From: Klaus Sembritzki Date: Sun, 25 Jan 2026 14:21:57 GMT Subject: Re: The SHA256 of "xy\n" (ASCII, no CRLF) contains 1337, ACBAD in za, and I am 1aa Message-ID: In-Reply-To: If vanilla-SHA256 continues to be used in vanilla-Git in the future, I suggest that platforms like GitHub hash internally. On Sun, Jan 25, 2026 at 12:25 PM Klaus Sembritzki wrote: > > On Sun, Jan 25, 2026 at 10:47 AM Klaus Sembritzki wrote: > > > > On Sat, Jan 24, 2026 at 10:25 AM Klaus Sembritzki wrote: > > > > > > On Sat, Jan 24, 2026 at 8:28 AM Jeff King wrote: > > > > > > > > On Fri, Jan 23, 2026 at 02:36:41PM -0800, Junio C Hamano wrote: > > > > > > > > > Jeff King writes: > > > > > > > > > > > On Fri, Jan 23, 2026 at 09:16:46PM +0100, Klaus Sembritzki wrote: > > > > > > > > > > > >> $ # My initials (ks): 1aa > > > > > >> $ echo ks | sha256sum > > > > > >> $ 1aa44e718d5bc9b7ff2003dbbb6f154e16636d5c2128ffce4751af5124b65337 > > > > > >> > > > > > >> $ # 50566750337 > > > > > >> $ echo thinking | sha256sum > > > > > >> $ 50566750337beb9e98e553fd9196d10576f9eb0cbc6b66e2586b9d73af4f352f > > > > > > > > > > > > Oh man, I've got deadbeef! > > > > > > > > > > > > $ echo jk35252822 | sha256sum > > > > > > 33f1a74529870456c56ad97c59cfed6bdeadbeef9b9bc3f4ff49bb203e36f96b > > > > > > > > > > > > What could it all mean? > > > > > > > > > > Sorry, but I have to admit that I completely lack humor receptor > > > > > cells. > > > > > > > > Probably because it was not that funny. :) > > > > > > > > The original message seemed to be looking for Numerology-style meanings > > > > in random data. I wasn't sure if it was serious or not, but I could not > > > > resist either playing along (if not) or trolling (if so). > > > > > > > > But here's my deadbeef brute-force program for fun. > > > > > > > > -Peff > > > > > > > > -- >8 -- > > > > #include > > > > #include > > > > #include > > > > > > > > int main(int argc, const char **argv) > > > > { > > > > const char needle[] = { 0xde, 0xad, 0xbe, 0xef }; > > > > const EVP_MD *algo = EVP_sha256(); > > > > EVP_MD_CTX *ctx = EVP_MD_CTX_new(); > > > > > > > > EVP_DigestInit_ex(ctx, algo, NULL); > > > > while (*++argv) > > > > EVP_DigestUpdate(ctx, *argv, strlen(*argv)); > > > > > > > > for (unsigned i = 0; ; i++) { > > > > char buf[16]; > > > > char *p; > > > > unsigned char digest[32]; > > > > EVP_MD_CTX *copy = EVP_MD_CTX_dup(ctx); > > > > > > > > p = buf + sizeof(buf); > > > > for (unsigned v = i; v; v /= 10) > > > > *--p = '0' + (v % 10); > > > > EVP_DigestUpdate(copy, p, buf + sizeof(buf) - p); > > > > EVP_DigestUpdate(copy, "\n", 1); > > > > EVP_DigestFinal_ex(copy, digest, NULL); > > > > EVP_MD_CTX_free(copy); > > > > if (memmem(digest, sizeof(digest), needle, sizeof(needle))) > > > > printf("%d\n", i); > > > > } > > > > } > > > > > > Incrementing the MSB instead of the LSB (indexing naturally starts > > > with 1 in that case) seems to improve the performance, and it finds > > > different solutions, if the program is terminated early. > > > The rationale is that there is autocorrelation in the observations, > > > though I cannot judge what that means in this concrete example. The > > > speedup is not that dramatic here, so SHA256 seems to be pretty > > > random. > > > > > > #include > > > #include > > > #include > > > #include > > > > > > #ifdef POLYFILL > > > EVP_MD_CTX *EVP_MD_CTX_dup(const EVP_MD_CTX *in) > > > { > > > EVP_MD_CTX *out = EVP_MD_CTX_new(); > > > > > > if (out != NULL && !EVP_MD_CTX_copy_ex(out, in)) { > > > EVP_MD_CTX_free(out); > > > out = NULL; > > > } > > > return out; > > > } > > > #endif > > > > > > #define REVERTED_BIT(n, position, width) (((n & (1 << position)) >> > > > position) << (width - position - 2)) > > > > > > uint32_t revert_bits(uint32_t n, uint32_t width) { > > > uint32_t result = 0; > > > for (uint32_t i = 0; i < width; ++i) { > > > result |= REVERTED_BIT(n, i, width); > > > } > > > return result; > > > } > > > > > > int main(int argc, const char **argv) > > > { > > > const char needle[] = { 0xde, 0xad, 0xbe, 0xef }; > > > const EVP_MD *algo = EVP_sha256(); > > > EVP_MD_CTX *ctx = EVP_MD_CTX_new(); > > > > > > EVP_DigestInit_ex(ctx, algo, NULL); > > > while (*++argv) > > > EVP_DigestUpdate(ctx, *argv, strlen(*argv)); > > > > > > for (uint32_t n = 1; ; n++) { > > > #if (COUNTER_WIDTH != 0) > > > uint32_t i = revert_bits(n, COUNTER_WIDTH); > > > #else > > > uint32_t i = n; > > > #endif > > > // printf("%u %u\n", n, i); > > > char buf[16]; > > > char *p; > > > unsigned char digest[32]; > > > EVP_MD_CTX *copy = EVP_MD_CTX_dup(ctx); > > > > > > p = buf + sizeof(buf); > > > for (uint32_t v = i; v; v /= 10) > > > *--p = '0' + (v % 10); > > > EVP_DigestUpdate(copy, p, buf + sizeof(buf) - p); > > > EVP_DigestUpdate(copy, "\n", 1); > > > EVP_DigestFinal_ex(copy, digest, NULL); > > > EVP_MD_CTX_free(copy); > > > if (memmem(digest, sizeof(digest), needle, sizeof(needle))) > > > { > > > printf("counter width: %2u | n: %10u | i: %10u\n", > > > COUNTER_WIDTH, n, i); > > > return 0; > > > } > > > } > > > return 1; > > > } > > > > > > cc -DPOLYFILL -DCOUNTER_WIDTH=32 jk_evp.c -lssl -lcrypto -o jk_evp_msb_32 > > > cc -DPOLYFILL -DCOUNTER_WIDTH=31 jk_evp.c -lssl -lcrypto -o jk_evp_msb_31 > > > cc -DPOLYFILL -DCOUNTER_WIDTH=30 jk_evp.c -lssl -lcrypto -o jk_evp_msb_30 > > > cc -DPOLYFILL -DCOUNTER_WIDTH=29 jk_evp.c -lssl -lcrypto -o jk_evp_msb_29 > > > cc -DPOLYFILL -DCOUNTER_WIDTH=0 jk_evp.c -lssl -lcrypto -o jk_evp_0 > > > counter width: 32 | n: 103832253 | i: 1588397616 > > > counter width: 31 | n: 62413559 | i: 1003915120 > > > counter width: 30 | n: 166340413 | i: 396135154 > > > counter width: 29 | n: 137077701 | i: 171728193 > > > counter width: 0 | n: 171728193 | i: 171728193 > > > > I have to admit there is a bug in my previous code, it only works > > correctly for even numbers. It should have been: > > > > #define REVERTED_BIT(n, position, width) (((n & (1 << position)) >> > > position) << (width - position - 2 + (width & 1))) > > > > uint32_t revert_bits(uint32_t n, uint32_t width) { > > uint32_t result = 0; > > for (uint32_t i = 0; i < width; ++i) { > > result |= REVERTED_BIT(n, i, width); > > } > > return result; > > } > > > > The actual performance measurements are: > > counter width: 32 | n: 103832253 | i: 1588397616 > > counter width: 31 | n: 103832253 | i: 1588397616 > > counter width: 30 | n: 166340413 | i: 396135154 > > counter width: 29 | n: 166340413 | i: 396135154 > > counter width: 28 | n: 268041599 | i: 2281045247 # 2**28 is close to > > the first finding when counting by incrementing the LSB, so this is > > slow. > > counter width: 27: This does not find anything, because 2**27 is lower > > than the first finding when counting by incrementing the LSB. > > counter width: 0 | n: 171728193 | i: 171728193 # > > math.log2(171728193) = 27.35555166834193 > > As an interim solution, it could make sense to leave small input > unmodified, and to permute large texts to get rid of the natural > language manifold.