[AMDGPU][NewPM] Port SIOptimizeVGPRLiveRange pass to NPM. (#117686)

Author: cdevadas

llvm/lib/Target/AMDGPU/AMDGPU.h

@@ -42,7 +42,7 @@ FunctionPass *createSILoadStoreOptimizerLegacyPass();
 FunctionPass *createSIWholeQuadModePass();
 FunctionPass *createSIFixControlFlowLiveIntervalsPass();
 FunctionPass *createSIOptimizeExecMaskingPreRAPass();
-FunctionPass *createSIOptimizeVGPRLiveRangePass();
+FunctionPass *createSIOptimizeVGPRLiveRangeLegacyPass();
 FunctionPass *createSIFixSGPRCopiesLegacyPass();
 FunctionPass *createLowerWWMCopiesPass();
 FunctionPass *createSIMemoryLegalizerPass();
@@ -359,8 +359,8 @@ struct AMDGPUUnifyMetadataPass : PassInfoMixin<AMDGPUUnifyMetadataPass> {
 void initializeSIOptimizeExecMaskingPreRAPass(PassRegistry&);
 extern char &SIOptimizeExecMaskingPreRAID;
 
-void initializeSIOptimizeVGPRLiveRangePass(PassRegistry &);
-extern char &SIOptimizeVGPRLiveRangeID;
+void initializeSIOptimizeVGPRLiveRangeLegacyPass(PassRegistry &);
+extern char &SIOptimizeVGPRLiveRangeLegacyID;
 
 void initializeAMDGPUAnnotateUniformValuesLegacyPass(PassRegistry &);
 extern char &AMDGPUAnnotateUniformValuesLegacyPassID;

llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def

@@ -101,6 +101,7 @@ MACHINE_FUNCTION_PASS("si-fold-operands", SIFoldOperandsPass());
 MACHINE_FUNCTION_PASS("gcn-dpp-combine", GCNDPPCombinePass())
 MACHINE_FUNCTION_PASS("si-load-store-opt", SILoadStoreOptimizerPass())
 MACHINE_FUNCTION_PASS("si-lower-sgpr-spills", SILowerSGPRSpillsPass())
+MACHINE_FUNCTION_PASS("si-opt-vgpr-liverange", SIOptimizeVGPRLiveRangePass())
 MACHINE_FUNCTION_PASS("si-peephole-sdwa", SIPeepholeSDWAPass())
 MACHINE_FUNCTION_PASS("si-pre-allocate-wwm-regs", SIPreAllocateWWMRegsPass())
 MACHINE_FUNCTION_PASS("si-shrink-instructions", SIShrinkInstructionsPass())

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -39,6 +39,7 @@
 #include "SILowerSGPRSpills.h"
 #include "SIMachineFunctionInfo.h"
 #include "SIMachineScheduler.h"
+#include "SIOptimizeVGPRLiveRange.h"
 #include "SIPeepholeSDWA.h"
 #include "SIPreAllocateWWMRegs.h"
 #include "SIShrinkInstructions.h"
@@ -472,7 +473,7 @@ extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
   initializeSIPeepholeSDWALegacyPass(*PR);
   initializeSIShrinkInstructionsLegacyPass(*PR);
   initializeSIOptimizeExecMaskingPreRAPass(*PR);
-  initializeSIOptimizeVGPRLiveRangePass(*PR);
+  initializeSIOptimizeVGPRLiveRangeLegacyPass(*PR);
   initializeSILoadStoreOptimizerLegacyPass(*PR);
   initializeAMDGPUCtorDtorLoweringLegacyPass(*PR);
   initializeAMDGPUAlwaysInlinePass(*PR);
@@ -1421,7 +1422,7 @@ void GCNPassConfig::addOptimizedRegAlloc() {
   // the register in LiveVariables, this would trigger a failure in verifier,
   // we should fix it and enable the verifier.
   if (OptVGPRLiveRange)
-    insertPass(&LiveVariablesID, &SIOptimizeVGPRLiveRangeID);
+    insertPass(&LiveVariablesID, &SIOptimizeVGPRLiveRangeLegacyID);
 
