topic Re: Trouble with Logical .AND. Operator in Intel® Fortran Compiler

Trouble with Logical .AND. Operator

jackosullivan1 — Sat, 25 Aug 2007 05:54:37 GMT

Compiling a purely Fortran program -- self-contained exe --with the following options generated through the MS VisualStudio 2003.NET (a la Visual C++.NET):

/nologo /Zi /Od /include:"Debug/" /fpscomp:filesfromcmd /fpscomp:ioformat
/fpscomp:logicals /fpscomp:general /error_limit:10 /warn:errors /warn:unused
/warn:interfaces /real_size:64 /Qsave /align:rec8byte /align:dcommons
/assume:dummy_aliases /Qzero /fpe:0 /iface:cvf /module:"Debug/" /object:"Debug/"
/traceback /check:bounds /check:format /libs:qwin /c

I'm getting an erroneous ".TRUE." with the following code which is in a .f90 source file. I am linking it witha large number of .for files, but all the action involved in this erroris local to the .f90source. (Line numbers are for reference only and not included in the code):

1 Logical a,b

2 x=1.
3 y=2.
4 a=.TRUE.
5 v=0.
6 if(v=0.)b=.FALSE.

7 if(a .AND. b)then
8 p=x/v
9 else
10 p=x/y
11 endif

Line 8 gets executed whenever a is .TRUE. (according to the cursor-hover help in MSVS2003.net), regardless of the value of b... not a nice thing when I have a zero in v.

Any ideas on what I might be doing wrong would be greatly appreciated.

Thanks and God bless!
Jack O'Sullivan

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Sat, 25 Aug 2007 12:28:38 GMT

I would ask that you verify that line 8 is actually being executed and that it isn't just the debugger making a mistake. PC-line correlation information can sometimes be flaky.

I will also say that sometimes I see paraphrasing of code that doesn't match the actual code. If you can show that the wrong line is actually being executed, please create a complete test case and submit it to Intel Premier Support.

Re: Trouble with Logical .AND. Operator

TimP — Sat, 25 Aug 2007 16:27:12 GMT

In fact, the way you have quoted the code, it should throw a compile error. If you wrote the same kind of thing in C, and disabled warnings, you would get the kind of behavior you report.

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Sat, 25 Aug 2007 19:39:20 GMT

Tim, I don't see the error. What is it?

Re: Trouble with Logical .AND. Operator

TimP — Sun, 26 Aug 2007 04:59:55 GMT

On my screen, it shows if(v=0.) ....
where if(v == 0.) ...
might have been meant.

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Sun, 26 Aug 2007 15:30:43 GMT

Good eyes! However, I am pretty sure that this is a "representation" and not the actual code. An easy typo to make (I've done that myself too many times.) You're right that in Fortran this would be a syntax error and in C a different meaning.

Re: Trouble with Logical .AND. Operator

Paul_Curtis — Sun, 26 Aug 2007 17:08:25 GMT

Logical quantity b is undefined (defaults to .FALSE.?) in the code as written.

Perhaps line 6 should read

b = (v /= 0.)

which always explicitly assigns the value of b.

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Sun, 26 Aug 2007 20:54:19 GMT

True enough, though if the code shown is an accurate representation, b should be set to .FALSE. because v is zero. There is no default initialization here. My guess is that the actual code does something different.

Re: Trouble with Logical .AND. Operator

jackosullivan1 — Mon, 27 Aug 2007 14:58:39 GMT

Thanks, folks, for your attention to this. Sorry to have put you all to such trouble.

In following Steve's initial "homework assignment" to create and submit a complete test case, I found the error -- one which I do believe should have been caught by the compiler.

Steve correctly recognized the bad paraphrase with the = rather than ==, and my paraphrasing of the problem area of the program was definitely incomplete.

It turns out that what caused the error -- even though I have called for the compiler to treat all warnings as errors -- was a duplicate specification of one of the logical variables.

A true extract of thesubroutine is shown below:

SUBROUTINE lngvalv [from .f90]
! -- removed from the CONTAINS in the subroutine which now calls
! this one because the IVF interface to Visual Studio does not
! permit debugging of the containER variables when in the
! containED portion of the source file

....

use mlimits
 [which is as follows from .for:
 module mLimits
 integer(4), parameter ::
 &maxsecno=20000
 &,maxvalv=300
 ....
 end module mLimits]

....

use mValpc
 [which is as follows from .for:
 module mValpc
 use mLimits
 COMMON/VALPC/T01(21,maxvalv),S01(21,maxvalv),STK(21,maxvalv)
 & ,COD(21,maxvalv),TQ,VALDIM(maxvalv)
 1 ,COEF,NCRV1(maxvalv),NCRV2(maxvalv) ,KKK,SX,ITIM,ISTK
 &,THETO(maxvalv),
 2 THETS(maxvalv),DELH,IOPG(maxvalv),TV(maxvalv),ICHKV(maxvalv)
 &,CDN(21,maxvalv),VVN(maxvalv),NCV2
 save /valpc/
 end module mValpc ]

