[AMDGPU][NewPM] Port AMDGPUInsertDelayAlu to NPM (#128003)

Author: optimisan

llvm/lib/Target/AMDGPU/AMDGPU.h

@@ -270,6 +270,12 @@ struct AMDGPUAtomicOptimizerPass : PassInfoMixin<AMDGPUAtomicOptimizerPass> {
   ScanOptions ScanImpl;
 };
 
+struct AMDGPUInsertDelayAluPass
+    : public PassInfoMixin<AMDGPUInsertDelayAluPass> {
+  PreservedAnalyses run(MachineFunction &F,
+                        MachineFunctionAnalysisManager &MFAM);
+};
+
 Pass *createAMDGPUStructurizeCFGPass();
 FunctionPass *createAMDGPUISelDag(TargetMachine &TM, CodeGenOptLevel OptLevel);
 ModulePass *createAMDGPUAlwaysInlinePass(bool GlobalOpt = true);
@@ -416,7 +422,7 @@ extern char &SIMemoryLegalizerID;
 void initializeSIModeRegisterPass(PassRegistry&);
 extern char &SIModeRegisterID;
 
-void initializeAMDGPUInsertDelayAluPass(PassRegistry &);
+void initializeAMDGPUInsertDelayAluLegacyPass(PassRegistry &);
 extern char &AMDGPUInsertDelayAluID;
 
 void initializeSIInsertHardClausesPass(PassRegistry &);

llvm/lib/Target/AMDGPU/AMDGPUInsertDelayAlu.cpp

@@ -23,22 +23,13 @@ using namespace llvm;
 
 namespace {
 
-class AMDGPUInsertDelayAlu : public MachineFunctionPass {
+class AMDGPUInsertDelayAlu {
 public:
-  static char ID;
-
   const SIInstrInfo *SII;
   const TargetRegisterInfo *TRI;
 
   const TargetSchedModel *SchedModel;
 
-  AMDGPUInsertDelayAlu() : MachineFunctionPass(ID) {}
-
-  void getAnalysisUsage(AnalysisUsage &AU) const override {
-    AU.setPreservesCFG();
-    MachineFunctionPass::getAnalysisUsage(AU);
-  }
-
   // Return true if MI waits for all outstanding VALU instructions to complete.
   static bool instructionWaitsForVALU(const MachineInstr &MI) {
     // These instruction types wait for VA_VDST==0 before issuing.
@@ -416,10 +407,7 @@ class AMDGPUInsertDelayAlu : public MachineFunctionPass {
     return Changed;
   }
 
-  bool runOnMachineFunction(MachineFunction &MF) override {
-    if (skipFunction(MF.getFunction()))
-      return false;
-
+  bool run(MachineFunction &MF) {
     LLVM_DEBUG(dbgs() << "AMDGPUInsertDelayAlu running on " << MF.getName()
                       << "\n");
 
@@ -454,11 +442,39 @@ class AMDGPUInsertDelayAlu : public MachineFunctionPass {
   }
 };
 
+class AMDGPUInsertDelayAluLegacy : public MachineFunctionPass {
+public:
+  static char ID;
+
+  AMDGPUInsertDelayAluLegacy() : MachineFunctionPass(ID) {}
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesCFG();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override {
+    if (skipFunction(MF.getFunction()))
+      return false;
+    AMDGPUInsertDelayAlu Impl;
+    return Impl.run(MF);
+  }
+};
 } // namespace
 
-char AMDGPUInsertDelayAlu::ID = 0;
+PreservedAnalyses
+AMDGPUInsertDelayAluPass::run(MachineFunction &MF,
+                              MachineFunctionAnalysisManager &MFAM) {
+  if (!AMDGPUInsertDelayAlu().run(MF))
+    return PreservedAnalyses::all();
+  auto PA = getMachineFunctionPassPreservedAnalyses();
+  PA.preserveSet<CFGAnalyses>();
+  return PA;
+} // end namespace llvm
+
+char AMDGPUInsertDelayAluLegacy::ID = 0;
 
-char &llvm::AMDGPUInsertDelayAluID = AMDGPUInsertDelayAlu::ID;
+char &llvm::AMDGPUInsertDelayAluID = AMDGPUInsertDelayAluLegacy::ID;
 
-INITIALIZE_PASS(AMDGPUInsertDelayAlu, DEBUG_TYPE, "AMDGPU Insert Delay ALU",
-                false, false)
+INITIALIZE_PASS(AMDGPUInsertDelayAluLegacy, DEBUG_TYPE,
+                "AMDGPU Insert Delay ALU", false, false)

llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def

@@ -96,6 +96,7 @@ FUNCTION_PASS_WITH_PARAMS(
 #ifndef MACHINE_FUNCTION_PASS
 #define MACHINE_FUNCTION_PASS(NAME, CREATE_PASS)
 #endif
+MACHINE_FUNCTION_PASS("amdgpu-insert-delay-alu", AMDGPUInsertDelayAluPass())
 MACHINE_FUNCTION_PASS("amdgpu-isel", AMDGPUISelDAGToDAGPass(*this))
 MACHINE_FUNCTION_PASS("amdgpu-pre-ra-long-branch-reg", GCNPreRALongBranchRegPass())
 MACHINE_FUNCTION_PASS("amdgpu-rewrite-partial-reg-uses", GCNRewritePartialRegUsesPass())
@@ -122,7 +123,6 @@ MACHINE_FUNCTION_PASS("si-wqm", SIWholeQuadModePass())
 #undef MACHINE_FUNCTION_PASS
 
