[Patch, Fortran] PR34342 - BOZ diagnostic, Fortran 2003 BOZ, BOZ extensions

[Tobias Burnus]

:ADDPATCH fortran:

This patch adds:

a) Fortran 2003 BOZ, i.e. real(boz), dble(boz),
cmplx(boz,y)/cmplx(x,boz) are treated as if the BOZ had been TRANSFERed.

b) Using BOZ in DATA statements to initialize non-integer variables
gives now an error with -std=f95/f2003

c) Using BOZ in DATA statement for non-integer variables, now TRANSFERs
the bit pattern and does not convert the integer value to be compatible
with several other compilers. (Only for REAL and COMPLEX variables.)

d) Using  "real r = boz" now also TRANSFERs the bit pattern as long as
there is no expression on the right-hand side; i.e. in "real r = boz +
1" the boz is regarded as integer. (For expressions with BOZ, different
compilers behave quite differently.)

e) Improve the BOZ documentation by (1) telling more about BOZ and (2)
by making clear how gfortran extends the standard syntax.

If anyone has an idea how to reject with -f2003 BOZ of the following
type I'd be happy:
   r = z'1234' + 1.0

Build and regression tested on x86-64-linux.
OK for the trunk?

Tobias

PS: I plan to compile/test a couple of applications to make sure we do
not regress.

PPS: boz_9.f90 will fail with -fdefault-integer-8/-fdefault-real-8. Has
anyone an idea how to fix it w/o specifying a kind= parameter for int,
real, cmplx (the kind parameter causes a different code path) and
without breaking dble?

PPPS: I just saw that there is now no range check for:
  DOUBLE PRECISION inf
  DATA inf / Z'7FF0000000000000' /
Please state whether it should be added.

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: boz-fix.diff
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20071206/a2e9ff63/attachment.ksh>