   // This must be run immediately after phi elimination and before
   // TwoAddressInstructions, otherwise the processing of the tied operand of

llvm/lib/Target/AMDGPU/SIOptimizeVGPRLiveRange.cpp

@@ -71,6 +71,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "SIOptimizeVGPRLiveRange.h"
 #include "AMDGPU.h"
 #include "GCNSubtarget.h"
 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
@@ -86,7 +87,7 @@ using namespace llvm;
 
 namespace {
 
-class SIOptimizeVGPRLiveRange : public MachineFunctionPass {
+class SIOptimizeVGPRLiveRange {
 private:
   const SIRegisterInfo *TRI = nullptr;
   const SIInstrInfo *TII = nullptr;
@@ -96,7 +97,10 @@ class SIOptimizeVGPRLiveRange : public MachineFunctionPass {
   MachineRegisterInfo *MRI = nullptr;
 
 public:
-  static char ID;
+  SIOptimizeVGPRLiveRange(LiveVariables *LV, MachineDominatorTree *MDT,
+                          MachineLoopInfo *Loops)
+      : LV(LV), MDT(MDT), Loops(Loops) {}
+  bool run(MachineFunction &MF);
 
   MachineBasicBlock *getElseTarget(MachineBasicBlock *MBB) const;
 
@@ -136,8 +140,13 @@ class SIOptimizeVGPRLiveRange : public MachineFunctionPass {
       Register Reg, MachineBasicBlock *LoopHeader,
       SmallSetVector<MachineBasicBlock *, 2> &LoopBlocks,
       SmallVectorImpl<MachineInstr *> &Instructions) const;
+};
 
-  SIOptimizeVGPRLiveRange() : MachineFunctionPass(ID) {}
+class SIOptimizeVGPRLiveRangeLegacy : public MachineFunctionPass {
+public:
+  static char ID;
+
+  SIOptimizeVGPRLiveRangeLegacy() : MachineFunctionPass(ID) {}
 
   bool runOnMachineFunction(MachineFunction &MF) override;
 
@@ -611,35 +620,59 @@ void SIOptimizeVGPRLiveRange::optimizeWaterfallLiveRange(
   }
 }
 
-char SIOptimizeVGPRLiveRange::ID = 0;
+char SIOptimizeVGPRLiveRangeLegacy::ID = 0;
 
-INITIALIZE_PASS_BEGIN(SIOptimizeVGPRLiveRange, DEBUG_TYPE,
+INITIALIZE_PASS_BEGIN(SIOptimizeVGPRLiveRangeLegacy, DEBUG_TYPE,
                       "SI Optimize VGPR LiveRange", false, false)
 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)
 INITIALIZE_PASS_DEPENDENCY(LiveVariablesWrapperPass)
-INITIALIZE_PASS_END(SIOptimizeVGPRLiveRange, DEBUG_TYPE,
+INITIALIZE_PASS_END(SIOptimizeVGPRLiveRangeLegacy, DEBUG_TYPE,
                     "SI Optimize VGPR LiveRange", false, false)
 
-char &llvm::SIOptimizeVGPRLiveRangeID = SIOptimizeVGPRLiveRange::ID;
+char &llvm::SIOptimizeVGPRLiveRangeLegacyID = SIOptimizeVGPRLiveRangeLegacy::ID;
+
+FunctionPass *llvm::createSIOptimizeVGPRLiveRangeLegacyPass() {
+  return new SIOptimizeVGPRLiveRangeLegacy();
+}
+
+bool SIOptimizeVGPRLiveRangeLegacy::runOnMachineFunction(MachineFunction &MF) {
+  if (skipFunction(MF.getFunction()))
+    return false;
 
-FunctionPass *llvm::createSIOptimizeVGPRLiveRangePass() {
-  return new SIOptimizeVGPRLiveRange();
+  LiveVariables *LV = &getAnalysis<LiveVariablesWrapperPass>().getLV();
+  MachineDominatorTree *MDT =
+      &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
+  MachineLoopInfo *Loops = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
+  return SIOptimizeVGPRLiveRange(LV, MDT, Loops).run(MF);
 }
 
