| Issue |
241864
|
| Summary |
[AArch64] wrong code at -O1: sign extension of a shifted 8-bit value is selected as `sbfiz`
|
| Labels |
generated by fuzzer
|
| Assignees |
|
| Reporter |
hazohelet
|
Live Reproducer: https://godbolt.org/z/KTGEEfbr8
Compiling the following C code for `aarch64` with `clang -O1` and running it prints the wrong result.
```c
#include <stdio.h>
static volatile int X = -1;
unsigned _BitInt(65) m;
__int128 sign(signed char a) { return a < 0 ? -1 : 0; }
int main(void) {
signed char x = X;
m = sign(x);
unsigned long c = x;
for (int i = 0; i < 8; i++)
c *= m;
printf("%lu\n", c);
}
```
clang command:
```sh
clang --target=aarch64-linux-gnu -static -O1 repro.c && qemu-aarch64 ./a.out
```
Output (wrong):
```text
0
```
Expected (what -O0 prints, and what x86-64 prints):
```text
18446744073709551615
```
`x` is -1, so `sign(x)` is -1 and `m` is all ones. The low word of `m` is computed as `(long)(signed char)X << 33` instead of the sign of `(signed char)X`:
```asm
ldr w9, [x8, :lo12:X]
sxtb w10, w9
mov w8, w9
sxtb x1, w9
sbfiz x8, x8, #33, #8
...
stp x8, x9, [x10] // m
```
Notes:
* Correct at -O0; -O1, -O2, -O3, -Os and -Oz are wrong
clang version:
```text
clang version 24.0.0git (https://github.com/llvm/llvm-project f5744e07d2dba7367ec68426fd6b6b07e2916b39)
```
This reproduces with current LLVM trunk.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs