https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89544
Bug ID: 89544
Summary: Argument marshalling incorrectly assumes stack slots
are naturally aligned.
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: rearnsha at gcc dot gnu.org
CC: bernd.edlinger at hotmail dot de, ebotcazou at gcc dot
gnu.org
Target Milestone: ---
Target: arm
The AAPCS requires that an object which is artificially overaligned is passed
at either its natural alignment (if <= 8) or on an 8-byte boundary if it is
more aligned than that. So a struct of the form
struct s
{ int a; int b;} __attribute__((aligned(8));
has a natural alignment of 4 and must be passed by value with 4-byte alignment.
The middle end code ignores this and assign_parm_find_stack_rtl ends up
generating invalid rtl of the form
(mem/c:DI (plus:SI (reg/f:SI 104 virtual-incoming-args)
(const_int 4 [0x4])) [1 f+0 S8 A32])
ie a DImode object with 32-bit alignment. It then proceeds to ignore the under
alignment when expanding to RTL and incorrectly calls gen_movdi directly,
rather than falling back to the misaligned move code.
Test case:
struct s {
int a, b;
} __attribute__((aligned(8)));
struct s f0;
int f(int a, int b, int c, int d, int e, struct s f)
{
f0 = f;
return __alignof(f);
}
When compiled with "-O -march=armv5te" generates:
@ args = 12, pretend = 0, frame = 0
@ frame_needed = 0, uses_anonymous_args = 0
@ link register save eliminated.
ldrd r0, [sp, #4] <========== Invalid alignment for ldrd
ldr r3, .L2
strd r0, [r3]
mov r0, #8
bx lr
