3 // Copyright (c) 2000, 2001, Intel Corporation
4 // All rights reserved.
6 // Contributed 2/2/2000 by John Harrison, Ted Kubaska, Bob Norin, Shane Story,
7 // and Ping Tak Peter Tang of the Computational Software Lab, Intel Corporation.
11 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
12 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
13 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
14 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
15 // CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
16 // EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
17 // PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
18 // PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
19 // OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
20 // NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
21 // SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
23 // Intel Corporation is the author of this code, and requests that all
24 // problem reports or change requests be submitted to it directly at
25 // http://developer.intel.com/opensource.
28 //==============================================================
29 // 2/02/00: Initial version
30 // 2/08/01 Corrected behavior for all rounding modes.
33 //==============================================================
34 // double rint(double x)
36 #include "libm_support.h"
39 // general registers used:
43 rint_GR_exponent = r16
48 rint_GR_rcs0_mask = r21
51 // predicate registers used:
54 // floating-point registers used:
59 RINT_FLOAT_INT_f8 = f12
62 // Overview of operation
63 //==============================================================
65 // double rint(double x)
66 // Return an integer value (represented as a double) that is x rounded to integer in current
68 // Inexact is set if x != rint(x)
69 // *******************************************************************************
71 // Set denormal flag for denormal input and
72 // and take denormal fault if necessary.
74 // Is the input an integer value already?
77 // if the exponent is >= 1003e => 3F(true) = 63(decimal)
78 // we have a significand of 64 bits 1.63-bits.
79 // If we multiply by 2^63, we no longer have a fractional part
80 // So input is an integer value already.
83 // if the exponent is >= 10033 => 34(true) = 52(decimal)
85 // we have a significand of 53 bits 1.52-bits. (implicit 1)
86 // If we multiply by 2^52, we no longer have a fractional part
87 // So input is an integer value already.
90 // if the exponent is >= 10016 => 17(true) = 23(decimal)
91 // we have a significand of 53 bits 1.52-bits. (implicit 1)
92 // If we multiply by 2^52, we no longer have a fractional part
93 // So input is an integer value already.
95 // If x is NAN, ZERO, or INFINITY, then return
97 // qnan snan inf norm unorm 0 -+
98 // 1 1 1 0 0 1 11 0xe7
112 .type __rint,@function
117 mov rint_GR_fpsr = ar40 // Read the fpsr--need to check rc.s0
118 fcvt.fx.s1 RINT_INT_f8 = f8
119 addl rint_GR_10033 = 0x10033, r0
122 mov rint_GR_FFFF = -1
123 fnorm.s1 RINT_NORM_f8 = f8
124 mov rint_GR_17ones = 0x1FFFF
129 setf.sig RINT_FFFF = rint_GR_FFFF
130 fclass.m.unc p6,p0 = f8, 0xe7
131 mov rint_GR_rcs0_mask = 0x0c00
138 (p6) br.ret.spnt b0 // Exit if x nan, inf, zero
144 fcvt.xf RINT_FLOAT_INT_f8 = RINT_INT_f8
150 getf.exp rint_GR_signexp = RINT_NORM_f8
151 fcmp.eq.s0 p8,p0 = f8,f0 // Dummy op to set denormal
160 and rint_GR_exponent = rint_GR_signexp, rint_GR_17ones
165 cmp.ge.unc p7,p6 = rint_GR_exponent, rint_GR_10033
166 and rint_GR_rcs0 = rint_GR_rcs0_mask, rint_GR_fpsr
171 // Check to see if s0 rounding mode is round to nearest. If not then set s2
172 // rounding mode to that of s0 and repeat conversions.
175 cmp.ne p11,p0 = rint_GR_rcs0, r0
176 (p6) fclass.m.unc p9,p10 = RINT_FLOAT_INT_f8, 0x07 // Test for result=0
177 (p11) br.cond.spnt L(RINT_NOT_ROUND_NEAREST) // Branch if not round to nearest
183 (p6) fcmp.eq.unc.s1 p0,p8 = RINT_FLOAT_INT_f8, RINT_NORM_f8
188 (p7) fnorm.d.s0 f8 = f8
193 // If result is zero, merge sign of input
196 (p9) fmerge.s f8 = f8, RINT_FLOAT_INT_f8
201 (p10) fnorm.d f8 = RINT_FLOAT_INT_f8
208 (p8) fmpy.s0 RINT_INEXACT = RINT_FFFF,RINT_FFFF // Dummy to set inexact
213 L(RINT_NOT_ROUND_NEAREST):
214 // Set rounding mode of s2 to that of s0
216 mov rint_GR_rcs0 = r0 // Clear so we don't come back here
224 fcvt.fx.s2 RINT_INT_f8 = f8
231 fcvt.xf RINT_FLOAT_INT_f8 = RINT_INT_f8
232 br.cond.sptk L(RINT_COMMON)
238 ASM_SIZE_DIRECTIVE(rint)
240 ASM_SIZE_DIRECTIVE(__rint)