Update CodeGenOpt

There is a flag-day change around CodeGen enums in llvm/llvm-project#66295.

Update the code to use both old and new schemes depending on the LLVM version used.

Change enum to uint32_t in some places to reduce the number of include guards.

lgc/include/lgc/patch/Patch.h

@@ -53,8 +53,7 @@ class Patch {
   virtual ~Patch() {}
 
   static void addPasses(PipelineState *pipelineState, lgc::PassManager &passMgr, llvm::Timer *patchTimer,
-                        llvm::Timer *optTimer, Pipeline::CheckShaderCacheFunc checkShaderCacheFunc,
-                        llvm::CodeGenOpt::Level optLevel);
+                        llvm::Timer *optTimer, Pipeline::CheckShaderCacheFunc checkShaderCacheFunc, uint32_t optLevel);
 
   // Register all the patching passes into the given pass manager
   static void registerPasses(lgc::PassManager &passMgr);
@@ -65,7 +64,7 @@ class Patch {
   static llvm::GlobalVariable *getLdsVariable(PipelineState *pipelineState, llvm::Module *module);
 
 protected:
-  static void addOptimizationPasses(lgc::PassManager &passMgr, llvm::CodeGenOpt::Level optLevel);
+  static void addOptimizationPasses(lgc::PassManager &passMgr, uint32_t optLevel);
 
   void init(llvm::Module *module);
 

lgc/interface/lgc/LgcContext.h

@@ -84,7 +84,15 @@ class LgcContext {
   // @param gpuName : LLVM GPU name (e.g. "gfx900"); empty to use -mcpu option setting
   // @param optLevel : LLVM optimization level used to initialize target machine
   static std::unique_ptr<llvm::TargetMachine> createTargetMachine(llvm::StringRef gpuName,
-                                                                  llvm::CodeGenOpt::Level optLevel);
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+                                                                  // Old version of the code
+                                                                  llvm::CodeGenOpt::Level optLevel
+#else
+                                                                  // New version of the code (also handles unknown
+                                                                  // version, which we treat as latest)
+                                                                  llvm::CodeGenOptLevel optLevel
+#endif
+  );
 
   // Create the LgcContext.
   //
@@ -129,11 +137,21 @@ class LgcContext {
   // Adds target passes to pass manager, depending on "-filetype" and "-emit-llvm" options
   void addTargetPasses(lgc::LegacyPassManager &passMgr, llvm::Timer *codeGenTimer, llvm::raw_pwrite_stream &outStream);
 
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+  // Old version of the code
   // Returns the optimization level for the context.
   llvm::CodeGenOpt::Level getOptimizationLevel() const;
 
   // Returns the optimization level used for context initialization.
   llvm::CodeGenOpt::Level getInitialOptimizationLevel() const { return m_initialOptLevel; }
+#else
+  // New version of the code (also handles unknown version, which we treat as latest)
+  // Returns the optimization level for the context.
+  llvm::CodeGenOptLevel getOptimizationLevel() const;
+
+  // Returns the optimization level used for context initialization.
+  llvm::CodeGenOptLevel getInitialOptimizationLevel() const { return m_initialOptLevel; }
+#endif
 
   // Utility method to create a start/stop timer pass
   static llvm::ModulePass *createStartStopTimer(llvm::Timer *timer, bool starting);
@@ -167,7 +185,13 @@ class LgcContext {
   TargetInfo *m_targetInfo = nullptr;                // Target info
   unsigned m_palAbiVersion = 0xFFFFFFFF;             // PAL pipeline ABI version to compile for
   PassManagerCache *m_passManagerCache = nullptr;    // Pass manager cache and creator
-  llvm::CodeGenOpt::Level m_initialOptLevel;         // Optimization level at initialization
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+  // Old version of the code
+  llvm::CodeGenOpt::Level m_initialOptLevel; // Optimization level at initialization
+#else
+  // New version of the code (also handles unknown version, which we treat as latest)
+  llvm::CodeGenOptLevel m_initialOptLevel; // Optimization level at initialization
+#endif
 };
 
