Extended Description

https://simd.godbolt.org/z/KrahKsdhe

clang -g0 -O3 -march=haswell

#include
#include

auto lz(unsigned x) {
int c = __builtin_clz( x );
return x ? c : sizeof(x) * CHAR_BIT;
}

auto tz(unsigned x) {
int c = __builtin_ctz( x );
return x ? c : sizeof(x) * CHAR_BIT;
}

lz:
lzcntl %edi, %eax
retq

tz:
tzcntl %edi, %eax
retq

auto lz_ubsan(unsigned x) {
int c = __builtin_clz( x ? x : ~0U );
return x ? c : sizeof(x) * CHAR_BIT;
}

auto tz_ubsan(unsigned x) {
int c = __builtin_ctz( x ? x : ~0U );
return x ? c : sizeof(x) * CHAR_BIT;
}

lz_ubsan:
xorl %eax, %eax
cmpl $1, %edi
sbbl %eax, %eax
orl %edi, %eax
lzcntl %eax, %ecx
testl %edi, %edi
movl $32, %eax
cmovnel %ecx, %eax
retq

tz_ubsan:
xorl %eax, %eax
cmpl $1, %edi
sbbl %eax, %eax
orl %edi, %eax
tzcntl %eax, %ecx
testl %edi, %edi
movl $32, %eax
cmovnel %ecx, %eax
retq

To prevent ubsan warnings, we mustn't call the ctz/clz builtins with a zero-value. So we typically insert an all-bits value to make it shutup

Unfortunately this conflicts with the post-combine that recognise cases where the zero-value input should be correctly handled (e.g tzcnt/lzcnt x86 instructions).

Funnily enough, gcc has better code from the ubsan-safe cases.....