 #define DUMMY_MACHINE_FUNCTION_PASS(NAME, CREATE_PASS)
-DUMMY_MACHINE_FUNCTION_PASS("amdgpu-insert-delay-alu", AMDGPUInsertDelayAluPass())
 DUMMY_MACHINE_FUNCTION_PASS("amdgpu-pre-ra-optimizations", GCNPreRAOptimizationsPass())
 DUMMY_MACHINE_FUNCTION_PASS("amdgpu-rewrite-partial-reg-uses", GCNRewritePartialRegUsesPass())
 DUMMY_MACHINE_FUNCTION_PASS("amdgpu-set-wave-priority", AMDGPUSetWavePriorityPass())

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -533,7 +533,7 @@ extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
   initializeAMDGPURewriteUndefForPHILegacyPass(*PR);
   initializeAMDGPUUnifyMetadataPass(*PR);
   initializeSIAnnotateControlFlowLegacyPass(*PR);
-  initializeAMDGPUInsertDelayAluPass(*PR);
+  initializeAMDGPUInsertDelayAluLegacyPass(*PR);
   initializeSIInsertHardClausesPass(*PR);
   initializeSIInsertWaitcntsPass(*PR);
   initializeSIModeRegisterPass(*PR);
@@ -2147,6 +2147,46 @@ void AMDGPUCodeGenPassBuilder::addPostRegAlloc(AddMachinePass &addPass) const {
   Base::addPostRegAlloc(addPass);
 }
 
+void AMDGPUCodeGenPassBuilder::addPreEmitPass(AddMachinePass &addPass) const {
+  if (isPassEnabled(EnableVOPD, CodeGenOptLevel::Less)) {
+    // TODO: addPass(GCNCreateVOPDPass());
+  }
+  // TODO: addPass(SIMemoryLegalizerPass());
+  // TODO: addPass(SIInsertWaitcntsPass());
+
+  // TODO: addPass(SIModeRegisterPass());
+
+  if (TM.getOptLevel() > CodeGenOptLevel::None) {
+    // TODO: addPass(SIInsertHardClausesPass());
+  }
+
+  // addPass(SILateBranchLoweringPass());
+  if (isPassEnabled(EnableSetWavePriority, CodeGenOptLevel::Less)) {
+    // TODO: addPass(AMDGPUSetWavePriorityPass());
+  }
+
+  if (TM.getOptLevel() > CodeGenOptLevel::None) {
+    // TODO: addPass(SIPreEmitPeepholePass());
+  }
+
+  // The hazard recognizer that runs as part of the post-ra scheduler does not
+  // guarantee to be able handle all hazards correctly. This is because if there
+  // are multiple scheduling regions in a basic block, the regions are scheduled
+  // bottom up, so when we begin to schedule a region we don't know what
+  // instructions were emitted directly before it.
+  //
+  // Here we add a stand-alone hazard recognizer pass which can handle all
+  // cases.
+  // TODO: addPass(PostRAHazardRecognizerPass());
+  addPass(AMDGPUWaitSGPRHazardsPass());
+
+  if (isPassEnabled(EnableInsertDelayAlu, CodeGenOptLevel::Less)) {
+    addPass(AMDGPUInsertDelayAluPass());
+  }
+
+  // TODO: addPass(BranchRelaxationPass());
+}
+
 bool AMDGPUCodeGenPassBuilder::isPassEnabled(const cl::opt<bool> &Opt,
                                              CodeGenOptLevel Level) const {
   if (Opt.getNumOccurrences())

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h

@@ -180,6 +180,7 @@ class AMDGPUCodeGenPassBuilder
   void addPreRewrite(AddMachinePass &) const;
   void addMachineSSAOptimization(AddMachinePass &) const;
   void addPostRegAlloc(AddMachinePass &) const;
+  void addPreEmitPass(AddMachinePass &) const;
 
   /// Check if a pass is enabled given \p Opt option. The option always
   /// overrides defaults if explicitly used. Otherwise its default will be used

llvm/test/CodeGen/AMDGPU/insert-delay-alu-literal.mir

@@ -1,5 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
 # RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize64 -verify-machineinstrs -run-pass=amdgpu-insert-delay-alu %s -o - | FileCheck %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize64 -verify-machineinstrs -passes=amdgpu-insert-delay-alu %s -o - | FileCheck %s
 
 ---
 name: valu_dep_1