[AMDGPU][AMDGPUDemoteSCCBranchToExecz] create new pass (boilerplate only)

Author: jmmartinez

llvm/lib/Target/AMDGPU/AMDGPU.h

@@ -373,6 +373,9 @@ extern char &AMDGPUCodeGenPrepareID;
 void initializeAMDGPURemoveIncompatibleFunctionsPass(PassRegistry &);
 extern char &AMDGPURemoveIncompatibleFunctionsID;
 
+void initializeAMDGPUDemoteSCCBranchToExeczLegacyPass(PassRegistry &);
+extern char &AMDGPUDemoteSCCBranchToExeczLegacyID;
+
 void initializeAMDGPULateCodeGenPrepareLegacyPass(PassRegistry &);
 extern char &AMDGPULateCodeGenPrepareLegacyID;
 

llvm/lib/Target/AMDGPU/AMDGPUDemoteSCCBranchToExecz.cpp

@@ -0,0 +1,54 @@
+#include "llvm/CodeGen/MachineFunctionPass.h"
+
+#include "AMDGPU.h"
+#include "AMDGPUDemoteSCCBranchToExecz.h"
+
+using namespace llvm;
+
+namespace {
+#define DEBUG_TYPE "amdgpu-demote-scc-to-execz"
+const char PassName[] = "AMDGPU if conversion";
+
+class AMDGPUDemoteSCCBranchToExecz {
+public:
+  AMDGPUDemoteSCCBranchToExecz() = default;
+
+  bool run() { return false; }
+};
+
+class AMDGPUDemoteSCCBranchToExeczLegacy : public MachineFunctionPass {
+public:
+  static char ID;
+
+  AMDGPUDemoteSCCBranchToExeczLegacy() : MachineFunctionPass(ID) {}
+
+  bool runOnMachineFunction(MachineFunction &MF) override {
+    AMDGPUDemoteSCCBranchToExecz IfCvt{};
+    return IfCvt.run();
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  StringRef getPassName() const override { return PassName; }
+};
+
+char AMDGPUDemoteSCCBranchToExeczLegacy::ID = 0;
+
+} // namespace
+
+PreservedAnalyses llvm::AMDGPUDemoteSCCBranchToExeczPass::run(
+    MachineFunction &MF, MachineFunctionAnalysisManager &MFAM) {
+  AMDGPUDemoteSCCBranchToExecz IfCvt{};
+  if (!IfCvt.run())
+    return PreservedAnalyses::all();
+  return PreservedAnalyses::none();
+}
+
+char &llvm::AMDGPUDemoteSCCBranchToExeczLegacyID =
+    AMDGPUDemoteSCCBranchToExeczLegacy::ID;
+INITIALIZE_PASS_BEGIN(AMDGPUDemoteSCCBranchToExeczLegacy, DEBUG_TYPE, PassName,
+                      false, false)
+INITIALIZE_PASS_END(AMDGPUDemoteSCCBranchToExeczLegacy, DEBUG_TYPE, PassName,
+                    false, false)

llvm/lib/Target/AMDGPU/AMDGPUDemoteSCCBranchToExecz.h

@@ -0,0 +1,31 @@
+//===- AMDGPURDemoteSCCBranchToExecz.h --- demote s_cbranch_scc -*- 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
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Pass used to demote s_cbranch_scc0/1 branches to s_cbranch_execz
+/// branches. These can be later removed by SIPreEmitPeephole.
+///
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUDEMOTESCCBRANCHTOEXECZ_H
+#define LLVM_LIB_TARGET_AMDGPU_AMDGPUDEMOTESCCBRANCHTOEXECZ_H
+
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/IR/PassManager.h"
+
+namespace llvm {
+class AMDGPUDemoteSCCBranchToExeczPass
+    : public PassInfoMixin<AMDGPUDemoteSCCBranchToExeczPass> {
+public:
+  AMDGPUDemoteSCCBranchToExeczPass() = default;
+  PreservedAnalyses run(MachineFunction &MF,
+                        MachineFunctionAnalysisManager &MFAM);
+};
+} // namespace llvm
+
+#endif

llvm/lib/Target/AMDGPU/AMDGPUPassRegistry.def

@@ -95,6 +95,7 @@ FUNCTION_PASS_WITH_PARAMS(
 #define MACHINE_FUNCTION_PASS(NAME, CREATE_PASS)
 #endif
 MACHINE_FUNCTION_PASS("amdgpu-isel", AMDGPUISelDAGToDAGPass(*this))
+MACHINE_FUNCTION_PASS("amdgpu-demote-scc-to-execz", AMDGPUDemoteSCCBranchToExeczPass())
 MACHINE_FUNCTION_PASS("si-fix-sgpr-copies", SIFixSGPRCopiesPass())
 MACHINE_FUNCTION_PASS("si-i1-copies", SILowerI1CopiesPass())
 MACHINE_FUNCTION_PASS("si-fold-operands", SIFoldOperandsPass());

