Test reflecting the LLVM 11 behavior for transforming reads of the form f[k], with k symbolic and f a 4-element vector into something along the lines:

 if k == 0 => f[0]
 elif k == 1 => f[1]
 elif k == 2 => f[2]
 elif k == 3 => f[3]
 else ==> undef

1 files changed, 46 insertions, 0 deletions

diff --git a/test/VectorInstructions/oob-read-llvm-geq11.c b/test/VectorInstructions/oob-read-llvm-geq11.c
new file mode 100644
index 00000000..309e05b1
--- /dev/null
+++ b/test/VectorInstructions/oob-read-llvm-geq11.c
@@ -0,0 +1,46 @@
+// REQUIRES: geq-llvm-11.0
+// XFAIL: geq-llvm-11.0
+
+/* The scalarizer pass in LLVM 11 was changed to generate, for a
+ read f[k], with k symbolic and f a 4-element vector:
+ if k == 0 => f[0]
+ elif k == 1 => f[1]
+ elif k == 2 => f[2]
+ elif k == 3 => f[3]
+ else ==> undef
+
+ Therefore, even though an OOB access might exist at the source code
+ level, no such OOB accesses exist anymore at the LLVM IR level.
+
+ And since undef is currently treated in KLEE as 0, an overflowing
+ access is always translated as f[0], which may lead to future
+ problems being missed.
+
+ This test is marked as XFAIL as a reminder that we need to fix this
+ behaviour, most likely by having undef return a new symbolic variable.
+*/
+
+// RUN: %clang %s -emit-llvm %O0opt -g -c -o %t1.bc
+// RUN: rm -rf %t.klee-out
+// NOTE: Have to pass `--optimize=false` to avoid vector operations being
+// constant folded away.
+// RUN: %klee --output-dir=%t.klee-out --optimize=false --exit-on-error %t1.bc 2>%t.log
+// RUN: FileCheck -input-file=%t.stderr.log %s
+
+#include "klee/klee.h"
+
+#include <assert.h>
+#include <stdint.h>
+#include <stdio.h>
+
+typedef uint32_t v4ui __attribute__((vector_size(16)));
+int main() {
+  v4ui f = {1, 2, 3, 4};
+  int k = klee_range(4, 10, "k");
+
+  uint32_t v = f[k]; // Symbolic out-of-bounds read
+  v = f[v];          // This should trigger an error, but currently this returns f[0] = 1
+  assert(v != 1);
+
+  return 0;
+}