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 Corrected behavior when both args are zero to conform to C99, and
51 // fixed flag settings for several cases
52 // 05/20/02 Cleaned up namespace and sf0 syntax
53 // 02/10/03 Reordered header: .section, .global, .proc, .align
56 //==============================================================
57 // float nextafterf( float x, float y );
58 // input floating point f8, f9
59 // output floating point f8
62 //==============================================================
63 nextafter_GR_max_pexp = r14
64 nextafter_GR_min_pexp = r15
65 nextafter_GR_exp = r16
66 nextafter_GR_sig = r17
67 nextafter_GR_lnorm_sig = r18
68 nextafter_GR_sign_mask = r19
69 nextafter_GR_exp_mask = r20
70 nextafter_GR_sden_sig = r21
71 nextafter_GR_new_sig = r22
72 nextafter_GR_new_exp = r23
73 nextafter_GR_lden_sig = r24
74 nextafter_GR_snorm_sig = r25
75 nextafter_GR_exp1 = r26
76 nextafter_GR_x_exp = r27
77 nextafter_GR_min_den_rexp = r28
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_sden_exp = f14
94 NEXTAFTER_save_f8 = f33
95 NEXTAFTER_new_exp = f34
96 NEXTAFTER_new_sig = f35
97 NEXTAFTER_lden_sig = f36
98 NEXTAFTER_snorm_sig = f37
103 // Overview of operation
104 //==============================================================
105 // nextafterf determines the next representable value
106 // after x in the direction of y.
110 GLOBAL_LIBM_ENTRY(nextafterf)
112 // Extract signexp from x
113 // Form smallest denormal significand = ulp size
115 getf.exp nextafter_GR_exp = f8
116 movl nextafter_GR_sden_sig = 0x0000010000000000
118 // Form largest normal exponent
119 // Is x < y ? p10 if yes, p11 if no
120 // Form smallest normal exponent
122 addl nextafter_GR_max_pexp = 0x1007e, r0
123 fcmp.lt.s1 p10,p11 = f8, f9
124 addl nextafter_GR_min_pexp = 0x0ff81, r0 ;;
129 getf.sig nextafter_GR_sig = f8
130 fcmp.eq.s0 p6,p0 = f8, f9
133 // Extract significand from x
134 // Form largest normal significand
137 movl nextafter_GR_lnorm_sig = 0xffffff0000000000 ;;
140 // Move largest normal significand to fp reg for special cases
142 setf.sig NEXTAFTER_lnorm_sig = nextafter_GR_lnorm_sig
144 addl nextafter_GR_sign_mask = 0x20000, r0 ;;
147 // Move smallest denormal significand and signexp to fp regs
149 // Set p12 and p13 based on whether significand increases or decreases
150 // It increases (p12 set) if x<y and x>=0 or if x>y and x<0
151 // It decreases (p13 set) if x<y and x<0 or if x>y and x>=0
153 setf.sig NEXTAFTER_sden_sig = nextafter_GR_sden_sig
154 fclass.m p8,p0 = f8, 0xc3
155 (p10) cmp.lt p12,p13 = nextafter_GR_exp, nextafter_GR_sign_mask
158 setf.exp NEXTAFTER_sden_exp = nextafter_GR_min_pexp
160 (p11) cmp.ge p12,p13 = nextafter_GR_exp, nextafter_GR_sign_mask ;;
163 .pred.rel "mutex",p12,p13
165 // Form expected new significand, adding or subtracting 1 ulp increment
166 // If x=y set result to y
167 // Form smallest normal significand and largest denormal significand
169 (p12) add nextafter_GR_new_sig = nextafter_GR_sig, nextafter_GR_sden_sig
170 (p6) fmerge.s f8=f9,f9
171 dep.z nextafter_GR_snorm_sig = 1,63,1 // 0x8000000000000000
174 (p13) sub nextafter_GR_new_sig = nextafter_GR_sig, nextafter_GR_sden_sig
175 movl nextafter_GR_lden_sig = 0x7fffff0000000000 ;;
178 // Move expected result significand and signexp to fp regs
180 // Form new exponent in case result exponent needs incrementing or decrementing
182 setf.exp NEXTAFTER_new_exp = nextafter_GR_exp
183 fclass.m p9,p0 = f9, 0xc3
