Hi Jerry,
thanks for your draft patch!
I don't mind implementing it step by step, but it does not
work here as I think it should (based on two other compilers).
Let's look at the testcases first:
- real(10) should be restricted to platforms where this is
available; it will be rejected on several others
- EXformat_1.f90
r4 = -huge(1.0_4/3.0_4)
this is identical to
r4 = -huge(1.0_4)
and should print as:
-0XF.FFFFFP+124
and not
-0X1.P+127
etc.
Similarly for EXformat_2.f90, the first write should be
> -0X9.D14707B63DFBP+3<
and not
> -0X1.P+6<
I would also recommend to extend the string length s1 so that
other brands do not complain about it being to short:
character(kind=1, len=16) :: s1
I also get a warning at compilation of io.cc:
../../gcc-trunk/gcc/fortran/io.cc: In function 'format_token format_lex()':
../../gcc-trunk/gcc/fortran/io.cc:449:11: warning: this statement may
fall through [-Wimplicit-fallthrough=]
449 | switch (c)
| ^~~~~~
../../gcc-trunk/gcc/fortran/io.cc:464:9: note: here
464 | case 'Z':
| ^~~~
Can you check the logic?
I stopped the review here.
Finally a few general comments: please try to stick to the
80 columns/line recommendation for commit messages as well
as for code.
Best,
Harald
On 1/29/26 20:55, Jerry D wrote:
The attached patch is the first step in implementing the EX format
specifier.
As noted in the Change Log, I chose to hold off on the KIND=16 hex float
output for a future patch. Likewise for the corresponding READ
implementation. Unfortunately I ran into a number of places in the files
where tabs were not used for indentation. I opted to fix those now as my
editor does it on the fly.
The two new test cases pass on x86_64. Some implementations of the %A in
snprintf may normalize the hex representation differently. I am hoping
we can get some broader testing of this so I can set the { target *-*-
* } for those that pass the tests.
I intend to take care of the READ and KIND=16 implementations in my next
round.
I Would like this to get into the wild if I may.
Regression tested on x86_64-linux-gnu.
Comments welcome.
OK for mainline?
Regards,
Jerry
Author: Jerry DeLisle <[email protected]>
Date: Thu Jan 29 11:07:22 2026 -0800
Fortran: Implement EX format writing of floats
This patch implements the Fortran 2018 Standard EX formatting for
WRITE of floating point. KIND=16 is not implemented by this. READ
is not implemented.
PR fortran/93727
gcc/fortran/ChangeLog:
* io.cc (enum format_token): Add FMT_EX token.
(format_lex): Add parsing for the FMT_EX and add
checks as needed.
libgfortran/ChangeLog:
* io/format.c (format_lex): Add use of FMT_EX.
(parse_format_list): Add parsing of FMT_EX.
(next_format): Use FMT_EX in reversion check.
* io/io.h (write_ex): Add prototype for write_ex.
(internal_proto): Make it internal.
* io/transfer.c (formatted_transfer_scalar_write): Use
FMT_EX token in the main loop processing.
* io/write.c (write_default_char4): Cleanup whitespace.
(write_a): Likewise.
(write_boz): Likewise.
(write_decimal): Likewise.
(otoa_big): Likewise.
(write_character): Likewise.
(write_float_0): Likewise.
(write_ex): New function that implements the hex float write,
padding with spaces as needed. Handles kind=1 and kind=4
character output. Uses the new function
get_float_hex_string ()
defined in io/write_float.def.
(write_real): Cleanup whitespace.
(write_complex): Likewise.
(nml_write_obj): Likewise.
(namelist_write): Likewise.
* io/write_float.def (get_float_hex_string): New function
that uses the
snprintf() function with %A to create the initial hex
formmatted float
string for later re-formatting in the write_ex function.
(build_float_string): Cleanup whitespace.
(quadmath_snprintf): Likewise.
(determine_en_precision): Likewise.
gcc/testsuite/ChangeLog:
* gfortran.dg/EXformat_1.f90: New test.
* gfortran.dg/EXformat_2.f90: New test.
