Add Patmos Stack Cache Optimization

.github/actions/build-test-all/action.yml

@@ -167,10 +167,10 @@ runs:
     - name: Download Newlib
       if: inputs.enable-package == 'true'
       env:
-        NEWLIB_COMMIT: 4a74fe0ca51e6f462999446f8adbfd8f6b8c91e5
+        NEWLIB_COMMIT: 79b0b16dfc224c2d63b31e5c314bc2a5c430f01e
       shell: bash
       run: |
-        git clone https://github.com/t-crest/patmos-newlib ${{env.NEWLIB_PATH}}
+        git clone https://github.com/R3freshReload/patmos-newlib.git ${{env.NEWLIB_PATH}}
         cd ${{env.NEWLIB_PATH}}
         git checkout $NEWLIB_COMMIT
     # Builds newlib to Patmos assembly (instead of bitcode)

.github/actions/install-dependencies/action.yml

@@ -8,7 +8,7 @@ runs:
       name: Install GNU-tar as 'tar'
       shell: bash
       run: |
-        brew install gnu-tar
+        brew install gnu-tar texinfo
         # Set gtar to be available as 'tar'
         echo "/usr/local/opt/gnu-tar/libexec/gnubin" >> $GITHUB_PATH
     - name: Create Installation Directory

.github/workflows/patmos-ci.yml

@@ -69,10 +69,16 @@ jobs:
           - platform: ubuntu-20.04
             cc: gcc
             cxx: g++
-          - platform: macos-latest
+          - platform: macos-latest # ARM
             cc: clang
             cxx: clang++
-          - platform: macos-12
+          - platform: macos-14 # ARM
+            cc: clang
+            cxx: clang++
+          - platform: macos-13 # x86
+            cc: clang
+            cxx: clang++
+          - platform: macos-12 # x86
             cc: clang
             cxx: clang++
 
@@ -105,10 +111,16 @@ jobs:
           - platform: ubuntu-20.04
             cc: gcc
             cxx: g++
-          - platform: macos-latest
+          - platform: macos-latest # ARM
             cc: clang
             cxx: clang++
-          - platform: macos-12
+          - platform: macos-14 # ARM
+            cc: clang
+            cxx: clang++
+          - platform: macos-13 # x86
+            cc: clang
+            cxx: clang++
+          - platform: macos-12 # x86
             cc: clang
             cxx: clang++
     needs: [package-patmos]     

llvm/cmake/platforms/Patmos.cmake

@@ -6,6 +6,8 @@ if (${DUMP_MACHINE} MATCHES "x86_64-linux-gnu")
 	set( TARGET_TRIPLE "x86_64-linux-gnu")
 elseif(${DUMP_MACHINE} MATCHES "x86_64-apple-darwin.*")
 	set( TARGET_TRIPLE "x86_64-apple-darwin")
+elseif(${DUMP_MACHINE} MATCHES "arm64-apple-darwin.*")
+	set( TARGET_TRIPLE "arm64-apple-darwin")
 else()
 	message(FATAL_ERROR "Unsupported platform for packaging")
 endif()

llvm/include/llvm/Support/Signals.h

@@ -15,6 +15,8 @@
 #define LLVM_SUPPORT_SIGNALS_H
 
 #include <string>
+#include <cstdint>
+
 
 namespace llvm {
 class StringRef;

llvm/lib/Target/Patmos/CMakeLists.txt

@@ -26,6 +26,7 @@ add_llvm_target(PatmosCodeGen
   PatmosSelectionDAGInfo.cpp
   PatmosAsmPrinter.cpp
   PatmosMCInstLower.cpp
+  PatmosStackCachePromotion.cpp
   PatmosDelaySlotFiller.cpp
   PatmosFunctionSplitter.cpp
   PatmosDelaySlotKiller.cpp

llvm/lib/Target/Patmos/PMLExport.cpp

@@ -61,7 +61,7 @@ template <> struct FlowGraphTrait<const BasicBlock> {
     return llvm::succ_end(BB);
   }
   static StringValue getName(const BasicBlock *BB) {
-    return BB->getName().str();
+    return BB->getNameOrAsOperand();
   }
 };
 template <> struct FlowGraphTrait<MachineBasicBlock> {
@@ -221,10 +221,10 @@ yaml::FlowFact<yaml::StringValue> *PMLBitcodeExport::createLoopFact(const BasicB
 
   auto *FF = new yaml::FlowFact<yaml::StringValue>(yaml::level_bitcode);
 
