From 39b2a52bd0cb6a6d58bd12e90466fdef7f3f6992 Mon Sep 17 00:00:00 2001
From: conrad <conrad@localhost.localdomain>
Date: Mon, 14 Jul 2025 14:02:19 +1000
Subject: [PATCH] allow to optionally disable optimised handling of pow()

---
 include/armadillo_bits/Cube_meat.hpp    | 12 ++++++------
 include/armadillo_bits/Mat_meat.hpp     | 12 ++++++------
 include/armadillo_bits/arma_config.hpp  |  7 +++++++
 include/armadillo_bits/config.hpp       |  9 +++++++++
 include/armadillo_bits/config.hpp.cmake |  9 +++++++++
 include/armadillo_bits/debug.hpp        |  1 +
 include/armadillo_bits/fn_accu.hpp      |  8 ++++----
 include/armadillo_bits/op_sum_meat.hpp  |  4 ++--
 8 files changed, 44 insertions(+), 18 deletions(-)

diff --git a/include/armadillo_bits/Cube_meat.hpp b/include/armadillo_bits/Cube_meat.hpp
index 41c6b30a..e9de1e22 100644
--- a/include/armadillo_bits/Cube_meat.hpp
+++ b/include/armadillo_bits/Cube_meat.hpp
@@ -2872,7 +2872,7 @@ Cube<eT>::Cube(const eOpCube<T1, eop_type>& X)
   
   init_cold();
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -2902,7 +2902,7 @@ Cube<eT>::operator=(const eOpCube<T1, eop_type>& X)
   
   init_warm(X.get_n_rows(), X.get_n_cols(), X.get_n_slices());
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -2932,7 +2932,7 @@ Cube<eT>::operator+=(const eOpCube<T1, eop_type>& X)
   
   if(bad_alias)  { const Cube<eT> tmp(X); return (*this).operator+=(tmp); }
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -2962,7 +2962,7 @@ Cube<eT>::operator-=(const eOpCube<T1, eop_type>& X)
   
   if(bad_alias)  { const Cube<eT> tmp(X); return (*this).operator-=(tmp); }
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -2992,7 +2992,7 @@ Cube<eT>::operator%=(const eOpCube<T1, eop_type>& X)
   
   if(bad_alias)  { const Cube<eT> tmp(X); return (*this).operator%=(tmp); }
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -3022,7 +3022,7 @@ Cube<eT>::operator/=(const eOpCube<T1, eop_type>& X)
   
   if(bad_alias)  { const Cube<eT> tmp(X); return (*this).operator/=(tmp); }
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
diff --git a/include/armadillo_bits/Mat_meat.hpp b/include/armadillo_bits/Mat_meat.hpp
index 5dc3dbce..e4e0d695 100644
--- a/include/armadillo_bits/Mat_meat.hpp
+++ b/include/armadillo_bits/Mat_meat.hpp
@@ -5201,7 +5201,7 @@ Mat<eT>::Mat(const eOp<T1, eop_type>& X)
   
   init_cold();
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -5231,7 +5231,7 @@ Mat<eT>::operator=(const eOp<T1, eop_type>& X)
   
   init_warm(X.get_n_rows(), X.get_n_cols());
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -5260,7 +5260,7 @@ Mat<eT>::operator+=(const eOp<T1, eop_type>& X)
   
   if(bad_alias)  { const Mat<eT> tmp(X); return (*this).operator+=(tmp); }
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -5289,7 +5289,7 @@ Mat<eT>::operator-=(const eOp<T1, eop_type>& X)
   
   if(bad_alias)  { const Mat<eT> tmp(X); return (*this).operator-=(tmp); }
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -5335,7 +5335,7 @@ Mat<eT>::operator%=(const eOp<T1, eop_type>& X)
   
   if(bad_alias)  { const Mat<eT> tmp(X); return (*this).operator%=(tmp); }
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
@@ -5364,7 +5364,7 @@ Mat<eT>::operator/=(const eOp<T1, eop_type>& X)
   
