InstrEmitter::CreateVirtualRegisters handling of CopyToReg


I wonder if anyone has any comment on a patch like:

diff --git a/lib/CodeGen/SelectionDAG/InstrEmitter.cpp b/lib/CodeGen/SelectionDAG/InstrEmitter.cpp
index 65ee3816f84..4780f6f0e59 100644
--- a/lib/CodeGen/SelectionDAG/InstrEmitter.cpp
+++ b/lib/CodeGen/SelectionDAG/InstrEmitter.cpp
@@ -243,18 +243,21 @@ void InstrEmitter::CreateVirtualRegisters(SDNode *Node,

  if \(\!VRBase && \!IsClone && \!IsCloned\)
    for \(SDNode \*User : Node\->uses\(\)\) \{
      if \(User\->getOpcode\(\) == ISD::CopyToReg &&
          User\->getOperand\(2\)\.getNode\(\) == Node &&
          User\->getOperand\(2\)\.getResNo\(\) == i\) \{
        unsigned Reg = cast<RegisterSDNode>\(User\->getOperand\(1\)\)\->getReg\(\);
        if \(TargetRegisterInfo::isVirtualRegister\(Reg\)\) \{

- const TargetRegisterClass *RegRC = MRI->getRegClass(Reg);
- if (RegRC == RC) {
+ // Allow constraining the virtual register's class within reason,
+ // just like what AddRegisterOperand will allow.
+ const TargetRegisterClass *ConstrainedRC
+ = MRI->constrainRegClass(Reg, RC, MinRCSize);
+ if (ConstrainedRC) {
VRBase = Reg;
MIB.addReg(VRBase, RegState::Define);

Why do the register classes currently have to match exactly in this case?

It seems that these COPYs that now remain may end up in the same register class, if the users require it. So why not constrain also here directly, if this is done generally when the register is used as input?



It would still be nice to get some comment about this patch which seems to help on SystemZ in the case where GRX32 pseudo instructions are used. This means that the register eventually ends up either in GR32 or GRH32 (low or high 32 bit subregs).

I haven't had the occasion to touch this particular part of InstrEmitter, but the patch looks reasonable.