This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[PATCH] PowerPC - Fix PPC64 floorl


The commit 2460d3aa21f04cdf28497683bd3e29183189f779 added some additional tests for powl intended
to check a correction for i386/x86_64. The tests however started to trigger some issues with powl in
PPC64. I tracked the issue an it is due the floorl assembly version for PPC64 that fails to correctly
floor some values large than 2**52. 

This patch removes it (which is not an optimization any longer) and replace it by the C version 
adding also an optimization to values lower than 2**52.

--

2012-04-17  Adhemerval Zanella  <azanella@linux.vnet.ibm.com>

	* sysdeps/ieee754/ldbl-128ibm/s_floorl.c: Add an optimization to values
	lower than 2**52.
	* sysdeps/powerpc/powerpc64/fpu/s_floorl.S: Delete file. Using the
	sysdeps/ieee754/ldbl-128ibm/s_floorl.c version. 

diff --git a/sysdeps/ieee754/ldbl-128ibm/s_floorl.c b/sysdeps/ieee754/ldbl-128ibm/s_floorl.c
index 2cee693..0021056 100644
--- a/sysdeps/ieee754/ldbl-128ibm/s_floorl.c
+++ b/sysdeps/ieee754/ldbl-128ibm/s_floorl.c
@@ -27,9 +27,13 @@ long double
 __floorl (long double x)
 {
   double xh, xl, hi, lo;
+  double two52 = 0x10000000000000LL;
 
   ldbl_unpack (x, &xh, &xl);
 
+  if (__builtin_expect(__builtin_fabs (x) < two52, 1))
+      return ldbl_pack (__floor (xh), -0.0);
+
   /* Return Inf, Nan, +/-0 unchanged.  */
   if (__builtin_expect (xh != 0.0
 			&& __builtin_isless (__builtin_fabs (xh),
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)
-- 
1.7.1


-- 
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


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]