184 (p12) add nextafter_GR_exp1 = 1, nextafter_GR_exp
187 setf.sig NEXTAFTER_new_sig = nextafter_GR_new_sig
188 (p13) add nextafter_GR_exp1 = -1, nextafter_GR_exp
189 (p6) br.ret.spnt b0 ;; // Exit if x=y
192 // Move largest normal signexp to fp reg for special cases
195 setf.exp NEXTAFTER_lnorm_exp = nextafter_GR_max_pexp
196 fclass.m p7,p0 = f8, 0x7
201 (p8) fma.s0 f8 = f8,f1,f9
202 (p8) br.ret.spnt b0 ;; // Exit if x=nan
205 // Move exp+-1 and smallest normal significand to fp regs for special cases
208 setf.exp NEXTAFTER_exp1 = nextafter_GR_exp1
209 fclass.m p6,p0 = f8, 0x23
210 addl nextafter_GR_exp_mask = 0x1ffff, r0
213 setf.sig NEXTAFTER_snorm_sig = nextafter_GR_snorm_sig
214 (p9) fma.s0 f8 = f8,f1,f9
215 (p9) br.ret.spnt b0 ;; // Exit if y=nan
218 // Move largest denormal significand to fp regs for special cases
221 setf.sig NEXTAFTER_lden_sig = nextafter_GR_lden_sig
222 mov NEXTAFTER_save_f8 = f8
223 (p7) br.cond.spnt NEXTAFTER_ZERO ;; // Exit if x=0
226 // Mask off the sign to get x_exp
228 and nextafter_GR_x_exp = nextafter_GR_exp_mask, nextafter_GR_exp
230 (p6) br.cond.spnt NEXTAFTER_INF ;; // Exit if x=inf
233 // Check 6 special cases when significand rolls over:
234 // 1 sig size incr, x_sig=max_sig, x_exp < max_exp
235 // Set p6, result is sig=min_sig, exp++
236 // 2 sig size incr, x_sig=max_sig, x_exp >= max_exp
237 // Set p7, result is inf, signal overflow
238 // 3 sig size decr, x_sig=min_sig, x_exp > min_exp
239 // Set p8, result is sig=max_sig, exp--
240 // 4 sig size decr, x_sig=min_sig, x_exp = min_exp
241 // Set p9, result is sig=max_den_sig, exp same, signal underflow and inexact
242 // 5 sig size decr, x_sig=min_den_sig, x_exp = min_exp
243 // Set p10, result is zero, sign of x, signal underflow and inexact
244 // 6 sig size decr, x_sig=min_sig, x_exp < min_exp
245 // Set p14, result is zero, sign of x, signal underflow and inexact
247 // Form exponent of smallest float denormal (if normalized register format)
249 adds nextafter_GR_min_den_rexp = -23, nextafter_GR_min_pexp
250 (p12) cmp.eq.unc p6,p0 = nextafter_GR_new_sig, r0
251 (p13) cmp.eq.unc p8,p10 = nextafter_GR_new_sig, nextafter_GR_lden_sig ;;
255 (p6) cmp.lt.unc p6,p7 = nextafter_GR_x_exp, nextafter_GR_max_pexp
256 (p8) cmp.gt.unc p8,p9 = nextafter_GR_x_exp, nextafter_GR_min_pexp
257 (p10) cmp.eq.unc p10,p0 = nextafter_GR_new_sig, r0 ;;
260 // Create small normal in case need to generate underflow flag
262 (p10) cmp.le.unc p10,p0 = nextafter_GR_x_exp, nextafter_GR_min_pexp
263 fmerge.se NEXTAFTER_tmp = NEXTAFTER_sden_exp, NEXTAFTER_lnorm_sig
264 (p9) cmp.gt.unc p9,p14 = nextafter_GR_x_exp, nextafter_GR_min_den_rexp
266 // Branch if cases 1, 2, 3
268 (p6) br.cond.spnt NEXTAFTER_EXPUP
269 (p7) br.cond.spnt NEXTAFTER_OVERFLOW
270 (p8) br.cond.spnt NEXTAFTER_EXPDOWN ;;
273 // Branch if cases 4, 5, 6
275 (p9) br.cond.spnt NEXTAFTER_NORM_TO_DENORM
276 (p10) br.cond.spnt NEXTAFTER_UNDERFLOW_TO_ZERO
277 (p14) br.cond.spnt NEXTAFTER_UNDERFLOW_TO_ZERO ;;
280 // Here if no special cases
281 // Set p6 if result will be a denormal, so can force underflow flag
282 // Case 1: x_exp=min_exp, x_sig=unnormalized
283 // Case 2: x_exp<min_exp
285 cmp.lt p6,p7 = nextafter_GR_x_exp, nextafter_GR_min_pexp
286 fmerge.se f8 = NEXTAFTER_new_exp, NEXTAFTER_new_sig
293 (p7) tbit.z p6,p0 = nextafter_GR_new_sig, 63 ;;
296 NEXTAFTER_COMMON_FINISH:
297 // Force underflow and inexact if denormal result
300 (p6) fma.s.s0 NEXTAFTER_tmp = NEXTAFTER_tmp,NEXTAFTER_tmp,f0
