📄 libm_tan.s
字号:
//// -cot( r + c ) =// / (1/[sin(B)*cos(B)]) * tan(x)// sgn_r * | -cot(B) + -------------------------------- +// \ tan(B) + tan(x)// \ // CORR |// ///// The values of tan(B), cot(B) and 1/(sin(B)*cos(B)) are// calculated beforehand and stored in a table. Specifically,// the table values are//// tan(B) as T_hi + T_lo;// cot(B) as C_hi + C_lo;// 1/[sin(B)*cos(B)] as SC_inv//// T_hi, C_hi are in double-precision memory format;// T_lo, C_lo are in single-precision memory format;// SC_inv is in extended-precision memory format.//// The value of tan(x) will be approximated by a short polynomial of// the form//// tan(x) as x + x * P, where// P = x^2 * (P2_1 + x^2 * (P2_2 + x^2 * P2_3))//// Because |x| <= 2^(-7), tan(B) + x approximates tan(B) + tan(x)// to a relative accuracy better than 2^(-18). Thus, a good// initial guess of 1/( tan(B) + tan(x) ) to initiate the iterative// division is://// 1/(tan(B) + tan(x)) is approximately// 1/(tan(B) + x) is// cot(B)/(1 + x*cot(B)) is approximately// C_hi / ( 1 + C_hi * x ) is approximately//// C_hi * [ 1 - (C_hi * x) + (C_hi * x)^2 ]//// The calculation of -cot(r+c) therefore proceed as follows://// Cx := C_hi * x// xsq := x * x//// V_hi := C_hi*(1 - Cx*(1 - Cx))// P := xsq * (P1_1 + xsq*(P1_2 + xsq*P1_3))// ...V_hi serves as an initial guess of 1/(tan(B) + tan(x))// ...good to about 18 bits of accuracy//// tanx := x + x*P// D := T_hi + tanx// ...D is a double precision denominator: tan(B) + tan(x)//// V_hi := V_hi + V_hi*(1 - V_hi*D)// ....V_hi approximates 1/(tan(B)+tan(x)) to 40 bits//// V_lo := V_hi * ( [ (1 - V_hi*T_hi) - V_hi*tanx ]// - V_hi*T_lo ) ...observe all order// ...V_hi + V_lo approximates 1/(tan(B) + tan(x))// ...to extra accuracy//// ... SC_inv(B) * (x + x*P)// ... -cot(B) + ------------------------- + CORR// ... tan(B) + (x + x*P)// ...// ... =-cot(B) + SC_inv(B)*(x + x*P)*(V_hi + V_lo) + CORR// ...//// Sx := SC_inv * x// CORR := sgn_r * c * SC_inv * C_hi//// ...put the ingredients together to compute// ... SC_inv(B) * (x + x*P)// ... -cot(B) + ------------------------- + CORR// ... tan(B) + (x + x*P)// ...// ... =-cot(B) + SC_inv(B)*(x + x*P)*(V_hi + V_lo) + CORR// ...// ... =-C_hi - C_lo + CORR +// ... Sx * V_hi + Sx * V_lo + Sx * P *(V_hi + V_lo)//// CORR := CORR - C_lo// tail := V_lo + P*(V_hi + V_lo)// tail := Sx * tail + CORR// tail := Sx * V_hi + tail// C_hi := -sgn_r * C_hi//// ...C_hi + sgn_r*tail now approximates// ...sgn_r*(-cot(B+x) + CORR) accurately//// Result := C_hi + sgn_r*tail in user-defined rounding control// ...It is crucial that independent paths be fully// ...exploited for performance's sake.