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

[Jay Foad via llvm-commits]

================
@@ -2434,6 +2445,105 @@ 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)) {
+        if (SingleUseIdx == -1 && CandMO.isUse()) {
+          SingleUseIdx = i;
+        } else {
+          SingleUseIdx = -1;
+          break;
+        }
+      }
+    }
+    if (SingleUseIdx == -1)
+      break;
+    MachineOperand *UseMO = &I->getOperand(SingleUseIdx);
+    if (UseMO->getSubReg() == AMDGPU::NoSubRegister)
+      break;
+
+    unsigned Offset = RI.getSubRegIdxOffset(UseMO->getSubReg());
+    unsigned SubregSize = RI.getSubRegIdxSize(UseMO->getSubReg());
+
+    MachineFunction *MF = MBB.getParent();
+    MachineRegisterInfo &MRI = MF->getRegInfo();
+    assert(MRI.hasAtMostUserInstrs(DestReg, 0) &&
----------------
jayfoad wrote:

`use_nodbg_empty()`?

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