Re: [PATCH][gdb/testsuite] Fix gdb.ada/float-bits.exp with -m32

[Tom de Vries via Gdb-patches <gdb-patches@sourceware.org>]

On 5/3/22 08:47, Luis Machado wrote:
> On 5/2/22 08:10, Tom de Vries wrote:
>> On 4/25/22 12:31, Luis Machado wrote:
>>> On 4/14/22 14:14, Tom de Vries via Gdb-patches wrote:
>>>> Hi,
>>>>
>>>> With test-case gdb.ada/float-bits.exp and native we get:
>>>> ...
>>>> (gdb) print 16llf#7FFFF7FF4054A56FA5B99019A5C8#^M
>>>> $9 = 5.0e+25^M
>>>> (gdb) PASS: gdb.ada/float-bits.exp: print 
>>>> 16llf#7FFFF7FF4054A56FA5B99019A5C8#
>>>> ...
>>>> but with target board unix/-m32 we have instead:
>>>> ...
>>>> (gdb) print 16llf#7FFFF7FF4054A56FA5B99019A5C8#^M
>>>> Cannot export value 2596145952482202326224873165792712 as 96-bits \
>>>>    unsigned integer (must be between 0 and 
>>>> 79228162514264337593543950335)^M
>>>> (gdb) FAIL: gdb.ada/float-bits.exp: print 
>>>> 16llf#7FFFF7FF4054A56FA5B99019A5C8#
>>>> ...
>>>>
>>>> Fix this by testing whether 16llf is supported by doing ptype 
>>>> long_long_float
>>>> which gets us either:
>>>> ...
>>>> type = <16-byte float>^M
>>>> ...
>>>> or:
>>>> ...
>>>> type = <12-byte float>^M
>>>> ...
>>>>
>>>> Tested on x86_64-linux with native and unix/-m32.
>>>
>>> Unfortunately not all targets support 128-bit long doubles. For arm 
>>> and aarch64 the compiler won't generate a 128-bit float, but a 64-bit 
>>> float, so the 16ll tests won't be meaningful.
>>>
>>
>> Right, but I'd expect those tests are skipped because 16llf_supported 
>> is 0 for 64-bit long double.
> 
> They are skipped, but the testcase still assumes some 16llf tests can be 
> executed:
> 
> gdb_test "print val_long_double" " = 5.0e\\+25"
> 
> gdb_test "print val_long_double" " = 5.0e\\+25" \
>      "print val_long_double after assignment"
> 

Ah, I see, you mean the val_long_double tests that do not use 16llf.

> The above couple tests won't work correctly, as you're trying to force a 
> 128-bit value into a 64-bit variable.
>>
>>> FAIL: gdb.ada/float-bits.exp: print val_long_double
>>> FAIL: gdb.ada/float-bits.exp: print val_long_double after assignment
>>>
>>
>> Can you show me the actual failure mode, that is, copy from gdb.log 
>> instead of gdb.sum?  I'm surprised that these fail because AFAICT, the 
>> used constant: 5.0e+25 is exactly representable in 64-bit ieee ( I 
>> used https://babbage.cs.qc.cuny.edu/ieee-754.old/decimal.html to check 
>> this ).
> 
> Sure, here it is:
> 
> print val_long_double
> $9 = 5.0000000000000002e+25
> (gdb) FAIL: gdb.ada/float-bits.exp: print val_long_double
> print val_long_double
> $10 = 5.0000000000000002e+25
> (gdb) FAIL: gdb.ada/float-bits.exp: print val_long_double after assignment
> 
>>

I see.  I can reproduce this using reproducer patch posted below, which 
gives me:
...
(gdb) print val_double_2^M
$9 = 5.0000000000000002e+25^M
(gdb) print /x val_double_2^M
$10 = 0x4544adf4b7320335^M
(gdb)
...

So, I realize now that my statement about it being exactly representable 
was incorrect, it's represented as 2^85 * 1.2924697071141058.  I was 
thrown off by the website presenting me a crisp '5.0000000000000000e+25' 
as the value, but that's just one way to print it, and glibc evidently 
uses a different rounding in printing, which is allowed AFAICT, as per 
'The low-order digit shall be rounded in an implementation-defined manner'.

So I suppose we could just accept
   " = (5.0e\+25|5.000000000000000\de\+25)"
for 64-bits long double.

>>> I wonder if it would be best to bail out as soon as we find out the 
>>> target has no support for 128-bit floats. I can write a patch for that.
>>>
>>
>> With a rewrite like this:
>> ...
>> -set 16llf_supported 0
>> +set long_double_bytes 0
>>   gdb_test_multiple "ptype long_long_float" "" {
>> -    -re -wrap "<16-byte float>" {
>> -       set 16llf_supported 1
>> -       pass $gdb_test_name
>> -    }
>> -    -re -wrap "<\\d+-byte float>" {
>> -       pass $gdb_test_name
>> +    -re -wrap "<\(\\d+\)-byte float>" {
>> +       set long_double_bytes $expect_out(1,string)
>>       }
>>   }
>>
>> +set 16llf_supported [expr $long_double_bytes >= 16]
>> ...
>> we can formulate the precondition for any test in terms number of long 
>> double bytes.
> 
> I've sent a patch to make this test a bit more generic: 
> https://sourceware.org/pipermail/gdb-patches/2022-April/188440.html
> 
> Did you see it?
> 

I did not, I've just come back from vacation and am still working 
through my inbox (and you didn't post the patch in reply to this thread, 
or cc-ed me, or updated this thread with a reference-to-post, any of 
which would mean I would have some pointer in my inbox) .  I'll take a look.

Thanks,
- Tom

>>
>> Thanks,
>> - Tom
>>
>>> I'm guessing some 16ll tests still work for 12-byte floats, right?
>>>
>>> What do you think?
> 

----------

diff --git a/gdb/testsuite/gdb.ada/float-bits.exp 
b/gdb/testsuite/gdb.ada/float-bits.exp
index 4ca8dbf88e5..937e307f8b4 100644
--- a/gdb/testsuite/gdb.ada/float-bits.exp
+++ b/gdb/testsuite/gdb.ada/float-bits.exp
@@ -42,6 +42,9 @@ gdb_test "print val_double := 16lf#bc0d83c94fb6d2ac#" 
" = -2.0e-19"

  gdb_test "print val_double" " = -2.0e-19" \
      "print val_double after assignment"

+gdb_test "print val_double_2"
+gdb_test "print /x val_double_2"
+
  set 16llf_supported 0
  gdb_test_multiple "ptype long_long_float" "" {
      -re -wrap "<16-byte float>" {
diff --git a/gdb/testsuite/gdb.ada/float-bits/prog.adb 
b/gdb/testsuite/gdb.ada/float-bits/p
rog.adb
index 0d8c18f8d47..001bf0bb3b9 100644
--- a/gdb/testsuite/gdb.ada/float-bits/prog.adb
+++ b/gdb/testsuite/gdb.ada/float-bits/prog.adb
@@ -16,6 +16,7 @@
  procedure Prog is
     Val_Float : Float := 23.0;
     Val_Double : Long_Float := -2.0e-19;
+   Val_Double_2 : Long_Float := 5.0e+25;
     Val_Long_Double : Long_Long_Float := 5.0e+25;
  begin
     null; -- BREAK