//// 3. Implementation Notes// =======================//// Table entries T_hi, T_lo; C_hi, C_lo; SC_inv//// Recall that 2^(-2) <= |r| <= pi/4;//// r = sgn_r * 2^k * 1.b_1 b_2 ... b_63//// and//// B = 2^k * 1.b_1 b_2 b_3 b_4 b_5 1//// Thus, for k = -2, possible values of B are//// B = 2^(-2) * ( 1 + index/32 + 1/64 ),// index ranges from 0 to 31//// For k = -1, however, since |r| <= pi/4 = 0.78...// possible values of B are//// B = 2^(-1) * ( 1 + index/32 + 1/64 )// index ranges from 0 to 19.////#include "libm_support.h"#ifdef _LIBC.rodata#else.data#endif.align 128TAN_BASE_CONSTANTS:ASM_TYPE_DIRECTIVE(TAN_BASE_CONSTANTS,@object)data4 0x4B800000, 0xCB800000, 0x38800000, 0xB8800000 // two**24, -two**24 // two**-14, -two**-14data4 0x4E44152A, 0xA2F9836E, 0x00003FFE, 0x00000000 // two_by_pidata4 0xCE81B9F1, 0xC84D32B0, 0x00004016, 0x00000000 // P_0data4 0x2168C235, 0xC90FDAA2, 0x00003FFF, 0x00000000 // P_1data4 0xFC8F8CBB, 0xECE675D1, 0x0000BFBD, 0x00000000 // P_2data4 0xACC19C60, 0xB7ED8FBB, 0x0000BF7C, 0x00000000 // P_3data4 0x5F000000, 0xDF000000, 0x00000000, 0x00000000 // two_to_63, -two_to_63data4 0x6EC6B45A, 0xA397E504, 0x00003FE7, 0x00000000 // Inv_P_0data4 0xDBD171A1, 0x8D848E89, 0x0000BFBF, 0x00000000 // d_1data4 0x18A66F8E, 0xD5394C36, 0x0000BF7C, 0x00000000 // d_2data4 0x2168C234, 0xC90FDAA2, 0x00003FFE, 0x00000000 // PI_BY_4data4 0x2168C234, 0xC90FDAA2, 0x0000BFFE, 0x00000000 // MPI_BY_4data4 0x3E800000, 0xBE800000, 0x00000000, 0x00000000 // two**-2, -two**-2data4 0x2F000000, 0xAF000000, 0x00000000, 0x00000000 // two**-33, -two**-33data4 0xAAAAAABD, 0xAAAAAAAA, 0x00003FFD, 0x00000000 // P1_1data4 0x88882E6A, 0x88888888, 0x00003FFC, 0x00000000 // P1_2data4 0x0F0177B6, 0xDD0DD0DD, 0x00003FFA, 0x00000000 // P1_3data4 0x646B8C6D, 0xB327A440, 0x00003FF9, 0x00000000 // P1_4data4 0x1D5F7D20, 0x91371B25, 0x00003FF8, 0x00000000 // P1_5data4 0x61C67914, 0xEB69A5F1, 0x00003FF6, 0x00000000 // P1_6data4 0x019318D2, 0xBEDD37BE, 0x00003FF5, 0x00000000 // P1_7data4 0x3C794015, 0x9979B146, 0x00003FF4, 0x00000000 // P1_8data4 0x8C6EB58A, 0x8EBD21A3, 0x00003FF3, 0x00000000 // P1_9data4 0xAAAAAAB4, 0xAAAAAAAA, 0x00003FFD, 0x00000000 // Q1_1data4 0x0B5FC93E, 0xB60B60B6, 0x00003FF9, 0x00000000 // Q1_2data4 0x0C9BBFBF, 