For many cases, this threshold bump may not increase inlining if the old threshold was already aggressively inlining everything. In such cases, no compile time impact will be observed. For your motivating testcase that needs this threshold bump, did you measure how this impacted the compile time?

Did you study the compile-time impact of bumping the inlining threshold? While I would expect perf benefits in some cases from such a change, in my experience this threshold increase can also cause significant compile-time overhead in some cases.

Was the option nvptx-emit-line-numbers working? I see it removed.

Note that PTX allows .file even before .text. The spec is accurate here:

What do you mean by this: