strbuf_setlen: don't write to strbuf_slopbuf

strbuf_setlen(., 0) writes '\0' to sb.buf[0], where buf is either
allocated and unique to sb, or the global slopbuf. The slopbuf is meant
to provide a guarantee that buf is not NULL and that a freshly
initialized buffer contains the empty string, but it is not supposed to
be written to. That strbuf_setlen writes to slopbuf has at least two
implications:

First, it's wrong in principle. Second, it might be hiding misuses which
are just waiting to wreak havoc. Third, ThreadSanitizer detects a race
when multiple threads write to slopbuf at roughly the same time, thus
potentially making any more critical races harder to spot.

Avoid writing to strbuf_slopbuf in strbuf_setlen. Let's instead assert
on the first byte of slopbuf being '\0', since it helps ensure the
promised invariant of buf[len] == '\0'. (We know that "len" was already
0, or someone has messed with "alloc". If someone has fiddled with the
fields that much beyond the correct interface, they're on their own.)

This is a function which is used in many places, possibly also in hot
code paths. There are two branches in strbuf_setlen already, and we are
adding a third and possibly a fourth (in the assert). In hot code paths,
we hopefully reuse the buffer in order to avoid continous reallocations.
Thus, after a start-up phase, we should always take the same path,
which might help branch prediction, and we would never make the assert.
If a hot code path continuously reallocates, we probably have bigger
performance problems than this new safety-check.

Simple measurements do not contradict this reasoning. 100000000 times
resetting a buffer and adding the empty string takes 5.29/5.26 seconds
with/without this patch (best of three). Releasing at every iteration
yields 18.01/17.87. Adding a 30-character string instead of the empty
string yields 5.61/5.58 and 17.28/17.28(!).

This patch causes the git binary emitted by gcc 5.4.0 -O2 on my machine
to grow from 11389848 bytes to 11497184 bytes, an increase of 0.9%.

I also tried to piggy-back on the fact that we already check alloc,
which should already tell us whether we are using the slopbuf:

        if (sb->alloc) {
                if (len > sb->alloc - 1)
                        die("BUG: strbuf_setlen() beyond buffer");
                sb->buf[len] = '\0';
        } else {
                if (len)
                        die("BUG: strbuf_setlen() beyond buffer");
                assert(!strbuf_slopbuf[0]);
        }
        sb->len = len;

That didn't seem to be much slower (5.38, 18.02, 5.70, 17.32 seconds),
but it does introduce some minor code duplication. The resulting git
binary was 11510528 bytes large (another 0.1% increase).

Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

1 files changed, 4 insertions, 1 deletions

diff --git a/strbuf.h b/strbuf.h
index 2075384e0b..1a77fe146a 100644
--- a/strbuf.h
+++ b/strbuf.h
@@ -147,7 +147,10 @@ static inline void strbuf_setlen(struct strbuf *sb, size_t len)
 	if (len > (sb->alloc ? sb->alloc - 1 : 0))
 		die("BUG: strbuf_setlen() beyond buffer");
 	sb->len = len;
-	sb->buf[len] = '\0';
+	if (sb->buf != strbuf_slopbuf)
+		sb->buf[len] = '\0';
+	else
+		assert(!strbuf_slopbuf[0]);
 }
 
 /**