Re: ARM floating point regs regression

Re: ARM floating point regs regression

[Andrew Cagney]

> Program received signal SIGSEGV, Segmentation fault.
> floatformat_is_nan (fmt=0x0, val=0xefbfcc14 "")
>     at /nfs/sun18//work/rearnsha/gnusrc/src/gdb/src/gdb/doublest.c:457
> 457       if (! fmt->exp_nan)
> (top-gdb) p fmt
> $1 = (struct floatformat *) 0x0
> (top-gdb) where
> #0  floatformat_is_nan (fmt=0x0, val=0xefbfcc14 "")
>     at /nfs/sun18//work/rearnsha/gnusrc/src/gdb/src/gdb/doublest.c:457
> #1  0x0007ba4c in c_val_print (type=0x1c0d80, valaddr=0xefbfcc14 "", 
>     embedded_offset=0, address=0, stream=0x1ae000, format=0, deref_ref=1, 
>     recurse=0, pretty=Val_no_prettyprint)
>     at /nfs/sun18//work/rearnsha/gnusrc/src/gdb/src/gdb/c-valprint.c:439
> #2  0x00016400 in val_print (type=0x1c0d80, valaddr=0xefbfcc14 "", 
>     embedded_offset=0, address=0, stream=0x1ae000, format=0, deref_ref=1, 
>     recurse=0, pretty=Val_pretty_default)
> ...
> 
> Breakpoint 3, print_floating (valaddr=0xefbfcc14 "", type=0x1c0d80, 
> stream=0x0)
>     at /nfs/sun18//work/rearnsha/gnusrc/src/gdb/src/gdb/valprint.c:551
> 551       const struct floatformat *fmt = floatformat_from_type (type);
> (top-gdb) p type
> $2 = (struct type *) 0x1c0d80
> (top-gdb) p *type
> $3 = {code = TYPE_CODE_FLT, 
>   name = 0x527c8 "builtin_type_arm_ext_littlebyte_bigword", tag_name = 
> 0x0,
>   length = 12, upper_bound_type = 0, lower_bound_type = 0, objfile = 0x0, 
>   target_type = 0x0, pointer_type = 0x0, reference_type = 0x0, 
>   cv_type = 0x1c0d80, as_type = 0x1c0d80, flags = 0, nfields = 0, 
>   fields = 0x0, vptr_basetype = 0x0, vptr_fieldno = -1, type_specific = {
>     arg_types = 0x0, cplus_stuff = 0x0, floatformat = 0x0}}


Hmm, type->type_specific.floatformat shouldn't be NULL here.  Checking 
gdbtypes.c:

   builtin_type_i387_ext =
     init_type (TYPE_CODE_FLT, floatformat_i387_ext.totalsize / 8,
	       0, "builtin_type_i387_ext", NULL);
   TYPE_FLOATFORMAT (builtin_type_i387_ext) = &floatformat_i387_ext;

which is good, but:

   builtin_type_arm_ext_littlebyte_bigword =
     init_type (TYPE_CODE_FLT, 
floatformat_arm_ext_littlebyte_bigword.totalsize / 8,
	       0, "builtin_type_arm_ext_littlebyte_bigword", NULL);

is bad.

I don't know how but almost all the lines initializing TYPE_FLOATFORMAT 
were lost.

Andrew

Re: ARM floating point regs regression

[Richard Earnshaw]

> Hmm, type->type_specific.floatformat shouldn't be NULL here.  Checking 
> gdbtypes.c:
> 
>    builtin_type_i387_ext =
>      init_type (TYPE_CODE_FLT, floatformat_i387_ext.totalsize / 8,
> 	       0, "builtin_type_i387_ext", NULL);
>    TYPE_FLOATFORMAT (builtin_type_i387_ext) = &floatformat_i387_ext;
> 
> which is good, but:
> 
>    builtin_type_arm_ext_littlebyte_bigword =
>      init_type (TYPE_CODE_FLT, 
> floatformat_arm_ext_littlebyte_bigword.totalsize / 8,
> 	       0, "builtin_type_arm_ext_littlebyte_bigword", NULL);
> 
> is bad.
> 
> I don't know how but almost all the lines initializing TYPE_FLOATFORMAT 
> were lost.
> 

Hmm, I think I understand why it is failing.

