lkml.org 
[lkml]   [2000]   [Sep]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: Alpha/Linux FP denormal processing
On Tue, Sep 19, 2000 at 09:52:15AM -0700, Richard Henderson wrote:
> > Instead of *= 0.5, try /= 2.0
>
> Yes indeed you've found a bug in the kernel's FP emulation.
> I'll see about fixing it.

Rather than fix the old udiv128 function, which was trying to do
128/128 bit division, I've pulled in a subroutine from libgcc
that does 128/64 bit division, which is all we need here.

So it should be a bit faster than the old routine, for what ever
that's worth to someone who is already trapping into the kernel...



r~
diff -ruNp linux/arch/alpha/math-emu/Makefile 2.4.0-9-4/arch/alpha/math-emu/Makefile
--- linux/arch/alpha/math-emu/Makefile Thu Dec 2 15:28:54 1999
+++ 2.4.0-9-4/arch/alpha/math-emu/Makefile Tue Sep 19 18:45:17 2000
@@ -8,7 +8,7 @@
# Note 2! The CFLAGS definition is now in the main makefile...

O_TARGET := math-emu.o
-O_OBJS := math.o
+O_OBJS := math.o qrnnd.o
CFLAGS += -I. -I$(TOPDIR)/include/math-emu -w

ifeq ($(CONFIG_MATHEMU),m)
diff -ruNp linux/arch/alpha/math-emu/math.c 2.4.0-9-4/arch/alpha/math-emu/math.c
--- linux/arch/alpha/math-emu/math.c Tue Mar 21 10:47:06 2000
+++ 2.4.0-9-4/arch/alpha/math-emu/math.c Tue Sep 19 18:26:38 2000
@@ -84,66 +84,6 @@ void cleanup_module(void)

#endif /* MODULE */

-/* For 128-bit division. */
-
-void
-udiv128(unsigned long divisor_f0, unsigned long divisor_f1,
- unsigned long dividend_f0, unsigned long dividend_f1,
- unsigned long *quot, unsigned long *remd)
-{
- _FP_FRAC_DECL_2(quo);
- _FP_FRAC_DECL_2(rem);
- _FP_FRAC_DECL_2(tmp);
- unsigned long i, num_bits, bit;
-
- _FP_FRAC_SET_2(rem, _FP_ZEROFRAC_2);
- _FP_FRAC_SET_2(quo, _FP_ZEROFRAC_2);
-
- if (_FP_FRAC_ZEROP_2(divisor))
- goto out;
-
- if (_FP_FRAC_GT_2(divisor, dividend)) {
- _FP_FRAC_COPY_2(rem, dividend);
- goto out;
- }
-
- if (_FP_FRAC_EQ_2(divisor, dividend)) {
- __FP_FRAC_SET_2(quo, 0, 1);
- goto out;
- }
-
- num_bits = 128;
- while (1) {
- bit = _FP_FRAC_NEGP_2(dividend);
- _FP_FRAC_COPY_2(tmp, rem);
- _FP_FRAC_SLL_2(tmp, 1);
- _FP_FRAC_LOW_2(tmp) |= bit;
- if (! _FP_FRAC_GE_2(tmp, divisor))
- break;
- _FP_FRAC_COPY_2(rem, tmp);
- _FP_FRAC_SLL_2(dividend, 1);
- num_bits--;
- }
-
- for (i = 0; i < num_bits; i++) {
- bit = _FP_FRAC_NEGP_2(dividend);
- _FP_FRAC_SLL_2(rem, 1);
- _FP_FRAC_LOW_2(rem) |= bit;
- _FP_FRAC_SUB_2(tmp, rem, divisor);
- bit = _FP_FRAC_NEGP_2(tmp);
- _FP_FRAC_SLL_2(dividend, 1);
- _FP_FRAC_SLL_2(quo, 1);
- if (!bit) {
- _FP_FRAC_LOW_2(quo) |= 1;
- _FP_FRAC_COPY_2(rem, tmp);
- }
- }
-
-out:
- *quot = quo_f1;
- *remd = rem_f1;
- return;
-}

