todo

很好用的内存泄漏检测工具。要配合linux的gcc使用,安装后将你的编译器gcc命令换成yamd-gcc就行了

Hey Emacs, this is a -*- text -*- file.

To do:

- yamd.c is getting large and messy.  Should be split into several
  files.  This will cause a bit of namespace pollution and force me to
  type messy names starting with underscores, but is probably worth it.  

- Figure out possible /tmp race with script.

- Use file version numbers (yamd0nn.zip etc) and change the website to
  show past versions.

  First part done.

- log_vprintf really ought to use vsnprintf when possible.  It looks
  like DJGPP should have this before too long; that's a good thing.
  Don't forget to correctly handle the overflow case.

- Audit for correct aliasing so that gcc 2.95 and its
  -fstrict-aliasing can be used.

  Have had a look and nothing seems horribly wrong.

- Write the "interpreting logfile" docs.

- Fill freshly allocated memory with some (user-specified) pattern.
  Currently we get random garbage (if regular malloc is used) or zeros
  (if mmap); neither is probably very good for finding uninitialized
  memory bugs.

- autoconf might be a good idea.  The manual HAVE_* macros are getting
  messy.

- Add more INFO-level logging, of normal events.  ??? Like what ???

- Output could be prettier.  Work on do-syms.  Also have it find the
  executable name automagically.

- Add checking for use of uninitialized memory.  386 has no write-only
  pages, so set all pages not-present and have the fault handler barf
  if they are read.  Handling each page individually would be slower
  and involve a lot more bookkeeping, so maybe just deal with the
  whole block.  But: This will probably catch only the grossest bugs,
  so I wonder if it's really worth it.  Also, I'd rather see the fault
  handler gone.

- I wonder if the signal handler should be removed entirely.  This
  would have the following advantages:

  * Eliminates system-specific fault handling code.

  * Traceback stuff could be much simplified, as we will always be on
    the call chain.  Much better portability.

    But: Under Linux, when dynamically loaded, postmortem debugging
    can't use YAMD.  This means we need to describe the related block
    at once; we won't be able to get that info later.

    OTOH, I could write a GDB script that emulates describe_address by
    walking the block chain itself.  This makes it GDB-dependent,
    however, and maybe depending on postmortem debugging working is
    unwise... hmm.

    This may also just be a GDB or glibc/ld.so bug.  Try it with 4.18
    and 2.1.1.

- Clean up some of the ugly address arithmetic.

- Fix signal handler under DJGPP.  Hmm... dang hard.

- Eliminate all the -D's on the compilation command line by making
  some sort of config.h that the Makefile creates.

- Check for suitable DPMI server on startup; else warn user.