Issue 244615
Summary Wrong code at -Os: `cttz(x, true)` lowered to BSF, then MOVD copies the undefined upper bits of its destination into a constant-0 vector lane
Labels new issue
Assignees
Reporter haoxintu
    ## Description

```c
typedef short v2s __attribute__((ext_vector_type(2)));
typedef unsigned short v2u __attribute__((ext_vector_type(2)));

__attribute__((noinline)) void
k (const short *pa, const short *pb, short *po, unsigned junk)   /* junk arrives in %ecx */
{
  (void) junk;
  v2s a = { pa[0], pa[1] }, b = { pb[0], pb[1] };
  v2s r = (v2s) __builtin_elementwise_ctzg ((v2u) __builtin_shufflevector (a, ((v2s) {32767, 32767}), 1, 3),
 (v2u) __builtin_shufflevector (a, b, 1, 2));
 po[0] = r[0];
  po[1] = r[1];
}

int main (void)
{
  /* a[1] == 0: lane 0 takes the fallback a[1] = 0, lane 1 is ctz (32767) = 0 */
  short a[2] = { 0, 0 }, b[2] = { 5, 0 }, o[2];
  k (a, b, o, 0x12340000u);
  if (o[0] != 0 || o[1] != 0)
    __builtin_abort ();
  return 0;
}
```

```
$ clang -Os repro.c && ./a.out      # Aborted; -O0/-O1/-O2/-O3 are fine
```

The IR is fine: lane 1 is the constant 0, and the `cttz(x, true)` result (poison for x == 0) only feeds lane 0, which the select replaces by the fallback when x == 0. The -Os code is:

```
	bsfl	%eax, %ecx          # x == 0: %ecx keeps its old value (here the argument junk = 0x12340000)
	movd	%ecx, %xmm1         # all 32 bits -> lanes 0 and 1
	...
	movd	%xmm2, (%rdx)
```

Building the `<2 x i16>` from the 32-bit BSF result assumes the upper 16 bits are zero, because `cttz` of an i16 is at most 16. That holds only for a nonzero input. For x == 0, lane 1 receives the upper half of `%ecx`: `o[1]` is 4660 = 0x1234.

`llc -O2` (trunk and LLVM 21) on this IR shows the same thing. The driver passes 0x12340000 in `%ecx` and prints `0 4660`:

```llvm
define void @k(ptr %pa, ptr %pb, ptr %po, i32 %junk) optsize {
  %p1 = getelementptr inbounds i8, ptr %pa, i64 2
  %x = load i16, ptr %p1, align 2
  %y = load i16, ptr %pb, align 2
  %c = call i16 @llvm.cttz.i16(i16 %x, i1 true)
  %v = insertelement <2 x i16> <i16 poison, i16 0>, i16 %c, i64 0
  %t = insertelement <2 x i16> <i16 poison, i16 32767>, i16 %x, i64 0
  %iszero = icmp eq <2 x i16> %t, zeroinitializer
 %f0 = insertelement <2 x i16> poison, i16 %x, i64 0
  %f = insertelement <2 x i16> %f0, i16 %y, i64 1
  %r = select <2 x i1> %iszero, <2 x i16> %f, <2 x i16> %v
  store <2 x i16> %r, ptr %po, align 2
  ret void
}
```

Without `optsize` the same IR is compiled correctly.

Reproduced in Godbolt (x86-64 clang trunk, -Os and -O1): https://godbolt.org/z/e8Efq8M3e

## Checked Versions

- Wrong at -Os with clang trunk 2012344e5da8 (2026-11-19). -O0/-O1/-O2/-O3 are fine.
- llc from LLVM 21.1.8 miscompiles the IR the same way, so this is not a recent regression.
- Sanitizers (ASan, UBSan) are clean.

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

Reply via email to