[AMDGPU][NPM] Port SIInsertWaitcnts to NPM

Author: optimisan

llvm/lib/Target/AMDGPU/AMDGPU.h

@@ -370,6 +370,13 @@ class SIMemoryLegalizerPass : public PassInfoMixin<SIMemoryLegalizerPass> {
   static bool isRequired() { return true; }
 };
 
+class SIInsertWaitcntsPass : public PassInfoMixin<SIInsertWaitcntsPass> {
+public:
+  PreservedAnalyses run(MachineFunction &MF,
+                        MachineFunctionAnalysisManager &MFAM);
+  static bool isRequired() { return true; }
+};
+
 FunctionPass *createAMDGPUAnnotateUniformValuesLegacy();
 
 ModulePass *createAMDGPUPrintfRuntimeBinding();
@@ -446,7 +453,7 @@ extern char &AMDGPUInsertDelayAluID;
 void initializeSIInsertHardClausesPass(PassRegistry &);
 extern char &SIInsertHardClausesID;
 
-void initializeSIInsertWaitcntsPass(PassRegistry&);
+void initializeSIInsertWaitcntsLegacyPass(PassRegistry &);
 extern char &SIInsertWaitcntsID;
 
 void initializeSIFormMemoryClausesLegacyPass(PassRegistry &);

llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def

@@ -109,6 +109,7 @@ MACHINE_FUNCTION_PASS("si-fix-vgpr-copies", SIFixVGPRCopiesPass())
 MACHINE_FUNCTION_PASS("si-fold-operands", SIFoldOperandsPass());
 MACHINE_FUNCTION_PASS("si-form-memory-clauses", SIFormMemoryClausesPass())
 MACHINE_FUNCTION_PASS("si-i1-copies", SILowerI1CopiesPass())
+MACHINE_FUNCTION_PASS("si-insert-waitcnts", SIInsertWaitcntsPass())
 MACHINE_FUNCTION_PASS("si-load-store-opt", SILoadStoreOptimizerPass())
 MACHINE_FUNCTION_PASS("si-lower-control-flow", SILowerControlFlowPass())
 MACHINE_FUNCTION_PASS("si-lower-sgpr-spills", SILowerSGPRSpillsPass())
@@ -131,7 +132,6 @@ DUMMY_MACHINE_FUNCTION_PASS("amdgpu-rewrite-partial-reg-uses", GCNRewritePartial
 DUMMY_MACHINE_FUNCTION_PASS("amdgpu-set-wave-priority", AMDGPUSetWavePriorityPass())
 
 DUMMY_MACHINE_FUNCTION_PASS("si-insert-hard-clauses", SIInsertHardClausesPass())
-DUMMY_MACHINE_FUNCTION_PASS("si-insert-waitcnts", SIInsertWaitcntsPass())
 DUMMY_MACHINE_FUNCTION_PASS("si-late-branch-lowering", SILateBranchLoweringPass())
 DUMMY_MACHINE_FUNCTION_PASS("si-pre-emit-peephole", SIPreEmitPeepholePass())
 // TODO: Move amdgpu-preload-kern-arg-prolog to MACHINE_FUNCTION_PASS since it

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -535,7 +535,7 @@ extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
   initializeSIAnnotateControlFlowLegacyPass(*PR);
   initializeAMDGPUInsertDelayAluLegacyPass(*PR);
   initializeSIInsertHardClausesPass(*PR);
-  initializeSIInsertWaitcntsPass(*PR);
+  initializeSIInsertWaitcntsLegacyPass(*PR);
   initializeSIModeRegisterLegacyPass(*PR);
   initializeSIWholeQuadModeLegacyPass(*PR);
   initializeSILowerControlFlowLegacyPass(*PR);
@@ -2153,7 +2153,7 @@ void AMDGPUCodeGenPassBuilder::addPreEmitPass(AddMachinePass &addPass) const {
   }
 
   addPass(SIMemoryLegalizerPass());
-  // TODO: addPass(SIInsertWaitcntsPass());
+  addPass(SIInsertWaitcntsPass());
 
