[PATCH] D113784: [AMDGPU][GlobalISel] Try to rematerialize when repairing regbank

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added a comment.
Herald added a subscriber: kosarev.
Herald added a project: All.

I'm currently thinking rematerialized in regbankselect is a good strategy. We would still want to have a post-isel operand optimizer pass try to re-scalarize constants when there's open constant bus use.



================
Comment at: llvm/include/llvm/CodeGen/GlobalISel/RegisterBankInfo.h:631
+  /// instead of repairing it by inserting copies.
+  virtual bool canRematerialize(Register Reg,
+                                const MachineRegisterInfo &MRI) const {
----------------
It would be better to report the rematerializability of an opcode as part of the reported mapping rather than add another callback for it


================
Comment at: llvm/lib/CodeGen/GlobalISel/RegBankSelect.cpp:166-170
+    if (RBI->canRematerialize(Src, *MRI)) {
+      MachineInstr *SrcMI = MRI->getVRegDef(Src);
+      MI = SrcMI->getMF()->CloneMachineInstr(SrcMI);
+      MI->getOperand(0).setReg(Dst);
+      LLVM_DEBUG(dbgs() << "Clone with new regbank: " << *SrcMI << '\n');
----------------
I think it would be cleaner to handle this as a separate RepairingPlacement case in applyMapping


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113784/new/

https://reviews.llvm.org/D113784