[PATCH] [AArch64] Mark single precision pseudo registers unavailable if invalid

[PATCH] [AArch64] Mark single precision pseudo registers unavailable if invalid

[Pierre Langlois]

Hi all,

I noticed two failure in gdb.trace/mi-trace-frame-collected.exp:

FAIL: gdb.trace/mi-trace-frame-collected.exp: live:
  -trace-frame-collected (register)
FAIL: gdb.trace/mi-trace-frame-collected.exp: tfile:
  -trace-frame-collected (register)

In these cases, we are not collecting registers so the MI command
-trace-frame-collected should only give us the value of the PC.
However, it also gives us all of the single precision pseudo registers,
initialized with 0x0.

We can reproduce this error by simply issuing the
'maint print cooked-register' when no inferior is connected:

~~~
...
(gdb) maint print cooked-register
 Name         Nr  Rel Offset    Size  Type            Cooked value
 x0            0    0      0       8 long            <unavailable>
 x1            1    1      8       8 long            <unavailable>
 ...
 d30         130   62   1540       8 *1              <unavailable>
 d31         131   63   1548       8 *1              <unavailable>
 s0          132   64   1556       4 *1              0x00000000
 s1          133   65   1560       4 *1              0x00000000
 s2          134   66   1564       4 *1              0x00000000
 ...
 s28         160   92   1668       4 *1              0x00000000
 s29         161   93   1672       4 *1              0x00000000
 s30         162   94   1676       4 *1              0x00000000
 s31         163   95   1680       4 *1              0x00000000
 h0          164   96   1684       2 *1              <unavailable>
 h1          165   97   1686       2 *1              <unavailable>
 h2          166   98   1688       2 *1              <unavailable>
 ...
~~~

It turns out GDB does not check if S registers are valid before returning
a value for them.  It should return <unavailable> in this case.

Thanks,
Pierre

gdb/ChangeLog:

	* aarch64-tdep.c (aarch64_pseudo_read_value): Mark S register as
	unavailable if invalid.
---
 gdb/aarch64-tdep.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/gdb/aarch64-tdep.c b/gdb/aarch64-tdep.c
index 2cecad0..cec4d3e 100644
--- a/gdb/aarch64-tdep.c
+++ b/gdb/aarch64-tdep.c
@@ -2482,7 +2482,11 @@ aarch64_pseudo_read_value (struct gdbarch *gdbarch,
 
       v_regnum = AARCH64_V0_REGNUM + regnum - AARCH64_S0_REGNUM;
       status = regcache_raw_read (regcache, v_regnum, reg_buf);
-      memcpy (buf, reg_buf, S_REGISTER_SIZE);
+      if (status != REG_VALID)
+	mark_value_bytes_unavailable (result_value, 0,
+				      TYPE_LENGTH (value_type (result_value)));
+      else
+	memcpy (buf, reg_buf, S_REGISTER_SIZE);
       return result_value;
     }
 
-- 
2.1.0

Re: [PATCH] [AArch64] Mark single precision pseudo registers unavailable if invalid

[Yao Qi]

Pierre Langlois <pierre.langlois@arm.com> writes:

> It turns out GDB does not check if S registers are valid before returning
> a value for them.  It should return <unavailable> in this case.

Yes, your fix is correct, however ...

>
> Thanks,
> Pierre
>
> gdb/ChangeLog:
>
> 	* aarch64-tdep.c (aarch64_pseudo_read_value): Mark S register as
> 	unavailable if invalid.

... when I read your patch, I am wondering why does aarch64 implement
gdbarch method pseudo_register_read_value rather than
pseudo_register_read.  If we implement the pseudo_register_read, the
caller will mark the value unavailable according to its return value.
pseudo_register_read_value was added to handle partially available
registers by https://sourceware.org/ml/gdb-patches/2011-07/msg00351.html
but I don't think of a case that some aarch64 register is partially
available.  Maybe, another fix to this problem is to implement
pseudo_register_read instead of pseudo_register_read_value.

-- 
Yao (齐尧)

Re: [PATCH] [AArch64] Mark single precision pseudo registers unavailable if invalid

[Yao Qi]

On 15/07/15 15:00, Yao Qi wrote:
> ... when I read your patch, I am wondering why does aarch64 implement
> gdbarch method pseudo_register_read_value rather than
> pseudo_register_read.  If we implement the pseudo_register_read, the
> caller will mark the value unavailable according to its return value.
> pseudo_register_read_value was added to handle partially available
> registers byhttps://sourceware.org/ml/gdb-patches/2011-07/msg00351.html
> but I don't think of a case that some aarch64 register is partially
> available.  Maybe, another fix to this problem is to implement
> pseudo_register_read instead of pseudo_register_read_value.

To be clear, your patch can be pushed in, as it is correct and fixes
fails in tests.  We can discuss and implement pseudo_register_read
rather than pseudo_register_read_value for aarch64 later in the
follow-up patch.

-- 
Yao (齐尧)

Re: [PATCH] [AArch64] Mark single precision pseudo registers unavailable if invalid

[Pierre Langlois]

On 15/07/15 15:41, Yao Qi wrote:
> On 15/07/15 15:00, Yao Qi wrote:
>> ... when I read your patch, I am wondering why does aarch64 implement
>> gdbarch method pseudo_register_read_value rather than
>> pseudo_register_read.  If we implement the pseudo_register_read, the
>> caller will mark the value unavailable according to its return value.
>> pseudo_register_read_value was added to handle partially available
>> registers byhttps://sourceware.org/ml/gdb-patches/2011-07/msg00351.html
>> but I don't think of a case that some aarch64 register is partially
>> available.  Maybe, another fix to this problem is to implement
>> pseudo_register_read instead of pseudo_register_read_value.

Yes, it makes sense to me.  I don't think we should have partially
available V registers in any cases.  For example, reading the S0 register
just ignores the top-level bits of the Q0 register, but they are still
available to read through Q0.  And writing to S0 will clear the top-level
bits of Q0.  As it is implemented in aarch64_pseudo_write.

> 
> To be clear, your patch can be pushed in, as it is correct and fixes
> fails in tests.  We can discuss and implement pseudo_register_read
> rather than pseudo_register_read_value for aarch64 later in the
> follow-up patch.
> 

OK, I'll push it in and make sure pseudo_register_read works as expected in
a follow-up patch.

Thanks,
Pierre