.file "rint.s"
-// Copyright (C) 2000, 2001, Intel Corporation
+
+// Copyright (c) 2000 - 2003, Intel Corporation
// All rights reserved.
-//
-// Contributed 2/2/2000 by John Harrison, Ted Kubaska, Bob Norin, Shane Story,
-// and Ping Tak Peter Tang of the Computational Software Lab, Intel Corporation.
+//
+// Contributed 2000 by the Intel Numerics Group, Intel Corporation
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// * The name of Intel Corporation may not be used to endorse or promote
// products derived from this software without specific prior written
// permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
-// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
-// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
-// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
// OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
-// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
// Intel Corporation is the author of this code, and requests that all
-// problem reports or change requests be submitted to it directly at
-// http://developer.intel.com/opensource.
+// problem reports or change requests be submitted to it directly at
+// http://www.intel.com/software/products/opensource/libraries/num.htm.
//
// History
//==============================================================
-// 2/02/00: Initial version
-// 2/08/01 Corrected behavior for all rounding modes.
-//
+// 02/02/00 Initial version
+// 02/08/01 Corrected behavior for all rounding modes.
+// 05/20/02 Cleaned up namespace and sf0 syntax
+// 01/20/03 Improved performance
+//==============================================================
+
// API
//==============================================================
// double rint(double x)
+//==============================================================
-#include "libm_support.h"
-
-//
-// general registers used:
-//
-rint_GR_FFFF = r14
-rint_GR_signexp = r15
-rint_GR_exponent = r16
-rint_GR_17ones = r17
-rint_GR_10033 = r18
-rint_GR_fpsr = r19
-rint_GR_rcs0 = r20
-rint_GR_rcs0_mask = r21
+// general input registers:
+// r14 - r21
+rSignexp = r14
+rExp = r15
+rExpMask = r16
+rBigexp = r17
+rM1 = r18
+rFpsr = r19
+rRcs0 = r20
+rRcs0Mask = r21
-// predicate registers used:
-// p6-11
+// floating-point registers:
+// f8 - f11
-// floating-point registers used:
+fXInt = f9
+fNormX = f10
+fTmp = f11
-RINT_NORM_f8 = f9
-RINT_FFFF = f10
-RINT_INEXACT = f11
-RINT_FLOAT_INT_f8 = f12
-RINT_INT_f8 = f13
+// predicate registers used:
+// p6 - p10
// Overview of operation
//==============================================================
-
// double rint(double x)
-// Return an integer value (represented as a double) that is x rounded to integer in current
-// rounding mode
+// Return an integer value (represented as a double) that is x
+// rounded to integer in current rounding mode
// Inexact is set if x != rint(x)
-// *******************************************************************************
-
-// Set denormal flag for denormal input and
-// and take denormal fault if necessary.
-
-// Is the input an integer value already?
+//==============================================================
// double_extended
-// if the exponent is >= 1003e => 3F(true) = 63(decimal)
+// if the exponent is > 1003e => 3F(true) = 63(decimal)
// we have a significand of 64 bits 1.63-bits.
// If we multiply by 2^63, we no longer have a fractional part
// So input is an integer value already.
// So input is an integer value already.
// single
-// if the exponent is >= 10016 => 17(true) = 23(decimal)
-// we have a significand of 53 bits 1.52-bits. (implicit 1)
-// If we multiply by 2^52, we no longer have a fractional part
+// if the exponent is > 10016 => 17(true) = 23(decimal)
+// we have a significand of 24 bits 1.23-bits. (implicit 1)
+// If we multiply by 2^23, we no longer have a fractional part
// So input is an integer value already.
