Fixed out-of-bound memory access issue for large inputs.

IlyaGrebnov · IlyaGrebnov · commit 918b4903647d · 2024-11-18T20:45:04.000-08:00
diff --git a/CHANGES b/CHANGES
@@ -1,3 +1,6 @@
+Changes in 2.8.6 (November 18, 2024)
+- Fixed out-of-bound memory access issue for large inputs.
+
 Changes in 2.8.5 (July 31, 2024)
 - Miscellaneous changes to reduce compiler warnings about implicit functions.
 
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.10)
 
-project(libsais VERSION 2.8.5 LANGUAGES C DESCRIPTION "libsais is a library for linear time suffix array, longest common prefix array and burrows wheeler transform construction based on induced sorting algorithm.")
+project(libsais VERSION 2.8.6 LANGUAGES C DESCRIPTION "libsais is a library for linear time suffix array, longest common prefix array and burrows wheeler transform construction based on induced sorting algorithm.")
 
 set(CMAKE_C_STANDARD 99)
 set(CMAKE_C_STANDARD_REQUIRED ON)
diff --git a/README.md b/README.md
@@ -23,6 +23,8 @@ The libsais provides simple C99 API to construct suffix array and Burrows-Wheele
 The libsais is released under the [Apache License Version 2.0](LICENSE "Apache license")
 
 ## Changes
+* November 18, 2024 (2.8.6)
+  * Fixed out-of-bound memory access issue for large inputs.
 * July 31, 2024 (2.8.5)
   * Miscellaneous changes to reduce compiler warnings about implicit functions.
 * June 13, 2024 (2.8.4)
diff --git a/VERSION b/VERSION
@@ -1 +1 @@
-2.8.5
+2.8.6
diff --git a/include/libsais.h b/include/libsais.h
@@ -26,8 +26,8 @@ Please see the file LICENSE for full copyright information.
 
 #define LIBSAIS_VERSION_MAJOR   2
 #define LIBSAIS_VERSION_MINOR   8
-#define LIBSAIS_VERSION_PATCH   5
-#define LIBSAIS_VERSION_STRING  "2.8.5"
+#define LIBSAIS_VERSION_PATCH   6
+#define LIBSAIS_VERSION_STRING  "2.8.6"
 
 #ifdef _WIN32
     #ifdef LIBSAIS_SHARED
diff --git a/include/libsais16.h b/include/libsais16.h
@@ -26,8 +26,8 @@ Please see the file LICENSE for full copyright information.
 
 #define LIBSAIS16_VERSION_MAJOR   2
 #define LIBSAIS16_VERSION_MINOR   8
-#define LIBSAIS16_VERSION_PATCH   5
-#define LIBSAIS16_VERSION_STRING  "2.8.5"
+#define LIBSAIS16_VERSION_PATCH   6
+#define LIBSAIS16_VERSION_STRING  "2.8.6"
 
 #ifdef _WIN32
     #ifdef LIBSAIS_SHARED
diff --git a/include/libsais16x64.h b/include/libsais16x64.h
@@ -26,8 +26,8 @@ Please see the file LICENSE for full copyright information.
 
 #define LIBSAIS16X64_VERSION_MAJOR   2
 #define LIBSAIS16X64_VERSION_MINOR   8
-#define LIBSAIS16X64_VERSION_PATCH   5
-#define LIBSAIS16X64_VERSION_STRING  "2.8.5"
+#define LIBSAIS16X64_VERSION_PATCH   6
+#define LIBSAIS16X64_VERSION_STRING  "2.8.6"
 
 #ifdef _WIN32
     #ifdef LIBSAIS_SHARED
diff --git a/include/libsais64.h b/include/libsais64.h
@@ -26,8 +26,8 @@ Please see the file LICENSE for full copyright information.
 
 #define LIBSAIS64_VERSION_MAJOR   2
 #define LIBSAIS64_VERSION_MINOR   8
-#define LIBSAIS64_VERSION_PATCH   5
-#define LIBSAIS64_VERSION_STRING  "2.8.5"
+#define LIBSAIS64_VERSION_PATCH   6
+#define LIBSAIS64_VERSION_STRING  "2.8.6"
 
 #ifdef _WIN32
     #ifdef LIBSAIS_SHARED
diff --git a/src/libsais.c b/src/libsais.c
@@ -6270,11 +6270,11 @@ static sa_sint_t libsais_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t *
 {
     fs = fs < (SAINT_MAX - n) ? fs : (SAINT_MAX - n);
 
