On 11/22/13 12:44, Eric Botcazou wrote:
Hi,
this is a regression present on the mainline and 4.8 branch. emit_group_store
happily writes past the end of a packed structure, thus accessing a distinct
variable stored there. Then the scheduler swaps a couple of writes, leading
to wrong code. Fixed by preventing emit_group_store from writing past the end
of the structure.
Tested on x86_64-suse-linux, OK for the mainline and 4.8 branch?
2013-11-22 Eric Botcazou <ebotca...@adacore.com>
PR middle-end/59138
* expr.c (emit_group_store): Do not write past the end of the structure.
(store_bit_field): Fix formatting.
2013-11-22 Eric Botcazou <ebotca...@adacore.com>
* gcc.c-torture/execute/20131122-1.c: New test.
It looks like this patch was for gcc-4.8; the code has changed a little
since then.
I'm having a hard time seeing why this change was made:
/* Optimize the access just a bit. */
- if (MEM_P (dest)
- && (! SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (dest))
- || MEM_ALIGN (dest) >= GET_MODE_ALIGNMENT (mode))
- && bytepos * BITS_PER_UNIT % GET_MODE_ALIGNMENT (mode) == 0
- && bytelen == GET_MODE_SIZE (mode))
+ else if (MEM_P (dest)
+ && (!SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (dest))
+ || MEM_ALIGN (dest) >= GET_MODE_ALIGNMENT (mode))
+ && bytepos * BITS_PER_UNIT % GET_MODE_ALIGNMENT (mode) == 0
+ && bytelen == GET_MODE_SIZE (mode))
emit_move_insn (adjust_address (dest, mode, bytepos), tmps[i]);
But that may be an artifact of looking at the trunk where the code is a
bit different already.
Presumably you've verified this is still a problem on the trunk?
jeff
