BER is all very well, but it turns out that it's better to store the

offset as either a byte (if <256), or a 0 byte with a STRLEN before.
"better" in that the reading can be inlined, and even then the object
code is smaller (function calls have space overhead). So goodbye
Perl_sv_read_offset() and hello SvOOK_offset().

p4raw-id: //depot/perl@32838

1 files changed, 63 insertions, 6 deletions

diff --git a/sv.h b/sv.h
index afa18dc527..e13b5ba3f4 100644
--- a/sv.h
+++ b/sv.h
@@ -653,10 +653,12 @@ Will also turn off the UTF-8 status.
 Returns a boolean indicating whether the SV contains a v-string.
 
 =for apidoc Am|U32|SvOOK|SV* sv
-Returns a U32 indicating whether the SvIVX is a valid offset value for
-the SvPVX.  This hack is used internally to speed up removal of characters
-from the beginning of a SvPV.  When SvOOK is true, then the start of the
-allocated string buffer is really (SvPVX - SvIVX).
+Returns a U32 indicating whether the pointer to the string buffer is offset.
+This hack is used internally to speed up removal of characters from the
+beginning of a SvPV.  When SvOOK is true, then the start of the
+allocated string buffer is actually C<SvOOK_offset()> bytes before SvPVX.
+This offset used to be stored in SvIVX, but is now stored within the spare
+part of the buffer.
 
 =for apidoc Am|U32|SvROK|SV* sv
 Tests if the SV is an RV.
@@ -1240,8 +1242,9 @@ the scalar's value cannot change unless written to.
 		     if (SvLEN(sv)) {					\
 			 assert(!SvROK(sv));				\
 			 if(SvOOK(sv)) {				\
-			     SvPV_set(sv, SvPVX_mutable(sv)		\
-					  - sv_read_offset(sv));	\
+			     STRLEN zok; 				\
+			     SvOOK_offset(sv, zok);			\
+			     SvPV_set(sv, SvPVX_mutable(sv) - zok);	\
 			     SvFLAGS(sv) &= ~SVf_OOK;			\
 			 }						\
 			 Safefree(SvPVX(sv));				\
@@ -1929,6 +1932,60 @@ C<SvUTF8_on> on the new SV.  Implemented as a wrapper around C<newSVpvn_flags>.
 #define newSVpvn_utf8(s, len, u) newSVpvn_flags((s), (len), (u) ? SVf_UTF8 : 0)
 
 /*
+=for apidoc Am|void|SvOOK_offset|NN SV*sv|STRLEN len
+
+Reads into I<len> the offset from SvPVX back to the true start of the
+allocated buffer, which will be non-zero if C<sv_chop> has been used to
+efficiently remove characters from start of the buffer. Implemented as a
+macro, which takes the address of I<len>, which must be of type C<STRLEN>.
+Evaluates I<sv> more than once. Sets I<len> to 0 if C<SvOOK(sv)> is false.
+
+=cut
+*/
+
+#ifdef DEBUGGING
+/* Does the bot know something I don't?
+10:28 <@Nicholas> metabatman
+10:28 <+meta> Nicholas: crash
+*/
+#  define SvOOK_offset(sv, offset) STMT_START {				\
+	assert(sizeof(offset) == sizeof(STRLEN));			\
+	if (SvOOK(sv)) {						\
+	    const U8 *crash = (U8*)SvPVX_const(sv);			\
+	    offset = *--crash;						\
+ 	    if (!offset) {						\
+		crash -= sizeof(STRLEN);				\
+		Copy(crash, (U8 *)&offset, sizeof(STRLEN), U8);		\
+	    }								\
+	    {								\
+		/* Validate the preceding buffer's sentinels to		\
+		   verify that no-one is using it.  */			\
+		const U8 *const bonk = (U8 *) SvPVX_const(sv) - offset;	\
+		while (crash > bonk) {					\
+		    --crash;						\
+		    assert (*crash == (U8)PTR2UV(crash));		\
+		}							\
+	    }								\
+	} else {							\
+	    offset = 0;							\
+	}								\
+    } STMT_END
+#else
+    /* This is the same code, but avoids using any temporary variables:  */
+#  define SvOOK_offset(sv, offset) STMT_START {				\
+	assert(sizeof(offset) == sizeof(STRLEN));			\
+	if (SvOOK(sv)) {						\
+	    offset = ((U8*)SvPVX_const(sv))[-1];			\
+	    if (!offset) {						\
+		Copy(SvPVX_const(sv) - 1 - sizeof(STRLEN),		\
+		     (U8 *)&offset, sizeof(STRLEN), U8);		\
+	    }								\
+	} else {							\
+	    offset = 0;							\
+	}								\
+    } STMT_END
+#endif
+/*
  * Local variables:
  * c-indentation-style: bsd
  * c-basic-offset: 4