[RegisterBankInfo] Ignore InstrMappings that create impossible to repair operands

[tstellar]

Summary

This is a follow-up to r303043. In computeMapping(), we need to disqualify an
InstrMapping if it would be impossible to repair one of the registers in the
instruction to match the mapping.

This change is needed in order to be able to define an instruction
mapping for G_SELECT for the AMDGPU target and will be tested
by test/CodeGen/AMDGPU/GlobalISel/regbankselect-select.mir

[qcolombet]

Silly me for not catching this during the first review.

LGTM!

llvm/trunk/lib/CodeGen/GlobalISel/RegBankSelect.cpp

for (unsigned OpIdx = 0, EndOpIdx = InstrMapping.getNumOperands();
     // For the greedy mode, we should choose greedily what is the best
     // choice based on the next use of MO.
 
     // Sums up the repairing cost of MO at each insertion point.
     uint64_t RepairCost = getRepairCost(MO, ValMapping);
 
     // This is an impossible to repair cost.
     if (RepairCost == std::numeric_limits<unsigned>::max())
-      continue;
+      return MappingCost::ImpossibleCost();
 
     // Bias used for splitting: 5%.
     const uint64_t PercentageForBias = 5;
     uint64_t Bias = (RepairCost * PercentageForBias + 99) / 100;
     // We should not need more than a couple of instructions to repair
     // an assignment. In other words, the computation should not
     // overflow because the repairing cost is free of basic block
     // frequency.