   // TODO: addPass(SIModeRegisterPass());
 

llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp

@@ -33,6 +33,7 @@
 #include "llvm/ADT/Sequence.h"
 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachinePassManager.h"
 #include "llvm/CodeGen/MachinePostDominators.h"
 #include "llvm/Support/DebugCounter.h"
 #include "llvm/TargetParser/TargetParser.h"
@@ -594,7 +595,7 @@ class WaitcntGeneratorGFX12Plus : public WaitcntGenerator {
   AMDGPU::Waitcnt getAllZeroWaitcnt(bool IncludeVSCnt) const override;
 };
 
-class SIInsertWaitcnts : public MachineFunctionPass {
+class SIInsertWaitcnts {
 private:
   const GCNSubtarget *ST = nullptr;
   const SIInstrInfo *TII = nullptr;
@@ -633,9 +634,9 @@ class SIInsertWaitcnts : public MachineFunctionPass {
   InstCounterType MaxCounter = NUM_NORMAL_INST_CNTS;
 
 public:
-  static char ID;
-
-  SIInsertWaitcnts() : MachineFunctionPass(ID) {
+  SIInsertWaitcnts(MachineLoopInfo *MLI, MachinePostDominatorTree *PDT,
+                   AliasAnalysis *AA)
+      : MLI(MLI), PDT(PDT), AA(AA) {
     (void)ForceExpCounter;
     (void)ForceLgkmCounter;
     (void)ForceVMCounter;
@@ -645,20 +646,7 @@ class SIInsertWaitcnts : public MachineFunctionPass {
   bool isPreheaderToFlush(MachineBasicBlock &MBB,
                           WaitcntBrackets &ScoreBrackets);
   bool isVMEMOrFlatVMEM(const MachineInstr &MI) const;
-  bool runOnMachineFunction(MachineFunction &MF) override;
-
-  StringRef getPassName() const override {
-    return "SI insert wait instructions";
-  }
-
-  void getAnalysisUsage(AnalysisUsage &AU) const override {
-    AU.setPreservesCFG();
-    AU.addRequired<MachineLoopInfoWrapperPass>();
-    AU.addRequired<MachinePostDominatorTreeWrapperPass>();
-    AU.addUsedIfAvailable<AAResultsWrapperPass>();
-    AU.addPreserved<AAResultsWrapperPass>();
-    MachineFunctionPass::getAnalysisUsage(AU);
-  }
+  bool run(MachineFunction &MF);
 
   bool isForceEmitWaitcnt() const {
     for (auto T : inst_counter_types())
@@ -742,6 +730,36 @@ class SIInsertWaitcnts : public MachineFunctionPass {
                             WaitcntBrackets &ScoreBrackets);
 };
 
+class SIInsertWaitcntsLegacy : public MachineFunctionPass {
+public:
+  static char ID;
+  SIInsertWaitcntsLegacy() : MachineFunctionPass(ID) {}
+
+  bool runOnMachineFunction(MachineFunction &MF) override {
+    auto *MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
+    auto *PDT =
+        &getAnalysis<MachinePostDominatorTreeWrapperPass>().getPostDomTree();
+    AliasAnalysis *AA = nullptr;
+    if (auto *AAR = getAnalysisIfAvailable<AAResultsWrapperPass>())
+      AA = &AAR->getAAResults();
+
+    return SIInsertWaitcnts(MLI, PDT, AA).run(MF);
+  }
+
+  StringRef getPassName() const override {
+    return "SI insert wait instructions";
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesCFG();
+    AU.addRequired<MachineLoopInfoWrapperPass>();
+    AU.addRequired<MachinePostDominatorTreeWrapperPass>();
+    AU.addUsedIfAvailable<AAResultsWrapperPass>();
+    AU.addPreserved<AAResultsWrapperPass>();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+};
+
 } // end anonymous namespace
 
 RegInterval WaitcntBrackets::getRegInterval(const MachineInstr *MI,
@@ -1124,19 +1142,19 @@ bool WaitcntBrackets::counterOutOfOrder(InstCounterType T) const {
   return hasMixedPendingEvents(T);
 }
 
-INITIALIZE_PASS_BEGIN(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false,
-                      false)
+INITIALIZE_PASS_BEGIN(SIInsertWaitcntsLegacy, DEBUG_TYPE, "SI Insert Waitcnts",
+                      false, false)
 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)
 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTreeWrapperPass)
-INITIALIZE_PASS_END(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false,
-                    false)
+INITIALIZE_PASS_END(SIInsertWaitcntsLegacy, DEBUG_TYPE, "SI Insert Waitcnts",
+                    false, false)
 
