CHG: test of re-adding @piwi's bucketsort to @blapost's crapt1 v3.3 imp.

diff --git a/client/nonce2key/crapto1.c b/client/nonce2key/crapto1.c
index 8754c3c924e5301c504bba27ae746b985d285356..59d6a8d6b545ffa65a2fdca1017204fb5562705d 100644 (file)
--- a/client/nonce2key/crapto1.c
+++ b/client/nonce2key/crapto1.c
@@ -24,36 +24,77 @@
 static uint8_t filterlut[1 << 20];\r
 static void __attribute__((constructor)) fill_lut()\r
 {\r
-        uint32_t i;\r
-        for(i = 0; i < 1 << 20; ++i)\r
-                filterlut[i] = filter(i);\r
+       uint32_t i;\r
+       for(i = 0; i < 1 << 20; ++i)\r
+                       filterlut[i] = filter(i);\r
 }\r
 #define filter(x) (filterlut[(x) & 0xfffff])\r
 #endif\r
 \r
-static void quicksort(uint32_t* const start, uint32_t* const stop)\r
-{\r
-       uint32_t *it = start + 1, *rit = stop, t;\r
 \r
-       if(it > rit)\r
-               return;\r
 \r
-       while(it < rit)\r
-               if(*it <= *start)\r
-                       ++it;\r
-               else if(*rit > *start)\r
-                       --rit;\r
-               else\r
-                       t = *it,  *it = *rit, *rit = t;\r
+typedef struct bucket {\r
+       uint32_t *head;\r
+       uint32_t *bp;\r
+} bucket_t;\r
+\r
+typedef bucket_t bucket_array_t[2][0x100];\r
+\r
+typedef struct bucket_info {\r
+       struct {\r
+               uint32_t *head, *tail;\r
+               } bucket_info[2][0x100];\r
+               uint32_t numbuckets;\r
+       } bucket_info_t;\r
+\r
+\r
+static void bucket_sort_intersect(uint32_t* const estart, uint32_t* const estop,\r
+                                                                 uint32_t* const ostart, uint32_t* const ostop,\r
+                                                                 bucket_info_t *bucket_info, bucket_array_t bucket)\r
+{\r
+       uint32_t *p1, *p2;\r
+       uint32_t *start[2];\r
+       uint32_t *stop[2];\r
+\r
+       start[0] = estart;\r
+       stop[0] = estop;\r
+       start[1] = ostart;\r
+       stop[1] = ostop;\r
+\r
+       // init buckets to be empty\r
+       for (uint32_t i = 0; i < 2; i++) {\r
+               for (uint32_t j = 0x00; j <= 0xff; j++) {\r
+                       bucket[i][j].bp = bucket[i][j].head;\r
+               }\r
+       }\r
+\r
+       // sort the lists into the buckets based on the MSB (contribution bits)\r
+       for (uint32_t i = 0; i < 2; i++) {\r
+               for (p1 = start[i]; p1 <= stop[i]; p1++) {\r
+                       uint32_t bucket_index = (*p1 & 0xff000000) >> 24;\r
+                       *(bucket[i][bucket_index].bp++) = *p1;\r
+               }\r
+       }\r
 \r
-       if(*rit >= *start)\r
-               --rit;\r
-       if(rit != start)\r
-               t = *rit,  *rit = *start, *start = t;\r
 \r
-       quicksort(start, rit - 1);\r
-       quicksort(rit + 1, stop);\r
+       // write back intersecting buckets as sorted list.\r
+       // fill in bucket_info with head and tail of the bucket contents in the list and number of non-empty buckets.\r
+       uint32_t nonempty_bucket;\r
+       for (uint32_t i = 0; i < 2; i++) {\r
+               p1 = start[i];\r
+               nonempty_bucket = 0;\r
+               for (uint32_t j = 0x00; j <= 0xff; j++) {\r
+                       if (bucket[0][j].bp != bucket[0][j].head && bucket[1][j].bp != bucket[1][j].head) { // non-empty intersecting buckets only\r
+                               bucket_info->bucket_info[i][nonempty_bucket].head = p1;\r
+                               for (p2 = bucket[i][j].head; p2 < bucket[i][j].bp; *p1++ = *p2++);\r
+                               bucket_info->bucket_info[i][nonempty_bucket].tail = p1 - 1;\r
+                               nonempty_bucket++;\r
+                       }\r
+               }\r
+               bucket_info->numbuckets = nonempty_bucket;\r
+               }\r
 }\r
+\r
 /** binsearch\r
