[llvm-branch-commits] [llvm] [AMDGPU][NPM] Port SIMemoryLegalizer to NPM (PR #130060)

[Matt Arsenault via llvm-branch-commits]

================
@@ -2767,11 +2775,26 @@ bool SIMemoryLegalizer::expandAtomicCmpxchgOrRmw(const SIMemOpInfo &MOI,
   return Changed;
 }
 
-bool SIMemoryLegalizer::runOnMachineFunction(MachineFunction &MF) {
-  bool Changed = false;
-
+bool SIMemoryLegalizerLegacy::runOnMachineFunction(MachineFunction &MF) {
   const MachineModuleInfo &MMI =
       getAnalysis<MachineModuleInfoWrapperPass>().getMMI();
+  return SIMemoryLegalizer(MMI).run(MF);
+}
+
+PreservedAnalyses
+SIMemoryLegalizerPass::run(MachineFunction &MF,
+                           MachineFunctionAnalysisManager &MFAM) {
+  auto *MMI = MFAM.getResult<ModuleAnalysisManagerMachineFunctionProxy>(MF)
+                  .getCachedResult<MachineModuleAnalysis>(
+                      *MF.getFunction().getParent());
+  assert(MMI && "MachineModuleAnalysis must be available");
+  if (!SIMemoryLegalizer(MMI->getMMI()).run(MF))
----------------
arsenm wrote:

This pass shouldn't really depend on MachineModuleInfo (not this patch's fault though) 

https://github.com/llvm/llvm-project/pull/130060