Re: [PATCH] Fix PR target/33923

[Jim Wilson <wilson at specifix dot com>]

Eric Botcazou wrote:
> 	PR target/33923
> 	* config/ia64/ia64.h (CANNOT_CHANGE_MODE_CLASS): Forbid mode changes
> 	for registers in BR_REGS class.

I found this a bit puzzling.  HARD_REGNO_MODE_OK says that BR_REGS can 
only hold DImode, so it seems strange that we need 
CANNOT_CHANGE_MODE_CLASS support here.  I tried debugging this myself.

The underlying problem here is that an SImode register got tied to a 
DImode register, and then the DImode register got allocated to BR_REGS, 
thus causing the invalid SImode BR_REGS reference.  This makes sense. 
The MODES_TIEABLE_P docs clearly say that MODES_TIEABLE_P must return 
false if there is any register class that accepts one mode but not the 
other.  BR_REGS accepts DImode but not SImode, so MODES_TIEABLE_P should 
return false in this case.  I did verify that this fixes the problem. 
However, the docs also say that returning false for integer mode 
combinations is a very bad idea, and that you should try an alternative 
solution if possible.  Which is what your patch does by using 
CANNOT_CHANGE_MODE_CLASS.

The only minor criticism I have is that I'd like the comment to mention 
the MODES_TIEABLE_P connection, to make it clear why the BR_REGS support 
in CANNOT_CHANGE_MODE_CLASS is necessary.  Otherwise the patch doesn't 
make much sense to me.  The patch is OK with that change.

In theory, we may have the same problem with AR_REGS, the application 
registers.  These registers also only allow DImode.  However, the 
register allocator shouldn't be using those, so it may not be a real 
problem.  I'm not going to worry about this.  If it pops up, we can deal 
with it then.
--
Jim Wilson, GNU Tools Support, http://www.specifix.com