[AArch64] wrong code at -O1: sign extension of a shifted 8-bit value is selected as sbfiz #241864

Live Reproducer: https://godbolt.org/z/KTGEEfbr8

Compiling the following C code for aarch64 with clang -O1 and running it prints the wrong result.

#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:

clang --target=aarch64-linux-gnu -static -O1 repro.c && qemu-aarch64 ./a.out

Output (wrong):

0

Expected (what -O0 prints, and what x86-64 prints):

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:

        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:

clang version 24.0.0git (https://github.com/llvm/llvm-project f5744e07d2dba7367ec68426fd6b6b07e2916b39)

This reproduces with current LLVM trunk.