[Flang] reads d* prefixed strings as floating-point values

[marikurz-amd]

Trying to read a (non-real) string into a real variable unexpectedly succeeds instead of throwing an error if both

• the string starts with d or e,
• and a floating-point format identifier without explicit width is used.

This behavior can be reproduced with

Flang Version: 21.0.0git commit 532facc78e

and the following snippet:

program reproducer
implicit none
character(len=4)  :: str = "dstr"
real              :: test
integer           :: istat
read(str,"(D)", iostat=istat) test
WRITE(*,*) "Is the string real? ", (istat .EQ. 0)
end program reproducer

which yields

Is the string real? T

Thus, flang compiles and succeeds in reading the string "dstr" into a real variable, which should not be allowed. If an explicit width is used, such as "(D4.0)", Flang throws istat != 0 as expected.

In my opinion, the expected behavior would be one of the following:

1. Flang should throw a compile-time error as no standard-conform, explicit field width in the format "(Dw.d)" is provided. This is how gfortran behaves.
2. Flang should correctly detect during runtime that the string cannot be transformed into a real in a meaningful way and provide iostat != 0. This is the behavior of Classic-Flang and the Cray compiler.

I would speculate this behavior is somehow caused by d and e are valid components of a floating-point number in exponential format.

[llvmbot]

@llvm/issue-subscribers-flang-frontend

Author: Marius Kurz (marikurz-amd)

Trying to read a (non-real) string into a `real` variable unexpectedly succeeds instead of throwing an error if both - the string starts with `d` or `e`, - _and_ a floating-point format identifier without explicit width is used.

This behavior can be reproduced with

Flang Version: 21.0.0git commit 532facc78e

and the following snippet:

program reproducer
implicit none
character(len=4)  :: str = "dstr"
real              :: test
integer           :: istat
read(str,"(D)", iostat=istat) test
WRITE(*,*) "Is the string real? ", (istat .EQ. 0)
end program reproducer

which yields

Is the string real? T

Thus, flang compiles and succeeds in reading the string "dstr" into a real variable, which should not be allowed. If an explicit width is used, such as "(D4.0)", Flang throws istat != 0 as expected.

In my opinion, the expected behavior would be one of the following:

1. Flang should throw a compile-time error as no standard-conform, explicit field width in the format "(Dw.d)" is provided. This is how gfortran behaves.
2. Flang should correctly detect during runtime that the string cannot be transformed into a real in a meaningful way and provide iostat != 0. This is the behavior of Classic-Flang and the Cray compiler.

I would speculate this behavior is somehow caused by d and e are valid components of a floating-point number in exponential format.

[klausler]

At least two compilers accept a numeric data edit descriptor with no width or digits, so I do as well for maximum portability. You'll get a compilation-time warning on this reproducer with -pedantic if you want one. But no compiler except flang-new accepts a real input field that starts with an exponent letter, so I'll make that an run-time error.

[llvmbot]

@llvm/issue-subscribers-flang-runtime

Author: Marius Kurz (marikurz-amd)

Trying to read a (non-real) string into a `real` variable unexpectedly succeeds instead of throwing an error if both - the string starts with `d` or `e`, - _and_ a floating-point format identifier without explicit width is used.

This behavior can be reproduced with

Flang Version: 21.0.0git commit 532facc78e

and the following snippet:

program reproducer
implicit none
character(len=4)  :: str = "dstr"
real              :: test
integer           :: istat
read(str,"(D)", iostat=istat) test
WRITE(*,*) "Is the string real? ", (istat .EQ. 0)
end program reproducer

which yields

Is the string real? T

Thus, flang compiles and succeeds in reading the string "dstr" into a real variable, which should not be allowed. If an explicit width is used, such as "(D4.0)", Flang throws istat != 0 as expected.

In my opinion, the expected behavior would be one of the following:

1. Flang should throw a compile-time error as no standard-conform, explicit field width in the format "(Dw.d)" is provided. This is how gfortran behaves.
2. Flang should correctly detect during runtime that the string cannot be transformed into a real in a meaningful way and provide iostat != 0. This is the behavior of Classic-Flang and the Cray compiler.

I would speculate this behavior is somehow caused by d and e are valid components of a floating-point number in exponential format.