304 // Final normalization to result precision and exit
314 cmp.lt p6,p7 = nextafter_GR_x_exp, nextafter_GR_min_pexp
315 fmerge.se f8 = NEXTAFTER_exp1, NEXTAFTER_snorm_sig
316 br.cond.sptk NEXTAFTER_COMMON_FINISH ;;
321 cmp.lt p6,p7 = nextafter_GR_x_exp, nextafter_GR_min_pexp
322 fmerge.se f8 = NEXTAFTER_exp1, NEXTAFTER_lnorm_sig
323 br.cond.sptk NEXTAFTER_COMMON_FINISH ;;
326 NEXTAFTER_NORM_TO_DENORM:
329 fmerge.se f8 = NEXTAFTER_new_exp, NEXTAFTER_lden_sig
332 // Force underflow and inexact
335 fma.s.s0 NEXTAFTER_tmp = NEXTAFTER_tmp,NEXTAFTER_tmp,f0
339 NEXTAFTER_UNDERFLOW_TO_ZERO:
342 fmerge.s f8 = NEXTAFTER_save_f8,f0
343 br.cond.sptk NEXTAFTER_COMMON_FINISH ;;
347 // Here if f8 is +- infinity
349 // if f8 is +inf, no matter what y is return largest float
350 // if f8 is -inf, no matter what y is return -largest float
354 fmerge.se NEXTAFTER_lnorm = NEXTAFTER_lnorm_exp,NEXTAFTER_lnorm_sig
360 fmerge.s f8 = f8,NEXTAFTER_lnorm
366 // Here if f8 is +- zero
368 // if f8 is zero and y is +, return + smallest float denormal
369 // if f8 is zero and y is -, return - smallest float denormal
373 fmerge.se NEXTAFTER_sden = NEXTAFTER_sden_exp,NEXTAFTER_sden_sig
377 // Create small normal to generate underflow flag
380 fmerge.se NEXTAFTER_tmp = NEXTAFTER_sden_exp, NEXTAFTER_lnorm_sig
384 // Add correct sign from direction arg
387 fmerge.s f8 = f9,NEXTAFTER_sden
391 // Force underflow and inexact flags
394 fma.s.s0 NEXTAFTER_tmp = NEXTAFTER_tmp,NEXTAFTER_tmp,f0
398 GLOBAL_LIBM_END(nextafterf)
399 // Stack operations when calling error support.
400 // (1) (2) (3) (call) (4)
401 // sp -> + psp -> + psp -> + sp -> +
403 // | | <- GR_Y R3 ->| <- GR_RESULT | -> f8
405 // | <-GR_Y Y2->| Y2 ->| <- GR_Y |
407 // | | <- GR_X X1 ->| |
409 // sp-64 -> + sp -> + sp -> + +
410 // save ar.pfs save b0 restore gp
411 // save gp restore ar.pfs
415 LOCAL_LIBM_ENTRY(__libm_error_region)
417 // Here if f8 is finite, but result will be infinite
418 // Use frcpa to generate infinity of correct sign
419 // Call error support to report possible range error
423 alloc r32=ar.pfs,2,2,4,0
424 frcpa.s1 f8,p6 = NEXTAFTER_save_f8, f0
428 // Create largest float
431 fmerge.se NEXTAFTER_lnorm = NEXTAFTER_lnorm_exp,NEXTAFTER_lnorm_sig
435 // Force overflow and inexact flags to be set
437 mov r39 = 155 // Error code
438 fma.s.s0 NEXTAFTER_tmp = NEXTAFTER_lnorm,NEXTAFTER_lnorm,f0
445 add GR_Parameter_Y=-32,sp // Parameter 2 value
447 .save ar.pfs,GR_SAVE_PFS
448 mov GR_SAVE_PFS=ar.pfs // Save ar.pfs
452 add sp=-64,sp // Create new stack
454 mov GR_SAVE_GP=gp // Save gp
460 stfs [GR_Parameter_Y] = f9,16 // STORE Parameter 2 on stack
461 add GR_Parameter_X = 16,sp // Parameter 1 address
463 mov GR_SAVE_B0=b0 // Save b0
469 stfs [GR_Parameter_X] = NEXTAFTER_save_f8 // STORE Parameter 1 on stack
470 add GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
474 stfs [GR_Parameter_Y] = f8 // STORE Parameter 3 on stack
475 add GR_Parameter_Y = -16,GR_Parameter_Y
476 br.call.sptk b0=__libm_error_support# // Call error handling function
481 add GR_Parameter_RESULT = 48,sp
486 ldfs f8 = [GR_Parameter_RESULT] // Get return result off stack
488 add sp = 64,sp // Restore stack pointer
489 mov b0 = GR_SAVE_B0 // Restore return address
492 mov gp = GR_SAVE_GP // Restore gp
493 mov ar.pfs = GR_SAVE_PFS // Restore ar.pfs
494 br.ret.sptk b0 // Return
497 LOCAL_LIBM_END(__libm_error_region)
500 .type __libm_error_support#,@function
501 .global __libm_error_support#