For s///r, don't call SvPV_force() on the original value. Resolves #97954.

8ca8a454f60a417f optimised the implementation of s///r by avoiding an
unconditional copy of the original value. However, it introduced a behaviour
regression where if original value happened to be one of a few particular
types, it could be modified by being forced to a string using SvPV_force().
The substitution was (correctly) performed on a copy of this string.

1 files changed, 8 insertions, 0 deletions

diff --git a/pp_hot.c b/pp_hot.c
index 6abbf19c8e..758d334e62 100644
--- a/pp_hot.c
+++ b/pp_hot.c
@@ -2294,6 +2294,14 @@ PP(pp_subst)
     else {
 	if (force_on_match) {
 	    force_on_match = 0;
+	    if (rpm->op_pmflags & PMf_NONDESTRUCT) {
+		/* I feel that it should be possible to avoid this mortal copy
+		   given that the code below copies into a new destination.
+		   However, I suspect it isn't worth the complexity of
+		   unravelling the C<goto force_it> for the small number of
+		   cases where it would be viable to drop into the copy code. */
+		TARG = sv_2mortal(newSVsv(TARG));
+	    }
 	    s = SvPV_force(TARG, len);
 	    goto force_it;
 	}