4 files changed, 34 insertions, 27 deletions

diff --git a/src/solvers/h48/gendata_cocsep.h b/src/solvers/h48/gendata_cocsep.h
index fd483ce..a8142b4 100644
--- a/src/solvers/h48/gendata_cocsep.h
+++ b/src/solvers/h48/gendata_cocsep.h
@@ -5,8 +5,7 @@ STATIC size_t gendata_cocsep(void *, uint64_t *, cube_t *);
 STATIC uint32_t gendata_cocsep_dfs(cocsep_dfs_arg_t *);
 STATIC void getdistribution_cocsep(const uint32_t *, uint64_t [static 21]);
 
-STATIC_INLINE int8_t get_h48_cdata(
-    cube_t, _Atomic const uint32_t *, uint32_t *);
+STATIC_INLINE int8_t get_h48_cdata(cube_t, const uint32_t *, uint32_t *);
 
 /*
 Each element of the cocsep table is a uint32_t used as follows:
@@ -159,7 +158,7 @@ set_visited(uint8_t *a, int64_t i)
 }
 
 STATIC_INLINE int8_t
-get_h48_cdata(cube_t cube, _Atomic const uint32_t *cocsepdata, uint32_t *cdata)
+get_h48_cdata(cube_t cube, const uint32_t *cocsepdata, uint32_t *cdata)
 {
         int64_t coord;
 
diff --git a/src/solvers/h48/gendata_h48.h b/src/solvers/h48/gendata_h48.h
index 391c97b..4124d56 100644
--- a/src/solvers/h48/gendata_h48.h
+++ b/src/solvers/h48/gendata_h48.h
@@ -31,13 +31,15 @@ STATIC tableinfo_t makeinfo_h48k2(gendata_h48_arg_t *);
 STATIC void getdistribution_h48(_Atomic const uint8_t *,
     uint64_t [static INFO_DISTRIBUTION_LEN], uint8_t, uint8_t);
 
-STATIC _Atomic const uint32_t *get_cocsepdata_constptr(const void *);
-STATIC _Atomic const uint8_t *get_h48data_constptr(const void *);
+STATIC const uint32_t *get_cocsepdata_constptr(const void *);
+STATIC const uint8_t *get_h48data_constptr(const void *);
 
-STATIC_INLINE uint8_t get_h48_pval(_Atomic const uint8_t *, int64_t, uint8_t);
+STATIC_INLINE uint8_t get_h48_pval_atomic(
+    _Atomic const uint8_t *, int64_t, uint8_t);
+STATIC_INLINE uint8_t get_h48_pval(const uint8_t *, int64_t, uint8_t);
 STATIC_INLINE void set_h48_pval(_Atomic uint8_t *, int64_t, uint8_t, uint8_t);
 STATIC_INLINE uint8_t get_h48_bound(
-    cube_t, uint32_t, uint8_t, uint8_t, _Atomic const uint8_t *);
+    cube_t, uint32_t, uint8_t, uint8_t, const uint8_t *);
 
 size_t gendata_h48_derive(uint8_t, const void *, void *);
 
@@ -196,7 +198,7 @@ gendata_h48h0k4(gendata_h48_arg_t *arg)
                         pthread_join(thread[t], NULL);
 
                 for (i = 0, cc = 0; i < h48max; i++) {
-                        val = get_h48_pval(table, i, 4);
+                        val = get_h48_pval_atomic(table, i, 4);
                         cc += val == d;
                 }
 
@@ -242,7 +244,7 @@ gendata_h48h0k4_runthread(void *arg)
          * Otherwise, scan all neighbors of unvisited coordinates.
          */
         for (i = bfsarg->start; i < bfsarg->end; i++) {
-                c = get_h48_pval(bfsarg->table, i, 4);
+                c = get_h48_pval_atomic(bfsarg->table, i, 4);
 
                 if ((bfsarg->depth < breakpoint && c != bfsarg->depth - 1) ||
                     (bfsarg->depth >= breakpoint && c != 0xF))
@@ -252,7 +254,7 @@ gendata_h48h0k4_runthread(void *arg)
                 for (m = 0; m < 18; m++) {
                         moved = move(cube, m);
                         j = coord_h48(moved, bfsarg->cocsepdata, 0);
-                        c = get_h48_pval(bfsarg->table, j, 4);
+                        c = get_h48_pval_atomic(bfsarg->table, j, 4);
                         if (bfsarg->depth < breakpoint) {
                                 if (c <= bfsarg->depth)
                                         continue;
@@ -387,7 +389,7 @@ gendata_h48k2(gendata_h48_arg_t *arg)
         h48map_destroy(&shortcubes);
 
         for (j = 0; j < H48_COORDMAX(arg->h); j++) {
-                t = get_h48_pval(table, j, 2);
+                t = get_h48_pval_atomic(table, j, 2);
                 arg->info.distribution[t]++;
         }
 
@@ -522,7 +524,7 @@ gendata_h48_mark(gendata_h48_mark_t *arg)
 
