Improve the efficiency of diff.

Author: zoziha

src/stdlib_math_diff.fypp

@@ -18,7 +18,6 @@ contains
         ${t1}$, intent(in), optional :: prepend(:), append(:)
         ${t1}$, allocatable :: y(:)
         integer :: size_prepend, size_append, size_x
-        ${t1}$, allocatable :: work(:)
         integer :: n_, i
 
         n_ = optval(n, 1)
@@ -38,16 +37,24 @@ contains
             return
         end if
 
-        allocate(work(size_x + size_prepend + size_append))
+        #! Use a quick exit for the common case, to avoid memory allocation.
+        if (size_prepend == 0 .and. size_append == 0 .and. n_ == 1) then
+            y = x(2:) - x(1:size_x-1)
+            return
+        end if
+
+        block
+        ${t1}$ :: work(size_x + size_prepend + size_append)
         if (size_prepend > 0) work(:size_prepend) = prepend
         work(size_prepend+1:size_prepend+size_x) = x
         if (size_append > 0) work(size_prepend+size_x+1:) = append
 
         do i = 1, n_
             work(1:size(work)-i) = work(2:size(work)-i+1) - work(1:size(work)-i)
         end do
-        
+
         y = work(1:size(work)-n_)
+        end block
 
     end function diff_1_${k1}$
 
@@ -58,7 +65,6 @@ contains
         ${t1}$, allocatable :: y(:, :)
         integer :: size_prepend, size_append, size_x
         integer :: n_, dim_, i
-        ${t1}$, allocatable :: work(:, :)
 
         n_ = optval(n, 1)
         if (n_ <= 0) then
@@ -87,19 +93,32 @@ contains
             return
         end if
 
+        #! Use a quick exit for the common case, to avoid memory allocation.
+        if (size_prepend == 0 .and. size_append == 0 .and. n_ == 1) then
+            if (dim_ == 1) then
+                y = x(2:, :) - x(1:size_x-1, :)
+            elseif (dim_ == 2) then
+                y = x(:, 2:) - x(:, 1:size_x-1)
+            end if
+            return
+        end if
+        
         if (dim_ == 1) then
-            allocate(work(size_x+size_prepend+size_append, size(x, 2)))
+            block
+            ${t1}$ :: work(size_x+size_prepend+size_append, size(x, 2))
             if (size_prepend > 0) work(1:size_prepend, :) = prepend
             work(size_prepend+1:size_x+size_prepend, :) = x
             if (size_append > 0) work(size_x+size_prepend+1:, :) = append
             do i = 1, n_
                 work(1:size(work,1)-i, :) = work(2:size(work)-i+1, :) - work(1:size(x, 1)-i, :)
             end do
 
-            y = work(1:size(work)-n_, :)
+            y = work(1:size(work,1)-n_, :)
+            end block
 
         elseif (dim_ == 2) then
-            allocate(work(size(x, 1), size_x+size_prepend+size_append))
+            block
+            ${t1}$ :: work(size(x, 1), size_x+size_prepend+size_append)
             if (size_prepend > 0) work(:, 1:size_prepend) = prepend
             work(:, size_prepend+1:size_x+size_prepend) = x
             if (size_append > 0) work(:, size_x+size_prepend+1:) = append
@@ -108,6 +127,7 @@ contains
             end do
 
             y = work(:, 1:size(work,2)-n_)
+            end block
 
         end if