4 // Copyright (c) 2000 - 2003, Intel Corporation
5 // All rights reserved.
7 // Contributed 2000 by the Intel Numerics Group, Intel Corporation
9 // Redistribution and use in source and binary forms, with or without
10 // modification, are permitted provided that the following conditions are
13 // * Redistributions of source code must retain the above copyright
14 // notice, this list of conditions and the following disclaimer.
16 // * Redistributions in binary form must reproduce the above copyright
17 // notice, this list of conditions and the following disclaimer in the
18 // documentation and/or other materials provided with the distribution.
20 // * The name of Intel Corporation may not be used to endorse or promote
21 // products derived from this software without specific prior written
24 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
27 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
28 // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
29 // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
30 // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
31 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
32 // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
33 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
34 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 // Intel Corporation is the author of this code, and requests that all
37 // problem reports or change requests be submitted to it directly at
38 // http://www.intel.com/software/products/opensource/libraries/num.htm.
40 //*********************************************************************
43 // 05/21/01 Extracted logl and log10l from log1pl.s file, and optimized
45 // 06/20/01 Fixed error tag for x=-inf.
46 // 05/20/02 Cleaned up namespace and sf0 syntax
47 // 02/10/03 Reordered header: .section, .global, .proc, .align;
48 // used data8 for long double table values
50 //*********************************************************************
52 //*********************************************************************
54 // Function: Combined logl(x) and log10l(x) where
55 // logl(x) = ln(x), for double-extended precision x values
56 // log10l(x) = log (x), for double-extended precision x values
59 //*********************************************************************
63 // Floating-Point Registers: f8 (Input and Return Value)
66 // General Purpose Registers:
68 // r53-r56 (Used to pass arguments to error handling routine)
70 // Predicate Registers: p6-p14
72 //*********************************************************************
74 // IEEE Special Conditions:
76 // Denormal fault raised on denormal inputs
77 // Overflow exceptions cannot occur
78 // Underflow exceptions raised when appropriate for log1p
79 // (Error Handling Routine called for underflow)
80 // Inexact raised when appropriate by algorithm
87 // logl(EM_special Values) = QNaN
89 // log10l(-inf) = QNaN
90 // log10l(+/-0) = -inf
91 // log10l(SNaN) = QNaN
92 // log10l(QNaN) = QNaN
93 // log10l(EM_special Values) = QNaN
95 //*********************************************************************
99 // The method consists of two cases.
101 // If |X-1| < 2^(-7) use case log_near1;
102 // else use case log_regular;
106 // logl( 1 + X ) can be approximated by a simple polynomial
107 // in W = X-1. This polynomial resembles the truncated Taylor
108 // series W - W^/2 + W^3/3 - ...
112 // Here we use a table lookup method. The basic idea is that in
113 // order to compute logl(Arg) for an argument Arg in [1,2), we
114 // construct a value G such that G*Arg is close to 1 and that
115 // logl(1/G) is obtainable easily from a table of values calculated
118 // logl(Arg) = logl(1/G) + logl(G*Arg)
119 // = logl(1/G) + logl(1 + (G*Arg - 1))
121 // Because |G*Arg - 1| is small, the second term on the right hand
122 // side can be approximated by a short polynomial. We elaborate
123 // this method in four steps.
125 // Step 0: Initialization
127 // We need to calculate logl( X ). Obtain N, S_hi such that
129 // X = 2^N * S_hi exactly
131 // where S_hi in [1,2)
133 // Step 1: Argument Reduction
135 // Based on S_hi, obtain G_1, G_2, G_3 from a table and calculate
137 // G := G_1 * G_2 * G_3
138 // r := (G * S_hi - 1)
140 // These G_j's have the property that the product is exactly
141 // representable and that |r| < 2^(-12) as a result.
143 // Step 2: Approximation
146 // logl(1 + r) is approximated by a short polynomial poly(r).
148 // Step 3: Reconstruction
151 // Finally, logl( X ) is given by
153 // logl( X ) = logl( 2^N * S_hi )
154 // ~=~ N*logl(2) + logl(1/G) + logl(1 + r)
155 // ~=~ N*logl(2) + logl(1/G) + poly(r).
157 // **** Algorithm ****
161 // Here we compute a simple polynomial. To exploit parallelism, we split
162 // the polynomial into two portions.