-// If x is NAN, ZERO, or INFINITY, then return
-
-// qnan snan inf norm unorm 0 -+
-// 1 1 1 0 0 1 11 0xe7
-
-
-.align 32
-.global rint#
-
.section .text
-.proc rint#
-.align 32
-
-
-rint:
-#ifdef _LIBC
-.global __rint
-.type __rint,@function
-__rint:
-#endif
+GLOBAL_IEEE754_ENTRY(rint)
{ .mfi
- mov rint_GR_fpsr = ar40 // Read the fpsr--need to check rc.s0
- fcvt.fx.s1 RINT_INT_f8 = f8
- addl rint_GR_10033 = 0x10033, r0
+ getf.exp rSignexp = f8 // Get signexp, recompute if unorm
+ fclass.m p7,p0 = f8, 0x0b // Test x unorm
+ addl rBigexp = 0x10033, r0 // Set exponent at which is integer
}
{ .mfi
- mov rint_GR_FFFF = -1
- fnorm.s1 RINT_NORM_f8 = f8
- mov rint_GR_17ones = 0x1FFFF
-;;
+ mov rM1 = -1 // Set all ones
+ fcvt.fx.s1 fXInt = f8 // Convert to int in significand
+ mov rExpMask = 0x1FFFF // Form exponent mask
}
+;;
{ .mfi
- setf.sig RINT_FFFF = rint_GR_FFFF
- fclass.m.unc p6,p0 = f8, 0xe7
- mov rint_GR_rcs0_mask = 0x0c00
-;;
+ mov rFpsr = ar40 // Read fpsr -- check rc.s0
+ fclass.m p6,p0 = f8, 0x1e3 // Test x natval, nan, inf
+ nop.i 0
}
-
{ .mfb
- nop.m 999
-(p6) fnorm.d f8 = f8
-(p6) br.ret.spnt b0 // Exit if x nan, inf, zero
-;;
+ setf.sig fTmp = rM1 // Make const for setting inexact
+ fnorm.s1 fNormX = f8 // Normalize input
+(p7) br.cond.spnt RINT_UNORM // Branch if x unorm
}
-
-{ .mfi
- nop.m 999
- fcvt.xf RINT_FLOAT_INT_f8 = RINT_INT_f8
- nop.i 999
;;
+
+
+RINT_COMMON:
+// Return here from RINT_UNORM
+{ .mfb
+ and rExp = rSignexp, rExpMask // Get biased exponent
+(p6) fma.d.s0 f8 = f8, f1, f0 // Result if x natval, nan, inf
+(p6) br.ret.spnt b0 // Exit if x natval, nan, inf
}
+;;
{ .mfi
- getf.exp rint_GR_signexp = RINT_NORM_f8
- fcmp.eq.s0 p8,p0 = f8,f0 // Dummy op to set denormal
- nop.i 999
-;;
+ mov rRcs0Mask = 0x0c00 // Mask for rc.s0
+ fcvt.xf f8 = fXInt // Result assume |x| < 2^52
+ cmp.ge p7,p8 = rExp, rBigexp // Is |x| >= 2^52?
}
-
-
-{ .mii
- nop.m 999
- nop.i 999
- and rint_GR_exponent = rint_GR_signexp, rint_GR_17ones
;;
-}
-{ .mmi
- cmp.ge.unc p7,p6 = rint_GR_exponent, rint_GR_10033
- and rint_GR_rcs0 = rint_GR_rcs0_mask, rint_GR_fpsr
- nop.i 999
-;;
+// We must correct result if |x| >= 2^52
+{ .mfi
+ nop.m 0
+(p7) fma.d.s0 f8 = fNormX, f1, f0 // If |x| >= 2^52, result x
+ nop.i 0
}
-
-// Check to see if s0 rounding mode is round to nearest. If not then set s2
-// rounding mode to that of s0 and repeat conversions.
