[Trifocal+P2Pt] Fix buffer overlow - clean memleak test

rfabbri · Nov 18, 2023 · 2e52031 · 2e52031
1 parent 4811c3b
commit 2e52031
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 10 deletions.
diff --git a/minus/chicago14a-HxH.hxx b/minus/chicago14a-HxH.hxx
@@ -146,7 +146,7 @@ HxH(const C<F>* __restrict ux /*x and t*/, const C<F> * __restrict uparams, C<F>
   static constexpr F C0 = 1.;
   static constexpr F C1 = -1.;
   static constexpr F C2 = 2.;
-  static constexpr F C3 = 0.;
+//  static constexpr F C3 = 0.;
   const C<F> G0  = C1 * t;
   const C<F> G1  = C0 + G0;
   const C<F> G2  = G1 * X15;
@@ -3321,7 +3321,7 @@ HxH_constants(const C<F>* __restrict ux /*x and t*/, const C<F> * __restrict upa
   static constexpr F C0 = 1.;
   static constexpr F C1 = -1.;
   static constexpr F C2 = 2.;
-  static constexpr F C3 = 0.;
+//  static constexpr F C3 = 0.;
   const C<F> G0  = C1 * t;
   const C<F> G1  = C0 + G0;
   const C<F> G2  = G1 * X15;
@@ -6526,7 +6526,7 @@ HxH_constants_all_sols(const C<F>* __restrict ux /*x and t*/, const C<F> * __res
   static constexpr F C0 = 1.;
   static constexpr F C1 = -1.;
   static constexpr F C2 = 2.;
-  static constexpr F C3 = 0.;
+//  static constexpr F C3 = 0.;
   const C<F> G0  = C1 * t;
   const C<F> G1  = C0 + G0;
   const C<F> G2  = G1 * X15;

diff --git a/minus/chicago14a-Hxt.hxx b/minus/chicago14a-Hxt.hxx
@@ -155,7 +155,7 @@ Hxt(const C<F> * __restrict ux, const C<F> * __restrict uparams, C<F> * __restri
   static constexpr F C0 = 1.;
   static constexpr F C1 = -1.;
   static constexpr F C2 = 2.;
-  static constexpr F C3 = 0.;
+//  static constexpr F C3 = 0.;
   const C<F> G0 = -t;
   const C<F> G1 = C0 + G0; // Packet2cd G1G3 = pload(G1G3); // Use Eigen arrays for AVX/SSE
   const C<F> G2 = G1 * X15;
@@ -3938,7 +3938,7 @@ Hxt_constants(const C<F> * __restrict ux, const C<F> * __restrict uparams, C<F>
   static constexpr F C0 = 1.;
   static constexpr F C1 = -1.;
   static constexpr F C2 = 2.;
-  static constexpr F C3 = 0.;
+//  static constexpr F C3 = 0.;
   const C<F> G0 = -t;
   const C<F> G1 = C0 + G0; // Packet2cd G1G3 = pload(G1G3); // Use Eigen arrays for AVX/SSE
   const C<F> G2 = G1 * X15;

diff --git a/minus/internal-util.h b/minus/internal-util.h
@@ -46,9 +46,9 @@ struct minus_array { // Speed critical -----------------------------------------
   }
 
   static inline __attribute__((always_inline)) void 
-  fadd_to_self(C<F> * __restrict a, const C<F> * __restrict b)
+  fadd_to_self(F * __restrict a, F * __restrict b)
   {
-    for (unsigned int i=0; i < N+1; ++i,++a,++b) *a += *b;
+    for (unsigned int i=0; i < 2*N+1; ++i,++a,++b) *a += *b;
   }
 
   static inline __attribute__((always_inline)) void 

diff --git a/minus/minus.hxx b/minus/minus.hxx
@@ -110,6 +110,7 @@ track(const track_settings &s, const C<F> s_sols_u[f::nve*f::nsols], const C<F>
   C<F> *const dxdt = (C<F> *) dxdtf;
   C<F> *const x0 = x0t0;
   F    *const t0 = (F *) (x0t0 + f::nve);
+  F    *const t  = (F *) (xt + f::nve);
   C<F> *const x1t1 = xt;
   C<F> *const dx = dxdt;
   C<F> *const dx4 = dx;
@@ -169,7 +170,7 @@ track(const track_settings &s, const C<F> s_sols_u[f::nve*f::nsols], const C<F>
 
       v::add_to_self(xt, dx4);
       v::multiply_scalar_to_self(dx4, 2.);
-      xt[f::nve] += one_half_dt;  // t0+.5dt
+      *t += one_half_dt;  // t0+.5dt
       evaluate_Hxt(xt, params, Hxt);
       memoize_Hxt<P,F>(Hxt);/*, ycHxt);*/
       lsolve<P,F>(AA, dxi);
@@ -190,7 +191,7 @@ track(const track_settings &s, const C<F> s_sols_u[f::nve*f::nsols], const C<F>
       v::add_to_self(xt, dxi);
       v::multiply_scalar_to_self(dxi, 2.);
       v::add_to_self(dx4, dxi);
-      xt[f::nve] = *t0 + *dt;               // t0+dt
+      *t = *t0 + *dt;               // t0+dt
       evaluate_Hxt(xt, params, Hxt);
       memoize_Hxt<P,F>(Hxt);/*, ycHxt);*/
       lsolve<P,F>(AA, dxi);
@@ -203,7 +204,7 @@ track(const track_settings &s, const C<F> s_sols_u[f::nve*f::nsols], const C<F>
 
       // make prediction
       v::fcopy(x0t0, x1t1);
-      v::fadd_to_self(x1t1, dxdt);
+      v::fadd_to_self((double *)x1t1, (double *)dxdt);
 
 
       /// CORRECTOR ///