168 // Y_hi := W + Wsq*(P_1 + W*(P_2 + W*(P_3 + W*P_4))
169 // Y_lo := W6*(P_5 + W*(P_6 + W*(P_7 + W*P_8)))
173 // We present the algorithm in four steps.
175 // Step 0. Initialization
176 // ----------------------
179 // N := unbaised exponent of Z
180 // S_hi := 2^(-N) * Z
182 // Step 1. Argument Reduction
183 // --------------------------
187 // Z = 2^N * S_hi = 2^N * 1.d_1 d_2 d_3 ... d_63
189 // We obtain G_1, G_2, G_3 by the following steps.
192 // Define X_0 := 1.d_1 d_2 ... d_14. This is extracted
195 // Define A_1 := 1.d_1 d_2 d_3 d_4. This is X_0 truncated
198 // Define index_1 := [ d_1 d_2 d_3 d_4 ].
200 // Fetch Z_1 := (1/A_1) rounded UP in fixed point with
201 // fixed point lsb = 2^(-15).
202 // Z_1 looks like z_0.z_1 z_2 ... z_15
203 // Note that the fetching is done using index_1.
204 // A_1 is actually not needed in the implementation
205 // and is used here only to explain how is the value
208 // Fetch G_1 := (1/A_1) truncated to 21 sig. bits.
209 // floating pt. Again, fetching is done using index_1. A_1
210 // explains how G_1 is defined.
212 // Calculate X_1 := X_0 * Z_1 truncated to lsb = 2^(-14)
213 // = 1.0 0 0 0 d_5 ... d_14
214 // This is accomplished by integer multiplication.
215 // It is proved that X_1 indeed always begin
216 // with 1.0000 in fixed point.
219 // Define A_2 := 1.0 0 0 0 d_5 d_6 d_7 d_8. This is X_1
220 // truncated to lsb = 2^(-8). Similar to A_1,
221 // A_2 is not needed in actual implementation. It
222 // helps explain how some of the values are defined.
224 // Define index_2 := [ d_5 d_6 d_7 d_8 ].
226 // Fetch Z_2 := (1/A_2) rounded UP in fixed point with
227 // fixed point lsb = 2^(-15). Fetch done using index_2.
228 // Z_2 looks like z_0.z_1 z_2 ... z_15
230 // Fetch G_2 := (1/A_2) truncated to 21 sig. bits.
233 // Calculate X_2 := X_1 * Z_2 truncated to lsb = 2^(-14)
234 // = 1.0 0 0 0 0 0 0 0 d_9 d_10 ... d_14
235 // This is accomplished by integer multiplication.
236 // It is proved that X_2 indeed always begin
237 // with 1.00000000 in fixed point.
240 // Define A_3 := 1.0 0 0 0 0 0 0 0 d_9 d_10 d_11 d_12 d_13 1.
241 // This is 2^(-14) + X_2 truncated to lsb = 2^(-13).
243 // Define index_3 := [ d_9 d_10 d_11 d_12 d_13 ].
245 // Fetch G_3 := (1/A_3) truncated to 21 sig. bits.
246 // floating pt. Fetch is done using index_3.
248 // Compute G := G_1 * G_2 * G_3.
250 // This is done exactly since each of G_j only has 21 sig. bits.
257 // Step 2. Approximation
258 // ---------------------
260 // This step computes an approximation to logl( 1 + r ) where r is the
261 // reduced argument just obtained. It is proved that |r| <= 1.9*2^(-13);
262 // thus logl(1+r) can be approximated by a short polynomial:
264 // logl(1+r) ~=~ poly = r + Q1 r^2 + ... + Q4 r^5
267 // Step 3. Reconstruction
268 // ----------------------
270 // This step computes the desired result of logl(X):
272 // logl(X) = logl( 2^N * S_hi )
273 // = N*logl(2) + logl( S_hi )
274 // = N*logl(2) + logl(1/G) +
275 // logl(1 + G*S_hi - 1 )
277 // logl(2), logl(1/G_j) are stored as pairs of (single,double) numbers:
278 // log2_hi, log2_lo, log1byGj_hi, log1byGj_lo. The high parts are
279 // single-precision numbers and the low parts are double precision
280 // numbers. These have the property that
282 // N*log2_hi + SUM ( log1byGj_hi )
284 // is computable exactly in double-extended precision (64 sig. bits).
