This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] PowerPC: remove wrong truncl implementation for PowerPC64
- From: Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>
- To: "GNU C. Library" <libc-alpha at sourceware dot org>
- Date: Tue, 07 Jan 2014 20:07:11 -0200
- Subject: Re: [PATCH] PowerPC: remove wrong truncl implementation for PowerPC64
- Authentication-results: sourceware.org; auth=none
- References: <52CC7A2A dot 3070006 at linux dot vnet dot ibm dot com>
I forgot to add that I'll create a bugzilla for the truncl on PPC64 and
send a patch to add some test inputs for 2.20.
On 07-01-2014 20:05, Adhemerval Zanella wrote:
> While seeking why some tgammal results shows wrong signal on PPC64, I found
> out that the assembly implementation truncl (sysdeps/powerpc/powerpc64/fpu/s_truncl.S)
> for is in fact returning wrong results for some inputs. More specifically,
> for number where the first double is a exact integer and the precision
> is determined by second long double.
>
> Checking on implementation comments and history, I am very confident the
> assembly implementation was based on a version before commit
> 5c68d401698a58cf7da150d9cce769fa6679ba5f that fixes BZ#2423 (Errors in
> long double (ldbl-128ibm) rounding functions in glibc-2.4), since
> the implementation indeed does not handle some inputs correctly.
>
> By just removing the implementation and make the build select
> sysdeps/ieee754/ldbl-128ibm/s_truncl.c instead it fixes the tgammal
> issues.
>
> Tested on PPC64.
>
> ---
>
> 2014-01-07 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
>
> * sysdeps/powerpc/powerpc64/fpu/s_truncl.S: Remove wrong
> implementation.
> * sysdeps/ieee754/ldbl-128ibm/s_truncl.c: Remove bogus comment.
>
> --
>
> diff --git a/sysdeps/ieee754/ldbl-128ibm/s_truncl.c b/sysdeps/ieee754/ldbl-128ibm/s_truncl.c
> index 490ad26..aca3f3e 100644
> --- a/sysdeps/ieee754/ldbl-128ibm/s_truncl.c
> +++ b/sysdeps/ieee754/ldbl-128ibm/s_truncl.c
> @@ -17,9 +17,6 @@
> License along with the GNU C Library; if not, see
> <http://www.gnu.org/licenses/>. */
>
> -/* This has been coded in assembler because GCC makes such a mess of it
> - when it's coded in C. */
> -
> #include <math.h>
> #include <math_ldbl_opt.h>
> #include <float.h>
> diff --git a/sysdeps/powerpc/powerpc64/fpu/s_truncl.S b/sysdeps/powerpc/powerpc64/fpu/s_truncl.S
> deleted file mode 100644
> index e5a3008..0000000
> --- a/sysdeps/powerpc/powerpc64/fpu/s_truncl.S
> +++ /dev/null
> @@ -1,120 +0,0 @@
> -/* long double trunc function.
> - IBM extended format long double version.
> - Copyright (C) 2004-2014 Free Software Foundation, Inc.
> - This file is part of the GNU C Library.
> -
> - The GNU C Library is free software; you can redistribute it and/or
> - modify it under the terms of the GNU Lesser General Public
> - License as published by the Free Software Foundation; either
> - version 2.1 of the License, or (at your option) any later version.
> -
> - The GNU C Library is distributed in the hope that it will be useful,
> - but WITHOUT ANY WARRANTY; without even the implied warranty of
> - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> - Lesser General Public License for more details.