Although the FPA has an extended float format (which is used for stacking 
regs across function calls), this type isn't used for "long double".  So I 
suspect the initialization code in gdbtypes.c is not setting it up 
correctly for this case.

R.

Re: ARM floating point regs regression

[Andrew Cagney]

[-- Attachment #1: Type: text/plain, Size: 1054 bytes --]

> I don't know how but almost all the lines initializing TYPE_FLOATFORMAT 
>> were lost.
>> 


(try the attached, if it works, I'll add a few extra assertions so that the coredump can't occure).


> 
> 
> Hmm, I think I understand why it is failing.
> 
> Although the FPA has an extended float format (which is used for stacking 
> regs across function calls), this type isn't used for "long double".  So I 
> suspect the initialization code in gdbtypes.c is not setting it up 
> correctly for this case.


It shouldn't have got that far.  the FPA extended float should have 
TYPE_FLOATFORMAT set so that GDB can correctly and directly manipulate 
the floating point value - no conversions being needed.  The above 
should stop this.  Only as a last resort should gdb be trying to map a 
floating point type onto one of float, double or long double.

BTW, it didn't work this way in the past.  In the ``good old days'' GDB 
would convert hardware floating point formats into float, double, or 
long double (no matter how lossy that process was).

Andrew

[-- Attachment #2: diffs --]
[-- Type: text/plain, Size: 4742 bytes --]

2002-01-22  Andrew Cagney  <ac131313@redhat.com>

	* gdbtypes.c (_initialize_gdbtypes): Initialize TYPE_FLOATFORMAT
	for builtin_type_ieee_single_little, builtin_type_ieee_double_big,
	builtin_type_ieee_double_little,
	builtin_type_ieee_double_littlebyte_bigword,
	builtin_type_m68881_ext, builtin_type_i960_ext,
	builtin_type_m88110_ext, builtin_type_m88110_harris_ext,
	builtin_type_arm_ext_big, builtin_type_arm_ext_littlebyte_bigword,
	builtin_type_ia64_spill_big, builtin_type_ia64_spill_little and
	builtin_type_ia64_quad_big, builtin_type_ia64_quad_little.

