RegAllocFast: Remove early selection loop, the spill calculation will report cost 0 anyway for free regs

[arsenm]

Summary

The 2nd loop calculates spill costs but reports free registers as cost 0 anyway, so there is little benefit from having a separate early loop.

SUrprisingly this is not NFC, as many register are marked regDisabled so the first loop often picks up later registers unnecessarily instead of the first one available in the allocation order...

[MatzeB]

This is part of my regallocfast rewrite series, see also D52010

[arsenm]

ping, I'm depending on this set of patches for D55301

[arsenm]

Test fixes

[arsenm]

Rebase test fixes

[arsenm]

r356499

[rupprecht]

This introduced a test failure in clang (CodeGen/arm64-microsoft-status-reg.cpp, part of check-clang-codegen). I sent a quick fix in D59557. I have commit access but I'm not confident it's the right fix, so I'd appreciate a review of that.

lib/CodeGen/RegAllocFast.cpp

if (TargetRegisterInfo::isPhysicalRegister(Hint) &&
     if (Cost < spillDirty) {
       if (Cost)
         definePhysReg(MI, Hint, regFree);
       assignVirtToPhysReg(LR, Hint);
       return;
     }
   }
 
-  // First try to find a completely free register.
-  ArrayRef<MCPhysReg> AllocationOrder = RegClassInfo.getOrder(&RC);
-  for (MCPhysReg PhysReg : AllocationOrder) {
-    if (PhysRegState[PhysReg] == regFree && !isRegUsedInInstr(PhysReg)) {
-      assignVirtToPhysReg(LR, PhysReg);
-      return;
-    }
-  }
-
   MCPhysReg BestReg = 0;
   unsigned BestCost = spillImpossible;
+  ArrayRef<MCPhysReg> AllocationOrder = RegClassInfo.getOrder(&RC);
   for (MCPhysReg PhysReg : AllocationOrder) {
     LLVM_DEBUG(dbgs() << "\tRegister: " << printReg(PhysReg, TRI) << ' ');
     unsigned Cost = calcSpillCost(PhysReg);
     LLVM_DEBUG(dbgs() << "Cost: " << Cost << " BestCost: " << BestCost << '\n');
     // Immediate take a register with cost 0.
     if (Cost == 0) {
       assignVirtToPhysReg(LR, PhysReg);
       return;