diff options
Diffstat (limited to 'sysdeps/ia64/fpu/s_roundf.S')
| -rw-r--r-- | sysdeps/ia64/fpu/s_roundf.S | 316 |
1 files changed, 144 insertions, 172 deletions
diff --git a/sysdeps/ia64/fpu/s_roundf.S b/sysdeps/ia64/fpu/s_roundf.S index 42ee60b218..1e8dc78777 100644 --- a/sysdeps/ia64/fpu/s_roundf.S +++ b/sysdeps/ia64/fpu/s_roundf.S @@ -1,11 +1,10 @@ .file "roundf.s" -// Copyright (C) 2000, 2001, Intel Corporation + +// Copyright (c) 2000 - 2003, Intel Corporation // All rights reserved. -// -// Contributed 10/25/2000 by John Harrison, Cristina Iordache, Ted Kubaska, -// Bob Norin, Tom Rowan, 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 @@ -21,229 +20,202 @@ // * 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 //============================================================== -// 10/25/2000: Created +// 10/25/00 Initial version +// 06/14/01 Changed cmp to an equivalent form +// 05/20/02 Cleaned up namespace and sf0 syntax +// 01/20/03 Improved performance and reduced code size +// 04/18/03 Eliminate possible WAW dependency warning //============================================================== -// + // API //============================================================== // float roundf(float x) -// +//============================================================== -#include "libm_support.h" +// general input registers: +// r14 - r19 -// general input registers: -// -roundf_GR_half = r14 -roundf_GR_big = r15 -roundf_GR_expmask = r16 -roundf_GR_signexp = r17 -roundf_GR_exp = r18 -roundf_GR_expdiff = r19 - -// predicate registers used: -// p6 - p10 +rSignexp = r14 +rExp = r15 +rExpMask = r16 +rBigexp = r17 +rExpHalf = r18 +rExpMHalf = r19 + +// floating-point registers: +// f8 - f13 -// floating-point registers used: +fXtruncInt = f9 +fNormX = f10 +fHalf = f11 +fMHalf = f12 +fRem = f13 -ROUNDF_NORM_f8 = f9 -ROUNDF_TRUNC_f8 = f10 -ROUNDF_RINT_f8 = f11 -ROUNDF_FLOAT_TRUNC_f8 = f12 -ROUNDF_FLOAT_RINT_f8 = f13 -ROUNDF_REMAINDER = f14 -ROUNDF_HALF = f15 +// predicate registers used: +// p6 - p10 // Overview of operation //============================================================== - // float roundf(float x) -// Return an integer value (represented as a float) that is x -// rounded to nearest integer, halfway cases rounded away from -// zero. +// Return an integer value (represented as a float) that is x +// rounded to nearest integer, halfway cases rounded away from +// zero. // if x>0 result = trunc(x+0.5) // if x<0 result = trunc(x-0.5) -// ******************************************************************************* - -// Set denormal flag for denormal input and -// and take denormal fault if necessary. +// +//============================================================== -// If x is NAN, ZERO, INFINITY, or >= 2^23 then return +// double_extended +// 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. -// qnan snan inf norm unorm 0 -+ -// 1 1 1 0 0 1 11 0xe7 +// double +// if the exponent is >= 10033 => 34(true) = 52(decimal) +// 34 + 3ff = 433 +// 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 +// So input is an integer value already. +// single +// 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. -.align 32 -.global roundf# .section .text -.proc roundf# -.align 32 - +GLOBAL_LIBM_ENTRY(roundf) -roundf: - -// Get exponent for +0.5 -// Truncate x to integer { .mfi - addl roundf_GR_half = 0x0fffe, r0 - fcvt.fx.trunc.s1 ROUNDF_TRUNC_f8 = f8 - nop.i 999 -} - -// Get signexp of x -// Normalize input -// Form exponent mask -{ .mfi - getf.exp roundf_GR_signexp = f8 - fnorm ROUNDF_NORM_f8 = f8 - addl roundf_GR_expmask = 0x1ffff, r0 ;; + getf.exp rSignexp = f8 // Get signexp, recompute if unorm + fcvt.fx.trunc.s1 fXtruncInt = f8 // Convert to int in significand + addl rBigexp = 0x10016, r0 // Set exponent at which is integer } - -// Form +0.5 -// Round x to integer { .mfi - setf.exp ROUNDF_HALF = roundf_GR_half - fcvt.fx.s1 ROUNDF_RINT_f8 = f8 - nop.i 999 ;; + mov rExpHalf = 0x0FFFE // Form sign and exponent of 0.5 + fnorm.s1 fNormX = f8 // Normalize input + mov rExpMask = 0x1FFFF // Form exponent mask } -// Get exp of x -// Test for NAN, INF, ZERO -// Get exponent at which input has no fractional part -{ .mfi - and roundf_GR_exp = roundf_GR_expmask, roundf_GR_signexp - fclass.m p8,p9 = f8,0xe7 - addl roundf_GR_big = 0x10016, r0 ;; -} - -// Get exp-bigexp -// If exp is so big there is no fractional part, then turn on p8, off p9 -{ .mmi - sub roundf_GR_expdiff = roundf_GR_exp, roundf_GR_big ;; -#ifdef _LIBC -(p9) cmp.lt.or.andcm p8,p9 = r0, roundf_GR_expdiff -#else -(p9) cmp.ge.or.andcm p8,p9 = roundf_GR_expdiff, r0 -#endif - nop.i 999 ;; -} - -// Set p6 if x<0, else set p7 -{ .mfi - nop.m 999 -(p9) fcmp.lt.unc p6,p7 = f8,f0 - nop.i 999 +;; + +{ .mmf + setf.exp fHalf = rExpHalf // Form 0.5 + mov rExpMHalf = 0x2FFFE // Form sign and exponent of -0.5 + fclass.m p7,p0 = f8, 0x0b // Test x unorm } - -// If NAN, INF, ZERO, or no fractional part, result is just normalized input -{ .mfi - nop.m 999 -(p8) fnorm.s.s0 f8 = f8 - nop.i 999 ;; +;; + +{ .mfb + setf.exp fMHalf = rExpMHalf // Form -0.5 + fclass.m p6,p0 = f8, 0x1e3 // Test x natval, nan, inf +(p7) br.cond.spnt ROUND_UNORM // Branch if x unorm } +;; -// Float the truncated integer +ROUND_COMMON: +// Return here from ROUND_UNORM { .mfi - nop.m 999 -(p9) fcvt.xf ROUNDF_FLOAT_TRUNC_f8 = ROUNDF_TRUNC_f8 - nop.i 999 ;; + nop.m 0 + fcmp.lt.s1 p8,p9 = f8, f0 // Test if x < 0 + nop.i 0 +} +{ .mfb + and rExp = rSignexp, rExpMask // Get biased exponent +(p6) fma.s.s0 f8 = f8, f1, f0 // Result if x natval, nan, inf +(p6) br.ret.spnt b0 // Exit if x natval, nan, inf } +;; -// Float the rounded integer to get preliminary result { .mfi - nop.m 999 -(p9) fcvt.xf ROUNDF_FLOAT_RINT_f8 = ROUNDF_RINT_f8 - nop.i 999 ;; -} - -// If x<0 and the difference of the truncated input minus the input is 0.5 -// then result = truncated input - 1.0 -// Else if x>0 and the difference of the input minus truncated input is 0.5 -// then result = truncated input + 1.0 -// Else -// result = rounded input -// Endif -{ .mfi - nop.m 999 -(p6) fsub.s1 ROUNDF_REMAINDER = ROUNDF_FLOAT_TRUNC_f8, ROUNDF_NORM_f8 - nop.i 999 + cmp.lt p6,p0 = rExp, rExpHalf // Is |x| < 0.5? + fcvt.xf f8 = fXtruncInt // Pre-Result if 0.5 <= |x| < 2^23 + cmp.ge p7,p0 = rExp, rBigexp // Is |x| >= 2^23? } - { .mfi - nop.m 999 -(p7) fsub.s1 ROUNDF_REMAINDER = ROUNDF_NORM_f8, ROUNDF_FLOAT_TRUNC_f8 - nop.i 999 ;; + cmp.lt p10,p0 = rExp, rExpHalf // Is |x| < 0.5? + nop.f 0 + nop.i 0 } +;; -// Assume preliminary result is rounded integer +// We must correct result if |x| < 0.5, or |x| >= 2^23 +.pred.rel "mutex",p6,p7 { .mfi - nop.m 999 -(p9) fnorm.s.s0 f8 = ROUNDF_FLOAT_RINT_f8 - nop.i 999 + nop.m 0 +(p6) fmerge.s f8 = fNormX, f0 // If |x| < 0.5, result sgn(x)*0 + nop.i 0 } - -// If x<0, test if result=0 -{ .mfi - nop.m 999 -(p6) fcmp.eq.unc p10,p0 = ROUNDF_FLOAT_RINT_f8,f0 - nop.i 999 ;; +{ .mfb +(p7) cmp.eq p10,p0 = r0, r0 // Also turn on p10 if |x| >= 2^23 +(p7) fma.s.s0 f8 = fNormX, f1, f0 // If |x| >= 2^23, result x +(p10) br.ret.spnt b0 // Exit |x| < 0.5 or |x| >= 2^23 } +;; -// If x<0 and result=0, set result=-0 +// Here if 0.5 <= |x| < 2^23 { .mfi - nop.m 999 -(p10) fmerge.ns f8 = f1,f8 - nop.i 999 + nop.m 0 + fms.s1 fRem = fNormX, f1, f8 // Get remainder = x - trunc(x) + nop.i 0 } - -// If x<0, test if remainder=0.5 +;; + { .mfi - nop.m 999 -(p6) fcmp.eq.unc p6,p0 = ROUNDF_REMAINDER, ROUNDF_HALF - nop.i 999 ;; + nop.m 0 +(p8) fcmp.le.s1 p8,p0 = fRem, fMHalf + nop.i 0 } - -// If x>0, test if remainder=0.5 { .mfi - nop.m 999 -(p7) fcmp.eq.unc p7,p0 = ROUNDF_REMAINDER, ROUNDF_HALF - nop.i 999 ;; + nop.m 0 +(p9) fcmp.ge.s1 p9,p0 = fRem, fHalf + nop.i 0 } +;; -// If x<0 and remainder=0.5, result=truncated-1.0 -// If x>0 and remainder=0.5, result=truncated+1.0 -// Exit -.pred.rel "mutex",p6,p7 +// If x < 0 and remainder <= -0.5, then subtract 1 from result +// If x > 0 and remainder >= +0.5, then add 1 to result +.pred.rel "mutex",p8,p9 { .mfi - nop.m 999 -(p6) fsub.s.s0 f8 = ROUNDF_FLOAT_TRUNC_f8,f1 - nop.i 999 + nop.m 0 +(p8) fms.s.s0 f8 = f8, f1, f1 + nop.i 0 } - { .mfb - nop.m 999 -(p7) fadd.s.s0 f8 = ROUNDF_FLOAT_TRUNC_f8,f1 - br.ret.sptk b0 ;; + nop.m 0 +(p9) fma.s.s0 f8 = f8, f1, f1 + br.ret.sptk b0 +} +;; + + +ROUND_UNORM: +// Here if x unorm +{ .mfb + 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 ROUND_COMMON // Return to main path } +;; -.endp roundf -ASM_SIZE_DIRECTIVE(roundf) +GLOBAL_LIBM_END(roundf) |