-  FF->setLoopScope(Fn->getName().str(), BB->getName().str());
+  FF->setLoopScope(Fn->getName().str(), BB->getNameOrAsOperand());
 
   yaml::ProgramPoint *Block =
-    yaml::ProgramPoint::CreateBlock(Fn->getName().str(), BB->getName().str());
+    yaml::ProgramPoint::CreateBlock(Fn->getName().str(), BB->getNameOrAsOperand());
 
   FF->addTermLHS(Block, 1LL);
   FF->RHS = RHS;
@@ -247,27 +247,24 @@ void PMLBitcodeExport::serialize(MachineFunction &MF)
   yaml::BitcodeFunction *F = new yaml::BitcodeFunction(Fn.getName().str());
   F->Level = yaml::level_bitcode;
   yaml::BitcodeBlock *B;
-  for (Function::const_iterator BI = Fn.begin(), BE = Fn.end(); BI != BE;
+  for (auto BI = Fn.begin(), BE = Fn.end(); BI != BE;
       ++BI) {
-    if (BI->getName().empty()) {
-      llvm::errs() << "warning: unnamed bit-code BB in PML export\n";
-    }
-    B = F->addBlock(new yaml::BitcodeBlock(BI->getName().str()));
+    B = F->addBlock(new yaml::BitcodeBlock(BI->getNameOrAsOperand()));
 
     Loop *Loop = LI.getLoopFor(&*BI);
     while (Loop) {
-      B->Loops.push_back(Loop->getHeader()->getName().str());
+      B->Loops.push_back(Loop->getHeader()->getNameOrAsOperand());
       Loop = Loop->getParentLoop();
     }
 
     /// B->MapsTo = (maybe C-source debug info?)
-    for (const_pred_iterator PI = pred_begin(&*BI), PE = pred_end(&*BI);
+    for (auto PI = pred_begin(&*BI), PE = pred_end(&*BI);
         PI != PE; ++PI) {
-      B->Predecessors.push_back((*PI)->getName().str());
+      B->Predecessors.push_back((*PI)->getNameOrAsOperand());
     }
     for (auto SI = succ_begin(&*BI), SE = succ_end(&*BI);
         SI != SE; ++SI) {
-      B->Successors.push_back((*SI)->getName().str());
+      B->Successors.push_back((*SI)->getNameOrAsOperand());
     }
 
     unsigned Index = 0;
@@ -849,7 +846,7 @@ void addProgressNodes(yaml::RelationGraph *RG,
           }
           else {
             RN = RG->addNode(yaml::rnt_progress);
-            RN->setSrcBlock(IQI->first->getName().str());
+            RN->setSrcBlock(IQI->first->getNameOrAsOperand());
             RN->setDstBlock(MQI->first->getNumber());
             RMap.insert(std::make_pair(PNID, RN));
             RTodo.push_back(std::make_pair(PNID, RN));
@@ -935,7 +932,7 @@ void PMLRelationGraphExport::serialize(MachineFunction &MF)
     auto *SrcScope = new yaml::RelationScope<yaml::StringValue>(
         BF.getName().str(), yaml::level_bitcode);
     RG = new yaml::RelationGraph(SrcScope, DstScope);
-    RG->getEntryNode()->setSrcBlock(BF.getEntryBlock().getName().str());
+    RG->getEntryNode()->setSrcBlock(BF.getEntryBlock().getNameOrAsOperand());
     RG->getEntryNode()->setDstBlock(MF.front().getNumber());
     UnmatchedEvents.clear();
 
@@ -974,7 +971,7 @@ void PMLRelationGraphExport::serialize(MachineFunction &MF)
       // MBB, resp.), which results in new src/dst nodes being created, and
       // two bitcode and machinecode-level maps from events to a list of
       // (bitcode/machine block, list of RG predecessor blocks) pairs
-      expandProgressNode(RG, RN, yaml::rnt_src, [](auto *node, auto* block){node->setSrcBlock(block->getName().str());}, IBB, IEventMap, IEvents);
+      expandProgressNode(RG, RN, yaml::rnt_src, [](auto *node, auto* block){node->setSrcBlock(block->getNameOrAsOperand());}, IBB, IEventMap, IEvents);
       expandProgressNode(RG, RN, yaml::rnt_dst, [](auto *node, auto* block){node->setDstBlock(block->getNumber());},  MBB, MEventMap, MEvents);
 