-char SIInsertWaitcnts::ID = 0;
+char SIInsertWaitcntsLegacy::ID = 0;
 
-char &llvm::SIInsertWaitcntsID = SIInsertWaitcnts::ID;
+char &llvm::SIInsertWaitcntsID = SIInsertWaitcntsLegacy::ID;
 
 FunctionPass *llvm::createSIInsertWaitcntsPass() {
-  return new SIInsertWaitcnts();
+  return new SIInsertWaitcntsLegacy();
 }
 
 static bool updateOperandIfDifferent(MachineInstr &MI, AMDGPU::OpName OpName,
@@ -2406,16 +2424,29 @@ bool SIInsertWaitcnts::shouldFlushVmCnt(MachineLoop *ML,
   return HasVMemLoad && UsesVgprLoadedOutside && ST->hasVmemWriteVgprInOrder();
 }
 
-bool SIInsertWaitcnts::runOnMachineFunction(MachineFunction &MF) {
+PreservedAnalyses
+SIInsertWaitcntsPass::run(MachineFunction &MF,
+                          MachineFunctionAnalysisManager &MFAM) {
+  auto *MLI = &MFAM.getResult<MachineLoopAnalysis>(MF);
+  auto *PDT = &MFAM.getResult<MachinePostDominatorTreeAnalysis>(MF);
+  auto *AA = MFAM.getResult<FunctionAnalysisManagerMachineFunctionProxy>(MF)
+                 .getManager()
+                 .getCachedResult<AAManager>(MF.getFunction());
+
+  if (!SIInsertWaitcnts(MLI, PDT, AA).run(MF))
+    return PreservedAnalyses::all();
+
+  return getMachineFunctionPassPreservedAnalyses()
+      .preserveSet<CFGAnalyses>()
+      .preserve<AAManager>();
+}
+
+bool SIInsertWaitcnts::run(MachineFunction &MF) {
   ST = &MF.getSubtarget<GCNSubtarget>();
   TII = ST->getInstrInfo();
   TRI = &TII->getRegisterInfo();
   MRI = &MF.getRegInfo();
   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
-  MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
-  PDT = &getAnalysis<MachinePostDominatorTreeWrapperPass>().getPostDomTree();
-  if (auto *AAR = getAnalysisIfAvailable<AAResultsWrapperPass>())
-    AA = &AAR->getAAResults();
 
   AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST->getCPU());
 

llvm/test/CodeGen/AMDGPU/call-waw-waitcnt.mir

@@ -1,5 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 # RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs -run-pass=si-insert-waitcnts %s -o - | FileCheck -check-prefix=GCN %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs -passes=si-insert-waitcnts %s -o - | FileCheck -check-prefix=GCN %s
 
 # $sgpr30_sgpr31 will hold the return address. We need a waitcnt before SI_CALL so
 # that the return address is not clobbered in the callee by the outstanding load.

llvm/test/CodeGen/AMDGPU/insert-waitcnts-hang.mir

@@ -1,5 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 4
 # RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -run-pass si-insert-waitcnts %s -o - | FileCheck %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -passes si-insert-waitcnts %s -o - | FileCheck %s
 
 ---
 name: test

llvm/test/CodeGen/AMDGPU/vccz-corrupt-bug-workaround.mir

@@ -2,6 +2,8 @@
 # RUN: llc -run-pass=si-insert-waitcnts -mtriple=amdgcn -mcpu=gfx900 -o - %s | FileCheck %s -check-prefixes=CHECK,GFX9
 # RUN: llc -run-pass=si-insert-waitcnts -mtriple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize64 -o - %s | FileCheck %s
 # RUN: llc -run-pass=si-insert-waitcnts -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize64 -o - %s | FileCheck %s
+
+# RUN: llc -passes=si-insert-waitcnts -mtriple=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize64 -o - %s | FileCheck %s
 ---
 # CHECK-LABEL: name: vccz_corrupt_workaround
 # CHECK: $vcc = V_CMP_EQ_F32