1.file "coshf.s" 2 3 4// Copyright (c) 2000 - 2005, Intel Corporation 5// All rights reserved. 6// 7// 8// Redistribution and use in source and binary forms, with or without 9// modification, are permitted provided that the following conditions are 10// met: 11// 12// * Redistributions of source code must retain the above copyright 13// notice, this list of conditions and the following disclaimer. 14// 15// * Redistributions in binary form must reproduce the above copyright 16// notice, this list of conditions and the following disclaimer in the 17// documentation and/or other materials provided with the distribution. 18// 19// * The name of Intel Corporation may not be used to endorse or promote 20// products derived from this software without specific prior written 21// permission. 22 23// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 24// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 25// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 26// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS 27// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 28// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 29// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 30// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY 31// OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING 32// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 33// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 34// 35// Intel Corporation is the author of this code, and requests that all 36// problem reports or change requests be submitted to it directly at 37// http://www.intel.com/software/products/opensource/libraries/num.htm. 38 39// History 40//********************************************************************* 41// 02/02/00 Initial version 42// 02/16/00 The error tag for coshf overflow changed to 65 (from 64). 43// 04/04/00 Unwind support added 44// 08/15/00 Bundle added after call to __libm_error_support to properly 45// set [the previously overwritten] GR_Parameter_RESULT. 46// 05/07/01 Reworked to improve speed of all paths 47// 05/20/02 Cleaned up namespace and sf0 syntax 48// 11/15/02 Improved algorithm based on expf 49// 03/31/05 Reformatted delimiters between data tables 50// 51// API 52//********************************************************************* 53// float coshf(float) 54// 55// Overview of operation 56//********************************************************************* 57// Case 1: 0 < |x| < 0.25 58// Evaluate cosh(x) by a 8th order polynomial 59// Care is take for the order of multiplication; and A2 is not exactly 1/4!, 60// A3 is not exactly 1/6!, etc. 61// cosh(x) = 1 + (A1*x^2 + A2*x^4 + A3*x^6 + A4*x^8) 62// 63// Case 2: 0.25 < |x| < 89.41598 64// Algorithm is based on the identity cosh(x) = ( exp(x) + exp(-x) ) / 2. 65// The algorithm for exp is described as below. There are a number of 66// economies from evaluating both exp(x) and exp(-x). Although we 67// are evaluating both quantities, only where the quantities diverge do we 68// duplicate the computations. The basic algorithm for exp(x) is described 69// below. 