- /* normally if we're about to execute code from the same
- * CV that we used previously, we just use the existing
- * CX stack entry. However, its possible that in the
- * meantime we may have backtracked, popped from the save
- * stack, and undone the SAVECOMPPAD(s) associated with
- * PUSH_MULTICALL; in which case PL_comppad no longer
- * points to newcv's pad. */
+ /* Some notes about MULTICALL and the context and save stacks.
+ *
+ * In something like
+ * /...(?{ my $x)}...(?{ my $y)}...(?{ my $z)}.../
+ * since codeblocks don't introduce a new scope (so that
+ * local() etc accumulate), at the end of a successful
+ * match there will be a SAVEt_CLEARSV on the savestack
+ * for each of $x, $y, $z. If the three code blocks above
+ * happen to have come from different CVs (e.g. via
+ * embedded qr//s), then we must ensure that during any
+ * savestack unwinding, PL_comppad always points to the
+ * right pad at each moment. We achieve this by
+ * interleaving SAVEt_COMPPAD's on the savestack whenever
+ * there is a change of pad.
+ * In theory whenever we call a code block, we should
+ * push a CXt_SUB context, then pop it on return from
+ * that code block. This causes a bit of an issue in that
+ * normally popping a context also clears the savestack
+ * back to cx->blk_oldsaveix, but here we specifically
+ * don't want to clear the save stack on exit from the
+ * code block.
+ * Also for efficiency we don't want to keep pushing and
+ * popping the single SUB context as we backtrack etc.
+ * So instead, we push a single context the first time
+ * we need, it, then hang onto it until the end of this
+ * function. Whenever we encounter a new code block, we
+ * update the CV etc if that's changed. During the times
+ * in this function where we're not executing a code
+ * block, having the SUB context still there is a bit
+ * naughty - but we hope that no-one notices.
+ * When the SUB context is initially pushed, we fake up
+ * cx->blk_oldsaveix to be as if we'd pushed this context
+ * on first entry to S_regmatch rather than at some random
+ * point during the regexe execution. That way if we
+ * croak, popping the context stack will ensure that
+ * *everything* SAVEd by this function is undone and then
+ * the context popped, rather than e.g., popping the
+ * context (and restoring the original PL_comppad) then
+ * popping more of the savestack and restoring a bad
+ * PL_comppad.
+ */
+
+ /* If this is the first EVAL, push a MULTICALL. On
+ * subsequent calls, if we're executing a different CV, or
+ * if PL_comppad has got messed up from backtracking
+ * through SAVECOMPPADs, then refresh the context.
+ */