#478 has_md5sum fails when first byte of md5sum is zero.

[kda]

This happens for me on Ubuntu 19.10. Both from repo and built from source.
This does not happen on Raspbian Stretch.

I'm guessing some subtlety about the compiler.
I've attached a patch.
I've attached an offending flac.

The patch inline:

--- flac-1.3.3/src/flac/decode.c        2018-08-25 01:56:06.666934412 -0700
+++ flac-1.3.3.kda/src/flac/decode.c    2020-03-05 09:44:42.825560660 -0800
@@ -1306,8 +1306,9 @@

        if(metadata->type == FLAC__METADATA_TYPE_STREAMINFO) {
                FLAC__uint64 skip, until;

+    FLAC__byte emptyMD5[16] =   { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
                decoder_session->got_stream_info = true;
-               decoder_session->has_md5sum = memcmp(metadata->data.stream_info.md5sum, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0", 16);
+               decoder_session->has_md5sum = memcmp(metadata->data.stream_info.md5sum, emptyMD5, 16);
                decoder_session->bps = metadata->data.stream_info.bits_per_sample;
                decoder_session->channels = metadata->data.stream_info.channels;
                decoder_session->sample_rate = metadata->data.stream_info.sample_rate;

[Till Schäfer]

I can confirm the bug on gentoo after rebuilding flac 1.3.3. with gcc 9.3.0 (worked with gcc 9.2.0).

[Till Schäfer]

downstream bug reference: https://bugs.gentoo.org/719792

[Till Schäfer]

my analysis about the gcc versions was wrong, working was gcc 8.3 and failing was gcc 9.{2,3}

[Till Schäfer]

an in depth analysis downstream (see gentoo bug https://bugs.gentoo.org/719792) has revealed, that this issue is occurring only with certain optimization levels.

Furthermore a minimal test example was created for assamber analysis (attached).
Compiled with gcc -O2 -S -o test.s test.c callback1 (the erroneous case), is reduced to
movzbl 56(%rsi), %eax
movl %eax, (%rdx)
ret

Calling gcc with -O1 actually calls memcmp.

[Till Schäfer]

Given the above analysis a more minmal patch, which produced a better assembly was created. (just adding a != 0 check to convert FLAC__bool to an actual bool (FLAC__bool is int)

[Martijn van Beurden]

I'm sorry this wasn't replied to earlier, but these trackers are pretty much abandoned. A workaround was included in the 1.3.4 release.

More information can be found here: https://github.com/xiph/flac/pull/227