287 // Y_hi := N*log2_hi + SUM ( log1byGj_hi )
288 // Y_lo := poly_hi + [ poly_lo +
289 // ( SUM ( log1byGj_lo ) + N*log2_lo ) ]
295 // ************* DO NOT CHANGE THE ORDER OF THESE TABLES *************
297 // P_8, P_7, P_6, P_5, P_4, P_3, P_2, and P_1
299 LOCAL_OBJECT_START(Constants_P)
300 data8 0xE3936754EFD62B15,0x00003FFB
301 data8 0x8003B271A5E56381,0x0000BFFC
302 data8 0x9249248C73282DB0,0x00003FFC
303 data8 0xAAAAAA9F47305052,0x0000BFFC
304 data8 0xCCCCCCCCCCD17FC9,0x00003FFC
305 data8 0x8000000000067ED5,0x0000BFFD
306 data8 0xAAAAAAAAAAAAAAAA,0x00003FFD
307 data8 0xFFFFFFFFFFFFFFFE,0x0000BFFD
308 LOCAL_OBJECT_END(Constants_P)
310 // log2_hi, log2_lo, Q_4, Q_3, Q_2, and Q_1
312 LOCAL_OBJECT_START(Constants_Q)
313 data8 0xB172180000000000,0x00003FFE
314 data8 0x82E308654361C4C6,0x0000BFE2
315 data8 0xCCCCCAF2328833CB,0x00003FFC
316 data8 0x80000077A9D4BAFB,0x0000BFFD
317 data8 0xAAAAAAAAAAABE3D2,0x00003FFD
318 data8 0xFFFFFFFFFFFFDAB7,0x0000BFFD
319 LOCAL_OBJECT_END(Constants_Q)
321 // 1/ln10_hi, 1/ln10_lo
323 LOCAL_OBJECT_START(Constants_1_by_LN10)
324 data8 0xDE5BD8A937287195,0x00003FFD
325 data8 0xD56EAABEACCF70C8,0x00003FBB
326 LOCAL_OBJECT_END(Constants_1_by_LN10)
331 LOCAL_OBJECT_START(Constants_Z_1)
348 LOCAL_OBJECT_END(Constants_Z_1)
350 // G1 and H1 - IEEE single and h1 - IEEE double
352 LOCAL_OBJECT_START(Constants_G_H_h1)
353 data4 0x3F800000,0x00000000
354 data8 0x0000000000000000
355 data4 0x3F70F0F0,0x3D785196
356 data8 0x3DA163A6617D741C
357 data4 0x3F638E38,0x3DF13843
358 data8 0x3E2C55E6CBD3D5BB
359 data4 0x3F579430,0x3E2FF9A0
360 data8 0xBE3EB0BFD86EA5E7
361 data4 0x3F4CCCC8,0x3E647FD6
362 data8 0x3E2E6A8C86B12760
363 data4 0x3F430C30,0x3E8B3AE7
364 data8 0x3E47574C5C0739BA
365 data4 0x3F3A2E88,0x3EA30C68
366 data8 0x3E20E30F13E8AF2F
367 data4 0x3F321640,0x3EB9CEC8
368 data8 0xBE42885BF2C630BD
369 data4 0x3F2AAAA8,0x3ECF9927
370 data8 0x3E497F3497E577C6
371 data4 0x3F23D708,0x3EE47FC5
372 data8 0x3E3E6A6EA6B0A5AB
373 data4 0x3F1D89D8,0x3EF8947D
374 data8 0xBDF43E3CD328D9BE
375 data4 0x3F17B420,0x3F05F3A1
376 data8 0x3E4094C30ADB090A
377 data4 0x3F124920,0x3F0F4303
378 data8 0xBE28FBB2FC1FE510
379 data4 0x3F0D3DC8,0x3F183EBF
380 data8 0x3E3A789510FDE3FA
381 data4 0x3F088888,0x3F20EC80
382 data8 0x3E508CE57CC8C98F
383 data4 0x3F042108,0x3F29516A
384 data8 0xBE534874A223106C
385 LOCAL_OBJECT_END(Constants_G_H_h1)
389 LOCAL_OBJECT_START(Constants_Z_2)
406 LOCAL_OBJECT_END(Constants_Z_2)
408 // G2 and H2 - IEEE single and h2 - IEEE double
410 LOCAL_OBJECT_START(Constants_G_H_h2)