0x8AB355E0, 0x00003FF6, 0x00000000 // Q1_3data4 0xCBEE3D4C, 0xDDEBBC89, 0x00003FF2, 0x00000000 // Q1_4data4 0x5F80BBB6, 0xB3548A68, 0x00003FEF, 0x00000000 // Q1_5data4 0x4CED5BF1, 0x91362560, 0x00003FEC, 0x00000000 // Q1_6data4 0x8EE92A83, 0xF189D95A, 0x00003FE8, 0x00000000 // Q1_7data4 0xAAAB362F, 0xAAAAAAAA, 0x00003FFD, 0x00000000 // P2_1data4 0xE97A6097, 0x88888886, 0x00003FFC, 0x00000000 // P2_2data4 0x25E716A1, 0xDD108EE0, 0x00003FFA, 0x00000000 // P2_3//// Entries T_hi double-precision memory format// Index = 0,1,...,31 B = 2^(-2)*(1+Index/32+1/64)// Entries T_lo single-precision memory format// Index = 0,1,...,31 B = 2^(-2)*(1+Index/32+1/64)//data4 0x62400794, 0x3FD09BC3, 0x23A05C32, 0x00000000data4 0xDFFBC074, 0x3FD124A9, 0x240078B2, 0x00000000data4 0x5BD4920F, 0x3FD1AE23, 0x23826B8E, 0x00000000data4 0x15E2701D, 0x3FD23835, 0x22D31154, 0x00000000data4 0x63739C2D, 0x3FD2C2E4, 0x2265C9E2, 0x00000000data4 0xAFEEA48B, 0x3FD34E36, 0x245C05EB, 0x00000000data4 0x7DBB35D1, 0x3FD3DA31, 0x24749F2D, 0x00000000data4 0x67321619, 0x3FD466DA, 0x2462CECE, 0x00000000data4 0x1F94A4D5, 0x3FD4F437, 0x246D0DF1, 0x00000000data4 0x740C3E6D, 0x3FD5824D, 0x240A85B5, 0x00000000data4 0x4CB1E73D, 0x3FD61123, 0x23F96E33, 0x00000000data4 0xAD9EA64B, 0x3FD6A0BE, 0x247C5393, 0x00000000data4 0xB804FD01, 0x3FD73125, 0x241F3B29, 0x00000000data4 0xAB53EE83, 0x3FD7C25E, 0x2479989B, 0x00000000data4 0xE6640EED, 0x3FD8546F, 0x23B343BC, 0x00000000data4 0xE8AF1892, 0x3FD8E75F, 0x241454D1, 0x00000000data4 0x53928BDA, 0x3FD97B35, 0x238613D9, 0x00000000data4 0xEB9DE4DE, 0x3FDA0FF6, 0x22859FA7, 0x00000000data4 0x99ECF92D, 0x3FDAA5AB, 0x237A6D06, 0x00000000data4 0x6D8F1796, 0x3FDB3C5A, 0x23952F6C, 0x00000000data4 0x9CFB8BE4, 0x3FDBD40A, 0x2280FC95, 0x00000000data4 0x87943100, 0x3FDC6CC3, 0x245D2EC0, 0x00000000data4 0xB736C500, 0x3FDD068C, 0x23C4AD7D, 0x00000000data4 0xE1DDBC31, 0x3FDDA16D, 0x23D076E6, 0x00000000data4 0xEB515A93, 0x3FDE3D6E, 0x244809A6, 0x00000000data4 0xE6E9E5F1, 0x3FDEDA97, 0x220856C8, 0x00000000data4 0x1963CE69, 0x3FDF78F1, 0x244BE993, 0x00000000data4 0x7D635BCE, 0x3FE00C41, 0x23D21799, 0x00000000data4 0x1C302CD3, 0x3FE05CAB, 0x248A1B1D, 0x00000000data4 0xDB6A1FA0, 0x3FE0ADB9, 0x23D53E33, 0x00000000data4 0x4A20BA81, 0x3FE0FF72, 0x24DB9ED5, 0x00000000data4 0x153FA6F5, 0x3FE151D9, 0x24E9E451, 0x00000000//// Entries T_hi double-precision memory format// Index = 0,1,...,19 B = 2^(-1)*(1+Index/32+1/64)// Entries T_lo