On 01/05/2017 03:57, Michael Eager wrote:
I'm seeing incorrect values when there is a write to a memory-mapped I/O
device when icount is set. What I see happening is that a TB with ~20
instructions is executed which contains a write to the MM I/O address.
When it gets to the io_write routine, can_do_io is false, which results
in a call to cpu_io_recompile.
cpu_io_recompile does what it (sort of) says it is supposed to do: it
builds a new TB with the I/O instruction as the last instruction in the
block, then re-issues the TB. The problem is that the new TB contains
the instructions before the I/O instruction, so they are executed a
second time.
They shouldn't. When called from cpu_io_recompile,
cpu_restore_state_from_tb should compute the I/O instruction's target PC
from the host PC (stored in retaddr).
Then what happens is the following:
- cpu_io_recompile generates a new TB ending with the I/O instruction.
This new TB has a hash table conflict with the old TB (same
PC/cs_base/flags) the old TB is implicitly removed
- cpu_io_recompile calls cpu_loop_exit_noexc, which goes back to the
execution loop with updated PC
- because the PC is different, a new TB is looked up for the I/O
instruction's PC. The TB probably is not there and translation starts
again, this time at the I/O instruction
- the new TB, when executed, causes cpu_io_recompile to fire again.
This is the inefficient part mentioned in cpu_io_recompile
- cpu_io_recompile now compiles a one-instruction TB and goes back to
the execution loop
- finally the execution loop executes the one-instruction TB for the I/O
instruction, then it can go on