abort: Unsupported regclass to store

[llvmbot]

Bugzilla Link	9135
Resolution	WONTFIX
Resolved on	Nov 25, 2011 14:14
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@asl,@d0k

Extended Description

This was produced by a very large program and failed in a large function with nested switch statements. The test case is very large, so in the spirt of Fermat, I have omitted it. It is curious that the *.ll only aborts if it is first optimized by "opt -std-compile-opts".

When I gdb llc, I can see that SystemZInstrInfo::storeRegToStackSlot is being called with a TargetRegisterClass of CCR (i.e. the PSW).

Here is the abort backtrace:
Unsupported regclass to store
UNREACHABLE executed at SystemZInstrInfo.cpp:63!
0 llc 0x0000000000ea643f
1 llc 0x0000000000ea6f7a
2 libpthread.so.0 0x0000003f4420f3c0
3 libc.so.6 0x0000003f43a34085 gsignal + 53
4 libc.so.6 0x0000003f43a35a36 abort + 390
5 llc 0x0000000000e90d24 llvm::llvm_unreachable_internal(char const*, char const*, unsigned int) + 356
6 llc 0x00000000005f9a05
7 llc 0x0000000000b9b48c llvm::TargetInstrInfo::foldMemoryOperand(llvm::ilist_iteratorllvm::MachineInstr, llvm::SmallVectorImpl const&, int) const + 1068
8 llc 0x0000000000bf8006 llvm::LiveIntervals::tryFoldMemoryOperand(llvm::MachineInstr*&, llvm::VirtRegMap&, llvm::MachineInstr*, llvm::SlotIndex, llvm::SmallVector<unsigned int, 2u>&, bool, int, unsigned int) + 438
9 llc 0x0000000000bf840a llvm::LiveIntervals::rewriteInstructionForSpills(llvm::LiveInterval const&, llvm::VNInfo const*, bool, llvm::SlotIndex, llvm::SlotIndex, llvm::MachineInstr*, llvm::MachineInstr*, llvm::MachineInstr*, unsigned int, int, bool, bool, bool, bool, llvm::VirtRegMap&, llvm::TargetRegisterClass const*, llvm::SmallVector<int, 4u>&, llvm::MachineLoopInfo const*, unsigned int&, unsigned int, bool&, bool&, llvm::DenseMap<unsigned int, unsigned int, llvm::DenseMapInfo, llvm::DenseMapInfo >&, std::vector<llvm::LiveInterval*, std::allocatorllvm::LiveInterval* >&) + 746
10 llc 0x0000000000bf9ffc llvm::LiveIntervals::rewriteInstructionsForSpills(llvm::LiveInterval const&, bool, llvm::LiveRange const*&, llvm::MachineInstr*, llvm::MachineInstr*, unsigned int, int, bool, bool, bool, bool, llvm::VirtRegMap&, llvm::TargetRegisterClass const*, llvm::SmallVector<int, 4u>&, llvm::MachineLoopInfo const*, llvm::BitVector&, llvm::DenseMap<unsigned int, std::vector<llvm::LiveIntervals::SRInfo, std::allocatorllvm::LiveIntervals::SRInfo >, llvm::DenseMapInfo, llvm::DenseMapInfo<std::vector<llvm::LiveIntervals::SRInfo, std::allocatorllvm::LiveIntervals::SRInfo > > >&, llvm::BitVector&, llvm::DenseMap<unsigned int, std::vector<llvm::LiveIntervals::SRInfo, std::allocatorllvm::LiveIntervals::SRInfo >, llvm::DenseMapInfo, llvm::DenseMapInfo<std::vector<llvm::LiveIntervals::SRInfo, std::allocatorllvm::LiveIntervals::SRInfo > > >&, llvm::DenseMap<unsigned int, unsigned int, llvm::DenseMapInfo, llvm::DenseMapInfo >&, std::vector<llvm::LiveInterval*, std::allocatorllvm::LiveInterval* >&) + 3020
11 llc 0x0000000000bfbfa6 llvm::LiveIntervals::addIntervalsForSpills(llvm::LiveInterval const&, llvm::SmallVectorImplllvm::LiveInterval* const&, llvm::MachineLoopInfo const*, llvm::VirtRegMap&) + 3078
12 llc 0x0000000000b6a14f
13 llc 0x0000000000b155d4
14 llc 0x0000000000b18dff
15 llc 0x0000000000b1a05d
16 llc 0x0000000000ddfc97 llvm::FPPassManager::runOnFunction(llvm::Function&) + 599
17 llc 0x0000000000ddfd5b llvm::FPPassManager::runOnModule(llvm::Module&) + 75
18 llc 0x0000000000ddf827 llvm::MPPassManager::runOnModule(llvm::Module&) + 551
19 llc 0x0000000000ddf977 llvm::PassManagerImpl::run(llvm::Module&) + 167
20 llc 0x000000000053ea3d main + 4701
21 libc.so.6 0x0000003f43a1ee7d __libc_start_main + 253
22 llc 0x000000000053c9a9
Stack dump:
0. Program arguments: /home/bgl/work/llvm/Release+Asserts/bin/llc -march=systemz main-s390x.bc

1. Running pass 'Function Pass Manager' on module 'main-s390x.bc'.
2. Running pass 'Linear Scan Register Allocator' on function '@lex_Next'
   Aborted

I also have the output of "llc -march=systemz -debug" if that would be helpful.

[asl]

Couldn't you use the bugpoint to reduce the testcase?

[llvmbot]

I've never used bugpoint and hoped to keep it that way:-) I thought perhaps the fact that something was trying to save the PSW in the stack might ring a bell. Perhaps something as simple as marking the register class containing the PSW with a "don't do that!".

But even so, unless bug 9134 is fixed, the backend is pretty useless except for very small, toy programs.

[llvmbot]

OK, I ran bugpoint but I fear it wasn't much help:

$ opt -std-compile-opts --disable-inlining main-s390x.ll >main-s390x.bc
$ llc -march=systemz main-s390x.bc
Aborts
$ bugpoint -run-llc main-s390x.bc --tool-args -march=systemz >bugpoint.out1 2>&1 &
Runs for 6.5 hours
File sizes:
143256 main-s390x.bc
0 bugpoint.reference.out
140652 bugpoint-test-program.bc
0 bugpoint.program_error_messages
142644 bugpoint-reduced-simplified.bc
2613540 bugpoint.out1

So reduced-simplified is very much reduced. I'll keep trying to construct a minimal testcase.

[asl]

Make sure you're not having debug info inside .bc file.

[llvmbot]

There is no debug information. The .ll comes from my front-end which doesn't yet know how to produce debug info. It's just a very large program, its the front-end compiling itself.

[asl]

Ok... The bugpoint might be slow, yes.

[asl]

Reduced testcase attached. The problem is due to someone's trying to copy save/restore PSW:

244L %vreg2 = COPY %vreg24:subreg_32bit; GR32:%vreg2 GR64:%vreg24
252L %vreg21 = COPY %PSW; CCR:%vreg21
268L %vreg25 = ADD32rr %vreg25, %vreg2, %PSW<imp-def,dead>; GR32:%vreg25,%vreg2
276L %PSW = COPY %vreg21; CCR:%vreg21
292L JE <BB#8>, %PSW<imp-use,kill>
300L JMP <BB#8>

Strictly speaking, this might be a reasonable operation here, since add32rr clobbers status bits and jump uses it...

[asl]

Testcase

[d0k]

The SystemZ backend was removed.