ftmp0; + + fmuld %f22,%f60,%f56 ! (Y1_1) yy0 *= dtmp0; + + fitod %f17,%f24 ! (Y2_1) dtmp1 = (double) exp0; + + fmuld %f52,%f28,%f52 ! (Y2_1) dtmp0 *= y0; + fdtoi %f58,%f10 ! (Y0_1) ind0 = (int) yy0; + + st %f10,[%fp+tmp0] ! (Y0_1) STORE ind0 + faddd %f62,KA2,%f22 ! (Y0_2) dtmp0 += KA2; + + fcmped %fcc0,HTHRESHOLD,%f56 ! (Y1_1) if (yy0 >= HTHRESH) + ldd [%l2+%o1],%f60 ! (Y2_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + + sra %i1,12,%o1 ! (Y1_2) ind0 = i0 >> 12; + add %o2,stridex,%i3 ! px += stridex + lda [stridex+%o2]0x82,%g1 ! (Y2_2) ax0 = *px; + + and %o1,-8,%o0 ! (Y1_2) ind0 &= -8; + add %i2,stridey,%i2 ! py += stridey + ld [%fp+tmp5],%f12 ! (Y1_2) LOAD i0 + faddd %f52,KA0,%f4 ! (Y2_1) dtmp0 += KA0; + + and %g1,MASK_0x7fffffff,%i1 ! (Y2_2) exp0 = ax0 & 0x7fffffff; + and %g1,MASK_0x007fffff,%o2 ! (Y2_2) ax0 &= 0x007fffff; + lda [%i2]0x82,%f0 ! (Y0_2) ftmp0 = *py0; + fitod %f10,%f52 ! (Y0_1) dtmp0 = (double)ind0; + + srl %i1,23,%o3 ! (Y2_2) exp0 >>= 23; + cmp %i1,%o5 ! (Y2_2) ax0 ? 0x7f800000 + faddd %f60,%f24,%f18 ! (Y2_1) yy0 = dtmp0 + dtmp1; + + fmuld %f22,%f48,%f26 ! (Y0_2) dtmp0 *= y0; + add %l2,%o0,%i1 ! (Y1_2) (char*)__mt_constlog4f + ind0 + sub %o3,127,%l7 ! (Y2_2) exp0 -= 127; + fcmped %fcc1,LTHRESHOLD,%f56 ! (Y1_1) if (yy0 <= LTHRESH) + + fmuld %f4,%f28,%f24 ! (Y2_1) dtmp0 *= y0; + add %o2,CONST_0x8000,%o1 ! (Y2_2) i0 = ax0 + 0x8000; + ldd [%i1+8],%f50 ! (Y1_2) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + fitod %f12,%f28 ! (Y1_2) dtmp0 = (double) i0; + + sll %l7,8,%l7 ! (Y2_2) exp0 <<= 8; + and %o1,%l6,%o1 ! (Y2_2) i0 &= 0xffff0000; + st %l7,[%fp+tmp6] ! (Y2_2) STORE exp0 + fsubd %f58,%f52,%f60 ! (Y0_1) y0 = yy0 - dtmp0; + + + sub %o2,%o1,%i1 ! (Y2_2) i0 = ax0 - i0; + st %i1,[%fp+tmp2] ! (Y2_2) STORE i0 + bge,pn %icc,.update10 ! (Y2_2) if(ax0 >= 0x7f800000) + nop +.cont10: + lda [%i2]0x82,%o2 ! (Y0_2) ay0 = *(int*)py0; + cmp %g1,MASK_0x007fffff ! (Y2_2) ux0 ? 0x800000 + fmovdl %fcc0,HTHRESHOLD,%f56 ! (Y1_1) yy0 = HTHRESH; + + fmuld %f28,%f50,%f46 ! (Y1_2) y0 = dtmp0 * dtmp1; + ble,pn %icc,.update11 ! (Y2_2) if(ux0 < 0x800000) + faddd %f26,KA1,%f50 ! (Y0_2) dtmp0 += KA1; +.cont11: + fmuld KB2,%f60,%f62 ! (Y0_1) dtmp0 = KB2 * y0; + and %o2,MASK_0x7fffffff,%o2 ! (Y0_2) ay0 &= 0x7fffffff + ld [%fp+tmp3],%f4 ! (Y0_2) dtmp1 = (double) exp0; + faddd %f18,%f24,%f52 ! (Y2_1) yy0 += dtmp0; + + ld [%fp+tmp0],%g1 ! (Y0_1) LAOD ind0 + cmp %o2,%o5 ! (Y0_2) ay0 ? 0x7f800000 + bge,pn %icc,.update12 ! (Y0_2) if( ay0 >= 0x7f800000) + nop +.cont12: + fstod %f0,%f24 ! (Y0_2) dtmp0 = (double)ftmp0; + + cmp counter,6 ! counter + bl,pn %icc,.tail + sub %i5,stridez,%o4 + + ba .main_loop + nop + + .align 16 +.main_loop: + fmuld KA3,%f46,%f28 ! (Y1_1) dtmp0 = KA3 * y0; + and %g1,255,%o2 ! (Y0_0) ind0 &= 255; + sub counter,3,counter ! counter + fmovdg %fcc1,LTHRESHOLD,%f56 ! (Y1_0) yy0 = LTHRESH; + + fmuld %f54,%f52,%f18 ! (Y2_0) yy0 *= dtmp0; + sll %o2,3,%i1 ! (Y0_0) ind0 <<= 3; + add %o4,stridez,%l7 ! pz += stridez + faddd %f62,KB1,%f62 ! (Y0_0) dtmp0 += KB1; + + fpackfix %f10,%f10 ! (Y0_0) dtmp1 = vis_fpackfix(dtmp1); + fitod %f4,%f26 ! (Y0_1) dtmp1 = (double) exp0; + ldd [%l0+%i1],%f58 ! (Y0_0) di0 = *(double*)((char*)__mt_constexp2f + ind0); + + fmuld %f50,%f48,%f50 ! (Y0_1) dtmp0 *= y0; + fdtoi %f56,%f20 ! (Y1_0) ind0 = (int) yy0; + st %f20,[%fp+tmp1] ! (Y1_0) STORE ind0 + + faddd %f28,KA2,%f28 ! (Y1_1) dtmp0 += KA2; + + fmuld %f62,%f60,%f62 ! (Y0_0) yy0 = dtmp0 * y0; + ldd [%l2+%g5],%f60 ! (Y0_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + add %i3,stridex,%o2 ! px += stridex + fcmped %fcc0,HTHRESHOLD,%f18 ! (Y2_0) if (yy0 >= HTHRESH) + + lda [%o2]0x82,%i1 ! (Y0_2) ax0 = *px; + sra %o1,12,%g5 ! (Y2_1) ind0 = i0 >> 12; + fpadd32 %f10,%f58,%f22 ! (Y0_0) di0 = vis_fpadd32(di0,dtmp1); + + faddd %f50,KA0,%f58 ! (Y0_1) dtmp0 += KA0; + and %g5,-8,%o1 ! (Y2_1) ind0 &= -8; + ld [%fp+tmp2],%f6 ! (Y2_1) dtmp0 = (double) i0; + + fitod %f20,%f52 ! (Y1_0) dtmp0 = (double)ind0; + and %i1,MASK_0x7fffffff,%i3 ! (Y0_2) exp0 = ax0 & 0x7fffffff; + and %i1,MASK_0x007fffff,%g5 ! (Y0_2) ax0 &= 0x007fffff; + + fmuld %f62,%f22,%f62 ! (Y0_0) yy0 *= di0; + srl %i3,23,%o3 ! (Y0_2) exp0 >>= 23; + add %l2,%o1,%g1 ! (Y2_1) (char*)__mt_constlog4f + ind0 + faddd %f60,%f26,%f26 ! (Y0_1) yy0 = dtmp0 + dtmp1; + + fmuld %f28,%f46,%f50 ! (Y1_1) dtmp0 *= y0; + sub %o3,127,%o3 ! (Y0_2) exp0 -= 127; + cmp %i3,%o5 ! (Y0_2) ax0 ? 0x7f800000 + fcmped %fcc1,LTHRESHOLD,%f18 ! (Y2_0) if (yy0 <= LTHRESH) + + fmuld %f58,%f48,%f48 ! (Y0_1) dtmp0 *= y0; + add %g5,CONST_0x8000,%i3 ! (Y0_2) i0 = ax0 + 0x8000; + ldd [%g1+8],%f58 ! (Y2_1) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + fitod %f6,%f54 ! (Y2_1) dtmp0 = (double) i0; + + sll %o3,8,%o4 ! (Y0_2) exp0 <<= 8; + and %i3,%l6,%i3 ! (Y0_2) i0 &= 0xffff0000; + st %o4,[%fp+tmp3] ! (Y0_2) STORE exp0 + fsubd %f56,%f52,%f52 ! (Y1_0) y0 = yy0 - dtmp0; + + sub %g5,%i3,%o4 ! (Y0_2) i0 = ax0 - i0; + st %o4,[%fp+tmp2] ! (Y0_2) STORE i0 + bge,pn %icc,.update13 ! (Y0_2) if( ax0 >= 0x7f800000 ) + faddd %f62,%f22,%f62 ! (Y0_0) yy0 += di0; +.cont13: + lda [stridey+%i2]0x82,%g1 ! (Y1_1) ay0 = *(unsigned*)py0 + add %i2,stridey,%o4 ! py += stridey + cmp %i1,MASK_0x007fffff ! (Y0_2) ux0 ? 0x800000 + fmovdl %fcc0,HTHRESHOLD,%f18 ! (Y2_0) yy0 = HTHRESH; + + fmuld %f54,%f58,%f28 ! (Y2_1) y0 = dtmp0 * dtmp1; + lda [stridey+%i2]0x82,%f2 ! (Y1_1) ftmp0 = *py0; + ble,pn %icc,.update14 ! (Y0_2) if(ux0 < 0x800000) + faddd %f50,KA1,%f54 ! (Y1_1) dtmp0 += KA1; +.cont14: + fmuld KB2,%f52,%f56 ! (Y1_0) dtmp0 = KB2 * y0; + and %g1,MASK_0x7fffffff,%g1 ! (Y1_1) ay0 &= 0x7fffffff; + ld [%fp+tmp4],%f1 ! (Y1_1) LOAD exp0 + faddd %f26,%f48,%f58 ! (Y0_1) yy0 += dtmp0; + + ld [%fp+tmp1],%g5 ! (Y1_0) ind0 = (int) yy0; + cmp %g1,%o5 ! (Y1_1) ay0 ? 0x7f800000 + bge,pn %icc,.update15 ! (Y1_1) if(ay0 >= 0x7f800000) + fdtos %f62,%f8 ! (Y0_0) ftmp0 = (float)yy0; +.cont15: + st %f8,[%l7] ! (Y0_0) *pz0 = ftmp0; + fmovdg %fcc1,LTHRESHOLD,%f18 ! (Y2_0) yy0 = LTHRESH; + + add %l7,stridez,%l7 ! pz += stridez + fmuld KA3,%f28,%f62 ! (Y2_1) dtmp0 = KA3 * y0; + and %g5,255,%g5 ! (Y1_0) ind0 &= 255; + fstod %f2,%f22 ! (Y1_1) dtmp0 = (double)ftmp0; + + fmuld %f24,%f58,%f58 ! (Y0_1) yy0 *= dtmp0; + sll %g5,3,%i2 ! (Y1_0) ind0 <<= 3; + faddd %f56,KB1,%f60 ! (Y1_0) dtmp0 += KB1; + + fpackfix %f20,%f20 ! (Y1_0) dtmp1 = vis_fpackfix(dtmp1); + fitod %f1,%f48 ! (Y1_1) dtmp1 = (double) exp0; + ldd [%l0+%i2],%f56 ! (Y1_0) di0 = *(double*)((char*)__mt_constexp2f + ind0); + + fmuld %f54,%f46,%f54 ! (Y1_1) dtmp0 *= y0; + fdtoi %f18,%f2 ! (Y2_0) ind0 = (int) yy0; + st %f2,[%fp+tmp1] ! (Y2_0) STORE ind0 + + faddd %f62,KA2,%f26 ! (Y2_1) dtmp0 += KA2; + + fmuld %f60,%f52,%f62 ! (Y1_0) yy0 = dtmp0 * y0; + add %o2,stridex,%o2 ! px += stridex + ldd [%l2+%o0],%f60 ! (Y1_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + fcmped %fcc0,HTHRESHOLD,%f58 ! (Y0_1) if (yy0 >= HTHRESH) + + fpadd32 %f20,%f56,%f52 ! (Y1_0) di0 = vis_fpadd32(di0,dtmp1); + sra %i3,12,%o0 ! (Y0_2) ind0 = i0 >> 12; + lda [%o2]0x82,%o3 ! (Y1_2) ax0 = *px; + + faddd %f54,KA0,%f56 ! (Y1_1) dtmp0 += KA0; + and %o0,-8,%g5 ! (Y0_2) ind0 &= -8; + ld [%fp+tmp2],%f14 ! (Y0_2) dtmp0 = (double) i0; + + fitod %f2,%f54 ! (Y2_0) dtmp0 = (double)ind0; + and %o3,MASK_0x7fffffff,%i3 ! (Y1_2) exp0 = ax0 & 0x7fffffff; + and %o3,MASK_0x007fffff,%o0 ! (Y1_2) ax0 &= 0x007fffff; + + fmuld %f62,%f52,%f62 ! (Y1_0) yy0 *= di0; + cmp %i3,%o5 ! (Y1_2) ax0 ? 