[llvm] [RISCV] Copy AVLs whose LiveIntervals aren't extendable in insertVSETVLI (PR #98342)

[Philip Reames via llvm-commits]

================
@@ -1149,15 +1143,26 @@ void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB,
                 .addImm(Info.encodeVTYPE());
   if (LIS) {
     LIS->InsertMachineInstrInMaps(*MI);
-    // Normally the AVL's live range will already extend past the inserted
-    // vsetvli because the pseudos below will already use the AVL. But this
-    // isn't always the case, e.g. PseudoVMV_X_S doesn't have an AVL operand or
-    // we've taken the AVL from the VL output of another vsetvli.
     LiveInterval &LI = LIS->getInterval(AVLReg);
     SlotIndex SI = LIS->getInstructionIndex(*MI).getRegSlot();
-    assert((LI.liveAt(SI) && LI.getVNInfoAt(SI) == Info.getAVLVNInfo()) ||
-           (!LI.liveAt(SI) && LI.containsOneValue()));
-    LIS->extendToIndices(LI, SI);
+    // If the AVL value isn't live at MI, do a quick check to see if it's easily
+    // extendable. Otherwise, we need to copy it.
+    if (LI.getVNInfoBefore(SI) != Info.getAVLVNInfo()) {
+      if (!LI.liveAt(SI) && LI.containsOneValue())
+        LIS->extendToIndices(LI, SI);
+      else {
+        Register AVLCopyReg =
+            MRI->createVirtualRegister(&RISCV::GPRNoX0RegClass);
+        MachineBasicBlock::iterator AVLDef =
----------------
preames wrote:

Do you know how a PHIDef is handled here?  Is the slot index simply mapped to the first instruction in the block?  Or something else?  Either way, we may need special handling for that case here. 

For correctness, I'd appreciate if you did a local patch which removes the previous extend check and ensures everything works.  The fallback path should be generic.  

https://github.com/llvm/llvm-project/pull/98342