Index: gdbtypes.c
===================================================================
RCS file: /cvs/src/src/gdb/gdbtypes.c,v
retrieving revision 1.37
diff -p -r1.37 gdbtypes.c
*** gdbtypes.c	2002/01/20 19:42:04	1.37
--- gdbtypes.c	2002/01/22 17:54:14
*************** _initialize_gdbtypes (void)
*** 3328,3342 ****
--- 3328,3346 ----
    builtin_type_ieee_single_little =
      init_type (TYPE_CODE_FLT, floatformat_ieee_single_little.totalsize / 8,
  	       0, "builtin_type_ieee_single_little", NULL);
+   TYPE_FLOATFORMAT (builtin_type_ieee_single_little) = &floatformat_ieee_single_little;
    builtin_type_ieee_double_big =
      init_type (TYPE_CODE_FLT, floatformat_ieee_double_big.totalsize / 8,
  	       0, "builtin_type_ieee_double_big", NULL);
+   TYPE_FLOATFORMAT (builtin_type_ieee_double_big) = &floatformat_ieee_double_big;
    builtin_type_ieee_double_little =
      init_type (TYPE_CODE_FLT, floatformat_ieee_double_little.totalsize / 8,
  	       0, "builtin_type_ieee_double_little", NULL);
+   TYPE_FLOATFORMAT (builtin_type_ieee_double_little) = &floatformat_ieee_double_little;
    builtin_type_ieee_double_littlebyte_bigword =
      init_type (TYPE_CODE_FLT, floatformat_ieee_double_littlebyte_bigword.totalsize / 8,
  	       0, "builtin_type_ieee_double_littlebyte_bigword", NULL);
+   TYPE_FLOATFORMAT (builtin_type_ieee_double_littlebyte_bigword) = &floatformat_ieee_double_littlebyte_bigword;
    builtin_type_i387_ext =
      init_type (TYPE_CODE_FLT, floatformat_i387_ext.totalsize / 8,
  	       0, "builtin_type_i387_ext", NULL);
*************** _initialize_gdbtypes (void)
*** 3344,3376 ****
--- 3348,3390 ----
    builtin_type_m68881_ext =
      init_type (TYPE_CODE_FLT, floatformat_m68881_ext.totalsize / 8,
  	       0, "builtin_type_m68881_ext", NULL);
+   TYPE_FLOATFORMAT (builtin_type_m68881_ext) = &floatformat_m68881_ext;
    builtin_type_i960_ext =
      init_type (TYPE_CODE_FLT, floatformat_i960_ext.totalsize / 8,
  	       0, "builtin_type_i960_ext", NULL);
+   TYPE_FLOATFORMAT (builtin_type_i960_ext) = &floatformat_i960_ext;
    builtin_type_m88110_ext =
      init_type (TYPE_CODE_FLT, floatformat_m88110_ext.totalsize / 8,
  	       0, "builtin_type_m88110_ext", NULL);
+   TYPE_FLOATFORMAT (builtin_type_m88110_ext) = &floatformat_m88110_ext;
    builtin_type_m88110_harris_ext =
      init_type (TYPE_CODE_FLT, floatformat_m88110_harris_ext.totalsize / 8,
  	       0, "builtin_type_m88110_harris_ext", NULL);
+   TYPE_FLOATFORMAT (builtin_type_m88110_harris_ext) = &floatformat_m88110_harris_ext;
    builtin_type_arm_ext_big =
      init_type (TYPE_CODE_FLT, floatformat_arm_ext_big.totalsize / 8,
  	       0, "builtin_type_arm_ext_big", NULL);
+   TYPE_FLOATFORMAT (builtin_type_arm_ext_big) = &floatformat_arm_ext_big;
    builtin_type_arm_ext_littlebyte_bigword =
      init_type (TYPE_CODE_FLT, floatformat_arm_ext_littlebyte_bigword.totalsize / 8,
  	       0, "builtin_type_arm_ext_littlebyte_bigword", NULL);
+   TYPE_FLOATFORMAT (builtin_type_arm_ext_littlebyte_bigword) = &floatformat_arm_ext_littlebyte_bigword;
    builtin_type_ia64_spill_big =
      init_type (TYPE_CODE_FLT, floatformat_ia64_spill_big.totalsize / 8,
  	       0, "builtin_type_ia64_spill_big", NULL);
+   TYPE_FLOATFORMAT (builtin_type_ia64_spill_big) = &floatformat_ia64_spill_big;
    builtin_type_ia64_spill_little =
      init_type (TYPE_CODE_FLT, floatformat_ia64_spill_little.totalsize / 8,
  	       0, "builtin_type_ia64_spill_little", NULL);
+   TYPE_FLOATFORMAT (builtin_type_ia64_spill_little) = &floatformat_ia64_spill_little;
    builtin_type_ia64_quad_big =
      init_type (TYPE_CODE_FLT, floatformat_ia64_quad_big.totalsize / 8,
  	       0, "builtin_type_ia64_quad_big", NULL);
+   TYPE_FLOATFORMAT (builtin_type_ia64_quad_big) = &floatformat_ia64_quad_big;
    builtin_type_ia64_quad_little =
      init_type (TYPE_CODE_FLT, floatformat_ia64_quad_little.totalsize / 8,
  	       0, "builtin_type_ia64_quad_little", NULL);
+   TYPE_FLOATFORMAT (builtin_type_ia64_quad_little) = &floatformat_ia64_quad_little;
  
    add_show_from_set (
  		     add_set_cmd ("overload", no_class, var_zinteger, (char *) &overload_debug,

Re: ARM floating point regs regression

[Richard Earnshaw]

> > I don't know how but almost all the lines initializing TYPE_FLOATFORMAT 
> >> were lost.
> >> 
> 
> 
> (try the attached, if it works, I'll add a few extra assertions so that the coredump can't occure).

Yes, that's solved the coredump problem.  I've restarted a testrun to 
check that that was the only problem.  I'll post the results tomorrow (I'm 
not hanging around all evening just to see what they are like ;-)

R.

