[rfa, vrp] fix eh rewrite acats cxb3011 failure

[Richard Henderson <rth at redhat dot com>]

The only remaining regression from mainline for my EH rewrite
is this odd failure of cxb3011, which looks for all the world
like a latent bug in VRP.

> (gdb) call debug_gimple_stmt ($)
> # .MEM_1028 = VDEF <.MEM_1026>
> tc_string.F[1 ...]{lb: 1 sz: 1} = VIEW_CONVERT_EXPR<character[1:5]>(*SR.131_874);

This statement triggers a vec_assert_fail here:

>       i = VEC_length (vn_reference_op_s, vr->operands) - 1;
>       j = VEC_length (vn_reference_op_s, lhs) - 1;
>       while (j >= 0 && i >= 0
> 	     && vn_reference_op_eq (VEC_index (vn_reference_op_s,
> 					       vr->operands, i),
> 				    VEC_index (vn_reference_op_s, lhs, j)))
> 	{
> 	  i--;
> 	  j--;
> 	}

with I=2, J=3 on entry to the loop.  This means we run out of 
VR->OPERANDS values before we run out of LHS values.

It looks like we'll DTRT if we just add another check for I>=0 there in 
the while loop.  That said, I don't know whether it's indicative of a 
bug elsewhere that this case occurs.

And while I'm at it, there's a memory leak just below where we fail to 
free the LHS vector.

Ok?


r~

Attachment: d-cxb3011
Description: Text document