411 data4 0x3F800000,0x00000000
412 data8 0x0000000000000000
413 data4 0x3F7F00F8,0x3B7F875D
414 data8 0x3DB5A11622C42273
415 data4 0x3F7E03F8,0x3BFF015B
416 data8 0x3DE620CF21F86ED3
417 data4 0x3F7D08E0,0x3C3EE393
418 data8 0xBDAFA07E484F34ED
419 data4 0x3F7C0FC0,0x3C7E0586
420 data8 0xBDFE07F03860BCF6
421 data4 0x3F7B1880,0x3C9E75D2
422 data8 0x3DEA370FA78093D6
423 data4 0x3F7A2328,0x3CBDC97A
424 data8 0x3DFF579172A753D0
425 data4 0x3F792FB0,0x3CDCFE47
426 data8 0x3DFEBE6CA7EF896B
427 data4 0x3F783E08,0x3CFC15D0
428 data8 0x3E0CF156409ECB43
429 data4 0x3F774E38,0x3D0D874D
430 data8 0xBE0B6F97FFEF71DF
431 data4 0x3F766038,0x3D1CF49B
432 data8 0xBE0804835D59EEE8
433 data4 0x3F757400,0x3D2C531D
434 data8 0x3E1F91E9A9192A74
435 data4 0x3F748988,0x3D3BA322
436 data8 0xBE139A06BF72A8CD
437 data4 0x3F73A0D0,0x3D4AE46F
438 data8 0x3E1D9202F8FBA6CF
439 data4 0x3F72B9D0,0x3D5A1756
440 data8 0xBE1DCCC4BA796223
441 data4 0x3F71D488,0x3D693B9D
442 data8 0xBE049391B6B7C239
443 LOCAL_OBJECT_END(Constants_G_H_h2)
445 // G3 and H3 - IEEE single and h3 - IEEE double
447 LOCAL_OBJECT_START(Constants_G_H_h3)
448 data4 0x3F7FFC00,0x38800100
449 data8 0x3D355595562224CD
450 data4 0x3F7FF400,0x39400480
451 data8 0x3D8200A206136FF6
452 data4 0x3F7FEC00,0x39A00640
453 data8 0x3DA4D68DE8DE9AF0
454 data4 0x3F7FE400,0x39E00C41
455 data8 0xBD8B4291B10238DC
456 data4 0x3F7FDC00,0x3A100A21
457 data8 0xBD89CCB83B1952CA
458 data4 0x3F7FD400,0x3A300F22
459 data8 0xBDB107071DC46826
460 data4 0x3F7FCC08,0x3A4FF51C
461 data8 0x3DB6FCB9F43307DB
462 data4 0x3F7FC408,0x3A6FFC1D
463 data8 0xBD9B7C4762DC7872
464 data4 0x3F7FBC10,0x3A87F20B
465 data8 0xBDC3725E3F89154A
466 data4 0x3F7FB410,0x3A97F68B
467 data8 0xBD93519D62B9D392
468 data4 0x3F7FAC18,0x3AA7EB86
469 data8 0x3DC184410F21BD9D
470 data4 0x3F7FA420,0x3AB7E101
471 data8 0xBDA64B952245E0A6
472 data4 0x3F7F9C20,0x3AC7E701
473 data8 0x3DB4B0ECAABB34B8
474 data4 0x3F7F9428,0x3AD7DD7B
475 data8 0x3D9923376DC40A7E
476 data4 0x3F7F8C30,0x3AE7D474
477 data8 0x3DC6E17B4F2083D3
478 data4 0x3F7F8438,0x3AF7CBED
479 data8 0x3DAE314B811D4394
480 data4 0x3F7F7C40,0x3B03E1F3
481 data8 0xBDD46F21B08F2DB1
482 data4 0x3F7F7448,0x3B0BDE2F
483 data8 0xBDDC30A46D34522B
484 data4 0x3F7F6C50,0x3B13DAAA
485 data8 0x3DCB0070B1F473DB
486 data4 0x3F7F6458,0x3B1BD766
487 data8 0xBDD65DDC6AD282FD
488 data4 0x3F7F5C68,0x3B23CC5C
489 data8 0xBDCDAB83F153761A
490 data4 0x3F7F5470,0x3B2BC997
491 data8 0xBDDADA40341D0F8F
492 data4 0x3F7F4C78,0x3B33C711
493 data8 0x3DCD1BD7EBC394E8
494 data4 0x3F7F4488,0x3B3BBCC6
