This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
1.0 INTRODUCTION
This describes changes to MS-FORTRAN Version 3.13 that
were made too late to be described in either the reference
manual or the User's Guide and its addendum. It also
outlines differences between versions 3.13 and 3.10.
2.0 DIFFERENCE FROM VERSION 3.13
2.1 Enhancements
The following changes are made to the way the compiler
and runtime behave:
a. The default setting of the 8087 control word enables
NAN, divide-by-zero and infinity exceptions. These will now
give a runtime error. Use LCWRQQ if you preferred the 3.10
behavior.
b. Most incorrect format strings in FORMAT, READ and WRITE
statements now give warning messages.
c. The compiler front end is smaller and, on systems with
limited memory, should compile larger programs.
d. NULR7.OBJ is no longer required, since floating point
code is only included with your program if it is actually
needed.
e. The compiler sets an exit status to MSDOS 2.0.
2.2 Corrections
The following errors in version 3.10 have been
corrected:
Incorrect code was being generated for comparisons of reals
in named common.
External and actual functions of the same type gave an ERROR
33
An error on the last line of a formatted input file would
'create' an extra (empty) line. Errors were also caused if
a file did not end in a <Control Z>.
List directed format choices for real numbers were not
triggered by the ranges documented.
INTEGER*4 conversions to real were sometimes incorrectly
rounded.
Page 2
G format choices were not triggered by the ranges
documented.
READ or WRITE of an adjustable array caused a compile time
error 151.
Reading a real with list directed input would not accept a
null item.
Statement function actual and formal parameter sizes were
sometimes incorrectly matched.
PAS2 looped indefinitely for c=amin0(-5,min0(0,1))
The last record output to the printer was not printed on
certain types of printer.
INTEGER*4 exponentiation was truncated to 16-bits under some
circumstances.
PAS2 sometimes gave incorrect integer overflow message.
FOR1 was not deleting intermediate files if it detected a
compile error.
Incorrect code might be generated for multiple
multi-dimensioned array references.
If the result of an expression was assigned to an INTEGER*2,
divides of INTEGER*4 values were done with 16-bit division.
DEXP(X) sometimes returned zero.
amin0(i,i) caused a PAS2 loop.
An erroneous compile time error 71 sometimes occurred.
3.0 NULR7.OBJ
This module is not supplied with this version since the
Real Math Package is only included in your program if you
actually use floating point operations.
4.0 COMPILE TIME FORMAT CHECKING.
A special warning message will be displayed by the
compiler if a format specification contains certain errors.
This has the form:
***** Warning -- Invalid Format, Error n, Line m
Page 3
Where n is one of the following runtime error codes:
1200
1204..1213
1215..1224
1226..1228
1232
and m is the line number of the offending statement.
5.0 EXIT STATUS
The compiler supplies an exit status to MSDOS 2.0 that
can be accessed via the "IF ERRORLEVEL n" batch command.
The vaules returned by the compiler are:
n Value Meaning
0 No warning or errors issued
1 Warnings messges were issued
2 Fatal errors were encountered
6.0 ALTERNATIVE LINKER
Two versions of the MS-LINK utility are provided with
this version of MS Fortran. The first, named LINK.V1 is the
most current linker for MS-DOS versions 1.25 and below. It
will run under MS-DOS 2.0 but cannot accept pathnames or
subdirectories. The other is named LINK.V2 and will run
only on MS-DOS 2.0. The limits on program size,
relocations, externals per module etc., are bigger for
LINK.V2 than for either LINK.V1 or the linker usually
supplied with MS-DOS 2.0. The interface to both linkers is
identical and is as described in the linker documentation.
You should rename the one you want to use LINK.EXE.
7.0 EXTENDED NON-DECIMAL NUMBERS
Although the maximum 32-bit integer value is defined as
2**31-1, the compiler and runtime will read greater values
which are nominally in the range upto 2**32 without giving
an error if (and only if) the radix is other than 10. They
will be interpreted as the negative numbers with the
corresponding internal representation. For example,
16#FFFFFFFF will result in all the bits in the 32-bit
integer result being set, and will have an arithmetic value
of -1.
Page 4
8.0 $MESSAGE METACOMMAND
The $MESSAGE metacommand can be used to send messages
to the standard output device when running the Fortran front
end, for example:
$MESSAGE: 'This is displayed when you run FOR1'
The message string must be 40 characters or fewer.
9.0 FLOATING POINT OPERATIONS
Most users of MS-Fortran will find the default behavior
of the floating point operations will provide extremely
accurate, consistent and efficient processing of their
algorithms, whether they have an 8087 installed or not. If
this is the case for you, you need not be concerned with the
issues described either below or in the addendum to the
User's Guide.
However, those who wish to take advantage of the full
power and flexiblity of the proposed IEEE Real Math Standard
should read both the addendum and the following description
carefully.
9.1 Math Package Size
The REAL arithmetic support routines contribute about
6.5k bytes to your program and not 4.5k as specified in the
addendum.