  * Binary search for the first occurence of *stop's MSB in sorted [start,stop]\r
  */\r
@@ -126,9 +167,10 @@ static inline void extend_table_simple(uint32_t *tbl, uint32_t **end, int bit)
 static struct Crypto1State*\r
 recover(uint32_t *o_head, uint32_t *o_tail, uint32_t oks,\r
        uint32_t *e_head, uint32_t *e_tail, uint32_t eks, int rem,\r
-       struct Crypto1State *sl, uint32_t in)\r
+       struct Crypto1State *sl, uint32_t in, bucket_array_t bucket)\r
 {\r
-       uint32_t *o, *e, i;\r
+       uint32_t *o, *e;\r
+       bucket_info_t bucket_info;\r
 \r
        if(rem == -1) {\r
                for(e = e_head; e <= e_tail; ++e) {\r
@@ -142,7 +184,7 @@ recover(uint32_t *o_head, uint32_t *o_tail, uint32_t oks,
                return sl;\r
        }\r
 \r
-       for(i = 0; i < 4 && rem--; i++) {\r
+       for(uint32_t i = 0; i < 4 && rem--; i++) {\r
                oks >>= 1;\r
                eks >>= 1;\r
                in >>= 2;\r
@@ -157,20 +199,13 @@ recover(uint32_t *o_head, uint32_t *o_tail, uint32_t oks,
                        return sl;\r
        }\r
 \r
-       quicksort(o_head, o_tail);\r
-       quicksort(e_head, e_tail);\r
+       bucket_sort_intersect(e_head, e_tail, o_head, o_tail, &bucket_info, bucket);\r
 \r
-       while(o_tail >= o_head && e_tail >= e_head)\r
-               if(((*o_tail ^ *e_tail) >> 24) == 0) {\r
-                       o_tail = binsearch(o_head, o = o_tail);\r
-                       e_tail = binsearch(e_head, e = e_tail);\r
-                       sl = recover(o_tail--, o, oks,\r
-                                    e_tail--, e, eks, rem, sl, in);\r
+       for (int i = bucket_info.numbuckets - 1; i >= 0; i--) {\r
+               sl = recover(bucket_info.bucket_info[1][i].head, bucket_info.bucket_info[1][i].tail, oks,\r
+                                        bucket_info.bucket_info[0][i].head, bucket_info.bucket_info[0][i].tail, eks,\r
+                                        rem, sl, in, bucket);\r
        }\r
-               else if(*o_tail > *e_tail)\r
-                       o_tail = binsearch(o_head, o_tail) - 1;\r
-               else\r
-                       e_tail = binsearch(e_head, e_tail) - 1;\r
 \r
        return sl;\r
 }\r
@@ -203,6 +238,17 @@ struct Crypto1State* lfsr_recovery32(uint32_t ks2, uint32_t in)
 \r
        statelist->odd = statelist->even = 0;\r
 \r
+       // allocate memory for out of place bucket_sort\r
+       bucket_array_t bucket;\r
+       for (uint32_t i = 0; i < 2; i++)\r
+               for (uint32_t j = 0; j <= 0xff; j++) {\r
+                       bucket[i][j].head = malloc(sizeof(uint32_t)<<14);\r
+                       if (!bucket[i][j].head) {\r
+                               goto out;\r
+                       }\r
+               }\r
+\r
+\r
        // initialize statelists: add all possible states which would result into the rightmost 2 bits of the keystream\r
        for(i = 1 << 20; i >= 0; --i) {\r
                if(filter(i) == (oks & 1))\r
@@ -220,13 +266,18 @@ struct Crypto1State* lfsr_recovery32(uint32_t ks2, uint32_t in)
        // the statelists now contain all states which could have generated the last 10 Bits of the keystream.\r
        // 22 bits to go to recover 32 bits in total. From now on, we need to take the "in"\r
        // parameter into account.\r
-       in = (in >> 16 & 0xff) | (in << 16) | (in & 0xff00);\r
+       in = (in >> 16 & 0xff) | (in << 16) | (in & 0xff00);            // Byte swapping\r
        recover(odd_head, odd_tail, oks,\r
-               even_head, even_tail, eks, 11, statelist, in << 1);\r
+               even_head, even_tail, eks, 11, statelist, in << 1, bucket);\r
+\r
 \r
 out:\r
        free(odd_head);\r
        free(even_head);\r
+       for (uint32_t i = 0; i < 2; i++)\r
+               for (uint32_t j = 0; j <= 0xff; j++)\r
+                       free(bucket[i][j].head);\r
+\r
        return statelist;\r
 }\r
 \r