There was no crash on 2.7, and the fix for crash was backported to 3.6. This PR adds a new feature (the previous implementation didn't work, it was a dead code except OpenBSD where it caused a crash) that can help to catch memory bugs. I don't want to backport it.

(...) the fix for crash was backported to 3.6. This PR adds a new feature (the previous implementation didn't work, (...)

Oh, I didn't see that you already removed the memset() from 3.6.

In that case, I would suggest to backport this PR to Python 3.6 as well, to not loose the detection of "use after free" by debug hooks.

Left this on to @ned-deily. But note that this detection doesn't work since 3.4, and this change is much more complex than #4145.

My reaction is that the risk due to complexity of the changes outweigh the benefits so I would agree with @serhiy-storchaka that we should not backport this.

My reaction is that the risk due to complexity of the changes outweigh the benefits so I would agree with @serhiy-storchaka that we should not backport this.

Ok, fine.

@serhiy-storchaka: For Python 3.6, what do you think of using Python 2.7 simple code?

...
    if (nbytes < original_nbytes) {
        /* shrinking:  mark old extra memory dead */
        memset(q + nbytes, DEADBYTE, original_nbytes - nbytes + 2*SST);
    }

    /* Resize and add decorations. We may get a new pointer here, in which
     * case we didn't get the chance to mark the old memory with DEADBYTE,
     * but we live with that.
     */
    q = (uchar *)PyObject_Realloc(q - 2*SST, total);
...

The code expect that realloc() never fails when shrinking. While Xavier De Gaye and me made realloc() failing anyway using tools to inject memory allocation failures, it seems like in the wild with system realloc(), realloc() never fails in that case.

If you agree, I can prepare a PR if you want. Or you can do, as you want :-)

For the master branch, I really prefer this PR since it's the "most correct" code: it supports hardcore testing as Xavier and me did ;-)

This is what the initial version of #3844 did. But I agree with your arguments, adding this in 3.6 is risky. The third-party code in 3.6 can expect that realloc() don't change the content on failure.