Issue 244520
Summary [SelectionDAG] Wrong code at every -O level: i65 `mul` + `add` stored with sext gets bit 64 set (`_BitInt(65)`)
Labels new issue
Assignees
Reporter haoxintu
    ## Description

```c
_BitInt(65) c;

__attribute__((noinline)) void
d (long e, long f)
{
  c = (_BitInt(65)) e * 11400714819323198485wb + ((_BitInt(65)) f + (f < -1)) * 30;
}

int main (void)
{
  d (1, 0);
  /* 1 * 11400714819323198485 + (0 + 0) * 30 fits in signed _BitInt(65) (< 2^64): it must stay positive. */
 if (c != 11400714819323198485wb)
    __builtin_abort ();
  return 0;
}
```

```
$ clang -O0 repro.c && ./a.out      # Aborted (also -O1, -O2, -O3, -Os)
$ gcc -O2 repro.c && ./a.out        # ok
```

The low 64 bits are right (`0x9e3779b97f4a7c15`), but bit 64, the sign bit of `_BitInt(65)`, is set. So `c` becomes negative.
`-fsanitize=address,undefined` is clean.

The bug is in the back end. `llc -O0` and `llc -O2` on this IR store bit 64 = 1 for `d(1, 0)`:

```llvm
define void @d(i64 %e, i64 %f) {
  %conv = sext i64 %e to i65
  %mul = mul i65 %conv, 11400714819323198485
  %a = sext i64 %f to i65
  %b = icmp slt i64 %f, -1
  %z = zext i1 %b to i65
  %c2 = add i65 %a, %z
  %m3 = mul i65 %c2, 30
  %a4 = add i65 %mul, %m3
  %s = sext i65 %a4 to i128
  store i128 %s, ptr @c, align 8
  ret void
}
```

The `(sext f + zext (f < -1)) * 30` term is 0 here. Removing the icmp/zext, or the `* 30`, makes the result correct.

Reproduced in Godbolt (x86-64 clang assertions trunk, -O2): https://godbolt.org/z/h1hjd9xMP


Thanks.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to