0x7f800000 + add %l2,%g5,%g1 ! (Y0_2) (char*)__mt_constlog4f + ind0 + faddd %f60,%f48,%f12 ! (Y1_1) yy0 = dtmp0 + dtmp1; + + fmuld %f26,%f28,%f50 ! (Y2_1) dtmp0 *= y0; + srl %i3,23,%i3 ! (Y1_2) exp0 >>= 23; + add %o0,CONST_0x8000,%i1 ! (Y1_2) i0 = ax0 + 0x8000; + fcmped %fcc1,LTHRESHOLD,%f58 ! (Y0_1) if (yy0 <= LTHRESH) + + fmuld %f56,%f46,%f46 ! (Y1_1) dtmp0 *= y0; + ldd [%g1+8],%f48 ! (Y0_2) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + sub %i3,127,%i3 ! (Y1_2) exp0 -= 127; + fitod %f14,%f60 ! (Y0_2) dtmp0 = (double) i0; + + sll %i3,8,%i2 ! (Y1_2) exp0 <<= 8; + and %i1,%l6,%i1 ! (Y1_2) i0 &= 0xffff0000; + st %i2,[%fp+tmp4] ! (Y1_2) STORE exp0 + fsubd %f18,%f54,%f26 ! (Y2_0) y0 = yy0 - dtmp0; + + sub %o0,%i1,%o0 ! (Y1_2) i0 = ax0 - i0; + st %o0,[%fp+tmp5] ! (Y1_2) STORE i0 + bge,pn %icc,.update16 ! (Y1_2) if(ax0 >= 0x7f800000) + faddd %f62,%f52,%f54 ! (Y1_0) yy0 += di0; +.cont16: + lda [stridey+%o4]0x82,%i3 ! Y(2_1) ay0 = *py0 + cmp %o3,MASK_0x007fffff ! (Y1_2) ux0 ? 0x800000 + add %o4,stridey,%i2 ! py += stridey; + fmovdl %fcc0,HTHRESHOLD,%f58 ! (Y0_1) yy0 = HTHRESH; + + fmuld %f60,%f48,%f48 ! (Y0_2) y0 = dtmp0 * dtmp1; + lda [stridey+%o4]0x82,%f16 ! (Y2_1) ftmp0 = *py0; + ble,pn %icc,.update17 ! (Y1_2) if(ux0 < 0x800000) + faddd %f50,KA1,%f52 ! (Y2_1) dtmp0 += KA1; +.cont17: + fmuld KB2,%f26,%f4 ! (Y2_0) dtmp0 = KB2 * y0; + and %i3,MASK_0x7fffffff,%i3 ! (Y2_1) ay0 &= 0x7fffffff + ld [%fp+tmp6],%f17 ! (Y2_1) dtmp1 = (double) exp0; + faddd %f12,%f46,%f60 ! (Y1_1) yy0 += dtmp0; + + ld [%fp+tmp1],%o0 + cmp %i3,%o5 ! (Y2_1) ay0 ? 0x7f800000 + bge,pn %icc,.update18 ! (Y2_1) if(ay0 >= 0x7f800000) + fdtos %f54,%f15 ! (Y1_0) ftmp0 = (float)yy0; +.cont18: + st %f15,[%l7] ! (Y1_0) *pz0 = ftmp0; + add %l7,stridez,%o4 ! pz += stridez + fmovdg %fcc1,LTHRESHOLD,%f58 ! (Y0_1) yy0 = LTHRESH; + + fmuld KA3,%f48,%f62 ! (Y0_2) dtmp0 = KA3 * y0; + and %o0,255,%o0 ! (Y2_0) ind0 &= 255; + fstod %f16,%f54 ! (Y2_1) dtmp0 = (double)ftmp0; + + fmuld %f22,%f60,%f56 ! (Y1_1) yy0 *= dtmp0; + sll %o0,3,%l7 ! (Y2_0) ind0 <<= 3; + faddd %f4,KB1,%f60 ! (Y2_0) dtmp0 += KB1; + + fpackfix %f2,%f2 ! (Y2_0) dtmp1 = vis_fpackfix(dtmp1); + fitod %f17,%f24 ! (Y2_1) dtmp1 = (double) exp0; + ldd [%l0+%l7],%f4 ! (Y2_0) di0 = *(double*)((char*)__mt_constexp2f + ind0); + + fmuld %f52,%f28,%f52 ! (Y2_1) dtmp0 *= y0; + fdtoi %f58,%f10 ! (Y0_1) ind0 = (int) yy0; + + st %f10,[%fp+tmp0] ! (Y0_1) STORE ind0 + faddd %f62,KA2,%f22 ! (Y0_2) dtmp0 += KA2; + + fmuld %f60,%f26,%f62 ! (Y2_0) yy0 = dtmp0 * y0; + fcmped %fcc0,HTHRESHOLD,%f56 ! (Y1_1) if (yy0 >= HTHRESH) + ldd [%l2+%o1],%f60 ! (Y2_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + + sra %i1,12,%o1 ! (Y1_2) ind0 = i0 >> 12; + add %o2,stridex,%i3 ! px += stridex + lda [stridex+%o2]0x82,%g1 ! (Y2_2) ax0 = *px; + fpadd32 %f2,%f4,%f46 ! (Y2_0) di0 = vis_fpadd32(di0,dtmp1); + + and %o1,-8,%o0 ! (Y1_2) ind0 &= -8; + add %i2,stridey,%i2 ! py += stridey + ld [%fp+tmp5],%f12 ! (Y1_2) LOAD i0 + faddd %f52,KA0,%f4 ! (Y2_1) dtmp0 += KA0; + + and %g1,MASK_0x7fffffff,%i1 ! (Y2_2) exp0 = ax0 & 0x7fffffff; + and %g1,MASK_0x007fffff,%o2 ! (Y2_2) ax0 &= 0x007fffff; + lda [%i2]0x82,%f0 ! (Y0_2) ftmp0 = *py0; + fitod %f10,%f52 ! (Y0_1) dtmp0 = (double)ind0; + + fmuld %f62,%f46,%f62 ! (Y2_0) yy0 *= di0; + srl %i1,23,%o3 ! (Y2_2) exp0 >>= 23; + cmp %i1,%o5 ! (Y2_2) ax0 ? 0x7f800000 + faddd %f60,%f24,%f18 ! (Y2_1) yy0 = dtmp0 + dtmp1; + + fmuld %f22,%f48,%f26 ! (Y0_2) dtmp0 *= y0; + add %l2,%o0,%i1 ! (Y1_2) (char*)__mt_constlog4f + ind0 + sub %o3,127,%l7 ! (Y2_2) exp0 -= 127; + fcmped %fcc1,LTHRESHOLD,%f56 ! (Y1_1) if (yy0 <= LTHRESH) + + fmuld %f4,%f28,%f24 ! (Y2_1) dtmp0 *= y0; + add %o2,CONST_0x8000,%o1 ! (Y2_2) i0 = ax0 + 0x8000; + ldd [%i1+8],%f50 ! (Y1_2) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + fitod %f12,%f28 ! (Y1_2) dtmp0 = (double) i0; + + sll %l7,8,%l7 ! (Y2_2) exp0 <<= 8; + and %o1,%l6,%o1 ! (Y2_2) i0 &= 0xffff0000; + st %l7,[%fp+tmp6] ! (Y2_2) STORE exp0 + fsubd %f58,%f52,%f60 ! (Y0_1) y0 = yy0 - dtmp0; + + sub %o2,%o1,%i1 ! (Y2_2) i0 = ax0 - i0; + st %i1,[%fp+tmp2] ! (Y2_2) STORE i0 + bge,pn %icc,.update19 ! (Y2_2) if(ax0 >= 0x7f800000) + faddd %f62,%f46,%f22 ! (Y2_0) yy0 += di0; +.cont19: + lda [%i2]0x82,%o2 ! (Y0_2) ay0 = *(int*)py0; + cmp %g1,MASK_0x007fffff ! (Y2_2) ux0 ? 0x800000 + fmovdl %fcc0,HTHRESHOLD,%f56 ! (Y1_1) yy0 = HTHRESH; + + fmuld %f28,%f50,%f46 ! (Y1_2) y0 = dtmp0 * dtmp1; + ble,pn %icc,.update20 ! (Y2_2) if(ux0 < 0x800000) + faddd %f26,KA1,%f50 ! (Y0_2) dtmp0 += KA1; +.cont20: + fmuld KB2,%f60,%f62 ! (Y0_1) dtmp0 = KB2 * y0; + and %o2,MASK_0x7fffffff,%o2 ! (Y0_2) ay0 &= 0x7fffffff + ld [%fp+tmp3],%f4 ! (Y0_2) dtmp1 = (double) exp0; + faddd %f18,%f24,%f52 ! (Y2_1) yy0 += dtmp0; + + ld [%fp+tmp0],%g1 ! (Y0_1) LAOD ind0 + cmp %o2,%o5 ! (Y0_2) ay0 ? 0x7f800000 + bge,pn %icc,.update21 ! (Y0_2) if( ay0 >= 0x7f800000) + fdtos %f22,%f12 ! (Y2_0) ftmp0 = (float)yy0; +.cont21: + st %f12,[%o4] ! (Y2_0) *pz0 = ftmp0; + cmp counter,6 ! counter + bge,pt %icc,.main_loop + fstod %f0,%f24 ! (Y0_2) dtmp0 = (double)ftmp0; + +.tail: + subcc counter,1,counter + bneg,pn %icc,.begin + add %o4,stridez,%i5 + + fmuld KA3,%f46,%f28 ! (Y1_1) dtmp0 = KA3 * y0; + and %g1,255,%o2 ! (Y0_0) ind0 &= 255; + fmovdg %fcc1,LTHRESHOLD,%f56 ! (Y1_0) yy0 = LTHRESH; + + fmuld %f54,%f52,%f18 ! (Y2_0) yy0 *= dtmp0; + sll %o2,3,%i1 ! (Y0_0) ind0 <<= 3; + add %o4,stridez,%l7 ! pz += stridez + faddd %f62,KB1,%f62 ! (Y0_0) dtmp0 += KB1; + + fpackfix %f10,%f10 ! (Y0_0) dtmp1 = vis_fpackfix(dtmp1); + fitod %f4,%f26 ! (Y0_1) dtmp1 = (double) exp0; + ldd [%l0+%i1],%f58 ! (Y0_0) di0 = *(double*)((char*)__mt_constexp2f + ind0); + + fmuld %f50,%f48,%f50 ! (Y0_1) dtmp0 *= y0; + fdtoi %f56,%f20 ! (Y1_0) ind0 = (int) yy0; + st %f20,[%fp+tmp1] ! (Y1_0) STORE ind0 + + faddd %f28,KA2,%f28 ! (Y1_1) dtmp0 += KA2; + + fmuld %f62,%f60,%f62 ! (Y0_0) yy0 = dtmp0 * y0; + ldd [%l2+%g5],%f60 ! (Y0_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + fcmped %fcc0,HTHRESHOLD,%f18 ! (Y2_0) if (yy0 >= HTHRESH) + + fpadd32 %f10,%f58,%f22 ! (Y0_0) di0 = vis_fpadd32(di0,dtmp1); + + faddd %f50,KA0,%f58 ! (Y0_1) dtmp0 += KA0; + + fitod %f20,%f52 ! (Y1_0) dtmp0 = (double)ind0; + + fmuld %f62,%f22,%f62 ! (Y0_0) yy0 *= di0; + faddd %f60,%f26,%f26 ! (Y0_1) yy0 = dtmp0 + dtmp1; + + fmuld %f28,%f46,%f50 ! (Y1_1) dtmp0 *= y0; + fcmped %fcc1,LTHRESHOLD,%f18 ! (Y2_0) if (yy0 <= LTHRESH) + + fmuld %f58,%f48,%f48 ! (Y0_1) dtmp0 *= y0; + + fsubd %f56,%f52,%f52 ! (Y1_0) y0 = yy0 - dtmp0; + + faddd %f62,%f22,%f62 ! (Y0_0) yy0 += di0; + + lda [stridey+%i2]0x82,%g1 ! (Y1_1) ay0 = *(unsigned*)py0 + add %i2,stridey,%o4 ! py += stridey + fmovdl %fcc0,HTHRESHOLD,%f18 ! (Y2_0) yy0 = HTHRESH; + + lda [stridey+%i2]0x82,%f2 ! (Y1_1) ftmp0 = *py0; + faddd %f50,KA1,%f54 ! (Y1_1) dtmp0 += KA1; + + fmuld KB2,%f52,%f56 ! (Y1_0) dtmp0 = KB2 * y0; + and %g1,MASK_0x7fffffff,%g1 ! (Y1_1) ay0 &= 0x7fffffff; + ld [%fp+tmp4],%f1 ! (Y1_1) LOAD exp0 + faddd %f26,%f48,%f58 ! (Y0_1) yy0 += dtmp0; + + ld [%fp+tmp1],%g5 ! (Y1_0) ind0 = (int) yy0; + cmp %g1,%o5 ! (Y1_1) ay0 ? 