This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] PowerPC - Fix PPC64 floorl
On 04/20/2012 05:54 PM, Adhemerval Zanella wrote:
> On 04/19/2012 06:26 PM, Joseph S. Myers wrote:
>> On Thu, 19 Apr 2012, Steven Munroe wrote:
>>
>>>> I don't see how that can be correct. If the high part is small, but an
>>>> integer, the result of this function should still depend on whether the
>>>> low part is positive, negative or zero. Presumably the testsuite failed
>>>> to detect that bug in this patch, meaning you also need tests (conditioned
>>>> on TEST_LDOUBLE - they can be written to be valid for all long double
>>>> formats) for values that when rounded to 53 bits are an integer, but
>>>> differ from the integer value in lower bits (covering at least the four
>>>> cases of positive and negative value, above and below the exact integer it
>>>> is close to).
>>> Are you concerned that xh may be "prerounded" away (from default
>>> rounding mode)? This is not what we want for floor. In which case we
>>> should "deround" xh before the __floor operation.
>>> Perhaps:
>>>
>>> + if (__builtin_expect(__builtin_fabs (x) < two52, 1))
>>> + return ldbl_pack (__floor (xh+xl), -0.0);
>> That won't help either.
>>
>> In a valid long double, xh is always the value of the long double, rounded
>> to nearest, independent of the current rounding direction; that's the
>> definition of the format. The result of floorl doesn't depend on the
>> current rounding direction either; it always rounds to an integer value in
>> the direction of negative infinity.
>>
>> If xh is an integer, and xl is positive or zero with absolute value less
>> than 1, xh is the correct result of floorl. If xh is an integer, and xl
>> is negative with absolute value less than 1, the correct result of floorl
>> is xh - 1. (If xl has absolute value at least 1, which means xh is large,
>> then the correct result is xh + floor(xl); that formula is indeed correct
>> whenever xh is an integer. And if xh is not an integer, the correct
>> result is floor (xh).)
>>
>> The sort of case where I expect this patch will get things wrong is floorl
>> (0x1p51L - 0x1p-10L) or floorl (-0x1p51L - 0x1p-10L).
>>
> Indeed the cases you noted failed with both the old assembly and optimization
> I added based on the assembly. This patch now just use the C version and
> remove the wrong assembly version.
>
> --
>
> [PATCH] PowerPC - Fix PPC64 floorl
>
> The assembly version fails to correctly floorl some edge cases. The patch
> removes it and use the C version.
>
>
> 2012-04-17 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
>
> [BZ #3866]
> * sysdeps/powerpc/powerpc64/fpu/s_floorl.S: Delete file. Using the
> sysdeps/ieee754/ldbl-128ibm/s_floorl.c version.
> * math/libm-test.inc: (floor_test): Add more tests.
> * sysdeps/powerpc/fpu/libm-test-ulps: Update for pow tests.
>
> --
>
> diff --git a/math/libm-test.inc b/math/libm-test.inc
> index 94d60bf..e91aa26 100644
> --- a/math/libm-test.inc
> +++ b/math/libm-test.inc
> @@ -3442,6 +3442,9 @@ floor_test (void)
> TEST_f_f (floor, 10141204801825835211973625643008.5L, 10141204801825835211973625643008.0L);
> TEST_f_f (floor, 10141204801825835211973625643008.75L, 10141204801825835211973625643008.0L);
> TEST_f_f (floor, 10141204801825835211973625643009.5L, 10141204801825835211973625643009.0L);
> +
> + TEST_f_f (floor, 0xf.ffffffffffffff8p+47L, 0xf.fffffffffffep+47L);
> + TEST_f_f (floor, -0x8.000000000000004p+48L, -0x8.000000000001p+48L);
> #endif
>
> END (floor);
> diff --git a/sysdeps/powerpc/fpu/libm-test-ulps b/sysdeps/powerpc/fpu/libm-test-ulps
> index 1ab3d4a..c555297 100644
