Initial parallel CPU

Author: wsmoses

src/core/polyhedral/codegen_llvm.cc

@@ -21,6 +21,7 @@
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Config/llvm-config.h"
 #include "llvm/ExecutionEngine/ExecutionEngine.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/LegacyPassManager.h"
@@ -42,6 +43,10 @@
 #include "tc/core/polyhedral/scop.h"
 #include "tc/core/scope_guard.h"
 
+#ifndef LLVM_VERSION_MAJOR
+#error LLVM_VERSION_MAJOR not set
+#endif
+
 using namespace Halide;
 
 namespace tc {
@@ -303,8 +308,8 @@ class CodeGen_TC : public Halide::Internal::CodeGen_X86 {
     Halide::Internal::debug(3) << "Optimizing module\n";
 
     if (Halide::Internal::debug::debug_level() >= 3) {
-        #if LLVM_VERSION >= 50
-        module->print(dbgs(), nullptr, false, true);
+        #if LLVM_VERSION_MAJOR >= 5
+        module->print(llvm::dbgs(), nullptr, false, true);
         #else
         module->dump();
         #endif
@@ -338,16 +343,16 @@ class CodeGen_TC : public Halide::Internal::CodeGen_X86 {
 
     llvm::PassManagerBuilder b;
     b.OptLevel = 3;
-    b.tapirTarget = new llvm::tapir::CilkABI();
-#if LLVM_VERSION >= 50
+    b.tapirTarget = new llvm::CilkABI();
+#if LLVM_VERSION_MAJOR >= 5
     b.Inliner = llvm::createFunctionInliningPass(b.OptLevel, 0, false);
 #else
     b.Inliner = llvm::createFunctionInliningPass(b.OptLevel, 0);
 #endif
     b.LoopVectorize = true;
     b.SLPVectorize = true;
 
-#if LLVM_VERSION >= 50
+#if LLVM_VERSION_MAJOR >= 5
     if (TM) {
         TM->adjustPassManager(b);
     }
@@ -366,8 +371,8 @@ class CodeGen_TC : public Halide::Internal::CodeGen_X86 {
 
     Halide::Internal::debug(3) << "After LLVM optimizations:\n";
     if (Halide::Internal::debug::debug_level() >= 2) {
-        #if LLVM_VERSION >= 50
-        module->print(dbgs(), nullptr, false, true);
+        #if LLVM_VERSION_MAJOR >= 5
+        module->print(llvm::dbgs(), nullptr, false, true);
         #else
         module->dump();
         #endif
@@ -530,7 +535,8 @@ class LLVMCodegen {
         llvm::BasicBlock::Create(llvmCtx, "loop_latch", function);
     auto* loopExitBB = llvm::BasicBlock::Create(llvmCtx, "loop_exit", function);
 
-    bool parallel = true;
+    //TODO: integrate query ISL as to whether the relevant loop ought be parallelized
+    bool parallel = false;
 
     llvm::Value* SyncRegion = nullptr;
     if (parallel) {

src/core/polyhedral/llvm_jit.cc

@@ -37,6 +37,8 @@ Jit::Jit()
       objectLayer_([]() { return std::make_shared<SectionMemoryManager>(); }),
       compileLayer_(objectLayer_, orc::SimpleCompiler(*TM_)) {
   sys::DynamicLibrary::LoadLibraryPermanently(nullptr);
+  std::string err;
+  sys::DynamicLibrary::LoadLibraryPermanently("cilkrts", &err);
 }
 
 void Jit::codegenScop(
@@ -82,42 +84,4 @@ JITTargetAddress Jit::getSymbolAddress(const std::string Name) {
   return *res;
 }
 
-DEFINE_bool(llvm_no_opt, false, "Disable LLVM optimizations");
-DEFINE_bool(llvm_debug_passes, false, "Print pass debug info");
-DEFINE_bool(llvm_dump_optimized_ir, false, "Print optimized IR");
-
-std::shared_ptr<Module> Jit::optimizeModule(std::shared_ptr<Module> M) {
-  if (FLAGS_llvm_no_opt) {
-    return M;
-  }
-
-  PassBuilder PB(TM_.get());
-  AAManager AA;
-  CHECK(PB.parseAAPipeline(AA, "default"))
-      << "Unable to parse AA pipeline description.";
-  LoopAnalysisManager LAM(FLAGS_llvm_debug_passes);
-  FunctionAnalysisManager FAM(FLAGS_llvm_debug_passes);
-  CGSCCAnalysisManager CGAM(FLAGS_llvm_debug_passes);
-  ModuleAnalysisManager MAM(FLAGS_llvm_debug_passes);
-  FAM.registerPass([&] { return std::move(AA); });
-  PB.registerModuleAnalyses(MAM);
-  PB.registerCGSCCAnalyses(CGAM);
-  PB.registerFunctionAnalyses(FAM);
-  PB.registerLoopAnalyses(LAM);
-  PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
-
-  ModulePassManager MPM(FLAGS_llvm_debug_passes);
-  MPM.addPass(VerifierPass());
-  CHECK(PB.parsePassPipeline(MPM, "default<O3>", true, FLAGS_llvm_debug_passes))
-      << "Unable to parse pass pipline description.";
-  MPM.addPass(VerifierPass());
-
-  MPM.run(*M, MAM);
-
-  if (FLAGS_llvm_dump_optimized_ir) {
-    // M->dump(); // does not link
-    M->print(llvm::errs(), nullptr);
-  }
-
-  return M;
 }

test/test_mapper_llvm.cc

@@ -68,7 +68,7 @@ TEST(LLVMCodegen, MultiStmt) {
  {
    O1(i, j) +=! A(i, j, rk, rl) * B(i, j)
    O2(i, j) = C(i, j) * D(i, j)
-   O3(i, j) = O1(i, j) + O2(i, j)
+   O3(i, j) = O1(i, j) + O2(i, j) 
  }
  )TC";