0x7f800000 + bge,pn %icc,.update22 ! (Y1_1) if(ay0 >= 0x7f800000) + fdtos %f62,%f8 ! (Y0_0) ftmp0 = (float)yy0; +.cont22: + st %f8,[%l7] ! (Y0_0) *pz0 = ftmp0; + fmovdg %fcc1,LTHRESHOLD,%f18 ! (Y2_0) yy0 = LTHRESH; + + subcc counter,1,counter + bneg,pn %icc,.begin + add %l7,stridez,%i5 + + add %l7,stridez,%l7 ! pz += stridez + and %g5,255,%g5 ! (Y1_0) ind0 &= 255; + fstod %f2,%f22 ! (Y1_1) dtmp0 = (double)ftmp0; + + fmuld %f24,%f58,%f58 ! (Y0_1) yy0 *= dtmp0; + sll %g5,3,%i2 ! (Y1_0) ind0 <<= 3; + faddd %f56,KB1,%f60 ! (Y1_0) dtmp0 += KB1; + + fpackfix %f20,%f20 ! (Y1_0) dtmp1 = vis_fpackfix(dtmp1); + fitod %f1,%f48 ! (Y1_1) dtmp1 = (double) exp0; + ldd [%l0+%i2],%f56 ! (Y1_0) di0 = *(double*)((char*)__mt_constexp2f + ind0); + + fmuld %f54,%f46,%f54 ! (Y1_1) dtmp0 *= y0; + fdtoi %f18,%f2 ! (Y2_0) ind0 = (int) yy0; + st %f2,[%fp+tmp1] ! (Y2_0) STORE ind0 + + + fmuld %f60,%f52,%f62 ! (Y1_0) yy0 = dtmp0 * y0; + ldd [%l2+%o0],%f60 ! (Y1_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + fcmped %fcc0,HTHRESHOLD,%f58 ! (Y0_1) if (yy0 >= HTHRESH) + + fpadd32 %f20,%f56,%f52 ! (Y1_0) di0 = vis_fpadd32(di0,dtmp1); + + faddd %f54,KA0,%f56 ! (Y1_1) dtmp0 += KA0; + + fitod %f2,%f54 ! (Y2_0) dtmp0 = (double)ind0; + + fmuld %f62,%f52,%f62 ! (Y1_0) yy0 *= di0; + faddd %f60,%f48,%f12 ! (Y1_1) yy0 = dtmp0 + dtmp1; + + fcmped %fcc1,LTHRESHOLD,%f58 ! (Y0_1) if (yy0 <= LTHRESH) + + fmuld %f56,%f46,%f46 ! (Y1_1) dtmp0 *= y0; + + fsubd %f18,%f54,%f26 ! (Y2_0) y0 = yy0 - dtmp0; + + faddd %f62,%f52,%f54 ! (Y1_0) yy0 += di0; + + fmovdl %fcc0,HTHRESHOLD,%f58 ! (Y0_1) yy0 = HTHRESH; + + + fmuld KB2,%f26,%f4 ! (Y2_0) dtmp0 = KB2 * y0; + faddd %f12,%f46,%f60 ! (Y1_1) yy0 += dtmp0; + + ld [%fp+tmp1],%o0 + fdtos %f54,%f15 ! (Y1_0) ftmp0 = (float)yy0; + + st %f15,[%l7] ! (Y1_0) *pz0 = ftmp0; + add %l7,stridez,%o4 ! pz += stridez + fmovdg %fcc1,LTHRESHOLD,%f58 ! (Y0_1) yy0 = LTHRESH; + + subcc counter,1,counter + bneg,pn %icc,.begin + or %g0,%o4,%i5 + + and %o0,255,%o0 ! (Y2_0) ind0 &= 255; + + fmuld %f22,%f60,%f56 ! (Y1_1) yy0 *= dtmp0; + sll %o0,3,%l7 ! (Y2_0) ind0 <<= 3; + faddd %f4,KB1,%f60 ! (Y2_0) dtmp0 += KB1; + + fpackfix %f2,%f2 ! (Y2_0) dtmp1 = vis_fpackfix(dtmp1); + ldd [%l0+%l7],%f4 ! (Y2_0) di0 = *(double*)((char*)__mt_constexp2f + ind0); + + fdtoi %f58,%f10 ! (Y0_1) ind0 = (int) yy0; + + st %f10,[%fp+tmp0] ! (Y0_1) STORE ind0 + + fmuld %f60,%f26,%f62 ! (Y2_0) yy0 = dtmp0 * y0; + fcmped %fcc0,HTHRESHOLD,%f56 ! (Y1_1) if (yy0 >= HTHRESH) + + fpadd32 %f2,%f4,%f46 ! (Y2_0) di0 = vis_fpadd32(di0,dtmp1); + + add %i2,stridey,%i2 ! py += stridey + + fitod %f10,%f52 ! (Y0_1) dtmp0 = (double)ind0; + + fmuld %f62,%f46,%f62 ! (Y2_0) yy0 *= di0; + + fcmped %fcc1,LTHRESHOLD,%f56 ! (Y1_1) if (yy0 <= LTHRESH) + + + fsubd %f58,%f52,%f60 ! (Y0_1) y0 = yy0 - dtmp0; + + faddd %f62,%f46,%f22 ! (Y2_0) yy0 += di0; + + fmovdl %fcc0,HTHRESHOLD,%f56 ! (Y1_1) yy0 = HTHRESH; + + fmuld KB2,%f60,%f62 ! (Y0_1) dtmp0 = KB2 * y0; + + ld [%fp+tmp0],%g1 ! (Y0_1) LAOD ind0 + fdtos %f22,%f12 ! (Y2_0) ftmp0 = (float)yy0; + + st %f12,[%o4] ! (Y2_0) *pz0 = ftmp0; + + subcc counter,1,counter + bneg,pn %icc,.begin + add %o4,stridez,%i5 + + and %g1,255,%o2 ! (Y0_0) ind0 &= 255; + fmovdg %fcc1,LTHRESHOLD,%f56 ! (Y1_0) yy0 = LTHRESH; + + sll %o2,3,%i1 ! (Y0_0) ind0 <<= 3; + add %o4,stridez,%l7 ! pz += stridez + faddd %f62,KB1,%f62 ! (Y0_0) dtmp0 += KB1; + + fpackfix %f10,%f10 ! (Y0_0) dtmp1 = vis_fpackfix(dtmp1); + ldd [%l0+%i1],%f58 ! (Y0_0) di0 = *(double*)((char*)__mt_constexp2f + ind0); + + fdtoi %f56,%f20 ! (Y1_0) ind0 = (int) yy0; + st %f20,[%fp+tmp1] ! (Y1_0) STORE ind0 + + fmuld %f62,%f60,%f62 ! (Y0_0) yy0 = dtmp0 * y0; + + fpadd32 %f10,%f58,%f22 ! (Y0_0) di0 = vis_fpadd32(di0,dtmp1); + + fitod %f20,%f52 ! (Y1_0) dtmp0 = (double)ind0; + + fmuld %f62,%f22,%f62 ! (Y0_0) yy0 *= di0; + + fsubd %f56,%f52,%f52 ! (Y1_0) y0 = yy0 - dtmp0; + + faddd %f62,%f22,%f62 ! (Y0_0) yy0 += di0; + + fmuld KB2,%f52,%f56 ! (Y1_0) dtmp0 = KB2 * y0; + + ld [%fp+tmp1],%g5 ! (Y1_0) ind0 = (int) yy0; + fdtos %f62,%f8 ! (Y0_0) ftmp0 = (float)yy0; + st %f8,[%l7] ! (Y0_0) *pz0 = ftmp0; + + subcc counter,1,counter + bneg .begin + add %l7,stridez,%i5 + + add %l7,stridez,%l7 ! pz += stridez + and %g5,255,%g5 ! (Y1_0) ind0 &= 255; + + sll %g5,3,%i2 ! (Y1_0) ind0 <<= 3; + faddd %f56,KB1,%f60 ! (Y1_0) dtmp0 += KB1; + + fpackfix %f20,%f20 ! (Y1_0) dtmp1 = vis_fpackfix(dtmp1); + ldd [%l0+%i2],%f56 ! (Y1_0) di0 = *(double*)((char*)__mt_constexp2f + ind0); + + fmuld %f60,%f52,%f62 ! (Y1_0) yy0 = dtmp0 * y0; + + fpadd32 %f20,%f56,%f52 ! (Y1_0) di0 = vis_fpadd32(di0,dtmp1); + + fmuld %f62,%f52,%f62 ! (Y1_0) yy0 *= di0; + + faddd %f62,%f52,%f54 ! (Y1_0) yy0 += di0; + + fdtos %f54,%f15 ! (Y1_0) ftmp0 = (float)yy0; + + st %f15,[%l7] ! (Y1_0) *pz0 = ftmp0; + ba .begin + add %l7,stridez,%i5 ! pz += stridez + +.exit: + ret + restore + + .align 16 +.specs_exit: + add %i1,stridex,%o2 + add %i3,stridey,%i2 + st %f4,[%i5] + + sub counter,1,counter + ba .begin1 + add %i5,stridez,%i5 + +.spec1: + ld [%l0+2048+64],%f0 ! LOAD 1.0f + or %g0,%i1,%o1 + or %g0,%i3,%o3 + + ld [%o2],%f4 ! *px + or %g0,%o2,%i1 + or %g0,%i2,%i3 + + ld [%i3],%f6 ! *py + or %g0,%l7,%o2 + fsubs %f0,%f0,%f5 ! 0.0f + + sethi %hi(0x7f800000),%l6 + cmp %o4,0 ! ay ? 0 + be,a,pn %icc,.specs_exit ! if(ay == 0) + fmovs %f0,%f4 ! return 1.0f + + cmp %o3,%l6 ! ax0 ? 0x7f800000 + bgu,a %icc,.specs_exit ! ax0 > 0x7f800000 + fmuls %f4,%f6,%f4 ! return *px * *py; /* |X| or |Y| = Nan */ + + cmp %o4,%l6 ! ay ? 0x7f800000 + bgu,a .specs_exit ! ay > 0x7f800000 + fmuls %f4,%f6,%f4 ! return *px * *py; /* |X| or |Y| = Nan */ + + sethi %hi(0x3f800000),%o5 + bne,a %icc,1f ! if (ay != 0x7f800000) { /* |Y| = Inf */ + srl %o1,31,%o1 ! sx = ux >> 31 + + cmp %o3,%o5 ! ax0 ? 0x3f800000 + be,a .specs_exit ! if (ax0 == 0x3f800000) + fmuls %f6,%f5,%f4 ! return *py * 0.0f; /* +-1 ** +-Inf = NaN */ + + sub %o3,%o5,%o3 ! ax0 - 0x3f800000 + srl %o2,31,%o2 ! uy >> 31 + + srlx %o3,63,%o3 ! (ax0 - 0x3f800000) << 63 + + cmp %o3,%o2 ! ((ax0 - 0x3f800000) << 63) ? (uy >> 31) + bne,a .specs_exit + fzeros %f4 ! return 0.f; + + ba .specs_exit + fabss %f6,%f4 ! return fabss(*py) +1: + cmp %o1,0 ! sx ? 0 + be,pn %icc,.spec1_exit ! if (sx == 0) + or %g0,%g0,%o5 ! yisint0 = 0; + + srl %o4,23,%l7 ! exp = ay >> 23; + cmp %l7,0x97 ! exp ? 0x97 + bge,a,pn %icc,.spec1_exit ! if (exp >= 0x97) /* |Y| >= 2^24 */ + add %g0,2,%o5 ! yisint = 2; + + cmp %l7,0x7f ! exp ? 0x7f + bl,pn %icc,.spec1_exit ! if (exp < 0x7f) + sub %g0,%l7,%l7 ! exp = -exp; + + add %l7,(0x7f + 23),%l7 ! exp += (0x07f + 23); + srl %o4,%l7,%l6 ! i0 = ay >> exp + sll %l6,%l7,%l7 ! i0 << exp + + cmp %l7,%o4 ! (i0 << exp) ? ay + bne,pn %icc,.spec1_exit ! if((i0 << exp) != ay) + and %l6,1,%l6 ! i0 &= 1 + + sub %g0,%l6,%l6 ! i0 = -i0; + add %l6,2,%o5 ! yisint0 = 2 + i0; + +.spec1_exit: + srl %o2,31,%o2 ! uy >> 31 + cmp %o2,0 ! (uy >> 31) ? 