> --- a/sysdeps/powerpc/fpu/libm-test-ulps
> +++ b/sysdeps/powerpc/fpu/libm-test-ulps
> @@ -1488,6 +1488,17 @@ Test "log2 (e) == M_LOG2El":
> ildouble: 1
> ldouble: 1
>
> +# pow
> +Test "pow (0x0.ffffffp0, -0x1p24) == 2.7182819094701610539628664526874952929416":
> +float: 1
> +ifloat: 1
> +Test "pow (0x0.ffffffp0, 0x1p24) == 0.3678794302077803437135155590023422899744":
> +float: 1
> +ifloat: 1
> +Test "pow (0x1.000002p0, 0x1p24) == 7.3890552180866447284268641248075832310141":
> +float: 1
> +ifloat: 1
> +
> # pow_downward
> Test "pow_downward (1.0625, 1.125) == 1.070582293028761362162622578677070098674":
> ildouble: 1
> @@ -2448,6 +2459,8 @@ ildouble: 1
> ldouble: 1
>
> Function: "pow":
> +float: 1
> +ifloat: 1
> ildouble: 1
> ldouble: 1
>
> diff --git a/sysdeps/powerpc/powerpc64/fpu/s_floorl.S b/sysdeps/powerpc/powerpc64/fpu/s_floorl.S
> deleted file mode 100644
> index 578be43..0000000
> --- a/sysdeps/powerpc/powerpc64/fpu/s_floorl.S
> +++ /dev/null
> @@ -1,133 +0,0 @@
> -/* long double floor function.
> - IBM extended format long double version.
> - Copyright (C) 2004, 2006 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
> -
> - .section ".text"
> -/* long double [fp1,fp2] floorl (long double x [fp1,fp2])
> - IEEE 1003.1 floor function.
> -
> - PowerPC64 long double uses the IBM extended format which is
> - represented two 64-floating point double values. The values are
> - non-overlapping giving an effective precision of 106 bits. The first
> - double contains the high order bits of mantisa and is always rounded
> - to represent a normal rounding of long double to double. Since the
> - long double value is sum of the high and low values, the low double
> - normally has the opposite sign to compensate for the this rounding.
> -
> - For long double there are two cases:
> - 1) |x| < 2**52, all the integer bits are in the high double.
> - floor the high double and set the low double to -0.0.
> - 2) |x| >= 2**52, Rounding involves both doubles.
> - See the comment before lable .L2 for details.
> - */
> -
> -ENTRY (__floorl)
> - 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 7,3 /* Set rounding mode toward -inf. */
> - fneg fp2,fp12 /* set low double to -0.0. */
> - ble- cr6,.L0
> - fadd fp1,fp1,fp13 /* x+= TWO52; */
> - fsub fp1,fp1,fp13 /* x-= TWO52; */
> - fcmpu cr5,fp1,fp12 /* if (x > 0.0) */
> - mtfsf 0x01,fp11 /* restore previous rounding mode. */
> - bnelr+ cr5
> - fmr fp1,fp12 /* x must be +0.0 for the 0.0 case. */
> - blr
> -.L0:
> - bge- cr6,.L1 /* if (x < 0.0) */
> - fsub fp1,fp1,fp13 /* x-= TWO52; */
> - fadd fp1,fp1,fp13 /* x+= TWO52; */
> -.L1:
> - 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 7,3 /* Set rounding mode toward -inf. */
> - 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); */
> - 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); */
> - 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 (__floorl)
> -
> -long_double_symbol (libm, __floorl, floorl)
Although I found out the issue after bz#3866, there is already a bz entry reporting
this issue (bz#13886). New ChangeLog entry (the patch is the same).
---
2012-04-17 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
[BZ #13886]
* sysdeps/powerpc/powerpc64/fpu/s_floorl.S: Delete file. Using the
sysdeps/ieee754/ldbl-128ibm/s_floorl.c version.
* math/libm-test.inc: (floor_test): Add more tests.
* sysdeps/powerpc/fpu/libm-test-ulps: Update for pow tests.
--
Adhemerval Zanella Netto
Software Engineer
Linux Technology Center Brazil
Toolchain / GLIBC on Power Architecture
azanella@linux.vnet.ibm.com / azanella@br.ibm.com
+55 61 8642-9890