jwp@uwmacc.UUCP (Jeffrey W Percival) (06/02/87)
I have a problem I have distilled down into a short C program.
The problem is that when I subtract a tiny number from 1,
I get 1 (the right answer) on a VAX 780 running 4.3BSD, but
I get 0 (the wrong answer) on a MicroVAX running 4.3BSD. Here's
the program:
---------------------------------------------------
#include <stdio.h>
#include <math.h>
main()
{
double a, b, c, d, e;
a = 1.000000;
b = 0.000001;
c = 0.193113;
d = (b * c);
e = a - d;
fprintf(stdout, "a=%f, b=%f, c=%f, d=%f, e=%f\n", a, b, c, d, e);
exit(0);
}
----------------------------------------------------------
and here's the MicroVAX output followed by the VAX 780 output:
-------------------------------------------------------
a=1.000000, b=0.000001, c=0.193113, d=0.000000, e=0.000000
a=1.000000, b=0.000001, c=0.193113, d=0.000000, e=1.000000
So, the MicroVAX says that (a-d) is 0, not 1.
What's the deal here?
--
Jeff Percival (jwp@salvax1)chris@mimsy.UUCP (06/02/87)
In article <1573@uwmacc.UUCP> jwp@uwmacc.UUCP (Jeffrey W Percival) writes: >The problem is that when I subtract a tiny number from 1, >I get 1 (the right answer) on a VAX 780 running 4.3BSD, but >I get 0 (the wrong answer) on a MicroVAX running 4.3BSD. Here's >the program: [the important line:] > fprintf(stdout, "a=%f, b=%f, c=%f, d=%f, e=%f\n", a, b, c, d, e); There is a bug in the emulation code in /sys/vax/emulate.s. Try main(){printf("%.0f\n", 0.9);exit(0);} I have not seen a fix on comp.bugs.4bsd.ucb-fixes. -- In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 7690) Domain: chris@mimsy.umd.edu Path: seismo!mimsy!chris
drw@cullvax.UUCP (Dale Worley) (06/04/87)
jwp@uwmacc.UUCP (Jeffrey W Percival) writes: > I have a problem I have distilled down into a short C program. > The problem is that when I subtract a tiny number from 1, > I get 1 (the right answer) on a VAX 780 running 4.3BSD, but > I get 0 (the wrong answer) on a MicroVAX running 4.3BSD. Here's > the program: This may have to do with the fact that the code in BSD that simulates the VAX instructions that the uVAX hardware doesn't have has problems with it. If you have VMS on a uVAX also, use the debuggers to see if the floating subtract instructions produce different results on the two OS's from the same arguments. (There's someone out in netland who knows more about this.) Dale -- Dale Worley Cullinet Software UUCP: ...!seismo!harvard!mit-eddie!cullvax!drw ARPA: cullvax!drw@eddie.mit.edu Un*x (a generic name for a class of OS's) != Unix (AT&T's brand of such)
ed@mtxinu.UUCP (06/04/87)
>>The problem is that when I subtract a tiny number from 1, > >There is a bug in the emulation code in /sys/vax/emulate.s. Try > > main(){printf("%.0f\n", 0.9);exit(0);} > >I have not seen a fix on comp.bugs.4bsd.ucb-fixes. I posted fixes for that problem to this group some months ago. mt Xinu customers - including educational institutions - with maintenance contracts also received a Maintenance Release with the correct code in the past several weeks. I don't have the diffs handy right now; if anyone needs them and can't get them either from an archive or from our product release, send me *mail* and I'll dig them out again. -- Ed Gould mt Xinu, 2560 Ninth St., Berkeley, CA 94710 USA {ucbvax,decvax}!mtxinu!ed +1 415 644 0146 "A man of quality is not threatened by a woman of equality."
rick@seismo.CSS.GOV (Rick Adams) (06/05/87)
From ed@mtxinu.UUCP Tue Jan 27 15:57:30 1987
Path: beno!seismo!rutgers!sri-spam!mordor!lll-lcc!unisoft!mtxinu!ed
From: ed@mtxinu.UUCP (Ed Gould)
Newsgroups: comp.bugs.4bsd
Subject: MicroVAX emulation code bug fixes
Message-ID: <289@mtxinu.UUCP>
Date: 27 Jan 87 20:57:30 GMT
Reply-To: ed@mtxinu.UUCP (Ed Gould)
Distribution: comp
Organization: mt Xinu, Berkeley, CA
Lines: 246
Index: sys/vax/emulate.s 4.3BSD
Status: RO
Description:
The MicroVAX emulation code for unimplemented instructions has
a few bugs that show themselves at the limits of "printf"s.
Also, there are numerous places where separate increment and
test instructions can be combined into one "sob" instruction.
Repeat-By:
On a uVAX-II running 4.3:
printf("%.0f\n", 0.9);
produces
0
And, in some cases, "df" will produce a capacity of 00% when
it should be 100%. Other mysterious things can happen to code
written by those strange folk who program in assembly.