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -18,6 +18,7 @@
 #include "AMDGPU.h"
 #include "AMDGPUAliasAnalysis.h"
 #include "AMDGPUCtorDtorLowering.h"
+#include "AMDGPUDemoteSCCBranchToExecz.h"
 #include "AMDGPUExportClustering.h"
 #include "AMDGPUIGroupLP.h"
 #include "AMDGPUISelDAGToDAG.h"
@@ -498,6 +499,7 @@ extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
   initializeAMDGPURewriteUndefForPHILegacyPass(*PR);
   initializeAMDGPUUnifyMetadataPass(*PR);
   initializeSIAnnotateControlFlowLegacyPass(*PR);
+  initializeAMDGPUDemoteSCCBranchToExeczLegacyPass(*PR);
   initializeAMDGPUInsertDelayAluPass(*PR);
   initializeSIInsertHardClausesPass(*PR);
   initializeSIInsertWaitcntsPass(*PR);
@@ -1336,7 +1338,7 @@ void GCNPassConfig::addMachineSSAOptimization() {
 bool GCNPassConfig::addILPOpts() {
   if (EnableEarlyIfConversion)
     addPass(&EarlyIfConverterID);
-
+  addPass(&AMDGPUDemoteSCCBranchToExeczLegacyID);
   TargetPassConfig::addILPOpts();
   return false;
 }

llvm/lib/Target/AMDGPU/CMakeLists.txt

@@ -59,6 +59,7 @@ add_llvm_target(AMDGPUCodeGen
   AMDGPUGlobalISelDivergenceLowering.cpp
   AMDGPUGlobalISelUtils.cpp
   AMDGPUHSAMetadataStreamer.cpp
+  AMDGPUDemoteSCCBranchToExecz.cpp
   AMDGPUInsertDelayAlu.cpp
   AMDGPUInstCombineIntrinsic.cpp
   AMDGPUInstrInfo.cpp

llvm/test/CodeGen/AMDGPU/llc-pipeline.ll

@@ -315,6 +315,7 @@
 ; GCN-O1-NEXT:        Merge disjoint stack slots
 ; GCN-O1-NEXT:        Local Stack Slot Allocation
 ; GCN-O1-NEXT:        Remove dead machine instructions
+; GCN-O1-NEXT:        AMDGPU s_cbranch_scc to s_cbranch_execz conversion
 ; GCN-O1-NEXT:        MachineDominator Tree Construction
 ; GCN-O1-NEXT:        Machine Natural Loop Construction
 ; GCN-O1-NEXT:        Machine Block Frequency Analysis
@@ -617,6 +618,7 @@
 ; GCN-O1-OPTS-NEXT:        Merge disjoint stack slots
 ; GCN-O1-OPTS-NEXT:        Local Stack Slot Allocation
 ; GCN-O1-OPTS-NEXT:        Remove dead machine instructions
+; GCN-O1-OPTS-NEXT:        AMDGPU s_cbranch_scc to s_cbranch_execz conversion
 ; GCN-O1-OPTS-NEXT:        MachineDominator Tree Construction
 ; GCN-O1-OPTS-NEXT:        Machine Natural Loop Construction
 ; GCN-O1-OPTS-NEXT:        Machine Block Frequency Analysis
@@ -932,6 +934,7 @@
 ; GCN-O2-NEXT:        Merge disjoint stack slots
 ; GCN-O2-NEXT:        Local Stack Slot Allocation
 ; GCN-O2-NEXT:        Remove dead machine instructions
+; GCN-O2-NEXT:        AMDGPU s_cbranch_scc to s_cbranch_execz conversion
 ; GCN-O2-NEXT:        MachineDominator Tree Construction
 ; GCN-O2-NEXT:        Machine Natural Loop Construction
 ; GCN-O2-NEXT:        Machine Block Frequency Analysis
@@ -1260,6 +1263,7 @@
 ; GCN-O3-NEXT:        Merge disjoint stack slots
 ; GCN-O3-NEXT:        Local Stack Slot Allocation
 ; GCN-O3-NEXT:        Remove dead machine instructions
+; GCN-O3-NEXT:        AMDGPU s_cbranch_scc to s_cbranch_execz conversion
 ; GCN-O3-NEXT:        MachineDominator Tree Construction
 ; GCN-O3-NEXT:        Machine Natural Loop Construction
 ; GCN-O3-NEXT:        Machine Block Frequency Analysis