[llvm] [RISCV][VLOPT] Support segmented store instructions (PR #155467)

[Luke Lau via llvm-commits]

================
@@ -1396,6 +1502,22 @@ RISCVVLOptimizer::checkUsers(const MachineInstr &MI) const {
       continue;
     }
 
+    if (isTupleInsertInstr(UserMI, *MRI)) {
+      LLVM_DEBUG(dbgs().indent(4) << "Peeking through uses of INSERT_SUBREG\n");
+      for (MachineOperand &UseOp :
+           MRI->use_operands(UserMI.getOperand(0).getReg())) {
+        const MachineInstr &CandidateMI = *UseOp.getParent();
+        // We should not propagate the VL if the user is not a segmented store
+        // or another INSERT_SUBREG, since VL just works differently
+        // between segmented operations (per-field) v.s. other RVV ops (on the
+        // whole register group).
+        if (isTupleInsertInstr(CandidateMI, *MRI) ||
+            isSegmentedStoreInstr(CandidateMI))
+          Worklist.insert(&UseOp);
----------------
lukel97 wrote:

I think https://github.com/llvm/llvm-project/pull/151285 should make this less confusing. Currently checkUsers does two things, it checks to see if the EEW/EMUL are compatible, and it also tries to compute the maximum VL of its users. 

#151285 splits it out so checkUsers only does the EEW/EMUL check and leaves the maximum VL computation to the transfer function.

https://github.com/llvm/llvm-project/pull/155467