[ELF] Implement a workaround for libbacktrace

This commit enables mold to work with libbacktrace before 2022-04-06.

https://github.com/rui314/mold/issues/402

compress.cc

@@ -54,9 +54,33 @@ static std::vector<u8> do_compress(std::string_view input) {
   // on the compression size. +16 for Z_SYNC_FLUSH.
   std::vector<u8> buf(deflateBound(&strm, strm.avail_in) + 16);
 
+  // Compress data. It writes all compressed bytes except the last
+  // partial byte, so up to 7 bits can be held to be written to the
+  // buffer.
   strm.avail_out = buf.size();
   strm.next_out = buf.data();
+  CHECK(deflate(&strm, Z_BLOCK));
 
+  // This is a workaround for libbacktrace before 2022-04-06.
+  //
+  // Zlib is a bit stream, and what Z_SYNC_FLUSH does is to write a
+  // three bit value indicating the start of an uncompressed data
+  // block followed by four byte data 00 00 ff ff which indicates that
+  // the length of the block is zero. libbacktrace uses its own zlib
+  // inflate routine, and it had a bug that if that particular three
+  // bit value happens to end at a byte boundary, it accidentally
+  // skipped the next byte.
+  //
+  // In order to avoid triggering that bug, we should avoid calling
+  // deflate() with Z_SYNC_FLUSH if the current bit position is 5.
+  // If it's 5, we insert an empty block consisting of 10 bits so
+  // that the bit position is 7 in the next byte.
+  //
+  // https://github.com/ianlancetaylor/libbacktrace/pull/87
+  int nbits;
+  deflatePending(&strm, Z_NULL, &nbits);
+  if (nbits == 5)
+    CHECK(deflatePrime(&strm, 10, 2));
   CHECK(deflate(&strm, Z_SYNC_FLUSH));
 
   assert(strm.avail_out > 0);