-bool SIOptimizeVGPRLiveRange::runOnMachineFunction(MachineFunction &MF) {
+PreservedAnalyses
+SIOptimizeVGPRLiveRangePass::run(MachineFunction &MF,
+                                 MachineFunctionAnalysisManager &MFAM) {
+  MFPropsModifier _(*this, MF);
+  LiveVariables *LV = &MFAM.getResult<LiveVariablesAnalysis>(MF);
+  MachineDominatorTree *MDT = &MFAM.getResult<MachineDominatorTreeAnalysis>(MF);
+  MachineLoopInfo *Loops = &MFAM.getResult<MachineLoopAnalysis>(MF);
+
+  bool Changed = SIOptimizeVGPRLiveRange(LV, MDT, Loops).run(MF);
+  if (!Changed)
+    return PreservedAnalyses::all();
+
+  auto PA = getMachineFunctionPassPreservedAnalyses();
+  PA.preserve<LiveVariablesAnalysis>();
+  PA.preserve<DominatorTreeAnalysis>();
+  PA.preserve<MachineLoopAnalysis>();
+  PA.preserveSet<CFGAnalyses>();
+  return PA;
+}
 
+bool SIOptimizeVGPRLiveRange::run(MachineFunction &MF) {
   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
   TII = ST.getInstrInfo();
   TRI = &TII->getRegisterInfo();
-  MDT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
-  Loops = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
-  LV = &getAnalysis<LiveVariablesWrapperPass>().getLV();
   MRI = &MF.getRegInfo();
 
-  if (skipFunction(MF.getFunction()))
-    return false;
-
   bool MadeChange = false;
 
   // TODO: we need to think about the order of visiting the blocks to get

llvm/lib/Target/AMDGPU/SIOptimizeVGPRLiveRange.h

@@ -0,0 +1,33 @@
+//===- SIOptimizeVGPRLiveRange.h --------------------------------*- C++- *-===//
+//
+// 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_SIOPTIMIZEVGPRLIVERANGE_H
+#define LLVM_LIB_TARGET_AMDGPU_SIOPTIMIZEVGPRLIVERANGE_H
+
+#include "llvm/CodeGen/MachinePassManager.h"
+
+namespace llvm {
+class SIOptimizeVGPRLiveRangePass
+    : public PassInfoMixin<SIOptimizeVGPRLiveRangePass> {
+public:
+  PreservedAnalyses run(MachineFunction &MF,
+                        MachineFunctionAnalysisManager &MFAM);
+
+  MachineFunctionProperties getRequiredProperties() const {
+    return MachineFunctionProperties().set(
+        MachineFunctionProperties::Property::IsSSA);
+  }
+
+  MachineFunctionProperties getClearedProperties() const {
+    return MachineFunctionProperties().set(
+        MachineFunctionProperties::Property::NoPHIs);
+  }
+};
+} // namespace llvm
+
+#endif // LLVM_LIB_TARGET_AMDGPU_SIOPTIMIZEVGPRLIVERANGE_H

llvm/test/CodeGen/AMDGPU/opt-vgpr-live-range-verifier-error.mir

@@ -1,4 +1,5 @@
 # RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -verify-machineinstrs -run-pass=si-opt-vgpr-liverange -o - %s | FileCheck %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -passes=si-opt-vgpr-liverange -o - %s | FileCheck %s
 #
 # This is a very rare case which comes from llvm-reduce. The SI_IF/SI_ELSE usage is quite different from normal.
 #

llvm/test/CodeGen/AMDGPU/si-opt-vgpr-liverange-bug-deadlanes.mir

@@ -1,5 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 2
 # RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs -run-pass=si-opt-vgpr-liverange -o - %s | FileCheck -check-prefixes=CHECK %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -passes=si-opt-vgpr-liverange -o - %s | FileCheck -check-prefixes=CHECK %s
 
 # Tests a case that used to assert in SIOptimizeVGPRLiveRange when trying to optimize %3 which still appears
 # (though in an undef operand) in the REG_SEQUENCE of the "endif block". This undef pattern was caused by

llvm/test/CodeGen/AMDGPU/si-optimize-vgpr-live-range-dbg-instr.mir

@@ -1,4 +1,5 @@
 # RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -run-pass=si-opt-vgpr-liverange  %s -o - | FileCheck -check-prefix=GCN %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -passes=si-opt-vgpr-liverange  %s -o - | FileCheck -check-prefix=GCN %s
 
 # SIOptimizeVGPRLiveRange shouldn't try to modify use of %5 in DBG_VALUE_LIST