-    if (k > 0 && ((fs / k >= 6) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6)))
+    if (k > 0 && ((fs / k >= 6) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 6 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 6 ? (sa_sint_t *)libsais_align_up(&SA[n + fs - 6 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 6 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais_count_and_gather_lms_suffixes_32s_4k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
@@ -6329,11 +6329,11 @@ static sa_sint_t libsais_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t *
 
         return 0;
     }
-    else if (k > 0 && ((fs / k >= 4) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4)))
+    else if (k > 0 && (n <= SAINT_MAX / 2) && ((fs / k >= 4) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 4 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 4 ? (sa_sint_t *)libsais_align_up(&SA[n + fs - 4 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 4 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
@@ -6374,11 +6374,11 @@ static sa_sint_t libsais_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t *
 
         return 0;
     }
-    else if (k > 0 && ((fs / k >= 2) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2)))
+    else if (k > 0 && ((fs / k >= 2) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 2 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 2 ? (sa_sint_t *)libsais_align_up(&SA[n + fs - 2 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 2 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
diff --git a/src/libsais16.c b/src/libsais16.c
@@ -6248,11 +6248,11 @@ static sa_sint_t libsais16_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t
 {
     fs = fs < (SAINT_MAX - n) ? fs : (SAINT_MAX - n);
 
-    if (k > 0 && ((fs / k >= 6) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6)))
+    if (k > 0 && ((fs / k >= 6) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 6 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 6 ? (sa_sint_t *)libsais16_align_up(&SA[n + fs - 6 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 6 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais16_count_and_gather_lms_suffixes_32s_4k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
@@ -6307,11 +6307,11 @@ static sa_sint_t libsais16_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t
 
         return 0;
     }
-    else if (k > 0 && ((fs / k >= 4) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4)))
+    else if (k > 0 && (n <= SAINT_MAX / 2) && ((fs / k >= 4) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 4 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 4 ? (sa_sint_t *)libsais16_align_up(&SA[n + fs - 4 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 4 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais16_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
@@ -6352,11 +6352,11 @@ static sa_sint_t libsais16_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t
 
         return 0;
     }
-    else if (k > 0 && ((fs / k >= 2) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2)))
+    else if (k > 0 && ((fs / k >= 2) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 2 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 2 ? (sa_sint_t *)libsais16_align_up(&SA[n + fs - 2 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 2 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais16_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
diff --git a/src/libsais16x64.c b/src/libsais16x64.c
@@ -6286,7 +6286,7 @@ static sa_sint_t libsais16x64_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint
     if (n <= INT32_MAX)
     {
         sa_sint_t new_fs = (fs + fs + n + n) <= INT32_MAX ? (fs + fs + n) : INT32_MAX - n;
-        if ((new_fs / k >= 4) || (new_fs >= fs))
+        if ((new_fs / k >= 6) || (new_fs / k >= 4 && n <= INT32_MAX / 2) || (new_fs / k < 4 && new_fs >= fs))
         {
             libsais16x64_convert_inplace_64u_to_32u((uint32_t *)(void *)T, 0, n);
 
@@ -6304,11 +6304,11 @@ static sa_sint_t libsais16x64_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint
         }
     }
 
-    if (k > 0 && ((fs / k >= 6) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6)))
+    if (k > 0 && ((fs / k >= 6) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 6 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 6 ? (sa_sint_t *)libsais16x64_align_up(&SA[n + fs - 6 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 6 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais16x64_count_and_gather_lms_suffixes_32s_4k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
@@ -6363,11 +6363,11 @@ static sa_sint_t libsais16x64_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint
 
         return 0;
     }
-    else if (k > 0 && ((fs / k >= 4) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4)))
+    else if (k > 0 && (n <= SAINT_MAX / 2) && ((fs / k >= 4) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 4 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 4 ? (sa_sint_t *)libsais16x64_align_up(&SA[n + fs - 4 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 4 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais16x64_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
@@ -6408,11 +6408,11 @@ static sa_sint_t libsais16x64_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint
 
         return 0;
     }
-    else if (k > 0 && ((fs / k >= 2) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2)))
+    else if (k > 0 && ((fs / k >= 2) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 2 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 2 ? (sa_sint_t *)libsais16x64_align_up(&SA[n + fs - 2 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 2 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais16x64_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
diff --git a/src/libsais64.c b/src/libsais64.c
@@ -6308,7 +6308,7 @@ static sa_sint_t libsais64_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t
     if (n <= INT32_MAX)
     {
         sa_sint_t new_fs = (fs + fs + n + n) <= INT32_MAX ? (fs + fs + n) : INT32_MAX - n;
-        if ((new_fs / k >= 4) || (new_fs >= fs))
+        if ((new_fs / k >= 6) || (new_fs / k >= 4 && n <= INT32_MAX / 2) || (new_fs / k < 4 && new_fs >= fs))
         {
             libsais64_convert_inplace_64u_to_32u((uint32_t *)(void *)T, 0, n);
 