         FOREACH_H48SIM(arg->cube, arg->cocsepdata, arg->selfsim,
                 coord = coord_h48(arg->cube, arg->cocsepdata, arg->h);
-                oldval = get_h48_pval(arg->table, coord, arg->k);
+                oldval = get_h48_pval_atomic(arg->table, coord, arg->k);
                 newval = (uint8_t)MAX(arg->depth, 0);
                 if (newval < oldval) {
                         mutex = H48_INDEX(coord, arg->k) % CHUNKS;
@@ -546,7 +548,7 @@ gendata_h48k2_dfs_stop(cube_t cube, int8_t depth, h48k2_dfs_arg_t *arg)
                 coord = coord_h48(cube, arg->cocsepdata, arg->h);
                 mutex = H48_INDEX(coord, arg->k) % CHUNKS;
                 pthread_mutex_lock(arg->table_mutex[mutex]);
-                oldval = get_h48_pval(arg->table, coord, arg->k);
+                oldval = get_h48_pval_atomic(arg->table, coord, arg->k);
                 pthread_mutex_unlock(arg->table_mutex[mutex]);
                 return oldval <= depth;
         } else {
@@ -613,25 +615,31 @@ getdistribution_h48(
 
         h48max = H48_COORDMAX(h);
         for (i = 0; i < h48max; i++) {
-                val = get_h48_pval(table, i, k);
+                val = get_h48_pval_atomic(table, i, k);
                 distr[val]++;
         }
 }
 
-STATIC _Atomic const uint32_t *
+STATIC const uint32_t *
 get_cocsepdata_constptr(const void *data)
 {
-        return (_Atomic uint32_t *)((char *)data + INFOSIZE);
+        return (uint32_t *)((char *)data + INFOSIZE);
 }
 
-STATIC _Atomic const uint8_t *
+STATIC const uint8_t *
 get_h48data_constptr(const void *data)
 {
-        return (_Atomic uint8_t *)data + COCSEP_FULLSIZE + INFOSIZE;
+        return (uint8_t *)data + COCSEP_FULLSIZE + INFOSIZE;
 }
 
 STATIC_INLINE uint8_t
-get_h48_pval(_Atomic const uint8_t *table, int64_t i, uint8_t k)
+get_h48_pval_atomic(_Atomic const uint8_t *table, int64_t i, uint8_t k)
+{
+        return (table[H48_INDEX(i, k)] & H48_MASK(i, k)) >> H48_SHIFT(i, k);
+}
+
+STATIC_INLINE uint8_t
+get_h48_pval(const uint8_t *table, int64_t i, uint8_t k)
 {
         return (table[H48_INDEX(i, k)] & H48_MASK(i, k)) >> H48_SHIFT(i, k);
 }
@@ -649,7 +657,7 @@ get_h48_bound(
         uint32_t cdata,
         uint8_t h,
         uint8_t k,
-        _Atomic const uint8_t *table
+        const uint8_t *table
 ) {
         int64_t coord;
 
@@ -662,7 +670,7 @@ gendata_h48_derive(uint8_t h, const void *fulltable, void *buf)
 {
         size_t cocsepsize, h48size;
         uint8_t val_full, val_derive;
-        _Atomic const uint8_t *h48full;
+        const uint8_t *h48full;
         _Atomic uint8_t *h48derive;
         int64_t i, j, h48max;
         gendata_h48_arg_t arg;
@@ -703,7 +711,7 @@ gendata_h48_derive(uint8_t h, const void *fulltable, void *buf)
                 goto gendata_h48_derive_error;
         }
 
-        h48full = (_Atomic const uint8_t *)fulltable + cocsepsize + INFOSIZE;
+        h48full = (const uint8_t *)fulltable + cocsepsize + INFOSIZE;
         h48derive = (_Atomic uint8_t *)arg.h48buf + INFOSIZE;
         memset(h48derive, 0xFF, H48_TABLESIZE(h, arg.k));
         memset(arg.info.distribution, 0,
@@ -715,7 +723,7 @@ gendata_h48_derive(uint8_t h, const void *fulltable, void *buf)
                         LOG("Processing %" PRId64 "th coordinate\n", i);
                 j = i >> (int64_t)(fulltableinfo.h48h - h);
                 val_full = get_h48_pval(h48full, i, arg.k);
-                val_derive = get_h48_pval(h48derive, j, arg.k);
+                val_derive = get_h48_pval_atomic(h48derive, j, arg.k);
                 set_h48_pval(
                     h48derive, j, arg.k, MIN(val_full, val_derive));
         }
diff --git a/src/solvers/h48/solve.h b/src/solvers/h48/solve.h
index c6fe8b7..3948b15 100644
--- a/src/solvers/h48/solve.h
+++ b/src/solvers/h48/solve.h
@@ -8,8 +8,8 @@ typedef struct {
         int64_t maxsolutions;
         uint8_t h;
         uint8_t k;
-        _Atomic const uint32_t *cocsepdata;
-        _Atomic const uint8_t *h48data;
+        const uint32_t *cocsepdata;
+        const uint8_t *h48data;
         char **nextsol;
         uint8_t nissbranch;
         int8_t npremoves;
diff --git a/src/solvers/h48/stats.h b/src/solvers/h48/stats.h
index 5f96189..ca4b72e 100644
--- a/src/solvers/h48/stats.h
+++ b/src/solvers/h48/stats.h
@@ -10,8 +10,8 @@ typedef struct {
         int8_t nmoves;
         int8_t depth;
         uint8_t moves[MAXLEN];
-        _Atomic const uint32_t *cocsepdata;
-        _Atomic const uint8_t *h48data;
+        const uint32_t *cocsepdata;
+        const uint8_t *h48data;
         char *s;
 } dfsarg_solveh48stats_t;