0 + movne %icc,%g0,%o3 ! if (uy >> 31) ax0 = 0; + + sll %o5,31,%o5 ! yisint0 <<= 31; + add %o5,%o3,%o5 ! ax0 += yisint0; + + add %i1,stridex,%o2 ! px += stridex; + add %i3,stridey,%i2 ! py += stridey; + st %o5,[%i5] ! return *(float*)&ax0; + + sub counter,1,counter ! counter--; + ba .begin1 + add %i5,stridez,%i5 ! pz += stridez; + +.spec2: + or %g0,%i1,%o1 + or %g0,%i3,%o3 + ld [%l0+2048+64],%f0 ! LOAD 1.0f + or %g0,%o2,%i1 + or %g0,%i2,%i3 + + or %g0,%l7,%o2 + cmp %o4,0 ! ay ? 0 + be,a,pn %icc,.specs_exit ! if(ay == 0) + fmovs %f0,%f4 ! return 1.0f + + srl %o3,23,%l7 ! exp0 = (ax0 >> 23); + sub %l7,127,%l7 ! exp = exp0 = exp0 - 127; + + or %g0,%g0,%o5 ! yisint = 0; + cmp %o3,MASK_0x007fffff ! (int)ax0 ? 0x00800000 + bg,pn %icc,1f ! if ((int)ax0 >= 0x00800000) + nop + + ! X = denormal or negative + st %o3,[%fp+tmp0] ! *((float*) &ax0) = (float) (int)ax0; + ld [%fp+tmp0],%f4 + fitos %f4,%f4 + st %f4,[%fp+tmp0] + ld [%fp+tmp0],%o3 + + srl %o3,23,%l7 ! exp = (ax0 >> 23) + sub %l7,127+149,%l7 ! exp -= (127+149) +1: + cmp %o1,0 ! ux ? 0 + bg,a %icc,.spec_proc ! if((int)ux > 0) + sethi %hi(0xffff0000),%l6 + + srl %o4,23,%o0 ! exp = ay >> 23; + cmp %o0,0x97 ! exp ? 0x97 + bge,a,pn %icc,2f ! if (exp >= 0x97) /* |Y| >= 2^24 */ + add %g0,2,%o5 ! yisint0 = 2; /* Y - even */ + + cmp %o0,0x7f ! exp ? 0x7f + bl,pn %icc,2f ! if(exp < 0x7f) + nop + + sub %g0,%o0,%o0 ! exp = -exp; + add %o0,(0x7f + 23),%o0 ! exp += (0x7f + 23) + srl %o4,%o0,%l6 ! i0 = ay >> ((0x7f + 23) - exp); + sll %l6,%o0,%o0 ! i0 << ((0x7f + 23) - exp + cmp %o0,%o4 ! (i0 << ((0x7f + 23) - exp)) ? ay + bne,pn %icc,2f ! if(i0 << ((0x7f + 23) - exp)) != ay) + nop + + and %l6,1,%l6 ! i0 &= 1; + sub %g0,%l6,%l6 ! i0 = -i0; + add %l6,2,%o5 ! yisint = i0 + 2; +2: + cmp %o3,0 ! ax0 ? 0 + bne,pn %icc,4f ! if(ax0 != 0) + nop + + srl %o1,31,%o1 ! sx = ux >> 31 + srl %o2,31,%o2 ! uy >> 31 + + cmp %o2,0 ! (uy >> 31) ? 0 + be,a,pn %icc,3f ! if((uy >> 31) == 0) + fzeros %f4 ! return ZERO + + fdivs %f0,%f3,%f4 ! fy = ONE/ZERO +3: + andcc %o1,%o5,%g0 ! sx & yisint0 + be,pn %icc,.specs_exit ! if( (sx & yisint0) == 0 ) + nop + + ba .specs_exit + fnegs %f4,%f4 ! fy = -fy; +4: + cmp %o5,0 ! ysisint0 ? 0 + be,a %icc,.specs_exit ! if(yisint0 == 0) + fdivs %f3,%f3,%f4 ! return ZERO/ZERO + + sethi %hi(0xffff0000),%l6 + +.spec_proc: + sll %l7,8,%l7 ! exp0 = exp0 << 8; + st %l7,[%fp+tmp1] ! STORE exp0 + and %o3,MASK_0x007fffff,%g5 ! ax0 &= 0x007fffff; + ld [%i3],%f14 ! ftmp0 = py[0] + sllx %o5,63,%o5 ! ysisint0 <<= 63; + add %g5,CONST_0x8000,%o3 ! i0 = ax0 + 0x8000; + stx %o5,[%fp+tmp5] ! STORE yisint0 + and %o3,%l6,%l7 ! i0 &= 0xffff0000; + sub %g5,%l7,%o1 ! i0 = ax0 - i0; + sra %l7,12,%g5 ! ind0 = i0 >> 12; + st %o1,[%fp+tmp2] ! STORE i0 + fstod %f14,%f54 ! dtmp1 = (double)ftmp0 + and %g5,-8,%g5 ! ind0 &= -8; + add %l2,%g5,%l7 ! (char*)__mt_constlog4f + ind0 + ld [%fp+tmp1],%f18 ! LOAD exp0 + ld [%fp+tmp2],%f16 ! LOAD i0 + ldd [%l7+8],%f62 ! dtmp2 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + ldd [%l2+%g5],%f56 ! dtmp3 = *(double *)((char*)__mt_constlog4f + ind0); + fitod %f18,%f58 ! dtmp4 = (double)exp0 + fitod %f16,%f60 ! dtmp5 = (double)i0 + fmuld %f60,%f62,%f60 ! y0 = dtmp5 * dtmp2; + faddd %f56,%f58,%f58 ! yy0 = dtmp3 + dtmp4; + fmuld KA3,%f60,%f52 ! dtmp0 = KA3 * y0; + faddd %f52,KA2,%f50 ! dtmp0 += KA2; + fmuld %f50,%f60,%f48 ! dtmp0 *= y0; + faddd %f48,KA1,%f46 ! dtmp0 += KA1; + fmuld %f46,%f60,%f62 ! dtmp0 *= y0; + ldd [%fp+tmp5],%f24 ! LOAD yisint0 + faddd %f62,KA0,%f56 ! dtmp0 += KA0; + fmuld %f56,%f60,%f52 ! dtmp0 *= y0; + faddd %f58,%f52,%f50 ! yy0 += dtmp1; + fmuld %f54,%f50,%f52 ! yy0 *= dtmp1; + fcmped %fcc0,HTHRESHOLD,%f52 ! if (yy0 >= HTHRESH) + fcmped %fcc1,LTHRESHOLD,%f52 ! yy0 = HTHRESH; + fmovdl %fcc0,HTHRESHOLD,%f52 ! if (yy0 <= LTHRESH) + fmovdg %fcc1,LTHRESHOLD,%f52 ! yy0 = LTHRESH; + fdtoi %f52,%f20 ! ind0 = (int) yy0; + st %f20,[%fp+tmp3] ! STORE ind0 + fitod %f20,%f58 ! dtmp0 = (double) ind0; + fpackfix %f20,%f20 ! dtmp1 = vis_fpackfix(dtmp1) + ld [%fp+tmp3],%g1 ! LOAD ind0 + fsubd %f52,%f58,%f46 ! y0 = yy0 - dtmp0; + fpadd32 %f20,%f24,%f56 ! dtmp1 += yisint0 + and %g1,255,%o4 ! ind0 &= 255; + sll %o4,3,%o3 ! ind0 <<= 3; + ldd [%l0+%o3],%f54 ! di0 = *(double*)((char*)__mt_constexp2f + ind0); + fmuld KB2,%f46,%f48 ! dtmp0 = KB2 * y0; + fpadd32 %f56,%f54,%f56 ! di0 = vis_fpadd32(di0,dtmp1); + faddd %f48,KB1,%f62 ! dtmp0 += KB1; + fmuld %f62,%f46,%f60 ! yy0 = dtmp0 * y0; + fmuld %f60,%f56,%f52 ! yy0 *= di0; + faddd %f52,%f56,%f58 ! yy0 += di0; + ba .specs_exit + fdtos %f58,%f4 ! ftmp0 = (float)yy0; + + .align 16 +.update0: + cmp counter,1 + ble .cont0 + nop + + add %i2,stridey,%o1 + stx %o2,[%fp+tmp_px] + + stx %o1,[%fp+tmp_py] + sub counter,1,counter + + st counter,[%fp+tmp_counter] + ba .cont0 + or %g0,1,counter + + .align 16 +.update1: + cmp counter,1 + ble .cont1 + nop + + add %i2,stridey,%o1 + stx %o2,[%fp+tmp_px] + + stx %o1,[%fp+tmp_py] + sub counter,1,counter + + st counter,[%fp+tmp_counter] + ba .cont1 + or %g0,1,counter + + .align 16 +.update2: + cmp counter,2 + ble .cont2 + nop + + add %i2,stridey,%o2 + stx %i3,[%fp+tmp_px] + + add %o2,stridey,%o2 + stx %o2,[%fp+tmp_py] + + sub counter,2,counter + st counter,[%fp+tmp_counter] + ba .cont2 + or %g0,2,counter + + .align 16 +.update3: + cmp counter,2 + ble .cont3 + nop + + add %i2,stridey,%o2 + stx %i3,[%fp+tmp_px] + + add %o2,stridey,%o2 + stx %o2,[%fp+tmp_py] + + sub counter,2,counter + st counter,[%fp+tmp_counter] + ba .cont3 + or %g0,2,counter + + .align 16 +.update4: + cmp counter,3 + ble .cont4 + nop + + sll stridey,1,%g5 + add %i2,stridey,%o3 + stx %o2,[%fp+tmp_px] + + add %o3,%g5,%o3 + stx %o3,[%fp+tmp_py] + + sub counter,3,counter + st counter,[%fp+tmp_counter] + ba .cont4 + or %g0,3,counter + + .align 16 +.update5: + cmp counter,3 + ble .cont5 + nop + + sll stridey,1,%g5 + add %i2,stridey,%o3 + stx %o2,[%fp+tmp_px] + + add %o3,%g5,%o3 + stx %o3,[%fp+tmp_py] + + sub counter,3,counter + st counter,[%fp+tmp_counter] + ba .cont5 + or %g0,3,counter + + .align 16 +.update6: + fzeros %f2 + cmp counter,1 + ble .cont6 + nop + + ld [%fp+tmp_counter],%g1 + + sub %o2,stridex,%o3 + stx %o4,[%fp+tmp_py] + + sub %o3,stridex,%o3 + add %g1,counter,counter + stx %o3,[%fp+tmp_px] + + sub counter,1,counter + st counter,[%fp+tmp_counter] + ba .cont6 + or %g0,1,counter + + .align 16 +.update7: + cmp counter,4 + ble .cont7 + nop + + sll stridey,1,%g1 + add %o4,stridey,%o0 + stx %o2,[%fp+tmp_px] + + add %o0,%g1,%o0 + stx %o0,[%fp+tmp_py] + + sub counter,4,counter + st counter,[%fp+tmp_counter] + ba .cont7 + or %g0,4,counter + + .align 16 +.update8: + cmp counter,4 + ble .cont8 + nop + + sll stridey,1,%g1 + add %o4,stridey,%o0 + stx %o2,[%fp+tmp_px] + + add %o0,%g1,%o0 + stx %o0,[%fp+tmp_py] + + sub counter,4,counter + st counter,[%fp+tmp_counter] + ba .cont8 + or %g0,4,counter + + .align 16 +.update9: + cmp counter,2 + ble .cont9 + fzeros %f16 + + ld [%fp+tmp_counter],%i3 + + sub %o2,stridex,%g1 + stx %i2,[%fp+tmp_py] + + sub %g1,stridex,%g1 + add %i3,counter,counter + stx %g1,[%fp+tmp_px] + + sub counter,2,counter + st counter,[%fp+tmp_counter] + ba .cont9 + or %g0,2,counter + + .align 16 +.update10: + cmp counter,5 + ble .cont10 + nop + + add %i2,stridey,%i1 + stx %i3,[%fp+tmp_px] + + add %i1,stridey,%i1 + stx %i1,[%fp+tmp_py] + + sub counter,5,counter + st counter,[%fp+tmp_counter] + ba .cont10 + or %g0,5,counter + + .align 16 +.update11: + cmp counter,5 + ble .cont11 + nop + + add %i2,stridey,%i1 + stx %i3,[%fp+tmp_px] + + add %i1,stridey,%i1 + stx %i1,[%fp+tmp_py] + + sub counter,5,counter + st counter,[%fp+tmp_counter] + ba .cont11 + or %g0,5,counter + + .align 16 +.update12: + fzeros %f0 + cmp counter,3 + ble .cont12 + nop + + ld [%fp+tmp_counter],%o2 + + sub %i3,stridex,%i1 + stx %i2,[%fp+tmp_py] + + sub %i1,stridex,%i1 + add %o2,counter,counter + stx %i1,[%fp+tmp_px] + + sub counter,3,counter + st counter,[%fp+tmp_counter] + ba .cont12 + or %g0,3,counter + + .align 16 +.update13: + cmp counter,3 + ble .cont13 + nop + + sll stridey,1,%g5 + add %i2,stridey,%o3 + stx %o2,[%fp+tmp_px] + + add %o3,%g5,%o3 + stx %o3,[%fp+tmp_py] + + sub counter,3,counter + st counter,[%fp+tmp_counter] + ba .cont13 + or %g0,3,counter + + .align 16 +.update14: + cmp counter,3 + ble .cont14 + nop + + sll stridey,1,%g5 + add %i2,stridey,%o3 + stx %o2,[%fp+tmp_px] + + add %o3,%g5,%o3 + stx %o3,[%fp+tmp_py] + + sub counter,3,counter + st counter,[%fp+tmp_counter] + ba .cont14 + or %g0,3,counter + + .align 16 +.update15: + cmp counter,1 + ble .cont15 + fzeros %f2 + + ld [%fp+tmp_counter],%g1 + + sub %o2,stridex,%o3 + stx %o4,[%fp+tmp_py] + + sub %o3,stridex,%o3 + add %g1,counter,counter + stx %o3,[%fp+tmp_px] + + sub counter,1,counter + st counter,[%fp+tmp_counter] + ba .cont15 + or %g0,1,counter + + .align 16 +.update16: + cmp counter,4 + ble .cont16 + nop + + sll stridey,1,%g1 + add %o4,stridey,%o0 + stx %o2,[%fp+tmp_px] + + add %o0,%g1,%o0 + stx %o0,[%fp+tmp_py] + + sub counter,4,counter + st counter,[%fp+tmp_counter] + ba .cont16 + or %g0,4,counter + + .align 16 +.update17: + cmp counter,4 + ble .cont17 + nop + + sll stridey,1,%g1 + add %o4,stridey,%o0 + stx %o2,[%fp+tmp_px] + + add %o0,%g1,%o0 + stx %o0,[%fp+tmp_py] + + sub counter,4,counter + st counter,[%fp+tmp_counter] + ba .cont17 + or %g0,4,counter + + .align 16 +.update18: + fzeros %f16 + cmp counter,2 + ble .cont18 + nop + + ld [%fp+tmp_counter],%i3 + + sub %o2,stridex,%g1 + stx %i2,[%fp+tmp_py] + + sub %g1,stridex,%g1 + add %i3,counter,counter + stx %g1,[%fp+tmp_px] + + sub counter,2,counter + st counter,[%fp+tmp_counter] + ba .cont18 + or %g0,2,counter + + .align 16 +.update19: + cmp counter,5 + ble .cont19 + nop + + add %i2,stridey,%i1 + stx %i3,[%fp+tmp_px] + + add %i1,stridey,%i1 + stx %i1,[%fp+tmp_py] + + sub counter,5,counter + st counter,[%fp+tmp_counter] + ba .cont19 + or %g0,5,counter + + .align 16 +.update20: + cmp counter,5 + ble .cont20 + nop + + add %i2,stridey,%i1 + stx %i3,[%fp+tmp_px] + + add %i1,stridey,%i1 + stx %i1,[%fp+tmp_py] + + sub counter,5,counter + st counter,[%fp+tmp_counter] + ba .cont20 + or %g0,5,counter + + .align 16 +.update21: + cmp counter,3 + ble .cont21 + fzeros %f0 + + ld [%fp+tmp_counter],%o2 + + sub %i3,stridex,%i1 + stx %i2,[%fp+tmp_py] + + sub %i1,stridex,%i1 + add %o2,counter,counter + stx %i1,[%fp+tmp_px] + + + sub counter,3,counter + st counter,[%fp+tmp_counter] + ba .cont21 + or %g0,3,counter + + .align 16 +.update22: + cmp counter,3 + ble .cont22 + fzeros %f2 + + ld [%fp+tmp_counter],%g1 + + sub %i3,stridex,%i2 + stx %i2,[%fp+tmp_px] + + add %g1,counter,counter + stx %o4,[%fp+tmp_py] + + sub counter,3,counter + st counter,[%fp+tmp_counter] + ba .cont22 + or %g0,3,counter + +.stridex_zero: + ld [%fp+tmp_counter],counter + + stx %i3,[%fp+tmp_py] + + cmp counter,0 + ble,pn %icc,.exit + lda [%i1]0x82,%i1 ! (Y0_2) ax0 = *px; + + and %i1,MASK_0x7fffffff,%i3 ! (Y0_2) exp0 = ax0 & 0x7fffffff; + sub %i3,%l6,%l6 + and %i1,MASK_0x007fffff,%g5 ! (Y0_2) ax0 &= 0x007fffff; + srl %i3,23,%o3 ! (Y0_2) exp0 >>= 23; + srl %l6,31,%l6 + st %l6,[%fp+tmp5] + add %g5,CONST_0x8000,%i3 ! (Y0_2) i0 = ax0 + 0x8000; + sethi %hi(0xffff0000),%l6 + sub %o3,127,%o3 ! (Y0_2) exp0 -= 127; + and %i3,%l6,%i3 ! (Y0_2) i0 &= 0xffff0000; + sll %o3,8,%o4 ! (Y0_2) exp0 <<= 8; + st %o4,[%fp+tmp3] ! (Y0_2) STORE exp0 + sra %i3,12,%o0 ! (Y0_2) ind0 = i0 >> 12; + sub %g5,%i3,%o4 ! (Y0_2) i0 = ax0 - i0; + st %o4,[%fp+tmp2] ! (Y0_2) STORE i0 + and %o0,-8,%g5 ! (Y0_2) ind0 &= -8; + ld [%fp+tmp2],%f14 ! (Y0_2) dtmp0 = (double) i0; + add %l2,%g5,%g1 ! (Y0_2) (char*)__mt_constlog4f + ind0 + ldd [%g1+8],%f48 ! (Y0_2) dtmp1 = *(double *)((char*)__mt_constlog4f + ind0 + 8); + fitod %f14,%f60 ! (Y0_2) dtmp0 = (double) i0; + fmuld %f60,%f48,%f48 ! (Y0_2) y0 = dtmp0 * dtmp1; + fmuld KA3,%f48,%f62 ! (Y0_2) dtmp0 = KA3 * y0; + faddd %f62,KA2,%f22 ! (Y0_2) dtmp0 += KA2; + fmuld %f22,%f48,%f26 ! (Y0_2) dtmp0 *= y0; + faddd %f26,KA1,%f50 ! (Y0_2) dtmp0 += KA1; + ld [%fp+tmp3],%f4 ! (Y0_2) dtmp1 = (double) exp0; + fitod %f4,%f26 ! (Y0_1) dtmp1 = (double) exp0; + fmuld %f50,%f48,%f50 ! (Y0_1) dtmp0 *= y0; + ldd [%l2+%g5],%f60 ! (Y0_1) dtmp0 = *(double *)((char*)__mt_constlog4f + ind0); + faddd %f50,KA0,%f58 ! (Y0_1) dtmp0 += KA0; + faddd %f60,%f26,%f26 ! (Y0_1) yy0 = dtmp0 + dtmp1; + fmuld %f58,%f48,%f48 ! (Y0_1) dtmp0 *= y0; + sub %l2,3200,%o4 + sub %l2,1152-600,%o3 + faddd %f26,%f48,%f46 ! (Y0_1) yy0 += dtmp0; + or %g0,%i5,%g1 + sethi %hi(0x7f800000),%o1 + +.xbegin: + ld [%fp+tmp_counter],counter + ldx [%fp+tmp_py],%o5 + st %g0,[%fp+tmp_counter] +.xbegin1: + subcc counter,1,counter + bneg,pn %icc,.exit + nop + + lda [%o5]0x82,%i5 ! (Y0_0) ay = py[0]; + + lda [%o5]0x82,%f5 ! (Y0_0) ftmp0 = py[0]; + + and %i5,MASK_0x7fffffff,%i3 ! (Y0_0) ay &= 0x7fffffff + + cmp %i3,%o1 + bge,pn %icc,.xspec + nop + + fstod %f5,%f52 ! (Y0_0) dtmp0 = (double)ftmp0; + + fmuld %f52,%f46,%f26 ! (Y0_0) yy0 = dtmp0 * yy; + add %o5,stridey,%o5 ! py += stridey + + lda [%o5]0x82,%i5 ! (Y1_0) ay = ((int*)py)[0]; + + lda [%o5]0x82,%f7 ! (Y1_0) ftmp0 = py[0]; + + and %i5,MASK_0x7fffffff,%i5 ! (Y1_0) ay &= 0x7fffffff + fcmped %fcc0,HTHRESHOLD,%f26 ! (Y0_0) if (yy0 >= HTHRESH) + + cmp %i5,%o1 + bge,pn %icc,.xupdate0 + nop + +.xcont0: + fstod %f7,%f48 ! (Y1_0) dtmp0 = (double)ftmp0; + + fcmped %fcc1,LTHRESHOLD,%f26 ! (Y0_1) if (yy0 <= LTHRESH) + + add %o5,stridey,%o5 ! py += stridey + fmuld %f48,%f46,%f28 ! (Y1_1) yy0 = dtmp0 * yy; + + lda [%o5]0x82,%i3 ! (Y0_0) ay = py[0]; + + lda [%o5]0x82,%f5 ! (Y0_0) ftmp0 = py[0]; + + and %i3,MASK_0x7fffffff,%i3 ! (Y0_0) ay &= 0x7fffffff + fmovdl %fcc0,HTHRESHOLD,%f26 ! (Y0_1) yy0 = HTHRESH; + + cmp %i3,%o1 + bge,pn %icc,.xupdate1 + fcmped %fcc2,HTHRESHOLD,%f28 ! (Y1_1) if (yy0 >= HTHRESH) +.xcont1: + fstod %f5,%f52 ! (Y0_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc1,LTHRESHOLD,%f26 ! (Y0_1) yy0 = LTHRESH; + + fcmped %fcc3,LTHRESHOLD,%f28 ! (Y1_1) if (yy0 <= LTHRESH) + + fmuld %f52,%f46,%f22 ! (Y0_0) yy0 = dtmp0 * yy; + + fdtoi %f26,%f0 ! (Y0_1) ii0 = (int) yy0; + + add %o5,stridey,%o5 ! py += stridey + st %f0,[%fp+tmp1] ! (Y0_1) STORE ii0 + + lda [%o5]0x82,%l7 ! (Y1_0) ay = ((int*)py)[0]; + + lda [%o5]0x82,%f7 ! (Y1_0) ftmp0 = py[0]; + fmovdl %fcc2,HTHRESHOLD,%f28 ! (Y1_1) yy0 = HTHRESH; + + and %l7,MASK_0x7fffffff,%l7 ! (Y1_0) ay &= 0x7fffffff + fcmped %fcc0,HTHRESHOLD,%f22 ! (Y0_0) if (yy0 >= HTHRESH) + + cmp %l7,%o1 + bge,pn %icc,.xupdate2 + nop +.xcont2: + fstod %f7,%f48 ! (Y1_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc3,LTHRESHOLD,%f28 ! (Y1_2) yy0 = LTHRESH; + + fcmped %fcc1,LTHRESHOLD,%f22 ! (Y0_1) if (yy0 <= LTHRESH) + + fitod %f0,%f52 ! (Y0_2) dtmp0 = (double)ii0; + + add %o5,stridey,%o5 ! py += stridey + fmuld %f48,%f46,%f24 ! (Y1_1) yy0 = dtmp0 * yy; + + fdtoi %f28,%f3 ! (Y1_2) ii0 = (int) yy0; + lda [%o5]0x82,%i3 ! (Y0_0) ay = py[0]; + + st %f3,[%fp+tmp0] ! (Y1_2) STORE ii0 + + fsubd %f26,%f52,%f40 ! (Y0_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%f5 ! (Y0_0) ftmp0 = py[0]; + + and %i3,MASK_0x7fffffff,%i3 ! (Y0_0) ay &= 0x7fffffff + fmovdl %fcc0,HTHRESHOLD,%f22 ! (Y0_1) yy0 = HTHRESH; + + cmp %i3,%o1 + bge,pn %icc,.xupdate3 + fcmped %fcc2,HTHRESHOLD,%f24 ! (Y1_1) if (yy0 >= HTHRESH) +.xcont3: + ld [%fp+tmp1],%i2 ! (Y0_2) LOAD ii0 + fmuld KB2,%f40,%f36 ! (Y0_2) dtmp0 = KB2 * y0; + fstod %f5,%f52 ! (Y0_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc1,LTHRESHOLD,%f22 ! (Y0_1) yy0 = LTHRESH; + + sra %i2,6,%l6 ! (Y0_2) i0 = ii0 >> 6; + and %i2,255,%l7 ! (Y0_2) ii0 &= 255; + fcmped %fcc3,LTHRESHOLD,%f24 ! (Y1_1) if (yy0 <= LTHRESH) + + fitod %f3,%f56 ! (Y1_2) dtmp0 = (double)ii0; + sll %l7,3,%o0 ! (Y0_2) ii0 <<= 3; + and %l6,-4,%g5 ! (Y0_2) i0 &= -4; + + faddd %f36,KB1,%f60 ! (Y0_2) dtmp0 += KB1; + fmuld %f52,%f46,%f26 ! (Y0_0) yy0 = dtmp0 * yy; + ld [%g5+%o3],%f10 ! (Y0_2) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0] + + fdtoi %f22,%f0 ! (Y0_1) ii0 = (int) yy0; + ldd [%o4+%o0],%f62 ! (Y0_2) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + add %o5,stridey,%o5 ! py += stridey + st %f0,[%fp+tmp1] ! (Y0_1) STORE ii0 + + fsubd %f28,%f56,%f56 ! (Y1_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%i5 ! (Y1_0) ay = ((int*)py)[0]; + + fmuld %f60,%f40,%f60 ! (Y0_2) yy0 = dtmp0 * y0; + fmovdl %fcc2,HTHRESHOLD,%f24 ! (Y1_1) yy0 = HTHRESH; + lda [%o5]0x82,%f7 ! (Y1_0) ftmp0 = py[0]; + + fmuld %f10,%f62,%f62 ! (Y0_2) di0 *= dtmp0; + ld [%fp+tmp0],%g5 ! (Y1_2) LOAD ii0 + and %i5,MASK_0x7fffffff,%i5 ! (Y1_0) ay &= 0x7fffffff + fcmped %fcc0,HTHRESHOLD,%f26 ! (Y0_0) if (yy0 >= HTHRESH) + + cmp %i5,%o1 + bge,pn %icc,.xupdate4 +.xcont4: + fmuld KB2,%f56,%f58 ! (Y1_2) dtmp0 = KB2 * y0; + fstod %f7,%f48 ! (Y1_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc3,LTHRESHOLD,%f24 ! (Y1_2) yy0 = LTHRESH; + sra %g5,6,%i0 ! (Y1_3) i0 = ii0 >> 6; + and %g5,255,%i1 ! (Y1_3) ii0 &= 255; + fmuld %f60,%f62,%f40 ! (Y0_3) dtmp0 = yy0 * di0; + + fcmped %fcc1,LTHRESHOLD,%f26 ! (Y0_1) if (yy0 <= LTHRESH) + sll %i1,3,%i3 ! (Y1_3) ii0 <<= 3; + and %i0,-4,%i0 ! (Y1_3) i0 &= -4; + + fitod %f0,%f52 ! (Y0_2) dtmp0 = (double)ii0; + ld [%i0+%o3],%f10 ! (Y1_3) di0 = ((double*)((char*)(__mt_constexp2fb + 150) + i0))[0]; + + faddd %f58,KB1,%f58 ! (Y1_3) dtmp0 += KB1; + add %o5,stridey,%o5 ! py += stridey + ldd [%o4+%i3],%f18 ! (Y1_3) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + fmuld %f48,%f46,%f28 ! (Y1_1) yy0 = dtmp0 * yy; + + fdtoi %f24,%f3 ! (Y1_2) ii0 = (int) yy0; + lda [%o5]0x82,%i3 ! (Y0_0) ay = py[0]; + + faddd %f40,%f62,%f60 ! (Y0_3) dtmp0 += di0; + st %f3,[%fp+tmp0] ! (Y1_2) STORE ii0 + + fsubd %f22,%f52,%f40 ! (Y0_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%f5 ! (Y0_0) ftmp0 = py[0]; + + fmuld %f58,%f56,%f56 ! (Y1_3) yy0 = dtmp0 * y0; + and %i3,MASK_0x7fffffff,%i3 ! (Y0_0) ay &= 0x7fffffff + fmovdl %fcc0,HTHRESHOLD,%f26 ! (Y0_1) yy0 = HTHRESH; + + fmuld %f10,%f18,%f50 ! (Y1_3) di0 *= dtmp0; + cmp %i3,%o1 + bge,pn %icc,.xupdate5 + fcmped %fcc2,HTHRESHOLD,%f28 ! (Y1_1) if (yy0 >= HTHRESH) +.xcont5: + fdtos %f60,%f1 ! (Y0_3) ftmp0 = (float)dtmp0; + add %g1,stridez,%i3 ! pz += stridez + st %f1,[%g1] ! (Y0_3) pz[0] = ftmp0; + + subcc counter,1,counter + bneg,pn %icc,.xbegin + or %g0,%i3,%g1 + + ld [%fp+tmp1],%i2 ! (Y0_2) LOAD ii0 + fmuld KB2,%f40,%f36 ! (Y0_2) dtmp0 = KB2 * y0; + fstod %f5,%f52 ! (Y0_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc1,LTHRESHOLD,%f26 ! (Y0_1) yy0 = LTHRESH; + + fmuld %f56,%f50,%f58 ! (Y1_3) dtmp0 = yy0 * di0; + sra %i2,6,%l6 ! (Y0_2) i0 = ii0 >> 6; + and %i2,255,%l7 ! (Y0_2) ii0 &= 255; + fcmped %fcc3,LTHRESHOLD,%f28 ! (Y1_1) if (yy0 <= LTHRESH) + + fitod %f3,%f56 ! (Y1_2) dtmp0 = (double)ii0; + sll %l7,3,%o0 ! (Y0_2) ii0 <<= 3; + and %l6,-4,%g5 ! (Y0_2) i0 &= -4; + + faddd %f36,KB1,%f60 ! (Y0_2) dtmp0 += KB1; + fmuld %f52,%f46,%f22 ! (Y0_0) yy0 = dtmp0 * yy; + ld [%g5+%o3],%f10 ! (Y0_2) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0] + + fdtoi %f26,%f0 ! (Y0_1) ii0 = (int) yy0; + ldd [%o4+%o0],%f62 ! (Y0_2) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + faddd %f58,%f50,%f58 ! (Y1_3) dtmp0 += di0; + add %o5,stridey,%o5 ! py += stridey + st %f0,[%fp+tmp1] ! (Y0_1) STORE ii0 + + fsubd %f24,%f56,%f56 ! (Y1_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%l7 ! (Y1_0) ay = ((int*)py)[0]; + + fmuld %f60,%f40,%f60 ! (Y0_2) yy0 = dtmp0 * y0; + add %i3,stridez,%i5 ! pz += stridez + lda [%o5]0x82,%f7 ! (Y1_0) ftmp0 = py[0]; + fmovdl %fcc2,HTHRESHOLD,%f28 ! (Y1_1) yy0 = HTHRESH; + + fmuld %f10,%f62,%f62 ! (Y0_2) di0 *= dtmp0; + and %l7,MASK_0x7fffffff,%l7 ! (Y1_0) ay &= 0x7fffffff + ld [%fp+tmp0],%g5 ! (Y1_2) LOAD ii0 + fcmped %fcc0,HTHRESHOLD,%f22 ! (Y0_0) if (yy0 >= HTHRESH) + + fdtos %f58,%f9 ! (Y1_3) ftmp0 = (float)dtmp0; + st %f9,[%i3] ! (Y1_3) pz[0] = ftmp0; + cmp %l7,%o1 + bge,pn %icc,.xupdate6 + +.xcont6: + fmuld KB2,%f56,%f58 ! (Y1_2) dtmp0 = KB2 * y0; + fstod %f7,%f48 ! (Y1_0) dtmp0 = (double)ftmp0; + + cmp counter,8 + bl,pn %icc,.xtail + nop + + ba .xmain_loop + nop + + .align 16 +.xmain_loop: + fmovdg %fcc3,LTHRESHOLD,%f28 ! (Y1_2) yy0 = LTHRESH; + sra %g5,6,%i0 ! (Y1_3) i0 = ii0 >> 6; + and %g5,255,%i1 ! (Y1_3) ii0 &= 255; + fmuld %f60,%f62,%f40 ! (Y0_3) dtmp0 = yy0 * di0; + + fcmped %fcc1,LTHRESHOLD,%f22 ! (Y0_1) if (yy0 <= LTHRESH) + sll %i1,3,%i3 ! (Y1_3) ii0 <<= 3; + and %i0,-4,%i0 ! (Y1_3) i0 &= -4; + + fitod %f0,%f52 ! (Y0_2) dtmp0 = (double)ii0; + sub counter,4,counter + ld [%i0+%o3],%f10 ! (Y1_3) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0]; + + faddd %f58,KB1,%f58 ! (Y1_3) dtmp0 += KB1; + add %o5,stridey,%o5 ! py += stridey + ldd [%o4+%i3],%f18 ! (Y1_3) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + fmuld %f48,%f46,%f24 ! (Y1_1) yy0 = dtmp0 * yy; + + fdtoi %f28,%f3 ! (Y1_2) ii0 = (int) yy0; + lda [%o5]0x82,%i3 ! (Y0_0) ay = py[0]; + + faddd %f40,%f62,%f60 ! (Y0_3) dtmp0 += di0; + st %f3,[%fp+tmp0] ! (Y1_2) STORE ii0 + + fsubd %f26,%f52,%f40 ! (Y0_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%f5 ! (Y0_0) ftmp0 = py[0]; + + fmuld %f58,%f56,%f56 ! (Y1_3) yy0 = dtmp0 * y0; + and %i3,MASK_0x7fffffff,%i3 ! (Y0_0) ay &= 0x7fffffff + fmovdl %fcc0,HTHRESHOLD,%f22 ! (Y0_1) yy0 = HTHRESH; + + fmuld %f10,%f18,%f50 ! (Y1_3) di0 *= dtmp0; + cmp %i3,%o1 + bge,pn %icc,.xupdate7 + fcmped %fcc2,HTHRESHOLD,%f24 ! (Y1_1) if (yy0 >= HTHRESH) +.xcont7: + fdtos %f60,%f1 ! (Y0_3) ftmp0 = (float)dtmp0; + add %i5,stridez,%i3 ! pz += stridez + st %f1,[%i5] ! (Y0_3) pz[0] = ftmp0; + + ld [%fp+tmp1],%i2 ! (Y0_2) LOAD ii0 + fmuld KB2,%f40,%f36 ! (Y0_2) dtmp0 = KB2 * y0; + fstod %f5,%f52 ! (Y0_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc1,LTHRESHOLD,%f22 ! (Y0_1) yy0 = LTHRESH; + + fmuld %f56,%f50,%f58 ! (Y1_3) dtmp0 = yy0 * di0; + sra %i2,6,%l6 ! (Y0_2) i0 = ii0 >> 6; + and %i2,255,%l7 ! (Y0_2) ii0 &= 255; + fcmped %fcc3,LTHRESHOLD,%f24 ! (Y1_1) if (yy0 <= LTHRESH) + + fitod %f3,%f56 ! (Y1_2) dtmp0 = (double)ii0; + sll %l7,3,%o0 ! (Y0_2) ii0 <<= 3; + and %l6,-4,%g5 ! (Y0_2) i0 &= -4; + + faddd %f36,KB1,%f60 ! (Y0_2) dtmp0 += KB1; + fmuld %f52,%f46,%f26 ! (Y0_0) yy0 = dtmp0 * yy; + ld [%g5+%o3],%f10 ! (Y0_2) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0] + + fdtoi %f22,%f0 ! (Y0_1) ii0 = (int) yy0; + ldd [%o4+%o0],%f62 ! (Y0_2) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + faddd %f58,%f50,%f58 ! (Y1_3) dtmp0 += di0; + add %o5,stridey,%o5 ! py += stridey + st %f0,[%fp+tmp1] ! (Y0_1) STORE ii0 + + fsubd %f28,%f56,%f56 ! (Y1_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%i5 ! (Y1_0) ay = ((int*)py)[0]; + + fmuld %f60,%f40,%f60 ! (Y0_2) yy0 = dtmp0 * y0; + fmovdl %fcc2,HTHRESHOLD,%f24 ! (Y1_1) yy0 = HTHRESH; + lda [%o5]0x82,%f7 ! (Y1_0) ftmp0 = py[0]; + + fmuld %f10,%f62,%f62 ! (Y0_2) di0 *= dtmp0; + ld [%fp+tmp0],%g5 ! (Y1_2) LOAD ii0 + and %i5,MASK_0x7fffffff,%i5 ! (Y1_0) ay &= 0x7fffffff + fcmped %fcc0,HTHRESHOLD,%f26 ! (Y0_0) if (yy0 >= HTHRESH) + + fdtos %f58,%f9 ! (Y1_3) ftmp0 = (float)dtmp0; + cmp %i5,%o1 + bge,pn %icc,.xupdate8 + +.xcont8: + fmuld KB2,%f56,%f58 ! (Y1_2) dtmp0 = KB2 * y0; + add %i3,stridez,%i5 ! pz += stridez + st %f9,[%i3] ! (Y1_3) pz[0] = ftmp0; + fstod %f7,%f48 ! (Y1_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc3,LTHRESHOLD,%f24 ! (Y1_2) yy0 = LTHRESH; + sra %g5,6,%i0 ! (Y1_3) i0 = ii0 >> 6; + and %g5,255,%i1 ! (Y1_3) ii0 &= 255; + fmuld %f60,%f62,%f40 ! (Y0_3) dtmp0 = yy0 * di0; + + fcmped %fcc1,LTHRESHOLD,%f26 ! (Y0_1) if (yy0 <= LTHRESH) + sll %i1,3,%i3 ! (Y1_3) ii0 <<= 3; + and %i0,-4,%i0 ! (Y1_3) i0 &= -4; + + fitod %f0,%f52 ! (Y0_2) dtmp0 = (double)ii0; + ld [%i0+%o3],%f10 ! (Y1_3) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0]; + + faddd %f58,KB1,%f58 ! (Y1_3) dtmp0 += KB1; + add %o5,stridey,%o5 ! py += stridey + ldd [%o4+%i3],%f18 ! (Y1_3) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + fmuld %f48,%f46,%f28 ! (Y1_1) yy0 = dtmp0 * yy; + + fdtoi %f24,%f3 ! (Y1_2) ii0 = (int) yy0; + lda [%o5]0x82,%i3 ! (Y0_0) ay = py[0]; + + faddd %f40,%f62,%f60 ! (Y0_3) dtmp0 += di0; + st %f3,[%fp+tmp0] ! (Y1_2) STORE ii0 + + fsubd %f22,%f52,%f40 ! (Y0_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%f5 ! (Y0_0) ftmp0 = py[0]; + + fmuld %f58,%f56,%f56 ! (Y1_3) yy0 = dtmp0 * y0; + and %i3,MASK_0x7fffffff,%i3 ! (Y0_0) ay &= 0x7fffffff + fmovdl %fcc0,HTHRESHOLD,%f26 ! (Y0_1) yy0 = HTHRESH; + + fmuld %f10,%f18,%f50 ! (Y1_3) di0 *= dtmp0; + cmp %i3,%o1 + bge,pn %icc,.xupdate9 + fcmped %fcc2,HTHRESHOLD,%f28 ! (Y1_1) if (yy0 >= HTHRESH) +.xcont9: + fdtos %f60,%f1 ! (Y0_3) ftmp0 = (float)dtmp0; + add %i5,stridez,%i3 ! pz += stridez + st %f1,[%i5] ! (Y0_3) pz[0] = ftmp0; + + ld [%fp+tmp1],%i2 ! (Y0_2) LOAD ii0 + fmuld KB2,%f40,%f36 ! (Y0_2) dtmp0 = KB2 * y0; + fstod %f5,%f52 ! (Y0_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc1,LTHRESHOLD,%f26 ! (Y0_1) yy0 = LTHRESH; + + fmuld %f56,%f50,%f58 ! (Y1_3) dtmp0 = yy0 * di0; + sra %i2,6,%l6 ! (Y0_2) i0 = ii0 >> 6; + and %i2,255,%l7 ! (Y0_2) ii0 &= 255; + fcmped %fcc3,LTHRESHOLD,%f28 ! (Y1_1) if (yy0 <= LTHRESH) + + fitod %f3,%f56 ! (Y1_2) dtmp0 = (double)ii0; + sll %l7,3,%o0 ! (Y0_2) ii0 <<= 3; + and %l6,-4,%g5 ! (Y0_2) i0 &= -4; + + faddd %f36,KB1,%f60 ! (Y0_2) dtmp0 += KB1; + fmuld %f52,%f46,%f22 ! (Y0_0) yy0 = dtmp0 * yy; + ld [%g5+%o3],%f10 ! (Y0_2) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0] + + fdtoi %f26,%f0 ! (Y0_1) ii0 = (int) yy0; + ldd [%o4+%o0],%f62 ! (Y0_2) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + faddd %f58,%f50,%f58 ! (Y1_3) dtmp0 += di0; + add %o5,stridey,%o5 ! py += stridey + st %f0,[%fp+tmp1] ! (Y0_1) STORE ii0 + + fsubd %f24,%f56,%f56 ! (Y1_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%l7 ! (Y1_0) ay = ((int*)py)[0]; + + fmuld %f60,%f40,%f60 ! (Y0_2) yy0 = dtmp0 * y0; + add %i3,stridez,%i5 ! pz += stridez + lda [%o5]0x82,%f7 ! (Y1_0) ftmp0 = py[0]; + fmovdl %fcc2,HTHRESHOLD,%f28 ! (Y1_1) yy0 = HTHRESH; + + fmuld %f10,%f62,%f62 ! (Y0_2) di0 *= dtmp0; + and %l7,MASK_0x7fffffff,%l7 ! (Y1_0) ay &= 0x7fffffff + ld [%fp+tmp0],%g5 ! (Y1_2) LOAD ii0 + fcmped %fcc0,HTHRESHOLD,%f22 ! (Y0_0) if (yy0 >= HTHRESH) + + fdtos %f58,%f9 ! (Y1_3) ftmp0 = (float)dtmp0; + st %f9,[%i3] ! (Y1_3) pz[0] = ftmp0; + cmp %l7,%o1 + bge,pn %icc,.xupdate10 +.xcont10: + fmuld KB2,%f56,%f58 ! (Y1_2) dtmp0 = KB2 * y0; + cmp counter,4 + bge,pt %icc,.xmain_loop + fstod %f7,%f48 ! (Y1_0) dtmp0 = (double)ftmp0; + +.xtail: + subcc counter,1,counter + bneg,pn %icc,.xbegin + or %g0,%i5,%g1 + + fmovdg %fcc3,LTHRESHOLD,%f28 ! (Y1_2) yy0 = LTHRESH; + sra %g5,6,%i0 ! (Y1_3) i0 = ii0 >> 6; + and %g5,255,%i1 ! (Y1_3) ii0 &= 255; + fmuld %f60,%f62,%f40 ! (Y0_3) dtmp0 = yy0 * di0; + + fcmped %fcc1,LTHRESHOLD,%f22 ! (Y0_1) if (yy0 <= LTHRESH) + sll %i1,3,%i3 ! (Y1_3) ii0 <<= 3; + and %i0,-4,%i0 ! (Y1_3) i0 &= -4; + + fitod %f0,%f52 ! (Y0_2) dtmp0 = (double)ii0; + ld [%i0+%o3],%f10 ! (Y1_3) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0]; + + faddd %f58,KB1,%f58 ! (Y1_3) dtmp0 += KB1; + add %o5,stridey,%o5 ! py += stridey + ldd [%o4+%i3],%f18 ! (Y1_3) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + fmuld %f48,%f46,%f24 ! (Y1_1) yy0 = dtmp0 * yy; + + fdtoi %f28,%f3 ! (Y1_2) ii0 = (int) yy0; + lda [%o5]0x82,%i3 ! (Y0_0) ay = py[0]; + + faddd %f40,%f62,%f60 ! (Y0_3) dtmp0 += di0; + st %f3,[%fp+tmp0] ! (Y1_2) STORE ii0 + + fsubd %f26,%f52,%f40 ! (Y0_2) y0 = yy0 - dtmp0; + lda [%o5]0x82,%f5 ! (Y0_0) ftmp0 = py[0]; + + fmuld %f58,%f56,%f56 ! (Y1_3) yy0 = dtmp0 * y0; + and %i3,MASK_0x7fffffff,%i3 ! (Y0_0) ay &= 0x7fffffff + fmovdl %fcc0,HTHRESHOLD,%f22 ! (Y0_1) yy0 = HTHRESH; + + fmuld %f10,%f18,%f50 ! (Y1_3) di0 *= dtmp0; + cmp %i3,%o1 + bge,pn %icc,.xupdate11 + fcmped %fcc2,HTHRESHOLD,%f24 ! (Y1_1) if (yy0 >= HTHRESH) +.xcont11: + fdtos %f60,%f1 ! (Y0_3) ftmp0 = (float)dtmp0; + add %i5,stridez,%i3 ! pz += stridez + st %f1,[%i5] ! (Y0_3) pz[0] = ftmp0; + + subcc counter,1,counter + bneg,pn %icc,.xbegin + or %g0,%i3,%g1 + + ld [%fp+tmp1],%i2 ! (Y0_2) LOAD ii0 + fmuld KB2,%f40,%f36 ! (Y0_2) dtmp0 = KB2 * y0; + fstod %f5,%f52 ! (Y0_0) dtmp0 = (double)ftmp0; + + fmovdg %fcc1,LTHRESHOLD,%f22 ! (Y0_1) yy0 = LTHRESH; + + fmuld %f56,%f50,%f58 ! (Y1_3) dtmp0 = yy0 * di0; + sra %i2,6,%l6 ! (Y0_2) i0 = ii0 >> 6; + and %i2,255,%l7 ! (Y0_2) ii0 &= 255; + fcmped %fcc3,LTHRESHOLD,%f24 ! (Y1_1) if (yy0 <= LTHRESH) + + fitod %f3,%f56 ! (Y1_2) dtmp0 = (double)ii0; + sll %l7,3,%o0 ! (Y0_2) ii0 <<= 3; + and %l6,-4,%g5 ! (Y0_2) i0 &= -4; + + faddd %f36,KB1,%f60 ! (Y0_2) dtmp0 += KB1; + fmuld %f52,%f46,%f26 ! (Y0_0) yy0 = dtmp0 * yy; + ld [%g5+%o3],%f10 ! (Y0_2) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0] + + fdtoi %f22,%f0 ! (Y0_1) ii0 = (int) yy0; + ldd [%o4+%o0],%f62 ! (Y0_2) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + faddd %f58,%f50,%f58 ! (Y1_3) dtmp0 += di0; + st %f0,[%fp+tmp1] ! (Y0_1) STORE ii0 + + fsubd %f28,%f56,%f56 ! (Y1_2) y0 = yy0 - dtmp0; + + fmuld %f60,%f40,%f60 ! (Y0_2) yy0 = dtmp0 * y0; + fmovdl %fcc2,HTHRESHOLD,%f24 ! (Y1_1) yy0 = HTHRESH; + + fmuld %f10,%f62,%f62 ! (Y0_2) di0 *= dtmp0; + ld [%fp+tmp0],%g5 ! (Y1_2) LOAD ii0 + fcmped %fcc0,HTHRESHOLD,%f26 ! (Y0_0) if (yy0 >= HTHRESH) + + fdtos %f58,%f9 ! (Y1_3) ftmp0 = (float)dtmp0; + + fmuld KB2,%f56,%f58 ! (Y1_2) dtmp0 = KB2 * y0; + add %i3,stridez,%i5 ! pz += stridez + st %f9,[%i3] ! (Y1_3) pz[0] = ftmp0; + + subcc counter,1,counter + bneg,pn %icc,.xbegin + or %g0,%i5,%g1 + + fmovdg %fcc3,LTHRESHOLD,%f24 ! (Y1_2) yy0 = LTHRESH; + sra %g5,6,%i0 ! (Y1_3) i0 = ii0 >> 6; + and %g5,255,%i1 ! (Y1_3) ii0 &= 255; + fmuld %f60,%f62,%f40 ! (Y0_3) dtmp0 = yy0 * di0; + + fcmped %fcc1,LTHRESHOLD,%f26 ! (Y0_1) if (yy0 <= LTHRESH) + sll %i1,3,%i3 ! (Y1_3) ii0 <<= 3; + and %i0,-4,%i0 ! (Y1_3) i0 &= -4; + + fitod %f0,%f52 ! (Y0_2) dtmp0 = (double)ii0; + ld [%i0+%o3],%f10 ! (Y1_3) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0]; + + faddd %f58,KB1,%f58 ! (Y1_3) dtmp0 += KB1; + ldd [%o4+%i3],%f18 ! (Y1_3) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + fdtoi %f24,%f3 ! (Y1_2) ii0 = (int) yy0; + + faddd %f40,%f62,%f60 ! (Y0_3) dtmp0 += di0; + st %f3,[%fp+tmp0] ! (Y1_2) STORE ii0 + + fsubd %f22,%f52,%f40 ! (Y0_2) y0 = yy0 - dtmp0; + + fmuld %f58,%f56,%f56 ! (Y1_3) yy0 = dtmp0 * y0; + fmovdl %fcc0,HTHRESHOLD,%f26 ! (Y0_1) yy0 = HTHRESH; + + fmuld %f10,%f18,%f50 ! (Y1_3) di0 *= dtmp0; + + fdtos %f60,%f1 ! (Y0_3) ftmp0 = (float)dtmp0; + add %i5,stridez,%i3 ! pz += stridez + st %f1,[%i5] ! (Y0_3) pz[0] = ftmp0; + + subcc counter,1,counter + bneg,pn %icc,.xbegin + or %g0,%i3,%g1 + + ld [%fp+tmp1],%i2 ! (Y0_2) LOAD ii0 + fmuld KB2,%f40,%f36 ! (Y0_2) dtmp0 = KB2 * y0; + + fmovdg %fcc1,LTHRESHOLD,%f26 ! (Y0_1) yy0 = LTHRESH; + + fmuld %f56,%f50,%f58 ! (Y1_3) dtmp0 = yy0 * di0; + sra %i2,6,%l6 ! (Y0_2) i0 = ii0 >> 6; + and %i2,255,%l7 ! (Y0_2) ii0 &= 255; + + fitod %f3,%f56 ! (Y1_2) dtmp0 = (double)ii0; + sll %l7,3,%o0 ! (Y0_2) ii0 <<= 3; + and %l6,-4,%g5 ! (Y0_2) i0 &= -4; + + faddd %f36,KB1,%f60 ! (Y0_2) dtmp0 += KB1; + ld [%g5+%o3],%f10 ! (Y0_2) di0 = ((double*)((char*)(__mt_constexp2fb + 150) + i0))[0]; + + fdtoi %f26,%f0 ! (Y0_1) ii0 = (int) yy0; + ldd [%o4+%o0],%f62 ! (Y0_2) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + faddd %f58,%f50,%f58 ! (Y1_3) dtmp0 += di0; + st %f0,[%fp+tmp1] ! (Y0_1) STORE ii0 + + fsubd %f24,%f56,%f56 ! (Y1_2) y0 = yy0 - dtmp0; + + fmuld %f60,%f40,%f60 ! (Y0_2) yy0 = dtmp0 * y0; + add %i3,stridez,%i5 ! pz += stridez + + fmuld %f10,%f62,%f62 ! (Y0_2) di0 *= dtmp0; + ld [%fp+tmp0],%g5 ! (Y1_2) LOAD ii0 + + fdtos %f58,%f9 ! (Y1_3) ftmp0 = (float)dtmp0; + st %f9,[%i3] ! (Y1_3) pz[0] = ftmp0; + + subcc counter,1,counter + bneg,pn %icc,.xbegin + or %g0,%i5,%g1 + + fmuld KB2,%f56,%f58 ! (Y1_2) dtmp0 = KB2 * y0; + + sra %g5,6,%i0 ! (Y1_3) i0 = ii0 >> 6; + and %g5,255,%i1 ! (Y1_3) ii0 &= 255; + fmuld %f60,%f62,%f40 ! (Y0_3) dtmp0 = yy0 * di0; + + sll %i1,3,%i3 ! (Y1_3) ii0 <<= 3; + and %i0,-4,%i0 ! (Y1_3) i0 &= -4; + + fitod %f0,%f52 ! (Y0_2) dtmp0 = (double)ii0; + ld [%i0+%o3],%f10 ! (Y1_3) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0]; + + faddd %f58,KB1,%f58 ! (Y1_3) dtmp0 += KB1; + ldd [%o4+%i3],%f18 ! (Y1_3) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + faddd %f40,%f62,%f60 ! (Y0_3) dtmp0 += di0; + + fsubd %f26,%f52,%f40 ! (Y0_2) y0 = yy0 - dtmp0; + + fmuld %f58,%f56,%f56 ! (Y1_3) yy0 = dtmp0 * y0; + + fmuld %f10,%f18,%f50 ! (Y1_3) di0 *= dtmp0; + + fdtos %f60,%f1 ! (Y0_3) ftmp0 = (float)dtmp0; + add %i5,stridez,%i3 ! pz += stridez + st %f1,[%i5] ! (Y0_3) pz[0] = ftmp0; + + subcc counter,1,counter + bneg,pn %icc,.xbegin + or %g0,%i3,%g1 + + ld [%fp+tmp1],%i2 ! (Y0_2) LOAD ii0 + fmuld KB2,%f40,%f36 ! (Y0_2) dtmp0 = KB2 * y0; + + fmuld %f56,%f50,%f58 ! (Y1_3) dtmp0 = yy0 * di0; + sra %i2,6,%l6 ! (Y0_2) i0 = ii0 >> 6; + and %i2,255,%l7 ! (Y0_2) ii0 &= 255; + + sll %l7,3,%o0 ! (Y0_2) ii0 <<= 3; + and %l6,-4,%g5 ! (Y0_2) i0 &= -4; + + faddd %f36,KB1,%f60 ! (Y0_2) dtmp0 += KB1; + ld [%g5+%o3],%f10 ! (Y0_2) di0 = ((double*)((char*)(__mt_constexp2fb + 150 ) + i0))[0] + + ldd [%o4+%o0],%f62 ! (Y0_2) dtmp0 = ((double*)((char*)__mt_constexp2fa + ii0))[0]; + + faddd %f58,%f50,%f58 ! (Y1_3) dtmp0 += di0; + + fmuld %f60,%f40,%f60 ! (Y0_2) yy0 = dtmp0 * y0; + + fmuld %f10,%f62,%f62 ! (Y0_2) di0 *= dtmp0; + + fdtos %f58,%f9 ! (Y1_3) ftmp0 = (float)dtmp0; + add %i3,stridez,%i5 ! pz += stridez + st %f9,[%i3] ! (Y1_3) pz[0] = ftmp0; + + subcc counter,1,counter + bneg,pn %icc,.xbegin + or %g0,%i5,%g1 + + fmuld %f60,%f62,%f40 ! (Y0_3) dtmp0 = yy0 * di0; + + faddd %f40,%f62,%f60 ! (Y0_3) dtmp0 += di0; + + fdtos %f60,%f1 ! (Y0_3) ftmp0 = (float)dtmp0; + add %i5,stridez,%i3 ! pz += stridez + st %f1,[%i5] ! (Y0_3) pz[0] = ftmp0; + + ba .xbegin + or %g0,%i3,%g1 + +.xspec: + bg,a,pn %icc,.yisnan ! if (ay > 0x7f800000) /* |Y| = Nan */ + ld [%o5],%f8 ! fy = *py; + + ld [%fp+tmp5],%l6 ! LOAD (ax-0x3f800000)<<63 + srl %i5,31,%i5 ! uy >> 31 + + cmp %l6,%i5 ! if((ax < 0x3f800000) != (uy >> 31)) + be,a,pn %icc,.xspec_exit ! if((ax < 0x3f800000) != (uy >> 31)) + st %i3,[%g1] ! fy = *(float*)&ay; + + st %g0,[%g1] ! fy = ZERO + add %g1,stridez,%g1 + ba .xbegin1 + add %o5,stridey,%o5 + +.yisnan: + fmuls %f8,%f8,%f8 ! fy = *py * *py; /* |Y| = Nan */ + st %f8,[%g1] + +.xspec_exit: + add %g1,stridez,%g1 + ba .xbegin1 + add %o5,stridey,%o5 + + .align 16 +.xupdate0: + cmp counter,0 + ble .xcont0 + fzeros %f7 + + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont0 + or %g0,0,counter + + .align 16 +.xupdate1: + cmp counter,1 + ble .xcont1 + fzeros %f5 + + sub counter,1,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont1 + or %g0,1,counter + + .align 16 +.xupdate2: + cmp counter,2 + ble .xcont2 + fzeros %f7 + + sub counter,2,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont2 + or %g0,2,counter + + .align 16 +.xupdate3: + cmp counter,3 + ble .xcont3 + fzeros %f5 + + sub counter,3,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont3 + or %g0,3,counter + + .align 16 +.xupdate4: + cmp counter,4 + ble .xcont4 + fzeros %f7 + + sub counter,4,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont4 + or %g0,4,counter + + .align 16 +.xupdate5: + cmp counter,5 + ble .xcont5 + fzeros %f5 + + sub counter,5,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont5 + or %g0,5,counter + + .align 16 +.xupdate6: + cmp counter,5 + ble .xcont6 + fzeros %f7 + + sub counter,5,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont6 + or %g0,5,counter + + .align 16 +.xupdate7: + cmp counter,2 + ble .xcont7 + fzeros %f5 + + sub counter,2,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont7 + or %g0,2,counter + + .align 16 +.xupdate8: + cmp counter,3 + ble .xcont8 + fzeros %f7 + + sub counter,3,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont8 + or %g0,3,counter + + .align 16 +.xupdate9: + cmp counter,4 + ble .xcont9 + fzeros %f5 + + sub counter,4,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont9 + or %g0,4,counter + + .align 16 +.xupdate10: + cmp counter,5 + ble .xcont10 + fzeros %f7 + + sub counter,5,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont10 + or %g0,5,counter + + .align 16 +.xupdate11: + cmp counter,5 + ble .xcont11 + fzeros %f5 + + sub counter,5,counter + stx %o5,[%fp+tmp_py] + + st counter,[%fp+tmp_counter] + ba .xcont11 + or %g0,5,counter + + SET_SIZE(__vpowf) + |