Duplicated branch on equal instructions when used in conjunction with count-trailing-ones

[Validark]

On targets where there is no direct count-trailing-zeros instruction, some insert a branch on 0 before running the code that calculates the trailing-zeros count. In my code, I want to commandeer this branch for added efficiency, however, LLVM gets confused when there is a bitwise negation present.

Here is some simplified Zig code which reproduces the problem (godbolt link here).

export fn foo(bitstring: u64, sum: u64) u64 {
    var acc: u64 = sum;
    const inverted_bitstring = ~bitstring;

    // Optimization: If @ctz implicitly inserts `if (inverted_bitstring == 0)`, make it more efficient.
    if (inverted_bitstring == 0) {
        acc += 64;
        return acc;
    }

    const str_len: u8 = @ctz(inverted_bitstring);
    acc += str_len;
    return acc;
}

On x86_64 (default target) we get:

foo:
        cmp     rdi, -1
        je      .LBB0_1
        xor     rdi, -1
        je      .LBB0_3
        rep       bsf rax, rdi
        add     rax, rsi
        ret
.LBB0_1:
        add     rsi, 64
        mov     rax, rsi
        ret
.LBB0_3:
        mov     eax, 64
        add     rax, rsi
        ret

On riscv64 (default target) we get the same branch-on-equal instruction twice.

foo:
        li      a2, -1
        beq     a0, a2, .LBB0_3
        beq     a0, a2, .LBB0_4
        ...
        ...
.LBB0_3:
        addi    a0, a1, 64
        ret
.LBB0_4:
        li      a0, 64
        add     a0, a0, a1
        ret

Also does the same thing on s390x, sparc64, and ve (default targets). If you would like to see some less-simplified code where I actually tried to use this idea, here is a godbolt link to a less-trimmed version of my actual code (go to line 461 in the source code).

Thank you to all LLVM contributors!

‒ Validark

[Validark]

Sounds like this is a problem with the is_zero_poison flag:

https://github.com/ziglang/zig/issues/17722#issuecomment-1780916722

Is this considered a bug in LLVM?

[nikic]

How do you dump LLVM IR with zig?