[RFC] Read pseudo registers from frame instead of regcache

From: Simon Marchi <simon.marchi@ericsson.com>

  From: Simon Marchi <simon.marchi@ericsson.com>

I would like to fix the problem described below.  The solution I
currently have would require me to change some gdbarch hooks and
therefore touch every architecture.  It will take quite some time and
it's not always trivial, so before doing this, I'd like to gather
comments to know if my approach looks right or wrong.

NOTE: Don't use you --enable-targets=all build for this, it will only
build for x86-64, i386, ARM and AVR.

The problem: Reading pseudo registers from an upper stack frame does not
work.  The raw registers needed to compose the pseudo registers are
always read from the current thread's regcache, which is effectively
frame #0's registers.

Here's an example using the x86-64 test included in this patch (with
some output remove for brevity):

  (gdb) tb break_here_asm
  (gdb) r
  (gdb) info registers rbx ebx
  rbx            0x2021222324252627  2315169217770759719
  ebx            0x24252627          606414375
  (gdb) up
  (gdb) info registers rbx ebx
  rbx            0x1011121314151617  1157726452361532951
  ebx            0x24252627          606414375

We would expect the last ebx value to be 0x14151617, half of the rbx value in
the same frame.

The gdbarch hooks

  - pseudo_register_read
  - pseudo_register_read_value
  - pseudo_register_write

receive a regcache as a parameter, as a way to get the values of the raw
registers required to compute the pseudo register value.  However, the
regcache object always contains the current register values, not the
values unwound to the frame we're interested in.  That's why ebx in
frame #1 appears to have the value it has in frame #0.  Instead, I think
the "read" hooks should take the "next frame" and read the register
values it needs from it.  In the example above,
amd64_pseudo_register_read_value would properly get the value of rbx for
frame #1 (which is at that point saved on the stack).

Similarly, the write hook would receive a frame and read/write raw
registers through that.

Another thing I'm wondering about is error handling.  In
amd64_pseudo_register_read_value, I changed functions that returned a
register_status for functions that throw errors when values are
unavailable.  I'm not sure if we should keep handling
unavailable/optimized out values in each arch hook, or let the hooks
throw and handle it in the callers.

Finally, this change affected the cooked_write_test and cooked_read_test
tests.  I had to set up a bit more fake global state to get them
to pass.  This is because reading/writing pseudo registers now involve
frames, and in particular the sentinel frame, which was not created
before.

Comments on any of this are very welcome!
---
 gdb/amd64-tdep.c                              |  60 ++---
 gdb/arm-tdep.c                                |  20 +-
 gdb/avr-tdep.c                                |  19 +-
 gdb/frame.c                                   |  83 ++++---
 gdb/gdbarch.c                                 |   8 +-
 gdb/gdbarch.h                                 |   8 +-
 gdb/gdbarch.sh                                |   4 +-
 gdb/i386-tdep.c                               | 209 ++++++------------
 gdb/i386-tdep.h                               |   4 +-
 gdb/regcache.c                                |  45 +++-
 .../gdb.arch/amd64-pseudo-unwind-asm.S        |  43 ++++
 gdb/testsuite/gdb.arch/amd64-pseudo-unwind.c  |  33 +++
 .../gdb.arch/amd64-pseudo-unwind.exp          |  81 +++++++
 gdb/value.c                                   |  16 +-
 gdb/value.h                                   |   3 +
 15 files changed, 397 insertions(+), 239 deletions(-)
 create mode 100644 gdb/testsuite/gdb.arch/amd64-pseudo-unwind-asm.S
 create mode 100644 gdb/testsuite/gdb.arch/amd64-pseudo-unwind.c
 create mode 100644 gdb/testsuite/gdb.arch/amd64-pseudo-unwind.exp

[RFC] Read pseudo registers from frame instead of regcache

Commit Message

Comments

Patch