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.
41 //==============================================================
42 // 02/02/00 Initial version
43 // 03/03/00 Modified to conform to C9X, and improve speed of main path
44 // 03/14/00 Fixed case where x is a power of 2, and x > y, improved speed
45 // 04/04/00 Unwind support added
46 // 05/12/00 Fixed erroneous denormal flag setting for exponent change cases 1,3
47 // 08/15/00 Bundle added after call to __libm_error_support to properly
48 // set [the previously overwritten] GR_Parameter_RESULT.
49 // 09/09/00 Updated fcmp so that qnans do not raise invalid.
50 // 12/15/00 Fixed case of smallest long double normal to largest denormal,
51 // now adhere to C99 for two zero args, and fixed flag settings
53 // 05/20/02 Cleaned up namespace and sf0 syntax
54 // 02/10/03 Reordered header: .section, .global, .proc, .align
57 //==============================================================
58 // long double nextafterl( long double x, long double y );
59 // input floating point f8, f9
60 // output floating point f8
63 //==============================================================
64 nextafter_GR_max_pexp = r14
65 nextafter_GR_min_pexp = r15
66 nextafter_GR_exp = r16
67 nextafter_GR_sig = r17
68 nextafter_GR_lnorm_sig = r18
69 nextafter_GR_sign_mask = r19
70 nextafter_GR_exp_mask = r20
71 nextafter_GR_sden_sig = r21
72 nextafter_GR_new_sig = r22
73 nextafter_GR_new_exp = r23
74 nextafter_GR_lden_sig = r24
75 nextafter_GR_snorm_sig = r25
76 nextafter_GR_exp1 = r26
77 nextafter_GR_x_exp = r27
78 // r36-39 parameters for libm_error_support
86 GR_Parameter_RESULT = r38
88 NEXTAFTER_lnorm_sig = f10
89 NEXTAFTER_lnorm_exp = f11
91 NEXTAFTER_sden_sig = f13
92 NEXTAFTER_den_exp = f14
94 NEXTAFTER_snorm_exp = f32
95 NEXTAFTER_save_f8 = f33
96 NEXTAFTER_new_exp = f34
97 NEXTAFTER_new_sig = f35
98 NEXTAFTER_lden_sig = f36
99 NEXTAFTER_snorm_sig = f37
104 // Overview of operation
105 //==============================================================
106 // nextafterl determines the next representable value
107 // after x in the direction of y.
111 GLOBAL_LIBM_ENTRY(nextafterl)
113 // Extract signexp from x
114 // Is x < y ? p10 if yes, p11 if no
115 // Form smallest denormal significand = ulp size
117 getf.exp nextafter_GR_exp = f8
118 fcmp.lt.s1 p10,p11 = f8, f9
119 addl nextafter_GR_sden_sig = 0x1, r0
121 // Form largest normal significand 0xffffffffffffffff
122 // Form smallest normal exponent
124 addl nextafter_GR_lnorm_sig = -0x1,r0
126 addl nextafter_GR_min_pexp = 0x0c001, r0 ;;
129 // Extract significand from x
130 // Is x=y? This fcmp also sets Invalid and Denormal if required
131 // Form largest normal exponent
133 getf.sig nextafter_GR_sig = f8
134 fcmp.eq.s0 p6,p0 = f8, f9
135 addl nextafter_GR_max_pexp = 0x13ffe, r0
137 // Move largest normal significand to fp reg for special cases
139 setf.sig NEXTAFTER_lnorm_sig = nextafter_GR_lnorm_sig
141 addl nextafter_GR_sign_mask = 0x20000, r0 ;;
144 // Move smallest denormal significand and exp to fp regs
146 // Set p12 and p13 based on whether significand increases or decreases
147 // It increases (p12 set) if x<y and x>=0 or if x>y and x<0
148 // It decreases (p13 set) if x<y and x<0 or if x>y and x>=0
150 setf.sig NEXTAFTER_sden_sig = nextafter_GR_sden_sig
151 fclass.m p8,p0 = f8, 0xc3
152 (p10) cmp.lt p12,p13 = nextafter_GR_exp, nextafter_GR_sign_mask
154 // Move smallest normal exp to fp regs
156 setf.exp NEXTAFTER_snorm_exp = nextafter_GR_min_pexp
158 (p11) cmp.ge p12,p13 = nextafter_GR_exp, nextafter_GR_sign_mask ;;
161 .pred.rel "mutex",p12,p13