70// 71// Take the input x. w is "how many log2/128 in x?" 72// w = x * 64/log2 73// NJ = int(w) 74// x = NJ*log2/64 + R 75 76// NJ = 64*n + j 77// x = n*log2 + (log2/64)*j + R 78// 79// So, exp(x) = 2^n * 2^(j/64)* exp(R) 80// 81// T = 2^n * 2^(j/64) 82// Construct 2^n 83// Get 2^(j/64) table 84// actually all the entries of 2^(j/64) table are stored in DP and 85// with exponent bits set to 0 -> multiplication on 2^n can be 86// performed by doing logical "or" operation with bits presenting 2^n 87 88// exp(R) = 1 + (exp(R) - 1) 89// P = exp(R) - 1 approximated by Taylor series of 3rd degree 90// P = A3*R^3 + A2*R^2 + R, A3 = 1/6, A2 = 1/2 91// 92 93// The final result is reconstructed as follows 94// exp(x) = T + T*P 95 96// Special values 97//********************************************************************* 98// coshf(+0) = 1.0 99// coshf(-0) = 1.0 100 101// coshf(+qnan) = +qnan 102// coshf(-qnan) = -qnan 103// coshf(+snan) = +qnan 104// coshf(-snan) = -qnan 105 106// coshf(-inf) = +inf 107// coshf(+inf) = +inf 108 109// Overflow and Underflow 110//********************************************************************* 111// coshf(x) = largest single normal when 112// x = 89.41598 = 0x42b2d4fc 113// 114// There is no underflow. 115 116// Registers used 117//********************************************************************* 118// Floating Point registers used: 119// f8 input, output 120// f6,f7, f9 -> f15, f32 -> f45 121 122// General registers used: 123// r2, r3, r16 -> r38 124 125// Predicate registers used: 126// p6 -> p15 127 128// Assembly macros 129//********************************************************************* 130// integer registers used 131// scratch 132rNJ = r2 133rNJ_neg = r3 134 135rJ_neg = r16 136rN_neg = r17 137rSignexp_x = r18 138rExp_x = r18 139rExp_mask = r19 140rExp_bias = r20 141rAd1 = r21 142rAd2 = r22 143rJ = r23 144rN = r24 145rTblAddr = r25 146rA3 = r26 147rExpHalf = r27 148rLn2Div64 = r28 149rGt_ln = r29 150r17ones_m1 = r29 151rRightShifter = r30 152rJ_mask = r30 153r64DivLn2 = r31 154rN_mask = r31 155// stacked 156GR_SAVE_PFS = r32 157GR_SAVE_B0 = r33 158GR_SAVE_GP = r34 159GR_Parameter_X = r35 160GR_Parameter_Y = r36 161GR_Parameter_RESULT = r37 162GR_Parameter_TAG = r38 163 164// floating point registers used 165FR_X = f10 166FR_Y = f1 167FR_RESULT = f8 168// scratch 169fRightShifter = f6 170f64DivLn2 = f7 171fNormX = f9 172fNint = f10 173fN = f11 174fR = f12 175fLn2Div64 = f13 176fA2 = f14 177fA3 = f15 178// stacked 179fP = f32 180fT = f33 181fMIN_SGL_OFLOW_ARG = f34 182fMAX_SGL_NORM_ARG = f35 183fRSqr = f36 184fA1 = f37 185fA21 = f37 186fA4 = f38 187fA43 = f38 188fA4321 = f38 189fX4 = f39 190fTmp = f39 191fGt_pln = f39 192fWre_urm_f8 = f40 193fXsq = f40 194fP_neg = f41 195fT_neg = f42 196fExp = f43 197fExp_neg = f44 198fAbsX = f45 199 200 201RODATA 202.align 16 203 204LOCAL_OBJECT_START(_coshf_table) 205data4 0x42b2d4fd // Smallest single arg to overflow single result 206data4 0x42b2d4fc // Largest single arg to give normal single result 207data4 0x00000000 // pad 208data4 0x00000000 // pad 209// 210// 2^(j/64) table, j goes from 0 to 63 211data8 0x0000000000000000 // 2^(0/64) 212data8 0x00002C9A3E778061 // 2^(1/64) 213data8 0x000059B0D3158574 // 2^(2/64) 214data8 0x0000874518759BC8 // 2^(3/64) 215data8 0x0000B5586CF9890F // 2^(4/64) 216data8 0x0000E3EC32D3D1A2 // 2^(5/64) 217data8 0x00011301D0125B51 // 2^(6/64) 218data8 0x0001429AAEA92DE0 // 2^(7/64) 219data8 0x000172B83C7D517B // 2^(8/64) 220data8 0x0001A35BEB6FCB75 // 2^(9/64) 221data8 0x0001D4873168B9AA // 2^(10/64) 222data8 0x0002063B88628CD6 // 2^(11/64) 223data8 0x0002387A6E756238 // 2^(12/64) 224data8 0x00026B4565E27CDD // 2^(13/64) 225data8 0x00029E9DF51FDEE1 // 2^(14/64) 226data8 0x0002D285A6E4030B // 2^(15/64) 227data8 0x000306FE0A31B715 // 2^(16/64) 228data8 0x00033C08B26416FF // 2^(17/64) 229data8 0x000371A7373AA9CB // 2^(18/64) 230data8 0x0003A7DB34E59FF7 // 2^(19/64) 231data8 0x0003DEA64C123422 // 2^(20/64) 232data8 0x0004160A21F72E2A // 2^(21/64) 233data8 0x00044E086061892D // 2^(22/64) 234data8 0x000486A2B5C13CD0 // 2^(23/64) 235data8 0x0004BFDAD5362A27 // 2^(24/64) 236data8 0x0004F9B2769D2CA7 // 2^(25/64) 237data8 0x0005342B569D4F82 // 2^(26/64) 238data8 0x00056F4736B527DA // 2^(27/64) 239data8 0x0005AB07DD485429 // 2^(28/64) 240data8 0x0005E76F15AD2148 // 2^(29/64) 241data8 0x0006247EB03A5585 // 2^(30/64) 242data8 0x0006623882552225 // 2^(31/64) 243data8 0x0006A09E667F3BCD // 2^(32/64) 244data8 0x0006DFB23C651A2F // 2^(33/64) 245data8 0x00071F75E8EC5F74 // 2^(34/64) 246data8 0x00075FEB564267C9 // 2^(35/64) 247data8 0x0007A11473EB0187 // 2^(36/64) 248data8 0x0007E2F336CF4E62 // 2^(37/64) 249data8 0x00082589994CCE13 // 2^(38/64) 250data8 0x000868D99B4492ED // 2^(39/64) 251data8 0x0008ACE5422AA0DB // 2^(40/64) 252data8 0x0008F1AE99157736 // 2^(41/64) 253data8 0x00093737B0CDC5E5 // 2^(42/64) 254data8 0x00097D829FDE4E50 // 2^(43/64) 255data8 0x0009C49182A3F090 // 2^(44/64) 256data8 0x000A0C667B5DE565 // 2^(45/64) 257data8 0x000A5503B23E255D // 2^(46/64) 258data8 0x000A9E6B5579FDBF // 2^(47/64) 259data8 0x000AE89F995AD3AD // 2^(48/64) 260data8 0x000B33A2B84F15FB // 2^(49/64) 261data8 0x000B7F76F2FB5E47 // 2^(50/64) 262data8 0x000BCC1E904BC1D2 // 2^(51/64) 263data8 0x000C199BDD85529C // 2^(52/64) 264data8 0x000C67F12E57D14B // 2^(53/64) 265data8 0x000CB720DCEF9069 // 2^(54/64) 266data8 0x000D072D4A07897C // 2^(55/64) 267data8 0x000D5818DCFBA487 // 2^(56/64) 268data8 0x000DA9E603DB3285 // 2^(57/64) 269data8 0x000DFC97337B9B5F // 2^(58/64) 270data8 0x000E502EE78B3FF6 // 2^(59/64) 271data8 0x000EA4AFA2A490DA // 2^(60/64) 272data8 0x000EFA1BEE615A27 // 2^(61/64) 273data8 0x000F50765B6E4540 // 2^(62/64) 274data8 0x000FA7C1819E90D8 // 2^(63/64) 275LOCAL_OBJECT_END(_coshf_table) 276 277LOCAL_OBJECT_START(cosh_p_table) 278data8 0x3efa3001dcf5905b // A4 279data8 0x3f56c1437543543e // A3 280data8 0x3fa5555572601504 // A2 281data8 0x3fdfffffffe2f097 // A1 282LOCAL_OBJECT_END(cosh_p_table) 283 284 285.section .text 286GLOBAL_IEEE754_ENTRY(coshf) 287 288{ .mlx 289 getf.exp rSignexp_x = f8 // Must recompute if x unorm 290 movl