495 data8 0xBDC3532B52E3E695
496 data4 0x3F7F3C90,0x3B43BAC0
497 data8 0xBDA3961EE846B3DE
498 data4 0x3F7F34A0,0x3B4BB0F4
499 data8 0xBDDADF06785778D4
500 data4 0x3F7F2CA8,0x3B53AF6D
501 data8 0x3DCC3ED1E55CE212
502 data4 0x3F7F24B8,0x3B5BA620
503 data8 0xBDBA31039E382C15
504 data4 0x3F7F1CC8,0x3B639D12
505 data8 0x3D635A0B5C5AF197
506 data4 0x3F7F14D8,0x3B6B9444
507 data8 0xBDDCCB1971D34EFC
508 data4 0x3F7F0CE0,0x3B7393BC
509 data8 0x3DC7450252CD7ADA
510 data4 0x3F7F04F0,0x3B7B8B6D
511 data8 0xBDB68F177D7F2A42
512 LOCAL_OBJECT_END(Constants_G_H_h3)
515 // Floating Point Registers
574 FR_Output_X_tmp = f76
581 // General Purpose Registers
609 // Added for unwind support
617 GR_Parameter_RESULT = r55
618 GR_Parameter_TAG = r56
622 GLOBAL_IEEE754_ENTRY(logl)
624 alloc r32 = ar.pfs,0,21,4,0
625 fclass.m p6, p0 = FR_Input_X, 0x1E3 // Test for natval, nan, inf
626 cmp.eq p7, p14 = r0, r0 // Set p7 if logl
629 addl GR_ad_z_1 = @ltoff(Constants_Z_1#),gp
630 fnorm.s1 FR_X_Prime = FR_Input_X // Normalize x
631 br.cond.sptk LOGL_BEGIN
635 GLOBAL_IEEE754_END(logl)
638 GLOBAL_IEEE754_ENTRY(log10l)
640 alloc r32 = ar.pfs,0,21,4,0
641 fclass.m p6, p0 = FR_Input_X, 0x1E3 // Test for natval, nan, inf
642 cmp.ne p7, p14 = r0, r0 // Set p14 if log10l
645 addl GR_ad_z_1 = @ltoff(Constants_Z_1#),gp
646 fnorm.s1 FR_X_Prime = FR_Input_X // Normalize x
652 // Common code for logl and log10
655 ld8 GR_ad_z_1 = [GR_ad_z_1] // Get pointer to Constants_Z_1
656 fclass.m p10, p0 = FR_Input_X, 0x0b // Test for denormal
657 mov GR_exp_2tom7 = 0x0fff8 // Exponent of 2^-7
662 getf.sig GR_signif = FR_Input_X // Get significand of x
663 fcmp.eq.s1 p9, p0 = FR_Input_X, f1 // Test for x=1.0
664 (p6) br.cond.spnt LOGL_64_special // Branch for nan, inf, natval
669 add GR_ad_tbl_1 = 0x040, GR_ad_z_1 // Point to Constants_G_H_h1
670 fcmp.lt.s1 p13, p0 = FR_Input_X, f0 // Test for x<0
671 add GR_ad_p = -0x100, GR_ad_z_1 // Point to Constants_P
674 add GR_ad_z_2 = 0x140, GR_ad_z_1 // Point to Constants_Z_2
675 add GR_ad_tbl_2 = 0x180, GR_ad_z_1 // Point to Constants_G_H_h2
676 (p10) br.cond.spnt LOGL_64_denormal // Branch for denormal
682 add GR_ad_q = 0x080, GR_ad_p // Point to Constants_Q
683 fcmp.eq.s1 p8, p0 = FR_Input_X, f0 // Test for x=0
684 extr.u GR_Index1 = GR_signif, 59, 4 // Get high 4 bits of signif
687 add GR_ad_tbl_3 = 0x280, GR_ad_z_1 // Point to Constants_G_H_h3
688 (p9) fma.s0 f8 = FR_Input_X, f0, f0 // If x=1, return +0.0
689 (p9) br.ret.spnt b0 // Exit if x=1
694 shladd GR_ad_z_1 = GR_Index1, 2, GR_ad_z_1 // Point to Z_1