 } // namespace lgc

lgc/patch/Patch.cpp

@@ -117,7 +117,7 @@ namespace lgc {
 // @param optLevel : The optimization level uses to adjust the aggressiveness of
 //                   passes and which passes to add.
 void Patch::addPasses(PipelineState *pipelineState, lgc::PassManager &passMgr, Timer *patchTimer, Timer *optTimer,
-                      Pipeline::CheckShaderCacheFunc checkShaderCacheFunc, CodeGenOpt::Level optLevel) {
+                      Pipeline::CheckShaderCacheFunc checkShaderCacheFunc, uint32_t optLevel) {
   // Start timer for patching passes.
   if (patchTimer)
     LgcContext::createAndAddStartStopTimer(passMgr, patchTimer, true);
@@ -340,7 +340,7 @@ void Patch::registerPasses(PassBuilder &passBuilder) {
 // @param [in/out] passMgr : Pass manager to add passes to
 // @param optLevel : The optimization level uses to adjust the aggressiveness of
 //                   passes and which passes to add.
-void Patch::addOptimizationPasses(lgc::PassManager &passMgr, CodeGenOpt::Level optLevel) {
+void Patch::addOptimizationPasses(lgc::PassManager &passMgr, uint32_t optLevel) {
   LLPC_OUTS("PassManager optimization level = " << optLevel << "\n");
 
   passMgr.addPass(ForceFunctionAttrsPass());

lgc/state/Compiler.cpp

@@ -219,7 +219,7 @@ bool PipelineState::generate(Module *pipelineModule, raw_pwrite_stream &outStrea
   } else {
     // Patching.
     Patch::addPasses(this, *passMgr, patchTimer, optTimer, std::move(checkShaderCacheFunc),
-                     getLgcContext()->getOptimizationLevel());
+                     static_cast<uint32_t>(getLgcContext()->getOptimizationLevel()));
 
     // Add pass to clear pipeline state from IR
     passMgr->addPass(PipelineStateClearer());

lgc/state/LgcContext.cpp

@@ -77,12 +77,23 @@ static cl::opt<bool> EmitLgc("emit-lgc", cl::desc("Emit LLVM assembly suitable f
 static cl::opt<bool> ShowEncoding("show-encoding", cl::desc("Show instruction encodings"), cl::init(false));
 
 // -opt: Override the optimization level passed in to LGC with the given one.
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+// Old version of the code
 static cl::opt<CodeGenOpt::Level> OptLevel("opt", cl::desc("Set the optimization level for LGC:"),
                                            cl::init(CodeGenOpt::Default),
                                            values(clEnumValN(CodeGenOpt::None, "none", "no optimizations"),
                                                   clEnumValN(CodeGenOpt::Less, "quick", "quick compilation time"),
                                                   clEnumValN(CodeGenOpt::Default, "default", "default optimizations"),
                                                   clEnumValN(CodeGenOpt::Aggressive, "fast", "fast execution time")));
+#else
+// New version of the code (also handles unknown version, which we treat as latest)
+static cl::opt<CodeGenOptLevel>
+    OptLevel("opt", cl::desc("Set the optimization level for LGC:"), cl::init(CodeGenOptLevel::Default),
+             values(clEnumValN(CodeGenOptLevel::None, "none", "no optimizations"),
+                    clEnumValN(CodeGenOptLevel::Less, "quick", "quick compilation time"),
+                    clEnumValN(CodeGenOptLevel::Default, "default", "default optimizations"),
+                    clEnumValN(CodeGenOptLevel::Aggressive, "fast", "fast execution time")));
+#endif
 
 // =====================================================================================================================
 // Set default for a command-line option, but only if command-line processing has not happened yet, or did not see
@@ -216,7 +227,14 @@ bool LgcContext::isGpuNameValid(llvm::StringRef gpuName) {
 //
 // @param gpuName : LLVM GPU name (e.g. "gfx900"); empty to use -mcpu option setting
 // @param optLevel : LLVM optimization level used to initialize target machine
-std::unique_ptr<TargetMachine> LgcContext::createTargetMachine(StringRef gpuName, CodeGenOpt::Level optLevel) {
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+// Old version of the code
+std::unique_ptr<TargetMachine> LgcContext::createTargetMachine(StringRef gpuName, CodeGenOpt::Level optLevel)
+#else
+// New version of the code (also handles unknown version, which we treat as latest)
+std::unique_ptr<TargetMachine> LgcContext::createTargetMachine(StringRef gpuName, CodeGenOptLevel optLevel)
+#endif
+{
   assert(Initialized && "Must call LgcContext::initialize before LgcContext::createTargetMachine");
 
   std::string mcpuName = codegen::getMCPU(); // -mcpu setting from llvm/CodeGen/CommandFlags.h
@@ -245,7 +263,7 @@ std::unique_ptr<TargetMachine> LgcContext::createTargetMachine(StringRef gpuName
   if (OptLevel.getPosition() != 0)
     optLevel = OptLevel;
 
-  LLPC_OUTS("TargetMachine optimization level = " << optLevel << "\n");
+  LLPC_OUTS("TargetMachine optimization level = " << static_cast<uint32_t>(optLevel) << "\n");
 
   return std::unique_ptr<TargetMachine>(target->createTargetMachine(triple, gpuName, "", targetOpts, {}, {}, optLevel));
 }
@@ -358,7 +376,14 @@ void LgcContext::addTargetPasses(lgc::LegacyPassManager &passMgr, Timer *codeGen
     passMgr.add(createStartStopTimer(codeGenTimer, false));
 }
 
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+// Old version of the code
 llvm::CodeGenOpt::Level LgcContext::getOptimizationLevel() const {
+#else
+// New version of the code (also handles unknown version, which we treat as latest)
+llvm::CodeGenOptLevel LgcContext::getOptimizationLevel() const {
+#endif
+
   return m_targetMachine->getOptLevel();
 }
 

lgc/tool/lgc/lgc.cpp

@@ -159,6 +159,8 @@ static bool runPassPipeline(Pipeline &pipeline, Module &module, raw_pwrite_strea
   passMgr->addPass(VerifierPass());
   passMgr->addPass(PipelineStateRecorder());
 
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+  // Old version of the code
   switch (codegen::getFileType()) {
   case CGFT_AssemblyFile:
     passMgr->addPass(PrintModulePass(outStream));
@@ -169,6 +171,19 @@ static bool runPassPipeline(Pipeline &pipeline, Module &module, raw_pwrite_strea
   case CGFT_Null:
     break;
   }
+#else
+  // New version of the code (also handles unknown version, which we treat as latest)
+  switch (codegen::getFileType()) {
+  case CodeGenFileType::AssemblyFile:
+    passMgr->addPass(PrintModulePass(outStream));
+    break;
+  case CodeGenFileType::ObjectFile:
+    passMgr->addPass(BitcodeWriterPass(outStream));
+    break;
+  case CodeGenFileType::Null:
+    break;
+  }
+#endif
 
   passMgr->run(module);
   return true;
@@ -239,11 +254,23 @@ int main(int argc, char **argv) {
     assert(optIterator != cl::getRegisteredOptions().end());
     cl::Option *opt = optIterator->second;
     if (opt->getNumOccurrences() == 0)
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+      // Old version of the code
       *static_cast<cl::opt<CodeGenFileType> *>(opt) = CGFT_AssemblyFile;
+#else
+      // New version of the code (also handles unknown version, which we treat as latest)
+      *static_cast<cl::opt<CodeGenFileType> *>(opt) = CodeGenFileType::AssemblyFile;
+#endif
   }
 
   // Create the LgcContext.
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+  // Old version of the code
   std::unique_ptr<TargetMachine> targetMachine(LgcContext::createTargetMachine(gpuName, CodeGenOpt::Level::Default));
+#else
+  // New version of the code (also handles unknown version, which we treat as latest)
+  std::unique_ptr<TargetMachine> targetMachine(LgcContext::createTargetMachine(gpuName, CodeGenOptLevel::Default));
+#endif
   if (!targetMachine) {
     errs() << progName << ": GPU type '" << gpuName << "' not recognized\n";
     return 1;

llpc/context/llpcContext.cpp

@@ -105,12 +105,15 @@ LgcContext *Context::getLgcContext() {
   return &*m_builderContext;
 }
 
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+// Old version of the code
 // =====================================================================================================================
 // Get optimization level. Also resets what getLastOptimizationLevel() returns.
 //
 // @returns: the optimization level for the context.
 CodeGenOpt::Level Context::getOptimizationLevel() {
-  uint32_t optLevel = CodeGenOpt::Level::Default;
+  uint32_t optLevel = static_cast<uint32_t>(CodeGenOpt::Level::Default);
+
   optLevel = getPipelineContext()->getPipelineOptions()->optimizationLevel;
   if (optLevel > 3)
     optLevel = 3;
@@ -126,6 +129,33 @@ CodeGenOpt::Level Context::getLastOptimizationLevel() const {
   return *m_lastOptLevel;
 }
 
+#else
+// New version of the code (also handles unknown version, which we treat as latest)
+
+// =====================================================================================================================
+// Get optimization level. Also resets what getLastOptimizationLevel() returns.
+//
+// @returns: the optimization level for the context.
+CodeGenOptLevel Context::getOptimizationLevel() {
+  uint32_t optLevel = static_cast<uint32_t>(CodeGenOptLevel::Default);
+
+  optLevel = getPipelineContext()->getPipelineOptions()->optimizationLevel;
+  if (optLevel > 3)
+    optLevel = 3;
+  else if (optLevel == 0) // Workaround for noopt bugs in the AMDGPU backend in LLVM.
+    optLevel = 1;
+  m_lastOptLevel = CodeGenOptLevel(optLevel);
+  return *m_lastOptLevel;
+}
+
+// =====================================================================================================================
+// Get the optimization level returned by the last getOptimizationLevel().
+CodeGenOptLevel Context::getLastOptimizationLevel() const {
+  return *m_lastOptLevel;
+}
+
+#endif
+
 // =====================================================================================================================
 // Loads library from external LLVM library.
 //

llpc/context/llpcContext.h

@@ -84,8 +84,15 @@ class Context : public llvm::LLVMContext {
   // Get (create if necessary) LgcContext
   lgc::LgcContext *getLgcContext();
 
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+  // Old version of the code
   llvm::CodeGenOpt::Level getOptimizationLevel();
   llvm::CodeGenOpt::Level getLastOptimizationLevel() const;
+#else
+  // New version of the code (also handles unknown version, which we treat as latest)
+  llvm::CodeGenOptLevel getOptimizationLevel();
+  llvm::CodeGenOptLevel getLastOptimizationLevel() const;
+#endif
 
   std::unique_ptr<llvm::Module> loadLibrary(const BinaryData *lib);
 
@@ -129,7 +136,14 @@ class Context : public llvm::LLVMContext {
   std::unique_ptr<llvm::TargetMachine> m_targetMachine; // Target machine for LGC context
   std::unique_ptr<lgc::LgcContext> m_builderContext;    // LGC context
 
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+  // Old version of the code
   std::optional<llvm::CodeGenOpt::Level> m_lastOptLevel{}; // What getOptimizationLevel() last returned
+#else
+  // New version of the code (also handles unknown version, which we treat as latest)
+  std::optional<llvm::CodeGenOptLevel> m_lastOptLevel{}; // What getOptimizationLevel() last returned
+#endif
+
   std::unique_ptr<llvm_dialects::DialectContext> m_dialectContext;
 
   unsigned m_useCount = 0; // Number of times this context is used.

llpc/tool/amdllpc.cpp

@@ -303,12 +303,24 @@ cl::opt<bool> DumpDuplicatePipelines(
 // -llpc_opt: Override the optimization level passed in to LGC with the given one.  This options is the same as the
 // `-opt` option in lgc.  The reason for the second option is to be able to test the LLPC API.  If both options are set
 // then `-opt` wins.
+
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+// Old version of the code
 cl::opt<CodeGenOpt::Level> LlpcOptLevel("llpc-opt", cl::desc("The optimization level for amdllpc to pass to LLPC:"),
                                         cl::init(CodeGenOpt::Default),
                                         values(clEnumValN(CodeGenOpt::None, "none", "no optimizations"),
                                                clEnumValN(CodeGenOpt::Less, "quick", "quick compilation time"),
                                                clEnumValN(CodeGenOpt::Default, "default", "default optimizations"),
                                                clEnumValN(CodeGenOpt::Aggressive, "fast", "fast execution time")));
+#else
+    // New version of the code (also handles unknown version, which we treat as latest)
+cl::opt<CodeGenOptLevel> LlpcOptLevel("llpc-opt", cl::desc("The optimization level for amdllpc to pass to LLPC:"),
+                                        cl::init(CodeGenOptLevel::Default),
+                                        values(clEnumValN(CodeGenOptLevel::None, "none", "no optimizations"),
+                                               clEnumValN(CodeGenOptLevel::Less, "quick", "quick compilation time"),
+                                               clEnumValN(CodeGenOptLevel::Default, "default", "default optimizations"),
+                                               clEnumValN(CodeGenOptLevel::Aggressive, "fast", "fast execution time")));
+#endif
 