single-precision memory format// Index = 0,1,...,19 B = 2^(-1)*(1+Index/32+1/64)//data4 0xBA1BE39E, 0x3FE1CEC4, 0x24B60F9E, 0x00000000data4 0x5ABD9B2D, 0x3FE277E4, 0x248C2474, 0x00000000data4 0x0272B110, 0x3FE32418, 0x247B8311, 0x00000000data4 0x890E2DF0, 0x3FE3D38B, 0x24C55751, 0x00000000data4 0x46236871, 0x3FE4866D, 0x24E5BC34, 0x00000000data4 0x45E044B0, 0x3FE53CEE, 0x24001BA4, 0x00000000data4 0x82EC06E4, 0x3FE5F742, 0x24B973DC, 0x00000000data4 0x25DF43F9, 0x3FE6B5A1, 0x24895440, 0x00000000data4 0xCAFD348C, 0x3FE77844, 0x240021CA, 0x00000000data4 0xCEED6B92, 0x3FE83F6B, 0x24C45372, 0x00000000data4 0xA34F3665, 0x3FE90B58, 0x240DAD33, 0x00000000data4 0x2C1E56B4, 0x3FE9DC52, 0x24F846CE, 0x00000000data4 0x27041578, 0x3FEAB2A4, 0x2323FB6E, 0x00000000data4 0x9DD8C373, 0x3FEB8E9F, 0x24B3090B, 0x00000000data4 0x65C9AA7B, 0x3FEC709B, 0x2449F611, 0x00000000data4 0xACCF8435, 0x3FED58F4, 0x23616A7E, 0x00000000data4 0x97635082, 0x3FEE480F, 0x24C2FEAE, 0x00000000data4 0xF0ACC544, 0x3FEF3E57, 0x242CE964, 0x00000000data4 0xF7E06E4B, 0x3FF01E20, 0x2480D3EE, 0x00000000data4 0x8A798A69, 0x3FF0A125, 0x24DB8967, 0x00000000//// Entries C_hi double-precision memory format// Index = 0,1,...,31 B = 2^(-2)*(1+Index/32+1/64)// Entries C_lo single-precision memory format// Index = 0,1,...,31 B = 2^(-2)*(1+Index/32+1/64)//data4 0xE63EFBD0, 0x400ED3E2, 0x259D94D4, 0x00000000data4 0xC515DAB5, 0x400DDDB4, 0x245F0537, 0x00000000data4 0xBE19A79F, 0x400CF57A, 0x25D4EA9F, 0x00000000data4 0xD15298ED, 0x400C1A06, 0x24AE40A0, 0x00000000data4 0x164B2708, 0x400B4A4C, 0x25A5AAB6, 0x00000000data4 0x5285B068, 0x400A855A, 0x25524F18, 0x00000000data4 0x3FFA549F, 0x4009CA5A, 0x24C999C0, 0x00000000data4 0x646AF623, 0x4009188A, 0x254FD801, 0x00000000data4 0x6084D0E7, 0x40086F3C, 0x2560F5FD, 0x00000000data4 0xA29A76EE, 0x4007CDD2, 0x255B9D19, 0x00000000data4 0x6C8ECA95, 0x400733BE, 0x25CB021B, 0x00000000data4 0x1F8DDC52, 0x4006A07E, 0x24AB4722, 0x00000000data4 0xC298AD58, 0x4006139B, 0x252764E2, 0x00000000data4 0xBAD7164B, 0x40058CAB, 0x24DAF5DB, 0x00000000data4 0xAE31A5D3, 0x40050B4B, 0x25EA20F4, 0x00000000data4 0x89F85A8A, 0x40048F21, 0x2583A3E8, 0x00000000data4 0xA862380D, 0x400417DA, 0x25DCC4CC, 0x00000000data4 0x1088FCFE, 0x4003A52B, 0x2430A492, 0x00000000data4 0xCD3527D5, 0x400336CC, 0x255F77CF, 0x00000000data4 0x5760766D, 0x4002CC7F, 0x25DA0BDA, 0x00000000data4 0x11CE02E3, 0x40026607, 0x256FF4A2, 0x00000000data4 0xD37BBE04, 0x4002032C, 