> -
> - You should have received a copy of the GNU Lesser General Public
> - License along with the GNU C Library; if not, see
> - <http://www.gnu.org/licenses/>. */
> -
> -#include <sysdep.h>
> -#include <math_ldbl_opt.h>
> -
> - .section ".toc","aw"
> -.LC0: /* 2**52 */
> - .tc FD_43300000_0[TC],0x4330000000000000
> -.LC1: /* 0.5 */
> - .tc FD_3fe00000_0[TC],0x3fe0000000000000
> - .section ".text"
> -
> -/* long double [fp1,fp2] truncl (long double x [fp1,fp2]) */
> -
> -ENTRY (__truncl)
> - mffs fp11 /* Save current FPU rounding mode. */
> - lfd fp13,.LC0@toc(2)
> - fabs fp0,fp1
> - fabs fp9,fp2
> - fsub fp12,fp13,fp13 /* generate 0.0 */
> - fcmpu cr7,fp0,fp13 /* if (fabs(x) > TWO52) */
> - fcmpu cr6,fp1,fp12 /* if (x > 0.0) */
> - bnl- cr7,.L2
> - mtfsfi 7,1 /* Set rounding mode toward 0. */
> - ble- cr6,.L1
> - fneg fp2,fp12
> - fadd fp1,fp1,fp13 /* x+= TWO52; */
> - fsub fp1,fp1,fp13 /* x-= TWO52; */
> - fabs fp1,fp1 /* if (x == 0.0) x = 0.0; */
> -.L0:
> - mtfsf 0x01,fp11 /* restore previous rounding mode. */
> - blr
> -.L1:
> - fneg fp2,fp12
> - bge- cr6,.L0 /* if (x < 0.0) */
> - fsub fp1,fp1,fp13 /* x-= TWO52; */
> - fadd fp1,fp1,fp13 /* x+= TWO52; */
> - fnabs fp1,fp1 /* if (x == 0.0) x = -0.0; */
> - mtfsf 0x01,fp11 /* restore previous rounding mode. */
> - blr
> -
> -/* The high double is > TWO52 so we need to round the low double and
> - perhaps the high double. In this case we have to round the low
> - double and handle any adjustment to the high double that may be
> - caused by rounding (up). This is complicated by the fact that the
> - high double may already be rounded and the low double may have the
> - opposite sign to compensate.This gets a bit tricky so we use the
> - following algorithm:
> -
> - tau = floor(x_high/TWO52);
> - x0 = x_high - tau;
> - x1 = x_low + tau;
> - r1 = rint(x1);
> - y_high = x0 + r1;
> - y_low = x0 - y_high + r1;
> - return y; */
> -.L2:
> - fcmpu cr7,fp9,fp13 /* if (|x_low| > TWO52) */
> - fcmpu cr0,fp9,fp12 /* || (|x_low| == 0.0) */
> - fcmpu cr5,fp2,fp12 /* if (x_low > 0.0) */
> - bgelr- cr7 /* return x; */
> - beqlr- cr0
> - mtfsfi 7,1 /* Set rounding mode toward 0. */
> - fdiv fp8,fp1,fp13 /* x_high/TWO52 */
> -
> - bng- cr6,.L6 /* if (x > 0.0) */
> - fctidz fp0,fp8
> - fcfid fp8,fp0 /* tau = floor(x_high/TWO52); */
> - fadd fp8,fp8,fp8 /* tau++; Make tau even */
> - bng cr5,.L4 /* if (x_low > 0.0) */
> - fmr fp3,fp1
> - fmr fp4,fp2
> - b .L5
> -.L4: /* if (x_low < 0.0) */
> - fsub fp3,fp1,fp8 /* x0 = x_high - tau; */
> - fadd fp4,fp2,fp8 /* x1 = x_low + tau; */
> -.L5:
> - fadd fp5,fp4,fp13 /* r1 = r1 + TWO52; */
> - fsub fp5,fp5,fp13 /* r1 = r1 - TWO52; */
> - b .L9
> -.L6: /* if (x < 0.0) */
> - fctidz fp0,fp8
> - fcfid fp8,fp0 /* tau = floor(x_high/TWO52); */
> - fadd fp8,fp8,fp8 /* tau++; Make tau even */
> - bnl cr5,.L7 /* if (x_low < 0.0) */
> - fmr fp3,fp1
> - fmr fp4,fp2
> - b .L8
> -.L7: /* if (x_low > 0.0) */
> - fsub fp3,fp1,fp8 /* x0 = x_high - tau; */
> - fadd fp4,fp2,fp8 /* x1 = x_low + tau; */
> -.L8:
> - fsub fp5,fp4,fp13 /* r1-= TWO52; */
> - fadd fp5,fp5,fp13 /* r1+= TWO52; */
> -.L9:
> - mtfsf 0x01,fp11 /* restore previous rounding mode. */
> - fadd fp1,fp3,fp5 /* y_high = x0 + r1; */
> - fsub fp2,fp3,fp1 /* y_low = x0 - y_high + r1; */
> - fadd fp2,fp2,fp5
> - blr
> -END (__truncl)
> -
> -long_double_symbol (libm, __truncl, truncl)
>