695 fclass.nm p10, p0 = FR_Input_X, 0x1FF // Test for unsupported
696 extr.u GR_X_0 = GR_signif, 49, 15 // Get high 15 bits of significand
699 ldfe FR_P8 = [GR_ad_p],16 // Load P_8 for near1 path
700 fsub.s1 FR_W = FR_X_Prime, f1 // W = x - 1
701 add GR_ad_ln10 = 0x060, GR_ad_q // Point to Constants_1_by_LN10
706 ld4 GR_Z_1 = [GR_ad_z_1] // Load Z_1
708 mov GR_exp_mask = 0x1FFFF // Create exponent mask
711 shladd GR_ad_tbl_1 = GR_Index1, 4, GR_ad_tbl_1 // Point to G_1
712 mov GR_Bias = 0x0FFFF // Create exponent bias
713 (p13) br.cond.spnt LOGL_64_negative // Branch if x<0
718 ldfps FR_G, FR_H = [GR_ad_tbl_1],8 // Load G_1, H_1
719 fmerge.se FR_S_hi = f1,FR_X_Prime // Form |x|
720 (p8) br.cond.spnt LOGL_64_zero // Branch if x=0
725 getf.exp GR_N = FR_X_Prime // Get N = exponent of x
726 ldfd FR_h = [GR_ad_tbl_1] // Load h_1
727 (p10) br.cond.spnt LOGL_64_unsupported // Branch for unsupported type
732 ldfe FR_log2_hi = [GR_ad_q],16 // Load log2_hi
733 fcmp.eq.s0 p8, p0 = FR_Input_X, f0 // Dummy op to flag denormals
734 pmpyshr2.u GR_X_1 = GR_X_0,GR_Z_1,15 // Get bits 30-15 of X_0 * Z_1
739 // For performance, don't use result of pmpyshr2.u for 4 cycles.
742 ldfe FR_log2_lo = [GR_ad_q],16 // Load log2_lo
743 (p14) ldfe FR_1LN10_hi = [GR_ad_ln10],16 // If log10l, load 1/ln10_hi
744 sub GR_N = GR_N, GR_Bias
749 ldfe FR_Q4 = [GR_ad_q],16 // Load Q4
750 (p14) ldfe FR_1LN10_lo = [GR_ad_ln10] // If log10l, load 1/ln10_lo
756 ldfe FR_Q3 = [GR_ad_q],16 // Load Q3
757 setf.sig FR_float_N = GR_N // Put integer N into rightmost significand
763 getf.exp GR_M = FR_W // Get signexp of w = x - 1
764 ldfe FR_Q2 = [GR_ad_q],16 // Load Q2
765 extr.u GR_Index2 = GR_X_1, 6, 4 // Extract bits 6-9 of X_1
770 ldfe FR_Q1 = [GR_ad_q] // Load Q1
771 shladd GR_ad_z_2 = GR_Index2, 2, GR_ad_z_2 // Point to Z_2
772 add GR_ad_p2 = 0x30,GR_ad_p // Point to P_4
777 ld4 GR_Z_2 = [GR_ad_z_2] // Load Z_2
778 shladd GR_ad_tbl_2 = GR_Index2, 4, GR_ad_tbl_2 // Point to G_2
779 and GR_M = GR_exp_mask, GR_M // Get exponent of w = x - 1
784 ldfps FR_G2, FR_H2 = [GR_ad_tbl_2],8 // Load G_2, H_2
785 cmp.lt p8, p9 = GR_M, GR_exp_2tom7 // Test |x-1| < 2^-7
791 // p8 is for the near1 path: |x-1| < 2^-7
792 // p9 is for regular path: |x-1| >= 2^-7
795 ldfd FR_h2 = [GR_ad_tbl_2] // Load h_2
802 (p8) ldfe FR_P7 = [GR_ad_p],16 // Load P_7 for near1 path
803 (p8) ldfe FR_P4 = [GR_ad_p2],16 // Load P_4 for near1 path
804 (p9) pmpyshr2.u GR_X_2 = GR_X_1,GR_Z_2,15 // Get bits 30-15 of X_1 * Z_2
809 // For performance, don't use result of pmpyshr2.u for 4 cycles.