   if(bad_alias)  { const Mat<eT> tmp(X); return (*this).operator/=(tmp); }
   
-  if(is_same_type<eop_type, eop_pow>::value)
+  if(arma_config::optimise_powexpr && is_same_type<eop_type, eop_pow>::value)
     {
     constexpr bool eT_non_int = is_non_integral<eT>::value;
     
diff --git a/include/armadillo_bits/arma_config.hpp b/include/armadillo_bits/arma_config.hpp
index b528219f..5467dc16 100644
--- a/include/armadillo_bits/arma_config.hpp
+++ b/include/armadillo_bits/arma_config.hpp
@@ -65,6 +65,13 @@ struct arma_config
   #endif
   
   
+  #if defined(ARMA_OPTIMISE_POWEXPR)
+    static constexpr bool optimise_powexpr = true;
+  #else
+    static constexpr bool optimise_powexpr = false;
+  #endif
+  
+  
   #if defined(ARMA_CHECK_CONFORMANCE)
     static constexpr bool check_conform = true;
   #else
diff --git a/include/armadillo_bits/config.hpp b/include/armadillo_bits/config.hpp
index a567629d..ea9e64fc 100644
--- a/include/armadillo_bits/config.hpp
+++ b/include/armadillo_bits/config.hpp
@@ -172,6 +172,11 @@
   //// of inv() and inv_sympd() within compound expressions
 #endif
 
+#if !defined(ARMA_OPTIMISE_POWEXPR)
+  #define ARMA_OPTIMISE_POWEXPR
+  //// Comment out the above line to disable optimised handling of pow()
+#endif
+
 #if !defined(ARMA_CHECK_CONFORMANCE)
   #define ARMA_CHECK_CONFORMANCE
   //// Comment out the above line to disable conformance checks for bounds and size.
@@ -346,6 +351,10 @@
   #undef ARMA_OPTIMISE_INVEXPR
 #endif
 
+#if defined(ARMA_DONT_OPTIMISE_POWEXPR)
+  #undef ARMA_OPTIMISE_POWEXPR
+#endif
+
 #if defined(ARMA_DONT_CHECK_CONFORMANCE)
   #if defined(ARMA_CHECK_CONFORMANCE) && (ARMA_WARN_LEVEL >= 2)
     #pragma message ("WARNING: conformance checks disabled")
diff --git a/include/armadillo_bits/config.hpp.cmake b/include/armadillo_bits/config.hpp.cmake
index 32036333..706da587 100644
--- a/include/armadillo_bits/config.hpp.cmake
+++ b/include/armadillo_bits/config.hpp.cmake
@@ -172,6 +172,11 @@
   //// of inv() and inv_sympd() within compound expressions
 #endif
 
+#if !defined(ARMA_OPTIMISE_POWEXPR)
+  #define ARMA_OPTIMISE_POWEXPR
+  //// Comment out the above line to disable optimised handling of pow()
+#endif
+
 #if !defined(ARMA_CHECK_CONFORMANCE)
   #define ARMA_CHECK_CONFORMANCE
   //// Comment out the above line to disable conformance checks for bounds and size.
@@ -346,6 +351,10 @@
   #undef ARMA_OPTIMISE_INVEXPR
 #endif
 
