[X86] Promote i32 CTTZ (BSF) instructions and remove speculation branch

[RKSimon]

Extends https://reviews.llvm.org/D132520 which did the same for i8/i16 cases.

Related to Issue #57810 - on some x64 targets without BMI, we'd be better off performing CTTZ (with zero handling) as a BTS + (REP) BSFq. Atom/Silvermont in particular would benefit from avoiding a i64 immediate, but anything with a fastish BTS as well.

auto ctz32(uint32_t x) {
    return x ? __builtin_ctz(x) : 32;
}
auto ctz32_64(uint64_t x) {
    return __builtin_ctzll(x | (1ULL << 32));
}

Current Codegen:

ctz32(unsigned int):                              # @ctz32(unsigned int)
        testl   %edi, %edi
        je      .LBB0_2
        rep bsfl        %edi, %eax
        retq
.LBB0_2:                                # %cond.end
        movl    $32, %eax
        retq
ctz32_64(unsigned long):                           # @ctz32_64(unsigned long)
        movabsq $4294967296, %rax               # imm = 0x100000000
        orq     %rdi, %rax
        rep bsfq        %rax, %rax
        retq

Could be:

ctz32_64(unsigned long):                           # @ctz32_64(unsigned long)
        btsq    $32, %rdi
        rep bsfq    %rdi, %rax
        retq

[llvmbot]

@llvm/issue-subscribers-backend-x86