812 (p8) ldfe FR_P6 = [GR_ad_p],16 // Load P_6 for near1 path
813 (p8) ldfe FR_P3 = [GR_ad_p2],16 // Load P_3 for near1 path
819 (p8) ldfe FR_P5 = [GR_ad_p],16 // Load P_5 for near1 path
820 (p8) ldfe FR_P2 = [GR_ad_p2],16 // Load P_2 for near1 path
821 (p8) fmpy.s1 FR_wsq = FR_W, FR_W // wsq = w * w for near1 path
826 (p8) ldfe FR_P1 = [GR_ad_p2],16 ;; // Load P_1 for near1 path
828 (p9) extr.u GR_Index3 = GR_X_2, 1, 5 // Extract bits 1-5 of X_2
833 (p9) shladd GR_ad_tbl_3 = GR_Index3, 4, GR_ad_tbl_3 // Point to G_3
834 (p9) fcvt.xf FR_float_N = FR_float_N
840 (p9) ldfps FR_G3, FR_H3 = [GR_ad_tbl_3],8 // Load G_3, H_3
847 (p9) ldfd FR_h3 = [GR_ad_tbl_3] // Load h_3
848 (p9) fmpy.s1 FR_G = FR_G, FR_G2 // G = G_1 * G_2
853 (p9) fadd.s1 FR_H = FR_H, FR_H2 // H = H_1 + H_2
861 (p9) fadd.s1 FR_h = FR_h, FR_h2 // h = h_1 + h_2
867 (p8) fmpy.s1 FR_w4 = FR_wsq, FR_wsq // w4 = w^4 for near1 path
872 (p8) fma.s1 FR_p87 = FR_W, FR_P8, FR_P7 // p87 = w * P8 + P7
879 (p8) fma.s1 FR_p43 = FR_W, FR_P4, FR_P3 // p43 = w * P4 + P3
886 (p9) fmpy.s1 FR_G = FR_G, FR_G3 // G = (G_1 * G_2) * G_3
891 (p9) fadd.s1 FR_H = FR_H, FR_H3 // H = (H_1 + H_2) + H_3
898 (p9) fadd.s1 FR_h = FR_h, FR_h3 // h = (h_1 + h_2) + h_3
903 (p8) fmpy.s1 FR_w6 = FR_w4, FR_wsq // w6 = w^6 for near1 path
910 (p8) fma.s1 FR_p432 = FR_W, FR_p43, FR_P2 // p432 = w * p43 + P2
915 (p8) fma.s1 FR_p876 = FR_W, FR_p87, FR_P6 // p876 = w * p87 + P6
922 (p9) fms.s1 FR_r = FR_G, FR_S_hi, f1 // r = G * S_hi - 1
927 (p9) fma.s1 FR_Y_hi = FR_float_N, FR_log2_hi, FR_H // Y_hi = N * log2_hi + H
934 (p9) fma.s1 FR_h = FR_float_N, FR_log2_lo, FR_h // h = N * log2_lo + h
941 (p8) fma.s1 FR_p4321 = FR_W, FR_p432, FR_P1 // p4321 = w * p432 + P1
946 (p8) fma.s1 FR_p8765 = FR_W, FR_p876, FR_P5 // p8765 = w * p876 + P5
953 (p9) fma.s1 FR_poly_lo = FR_r, FR_Q4, FR_Q3 // poly_lo = r * Q4 + Q3
958 (p9) fmpy.s1 FR_rsq = FR_r, FR_r // rsq = r * r
965 (p8) fma.s1 FR_Y_lo = FR_wsq, FR_p4321, f0 // Y_lo = wsq * p4321
970 (p8) fma.s1 FR_Y_hi = FR_W, f1, f0 // Y_hi = w for near1 path
977 (p9) fma.s1 FR_poly_lo = FR_poly_lo, FR_r, FR_Q2 // poly_lo = poly_lo * r + Q2
982 (p9) fma.s1 FR_rcub = FR_rsq, FR_r, f0 // rcub = r^3
989 (p8) fma.s1 FR_Y_lo = FR_w6, FR_p8765,FR_Y_lo // Y_lo = w6 * p8765 + w2 * p4321
996 (p9) fma.s1 FR_poly_hi = FR_Q1, FR_rsq, FR_r // poly_hi = Q1 * rsq + r
1003 (p9) fma.s1 FR_poly_lo = FR_poly_lo, FR_rcub, FR_h // poly_lo = poly_lo*r^3 + h
1010 (p9) fadd.s1 FR_Y_lo = FR_poly_hi, FR_poly_lo // Y_lo = poly_hi + poly_lo
1015 // Remainder of code is common for near1 and regular paths
1018 (p7) fadd.s0 f8 = FR_Y_lo,FR_Y_hi // If logl, result=Y_lo+Y_hi
1023 (p14) fmpy.s1 FR_Output_X_tmp = FR_Y_lo,FR_1LN10_hi
1030 (p14) fma.s1 FR_Output_X_tmp = FR_Y_hi,FR_1LN10_lo,FR_Output_X_tmp
1037 (p14) fma.s0 f8 = FR_Y_hi,FR_1LN10_hi,FR_Output_X_tmp
1038 br.ret.sptk b0 // Common exit for 0 < x < inf
1046 // If x=+-0 raise divide by zero and return -inf
1049 (p7) mov GR_Parameter_TAG = 0
1050 fsub.s1 FR_Output_X_tmp = f0, f1
1056 (p14) mov GR_Parameter_TAG = 6
1057 frcpa.s0 FR_Output_X_tmp, p8 = FR_Output_X_tmp, f0
1058 br.cond.sptk __libm_error_region
1065 fclass.m.unc p8, p0 = FR_Input_X, 0x1E1 // Test for natval, nan, +inf
1071 // For SNaN raise invalid and return QNaN.