r64DivLn2 = 0x40571547652B82FE // 64/ln(2) 291} 292{ .mlx 293 addl rTblAddr = @ltoff(_coshf_table),gp 294 movl rRightShifter = 0x43E8000000000000 // DP Right Shifter 295} 296;; 297 298{ .mfi 299 // point to the beginning of the table 300 ld8 rTblAddr = [rTblAddr] 301 fclass.m p6, p0 = f8, 0x0b // Test for x=unorm 302 addl rA3 = 0x3E2AA, r0 // high bits of 1.0/6.0 rounded to SP 303} 304{ .mfi 305 nop.m 0 306 fnorm.s1 fNormX = f8 // normalized x 307 addl rExpHalf = 0xFFFE, r0 // exponent of 1/2 308} 309;; 310 311{ .mfi 312 setf.d f64DivLn2 = r64DivLn2 // load 64/ln(2) to FP reg 313 fclass.m p15, p0 = f8, 0x1e3 // test for NaT,NaN,Inf 314 nop.i 0 315} 316{ .mlx 317 // load Right Shifter to FP reg 318 setf.d fRightShifter = rRightShifter 319 movl rLn2Div64 = 0x3F862E42FEFA39EF // DP ln(2)/64 in GR 320} 321;; 322 323{ .mfi 324 mov rExp_mask = 0x1ffff 325 fcmp.eq.s1 p13, p0 = f0, f8 // test for x = 0.0 326 shl rA3 = rA3, 12 // 0x3E2AA000, approx to 1.0/6.0 in SP 327} 328{ .mfb 329 nop.m 0 330 nop.f 0 331(p6) br.cond.spnt COSH_UNORM // Branch if x=unorm 332} 333;; 334 335COSH_COMMON: 336{ .mfi 337 setf.exp fA2 = rExpHalf // load A2 to FP reg 338 nop.f 0 339 mov rExp_bias = 0xffff 340} 341{ .mfb 342 setf.d fLn2Div64 = rLn2Div64 // load ln(2)/64 to FP reg 343(p15) fma.s.s0 f8 = f8, f8, f0 // result if x = NaT,NaN,Inf 344(p15) br.ret.spnt b0 // exit here if x = NaT,NaN,Inf 345} 346;; 347 348{ .mfi 349 // min overflow and max normal threshold 350 ldfps fMIN_SGL_OFLOW_ARG, fMAX_SGL_NORM_ARG = [rTblAddr], 8 351 nop.f 0 352 and rExp_x = rExp_mask, rSignexp_x // Biased exponent of x 353} 354{ .mfb 355 setf.s fA3 = rA3 // load A3 to FP reg 356(p13) fma.s.s0 f8 = f1, f1, f0 // result if x = 0.0 357(p13) br.ret.spnt b0 // exit here if x =0.0 358} 359;; 360 361{ .mfi 362 sub rExp_x = rExp_x, rExp_bias // True exponent of x 363 fmerge.s fAbsX = f0, fNormX // Form |x| 364 nop.i 0 365} 366;; 367 368{ .mfi 369 nop.m 0 370 // x*(64/ln(2)) + Right Shifter 371 fma.s1 fNint = fNormX, f64DivLn2, fRightShifter 372 add rTblAddr = 8, rTblAddr 373} 374{ .mfb 375 cmp.gt p7, p0 = -2, rExp_x // Test |x| < 2^(-2) 376 fma.s1 fXsq = fNormX, fNormX, f0 // x*x for small path 377(p7) br.cond.spnt COSH_SMALL // Branch if 0 < |x| < 2^-2 378} 379;; 380 381{ .mfi 382 nop.m 0 383 // check for overflow 384 fcmp.ge.s1 p12, p13 = fAbsX, fMIN_SGL_OFLOW_ARG 385 mov rJ_mask = 0x3f // 6-bit mask for J 386} 387;; 388 389{ .mfb 390 nop.m 0 391 fms.s1 fN = fNint, f1, fRightShifter // n in FP register 392 // branch out if overflow 393(p12) br.cond.spnt COSH_CERTAIN_OVERFLOW 394} 395;; 396 397{ .mfi 398 getf.sig rNJ = fNint // bits of n, j 399 // check for possible overflow 400 fcmp.gt.s1 p13, p0 = fAbsX, fMAX_SGL_NORM_ARG 401 nop.i 0 402} 403;; 404 405{ .mfi 406 addl rN = 0xFFBF - 63, rNJ // biased and shifted n-1,j 407 fnma.s1 fR = fLn2Div64, fN, fNormX // R = x - N*ln(2)/64 408 and rJ = rJ_mask, rNJ // bits of j 409} 410{ .mfi 411 sub rNJ_neg = r0, rNJ // bits of n, j for -x 412 nop.f 0 413 andcm rN_mask = -1, rJ_mask // 0xff...fc0 to mask N 414} 415;; 416 417{ .mfi 418 shladd rJ = rJ, 