Description
Compiling the following test case on SystemZ:
declare i64 @personality(i64, i64, i64, ptr, ptr)
declare void @callee(i64, i64, i64, i64, i64)
declare void @panic()
define void @test() uwtable personality ptr @personality {
start:
br label %bb1
bb1:
call void @callee(i64 0, i64 0, i64 0, i64 0, i64 0)
invoke void @panic()
to label %bb2 unwind label %bb3
bb2:
ret void
bb3:
%lp = landingpad { ptr, i32 }
catch ptr null
br label %bb1
}
results in:
stmg %r6, %r15, 48(%r15)
aghi %r15, -160
lghi %r6, 0
.LBB0_1: # %bb1
# =>This Inner Loop Header: Depth=1
lghi %r2, 0
lghi %r3, 0
lghi %r4, 0
lghi %r5, 0
brasl %r14, callee@PLT
.Ltmp0:
brasl %r14, panic@PLT
.Ltmp1:
j .LBB0_3
.LBB0_2: # %bb3
# in Loop: Header=BB0_1 Depth=1
.Ltmp2:
j .LBB0_1
.LBB0_3: # %bb2
lmg %r6, %r15, 208(%r15)
br %r14
The problem is that the initalization of %r6 was pulled out of the loop. However, the loop includes a path where the panic routine throws, and the exception is caught at .Ltmp2. At this point the register %r6 was clobbered by unwind library code - it is used as the "exception pointer register".
Looking at MachineLICMImpl::HoistRegionPostRA, there is special code to handle "funclet entry block" by acknowledging they'll implicitly clobber some registers:
// Funclet entry blocks will clobber all registers
if (const uint32_t *Mask = BB->getBeginClobberMask(TRI))
applyBitsNotInRegMaskToRegUnitsMask(*TRI, RUClobbers, Mask);
However, this does not handle Linux-style EH landing pad blocks.
If I add the following code immediately after the above:
// EH landing pads clobber exception pointer/selector registers
if (BB->isEHPad()) {
const MachineFunction &MF = *BB->getParent();
if (MF.getFunction().hasPersonalityFn()) {
auto PersonalityFn = MF.getFunction().getPersonalityFn();
const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
if (unsigned Reg = TLI.getExceptionPointerRegister(PersonalityFn))
for (MCRegUnitIterator RUI(Reg, TRI); RUI.isValid(); ++RUI)
RUClobbers.set(*RUI);
if (unsigned Reg = TLI.getExceptionSelectorRegister(PersonalityFn))
for (MCRegUnitIterator RUI(Reg, TRI); RUI.isValid(); ++RUI)
RUClobbers.set(*RUI);
}
}
the problem disappears. Not sure if this is the best fix - maybe this case should be considered inside getBeginClobberMask in the first place?