more "homogeneous" speed for different CXX standards.

Author: pachadotdev

inst/include/cpp4r/R.hpp

@@ -33,6 +33,7 @@
 // clang-format on
 
 #include <type_traits>
+#include "cpp4r/cpp_version.hpp"  // for CPP4R optimization macros
 
 #if defined(R_VERSION) && R_VERSION >= R_Version(4, 4, 0)
 // Use R's new macro
@@ -119,14 +120,14 @@ template <typename T>
 inline T na();
 
 template <typename T>
-inline typename std::enable_if<!std::is_same<typename std::decay<T>::type, double>::value,
+CPP4R_HOT inline typename std::enable_if<!std::is_same<typename std::decay<T>::type, double>::value,
                                bool>::type
 is_na(const T& value) {
   return value == na<T>();
 }
 
 template <typename T>
-inline typename std::enable_if<std::is_same<typename std::decay<T>::type, double>::value,
+CPP4R_HOT inline typename std::enable_if<std::is_same<typename std::decay<T>::type, double>::value,
                                bool>::type
 is_na(const T& value) {
   return ISNA(value);

inst/include/cpp4r/cpp_version.hpp

@@ -112,4 +112,39 @@
 #else
 #define CPP4R_LIKELY(x) (x)
 #define CPP4R_UNLIKELY(x) (x)
+#endif
+
+// CPP4R_RESTRICT: pointer aliasing hints for better optimization
+#if defined(__GNUC__) || defined(__clang__)
+#define CPP4R_RESTRICT __restrict__
+#elif defined(_MSC_VER)
+#define CPP4R_RESTRICT __restrict
+#else
+#define CPP4R_RESTRICT
+#endif
+
+// CPP4R_ALWAYS_INLINE: force inlining of hot path functions
+// CPP4R_NOINLINE: prevent inlining of cold path functions
+#if CPP4R_HAS_CXX17 && (defined(__GNUC__) || defined(__clang__))
+#define CPP4R_ALWAYS_INLINE [[gnu::always_inline]] inline
+#define CPP4R_NOINLINE [[gnu::noinline]]
+#elif defined(__GNUC__) || defined(__clang__)
+#define CPP4R_ALWAYS_INLINE __attribute__((always_inline)) inline
+#define CPP4R_NOINLINE __attribute__((noinline))
+#elif defined(_MSC_VER)
+#define CPP4R_ALWAYS_INLINE __forceinline
+#define CPP4R_NOINLINE __declspec(noinline)
+#else
+#define CPP4R_ALWAYS_INLINE inline
+#define CPP4R_NOINLINE
+#endif
+
+// CPP4R_HOT: mark frequently executed functions for better optimization
+// CPP4R_COLD: mark rarely executed functions (errors, initialization)
+#if defined(__GNUC__) || defined(__clang__)
+#define CPP4R_HOT __attribute__((hot))
+#define CPP4R_COLD __attribute__((cold))
+#else
+#define CPP4R_HOT
+#define CPP4R_COLD
 #endif

inst/include/cpp4r/matrix.hpp

@@ -101,7 +101,9 @@ class matrix : public matrix_slices<S> {
     }
     bool operator!=(const slice& rhs) const noexcept { return !operator==(rhs); }
 
-    T operator[](int pos) const { return parent_.vector_[offset_ + stride() * pos]; }
+    CPP4R_ALWAYS_INLINE T operator[](int pos) const { 
+      return parent_.vector_[offset_ + stride() * pos]; 
+    }
 
     // iterates elements of a slice
     class iterator {
@@ -239,7 +241,9 @@ class matrix : public matrix_slices<S> {
 
   r_vector<r_string> names() const { return r_vector<r_string>(vector_.names()); }
 
-  T operator()(int row, int col) const { return vector_[row + (col * nrow())]; }
+  CPP4R_ALWAYS_INLINE T operator()(int row, int col) const { 
+    return vector_[row + (col * nrow())]; 
+  }
 
   slice operator[](int index) const { return {*this, index}; }
 

inst/include/cpp4r/r_vector_impl.hpp

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "cpp4r/r_vector_fwd.hpp"
+#include "cpp4r/cpp_version.hpp"  // for CPP4R optimization macros
 
 namespace cpp4r {
 
@@ -252,18 +253,18 @@ inline r_vector<r_string> r_vector<T>::names() const {
 }
 
 template <typename T>
-inline T r_vector<T>::get_oob() {
+CPP4R_COLD inline T r_vector<T>::get_oob() {
   throw std::out_of_range("r_vector");
 }
 
 template <typename T>
 inline SEXP r_vector<T>::valid_type(SEXP x) {
   const SEXPTYPE type = get_sexptype();
 
-  if (x == nullptr) {
+  if (CPP4R_UNLIKELY(x == nullptr)) {
     throw type_error(type, NILSXP);
   }
-  if (detail::r_typeof(x) != type) {
+  if (CPP4R_UNLIKELY(detail::r_typeof(x) != type)) {
     throw type_error(type, detail::r_typeof(x));
   }
 
@@ -316,9 +317,9 @@ r_vector<T>::const_iterator::const_iterator(const r_vector* data, R_xlen_t pos)
 }
 
 template <typename T>
-inline typename r_vector<T>::const_iterator& r_vector<T>::const_iterator::operator++() {
+CPP4R_ALWAYS_INLINE typename r_vector<T>::const_iterator& r_vector<T>::const_iterator::operator++() {
   ++pos_;
-  if (use_buf(data_->is_altrep()) && pos_ >= block_start_ + length_) {
+  if (CPP4R_UNLIKELY(use_buf(data_->is_altrep()) && pos_ >= block_start_ + length_)) {
     fill_buf(pos_);
   }
   return *this;
@@ -396,12 +397,12 @@ inline typename r_vector<T>::const_iterator r_vector<T>::find(
 }
 
 template <typename T>
-inline T r_vector<T>::const_iterator::operator*() const {
-  if (use_buf(data_->is_altrep())) {
+CPP4R_ALWAYS_INLINE T r_vector<T>::const_iterator::operator*() const {
+  if (CPP4R_UNLIKELY(use_buf(data_->is_altrep()))) {
     // Use pre-loaded buffer for compatible ALTREP types
     return static_cast<T>(buf_[pos_ - block_start_]);
   } else {
-    // Otherwise pass through to normal retrieval method
+    // Otherwise pass through to normal retrieval method (common case)
     return data_->operator[](pos_);
   }
 }

scripts/combine-benchmarks.R

@@ -171,6 +171,14 @@ res <- res %>%
         rel_cpp4r_Rcpp = signif(rel_time_cpp4r / rel_time_Rcpp, 3)
     )
 
+# add rel time for cpp4r within the same test group
+res <- res %>%
+    group_by(test) %>%
+    mutate(
+        rel_cpp4r_within_test = signif(rel_time_cpp4r / min(rel_time_cpp4r, na.rm = TRUE), 3)
+    ) %>%
+    ungroup()
+
 res <- res %>%
     mutate(
         avg_time_cpp11 = paste(avg_time_cpp11, paste0("(", rel_time_cpp11, ")")),