/*
* Emulate the floating point instruction at address PC. Returns 0 if
diff -ruNp linux/arch/alpha/math-emu/qrnnd.S 2.4.0-9-4/arch/alpha/math-emu/qrnnd.S
--- linux/arch/alpha/math-emu/qrnnd.S Wed Dec 31 16:00:00 1969
+++ 2.4.0-9-4/arch/alpha/math-emu/qrnnd.S Tue Sep 19 18:41:26 2000
@@ -0,0 +1,163 @@
+ # Alpha 21064 __udiv_qrnnd
+ # Copyright (C) 1992, 1994, 1995, 2000 Free Software Foundation, Inc.
+
+ # This file is part of GCC.
+
+ # The GNU MP Library is free software; you can redistribute it and/or modify
+ # it under the terms of the GNU General Public License as published by
+ # the Free Software Foundation; either version 2 of the License, or (at your
+ # option) any later version.
+
+ # In addition to the permissions in the GNU General Public License, the
+ # Free Software Foundation gives you unlimited permission to link the
+ # compiled version of this file with other programs, and to distribute
+ # those programs without any restriction coming from the use of this
+ # file. (The General Public License restrictions do apply in other
+ # respects; for example, they cover modification of the file, and
+ # distribution when not linked into another program.)
+
+ # This file 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 Library General Public
+ # License for more details.
+
+ # You should have received a copy of the GNU General Public License
+ # along with GCC; see the file COPYING. If not, write to the
+ # Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
+ # MA 02111-1307, USA.
+
+ .set noreorder
+ .set noat
+
+ .text
+
+ .globl __udiv_qrnnd
+ .ent __udiv_qrnnd
+__udiv_qrnnd:
+ .frame $30,0,$26,0
+ .prologue 0
+
+#define cnt $2
+#define tmp $3
+#define rem_ptr $16
+#define n1 $17
+#define n0 $18
+#define d $19
+#define qb $20
+#define AT $at
+
+ ldiq cnt,16
+ blt d,$largedivisor
+
+$loop1: cmplt n0,0,tmp
+ addq n1,n1,n1
+ bis n1,tmp,n1
+ addq n0,n0,n0
+ cmpule d,n1,qb
+ subq n1,d,tmp
+ cmovne qb,tmp,n1
+ bis n0,qb,n0
+ cmplt n0,0,tmp
+ addq n1,n1,n1
+ bis n1,tmp,n1
+ addq n0,n0,n0
+ cmpule d,n1,qb
+ subq n1,d,tmp
+ cmovne qb,tmp,n1
+ bis n0,qb,n0
+ cmplt n0,0,tmp
+ addq n1,n1,n1
+ bis n1,tmp,n1
+ addq n0,n0,n0
+ cmpule d,n1,qb
+ subq n1,d,tmp
+ cmovne qb,tmp,n1
+ bis n0,qb,n0
+ cmplt n0,0,tmp
+ addq n1,n1,n1
+ bis n1,tmp,n1
+ addq n0,n0,n0
+ cmpule d,n1,qb
+ subq n1,d,tmp
+ cmovne qb,tmp,n1
+ bis n0,qb,n0
+ subq cnt,1,cnt
+ bgt cnt,$loop1
+ stq n1,0(rem_ptr)
+ bis $31,n0,$0
+ ret $31,($26),1
+
+$largedivisor:
+ and n0,1,$4
+
+ srl n0,1,n0
+ sll n1,63,tmp
+ or tmp,n0,n0
+ srl n1,1,n1
+
+ and d,1,$6
+ srl d,1,$5
+ addq $5,$6,$5
+
+$loop2: cmplt n0,0,tmp
+ addq n1,n1,n1
+ bis n1,tmp,n1
+ addq n0,n0,n0
+ cmpule $5,n1,qb
+ subq n1,$5,tmp
+ cmovne qb,tmp,n1
+ bis n0,qb,n0
+ cmplt n0,0,tmp
+ addq n1,n1,n1
+ bis n1,tmp,n1
+ addq n0,n0,n0
+ cmpule $5,n1,qb
+ subq n1,$5,tmp
+ cmovne qb,tmp,n1
+ bis n0,qb,n0
+ cmplt n0,0,tmp
+ addq n1,n1,n1
+ bis n1,tmp,n1
+ addq n0,n0,n0
+ cmpule $5,n1,qb
+ subq n1,$5,tmp
+ cmovne qb,tmp,n1
+ bis n0,qb,n0
+ cmplt n0,0,tmp
+ addq n1,n1,n1
+ bis n1,tmp,n1
+ addq n0,n0,n0
+ cmpule $5,n1,qb
+ subq n1,$5,tmp
+ cmovne qb,tmp,n1
+ bis n0,qb,n0
+ subq cnt,1,cnt
+ bgt cnt,$loop2
+
+ addq n1,n1,n1
+ addq $4,n1,n1
+ bne $6,$Odd
+ stq n1,0(rem_ptr)
+ bis $31,n0,$0
+ ret $31,($26),1
+
+$Odd:
+ /* q' in n0. r' in n1 */
+ addq n1,n0,n1
+
+ cmpult n1,n0,tmp # tmp := carry from addq
+ subq n1,d,AT
+ addq n0,tmp,n0
+ cmovne tmp,AT,n1
+
+ cmpult n1,d,tmp
+ addq n0,1,AT
+ cmoveq tmp,AT,n0
+ subq n1,d,AT
+ cmoveq tmp,AT,n1
+
+ stq n1,0(rem_ptr)
+ bis $31,n0,$0
+ ret $31,($26),1
+
+ .end __udiv_qrnnd
diff -ruNp linux/arch/alpha/math-emu/sfp-util.h 2.4.0-9-4/arch/alpha/math-emu/sfp-util.h
--- linux/arch/alpha/math-emu/sfp-util.h Thu Dec 2 15:28:54 1999
+++ 2.4.0-9-4/arch/alpha/math-emu/sfp-util.h Tue Sep 19 18:41:05 2000
@@ -17,18 +17,13 @@
: "r" ((UDItype)(u)), \
"r" ((UDItype)(v)))

-extern void udiv128(unsigned long, unsigned long,
- unsigned long, unsigned long,
- unsigned long *,
- unsigned long *);
-
-#define udiv_qrnnd(q, r, n1, n0, d) \
- do { \
- unsigned long xr, xi; \
- udiv128((n0), (n1), 0, (d), &xr, &xi); \
- (q) = xr; \
- (r) = xi; \
+#define udiv_qrnnd(q, r, n1, n0, d) \
+ do { unsigned long __r; \
+ (q) = __udiv_qrnnd (&__r, (n1), (n0), (d)); \
+ (r) = __r; \
} while (0)
+extern unsigned long __udiv_qrnnd (unsigned long *, unsigned long,
+ unsigned long , unsigned long);

#define UDIV_NEEDS_NORMALIZATION 1
\
 
 \ /
  Last update: 2005-03-22 12:39    [W:0.114 / U:0.008 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site