1072 // For QNaN raise invalid and return QNaN.
1073 // For +Inf return +Inf.
1077 (p8) fmpy.s0 f8 = FR_Input_X, f1
1078 (p8) br.ret.sptk b0 // Return for natval, nan, +inf
1083 // For -Inf raise invalid and return QNaN.
1086 (p7) mov GR_Parameter_TAG = 1
1093 (p14) mov GR_Parameter_TAG = 7
1094 fmpy.s0 FR_Output_X_tmp = FR_Input_X, f0
1095 br.cond.sptk __libm_error_region
1099 // Here if x denormal or unnormal
1102 getf.sig GR_signif = FR_X_Prime // Get significand of normalized input
1109 getf.exp GR_N = FR_X_Prime // Get exponent of normalized input
1111 br.cond.sptk LOGL_64_COMMON // Branch back to common code
1115 LOGL_64_unsupported:
1117 // Return generated NaN or other value.
1121 fmpy.s0 f8 = FR_Input_X, f0
1126 // Here if -inf < x < 0
1129 // Deal with x < 0 in a special way - raise
1130 // invalid and produce QNaN indefinite.
1133 (p7) mov GR_Parameter_TAG = 1
1134 frcpa.s0 FR_Output_X_tmp, p8 = f0, f0
1140 (p14) mov GR_Parameter_TAG = 7
1142 br.cond.sptk __libm_error_region
1147 GLOBAL_IEEE754_END(log10l)
1149 LOCAL_LIBM_ENTRY(__libm_error_region)
1152 add GR_Parameter_Y=-32,sp // Parameter 2 value
1154 .save ar.pfs,GR_SAVE_PFS
1155 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
1159 add sp=-64,sp // Create new stack
1161 mov GR_SAVE_GP=gp // Save gp
1164 stfe [GR_Parameter_Y] = FR_Y,16 // Save Parameter 2 on stack
1165 add GR_Parameter_X = 16,sp // Parameter 1 address
1166 .save b0, GR_SAVE_B0
1167 mov GR_SAVE_B0=b0 // Save b0
1171 stfe [GR_Parameter_X] = FR_X // Store Parameter 1 on stack
1172 add GR_Parameter_RESULT = 0,GR_Parameter_Y
1173 nop.b 0 // Parameter 3 address
1176 stfe [GR_Parameter_Y] = FR_RESULT // Store Parameter 3 on stack
1177 add GR_Parameter_Y = -16,GR_Parameter_Y
1178 br.call.sptk b0=__libm_error_support# // Call error handling function
1183 add GR_Parameter_RESULT = 48,sp
1186 ldfe f8 = [GR_Parameter_RESULT] // Get return result off stack
1188 add sp = 64,sp // Restore stack pointer
1189 mov b0 = GR_SAVE_B0 // Restore return address
1192 mov gp = GR_SAVE_GP // Restore gp
1193 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
1194 br.ret.sptk b0 // Return
1197 LOCAL_LIBM_END(__libm_error_region#)
1199 .type __libm_error_support#,@function
1200 .global __libm_error_support#
projects / thirdparty / glibc.git / blob