-L(RINT_COMMON):
-{ .mfb
- cmp.ne p11,p0 = rint_GR_rcs0, r0
-(p6) fclass.m.unc p9,p10 = RINT_FLOAT_INT_f8, 0x07 // Test for result=0
-(p11) br.cond.spnt L(RINT_NOT_ROUND_NEAREST) // Branch if not round to nearest
;;
-}
{ .mfi
- nop.m 999
-(p6) fcmp.eq.unc.s1 p0,p8 = RINT_FLOAT_INT_f8, RINT_NORM_f8
- nop.i 999
+ nop.m 0
+ fcmp.eq.unc.s1 p0, p9 = f8, fNormX // Is result = x ?
+ nop.i 0
}
{ .mfi
- nop.m 999
-(p7) fnorm.d.s0 f8 = f8
- nop.i 999
-;;
+ nop.m 0
+(p8) fmerge.s f8 = fNormX, f8 // Make sure sign rint(x) = sign x
+ nop.i 0
}
+;;
-// If result is zero, merge sign of input
{ .mfi
- nop.m 999
-(p9) fmerge.s f8 = f8, RINT_FLOAT_INT_f8
- nop.i 999
+(p8) and rRcs0 = rFpsr, rRcs0Mask // Get rounding mode for sf0
+ nop.f 0
+ nop.i 0
}
-{ .mfi
- nop.m 999
-(p10) fnorm.d f8 = RINT_FLOAT_INT_f8
- nop.i 999
;;
+
+// If |x| < 2^52 we must test for other rounding modes
+{ .mfi
+(p8) cmp.ne.unc p10,p0 = rRcs0, r0 // Test for other rounding modes
+(p9) fmpy.s0 fTmp = fTmp, fTmp // Dummy to set inexact
+ nop.i 0
+}
+{ .mbb
+ nop.m 0
+(p10) br.cond.spnt RINT_NOT_ROUND_NEAREST // Branch if not round nearest
+ br.ret.sptk b0 // Exit main path if round nearest
}
+;;
+
+
+RINT_UNORM:
+// Here if x unorm
{ .mfb
- nop.m 999
-(p8) fmpy.s0 RINT_INEXACT = RINT_FFFF,RINT_FFFF // Dummy to set inexact
- br.ret.sptk b0
-;;
+ getf.exp rSignexp = fNormX // Get signexp, recompute if unorm
+ fcmp.eq.s0 p7,p0 = f8, f0 // Dummy op to set denormal flag
+ br.cond.sptk RINT_COMMON // Return to main path
}
+;;
-L(RINT_NOT_ROUND_NEAREST):
-// Set rounding mode of s2 to that of s0
+RINT_NOT_ROUND_NEAREST:
+// Here if not round to nearest, and |x| < 2^52
+// Set rounding mode of s2 to that of s0, and repeat the conversion using s2
{ .mfi
- mov rint_GR_rcs0 = r0 // Clear so we don't come back here
- fsetc.s2 0x7f, 0x40
- nop.i 999
-;;
+ nop.m 0
+ fsetc.s2 0x7f, 0x40
+ nop.i 0
}
+;;
{ .mfi
- nop.m 999
- fcvt.fx.s2 RINT_INT_f8 = f8
- nop.i 999
+ nop.m 0
+ fcvt.fx.s2 fXInt = fNormX // Convert to int in significand
+ nop.i 0
+}
;;
+
+{ .mfi
+ nop.m 0
+ fcvt.xf f8 = fXInt // Expected result
+ nop.i 0
}
+;;
+// Be sure sign of result = sign of input. Fixes cases where result is 0.
{ .mfb
- nop.m 999
- fcvt.xf RINT_FLOAT_INT_f8 = RINT_INT_f8
- br.cond.sptk L(RINT_COMMON)
-;;
+ nop.m 0
+ fmerge.s f8 = fNormX, f8
+ br.ret.sptk b0 // Exit main path
}
+;;
-
-.endp rint
-ASM_SIZE_DIRECTIVE(rint)
-#ifdef _LIBC
-ASM_SIZE_DIRECTIVE(__rint)
-#endif
+GLOBAL_IEEE754_END(rint)
git.samba.org / jlayton / glibc.git / blobdiff