3, rTblAddr // address in the 2^(j/64) table 419 nop.f 0 420 and rN = rN_mask, rN // biased, shifted n-1 421} 422{ .mfi 423 addl rN_neg = 0xFFBF - 63, rNJ_neg // -x biased, shifted n-1,j 424 nop.f 0 425 and rJ_neg = rJ_mask, rNJ_neg // bits of j for -x 426} 427;; 428 429{ .mfi 430 ld8 rJ = [rJ] // Table value 431 nop.f 0 432 shl rN = rN, 46 // 2^(n-1) bits in DP format 433} 434{ .mfi 435 shladd rJ_neg = rJ_neg, 3, rTblAddr // addr in 2^(j/64) table -x 436 nop.f 0 437 and rN_neg = rN_mask, rN_neg // biased, shifted n-1 for -x 438} 439;; 440 441{ .mfi 442 ld8 rJ_neg = [rJ_neg] // Table value for -x 443 nop.f 0 444 shl rN_neg = rN_neg, 46 // 2^(n-1) bits in DP format for -x 445} 446;; 447 448{ .mfi 449 or rN = rN, rJ // bits of 2^n * 2^(j/64) in DP format 450 nop.f 0 451 nop.i 0 452} 453;; 454 455{ .mmf 456 setf.d fT = rN // 2^(n-1) * 2^(j/64) 457 or rN_neg = rN_neg, rJ_neg // -x bits of 2^n * 2^(j/64) in DP 458 fma.s1 fRSqr = fR, fR, f0 // R^2 459} 460;; 461 462{ .mfi 463 setf.d fT_neg = rN_neg // 2^(n-1) * 2^(j/64) for -x 464 fma.s1 fP = fA3, fR, fA2 // A3*R + A2 465 nop.i 0 466} 467{ .mfi 468 nop.m 0 469 fnma.s1 fP_neg = fA3, fR, fA2 // A3*R + A2 for -x 470 nop.i 0 471} 472;; 473 474{ .mfi 475 nop.m 0 476 fma.s1 fP = fP, fRSqr, fR // P = (A3*R + A2)*R^2 + R 477 nop.i 0 478} 479{ .mfi 480 nop.m 0 481 fms.s1 fP_neg = fP_neg, fRSqr, fR // P = (A3*R + A2)*R^2 + R, -x 482 nop.i 0 483} 484;; 485 486{ .mfi 487 nop.m 0 488 fmpy.s0 fTmp = fLn2Div64, fLn2Div64 // Force inexact 489 nop.i 0 490} 491;; 492 493{ .mfi 494 nop.m 0 495 fma.s1 fExp = fP, fT, fT // exp(x)/2 496 nop.i 0 497} 498{ .mfb 499 nop.m 0 500 fma.s1 fExp_neg = fP_neg, fT_neg, fT_neg // exp(-x)/2 501 // branch out if possible overflow result 502(p13) br.cond.spnt COSH_POSSIBLE_OVERFLOW 503} 504;; 505 506{ .mfb 507 nop.m 0 508 // final result in the absence of overflow 509 fma.s.s0 f8 = fExp, f1, fExp_neg // result = (exp(x)+exp(-x))/2 510 // exit here in the absence of overflow 511 br.ret.sptk b0 // Exit main path, 0.25 <= |x| < 89.41598 512} 513;; 514 515// Here if 0 < |x| < 0.25. Evaluate 8th order polynomial. 516COSH_SMALL: 517{ .mmi 518 add rAd1 = 0x200, rTblAddr 519 add rAd2 = 0x210, rTblAddr 520 nop.i 0 521} 522;; 523 524{ .mmi 525 ldfpd fA4, fA3 = [rAd1] 526 ldfpd fA2, fA1 = [rAd2] 527 nop.i 0 528} 529;; 530 531{ .mfi 532 nop.m 0 533 fma.s1 fX4 = fXsq, fXsq, f0 534 nop.i 0 535} 536;; 537 538{ .mfi 539 nop.m 0 540 fma.s1 fA43 = fXsq, fA4, fA3 541 nop.i 0 542} 543{ .mfi 544 nop.m 0 545 fma.s1 fA21 = fXsq, fA2, fA1 546 nop.i 0 547} 548;; 549 550{ .mfi 551 nop.m 0 552 fma.s1 fA4321 = fX4, fA43, fA21 553 nop.i 0 554} 555;; 556 557// Dummy multiply to generate inexact 558{ .mfi 559 nop.m 0 560 fmpy.s0 fTmp = fA4, fA4 561 nop.i 0 562} 563{ .mfb 564 nop.m 0 565 fma.s.s0 f8 = fA4321, fXsq, f1 566 br.ret.sptk b0 // Exit if 0 < |x| < 0.25 567} 568;; 569 570COSH_POSSIBLE_OVERFLOW: 571 572// Here if fMAX_SGL_NORM_ARG < x < fMIN_SGL_OFLOW_ARG 573// This cannot happen if input is a single, only if input higher precision. 574// Overflow is a possibility, not a certainty. 