@@ -6326,11 +6326,11 @@ static sa_sint_t libsais64_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t
         }
     }
 
-    if (k > 0 && ((fs / k >= 6) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6)))
+    if (k > 0 && ((fs / k >= 6) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 6 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 6 ? (sa_sint_t *)libsais64_align_up(&SA[n + fs - 6 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 6 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais64_count_and_gather_lms_suffixes_32s_4k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
@@ -6385,11 +6385,11 @@ static sa_sint_t libsais64_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t
 
         return 0;
     }
-    else if (k > 0 && ((fs / k >= 4) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4)))
+    else if (k > 0 && (n <= SAINT_MAX / 2) && ((fs / k >= 4) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 4 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 4 ? (sa_sint_t *)libsais64_align_up(&SA[n + fs - 4 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 4 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais64_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)
@@ -6430,11 +6430,11 @@ static sa_sint_t libsais64_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t
 
         return 0;
     }
-    else if (k > 0 && ((fs / k >= 2) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2)))
+    else if (k > 0 && ((fs / k >= 2) || (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1)))
     {
         sa_sint_t alignment = (fs - 1024) / k >= 2 ? (sa_sint_t)1024 : (sa_sint_t)16;
         sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 2 ? (sa_sint_t *)libsais64_align_up(&SA[n + fs - 2 * (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 2 * (fast_sint_t)k];
-        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2) ? local_buffer : buckets;
+        buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1) ? local_buffer : buckets;
 
         sa_sint_t m = libsais64_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);
         if (m > 1)

