[llvm] Greedy: Make trySplitAroundHintReg try to match hints with subreg copies (PR #160294)

[Quentin Colombet via llvm-commits]

================
@@ -1383,21 +1383,35 @@ bool RAGreedy::trySplitAroundHintReg(MCPhysReg Hint,
   // Compute the cost of assigning a non Hint physical register to VirtReg.
   // We define it as the total frequency of broken COPY instructions to/from
   // Hint register, and after split, they can be deleted.
-  for (const MachineInstr &Instr : MRI->reg_nodbg_instructions(Reg)) {
-    if (!TII->isFullCopyInstr(Instr))
+
+  // FIXME: This is miscounting the costs with subregisters. In particular, this
+  // should support recognizing SplitKit formed copy bundles instead of direct
+  // copy instructions.
+  for (const MachineOperand &Opnd : MRI->reg_nodbg_operands(Reg)) {
+    const MachineInstr &Instr = *Opnd.getParent();
+    if (!Instr.isCopy() || Opnd.isImplicit())
       continue;
-    Register OtherReg = Instr.getOperand(1).getReg();
-    if (OtherReg == Reg) {
-      OtherReg = Instr.getOperand(0).getReg();
-      if (OtherReg == Reg)
-        continue;
+
+    // Look for the other end of the copy.
+    const bool IsDef = Opnd.isDef();
+    const MachineOperand &OtherOpnd = Instr.getOperand(IsDef);
+    Register OtherReg = OtherOpnd.getReg();
+    assert(Reg == Opnd.getReg());
+    if (OtherReg == Reg)
----------------
qcolombet wrote:

> the check I commented on would ignore this copy, even though it likely needs to be materialized to a real instruction

Not necessarily. E.g., let's say we have a v2x64 and a 64 and we copy the low 32-bit, the copy would be:
v1.vsub0_sub0 = v2.sub0

The copy uses different subregs indices, but it can still be a no-op if the low 64 bit of v1 is coalesced on v2.

That said, I haven't checked the code yet, so I don't know if there's a problem at this point :).

https://github.com/llvm/llvm-project/pull/160294