+#if defined(ARMA_DONT_OPTIMISE_POWEXPR)
+  #undef ARMA_OPTIMISE_POWEXPR
+#endif
+
 #if defined(ARMA_DONT_CHECK_CONFORMANCE)
   #if defined(ARMA_CHECK_CONFORMANCE) && (ARMA_WARN_LEVEL >= 2)
     #pragma message ("WARNING: conformance checks disabled")
diff --git a/include/armadillo_bits/debug.hpp b/include/armadillo_bits/debug.hpp
index 3244d0ad..f0d435f9 100644
--- a/include/armadillo_bits/debug.hpp
+++ b/include/armadillo_bits/debug.hpp
@@ -1440,6 +1440,7 @@ inline void arma_debug_check(bool state, const char* msg)  { arma_conform_check(
         out << "\n@ arma_config::optimise_band    = " << arma_config::optimise_band;
         out << "\n@ arma_config::optimise_sym     = " << arma_config::optimise_sym;
         out << "\n@ arma_config::optimise_invexpr = " << arma_config::optimise_invexpr;
+        out << "\n@ arma_config::optimise_powexpr = " << arma_config::optimise_powexpr;
         out << "\n@ arma_config::check_conform    = " << arma_config::check_conform;
         out << "\n@ arma_config::check_nonfinite  = " << arma_config::check_nonfinite;
         out << "\n@ arma_config::fast_math        = " << arma_config::fast_math;
diff --git a/include/armadillo_bits/fn_accu.hpp b/include/armadillo_bits/fn_accu.hpp
index 250f2785..46dfa02f 100644
--- a/include/armadillo_bits/fn_accu.hpp
+++ b/include/armadillo_bits/fn_accu.hpp
@@ -407,14 +407,14 @@ accu(const eOp<T1,eop_pow>& expr)
   
   typedef eOp<T1,eop_pow> expr_type;
   
-  if(expr.aux == eT(2))
+  if(arma_config::optimise_powexpr && (expr.aux == eT(2)))
     {
     typedef eOp<T1,eop_square> modified_expr_type;
     
     return accu( reinterpret_cast< const modified_expr_type& >(expr) );
     }
   
-  if((expr.aux == eT(0.5)) && is_non_integral<eT>::value)
+  if(arma_config::optimise_powexpr && (expr.aux == eT(0.5)) && is_non_integral<eT>::value)
     {
     typedef eOp<T1,eop_sqrt> modified_expr_type;
     
@@ -1048,14 +1048,14 @@ accu(const eOpCube<T1,eop_pow>& expr)
   
   typedef eOpCube<T1,eop_pow> expr_type;
   
-  if(expr.aux == eT(2))
+  if(arma_config::optimise_powexpr && (expr.aux == eT(2)))
     {
     typedef eOpCube<T1,eop_square> modified_expr_type;
     
     return accu( reinterpret_cast< const modified_expr_type& >(expr) );
     }
   
-  if((expr.aux == eT(0.5)) && is_non_integral<eT>::value)
+  if(arma_config::optimise_powexpr && (expr.aux == eT(0.5)) && is_non_integral<eT>::value)
     {
     typedef eOpCube<T1,eop_sqrt> modified_expr_type;
     
diff --git a/include/armadillo_bits/op_sum_meat.hpp b/include/armadillo_bits/op_sum_meat.hpp
index d2f1eb8a..7eb044d5 100644
--- a/include/armadillo_bits/op_sum_meat.hpp
+++ b/include/armadillo_bits/op_sum_meat.hpp
@@ -154,7 +154,7 @@ op_sum::apply(Mat<typename T1::elem_type>& out, const Op< eOp<T1,eop_pow>, op_su
   
   typedef typename T1::elem_type eT;
   
-  if(in.m.aux == eT(2))
+  if(arma_config::optimise_powexpr && (in.m.aux == eT(2)))
     {
     typedef Op< eOp<T1,eop_square>, op_sum > modified_whole_expr_type;
     
@@ -163,7 +163,7 @@ op_sum::apply(Mat<typename T1::elem_type>& out, const Op< eOp<T1,eop_pow>, op_su
     return;
     }
   
-  if((in.m.aux == eT(0.5)) && is_non_integral<eT>::value)
+  if(arma_config::optimise_powexpr && (in.m.aux == eT(0.5)) && is_non_integral<eT>::value)
     {
     typedef Op< eOp<T1,eop_sqrt>, op_sum > modified_whole_expr_type;