AMDGPU/NewPM Port GCNDPPCombine to NPM (#105816)

Co-authored-by: Akshat Oke <[email protected]>

Author: optimisan

llvm/lib/Target/AMDGPU/AMDGPU.h

@@ -157,8 +157,8 @@ struct AMDGPULowerBufferFatPointersPass
 void initializeAMDGPURewriteOutArgumentsPass(PassRegistry &);
 extern char &AMDGPURewriteOutArgumentsID;
 
-void initializeGCNDPPCombinePass(PassRegistry &);
-extern char &GCNDPPCombineID;
+void initializeGCNDPPCombineLegacyPass(PassRegistry &);
+extern char &GCNDPPCombineLegacyID;
 
 void initializeSIFoldOperandsLegacyPass(PassRegistry &);
 extern char &SIFoldOperandsLegacyID;

llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def

@@ -98,4 +98,5 @@ MACHINE_FUNCTION_PASS("amdgpu-isel", AMDGPUISelDAGToDAGPass(*this))
 MACHINE_FUNCTION_PASS("si-fix-sgpr-copies", SIFixSGPRCopiesPass())
 MACHINE_FUNCTION_PASS("si-i1-copies", SILowerI1CopiesPass())
 MACHINE_FUNCTION_PASS("si-fold-operands", SIFoldOperandsPass());
+MACHINE_FUNCTION_PASS("gcn-dpp-combine", GCNDPPCombinePass())
 #undef MACHINE_FUNCTION_PASS

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -28,6 +28,7 @@
 #include "AMDGPUTargetObjectFile.h"
 #include "AMDGPUTargetTransformInfo.h"
 #include "AMDGPUUnifyDivergentExitNodes.h"
+#include "GCNDPPCombine.h"
 #include "GCNIterativeScheduler.h"
 #include "GCNSchedStrategy.h"
 #include "GCNVOPDUtils.h"
@@ -403,7 +404,7 @@ extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
   initializeR600VectorRegMergerPass(*PR);
   initializeGlobalISel(*PR);
   initializeAMDGPUDAGToDAGISelLegacyPass(*PR);
-  initializeGCNDPPCombinePass(*PR);
+  initializeGCNDPPCombineLegacyPass(*PR);
   initializeSILowerI1CopiesLegacyPass(*PR);
   initializeAMDGPUGlobalISelDivergenceLoweringPass(*PR);
   initializeSILowerWWMCopiesPass(*PR);
@@ -1273,7 +1274,7 @@ void GCNPassConfig::addMachineSSAOptimization() {
   // XXX - Can we get away without running DeadMachineInstructionElim again?
   addPass(&SIFoldOperandsLegacyID);
   if (EnableDPPCombine)
-    addPass(&GCNDPPCombineID);
+    addPass(&GCNDPPCombineLegacyID);
   addPass(&SILoadStoreOptimizerID);
   if (isPassEnabled(EnableSDWAPeephole)) {
     addPass(&SIPeepholeSDWAID);

llvm/lib/Target/AMDGPU/GCNDPPCombine.cpp

@@ -37,6 +37,7 @@
 // The mov_dpp instruction should reside in the same BB as all its uses
 //===----------------------------------------------------------------------===//
 
+#include "GCNDPPCombine.h"
 #include "AMDGPU.h"
 #include "GCNSubtarget.h"
 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
@@ -51,7 +52,7 @@ STATISTIC(NumDPPMovsCombined, "Number of DPP moves combined.");
 
 namespace {
 
-class GCNDPPCombine : public MachineFunctionPass {
+class GCNDPPCombine {
   MachineRegisterInfo *MRI;
   const SIInstrInfo *TII;
   const GCNSubtarget *ST;
@@ -76,12 +77,18 @@ class GCNDPPCombine : public MachineFunctionPass {
 
   bool combineDPPMov(MachineInstr &MI) const;
 
+  int getDPPOp(unsigned Op, bool IsShrinkable) const;
+  bool isShrinkable(MachineInstr &MI) const;
+
+public:
+  bool run(MachineFunction &MF);
+};
+
+class GCNDPPCombineLegacy : public MachineFunctionPass {
 public:
   static char ID;
 
-  GCNDPPCombine() : MachineFunctionPass(ID) {
-    initializeGCNDPPCombinePass(*PassRegistry::getPassRegistry());
-  }
+  GCNDPPCombineLegacy() : MachineFunctionPass(ID) {}
 
   bool runOnMachineFunction(MachineFunction &MF) override;
 