0x25208AED, 0x00000000data4 0x7F050775, 0x4001A3BD, 0x24B72DD6, 0x00000000data4 0xA554848A, 0x40014789, 0x24AB4DAA, 0x00000000data4 0x323E81B7, 0x4000EE65, 0x2584C440, 0x00000000data4 0x21CF1293, 0x40009827, 0x25C9428D, 0x00000000data4 0x3D415EEB, 0x400044A9, 0x25DC8482, 0x00000000data4 0xBD72C577, 0x3FFFE78F, 0x257F5070, 0x00000000data4 0x75EFD28E, 0x3FFF4AC3, 0x23EBBF7A, 0x00000000data4 0x60B52DDE, 0x3FFEB2AF, 0x22EECA07, 0x00000000data4 0x35204180, 0x3FFE1F19, 0x24191079, 0x00000000data4 0x54F7E60A, 0x3FFD8FCA, 0x248D3058, 0x00000000//// Entries C_hi double-precision memory format// Index = 0,1,...,19 B = 2^(-1)*(1+Index/32+1/64)// Entries C_lo single-precision memory format// Index = 0,1,...,19 B = 2^(-1)*(1+Index/32+1/64)//data4 0x79F6FADE, 0x3FFCC06A, 0x239C7886, 0x00000000data4 0x891662A6, 0x3FFBB91F, 0x250BD191, 0x00000000data4 0x529F155D, 0x3FFABFB6, 0x256CC3E6, 0x00000000data4 0x2E964AE9, 0x3FF9D300, 0x250843E3, 0x00000000data4 0x89DCB383, 0x3FF8F1EF, 0x2277C87E, 0x00000000data4 0x7C87DBD6, 0x3FF81B93, 0x256DA6CF, 0x00000000data4 0x1042EDE4, 0x3FF74F14, 0x2573D28A, 0x00000000data4 0x1784B360, 0x3FF68BAF, 0x242E489A, 0x00000000data4 0x7C923C4C, 0x3FF5D0B5, 0x2532D940, 0x00000000data4 0xF418EF20, 0x3FF51D88, 0x253C7DD6, 0x00000000data4 0x02F88DAE, 0x3FF4719A, 0x23DB59BF, 0x00000000data4 0x49DA0788, 0x3FF3CC66, 0x252B4756, 0x00000000data4 0x0B980DB8, 0x3FF32D77, 0x23FE585F, 0x00000000data4 0xE56C987A, 0x3FF2945F, 0x25378A63, 0x00000000data4 0xB16523F6, 0x3FF200BD, 0x247BB2E0, 0x00000000data4 0x8CE27778, 0x3FF17235, 0x24446538, 0x00000000data4 0xFDEFE692, 0x3FF0E873, 0x2514638F, 0x00000000data4 0x33154062, 0x3FF0632C, 0x24A7FC27, 0x00000000data4 0xB3EF115F, 0x3FEFC42E, 0x248FD0FE, 0x00000000data4 0x135D26F6, 0x3FEEC9E8, 0x2385C719, 0x00000000//// Entries SC_inv in Swapped IEEE format (extended)// Index = 0,1,...,31 B = 2^(-2)*(1+Index/32+1/64)//data4 0x1BF30C9E, 0x839D6D4A, 0x00004001, 0x00000000data4 0x554B0EB0, 0x80092804, 0x00004001, 0x00000000data4 0xA1CF0DE9, 0xF959F94C, 0x00004000, 0x00000000data4 0x77378677, 0xF3086BA0, 0x00004000, 0x00000000data4 0xCCD4723C, 0xED154515, 0x00004000, 0x00000000data4 0x1C27CF25, 0xE7790944, 0x00004000, 0x00000000data4 0x8DDACB88, 0xE22D037D, 0x00004000, 0x00000000data4 0x89C73522, 0xDD2B2D8A, 0x00004000, 0x00000000data4 0xBB2C1171, 0xD86E1A23, 0x00004000, 0x00000000data4 0xDFF5E0F9, 0xD3F0E288, 0x00004000, 0x00000000data4 0x283BEBD5, 0xCFAF16B1, 