ARM floating point regs regression

[Richard Earnshaw]

Andrew,

The following regressions in the testsuite seem to be coming from this 
patch:

2002-01-20  Andrew Cagney  <ac131313@redhat.com>

        * doublest.h (floatformat_from_type): Declare.
        * doublest.c (floatformat_from_type): New function.
        (convert_typed_floating): Use.

        * valprint.c (print_floating): Replace checks for IEEE_FLOAT with
        call to function floatformat_from_type.

All these occur after an "info all-registers", which causes gdb to 
segfault:

ERROR: Couldn't send p t_char_values(0,0) to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_char_values(0,0)
ERROR: Couldn't send p t_char_values('a','b') to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_char_values('a','b')
ERROR: Couldn't send p t_char_values(char_val1,char_val2) to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_char_values(char_val1,char_val2)
ERROR: Couldn't send p t_char_values('a',char_val2) to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_char_values('a',char_val2)
ERROR: Couldn't send p t_char_values(char_val1,'b') to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_char_values(char_val1,'b')
ERROR: Couldn't send p t_short_values(0,0) to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_short_values(0,0)
ERROR: Couldn't send p t_short_values(10,-23) to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_short_values(10,-23)
ERROR: Couldn't send p t_short_values(short_val1,short_val2) to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_short_values(short_val1,short_val2)
ERROR: Couldn't send p t_short_values(10,short_val2) to GDB.
UNRESOLVED: gdb.base/callfuncs.exp: p t_short_values(10,short_val2)


(gdb) info all-reg
r0             0x0      0
r1             0x200c8774       537692020
r2             0x2      2
r3             0x200c877c       537692028
r4             0xefbfd334       -272641228
r5             0xefbfd264       -272641436
r6             0xefbfd26c       -272641428
r7             0x1      1
r8             0x0      0
r9             0x0      0
r10            0x0      0
r11            0xefbfd238       -272641480
r12            0xefbfd204       -272641532
sp             0xefbfd22c       -272641492
lr             0x1ccc   7372
pc             0x1bf8   7160

Program received signal SIGSEGV, Segmentation fault.
floatformat_is_nan (fmt=0x0, val=0xefbfcc14 "")
    at /nfs/sun18//work/rearnsha/gnusrc/src/gdb/src/gdb/doublest.c:457
457       if (! fmt->exp_nan)
(top-gdb) p fmt
$1 = (struct floatformat *) 0x0
(top-gdb) where
#0  floatformat_is_nan (fmt=0x0, val=0xefbfcc14 "")
    at /nfs/sun18//work/rearnsha/gnusrc/src/gdb/src/gdb/doublest.c:457
#1  0x0007ba4c in c_val_print (type=0x1c0d80, valaddr=0xefbfcc14 "", 
    embedded_offset=0, address=0, stream=0x1ae000, format=0, deref_ref=1, 
    recurse=0, pretty=Val_no_prettyprint)
    at /nfs/sun18//work/rearnsha/gnusrc/src/gdb/src/gdb/c-valprint.c:439
#2  0x00016400 in val_print (type=0x1c0d80, valaddr=0xefbfcc14 "", 
    embedded_offset=0, address=0, stream=0x1ae000, format=0, deref_ref=1, 
    recurse=0, pretty=Val_pretty_default)
...

Breakpoint 3, print_floating (valaddr=0xefbfcc14 "", type=0x1c0d80, 
stream=0x0)
    at /nfs/sun18//work/rearnsha/gnusrc/src/gdb/src/gdb/valprint.c:551
551       const struct floatformat *fmt = floatformat_from_type (type);
(top-gdb) p type
$2 = (struct type *) 0x1c0d80
(top-gdb) p *type
$3 = {code = TYPE_CODE_FLT, 
  name = 0x527c8 "builtin_type_arm_ext_littlebyte_bigword", tag_name = 
0x0,
  length = 12, upper_bound_type = 0, lower_bound_type = 0, objfile = 0x0, 
  target_type = 0x0, pointer_type = 0x0, reference_type = 0x0, 
  cv_type = 0x1c0d80, as_type = 0x1c0d80, flags = 0, nfields = 0, 
  fields = 0x0, vptr_basetype = 0x0, vptr_fieldno = -1, type_specific = {
    arg_types = 0x0, cplus_stuff = 0x0, floatformat = 0x0}}