....

use mvalvlng
 [which is as follows from .for:
 module mValvLNG
 use mLimits
 DOUBLE PRECISION BB,CC,C1R,C1S,C3,C4,TA,TA2
 1 ,vv,vu,vd,pvv,voidu,voidd,hvvap,hu,hd,hv

 logical undecided,separate

 common/ValvLNG/bb,cc,ta,ta2,c3,c1r,c4,c1s,vu,vd
 1,vv,pvv,voidu,voidd,hvvap,hu,hd,hhv,tjlngmax
 1,i7,ii7,L7,LL7,iop9,jlngmax,idin,kk,undecided,separate(maxvalv)
 1,i1,i2,i3,i4,i5
 1,ii7vs,i7vs
 save /ValvLNG/

&nbs
p; end module mValvLNG]

....

LOGICAL :: upst,dnst,joined,separate(kkk),previouslyseparated

....

joined=.true.

....

previouslyseparated=separate(kkk)

separate(kkk)=.false.

....

if(joined .AND. previouslyseparated)then
 [divide by zero] <<<<
else
 [use another relationship]
endif

....

You'llnotice that the highlighted LOGICAL "separate" is specified both
in the module mvalvlng and in the subroutine itself -- the accidental 
result of an unthinking Copy/Paste.

The only hint of trouble came when run-time stopped me with a divide-
by-zero at the line I've flagged with <<<<

Question 1: Would this error have been caught had I used the
old-fashioned INCLUDE-a-source-file rather than the de rigueur
USE-a-module to get my variables specified?

I suspect this would have been caught had I been able to follow Steve's
earlier advice to run the Static Verification.  But all that does is bomb
out -- without any 'scuse me's at all! -- after trying to grab more RAM
than the IA-32 system has available.

This last problem I have reported to Intel Premier Support, only to be
politely advised by Xiaoping Duan, here has already been a known
report on huge memory usage by SV leading to out of memory. The main 
reason is that we keep info in memory for all routines simultaneously.
Keeping unneeded info results in 'out of memory' error. The developers
have already fixed it and the fix will be available in future 10.1 update.
I will associate your problem report to the existing bug report and let
you know when the new compiler containing the fix is available.
[Emphasis added.]

Question 2: When will 10.1 be available? Even a Beta would be lots
better than flying blind like this!

Quest ion 3: When, oh when, are They going to fix this Message
dialog so that you can force it to return to automatic word-wrap
after pasting code??!! Shift-Return is not one of my favorite
punctuation marks!

Question 4: When are "They" going to make this Message dialog
the same width as the Thread presentation window??!! Sliding
the window from right to left is not my favorite way to read things!

Thanks again to all, and of course, God bless!

Jack OSullivan

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Mon, 27 Aug 2007 15:17:42 GMT

Unfortunately, the mistake you made is perfectly legal Fortran so the compiler could not "catch" it. Uninitialized variable checking might have found it.

I'm glad to hear that you solved the problem. (Note - it was Tim who spotted the = typo.)

This is another illustration of the confusion caused by paraphrasing the code - in almost all cases the actual error is in code you did not show.

Re: Trouble with Logical .AND. Operator

jackosullivan1 — Mon, 27 Aug 2007 15:38:57 GMT

Steve,

If this is "perfectly legal Fortran", how is it that during a debug session I would get the divide by zero when the values of joined and previouslyseparated are revealed via cursor hovering to be .TRUE. and .FALSE. respectively, and the hover-revealed value of "(joined.and.previouslyseparated)" shows as .FALSE.?

The parentheses served as a proper gate, preventing the divide-by-zero as soon as I removed separate(kkk) from the LOGICAL statement.

Thanks again, asnd God bless!

Jack

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Mon, 27 Aug 2007 16:05:45 GMT

Upon rereading your expanded code, I'm still not quite certain of what the problem was. Ignore my comment about "perfectly legal". If you have a locally declared name that is the same as a USE-associated name, the compiler will complain. If I saw the whole application, I might then understand the issue and perhaps be able to explain it.

Re: Trouble with Logical .AND. Operator

jackosullivan1 — Mon, 27 Aug 2007 20:59:27 GMT

Steve,

I have created a short version that makes the same errors:

1) It does not catch the local plus common/module specification.

2) It does not treat the "a.and.b" as false when its contents are ".true..and..false."

It's a bit long to put here -- 65 lines or so in 5 source files -- so I will submit it to Premier Support.

Thanks and God bless!

