https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69838
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2016-02-19
CC| |jakub at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I think the regression is since r199754 when LRA got enabled on s390x.
Anyway, this is reproduceable even in C++ (if one looks at the RTL dumps, not
as an executable testcase) with -O2 -fexceptions -fnon-call-exceptions:
int
foo ()
{
try
{
*(int *)0 = 1;
}
catch (...)
{
return 1;
}
return 0;
}
The difference between reload and LRA is that reload reloads:
(insn 7 2 30 2 (set (mem:SI (const_int 0 [0]) [2 MEM[(int *)0B]+0 S4 A64])
(const_int 1 [0x1])) a.c:6 67 {*movsi_zarch}
(expr_list:REG_EH_REGION (const_int 1 [0x1])
(nil)))
as:
(insn 62 2 7 2 (set (reg:SI 1 %r1)
(const_int 1 [0x1])) a.c:6 67 {*movsi_zarch}
(nil))
(insn 7 62 30 2 (set (mem:SI (const_int 0 [0]) [2 MEM[(int *)0B]+0 S4 A64])
(reg:SI 1 %r1)) a.c:6 67 {*movsi_zarch}
(expr_list:REG_EH_REGION (const_int 1 [0x1])
(nil)))
i.e. the memory store is the original instruction and preserves the
REG_EH_REGION note on it, while lra reloads it as:
(insn 7 2 61 2 (set (reg:SI 1 %r1 [50])
(const_int 1 [0x1])) a.c:6 67 {*movsi_zarch}
(nil))
(insn 61 7 3 2 (set (mem:SI (const_int 0 [0]) [2 MEM[(int *)0B]+0 S4 A64])
(reg:SI 1 %r1 [50])) a.c:6 67 {*movsi_zarch}
(expr_list:REG_DEAD (reg:SI 1 %r1 [50])
(nil)))
where the original insn doesn't contain the MEM anymore, and the new one does.
Perhaps LRA should check if the insn it is reloading contains REG_EH_REGION
note, and if yes, for all the reloading instructions it has added for the insn
check if they might trap and if yes, copy (or move) the REG_EH_REGION note(s)
to them.
