> Lets go with this patch and hopefully stabilize the tree.  I don't think the
> vector conversions represent an important case.

Unfortunately the patch introduces GIMPLE checking failures in Ada so it will 
need to be completed/improved.  But let's postpone it because we have another 
class of GIMPLE checking failures introduced by the useless_type_conversion_p 
change itself:

c37213j.adb:21:05: warning: variable "X" is read but never assigned
c37213j.adb: In function 'C37213J.PROC.CONSTPROP':
c37213j.adb:41:4: error: invalid conversion in gimple call
struct c37213j__proc__value___PAD

struct c37213j__proc__value___PAD

# .MEM_38 = VDEF <.MEM_37>
MEM[(struct c37213j__proc__value___PAD *)R.12_25] = c37213j.proc.value (); 
[static-chain: &FRAME.39] [return slot optimization]

and:

eric@polaris:~/build/gcc/native> ~/install/gnat-head/bin/gcc -S c37213j.adb -
O2
c37213j.adb:21:05: warning: variable "X" is read but never assigned
c37213j.adb: In function 'C37213J.PROC.VALUE':
c37213j.adb:26:5: error: invalid conversion in return statement
struct c37213j__proc__value___PAD

struct c37213j__proc__value___PAD

# VUSE <.MEM_11>
return _9(D);

What happens here is that GIMPLE statements are remapped through cloning and 
we have a variably-modified type returned by a nested function, so the type of 
the LHS of a GIMPLE_CALL or that of the RHS of a GIMPLE_RETURN is remapped but 
of course not the return type of the function.  This used to be OK because 
remapping is done by means of copy_node and preserves TYPE_CANONICAL, so the 
conversion between remapped and original type was deemed useless; now the 
TYPE_CANONICAL check is gone so the conversion is not useless anymore...

I don't think that we want to introduce an artificial VCE to fix this so we 
probably need a couple of kludges in the GIMPLE verifier instead.

In any case, the more I look into the fallout of the useless_type_conversion_p 
change, the more I find it ill-advised.  We used to have a solid type system 
in the middle-end by means of the predicate and now we have cases for which it 
ought to return false and returns true (e.g. non-structurally equivalent types 
with different calling conventions) and cases for which it can return true and 
returns false (remapped types or types deemed equivalent by the languages).
I don't really know what it was made for, but there must be a better way...


	* gnat.dg/discr45.adb: New test.

-- 
Eric Botcazou