[llvm-branch-commits] [WebAssembly] Port WebAssemblyOptimizeLiveIntervalsPass (PR #209958)

Aiden Grossman via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Wed Jul 22 07:18:16 PDT 2026


================
@@ -121,3 +124,21 @@ bool WebAssemblyOptimizeLiveIntervals::runOnMachineFunction(
 
   return true;
 }
+
+bool WebAssemblyOptimizeLiveIntervalsLegacy::runOnMachineFunction(
+    MachineFunction &MF) {
+  return optimizeLiveIntervals(
+      MF, getAnalysis<LiveIntervalsWrapperPass>().getLIS());
+}
+
+PreservedAnalyses WebAssemblyOptimizeLiveIntervalsPass::run(
+    MachineFunction &MF, MachineFunctionAnalysisManager &MFAM) {
+  bool Changed =
+      optimizeLiveIntervals(MF, MFAM.getResult<LiveIntervalsAnalysis>(MF));
+  if (!Changed)
+    return PreservedAnalyses::all();
+  return getMachineFunctionPassPreservedAnalyses()
+      .preserveSet<CFGAnalyses>()
+      .preserve<LiveIntervalsAnalysis>()
+      .preserve<SlotIndexesAnalysis>();
----------------
boomanaiden154 wrote:

`MachineBlockFrequencyInfo` and `MachineDominatorTree` are both preserved by marking the CFG as preserved.

I don't really understand why `LiveVariables` is preserved in the Legacy version. We don't do anything to maintain it, and it also isn't computed/invalidated anywhere around this pass in the pipeline, so I've left it omitted.

https://gist.github.com/boomanaiden154/7155128f498a4275079a6c6d870b60ef is the `-debug-pass-manager` output which shows that preserving `LiveVariablesAnalysis` shouldn't matter.

https://github.com/llvm/llvm-project/pull/209958


More information about the llvm-branch-commits mailing list