gdb won't single-step over ARM integer divide opcode

gdb won't single-step over ARM integer divide opcode

[John Breitenbach]

I'm trying to debug a program compiled for a cortex-a53 using gcc-4.9.3 
with gdb-7.9.1, and when I attempt to single-step through an sdiv 
opcode, gdb won't step.  And it gives me a message about being unable to 
access memory:

1: x/i $pc
=> 0x8310 <main+32>:    sdiv    r2, r2, r4
Cannot access memory at address 0xffffd8f0

I did some debugging and have found that the function, 
arm_get_next_pc_raw, inappropriately decodes this opcode as a load into 
the PC register.  (bits 24..27 are 7, bit 20 is set, and the dest 
register appears to be the PC.

Binutils's logic to disassemble the sdiv/udiv opcodes has opcode & 
0x0ff0f0f0 = 0x0710f010 (with bit 21 distinguishing between udiv and sdiv)

I've come up with the following patch which makes my situation work. But 
I don't know how complete it is, as there may be other newer opcodes 
which fall into the formerly undefined instruction space.

Also, the comment "byte write to PC" around line 4930 seems wrong, as 
the check for bit 22 a few lines earlier catches that situation, and 
what's left is word writes to the PC."

John

---
  gdb/arm-tdep.c | 9 ++++++++-
  1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/gdb/arm-tdep.c b/gdb/arm-tdep.c
index 8e9552a..79da43e 100644
--- a/gdb/arm-tdep.c
+++ b/gdb/arm-tdep.c
@@ -4912,10 +4912,17 @@ arm_get_next_pc_raw (struct frame_info *frame, 
CORE_ADDR pc)
          break;
        }

+    case 0x7:
+        /* sdiv/udiv using masks from binutils/opcodes/arm-dis.c,
+          * otherwise, we think sdiv/udiv is a write to pc reg
+          */
+        if( (this_instr & 0x0fd0f0f0) == 0x0710f010)
+            break;
+        /* intentional fall-though to case 4..6 */
+
      case 0x4:
      case 0x5:        /* data transfer */
      case 0x6:
-    case 0x7:
        if (bit (this_instr, 20))
          {
            /* load */
-- 
1.9.3

Re: gdb won't single-step over ARM integer divide opcode

[Yao Qi]

On 05/08/15 19:02, John Breitenbach wrote:
> I did some debugging and have found that the function,
> arm_get_next_pc_raw, inappropriately decodes this opcode as a load into
> the PC register.  (bits 24..27 are 7, bit 20 is set, and the dest
> register appears to be the PC.
>
> Binutils's logic to disassemble the sdiv/udiv opcodes has opcode &
> 0x0ff0f0f0 = 0x0710f010 (with bit 21 distinguishing between udiv and sdiv)
>
> I've come up with the following patch which makes my situation work. But
> I don't know how complete it is, as there may be other newer opcodes
> which fall into the formerly undefined instruction space.

Hi John,
Thanks for your patch, however, I think I've fixed this problem in
patch https://sourceware.org/ml/gdb-patches/2015-06/msg00610.html
Does it work for you?  This commit will be in 7.10 release.

>
> Also, the comment "byte write to PC" around line 4930 seems wrong, as
> the check for bit 22 a few lines earlier catches that situation, and
> what's left is word writes to the PC."
>

Yeah, it is possible.  Patch is welcome :)

-- 
Yao (齐尧)

Re: gdb won't single-step over ARM integer divide opcode

[John Breitenbach]

Yao,
   Yes, your patch does work for me.  Sorry for not noticing your patch
earlier.

John

On 08/09/2015 07:39 AM, Yao Qi wrote:
> On 05/08/15 19:02, John Breitenbach wrote:
>> I did some debugging and have found that the function,
>> arm_get_next_pc_raw, inappropriately decodes this opcode as a load into
>> the PC register.  (bits 24..27 are 7, bit 20 is set, and the dest
>> register appears to be the PC.
>>
>> Binutils's logic to disassemble the sdiv/udiv opcodes has opcode &
>> 0x0ff0f0f0 = 0x0710f010 (with bit 21 distinguishing between udiv and 
>> sdiv)
>>
>> I've come up with the following patch which makes my situation work. But
>> I don't know how complete it is, as there may be other newer opcodes
>> which fall into the formerly undefined instruction space.
>
> Hi John,
> Thanks for your patch, however, I think I've fixed this problem in
> patch https://sourceware.org/ml/gdb-patches/2015-06/msg00610.html
> Does it work for you?  This commit will be in 7.10 release.
>
>>
>> Also, the comment "byte write to PC" around line 4930 seems wrong, as
>> the check for bit 22 a few lines earlier catches that situation, and
>> what's left is word writes to the PC."
>>
>
> Yeah, it is possible.  Patch is welcome :)
>