Here's the list of failing tests that I got:

Failing Tests (19):
LLVM :: CodeGen/AArch64/aarch64_f16_be.ll
LLVM :: CodeGen/AArch64/and-mask-removal.ll
LLVM :: CodeGen/ARM/2010-05-18-LocalAllocCrash.ll
LLVM :: CodeGen/ARM/Windows/alloca.ll
LLVM :: CodeGen/ARM/alloc-no-stack-realign.ll
LLVM :: CodeGen/ARM/big-endian-ret-f64.ll
LLVM :: CodeGen/ARM/vst3.ll
LLVM :: CodeGen/X86/atomic16.ll
LLVM :: CodeGen/X86/atomic32.ll
LLVM :: CodeGen/X86/atomic6432.ll
LLVM :: CodeGen/X86/avx-load-store.ll
LLVM :: CodeGen/X86/fast-isel-gep.ll
LLVM :: CodeGen/X86/inline-asm-tied.ll
LLVM :: CodeGen/X86/musttail.ll
LLVM :: CodeGen/X86/win32_sret.ll
LLVM :: CodeGen/X86/win64_eh.ll
LLVM :: CodeGen/XCore/threads.ll
LLVM :: DebugInfo/X86/op_deref.ll
LLVM :: DebugInfo/X86/vla.ll

CodeGen/ARM/2010-05-18-LocalAllocCrash.ll crashes in llc:
/home/probinson/projects/upstream/writeable/llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp:776: void llvm::InstrEmitter::EmitMachineNode(llvm::SDNode*, bool, bool, llvm::DenseMap<llvm::SDValue, unsigned int>&): Assertion `NumMIOperands >= II.getNumOperands() && NumMIOperands <= II.getNumOperands() + II.getNumImplicitDefs() + NumImpUses && "#operands for dag node doesn't match .td file!"' failed.

CodeGen/X86/avx-load-store.ll gets a selection failure:
LLVM ERROR: Cannot select: 0x44b4df0: v2f64 = vector_shuffle 0x44b5230, 0x44b74c0<0,3> [ORD=3] [ID=14]
0x44b5230: v2f64 = scalar_to_vector 0x44b48a0 [ORD=3] [ID=13]
0x44b48a0: f64,ch = load 0x44a12e0, 0x44b4bd0, 0x44b5120<LD8[%ptr]> [ORD=2] [ID=10]
0x44b4bd0: i64,ch = CopyFromReg 0x44a12e0, 0x44b73b0 [ORD=2] [ID=7]
0x44b73b0: i64 = Register %vreg2 [ID=1]
0x44b5120: i64 = undef [ID=3]
0x44b74c0: v2f64 = bitcast 0x44b63a0 [ORD=3] [ID=12]
0x44b63a0: v4i32 = BUILD_VECTOR 0x44b6c20, 0x44b6c20, 0x44b6c20, 0x44b6c20 [ORD=3] [ID=9]
0x44b6c20: i32 = Constant<0> [ID=6]
0x44b6c20: i32 = Constant<0> [ID=6]
0x44b6c20: i32 = Constant<0> [ID=6]
0x44b6c20: i32 = Constant<0> [ID=6]
In function: mov01

I don't think this is an entirely reasonable goal. DAGCombiner performs various canonicalizations the targets rely on, and being able to disable it increases the complexity they will need to deal with. There isn't really a way to have separate phases with the current DAG, so these aren't cleanly separated from "real" optimizations.