       // For each event and corresponding bitcode list IList and machinecode

llvm/lib/Target/Patmos/Patmos.h

@@ -41,6 +41,7 @@ namespace llvm {
   FunctionPass *createPatmosSPBundlingPass(const PatmosTargetMachine &tm);
   FunctionPass *createOppositePredicateCompensationPass(const PatmosTargetMachine &tm);
   FunctionPass *createDataCacheAccessEliminationPass(const PatmosTargetMachine &tm);
+  FunctionPass *createPatmosStackCachePromotionPass(const PatmosTargetMachine &tm);
   FunctionPass *createMemoryAccessNormalizationPass(const PatmosTargetMachine &tm);
   FunctionPass *createPatmosSPReducePass(const PatmosTargetMachine &tm);
   FunctionPass *createPreRegallocReduce(const PatmosTargetMachine &tm);

llvm/lib/Target/Patmos/PatmosExport.cpp

@@ -367,7 +367,7 @@ namespace llvm {
           // FIXME
           // we don't add it immediately to PMF, as we only support
           // arguments in registers at this point
-          std::string ArgName = ("%" + I->getName()).str();
+          std::string ArgName = I->getNameOrAsOperand();
           yaml::Argument *Arg = new yaml::Argument(ArgName, FAIdx);
           bool allInRegs = true;
 

llvm/lib/Target/Patmos/PatmosFrameLowering.cpp

@@ -128,6 +128,12 @@ void PatmosFrameLowering::assignFIsToStackCache(MachineFunction &MF,
     // find all spill slots and locations for callee saved registers
     if (MFI.isSpillSlotObjectIndex(FI))
       SCFIs[FI] = true;
+
+
+    const std::vector<int> &StackCacheAllocatable = PMFI.getStackCacheAnalysisFIs();
+    for(unsigned i=0; i<StackCacheAllocatable.size(); i++) {
+      SCFIs[StackCacheAllocatable[i]] = true;
+    }
   }
 }
 
@@ -330,7 +336,6 @@ void PatmosFrameLowering::emitPrologue(MachineFunction &MF, MachineBasicBlock &M
       MI->setFlag(MachineInstr::FrameSetup);
     }
   }
-
 }
 
 void PatmosFrameLowering::emitEpilogue(MachineFunction &MF,

llvm/lib/Target/Patmos/PatmosInstrInfo.cpp

@@ -575,7 +575,6 @@ PatmosII::MemType PatmosInstrInfo::getMemType(const MachineInstr &MI) {
     case  SWM: case  SHM: case  SBM:
       return PatmosII::MEM_M;
     default:
-    	MI.dump();
     	llvm_unreachable("Unexpected memory access instruction!");
   }
 

llvm/lib/Target/Patmos/PatmosMachineFunctionInfo.h

@@ -101,6 +101,9 @@ class PatmosMachineFunctionInfo : public MachineFunctionInfo {
   /// | LoopCnts | S0Spills | ExcessSpills |
   std::vector<int> SinglePathFIs;
 
+  /// FIs determined to be movable to stack cache
+  std::vector<int> StackCacheAnalysisFIs;
+
   // Index to the SinglePathFIs where the S0 spill slots start
   unsigned SPS0SpillOffset;
 
@@ -259,6 +262,21 @@ class PatmosMachineFunctionInfo : public MachineFunctionInfo {
     return SinglePathFIs.size();
   }
 
+  const std::vector<int>& getStackCacheAnalysisFIs(void) const {
+    return StackCacheAnalysisFIs;
+  }
+
+  void removeStackCacheAnalysisFIs(int fi) {
+    std::vector<int>::iterator position = std::find(StackCacheAnalysisFIs.begin(), StackCacheAnalysisFIs.end(), fi);
+
+    if (position != StackCacheAnalysisFIs.end()) // == myVector.end() means the element was not found
+      StackCacheAnalysisFIs.erase(position);
+  }
+
+  void addStackCacheAnalysisFI(int fi) {
+    StackCacheAnalysisFIs.push_back(fi);
+  }
+
   PatmosAnalysisInfo &getAnalysisInfo() { return AnalysisInfo; }
 
   const PatmosAnalysisInfo &getAnalysisInfo() const { return AnalysisInfo; }