Original file line number	Diff line number	Diff line change
`@@ -6270,11 +6270,11 @@ static sa_sint_t libsais_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t *`
`6270`	`6270`	`{`
`6271`	`6271`	`fs = fs < (SAINT_MAX - n) ? fs : (SAINT_MAX - n);`
`6272`	`6272`
`6273`		`- if (k > 0 && ((fs / k >= 6) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6)))`
	`6273`	`+ if (k > 0 && ((fs / k >= 6) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1)))`
`6274`	`6274`	`{`
`6275`	`6275`	`sa_sint_t alignment = (fs - 1024) / k >= 6 ? (sa_sint_t)1024 : (sa_sint_t)16;`
`6276`	`6276`	`sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 6 ? (sa_sint_t )libsais_align_up(&SA[n + fs - 6 (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 6 * (fast_sint_t)k];`
`6277`		`- buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6) ? local_buffer : buckets;`
	`6277`	`+ buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1) ? local_buffer : buckets;`
`6278`	`6278`
`6279`	`6279`	`sa_sint_t m = libsais_count_and_gather_lms_suffixes_32s_4k_omp(T, SA, n, k, buckets, threads, thread_state);`
`6280`	`6280`	`if (m > 1)`
`@@ -6329,11 +6329,11 @@ static sa_sint_t libsais_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t *`
`6329`	`6329`
`6330`	`6330`	`return 0;`
`6331`	`6331`	`}`
`6332`		`- else if (k > 0 && ((fs / k >= 4) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4)))`
	`6332`	`+ else if (k > 0 && (n <= SAINT_MAX / 2) && ((fs / k >= 4) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1)))`
`6333`	`6333`	`{`
`6334`	`6334`	`sa_sint_t alignment = (fs - 1024) / k >= 4 ? (sa_sint_t)1024 : (sa_sint_t)16;`
`6335`	`6335`	`sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 4 ? (sa_sint_t )libsais_align_up(&SA[n + fs - 4 (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 4 * (fast_sint_t)k];`
`6336`		`- buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4) ? local_buffer : buckets;`
	`6336`	`+ buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1) ? local_buffer : buckets;`
`6337`	`6337`
`6338`	`6338`	`sa_sint_t m = libsais_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);`
`6339`	`6339`	`if (m > 1)`
`@@ -6374,11 +6374,11 @@ static sa_sint_t libsais_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t *`
`6374`	`6374`
`6375`	`6375`	`return 0;`
`6376`	`6376`	`}`
`6377`		`- else if (k > 0 && ((fs / k >= 2) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2)))`
	`6377`	`+ else if (k > 0 && ((fs / k >= 2) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1)))`
`6378`	`6378`	`{`
`6379`	`6379`	`sa_sint_t alignment = (fs - 1024) / k >= 2 ? (sa_sint_t)1024 : (sa_sint_t)16;`
`6380`	`6380`	`sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 2 ? (sa_sint_t )libsais_align_up(&SA[n + fs - 2 (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 2 * (fast_sint_t)k];`
`6381`		`- buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2) ? local_buffer : buckets;`
	`6381`	`+ buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1) ? local_buffer : buckets;`
`6382`	`6382`
`6383`	`6383`	`sa_sint_t m = libsais_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);`
`6384`	`6384`	`if (m > 1)`
Original file line number	Diff line number	Diff line change
`@@ -6248,11 +6248,11 @@ static sa_sint_t libsais16_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t`
`6248`	`6248`	`{`
`6249`	`6249`	`fs = fs < (SAINT_MAX - n) ? fs : (SAINT_MAX - n);`
`6250`	`6250`
`6251`		`- if (k > 0 && ((fs / k >= 6) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6)))`
	`6251`	`+ if (k > 0 && ((fs / k >= 6) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1)))`
`6252`	`6252`	`{`
`6253`	`6253`	`sa_sint_t alignment = (fs - 1024) / k >= 6 ? (sa_sint_t)1024 : (sa_sint_t)16;`
`6254`	`6254`	`sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 6 ? (sa_sint_t )libsais16_align_up(&SA[n + fs - 6 (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 6 * (fast_sint_t)k];`
`6255`		`- buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6) ? local_buffer : buckets;`
	`6255`	`+ buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 6 && threads == 1) ? local_buffer : buckets;`
`6256`	`6256`
`6257`	`6257`	`sa_sint_t m = libsais16_count_and_gather_lms_suffixes_32s_4k_omp(T, SA, n, k, buckets, threads, thread_state);`
`6258`	`6258`	`if (m > 1)`
`@@ -6307,11 +6307,11 @@ static sa_sint_t libsais16_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t`
`6307`	`6307`
`6308`	`6308`	`return 0;`
`6309`	`6309`	`}`
`6310`		`- else if (k > 0 && ((fs / k >= 4) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4)))`
	`6310`	`+ else if (k > 0 && (n <= SAINT_MAX / 2) && ((fs / k >= 4) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1)))`
`6311`	`6311`	`{`
`6312`	`6312`	`sa_sint_t alignment = (fs - 1024) / k >= 4 ? (sa_sint_t)1024 : (sa_sint_t)16;`
`6313`	`6313`	`sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 4 ? (sa_sint_t )libsais16_align_up(&SA[n + fs - 4 (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 4 * (fast_sint_t)k];`
`6314`		`- buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4) ? local_buffer : buckets;`
	`6314`	`+ buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 4 && threads == 1) ? local_buffer : buckets;`
`6315`	`6315`
`6316`	`6316`	`sa_sint_t m = libsais16_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);`
`6317`	`6317`	`if (m > 1)`
`@@ -6352,11 +6352,11 @@ static sa_sint_t libsais16_main_32s_recursion(sa_sint_t * RESTRICT T, sa_sint_t`
`6352`	`6352`
`6353`	`6353`	`return 0;`
`6354`	`6354`	`}`
`6355`		`- else if (k > 0 && ((fs / k >= 2) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2)))`
	`6355`	`+ else if (k > 0 && ((fs / k >= 2) \|\| (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1)))`
`6356`	`6356`	`{`
`6357`	`6357`	`sa_sint_t alignment = (fs - 1024) / k >= 2 ? (sa_sint_t)1024 : (sa_sint_t)16;`
`6358`	`6358`	`sa_sint_t * RESTRICT buckets = (fs - alignment) / k >= 2 ? (sa_sint_t )libsais16_align_up(&SA[n + fs - 2 (fast_sint_t)k - alignment], (size_t)alignment * sizeof(sa_sint_t)) : &SA[n + fs - 2 * (fast_sint_t)k];`
`6359`		`- buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2) ? local_buffer : buckets;`
	`6359`	`+ buckets = (LIBSAIS_LOCAL_BUFFER_SIZE / k >= 2 && threads == 1) ? local_buffer : buckets;`
`6360`	`6360`
`6361`	`6361`	`sa_sint_t m = libsais16_count_and_gather_lms_suffixes_32s_2k_omp(T, SA, n, k, buckets, threads, thread_state);`
`6362`	`6362`	`if (m > 1)`