Confused by watchpoints on remote protocol

[Denio, Mike]

You can ignore my second question in the original e-mail. I was not reporting the stop correctly. GDB behavior was as expected once I fixed that.

The only question I have is that when I type:
"watch *(uint32_t *)(0x80000018)"
I get a hardware watch point.

And when I type:
"watch (uint32_t)ram_table"
I get a software watch point.

Anyone know why? And how to change it to be always hardware?

Thanks,
Mike

From: Denio, Mike <>
Sent: Saturday, August 06, 2022 3:43 PM
To: 'gdb@sourceware.org' <gdb@sourceware.org>
Subject: Confused by watchpoints on remote protocol

In GDB 11.2 for RISCV, I am trying to support hardware watchpoints. I think I'm close, but I have 2 questions:

1. I noticed that whether or not GDB will even attempt to use a hardware watch point depends on the how the command it entered. For example:

(gdb)  watch *(uint32_t *)(0x80000018)
Hardware watchpoint 1: *(uint32_t *)(0x80000018)
(gdb)  watch (uint32_t)ram_table
Watchpoint 2: (uint32_t)ram_table
(gdb) i b
Num     Type           Disp Enb Address    What
1       hw watchpoint  keep y              *(uint32_t *)(0x80000018)
2       watchpoint     keep y              (uint32_t)ram_table

When I enter is as "watch *(uint32_t *)(0x80000018)", a hardware watchpoint is used, but when I enter it as
"watch (uint32_t)ram_table", a soft watchpoint is used.

I don't understand this. How can I tell GDB to always use a hardware watchpoint?

Also, FYI, the order doesn't matter (it's not as if GDB assumes only 1 hardware watchpoint). Below we see GDB immediately uses a software watchpoint:

(gdb) watch (uint32_t)ram_table
Watchpoint 1: (uint32_t)ram_table
(gdb) watch *(uint32_t *)(0x80000018)
Hardware watchpoint 2: *(uint32_t *)(0x80000018)
(gdb) i b
Num     Type           Disp Enb Address    What
1       watchpoint     keep y              (uint32_t)ram_table
2       hw watchpoint  keep y              *(uint32_t *)(0x80000018)

Now that I think about it, how do I tell GDB how many hardware watchpoints I support?

2. Also a second related question. I noticed that unlike breakpoints, GDB will not remove, single step past, and then replace hardware watchpoints. Is it always assumed that hardware watchpoints break AFTER the operation in question? It makes sense to break after the operation for writes since the user wants to watch a value change, but it wasn't obvious to me for read and access watchpoints. Should all of them break on the following instruction?


Thanks in advance for any help. Please keep my e-mail on the reply as I am not subscribed.

Mike