RE: [PATCH 1/1] gdb, fortran: Fix quad floating-point type for ifort compiler.

["Willgerodt, Felix" <felix.willgerodt@intel.com>]

> -----Original Message-----
> From: Tom Tromey <tom@tromey.com>
> Sent: Freitag, 24. Februar 2023 15:11
> To: Willgerodt, Felix via Gdb-patches <gdb-patches@sourceware.org>
> Cc: Tom Tromey <tom@tromey.com>; Willgerodt, Felix
> <felix.willgerodt@intel.com>
> Subject: Re: [PATCH 1/1] gdb, fortran: Fix quad floating-point type for ifort
> compiler.
> 
> > When examining memory for this Fortran test, I see 128-bit values in
> > binary128 format, so I assume that this is just a software/compiler feature.
> > There is also decimal128 support in e.g. gcc.  I tried it only briefly,
> > we don't end up in i386_floatformat_for_type for those types, but
> > GDB prints them just fine.  I assume as they are marked as
> > DW_ATE_decimal_float instead of DW_ATE_float, but I didn't look into
> > it in detail.
> 
> > I am not aware of any other float types that exist with this length.
> 
> I wonder, then, if this patch would work?
> This way we would not have to curate a list of names.
> 
> Tom
> 
> diff --git a/gdb/i386-tdep.c b/gdb/i386-tdep.c
> index 476d17f1efd..57e50ffdd73 100644
> --- a/gdb/i386-tdep.c
> +++ b/gdb/i386-tdep.c
> @@ -8277,20 +8277,8 @@ static const struct floatformat **
>  i386_floatformat_for_type (struct gdbarch *gdbarch,
>  			   const char *name, int len)
>  {
> -  if (len == 128 && name)
> -    if (strcmp (name, "__float128") == 0
> -	|| strcmp (name, "_Float128") == 0
> -	|| strcmp (name, "complex _Float128") == 0
> -	|| strcmp (name, "complex(kind=16)") == 0
> -	|| strcmp (name, "COMPLEX(16)") == 0
> -	|| strcmp (name, "complex*32") == 0
> -	|| strcmp (name, "COMPLEX*32") == 0
> -	|| strcmp (name, "quad complex") == 0
> -	|| strcmp (name, "real(kind=16)") == 0
> -	|| strcmp (name, "real*16") == 0
> -	|| strcmp (name, "REAL*16") == 0
> -	|| strcmp (name, "REAL(16)") == 0)
> -      return floatformats_ieee_quad;
> +  if (len == 128)
> +    return floatformats_ieee_quad;
> 
>    return default_floatformat_for_type (gdbarch, name, len);
>  }


I tested this quickly and see some regressions in gdb.base/complex-.parts.exp,
gdb.base/bfp-test.exp and gdb.base/structs.exp (and a handful of others) on x86,
with multiple compilers. Also gcc on Fedora 37 and Ubuntu 22.04.

I didn't have the time to dig too deep. It seems like we return ieee_quad for
some cases where we shouldn’t.

(gdb) bt 5
#0  i386_floatformat_for_type (gdbarch=0x55555848f520, name=0x555558b5c760 "long double", 
    len=128) at sources/gdb/gdb/i386-tdep.c:8281
#1  0x0000555555cd9e7e in gdbarch_floatformat_for_type (gdbarch=0x55555848f520, 
    name=0x555558b5c760 "long double", length=128)
    at sources/gdb/gdb/gdbarch.c:1679
#2  0x0000555555f6fc9f in dwarf2_init_float_type (objfile=0x7fffe8021d00, bits=128, 
    name=0x555558b5c760 "long double", name_hint=0x555558b5c760 "long double", 
    byte_order=BFD_ENDIAN_LITTLE) at sources/gdb/gdb/dwarf2/read.c:15031
#3  0x0000555555f70543 in read_base_type (die=0x555558908f10, cu=0x7fffd8006970)
    at sources/gdb/gdb/dwarf2/read.c:15270
#4  0x0000555555f7b2be in read_type_die_1 (die=0x555558908f10, cu=0x7fffd8006970)
    at sources/gdb/gdb/dwarf2/read.c:19631

This seems to pass 128 bits for long double, where your patch did expect 80 bits.
The 80 bit length for x86 is only handled in arch-utils.c:default_floatformat_for_type.
I wonder if that could be streamlined a bit, not having to call gdbarch functions twice
to figure out the format.

I don't know why GDB passes 128 bit from the dwarf reader. I was debugging
complex-parts only. In dwarf, we see the type having 256 bits. So maybe GDB splits
it to 128 bit, as it is a complex?

Felix
Intel Deutschland GmbH
Registered Address: Am Campeon 10, 85579 Neubiberg, Germany
Tel: +49 89 99 8853-0, www.intel.de <http://www.intel.de>
Managing Directors: Christin Eisenschmid, Sharon Heck, Tiffany Doon Silva  
Chairperson of the Supervisory Board: Nicole Lau
Registered Office: Munich
Commercial Register: Amtsgericht Muenchen HRB 186928