9.2 $FLOATCALLS And The 8087
Contrary to the description in the addendum to the
Fortran User's Guide, programs compiled with the $FLOATCALLS
option, linked with the emulator library and run on a
machine with an 8087, will use the 8087 to do the actual
arithmetic. This means that using this option will result
in a much smaller performance penalty, when you have an
8087, than suggested in the addendum.
9.3 Environment Control And Exception Handling For 8087 Math
There are five exceptions required by the IEEE
standard. The "Control Word" defines what response is made
to each exception (see below for a description of the
Control Word and how to access it). The exceptions and the
default and alternative responses are as follows:
Page 5
1. Invalid Operation - any operation involving a NAN (Not A
a number), SQRT (-1), 0*infinity etc.
Default Action - Enabled: gives runtime error 2136.
Alternate Action - Disabled: returns a NAN
2. Divide by Zero - r/0.0
Default Action - Enabled: gives runtime error 2100
Alternate Action - Disabled: returns a properly signed
INF (infinity)
3. Overflow - Operation results in a number greater
than maximum representable number.
Default Action - Enabled: gives runtime error 2101.
Alternate Action - Disabled: returns INF.
4. Underflow - Operation results in a number smaller
than minimum representable number.
Default Action - Disabled: returns zero.
Alternate Action - Enabled: gives runtime error 2135.
5. Precision Loss - When a result is subject to rounding
error, means result is not exact.
Default Action - Disabled: Returns properly rounded
result.
Alternate Action - Enabled: gives runtime error 2139
Note that if you have masked the exceptions, you may
get not-a-number values. If you write these values with
formatted i/o, you will either get asterisks, or, more
usually, the letters INF - for infinity, IND - for
indefinite, and NAN - for not-a-number will appear, usually
in the fractional part of the field.
Contrary to the description in the MS-Fortran reference
(Section 6.2.1) the $DEBUG metacommand does not control the
handling of the exceptions (you should continue to use it to
control Integer aritmetic errors, however). Instead, there
are two memory locations that control both processors.
These are called the CONTROL and STATUS words. (You can
access or change the control and status words by using one
of the procedures described below). When one of the
exceptional conditions occurs, the appropriate bit in the
status word is set. This flag will remain set to indicate
that the exception occurred until cleared by the user. If
the bit in the control word relating to a given exception is
set then that exception is masked and the operation proceeds
with a supplied default. If the bit is unset any exception
of that type generates an error message, halts the operation
and your program will stop. In either case the exception is
ORed into the STATUS word.
The CONTROL word is also used to set modes for the
internal arithmetic required by the IEEE standard. These
are:
Page 6
Rounding Control - round to nearest (or even), Up,
Down, or Chop
Precision Control - Determines at which bit of the
manstissa rounding should take place. (24, 53,
or 64). Note all results are done to 64 bits
regardless of the precision control. It only
affects the rounding in the internal form.
On storage any result is again rounded to the
storage precision.
Infinity Control - Affine mode is the familar + and
- INF style of arithmetic. Projective mode is a
mode where + and - INF are considered to be the
same number. The principal effect is to change the
nature of comparisons (Projective INF does not
compare with anything but itself).
Format for STATUS BYTE and CONTROL WORD
15 8 7 6 5 4 3 2 1 0
STATUS | hi byte unused | | |PE|UE|OE|ZE| |IE|
| | | | |
Precision Exception -------------------+ | | | |
Underflow Exception ----------------------+ | | |
Overflow Exception --------------------------+ | |
Zero Divide Exception --------------------------+ |
Invalid Exception ------------------------------------+
(All other bits unused, may be either 1 or 0)
15 14 15 12 11-10 9-8 7 6 5 4 3 2 1 0
CONTROL | | | |IC| RC | PC | | |PM|UM|OM|ZM| |IM|
| | | | | | | |
Infinity Control -+ | | | | | | |
Round Control --------+ | | | | | |
Precision Control ----------+ | | | | |
| | | | |
Precision Mask ------------------------+ | | | |
Underflow Mask ---------------------------+ | | |
Overflow Mask -------------------------------+ | |
Zero Divide Mask -------------------------------+ |
Invalid Mask -----------------------------------------+
(All other bits unused, may be either 1 or 0)
Infinity Control
0 = Projective
1 = Affine
Round Control
00 = Round nearest or even
01 = Round down (toward -INF)
Page 7
10 = Round up (toward +INF)
11 = Chop (Truncate toward 0)
Precision Control
00 = 24 bits of mantissa
01 = (reserved)
10 = 53 bits of mantissa
11 = 64 bits of mantissa
You can read or set the value of the control word and
read the status word using the following subroutines:
C Load Control Word
SUBROUTINE LCWRQQ (CW)
INTEGER*2 CW
C Sets the control word to the value in CW
C Store Control Word
INTEGER*2 FUNCTION SCWRQQ
C Returns the value of the control word
C Store Status Word
INTEGER*2 FUNCTION SSWRQQ
C Returns the value of the status word
