I was investigating a perf regression with the passthrough command buffer. I ended up coding a solution implemented in http://crrev.com/c/1247941 . During implementation of my solution I found:

a) calling base::EraseIf was slower than calling std::erase/std::remove_if.
b) forcing ALWAYS_INLINE on base::EraseIf seemed to improve things.
c) writing a custom iterating deleter over a std::vector was faster than inlined base::EraseIf or the inlined std::erase/std::remove_if. (~5% improved score)

It does seem like Clang can be reluctant to inline functions at times. Using ALWAYS_INLINE does always work. Otherwise it may decide not to inline functions even if marked inline.

Also there could be potentially some slowness in the STL functions for iterate-based erase. dcheng assigning to you for triage. I can provide more info on my experiments if needed. Not sure what components to use for this issue.