On Sat, Sep 19, 2020 at 1:46 PM Mark Dilger <mark.dil...@enterprisedb.com> wrote:
> 0002 and 0003 look good to me. I like the way you cleaned up a bit with the > unicode_norm_props struct, which makes the code a bit more tidy, and on my > compiler under -O2 it does not generate any extra runtime dereferences, as > the compiler can see through the struct just fine. My only concern would be > if some other compilers might not see through the struct, resulting in a > runtime performance cost? I wouldn't even ask, except that qc_hash_lookup is > called in a fairly tight loop. (I assume you mean unicode_norm_info) Yeah, that usage was copied from the keyword list code. I believe it was done for the convenience of the callers. That is worth something, and so is consistency. That said, I'd be curious if there is a measurable impact for some platforms. > In your commit message for 0001 you mentioned testing on a multiplicity of > compilers. Did you do that for 0002 and 0003 as well? For that, I was simply using godbolt.org to test compiling the multiplications down to shift-and-adds. Very widespread, I only remember MSVC as not doing it. I'm not sure a few extra cycles would have been noticeable here, but it can't hurt to have that guarantee. -- John Naylor https://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
