diff --git a/CMakeLists.txt b/CMakeLists.txt
index cfaf8c88..fcfb5d24 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -2,7 +2,7 @@ cmake_minimum_required (VERSION 2.6)
 project (cint C)
 set(cint_VERSION_MAJOR "4")
 set(cint_VERSION_MINOR "0")
-set(cint_VERSION_PATCH "3")
+set(cint_VERSION_PATCH "4")
 set(cint_VERSION_TWEAK "0")
 set(cint_VERSION "${cint_VERSION_MAJOR}.${cint_VERSION_MINOR}.${cint_VERSION_PATCH}")
 set(cint_SOVERSION "${cint_VERSION_MAJOR}")
diff --git a/ChangeLog b/ChangeLog
index cca2540e..fe087dec 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,5 @@
+Version 4.0.4  (2020-10-29):
+	* Fix cache size bug in int3c1e
 Version 4.0.3  (2020-10-27):
 	* Fix memory allocation in int3c1e
 	* Improve macro MALLOC_INSTACK
diff --git a/README b/README
index d84d08ce..40602ab4 100644
--- a/README
+++ b/README
@@ -1,8 +1,8 @@
 libcint
 =======
 
-version 4.0.3
-2020-10-27
+version 4.0.4
+2020-10-29
 
 
 What is libcint
diff --git a/src/cint3c1e.c b/src/cint3c1e.c
index fd943a99..335753aa 100644
--- a/src/cint3c1e.c
+++ b/src/cint3c1e.c
@@ -377,6 +377,16 @@ FINT CINT3c1e_nuc_loop_nopt(double *gctr, CINTEnvVars *envs,
         return !*kempty;
 }
 
+#define PAIRDATA_NON0IDX_SIZE(ps) \
+                FINT *bas = envs->bas; \
+                FINT *shls  = envs->shls; \
+                FINT i_prim = bas(NPRIM_OF, shls[0]); \
+                FINT j_prim = bas(NPRIM_OF, shls[1]); \
+                FINT k_prim = bas(NPRIM_OF, shls[2]); \
+                FINT ps = (i_prim * x_ctr[0] \
+                           + j_prim * x_ctr[1] \
+                           + k_prim * x_ctr[2] \
+                           + envs->nf*3);
 
 FINT CINT3c1e_cart_drv(double *out, FINT *dims, CINTEnvVars *envs, CINTOpt *opt,
                        double *cache, FINT int_type)
@@ -385,16 +395,18 @@ FINT CINT3c1e_cart_drv(double *out, FINT *dims, CINTEnvVars *envs, CINTOpt *opt,
         FINT nc = envs->nf * x_ctr[0] * x_ctr[1] * x_ctr[2];
         FINT n_comp = envs->ncomp_e1 * envs->ncomp_tensor;
         if (out == NULL) {
+                PAIRDATA_NON0IDX_SIZE(pdata_size);
                 FINT leng = envs->g_size*3*((1<<envs->gbits)+1);
                 FINT len0 = envs->nf*n_comp;
-                FINT cache_size = leng + len0 + nc*n_comp*4;
+                FINT cache_size = leng + len0 + nc*n_comp*4 + pdata_size;
                 return cache_size;
         }
         double *stack = NULL;
         if (cache == NULL) {
+                PAIRDATA_NON0IDX_SIZE(pdata_size);
                 FINT leng = envs->g_size*3*((1<<envs->gbits)+1);
                 FINT len0 = envs->nf*n_comp;
-                FINT cache_size = leng + len0 + nc*n_comp*4;
+                FINT cache_size = leng + len0 + nc*n_comp*4 + pdata_size;
                 stack = malloc(sizeof(double)*cache_size);
                 cache = stack;
         }
@@ -472,16 +484,20 @@ FINT CINT3c1e_spheric_drv(double *out, FINT *dims, CINTEnvVars *envs, CINTOpt *o
         FINT nc = envs->nf * x_ctr[0] * x_ctr[1] * x_ctr[2];
         FINT n_comp = envs->ncomp_e1 * envs->ncomp_tensor;
         if (out == NULL) {
+                PAIRDATA_NON0IDX_SIZE(pdata_size);
                 FINT leng = envs->g_size*3*((1<<envs->gbits)+1);
                 FINT len0 = envs->nf*n_comp;
-                FINT cache_size = MAX(leng+len0+nc*n_comp*4, nc*n_comp+envs->nf*3);
+                FINT cache_size = MAX(leng+len0+nc*n_comp*4 + pdata_size,
+                                      nc*n_comp+envs->nf*3);
                 return cache_size;
         }
         double *stack = NULL;
         if (cache == NULL) {
+                PAIRDATA_NON0IDX_SIZE(pdata_size);
                 FINT leng = envs->g_size*3*((1<<envs->gbits)+1);
                 FINT len0 = envs->nf*n_comp;
-                FINT cache_size = MAX(leng+len0+nc*n_comp*4, nc*n_comp+envs->nf*3);
+                FINT cache_size = MAX(leng+len0+nc*n_comp*4 + pdata_size,
+                                      nc*n_comp+envs->nf*3);
                 stack = malloc(sizeof(double)*cache_size);
                 cache = stack;
         }
diff --git a/src/cint4c1e.c b/src/cint4c1e.c
index b9a64c5b..9fad988f 100644
--- a/src/cint4c1e.c
+++ b/src/cint4c1e.c
@@ -252,12 +252,11 @@ k_contracted: ;
                 FINT j_prim = bas(NPRIM_OF, shls[1]); \
                 FINT k_prim = bas(NPRIM_OF, shls[2]); \
                 FINT l_prim = bas(NPRIM_OF, shls[3]); \
-                FINT ps = ((i_prim*j_prim + k_prim*l_prim) * 5 \
-                           + i_prim * x_ctr[0] \
+                FINT ps = (i_prim * x_ctr[0] \
                            + j_prim * x_ctr[1] \
                            + k_prim * x_ctr[2] \
                            + l_prim * x_ctr[3] \
-                           +(i_prim+j_prim+k_prim+l_prim)*2 + envs->nf*3);
+                           + envs->nf*3);
 
 FINT CINT4c1e_cart_drv(double *out, FINT *dims, CINTEnvVars *envs, CINTOpt *opt,
                        double *cache)