Optimization regression after llvm 17

[alex-s168]

All clang versions on godbolt after 17.0.1 fail to optimize this lookup table into simple shift instructions

typedef struct {
#define X(a) int x##a : 1;
    X(0) X(1) X(2) X(3)
    X(4) X(5) X(6) X(7)
    X(8) X(9) X(10) X(11)
#undef X
} bits;

int get(bits b, int idx) {
    switch(idx){
#define X(a) case a: return b.x##a; break;
    X(0) X(1) X(2) X(3)
    X(4) X(5) X(6) X(7)
    X(8) X(9) X(10) X(11)
#undef X
    default: __builtin_unreachable();
    }
}

with clang 17.0.1 -target x86_64 -O3:

get:
        mov     cl, 15
        sub     cl, sil
        shl     edi, cl
        movsx   eax, di
        sar     eax, 15
        ret

but later clang versions don't do this anymore, and instead emit a jump table, with a shift in each case

See: https://godbolt.org/z/c79q7EY4P
(I opened the opt pipeline viewer and noticed that the output of InstCombinerPass is different.
I don't have much knowledge about the internals of llvm tho.

[nikic]

The problem is that InstCombine transforms one of the branches, so that the common part can no longer be sunk.