.file "floorf.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.
//
-.align 32
-.global floorf#
-
-.section .text
-.proc floorf#
-.align 32
-
// History
//==============================================================
-// 2/02/00: Initial version
-// 6/13/00: Improved speed
-// 6/27/00: Eliminated incorrect invalid flag setting
-// 2/07/01: Corrected sign of zero result in round to -inf mode
+// 02/02/00 Initial version
+// 06/13/00 Improved speed
+// 06/27/00 Eliminated incorrect invalid flag setting
+// 02/07/01 Corrected sign of zero result in round to -inf mode
+// 05/20/02 Cleaned up namespace and sf0 syntax
+// 01/28/03 Improved performance
+//==============================================================
// API
//==============================================================
// float floorf(float x)
+//==============================================================
-// general input registers:
-
-floor_GR_FFFF = r14
-floor_GR_signexp = r15
-floor_GR_exponent = r16
-floor_GR_expmask = r17
-floor_GR_bigexp = r18
-
-
-// predicate registers used:
+// general input registers:
+// r14 - r18
-// p6 ==> Input is NaN, infinity, zero
-// p7 ==> Input is denormal
-// p8 ==> Input is <0
-// p9 ==> Input is >=0
-// p10 ==> Input is already an integer (bigger than largest integer)
-// p11 ==> Input is not a large integer
-// p12 ==> Input is a smaller integer
-// p13 ==> Input is not an even integer, so inexact must be set
+rSignexp = r14
+rExp = r15
+rExpMask = r16
+rBigexp = r17
+rM1 = r18
+// floating-point registers:
+// f8 - f13
-// floating-point registers used:
+fXInt = f9
+fNormX = f10
+fTmp = f11
+fAdj = f12
+fPreResult = f13
-FLOOR_NORM_f8 = f9
-FLOOR_FFFF = f10
-FLOOR_INEXACT = f11
-FLOOR_FLOAT_INT_f8 = f12
-FLOOR_INT_f8 = f13
-FLOOR_adj = f14
+// predicate registers used:
+// p6 - p9
// Overview of operation
//==============================================================
-
// float floorf(float x)
-// Return an integer value (represented as a float) that is the largest
+// Return an integer value (represented as a float) that is the largest
// value not greater than x
// This is x rounded toward -infinity to an integral value.
// Inexact is set if x != floorf(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 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
-
-#include "libm_support.h"
-floorf:
-#ifdef _LIBC
-.global __floorf
-__floorf:
-#endif
+.section .text
+GLOBAL_IEEE754_ENTRY(floorf)
{ .mfi
- getf.exp floor_GR_signexp = f8
- fcvt.fx.trunc.s1 FLOOR_INT_f8 = f8
- addl floor_GR_bigexp = 0x10016, r0
+ getf.exp rSignexp = f8 // Get signexp, recompute if unorm
+ fclass.m p7,p0 = f8, 0x0b // Test x unorm
+ addl rBigexp = 0x10016, r0 // Set exponent at which is integer
}
{ .mfi
- addl floor_GR_FFFF = -1,r0
- fcmp.lt.s1 p8,p9 = f8,f0
- mov floor_GR_expmask = 0x1FFFF ;;
+ mov rM1 = -1 // Set all ones
+ fcvt.fx.trunc.s1 fXInt = f8 // Convert to int in significand
+ mov rExpMask = 0x1FFFF // Form exponent mask
}
+;;
-// p7 ==> denorm
{ .mfi
- setf.sig FLOOR_FFFF = floor_GR_FFFF
- fclass.m p7,p0 = f8, 0x0b
- nop.i 999
+ nop.m 0
+ fcmp.lt.s1 p8,p9 = f8, f0 // Test x < 0
+ nop.i 0
}
-{ .mfi
- nop.m 999
- fnorm.s1 FLOOR_NORM_f8 = f8
- nop.i 999 ;;
+{ .mfb
+ setf.sig fTmp = rM1 // Make const for setting inexact
+ fnorm.s1 fNormX = f8 // Normalize input
+(p7) br.cond.spnt FLOOR_UNORM // Branch if x unorm
}
+;;
-// p6 ==> NAN, INF, ZERO
-{ .mfb
- nop.m 999
- fclass.m p6,p10 = f8, 0xe7
-(p7) br.cond.spnt L(FLOOR_DENORM) ;;
+FLOOR_COMMON:
+// Return here from FLOOR_UNORM
+{ .mfi
+ nop.m 0
+ fclass.m p6,p0 = f8, 0x1e7 // Test x natval, nan, inf, 0
+ nop.i 0
}
+;;
-L(FLOOR_COMMON):
.pred.rel "mutex",p8,p9
-// Set adjustment to subtract from trunc(x) for result
-// If x<0, adjustment is -1.0
-// If x>=0, adjustment is 0.0
{ .mfi
- and floor_GR_exponent = floor_GR_signexp, floor_GR_expmask
-(p8) fnma.s1 FLOOR_adj = f1,f1,f0
- nop.i 999
+ nop.m 0
+(p8) fnma.s1 fAdj = f1, f1, f0 // If x < 0, adjustment is -1
+ nop.i 0
}
{ .mfi
- nop.m 999
-(p9) fadd.s1 FLOOR_adj = f0,f0
- nop.i 999 ;;
+ nop.m 0
+(p9) fma.s1 fAdj = f0, f0, f0 // If x > 0, adjustment is 0
+ nop.i 0
}
+;;
{ .mfi
- nop.m 999
- fcmp.eq.s0 p12,p0 = f8,f0 // Dummy op to set denormal and invalid flag
- nop.i 999
+ nop.m 0
+ fcvt.xf fPreResult = fXInt // trunc(x)
+ nop.i 0
}
-{ .mfi
-(p10) cmp.ge.unc p10,p11 = floor_GR_exponent, floor_GR_bigexp
-(p6) fnorm.s f8 = f8
- nop.i 999 ;;
+{ .mfb
+ nop.m 0
+(p6) fma.s.s0 f8 = f8, f1, f0 // Result if x natval, nan, inf, 0
+(p6) br.ret.spnt b0 // Exit if x natval, nan, inf, 0
}
+;;
-{ .mfi
- nop.m 999
-(p11) fcvt.xf FLOOR_FLOAT_INT_f8 = FLOOR_INT_f8
- nop.i 999 ;;
+{ .mmi
+ and rExp = rSignexp, rExpMask // Get biased exponent
+;;
+ cmp.ge p7,p6 = rExp, rBigexp // Is |x| >= 2^23?
+ nop.i 0
}
+;;
{ .mfi
- nop.m 999
-(p10) fnorm.s f8 = FLOOR_NORM_f8
- nop.i 999 ;;
+ nop.m 0
+(p6) fma.s.s0 f8 = fPreResult, f1, fAdj // Result if !int, |x| < 2^23
+ nop.i 0
}
-
{ .mfi
- nop.m 999
-(p11) fadd.s f8 = FLOOR_FLOAT_INT_f8,FLOOR_adj
- nop.i 999 ;;
+ nop.m 0
+(p7) fma.s.s0 f8 = fNormX, f1, f0 // Result, if |x| >= 2^23
+ nop.i 0
}
+;;
+
{ .mfi
- nop.m 999
-(p11) fcmp.eq.unc.s1 p12,p13 = FLOOR_FLOAT_INT_f8, FLOOR_NORM_f8
- nop.i 999 ;;
+ nop.m 0
+(p6) fcmp.eq.unc.s1 p8, p9 = fPreResult, fNormX // Is trunc(x) = x ?
+ nop.i 0
}
+;;
-// Set inexact if result not equal to input
{ .mfi
- nop.m 999
-(p13) fmpy.s0 FLOOR_INEXACT = FLOOR_FFFF,FLOOR_FFFF
- nop.i 999
+ nop.m 0
+(p9) fmpy.s0 fTmp = fTmp, fTmp // Dummy to set inexact
+ nop.i 0
}
-// Set result to input if integer
{ .mfb
- nop.m 999
-(p12) fnorm.s f8 = FLOOR_NORM_f8
- br.ret.sptk b0 ;;
+ nop.m 0
+(p8) fma.s.s0 f8 = fNormX, f1, f0 // If x int, result normalized x
+ br.ret.sptk b0 // Exit main path, 0 < |x| < 2^23
}
+;;
+
-// Here if input denorm
-L(FLOOR_DENORM):
+FLOOR_UNORM:
+// Here if x unorm
{ .mfb
- getf.exp floor_GR_signexp = FLOOR_NORM_f8
- fcvt.fx.trunc.s1 FLOOR_INT_f8 = FLOOR_NORM_f8
- br.cond.sptk L(FLOOR_COMMON) ;;
+ 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 FLOOR_COMMON // Return to main path
}
+;;
-.endp floorf
-ASM_SIZE_DIRECTIVE(floorf)
+GLOBAL_IEEE754_END(floorf)
git.samba.org / jlayton / glibc.git / blobdiff