Correct lzcnt results

[AlexGuteniev]

Fixes #1103

[StephanTLavavej]

Looks great, just a couple of questions about naming. If you or @fsb4000 can confirm that the tests for countl_zero fail on the affected processors before your change, and pass after, I think that should be sufficient for validation. (Note that you can compile a test program on one machine and just copy it to another if you statically link.)

[fsb4000]

If you or @fsb4000 can confirm that the tests for countl_zero fail on the affected processors before your change, and pass after, I think that should be sufficient for validation.

On my CPU(AMD FX8300) both versions(current master and the pull request) print same(correct) result:

[AlexGuteniev]

I confirm the fix on pre AMB-LZCNT-POPCNT machine

[StephanTLavavej]

Thanks @fsb4000 and @AlexGuteniev! I think this all makes sense - the FX-8300 worked because it physically supported Advanced Bit Manipulation, so the previous code said "oh, this doesn't support AVX2, we need to special-case 0" but otherwise used lzcnt which works. The test previously failed on Alex's even older processor which doesn't physically support lzcnt.

[barcharcraz]

it sucks that we can't test this by modifying __isa_available before running the functions.

[barcharcraz]

approved with a testing suggestion, I don't feel that suggestion should block merging this though.

[StephanTLavavej]

Thanks @statementreply for noticing this bug, @AlexGuteniev for fixing it, and @fsb4000 for helping to validate the fix! 😺 😸 🐈