[llvm-branch-commits] [RISCV] Port RISCVInsertVSETVLI to NewPM (PR #235407)

[Luke Lau via llvm-branch-commits]

================
@@ -1219,6 +1228,29 @@ bool RISCVInsertVSETVLI::runOnMachineFunction(MachineFunction &MF) {
 }
 
 /// Returns an instance of the Insert VSETVLI pass.
-FunctionPass *llvm::createRISCVInsertVSETVLIPass() {
-  return new RISCVInsertVSETVLI();
+FunctionPass *llvm::createRISCVInsertVSETVLILegacyPass() {
+  return new RISCVInsertVSETVLILegacy();
+}
+
+bool RISCVInsertVSETVLILegacy::runOnMachineFunction(MachineFunction &MF) {
+  RISCVInsertVSETVLIImpl Impl;
+  auto *LISWrapper = getAnalysisIfAvailable<LiveIntervalsWrapperPass>();
+  LiveIntervals *LIS = LISWrapper ? &LISWrapper->getLIS() : nullptr;
+  return Impl.runOnMachineFunction(MF, LIS);
+}
+
+PreservedAnalyses
+RISCVInsertVSETVLIPass::run(MachineFunction &MF,
+                            MachineFunctionAnalysisManager &MFAM) {
+  RISCVInsertVSETVLIImpl Impl;
----------------
lukel97 wrote:

The allocations probably aren't a massive deal, but another reason I'd rather we keep the subtarget bail logic in the run function is that it makes easier to see that we're bailing before calling `MFAM.getResult` etc. 
For getCachedResult in RISCVInsertVSETVLI it doesn't make a difference but I'm noticing in e.g. RISCVVLOptimizer we're unnecessarily computing MachineDominatorTreeAnalysis because we bail inside the Impl wrapper. It would be good to establish a pattern here so we don't accidentally forget

https://github.com/llvm/llvm-project/pull/235407