Simplify compiler and add cpu option

README.md

@@ -38,7 +38,7 @@ a & b
 The *jitmap-ir* tool takes an expression as first input argument and dumps the
 generated LLVM ir to stdout. By default, this will not use vectorized instruction.
 
-```
+```llvm
 # tools/jitmap-ir '(a & b & c & d | e ^ f)'
 ; ModuleID = 'jitmap-ir-module'
 source_filename = "jitmap-ir-module"
@@ -82,7 +82,7 @@ attributes #0 = { argmemonly }
 
 We can then use LLVM's `llc` to transform the IR into native assembly.
 
-```
+```objdump
 # tools/jitmap-ir '(a & b & c & d | e ^ f)' | llc-8 -O3 -mcpu=core-avx2
 ninja: no work to do.
  .text
@@ -127,7 +127,7 @@ query: # @query
 
 This code is still not fully optimized, `opt` is used for this.
 
-```
+```assembly
 # tools/jitmap-ir '(a & b & c & d | e ^ f)' | opt-8 -S -O3 -mcpu=core-avx2 -mtriple=x86_64-unknown-linux-gnu | llc-8 -O3 -mcpu=core-avx2
  .text
  .file "jitmap-ir-module"

include/jitmap/query/compiler.h

@@ -23,22 +23,15 @@ struct CompilerOptions {
  // will be clamped to 3.
  uint8_t optimization_level = 3;
 
- // Number of background compiler threads.
- uint8_t compiler_threads = 0;
-
- // Controls the number of scalar, i.e. width of the vector aggregate value in
- // the loop. A value of 1 will emit a scalar value instead of a vector value.
- //
- // Usually a small power of 2, e.g. 1, 2, 4, 8 or 16. See the documentation
- // of your hardware platform for the optimal value.
+ // CPU architecture to optimize for. This will dictate the "best" vector
+ // instruction set to compile with. If unspecified or empty, llvm will
+ // auto-detect the host cpu architecture.
  //
- // LLVM's optimizer is able to perform the auto-vectorization of the loop. In
- // cases where it can't, change the vector width here.
- uint8_t vector_width = 1;
-
- // Controls the width of each scalar (in bits). See the documentation
- // of your hardware platform for the optimal value.
- uint8_t scalar_width = kBitsPerBitsetWord;
+ // Invoke clang with `-mcpu=?` options to get a list of supported strings, e.g.
+ // - core-avx-i
+ // - core-avx2
+ // - skylake-avx512
+ std::string cpu = "";
 };
 
 // The QueryIRCodeGen class generates LLVM IR for query expression on dense

include/jitmap/query/expr.h

@@ -138,7 +138,7 @@ class ExprBuilder {
  return &full;
  }
 
- Expr* Var(std::string name) { return Build<VariableExpr>(name); }
+ Expr* Var(const std::string& name) { return Build<VariableExpr>(name); }
  Expr* Var(std::string_view name) { return Build<VariableExpr>(std::string(name)); }
 
  Expr* Not(Expr* expr) { return Build<NotOpExpr>(expr); }

include/jitmap/query/jit.h

@@ -30,13 +30,6 @@ class JitEngine {
  // https://llvm.org/doxygen/Host_8h_source.html for more information.
  std::string GetTargetCPU() const;
 
- // Return the LLVM features string for the host CPU.
- //
- // An array delimited by comma of symbols referencing a specific cpu feature.
- // The feature supported are prefixed by `+`, and unsupported by `-`. See
- // https://llvm.org/doxygen/Host_8h_source.html for more information.
- std::string GetTargetFeatureString() const;
-
  // Return the LLVM target triple for the host.
  //
  // The format is ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT. See

src/jitmap/query/compiler.cc

@@ -92,7 +92,7 @@ class QueryIRCodeGen::Impl {
  // Constants
  auto induction_type = llvm::Type::getInt64Ty(*ctx_);
  auto zero = llvm::ConstantInt::get(induction_type, 0);
- auto step = llvm::ConstantInt::get(induction_type, vector_width());
+ auto step = llvm::ConstantInt::get(induction_type, 1);
  auto n_words = llvm::ConstantInt::get(induction_type, word_size());
 
  auto loop_block = llvm::BasicBlock::Create(*ctx_, "loop", fn);
@@ -153,10 +153,8 @@ class QueryIRCodeGen::Impl {
  auto namify = [&i](std::string key) { return key + "_" + std::to_string(i); };
  // Compute the address to load
  auto gep = builder_.CreateInBoundsGEP(bitmap_addr, {loop_idx}, namify("gep"));
- // Cast previous address as a vector-type
- auto bitcast = builder_.CreateBitCast(gep, VectorPtrType(), namify("bitcast"));
  // Load in a register
- return builder_.CreateLoad(bitcast, namify("load"));
+ return builder_.CreateLoad(gep, namify("load"));
  };
 
  // Bind the variable bitmaps by name to inputs of the function
@@ -167,12 +165,12 @@ class QueryIRCodeGen::Impl {
  }
 
