From 27f50ccc77ac2eb9db62e7c1c5550af090234c19 Mon Sep 17 00:00:00 2001
From: conrad <conrad@localhost.localdomain>
Date: Thu, 9 Oct 2025 11:10:33 +1000
Subject: [PATCH] experiment with avoiding unnecessary alias checks

---
 include/armadillo_bits/Mat_bones.hpp       |  4 ++
 include/armadillo_bits/Mat_meat.hpp        |  8 ++--
 include/armadillo_bits/arma_forward.hpp    |  2 +
 include/armadillo_bits/op_htrans_bones.hpp |  6 +++
 include/armadillo_bits/op_htrans_meat.hpp  | 48 ++++++++++++++++++++++
 include/armadillo_bits/op_mean_bones.hpp   |  3 ++
 include/armadillo_bits/op_mean_meat.hpp    | 18 ++++++++
 7 files changed, 86 insertions(+), 3 deletions(-)

diff --git a/include/armadillo_bits/Mat_bones.hpp b/include/armadillo_bits/Mat_bones.hpp
index b8997272..d0673160 100644
--- a/include/armadillo_bits/Mat_bones.hpp
+++ b/include/armadillo_bits/Mat_bones.hpp
@@ -958,4 +958,8 @@ class Mat_aux
 
 
 
+template<typename eT> class Mat_noalias : public Mat<eT> {};
+
+
+
 //! @}
diff --git a/include/armadillo_bits/Mat_meat.hpp b/include/armadillo_bits/Mat_meat.hpp
index 3bb32ade..b7719217 100644
--- a/include/armadillo_bits/Mat_meat.hpp
+++ b/include/armadillo_bits/Mat_meat.hpp
@@ -5066,10 +5066,12 @@ Mat<eT>::Mat(const Op<T1, op_type>& X)
   , mem()
   {
   arma_debug_sigprint_this(this);
-
+  
   arma_type_check(( is_same_type< eT, typename T1::elem_type >::no ));
   
-  op_type::apply(*this, X);
+  // op_type::apply(*this, X);
+  
+  op_type::apply(static_cast< Mat_noalias<eT>& >(*this), X);
   }
 
 
@@ -5082,7 +5084,7 @@ Mat<eT>&
 Mat<eT>::operator=(const Op<T1, op_type>& X)
   {
   arma_debug_sigprint();
-
+  
   arma_type_check(( is_same_type< eT, typename T1::elem_type >::no ));
   
   op_type::apply(*this, X);
diff --git a/include/armadillo_bits/arma_forward.hpp b/include/armadillo_bits/arma_forward.hpp
index 981776ad..83c484c3 100644
--- a/include/armadillo_bits/arma_forward.hpp
+++ b/include/armadillo_bits/arma_forward.hpp
@@ -31,6 +31,8 @@ template<typename eT> class Row;
 template<typename eT> class Cube;
 template<typename oT> class field;
 
+template<typename eT> class Mat_noalias;
+
 template<typename eT>               struct xvec_htrans;
 template<typename eT, bool do_conj> struct xtrans_mat;
 
diff --git a/include/armadillo_bits/op_htrans_bones.hpp b/include/armadillo_bits/op_htrans_bones.hpp
index 0d86f275..c4293d46 100644
--- a/include/armadillo_bits/op_htrans_bones.hpp
+++ b/include/armadillo_bits/op_htrans_bones.hpp
@@ -77,6 +77,12 @@ struct op_htrans
   
   template<typename T1>
   inline static void apply(Mat<typename T1::elem_type>& out, const Op<T1,op_htrans>& in, const typename arma_cx_only<typename T1::elem_type>::result* junk = nullptr);
+  
+  template<typename T1>
+  inline static void apply(Mat_noalias<typename T1::elem_type>& out, const Op<T1,op_htrans>& in, const typename arma_not_cx<typename T1::elem_type>::result* junk = nullptr);
+  
+  template<typename T1>
+  inline static void apply(Mat_noalias<typename T1::elem_type>& out, const Op<T1,op_htrans>& in, const typename arma_cx_only<typename T1::elem_type>::result* junk = nullptr);
   };
 