0x00004000, 0x00000000data4 0x0D88DD53, 0xCBA4AFAA, 0x00004000, 0x00000000data4 0xCA67C43D, 0xC7CE03CC, 0x00004000, 0x00000000data4 0x0CA0DDB0, 0xC427BC82, 0x00004000, 0x00000000data4 0xF13D8CAB, 0xC0AECD57, 0x00004000, 0x00000000data4 0x71ECE6B1, 0xBD606C38, 0x00004000, 0x00000000data4 0xA44C4929, 0xBA3A0A96, 0x00004000, 0x00000000data4 0xE5CCCEC1, 0xB7394F6F, 0x00004000, 0x00000000data4 0x9637D8BC, 0xB45C1203, 0x00004000, 0x00000000data4 0x92CB051B, 0xB1A05528, 0x00004000, 0x00000000data4 0x6BA2FFD0, 0xAF04432B, 0x00004000, 0x00000000data4 0x7221235F, 0xAC862A23, 0x00004000, 0x00000000data4 0x5F00A9D1, 0xAA2478AF, 0x00004000, 0x00000000data4 0x81E082BF, 0xA7DDBB0C, 0x00004000, 0x00000000data4 0x45684FEE, 0xA5B0987D, 0x00004000, 0x00000000data4 0x627A8F53, 0xA39BD0F5, 0x00004000, 0x00000000data4 0x6EC5C8B0, 0xA19E3B03, 0x00004000, 0x00000000data4 0x91CD7C66, 0x9FB6C1F0, 0x00004000, 0x00000000data4 0x1FA3DF8A, 0x9DE46410, 0x00004000, 0x00000000data4 0xA8F6B888, 0x9C263139, 0x00004000, 0x00000000data4 0xC27B0450, 0x9A7B4968, 0x00004000, 0x00000000data4 0x5EE614EE, 0x98E2DB7E, 0x00004000, 0x00000000//// Entries SC_inv in Swapped IEEE format (extended)// Index = 0,1,...,19 B = 2^(-1)*(1+Index/32+1/64)//data4 0x13B2B5BA, 0x969F335C, 0x00004000, 0x00000000data4 0xD4C0F548, 0x93D446D9, 0x00004000, 0x00000000data4 0x61B798AF, 0x9147094F, 0x00004000, 0x00000000data4 0x758787AC, 0x8EF317CC, 0x00004000, 0x00000000data4 0xB99EEFDB, 0x8CD498B3, 0x00004000, 0x00000000data4 0xDFF8BC37, 0x8AE82A7D, 0x00004000, 0x00000000data4 0xE3C55D42, 0x892AD546, 0x00004000, 0x00000000data4 0xD15573C1, 0x8799FEA9, 0x00004000, 0x00000000data4 0x435A4B4C, 0x86335F88, 0x00004000, 0x00000000data4 0x3E93A87B, 0x84F4FB6E, 0x00004000, 0x00000000data4 0x80A382FB, 0x83DD1952, 0x00004000, 0x00000000data4 0xA4CB8C9E, 0x82EA3D7F, 0x00004000, 0x00000000data4 0x6861D0A8, 0x821B247C, 0x00004000, 0x00000000data4 0x63E8D244, 0x816EBED1, 0x00004000, 0x00000000data4 0x27E4CFC6, 0x80E42D91, 0x00004000, 0x00000000data4 0x28E64AFD, 0x807ABF8D, 0x00004000, 0x00000000data4 0x863B4FD8, 0x8031EF26, 0x00004000, 0x00000000data4 0xAE8C11FD, 0x800960AD, 0x00004000, 0x00000000data4 0x5FDBEC21, 0x8000E147, 0x00004000, 0x00000000data4 0xA07791FA, 0x80186650, 0x00004000, 0x00000000Arg = f8 Result = f8fp_tmp = f9U_2 = f10rsq = f11C_hi = f12C_lo = f13T_hi = f14T_lo = f15N_0 = f32d_1 = f33MPI_BY_4 = f34tail = f35tanx = f36Cx = f37Sx = f38sgn_r = f39CORR = f40P = f41D = f42ArgPrime = f43P_0 = f44P2_1 = f45P2_2 = f46P2_3 = f47P1_1 = f45P1_2 = f46P1_3 = f47P1_4 = f48P1_5 = f49P1_6 = f50P1_7 = f51P1_8 = f52P1_9 = f53TWO_TO_63 = f54NEGTWO_TO_63 = f55x = f56xsq = f57Tx = f58Tx1 = f59Set = f60poly1 = f61poly2 = f62Poly = f63Poly1 = f64Poly2 = f65r_to_the_8 = f66B = f67SC_inv = f68Pos_r = f69N_0_fix = f70PI_BY_4 = f71NEGTWO_TO_NEG2 = f72TWO_TO_24 = f73TWO_TO_NEG14 = f74TWO_TO_NEG33 = f75NEGTWO_TO_24 = f76NEGTWO_TO_NEG14 = f76NEGTWO_TO_NEG33 = f77two_by_PI = f78N = f79N_fix = f80P_1 = f81P_2 = f82P_3 = f83s_val = f84w = f85c = f86r = f87Z = f88A = f89a = f90t = f91U_1 = f92d_2 = f93TWO_TO_NEG2 = f94Q1_1 = f95Q1_2 = f96Q1_3 = f97Q1_4 = f98Q1_5 = f99Q1_6 = f100Q1_7 = f101Q1_8 = f102S_hi = f103S_lo = f104V_hi = f105V_lo = f106U_hi = f107U_lo = f108U_hiabs = f109V_hiabs = f110V = f111Inv_P_0 = f112GR_SAVE_B0 = r33GR_SAVE_GP = r34GR_SAVE_PFS = r35delta1 = r36table_ptr1 = r37table_ptr2 = r38i_0 = r39i_1 = r40 N_fix_gr = r41 N_inc = r42 exp_Arg = r43 exp_r = r44 sig_r = r45 lookup = r46 table_offset = r47 Create_B = r48 gr_tmp = r49GR_Parameter_X = r49GR_Parameter_r = r50.global __libm_tan.section .text.proc __libm_tan__libm_tan: { .mfialloc r32 = ar.pfs, 0,17,2,0(p0) fclass.m.unc p6,p0 = Arg, 0x1E7 addl gr_tmp = -1,r0 };;{ .mfi nop.m 999(p0) fclass.nm.unc p7,p0 = Arg, 0x1FF nop.i 999};;{ .mfi(p0) addl table_ptr1 = @ltoff(TAN_BASE_CONSTANTS), gp nop.f 999 nop.i 999};;{ .mmi ld8 table_ptr1 = [table_ptr1] setf.sig fp_tmp = gr_tmp // Make a constant so fmpy produces inexact nop.i 999};;//// Check for NatVals, Infs , NaNs, and Zeros // Check for everything - if false, then must be pseudo-zero// or pseudo-nan.// Local table pointer//{ .mbb(p0) add table_ptr2 = 96, table_ptr1(p6) br.cond.spnt __libm_TAN_SPECIAL (p7) br.cond.spnt __libm_TAN_SPECIAL ;;}//// Point to Inv_P_0// Branch out to deal with unsupporteds and special values. //{ .mmf(p0) ldfs TWO_TO_24 = [table_ptr1],4(p0) ldfs TWO_TO_63 = [table_ptr2],4//// Load -2**24, load -2**63.//(p0) fcmp.eq.s0 p0, p6 = Arg, f1 ;;}{ .mfi(p0) ldfs NEGTWO_TO_63 = [table_ptr2],12(p0) fnorm.s1 Arg = Arg nop.i 999}//// Load 2**24, Load 2**63.//{ .mmi(p0) ldfs NEGTWO_TO_24 = [table_ptr1],12 ;;//// Do fcmp to generate Denormal exception // - can't do FNORM (will generate Underflow when U is unmasked!)// Normalize input argument.//(p0) ldfe two_by_PI = [table_ptr1],16 nop.i 999⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -