From a334319f6530564d22e775935d9c91663623a1b4 Mon Sep 17 00:00:00 2001 From: Ulrich Drepper Date: Wed, 22 Dec 2004 20:10:10 +0000 Subject: (CFLAGS-tst-align.c): Add -mpreferred-stack-boundary=4. --- sysdeps/ia64/fpu/s_truncf.S | 207 ++++++++++++++++++++++++++------------------ 1 file changed, 121 insertions(+), 86 deletions(-) (limited to 'sysdeps/ia64/fpu/s_truncf.S') diff --git a/sysdeps/ia64/fpu/s_truncf.S b/sysdeps/ia64/fpu/s_truncf.S index ff40bc7101..0ac4181209 100644 --- a/sysdeps/ia64/fpu/s_truncf.S +++ b/sysdeps/ia64/fpu/s_truncf.S @@ -1,10 +1,11 @@ .file "truncf.s" - -// Copyright (c) 2000 - 2003, Intel Corporation +// Copyright (C) 2000, 2001, Intel Corporation // All rights reserved. -// -// Contributed 2000 by the Intel Numerics Group, Intel Corporation +// +// Contributed 7/7/2000 by John Harrison, Cristina Iordache, Ted Kubaska, +// Bob Norin, Shane Story, and Ping Tak Peter Tang of the +// Computational Software Lab, Intel Corporation. // // Redistribution and use in source and binary forms, with or without // modification, are permitted provided that the following conditions are @@ -20,28 +21,33 @@ // * 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://www.intel.com/software/products/opensource/libraries/num.htm. +// problem reports or change requests be submitted to it directly at +// http://developer.intel.com/opensource. // +.align 32 +.global truncf# + +.section .text +.proc truncf# +.align 32 + // History //============================================================== -// 07/07/00 Created -// 05/20/02 Cleaned up namespace and sf0 syntax -// 01/20/03 Improved performance and reduced code size +// 7/7/00: Created //============================================================== // API @@ -49,28 +55,25 @@ // float truncf(float x) //============================================================== -// general input registers: -// r14 - r18 +#include "libm_support.h" -rExpBias = r14 -rSignexp = r15 -rExp = r16 -rExpMask = r17 -rBigexp = r18 +// general input registers: +TRUNC_GR_FFFF = r14 +TRUNC_GR_signexp = r15 +TRUNC_GR_exponent = r16 +TRUNC_GR_expmask = r17 +TRUNC_GR_bigexp = r18 // floating-point registers: -// f8 - f10 - -fXtruncInt = f9 -fNormX = f10 +// f8, f9, f11, f12 -// predicate registers used: -// p6, p7 +// predicate registers used: +// p6, p7, p8, p9, p10, p11 // Overview of operation //============================================================== // float truncf(float x) -// Return an integer value (represented as a float) less than or +// Return an integer value (represented as a float) less than or // equal to x in magnitude. // This is x rounded toward zero to an integral value. //============================================================== @@ -94,73 +97,105 @@ fNormX = f10 // If we multiply by 2^23, we no longer have a fractional part // So input is an integer value already. -.section .text -GLOBAL_LIBM_ENTRY(truncf) +truncf: { .mfi - 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 + getf.exp TRUNC_GR_signexp = f8 + fcvt.fx.trunc.s1 f9 = f8 + addl TRUNC_GR_bigexp = 0x10016, r0 } { .mfi - mov rExpBias = 0x0FFFF // Form exponent bias - fnorm.s1 fNormX = f8 // Normalize input - mov rExpMask = 0x1FFFF // Form exponent mask -} -;; + mov TRUNC_GR_FFFF = 0x0FFFF + fnorm.s f11 = f8 + mov TRUNC_GR_expmask = 0x1FFFF +};; +// get the exponent of x +// convert x to integer in signficand of f9 +// Normalize x - this will raise invalid on SNaNs, the +// denormal operand flag - and possibly a spurious U flag +// get exponent only mask (will exclude sign bit) { .mfi nop.m 0 - fclass.m p7,p0 = f8, 0x0b // Test x unorm + fclass.m p7,p8 = f8, 0x0b nop.i 0 } -;; - -{ .mfb +{ .mfi nop.m 0 - fclass.m p6,p0 = f8, 0x1e3 // Test x natval, nan, inf -(p7) br.cond.spnt TRUNC_UNORM // Branch if x unorm -} -;; - -TRUNC_COMMON: -// Return here from TRUNC_UNORM + fcmp.eq.unc.s1 p9,p0 = f8,f0 + nop.i 0 +};; +// fclass to set p7 if unnorm +{ .mmi + and TRUNC_GR_exponent = TRUNC_GR_signexp, TRUNC_GR_expmask ;; +(p8) cmp.ge.unc p10,p11 = TRUNC_GR_exponent, TRUNC_GR_bigexp +(p8) cmp.ne.unc p6,p0 = TRUNC_GR_exponent, TRUNC_GR_signexp +};; +// Get the exponent of x +// Test if exponent such that result already an integer +// Test if x < 0 +{ .mmi +(p9) cmp.eq.andcm p10,p11 = r0, r0 +(p6) cmp.lt.unc p6,p0 = TRUNC_GR_exponent, TRUNC_GR_FFFF + nop.i 0 +};; +// If -1 < x < 0, set p6, turn off p10 and p11, and set result to -0.0 +{ .mfb +(p6) cmp.eq.andcm p10,p11 = r0, r0 +(p6) fmerge.s f8 = f8, f0 + nop.b 0 +};; +// If not a unnorm, set p10 if x already is a big int, nan, or inf? +// If not a unnorm, set p10 if x already is a big int, nan, or inf? +.pred.rel "mutex",p10,p11 { .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 -} -;; - -{ .mfi - cmp.lt p6,p0 = rExp, rExpBias // Is |x| < 1? - fcvt.xf f8 = fXtruncInt // Result, assume 1 <= |x| < 2^23 - cmp.ge p7,p0 = rExp, rBigexp // Is |x| >= 2^23? -} -;; - -// We must correct result if |x| < 1, or |x| >= 2^23 -.pred.rel "mutex",p6,p7 -{ .mfi nop.m 0 -(p6) fmerge.s f8 = fNormX, f0 // If |x| < 1, result sgn(x)*0 - nop.i 0 +(p11) fcvt.xf f8 = f9 + nop.b 0 } { .mfb nop.m 0 -(p7) fma.s.s0 f8 = fNormX, f1, f0 // If |x| >= 2^23, result x - br.ret.sptk b0 // Exit main path -} -;; - - -TRUNC_UNORM: -// Here if x unorm +(p10) fma.s.s1 f8 = f11,f1,f0 +(p8) br.ret.sptk b0 +};; +// If not a unnorm and not an big int, nan,or +/-inf convert signficand +// back to f8. +// If not a unorm and a big int, nan, or +/-inf, return fnorm'd x +// If not a unorm, Return +// If unnorm, get the exponent again - perhaps it wasn't a denorm. +{ .mfb +(p7) getf.exp TRUNC_GR_signexp = f11 +(p7) fcvt.fx.trunc.s1 f12 = f11 + nop.b 0 +};; +{ .mfb + and TRUNC_GR_exponent = TRUNC_GR_signexp, TRUNC_GR_expmask + fcmp.lt.unc.s1 p9,p0 = f8,f0 + nop.b 0 +};; { .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 TRUNC_COMMON // Return to main path + cmp.ge.unc p10,p11 = TRUNC_GR_exponent, TRUNC_GR_bigexp + nop.f 0 + nop.b 0 +};; +// If a unnorm, check to see if value is already a big int. +{ .mfb + nop.m 0 +(p11) fcvt.xf f8 = f12 + nop.b 0 } -;; - -GLOBAL_LIBM_END(truncf) +{ .mfi + nop.m 0 +(p10) fma.s.s1 f8 = f11,f1,f0 + nop.i 0 +};; +{ .mfb + nop.m 0 +(p9) fmerge.ns f8 = f1,f8 + br.ret.sptk b0 +};; +// If so return it. Otherwise, return (fcvt.xf(fcvt.fx.trunc(x))) +// Make sure the result is negative if it should be - that is +// negative(denormal) -> -0. +.endp truncf +ASM_SIZE_DIRECTIVE(truncf) -- cgit 1.4.1