@@ -96,22 +103,19 @@ class GCNDPPCombine : public MachineFunctionPass {
     return MachineFunctionProperties()
       .set(MachineFunctionProperties::Property::IsSSA);
   }
-
-private:
-  int getDPPOp(unsigned Op, bool IsShrinkable) const;
-  bool isShrinkable(MachineInstr &MI) const;
 };
 
 } // end anonymous namespace
 
-INITIALIZE_PASS(GCNDPPCombine, DEBUG_TYPE, "GCN DPP Combine", false, false)
+INITIALIZE_PASS(GCNDPPCombineLegacy, DEBUG_TYPE, "GCN DPP Combine", false,
+                false)
 
-char GCNDPPCombine::ID = 0;
+char GCNDPPCombineLegacy::ID = 0;
 
-char &llvm::GCNDPPCombineID = GCNDPPCombine::ID;
+char &llvm::GCNDPPCombineLegacyID = GCNDPPCombineLegacy::ID;
 
 FunctionPass *llvm::createGCNDPPCombinePass() {
-  return new GCNDPPCombine();
+  return new GCNDPPCombineLegacy();
 }
 
 bool GCNDPPCombine::isShrinkable(MachineInstr &MI) const {
@@ -749,9 +753,16 @@ bool GCNDPPCombine::combineDPPMov(MachineInstr &MovMI) const {
   return !Rollback;
 }
 
-bool GCNDPPCombine::runOnMachineFunction(MachineFunction &MF) {
+bool GCNDPPCombineLegacy::runOnMachineFunction(MachineFunction &MF) {
+  if (skipFunction(MF.getFunction()))
+    return false;
+
+  return GCNDPPCombine().run(MF);
+}
+
+bool GCNDPPCombine::run(MachineFunction &MF) {
   ST = &MF.getSubtarget<GCNSubtarget>();
-  if (!ST->hasDPP() || skipFunction(MF.getFunction()))
+  if (!ST->hasDPP())
     return false;
 
   MRI = &MF.getRegInfo();
@@ -781,3 +792,19 @@ bool GCNDPPCombine::runOnMachineFunction(MachineFunction &MF) {
   }
   return Changed;
 }
+
+PreservedAnalyses GCNDPPCombinePass::run(MachineFunction &MF,
+                                         MachineFunctionAnalysisManager &) {
+  if (MF.getFunction().hasOptNone())
+    return PreservedAnalyses::all();
+
+  MFPropsModifier _(*this, MF);
+
+  bool Changed = GCNDPPCombine().run(MF);
+  if (!Changed)
+    return PreservedAnalyses::all();
+
+  auto PA = getMachineFunctionPassPreservedAnalyses();
+  PA.preserveSet<CFGAnalyses>();
+  return PA;
+}

llvm/lib/Target/AMDGPU/GCNDPPCombine.h

@@ -0,0 +1,28 @@
+//=======--- GCNDPPCombine.h - optimization for DPP instructions ---==========//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_TARGET_AMDGPU_GCNDPPCOMBINE_H
+#define LLVM_LIB_TARGET_AMDGPU_GCNDPPCOMBINE_H
+
+#include "llvm/CodeGen/MachinePassManager.h"
+
+namespace llvm {
+class GCNDPPCombinePass : public PassInfoMixin<GCNDPPCombinePass> {
+public:
+  PreservedAnalyses run(MachineFunction &MF,
+                        MachineFunctionAnalysisManager &MAM);
+
+  MachineFunctionProperties getRequiredProperties() {
+    return MachineFunctionProperties().set(
+        MachineFunctionProperties::Property::IsSSA);
+  }
+};
+
+} // end namespace llvm
+
+#endif // LLVM_LIB_TARGET_AMDGPU_GCNDPPCOMBINE_H

llvm/test/CodeGen/AMDGPU/dpp_combine.mir

@@ -1,4 +1,5 @@
 # RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=gcn-dpp-combine -verify-machineinstrs -o - %s | FileCheck %s -check-prefix=GCN
+# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -passes=gcn-dpp-combine -verify-machineinstrs -o - %s | FileCheck %s -check-prefix=GCN
 
 ---
 # old is undefined: only combine when masks are fully enabled and

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.fp8.dpp.mir

@@ -1,5 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 # RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -run-pass=gcn-dpp-combine %s -o - | FileCheck -check-prefix=GFX12 %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -passes=gcn-dpp-combine %s -o - | FileCheck -check-prefix=GFX12 %s
 
 ---
 name:            test_cvt_f32_bf8_byte0