Jack

Re: Trouble with Logical .AND. Operator

Les_Neilson — Tue, 28 Aug 2007 13:37:17 GMT

Just a small item tonote : (may not be relevant in this instance but just in case it is)

In the first sample code there is if (v==0)

If "v" in the actual codeisReal and is the result of a calculation then it may not be exactly zero. Tests for equalitywith real/double precision variablesshould be avoided and a tolerance used instead.

Les

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Tue, 28 Aug 2007 15:01:44 GMT

Jack,

Thanks so much for sending the test program AND your build settings. These were critical to understanding the problems.

First, you are absolutely correct - the compiler should have complained about the duplicate declaration of "separate". It turns out that it does so in most cases, but not when the variable is declared as part of a COMMON in the module. We'll fix that.

The second issue is more interesting and requires your having set /fpscomp:logicals. In the test program, logical variable b is set to an unpredictable value because the local array separate is uninitialized. In the case in question, the value it gets (in my test run) is Z'0040105A'. Because /fpscomp:logicals was set, any non-zero value is TRUE, but the debugger is assuming the default logical semantics where even values (low bit clear) are false, so it displayed the value as .FALSE. even though the code treated it as .TRUE..

I'm not sure how we'll address this second problem - there isn't currently a mechanism to inform the debugger that /fpscomp:logicals was used. There's a couple of possibilities I've thought of and we'll try to figure out a solution for the future.

Re: Trouble with Logical .AND. Operator

jimdempseyatthecove — Tue, 28 Aug 2007 15:26:48 GMT

Jack,

I believe what you are experiencing is a problem of (feature of) name space (scope)and as a result uninitialized variables being used.

module mValvLNG
...
  logical ...,separate
! named common block   ValvLNK::separate
! not module name     mValvLNG::separate
common/ValvLNG/...
 ...,separate(maxvalv)
 ...
 save /ValvLNG/
end module mValvLNG]
...
(your subroutine/funciton)
use mValvLNG
...
! stack local (subprogram scope) separate
LOGICAL :: ...,separate(kkk),previouslyseparated
....
joined=.true.
....
! using uninitialized stack local
previouslyseparated=separate(kkk)

separate(kkk)=.false.
....
! previouslyseparated contains copy of uninitialized data
if(joined .AND. previouslyseparated)then
 [divide by zero] <<<<
else
 [use another relationship]
endif

I believe this qualifies as a nested scoping situation where it is valid to re-use the name to declare a different data area of the same name for use for different purpose.

I haven't tried it but I think if the module contained non-COMMON variables that the compiler error you sought would have occured.

Experiment byadding

logical :: separateFOO(maxvalv)

to module mValvLNG on a line by itself and without also being declared in a COMMON

Then in your program where the problem occured insert

logical :: separateFOO(kkk)

immediately following the logical that declared the local copy of separate.

When you compile I think you will see the error message that you were expecting.

Jim Dempsey

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Tue, 28 Aug 2007 15:33:47 GMT

Jim,

Fortran does not allow the sort of nested scoping you describe. As I posted above, it is a compiler bug that it does not detect the error when the use-inherited name is a COMMON variable.

Re: Trouble with Logical .AND. Operator

jimdempseyatthecove — Tue, 28 Aug 2007 15:36:21 GMT

Steve,

>> It turns out that it does so in most cases, but not when the variable is declared as part of a COMMON in the module. We'll fix that.

In this particular case it is a named COMMON block (a seperate scoping unit). I believe it should not be fixed as you would be breaking an F90 standard regarding scoping. Before you put this in for a fix check what the standards committee suggested regarding this situation.

Jim Dempsey

Re: Trouble with Logical .AND. Operator

jackosullivan1 — Tue, 28 Aug 2007 15:51:18 GMT

Jim ... and everyone else who has joined into this ...

Thanks so much for taking the time and effort to guide this poor ol' hydraulic engineer in the ways and foibles of Fortran. This site truly has the correct name: softwarecommunity!

Steve,

Thanks in particular for your speedy attention to this problem.

Perhaps while you are working on making COMMONvariables visible to the compiler's error-trapper, you might find a way to make the Go To Definition right-click function in MS Visual Studio actually do that for IVF?

Also, any idea of when 10.1 is coming out? I could sure use that Static Verification!

All,

Thanks again so much, and of course, God bless!

Jack

Re: Trouble with Logical .AND. Operator

Steven_L_Intel1 — Tue, 28 Aug 2007 16:04:26 GMT

Jack, I noticed that formatted and list-directed I/O will also assume the default interpretation and not honor /fpscomp:logicals. We're discussing this - one option may be to simply say that /fpscomp:logicals affects how logical tests are done computationally but not display of values.

Jim, my reading of the Fortran 2003 standard makes it clear that the standard does not allow this. Do you have something that shows otherwise?