[llvm] [AMDGPU] Rematerialize scalar loads (PR #68778)

[Piotr Sobczak via llvm-commits]

================
@@ -2434,6 +2445,103 @@ bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
   return true;
 }
 
+void SIInstrInfo::reMaterialize(MachineBasicBlock &MBB,
+                                MachineBasicBlock::iterator I, Register DestReg,
+                                unsigned SubIdx, const MachineInstr &Orig,
+                                const TargetRegisterInfo &RI) const {
+
+  // Try shrinking the instruction to remat only the part needed for current
+  // context.
+  // TODO: Handle more cases.
+  unsigned Opcode = Orig.getOpcode();
+  switch (Opcode) {
+  case AMDGPU::S_LOAD_DWORDX16_IMM:
+  case AMDGPU::S_LOAD_DWORDX8_IMM: {
+    if (SubIdx != 0)
+      break;
+
+    if (I == MBB.end())
+      break;
+
+    if (I->isBundled())
+      break;
+
+    // Look for a single use of the register that is also a subreg.
+    Register RegToFind = Orig.getOperand(0).getReg();
+    int SingleUseIdx = -1;
+    for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
+      const MachineOperand &CandMO = I->getOperand(i);
+      if (!CandMO.isReg())
+        continue;
+      Register CandReg = CandMO.getReg();
+      if (!CandReg)
+        continue;
+
+      if (CandReg == RegToFind || RI.regsOverlap(CandReg, RegToFind)) {
----------------
piotrAMD wrote:

Fixed, thanks.

https://github.com/llvm/llvm-project/pull/68778