  // Execute the expression tree on the input
- ExprCodeGenVisitor visitor{keyed_bitmaps, builder_, VectorType()};
+ ExprCodeGenVisitor visitor{keyed_bitmaps, builder_, ElementType()};
  auto result = query.expr().Visit(visitor);
 
  // Store the result in the output bitmap.
  auto gep = builder_.CreateInBoundsGEP(output, {loop_idx}, "gep_output");
- auto bitcast = builder_.CreateBitCast(gep, VectorPtrType(), "bitcast_output");
+ auto bitcast = builder_.CreateBitCast(gep, ElementPtrType(), "bitcast_output");
  builder_.CreateStore(result, bitcast);
  }
 
@@ -224,15 +222,8 @@ class QueryIRCodeGen::Impl {
 
  llvm::Type* ElementType() { return llvm::Type::getIntNTy(*ctx_, scalar_width()); }
  llvm::Type* ElementPtrType() { return ElementType()->getPointerTo(); }
- llvm::Type* VectorType() {
- // Reverts to scalar for debugging purposes.
- if (vector_width() == 1) return ElementType();
- return llvm::VectorType::get(ElementType(), vector_width());
- }
- llvm::Type* VectorPtrType() { return VectorType()->getPointerTo(); }
 
- uint8_t vector_width() const { return options_.vector_width; }
- uint8_t scalar_width() const { return options_.scalar_width; }
+ uint8_t scalar_width() const { return kBitsPerBitsetWord; }
  uint32_t word_size() const { return kBitsPerContainer / scalar_width(); }
 
  std::unique_ptr<llvm::LLVMContext> ctx_;

src/jitmap/query/jit.cc

@@ -62,14 +62,12 @@ llvm::CodeGenOpt::Level CodeGetOptFromNumber(uint8_t level) {
  }
 }
 
-llvm::SubtargetFeatures DetectHostCPUFeatures() {
- llvm::SubtargetFeatures features;
- llvm::StringMap<bool> map;
- llvm::sys::getHostCPUFeatures(map);
- for (auto& feature : map) {
- features.AddFeature(feature.first(), feature.second);
+std::string DetectCPU(const CompilerOptions& opts) {
+ if (opts.cpu.empty()) {
+ return llvm::sys::getHostCPUName();
  }
- return features;
+
+ return opts.cpu;
 }
 
 auto InitHostTargetMachineBuilder(const CompilerOptions& opts) {
@@ -79,9 +77,7 @@ auto InitHostTargetMachineBuilder(const CompilerOptions& opts) {
 
  auto machine_builder = ExpectOrRaise(orc::JITTargetMachineBuilder::detectHost());
  machine_builder.setCodeGenOptLevel(CodeGetOptFromNumber(opts.optimization_level));
- // In newer LLVM version, cpu and features are detected by detectHost();
- machine_builder.setCPU(llvm::sys::getHostCPUName());
- machine_builder.addFeatures(DetectHostCPUFeatures().getFeatures());
+ machine_builder.setCPU(DetectCPU(opts));
 
  return machine_builder;
 }
@@ -121,7 +117,6 @@ std::unique_ptr<orc::LLJIT> InitLLJIT(orc::JITTargetMachineBuilder machine_build
  const CompilerOptions& options) {
  return ExpectOrRaise(orc::LLJITBuilder()
  .setJITTargetMachineBuilder(machine_builder)
- .setNumCompileThreads(options.compiler_threads)
  .setObjectLinkingLayerCreator(ObjectLinkingLayerFactory)
  .create());
 }
@@ -150,7 +145,6 @@ class JitEngine::Impl {
 
  // Introspection
  std::string GetTargetCPU() const { return host_->getTargetCPU(); }
- std::string GetTargetFeatureString() const { return host_->getTargetFeatureString(); }
  std::string GetTargetTriple() const { return host_->getTargetTriple().normalize(); }
  const llvm::DataLayout layout() const { return host_->createDataLayout(); }
 
@@ -190,9 +184,6 @@ JitEngine::~JitEngine() {}
 JitEngine::JitEngine(JitEngine&& other) { std::swap(impl_, other.impl_); }
 
 std::string JitEngine::GetTargetCPU() const { return impl_->GetTargetCPU(); }
-std::string JitEngine::GetTargetFeatureString() const {
- return impl_->GetTargetFeatureString();
-}
 std::string JitEngine::GetTargetTriple() const { return impl_->GetTargetTriple(); }
 
 void JitEngine::Compile(QueryIRCodeGen query) { impl_->AddModule(std::move(query)); }

tests/query/jit_test.cc

@@ -53,7 +53,6 @@ class JitTest : public QueryTest {
 
 TEST_F(JitTest, CpuDetection) {
  EXPECT_NE(engine_.GetTargetCPU(), "");
- EXPECT_NE(engine_.GetTargetFeatureString(), "");
  EXPECT_NE(engine_.GetTargetTriple(), "");
 }