163 // Form expected new significand, adding or subtracting 1 ulp increment
164 // If x=y set result to y
165 // Form smallest normal significand and largest denormal significand
167 (p12) add nextafter_GR_new_sig = nextafter_GR_sig, nextafter_GR_sden_sig
168 (p6) fmerge.s f8=f9,f9
169 dep.z nextafter_GR_snorm_sig = 1,63,1 // 0x8000000000000000
172 (p13) sub nextafter_GR_new_sig = nextafter_GR_sig, nextafter_GR_sden_sig
173 movl nextafter_GR_lden_sig = 0x7fffffffffffffff ;;
176 // Move expected result significand and signexp to fp regs
178 // Form new exponent in case result exponent needs incrementing or decrementing
180 setf.exp NEXTAFTER_new_exp = nextafter_GR_exp
181 fclass.m p9,p0 = f9, 0xc3
182 (p12) add nextafter_GR_exp1 = 1, nextafter_GR_exp
185 setf.sig NEXTAFTER_new_sig = nextafter_GR_new_sig
186 (p13) add nextafter_GR_exp1 = -1, nextafter_GR_exp
187 (p6) br.ret.spnt b0 ;; // Exit if x=y
190 // Move largest normal signexp to fp reg for special cases
193 setf.exp NEXTAFTER_lnorm_exp = nextafter_GR_max_pexp
194 fclass.m p7,p0 = f8, 0x7
198 setf.exp NEXTAFTER_den_exp = nextafter_GR_min_pexp
199 (p8) fma.s0 f8 = f8,f1,f9
200 (p8) br.ret.spnt b0 ;; // Exit if x=nan
203 // Move exp+-1 and smallest normal significand to fp regs for special cases
206 setf.exp NEXTAFTER_exp1 = nextafter_GR_exp1
207 fclass.m p6,p0 = f8, 0x23
208 addl nextafter_GR_exp_mask = 0x1ffff, r0
211 setf.sig NEXTAFTER_snorm_sig = nextafter_GR_snorm_sig
212 (p9) fma.s0 f8 = f8,f1,f9
213 (p9) br.ret.spnt b0 ;; // Exit if y=nan
216 // Move largest denormal significand to fp regs for special cases
219 setf.sig NEXTAFTER_lden_sig = nextafter_GR_lden_sig
220 mov NEXTAFTER_save_f8 = f8
221 (p7) br.cond.spnt NEXTAFTER_ZERO ;; // Exit if x=0
224 // Mask off the sign to get x_exp
226 and nextafter_GR_x_exp = nextafter_GR_exp_mask, nextafter_GR_exp
228 (p6) br.cond.spnt NEXTAFTER_INF ;; // Exit if x=inf
231 // Check 5 special cases when significand rolls over:
232 // 1 sig size incr, x_sig=max_sig, x_exp < max_exp
233 // Set p6, result is sig=min_sig, exp++
234 // 2 sig size incr, x_sig=max_sig, x_exp >= max_exp
235 // Set p7, result is inf, signal overflow
236 // 3 sig size decr, x_sig=min_sig, x_exp > min_exp
237 // Set p8, result is sig=max_sig, exp--
238 // 4 sig size decr, x_sig=min_sig, x_exp = min_exp
239 // Set p9, result is sig=max_den_sig, exp same, signal underflow and inexact
240 // 5 sig size decr, x_sig=min_den_sig, x_exp = min_exp
241 // Set p10, result is zero, sign of x, signal underflow and inexact
244 (p12) cmp.eq.unc p6,p0 = nextafter_GR_new_sig, r0
245 (p13) cmp.eq.unc p9,p10 = nextafter_GR_new_sig, nextafter_GR_lden_sig
251 (p6) cmp.lt.unc p6,p7 = nextafter_GR_x_exp, nextafter_GR_max_pexp
252 (p10) cmp.eq.unc p10,p0 = nextafter_GR_new_sig, r0
253 (p9) cmp.le.unc p9,p8 = nextafter_GR_x_exp, nextafter_GR_min_pexp
257 // Create small normal in case need to generate underflow flag
260 fmerge.se NEXTAFTER_tmp = NEXTAFTER_snorm_exp, NEXTAFTER_lnorm_sig
263 // Branch if cases 1, 2, 3
265 (p6) br.cond.spnt NEXTAFTER_EXPUP
266 (p7) br.cond.spnt NEXTAFTER_OVERFLOW
267 (p8) br.cond.spnt NEXTAFTER_EXPDOWN ;;
270 // Branch if cases 4, 5
273 (p9) br.cond.spnt NEXTAFTER_NORM_TO_DENORM
274 (p10) br.cond.spnt NEXTAFTER_UNDERFLOW_TO_ZERO
278 // Here if no special cases
279 // Set p6 if result will be a denormal, so can force underflow flag
280 // Case 1: x_exp=min_exp, x_sig=unnormalized
281 // Case 2: x_exp<min_exp
283 cmp.lt p6,p7 = nextafter_GR_x_exp, nextafter_GR_min_pexp
284 fmerge.se f8 = NEXTAFTER_new_exp, NEXTAFTER_new_sig
291 (p6) tbit.z p6,p0 = nextafter_GR_new_sig, 63 ;;
294 NEXTAFTER_COMMON_FINISH:
295 // Force underflow and inexact if denormal result
298 (p6) fma.s0 NEXTAFTER_tmp = NEXTAFTER_tmp,NEXTAFTER_tmp,f0
302 // Final normalization to result precision and exit
312 cmp.lt p6,p7 = nextafter_GR_x_exp, nextafter_GR_min_pexp
313 fmerge.se f8 = NEXTAFTER_exp1, NEXTAFTER_snorm_sig
314 br.cond.sptk NEXTAFTER_COMMON_FINISH ;;
319 cmp.lt p6,p7 = nextafter_GR_x_exp, nextafter_GR_min_pexp
320 fmerge.se f8 = NEXTAFTER_exp1, NEXTAFTER_lnorm_sig
321 br.cond.sptk NEXTAFTER_COMMON_FINISH ;;
324 NEXTAFTER_NORM_TO_DENORM:
327 fmerge.se f8 = NEXTAFTER_exp1, NEXTAFTER_lden_sig
330 // Force underflow and inexact
333 fma.s0 NEXTAFTER_tmp = NEXTAFTER_tmp,NEXTAFTER_tmp,f0
337 NEXTAFTER_UNDERFLOW_TO_ZERO:
340 fmerge.s f8 = NEXTAFTER_save_f8,f0
341 br.cond.sptk NEXTAFTER_COMMON_FINISH ;;
345 // Here if f8 is +- infinity
347 // if f8 is +inf, no matter what y is return largest long double
348 // if f8 is -inf, no matter what y is return -largest long double
350 // Create largest long double
353 fmerge.se NEXTAFTER_lnorm = NEXTAFTER_lnorm_exp,NEXTAFTER_lnorm_sig
359 fmerge.s f8 = f8,NEXTAFTER_lnorm
365 // Here if f8 is +- zero
367 // if f8 is zero and y is +, return + smallest long double denormal
368 // if f8 is zero and y is -, return - smallest long double denormal
372 fmerge.se NEXTAFTER_sden = f0,NEXTAFTER_sden_sig
376 // Create small normal to generate underflow flag
379 fmerge.se NEXTAFTER_tmp = NEXTAFTER_snorm_exp, NEXTAFTER_lnorm_sig
383 // Add correct sign from direction arg
386 fmerge.s f8 = f9,NEXTAFTER_sden
390 // Force underflow and inexact flags
393 fma.s0 NEXTAFTER_tmp = NEXTAFTER_tmp,NEXTAFTER_tmp,f0
397 GLOBAL_LIBM_END(nextafterl)
398 // Stack operations when calling error support.
399 // (1) (2) (3) (call) (4)
400 // sp -> + psp -> + psp -> + sp -> +
402 // | | <- GR_Y R3 ->| <- GR_RESULT | -> f8
404 // | <-GR_Y Y2->| Y2 ->| <- GR_Y |
406 // | | <- GR_X X1 ->| |
408 // sp-64 -> + sp -> + sp -> + +
409 // save ar.pfs save b0 restore gp
410 // save gp restore ar.pfs
414 LOCAL_LIBM_ENTRY(__libm_error_region)
416 // Here if f8 is finite, but result will be infinite
417 // Use frcpa to generate infinity of correct sign
418 // Call error support to report possible range error
422 alloc r32=ar.pfs,2,2,4,0
423 frcpa.s1 f8,p6 = NEXTAFTER_save_f8, f0
427 // Create largest long double
430 fmerge.se NEXTAFTER_lnorm = NEXTAFTER_lnorm_exp,NEXTAFTER_lnorm_sig
434 // Force overflow and inexact flags to be set
436 mov r39 = 153 // Error code
437 fma.s0 NEXTAFTER_tmp = NEXTAFTER_lnorm,NEXTAFTER_lnorm,f0
444 add GR_Parameter_Y=-32,sp // Parameter 2 value
446 .save ar.pfs,GR_SAVE_PFS
447 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
451 add sp=-64,sp // Create new stack
453 mov GR_SAVE_GP=gp // Save gp
459 stfe [GR_Parameter_Y] = f9,16 // STORE Parameter 2 on stack
460 add GR_Parameter_X = 16,sp // Parameter 1 address
462 mov GR_SAVE_B0=b0 // Save b0
468 stfe [GR_Parameter_X] = NEXTAFTER_save_f8 // STORE Parameter 1 on stack
469 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
473 stfe [GR_Parameter_Y] = f8 // STORE Parameter 3 on stack
474 add GR_Parameter_Y = -16,GR_Parameter_Y
475 br.call.sptk b0=__libm_error_support# // Call error handling function
480 add GR_Parameter_RESULT = 48,sp
485 ldfe f8 = [GR_Parameter_RESULT] // Get return result off stack
487 add sp = 64,sp // Restore stack pointer
488 mov b0 = GR_SAVE_B0 // Restore return address
491 mov gp = GR_SAVE_GP // Restore gp
492 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
493 br.ret.sptk b0 // Return
496 LOCAL_LIBM_END(__libm_error_region)
499 .type __libm_error_support#,@function
500 .global __libm_error_support#