Fix:
A diff of emulate.s follows:
58,61c56,57
< toarg(r8,1) # save r8 in arg1 spot
< argl(4,r8) # (4) source address == r8
< toarg(r1,4) # save r1 in arg4 spot
< tstl arg3 # (3) source length == "arg3"
---
> argl(4,r3) # (4) source address == r3
> arguw(3,r2) # (3) source length == r2
64c60
< xorb2 (r8)+,r0
---
> xorb2 (r3)+,r0
71,72c67
< decl arg3
< jneq Lcrc_loop
---
> sobgtr r2,Lcrc_loop
76,77c71
< argl(1,r8)
< argl(4,r1)
---
> clrl r1
97,99c91,92
< decl r4
< jeql Lmovtc_out
< jbr Lmovtc_loop
---
> sobgtr r4,Lmovtc_loop
> jbr Lmovtc_out
102,103c95
< decl r4
< jneq Lmovtc_2loop
---
> sobgtr r4,Lmovtc_2loop
131,132c123
< decl r4
< jneq Lmovtuc_loop
---
> sobgtr r4,Lmovtuc_loop
163,164c154
< decl r2
< jneq Lmatchc_loop
---
> sobgtr r2,Lmatchc_loop
168a159
> movl r1,r3
180d170
< argl(3,r3) # (3) table address == r3
181a172
> argl(3,r3) # (3) table address == r3
190,191c181
< decl r0
< jneq Lspanc_loop
---
> sobgtr r0,Lspanc_loop
202d191
< argl(3,r3) # (3) table address == r3
203a193
> argl(3,r3) # (3) table address == r3
212,213c202
< decl r0
< jneq Lscanc_loop
---
> sobgtr r0,Lscanc_loop
226c215
< incl r0
---
> jeql Lskpc_out # forget zero length strings
228,233c217,220
< decl r0
< jeql Lskpc_out
< cmpb (r1)+,r11
< jeql Lskpc_loop
< decl r1
< tstl r0
---
> cmpb (r1),r11
> jneq Lskpc_out
> incl r1
> sobgtr r0,Lskpc_loop
234a222
> tstl r0 # be sure of condition codes
245c233
< incl r0
---
> jeql Lskpc_out # forget zero length strings
247c235
< decl r0
---
> cmpb (r1),r11
249,252c237,238
< cmpb (r1)+,r11
< jneq Llocc_loop
< decl r1
< tstl r0
---
> incl r1
> sobgtr r0,Llocc_loop
253a240
> tstl r0 # be sure of condition codes
270,271c257
< decl r0
< jneq Lcmpc3_loop
---
> sobgtr r0,Lcmpc3_loop
295,296c281
< decl r0
< jneq Lcmpc5_loop
---
> sobgtr r0,Lcmpc5_loop
304,305c289
< decl r2
< jneq Lcmpc5_str2loop
---
> sobgtr r2,Lcmpc5_str2loop
311,312c295
< decl r0
< jneq Lcmpc5_str1loop
---
> sobgtr r0,Lcmpc5_str1loop
514,515c497
< decl r11 # while (--source length)
< jneq Laddp4_same_loop
---
> sobgtr r11,Laddp4_same_loop # while (--source length)
583,584c565
< decl r11 # loop for length of source
< jneq Lmovp_copy
---
> sobgtr r11,Lmovp_copy # loop for length of source
797,798c778
< decl r11
< jneq L200
---
> sobgtr r11,L200
830,831c810
< decl r1
< jneq Le_fill_loop
---
> sobgtr r1,Le_fill_loop
856,857c835
< decl r1
< jneq L214
---
> sobgtr r1,L214
885,886c863
< decl r1
< jneq L221
---
> sobgtr r1,L221
901a879
> # arg2 holds "even or odd" destination length
918,919c896,901
< L245: # r3<0> counts digits going into destination
< bisl2 $1,r3 # and is flip-flop for which nibble to
---
> L245:
> clrl arg2 # arg2 is 1 if dstlen is even, 0 if odd
> blbs r3,L246
> incl arg2
> bisl2 $1,r3 # r3<0> counts digits going into destination
> L246: # and is flip-flop for which nibble to
946,948c928,929
< decl r3 # move to next nibble in destination, but
< jneq L258 # don't go beyond the end.
< incl r3
---
> sobgtr r3,L258 # move to next nibble in destination, but
> incl r3 # don't go beyond the end.
959c940
< jeql Lashp_sethigh
---
> jeql Lashp_round
981,984c962,966
< jgeq Lashp_noovfl # if we've moved passed destination limits
< clrl r3 # test the result for possible overflow
< tstl arg5 # ignore zero nibbles
< jeql L265 # if the nibble was non-zero, overflow
---
> cmpl r3,arg2 # if we've moved passed destination limits
> jgeq Lashp_noovfl # test the result for possible overflow
> movl arg2,r3 # ignore zero nibbles
> tstl arg5 # if the nibble was non-zero, overflow
> jeql L265
991,992c973
< decl r6
< insv arg5,$0,$4,(r6)
---
> movb arg5,-(r6)
994,995c975
< decl r10 # loop for length of source
< jneq Lashp_shloop
---
> sobgtr r10,Lashp_shloop # loop for length of source
997,1000c977,985
< Lashp_sethigh:
< jlbc r3,L266 # if we haven't set the high nibble,
< insv r2,$4,$4,(r6) # carry the round into the high nibble
< clrl r2
---
> Lashp_round:
> tstl r2 # take care of round out of high nibble
> jeql Lashp_zeroround
> decl r3
> cmpl r3,arg2 # if we've moved passed destination limits
> jlss Lashp_overfl # then overflow
> jlbs r3,L266
> insv arg5,$4,$4,(r6) # put the round into destination (high or low)
> jbr Lashp_zeroround
1001a987,989
> movb arg5,-(r6)
>
> Lashp_zeroround:
1008,1009c996
< cvtlb r2,-(r6) # fill up MSNs of destination with carry or zero
< clrl r2
---
> clrb -(r6) # fill up MSNs of destination with zeros
1013,1014d999
< tstl r2 # if carry beyond destination, overflow
< jneq Lashp_overfl
1102,1103c1087
< decl r11
< jneq Lcvtpl_loop
---
> sobgtr r11,Lcvtpl_loop
--
Ed Gould mt Xinu, 2560 Ninth St., Berkeley, CA 94710 USA
{ucbvax,decvax}!mtxinu!ed +1 415 644 0146
"A man of quality is not threatened by a woman of equality."