On Wed, 2019-01-23 at 16:54 +0000, Richard Sandiford wrote:
>
> IMO we shouldn't remove the assert. See:
>
> https://gcc.gnu.org/ml/gcc-patches/2018-08/msg01969.html
>
> and the thread leading up to it.
>
> Thanks,
> Richard
OK, I hadn't seen that thread. I didn't see any patch submitted
in response to your comment there so I created a new patch.
This patch leaves the assert in aarch64.c and changes the check
for the 'p' constraint in constain_operands, this version
fixes the pr84682-2.c test failure and causes no regressions
on aarch64 or x86.
Steve Ellcey
sell...@marvell.com
2019-01-23 Bin Cheng <bin.ch...@arm.com>
Steve Ellcey <sell...@marvell.com>
PR target/85711
* recog.c (address_operand): Return false on wrong mode for address.
(constrain_operands): Check for mode with 'p' constraint.
diff --git a/gcc/recog.c b/gcc/recog.c
index d0c498fced2..a9f584bc0dc 100644
--- a/gcc/recog.c
+++ b/gcc/recog.c
@@ -1070,6 +1070,11 @@ general_operand (rtx op, machine_mode mode)
int
address_operand (rtx op, machine_mode mode)
{
+ /* Wrong mode for an address expr. */
+ if (GET_MODE (op) != VOIDmode
+ && ! SCALAR_INT_MODE_P (GET_MODE (op)))
+ return false;
+
return memory_address_p (mode, op);
}
@@ -2696,10 +2701,13 @@ constrain_operands (int strict, alternative_mask alternatives)
/* p is used for address_operands. When we are called by
gen_reload, no one will have checked that the address is
strictly valid, i.e., that all pseudos requiring hard regs
- have gotten them. */
- if (strict <= 0
- || (strict_memory_address_p (recog_data.operand_mode[opno],
- op)))
+ have gotten them. We also want to make sure we have a
+ valid mode. */
+ if ((GET_MODE (op) == VOIDmode
+ || SCALAR_INT_MODE_P (GET_MODE (op)))
+ && (strict <= 0
+ || (strict_memory_address_p
+ (recog_data.operand_mode[opno], op))))
win = 1;
break;
