setjmp/longjmp control transfer is missliced

[Pipi9221]

[llvm-slicer] setjmp/longjmp control transfer is missliced

Hi, thanks for maintaining dg and llvm-slicer.
We may have found a slicing correctness issue and wanted to report it with a
minimal reproducer below.

Summary

For a setjmp/longjmp program, slicing with -c __assert_fail (or line
criterion 17:x) changes semantics: original/unsliced passes, but sliced
fails assertion.

The slice keeps _setjmp and assertion checks, but drops the longjmp
control-transfer path (jumpit()), which makes the remaining path assign
x = 2 and fail assert(x == 1).

Environment

• dg commit: dc3615d03fedfce62f37604b45d22e31c9d7e557
• slicer: llvm-slicer (built from this dg checkout)
• clang: Ubuntu clang version 14.0.0-1ubuntu1.1
• lli: Ubuntu LLVM version 14.0.0
• OS: Linux ... WSL2 x86_64 (kernel 5.15.167.4-microsoft-standard-WSL2)

Minimal Reproducer

#include <assert.h>
#include <setjmp.h>

static jmp_buf jb;

static void jumpit(void) {
  longjmp(jb, 1);
}

int main(void) {
  volatile int x = 0;
  if (setjmp(jb) == 0) {
    x = 1;
    jumpit();
    x = 2;
  }
  assert(x == 1);
  return 0;
}

Reproduction Steps

cat > /tmp/setjmp_longjmp_misslice.c <<'EOF'
#include <assert.h>
#include <setjmp.h>
static jmp_buf jb;
static void jumpit(void) { longjmp(jb, 1); }
int main(void) {
  volatile int x = 0;
  if (setjmp(jb) == 0) {
    x = 1;
    jumpit();
    x = 2;
  }
  assert(x == 1);
  return 0;
}
EOF

clang-14 -O0 -g -c -emit-llvm /tmp/setjmp_longjmp_misslice.c -o /tmp/setjmp_longjmp_misslice.bc
lli-14 /tmp/setjmp_longjmp_misslice.bc

llvm-slicer -c __assert_fail /tmp/setjmp_longjmp_misslice.bc -o /tmp/setjmp_longjmp_misslice.assert.sliced.bc
lli-14 /tmp/setjmp_longjmp_misslice.assert.sliced.bc

llvm-slicer -c 17:x /tmp/setjmp_longjmp_misslice.bc -o /tmp/setjmp_longjmp_misslice.line.sliced.bc
lli-14 /tmp/setjmp_longjmp_misslice.line.sliced.bc

Actual Result

• Unsliced run passes.
• Sliced runs fail:

Assertion `x == 1' failed.

Expected Result

Slicing should preserve semantics for the selected criterion.
The sliced program should not change setjmp/longjmp behavior and should not
turn a passing program into an assertion failure.

IR Evidence

Original IR (relevant):

store volatile i32 1, i32* %x
call void @jumpit()
store volatile i32 2, i32* %x
...
%v = load volatile i32, i32* %x
%ok = icmp eq i32 %v, 1

Sliced IR (relevant):

; call to jumpit removed
store volatile i32 2, i32* %x
...
%v = load volatile i32, i32* %x
%ok = icmp eq i32 %v, 1

This is a clear semantic mismatch introduced by slicing.

Additional Note

A repository grep for setjmp/longjmp handling (titles/bodies in local issue
snapshot and source-level modeling) shows no direct dedicated modeling path for
this control transfer, which may be related.