575 576// Recompute result using status field 2 with user's rounding mode, 577// and wre set. If result is larger than largest single, then we have 578// overflow 579 580{ .mfi 581 mov rGt_ln = 0x1007f // Exponent for largest single + 1 ulp 582 fsetc.s2 0x7F,0x42 // Get user's round mode, set wre 583 nop.i 0 584} 585;; 586 587{ .mfi 588 setf.exp fGt_pln = rGt_ln // Create largest single + 1 ulp 589 fma.s.s2 fWre_urm_f8 = fP, fT, fT // Result with wre set 590 nop.i 0 591} 592;; 593 594{ .mfi 595 nop.m 0 596 fsetc.s2 0x7F,0x40 // Turn off wre in sf2 597 nop.i 0 598} 599;; 600 601{ .mfi 602 nop.m 0 603 fcmp.ge.s1 p6, p0 = fWre_urm_f8, fGt_pln // Test for overflow 604 nop.i 0 605} 606;; 607 608{ .mfb 609 nop.m 0 610 nop.f 0 611(p6) br.cond.spnt COSH_CERTAIN_OVERFLOW // Branch if overflow 612} 613;; 614 615{ .mfb 616 nop.m 0 617 fma.s.s0 f8 = fP, fT, fT 618 br.ret.sptk b0 // Exit if really no overflow 619} 620;; 621 622// here if overflow 623COSH_CERTAIN_OVERFLOW: 624{ .mmi 625 addl r17ones_m1 = 0x1FFFE, r0 626;; 627 setf.exp fTmp = r17ones_m1 628 nop.i 0 629} 630;; 631 632{ .mfi 633 alloc r32 = ar.pfs, 0, 3, 4, 0 // get some registers 634 fmerge.s FR_X = f8,f8 635 nop.i 0 636} 637{ .mfb 638 mov GR_Parameter_TAG = 65 639 fma.s.s0 FR_RESULT = fTmp, fTmp, f0 // Set I,O and +INF result 640 br.cond.sptk __libm_error_region 641} 642;; 643 644// Here if x unorm 645COSH_UNORM: 646{ .mfb 647 getf.exp rSignexp_x = fNormX // Must recompute if x unorm 648 fcmp.eq.s0 p6, p0 = f8, f0 // Set D flag 649 br.cond.sptk COSH_COMMON // Return to main path 650} 651;; 652 653GLOBAL_IEEE754_END(coshf) 654libm_alias_float_other (__cosh, cosh) 655 656 657LOCAL_LIBM_ENTRY(__libm_error_region) 658.prologue 659{ .mfi 660 add GR_Parameter_Y=-32,sp // Parameter 2 value 661 nop.f 0 662.save ar.pfs,GR_SAVE_PFS 663 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs 664} 665{ .mfi 666.fframe 64 667 add sp=-64,sp // Create new stack 668 nop.f 0 669 mov GR_SAVE_GP=gp // Save gp 670};; 671{ .mmi 672 stfs [GR_Parameter_Y] = FR_Y,16 // Store Parameter 2 on stack 673 add GR_Parameter_X = 16,sp // Parameter 1 address 674.save b0, GR_SAVE_B0 675 mov GR_SAVE_B0=b0 // Save b0 676};; 677.body 678{ .mfi 679 stfs [GR_Parameter_X] = FR_X // Store Parameter 1 on stack 680 nop.f 0 681 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address 682} 683{ .mib 684 stfs [GR_Parameter_Y] = FR_RESULT // Store Parameter 3 on stack 685 add GR_Parameter_Y = -16,GR_Parameter_Y 686 br.call.sptk b0=__libm_error_support# // Call error handling function 687};; 688 689{ .mmi 690 add GR_Parameter_RESULT = 48,sp 691 nop.m 0 692 nop.i 0 693};; 694 695{ .mmi 696 ldfs f8 = [GR_Parameter_RESULT] // Get return result off stack 697.restore sp 698 add sp = 64,sp // Restore stack pointer 699 mov b0 = GR_SAVE_B0 // Restore return address 700};; 701{ .mib 702 mov gp = GR_SAVE_GP // Restore gp 703 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs 704 br.ret.sptk b0 // Return 705};; 706 707LOCAL_LIBM_END(__libm_error_region) 708 709 710.type __libm_error_support#,@function 711.global __libm_error_support# 712