https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65773
Bug ID: 65773
Summary: [5.1 regression] GCC 5.1 miscompiles LLVM function
AArch64InstrInfo::loadRegFromStackSlot()
Product: gcc
Version: 5.1.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: wschmidt at gcc dot gnu.org
CC: bergner at gcc dot gnu.org, jakub at gcc dot gnu.org,
rguenth at gcc dot gnu.org
Host: powerpc64*-linux-gnu
Target: powerpc64*-linux-gnu
Build: powerpc64*-linux-gnu
I built gcc from branches/gcc-5-branch yesterday and used it to build
Clang/LLVM as an acid test. Unfortunately the test suite shows a number of
failures. I've traced these down to a miscompile of
AArch64InstrInfo::loadRegFromStackSlot( <long argument list omitted> ). The
bad code can be seen here:
105eb74c: 00 00 60 38 li r3,0
105eb750: 28 00 41 f9 std r10,40(r1)
105eb754: e4 06 29 79 rldicr r9,r9,0,59
105eb758: 30 00 41 f9 std r10,48(r1)
105eb75c: 21 00 01 99 stb r8,33(r1)
105eb760: 22 00 21 99 stb r9,34(r1)
105eb764: a5 ec 79 48 bl 10d8a408
<_ZN4llvm12MachineInstr10addOp\
erandERNS_15MachineFunctionERKNS_14MachineOperandE+0x8>
This calls MachineInstr::addOperand(MachineFunction&, MachineOperand const&)
with r3 as the this pointer. Loading zero into the this pointer ahead of the
call seems somewhat unhelpful.
The same code compiled with the 4.8 gcc I usually use shows:
106a8fa0: 78 cb 25 7f mr r5,r25
106a8fa4: 78 f3 c3 7f mr r3,r30
106a8fa8: 20 00 9e eb ld r28,32(r30)
106a8fac: 78 eb a4 7f mr r4,r29
106a8fb0: 78 d3 46 7f mr r6,r26
106a8fb4: 65 d4 7c 48 bl 10e76418
<_ZNK4llvm12MachineInstr21getR\
egClassConstraintEjPKNS_15TargetInstrInfoEPKNS_18TargetRegisterInfoE+0x8>
which is clearly much more reasonable.
I'll gather reproduction materials, but wanted to get the bug filed quickly for
awareness.
