Missed optimization of loopDeletion probably because of a missing loopUnswitch

[ZY546]

https://godbolt.org/z/TW9b71zPM

int b, c;
int n;
void test() {
  for (int i = 0; i < 1000; i += b) {
    if(c < 0){
        c = n;
    }
  }
}

In the test code, the value of c only gets a chance to change when it first enters the loop, so from there we can make the following optimizations:

If initially c<0 holds, then the loop is not executed.

Otherwise, c is a loop invariant for all subsequent iterations of the loop, which can be optimized using loopUnswitch.

Furthermore, if we optimize via loopUnswitch, we see that the entire loop is unnecessary.

Finally, we get the expected code:

test():                               # @test()
        cmp     dword ptr [rip + c], 0
        js      .LBB0_1
        ret
.LBB0_1:                                # %if.then
        mov     eax, dword ptr [rip + n]
        mov     dword ptr [rip + c], eax
        ret

But Clang:

test():                               # @test()
        mov     eax, dword ptr [rip + b]
        mov     esi, dword ptr [rip + c]
        xor     ecx, ecx
        mov     edx, dword ptr [rip + n]
        jmp     .LBB0_1
.LBB0_3:                                # %for.inc
        add     ecx, eax
        cmp     ecx, 1000
        jge     .LBB0_4
.LBB0_1:                                # %for.body
        test    esi, esi
        jns     .LBB0_3
        mov     dword ptr [rip + c], edx
        mov     esi, edx
        jmp     .LBB0_3
.LBB0_4:                                # %for.cond.cleanup
        ret

[XChy]

It's an unremovable infinite loop if b = 0. I'm interested in some features of its phi version. I've opened an issue about it: #70767