 
diff --git a/include/armadillo_bits/op_htrans_meat.hpp b/include/armadillo_bits/op_htrans_meat.hpp
index 964c885e..5c8248b6 100644
--- a/include/armadillo_bits/op_htrans_meat.hpp
+++ b/include/armadillo_bits/op_htrans_meat.hpp
@@ -377,6 +377,54 @@ op_htrans::apply(Mat<typename T1::elem_type>& out, const Op<T1,op_htrans>& in, c
 
 
 
+template<typename T1>
+inline
+void
+op_htrans::apply(Mat_noalias<typename T1::elem_type>& out, const Op<T1,op_htrans>& in, const typename arma_not_cx<typename T1::elem_type>::result* junk)
+  {
+  arma_debug_sigprint();
+  arma_ignore(junk);
+  
+  if((is_Mat<typename Proxy<T1>::stored_type>::value) || (arma_config::openmp && Proxy<T1>::use_mp))
+    {
+    const quasi_unwrap<T1> U(in.m);
+    
+    op_strans::apply_mat_noalias(out, U.M);
+    }
+  else
+    {
+    const Proxy<T1> P(in.m);
+    
+    op_strans::apply_proxy(out, P);
+    }
+  }
+
+
+
+template<typename T1>
+inline
+void
+op_htrans::apply(Mat_noalias<typename T1::elem_type>& out, const Op<T1,op_htrans>& in, const typename arma_cx_only<typename T1::elem_type>::result* junk)
+  {
+  arma_debug_sigprint();
+  arma_ignore(junk);
+  
+  if((is_Mat<typename Proxy<T1>::stored_type>::value) || (arma_config::openmp && Proxy<T1>::use_mp))
+    {
+    const quasi_unwrap<T1> U(in.m);
+    
+    op_htrans::apply_mat_noalias(out, U.M);
+    }
+  else
+    {
+    const Proxy<T1> P(in.m);
+    
+    op_htrans::apply_proxy(out, P);
+    }
+  }
+
+
+
 //
 // op_htrans2
 
diff --git a/include/armadillo_bits/op_mean_bones.hpp b/include/armadillo_bits/op_mean_bones.hpp
index 33d50e47..e30de199 100644
--- a/include/armadillo_bits/op_mean_bones.hpp
+++ b/include/armadillo_bits/op_mean_bones.hpp
@@ -29,6 +29,9 @@ struct op_mean
   template<typename T1>
   inline static void apply(Mat<typename T1::elem_type>& out, const Op<T1,op_mean>& in);
   
+  template<typename T1>
+  inline static void apply(Mat_noalias<typename T1::elem_type>& out, const Op<T1,op_mean>& in);
+  
   template<typename eT>
   inline static void apply_noalias(Mat<eT>& out, const Mat<eT>& X, const uword dim);
   
diff --git a/include/armadillo_bits/op_mean_meat.hpp b/include/armadillo_bits/op_mean_meat.hpp
index 39a6904f..4760e2e0 100644
--- a/include/armadillo_bits/op_mean_meat.hpp
+++ b/include/armadillo_bits/op_mean_meat.hpp
@@ -52,6 +52,24 @@ op_mean::apply(Mat<typename T1::elem_type>& out, const Op<T1,op_mean>& in)
 
 
 
+template<typename T1>
+inline
+void
+op_mean::apply(Mat_noalias<typename T1::elem_type>& out, const Op<T1,op_mean>& in)
+  {
+  arma_debug_sigprint();
+  
+  const uword dim = in.aux_uword_a;
+  
+  arma_conform_check( (dim > 1), "mean(): parameter 'dim' must be 0 or 1" );
+  
+  const quasi_unwrap<T1> U(in.m);
+  
+  op_mean::apply_noalias(out, U.M, dim);
+  }
+
+
+
 template<typename eT>
 inline
 void