https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128739
Bug ID: 128739
Summary: [17 Regression] vectorizer: peeled early-break loop
drops final vector iteration's contribution to
reductions
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: ted.d.rodgers at gmail dot com
Target Milestone: ---
Created attachment 66172
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=66172&action=edit
longer test case
Real-world impact: miscompiles read_variable_length in liblz4 (inlined into
LZ4_decompress_generic), silently truncating decompressed output
Test case:
Fails with: gcc -O2 -mavx2, gcc -O2 -msse4.1 (returns 17347 instead of
25507 — exactly one vector of 255s missing).
Works with: baseline -O2/-O3 (loop not vectorized),
-fno-tree-loop-vectorize, or gcc-16
static unsigned char buf[256];
__attribute__ ((noinline)) static unsigned
rd (const unsigned char **pp, const unsigned char *lim)
{
const unsigned char *p = *pp;
unsigned int s = *p++;
unsigned int len = s;
while (s == 255)
{
if (p >= lim)
return 0;
s = *p++;
len += s;
}
*pp = p;
return len;
}
int
main (void)
{
for (int i = 0; i < 100; i++)
buf[31 + i] = 255;
buf[131] = 7;
const unsigned char *p = buf + 31;
if (rd (&p, buf + 132) != 255 * 100 + 7)
__builtin_abort();
}
Bisection:
r17-5740-g1b7a552c8dbb71 — gcc/tree-vect-loop-manip.cc, around line 1469
(vect_set_loop_condition, the LOOP_VINFO_EARLY_BREAKS_VECT_PEELED path that
creates latch_exit_edge).
A longer standalone test case reduced from lz4 is attached (prints len, showing
the 255×32-short fingerprint)