https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89228
Bug ID: 89228
Summary: ARM unwinder fails to restore vfp reg $d8
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libbacktrace
Assignee: unassigned at gcc dot gnu.org
Reporter: bernd.edlinger at hotmail dot de
CC: ian at gcc dot gnu.org
Target Milestone: ---
This makes gdb enter an endless Loop, which makes the gcc test Suite hang,
becaue the gdb process does not terminate, and the Time-out only kills
the debuggee.
A debug session is in gdb bugzilla at
https://sourceware.org/bugzilla/show_bug.cgi?id=24178
It is a bit weird, but I think the reason for this is
that __gxx_personality_v0 is saving d8 on the stack,
and restores the old d8 on return.
0036f864 <__gxx_personality_v0>:
36f864: e200c003 and ip, r0, #3
36f868: e92d4ff0 push {r4, r5, r6, r7, r8, r9, sl, fp, lr}
36f86c: e35c0001 cmp ip, #1
36f870: ed2d8b02 vpush {d8}
36f874: e3a03000 mov r3, #0
36f878: e24dd04c sub sp, sp, #76 ; 0x4c
36f87c: e1a07002 mov r7, r2
36f880: e58d100c str r1, [sp, #12]
36f884: e58d3020 str r3, [sp, #32]
36f888: 0a000082 beq 36fa98 <__gxx_personality_v0+0x234>
36f88c: e35c0002 cmp ip, #2
36f890: 1a000008 bne 36f8b8 <__gxx_personality_v0+0x54>
36f894: e59d000c ldr r0, [sp, #12]
36f898: e1a01007 mov r1, r7
36f89c: eb005e60 bl 387224 <__gnu_unwind_frame>
36f8a0: e3500000 cmp r0, #0
36f8a4: 13a00009 movne r0, #9
36f8a8: 03a00008 moveq r0, #8
36f8ac: e28dd04c add sp, sp, #76 ; 0x4c
36f8b0: ecbd8b02 vpop {d8}
36f8b4: e8bd8ff0 pop {r4, r5, r6, r7, r8, r9, sl, fp, pc}
But __gnu_unwind_frame restores the live vfp registers in violation of the ABI,
which would work, if there was a a way to avoid using vfp Registers in
the personality function.
This function is called from here:
static void __attribute__((noreturn))
unwind_phase2 (_Unwind_Control_Block * ucbp, phase2_vrs * vrs)
{
_Unwind_Reason_Code pr_result;
do
{
/* Find the entry for this routine. */
if (get_eit_entry (ucbp, VRS_PC(vrs)) != _URC_OK)
abort ();
UCB_SAVED_CALLSITE_ADDR (ucbp) = VRS_PC(vrs);
/* Call the pr to decide what to do. */
pr_result = ((personality_routine) UCB_PR_ADDR (ucbp))
(_US_UNWIND_FRAME_STARTING, ucbp, (_Unwind_Context *) vrs);
}
while (pr_result == _URC_CONTINUE_UNWIND);
if (pr_result != _URC_INSTALL_CONTEXT)
abort();
uw_restore_core_regs (vrs, &vrs->core);
}
my conclusion so far is:
The last uw_restore_core_regs restores only r0..r15, but it should
also restore vfp regs on demand. Note that phase2_vrs does not hold
enough space for the vfp Registers.
