| Message ID | 201106131759.p5DHxcB08717@lucas.cup.hp.com |
|---|---|
| State | New |
| Headers | show |
OK. Jason
Index: decl.c =================================================================== --- decl.c (revision 174979) +++ decl.c (working copy) @@ -3672,7 +3672,7 @@ TYPE_SIZE_UNIT (nullptr_type_node) = size_int (GET_MODE_SIZE (ptr_mode)); TYPE_UNSIGNED (nullptr_type_node) = 1; TYPE_PRECISION (nullptr_type_node) = GET_MODE_BITSIZE (ptr_mode); - SET_TYPE_MODE (nullptr_type_node, Pmode); + SET_TYPE_MODE (nullptr_type_node, ptr_mode); record_builtin_type (RID_MAX, "decltype(nullptr)", nullptr_type_node); nullptr_node = build_int_cst (nullptr_type_node, 0); }
The test g++.dg/cpp0x/nullptr21.C fails on IA64 HP-UX in 32 bit mode and in looking at the problem I tracked it down to some code in cxx_init_decl_processing that was added in r159131 to create nullptr_node: 2010-05-06 Magnus Fromreide <magfr@lysator.liu.se> Jason Merrill <jason@redhat.com> Add support for C++0x nullptr. The type for nullptr_node is nullptr_type_node and it is set up with a type, size, and alignment of ptr_mode. But the actual mode set with SET_TYPE_MODE is set to Pmode instead of ptr_mode. I don't know why, it may just be a typo. This causes problems on IA64 HP-UX in 32 bit mode because ptr_mode is 4 bytes (and 4 byte aligned) but Pmode is 8 bytes and 8 byte aligned. The code generated for nullptr21.C winds up doing an 8 byte load for nullptr on an address that is only 4 btye aligned and the code seg faults with a misaligned read. This patch changes Pmode to ptr_mode in the SET_TYPE_MODE call and that fixes the problem. It also caused no regressions on IA64 HP-UX, IA64 Linux, or x86 Linux. OK for checkin? Steve Ellcey sje@cup.hp.com 2011-06-13 Steve Ellcey <sje@cup.hp.com> * decl.c (cxx_init_decl_processing): Use ptr_mode instead of Pmode.