 // -resource-layout-scheme: specifies the layout scheme of the resource
 cl::opt<ResourceLayoutScheme> LayoutScheme("resource-layout-scheme", cl::desc("The resource layout scheme:"),
@@ -534,8 +546,16 @@ static void initCompileInfo(CompileInfo *compileInfo) {
   }
 
   // We want the default optimization level to be "Default" which is not 0.
-  compileInfo->gfxPipelineInfo.options.optimizationLevel = CodeGenOpt::Level::Default;
-  compileInfo->compPipelineInfo.options.optimizationLevel = CodeGenOpt::Level::Default;
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+  // Old version of the code
+  compileInfo->gfxPipelineInfo.options.optimizationLevel = static_cast<uint32_t>(CodeGenOpt::Level::Default);
+  compileInfo->compPipelineInfo.options.optimizationLevel = static_cast<uint32_t>(CodeGenOpt::Level::Default);
+#else
+  // New version of the code (also handles unknown version, which we treat as latest)
+  compileInfo->gfxPipelineInfo.options.optimizationLevel = static_cast<uint32_t>(CodeGenOptLevel::Default);
+  compileInfo->compPipelineInfo.options.optimizationLevel = static_cast<uint32_t>(CodeGenOptLevel::Default);
+#endif
+
   compileInfo->gfxPipelineInfo.options.resourceLayoutScheme = LayoutScheme;
   compileInfo->compPipelineInfo.options.forceCsThreadIdSwizzling = ForceCsThreadIdSwizzling;
   compileInfo->compPipelineInfo.options.overrideThreadGroupSizeX = OverrideThreadGroupSizeX;

llpc/tool/llpcCompilationUtils.h

@@ -103,8 +103,14 @@ struct CompileInfo {
   bool scratchAccessBoundsChecks;      // Whether to enable scratch access bounds checks
   bool enableImplicitInvariantExports; // Whether to enable implicit marking of position exports as invariant
   VfxPipelineType pipelineType;        // Pipeline type
+#if LLVM_MAIN_REVISION && LLVM_MAIN_REVISION < 474768
+  // Old version of the code
   std::optional<llvm::CodeGenOpt::Level> optimizationLevel; // The optimization level to pass the compiler
-  bool internalRtShaders;                                   // Whether to enable intrinsics for internal RT shaders
+#else
+                                // New version of the code (also handles unknown version, which we treat as latest)
+  std::optional<llvm::CodeGenOptLevel> optimizationLevel; // The optimization level to pass the compiler
+#endif
+  bool internalRtShaders;       // Whether to enable intrinsics for internal RT shaders
   bool enableColorExportShader; // Enable color export shader, only compile each stage of the pipeline without linking
 };
 

llpc/tool/llpcComputePipelineBuilder.cpp

@@ -138,7 +138,7 @@ Expected<BinaryData> ComputePipelineBuilder::buildComputePipeline() {
   pipelineInfo->options.overrideThreadGroupSizeY = compileInfo.compPipelineInfo.options.overrideThreadGroupSizeY;
   pipelineInfo->options.overrideThreadGroupSizeZ = compileInfo.compPipelineInfo.options.overrideThreadGroupSizeZ;
   if (compileInfo.optimizationLevel.has_value()) {
-    pipelineInfo->options.optimizationLevel = compileInfo.optimizationLevel.value();
+    pipelineInfo->options.optimizationLevel = static_cast<uint32_t>(compileInfo.optimizationLevel.value());
   }
   pipelineInfo->options.threadGroupSwizzleMode = compileInfo.compPipelineInfo.options.threadGroupSwizzleMode;
   pipelineInfo->options.reverseThreadGroup = compileInfo.compPipelineInfo.options.reverseThreadGroup;