Following r263086, we are replacing this by a runtime check.
More cleanup will follow on the IRBuilder itself, but I submitted
this patch separately as SROA has a fancy "prefixInserter" class
that needs extra-love.
Details
Details
- Reviewers
chandlerc - Commits
- rG1e9c9251824e: Do not specialize IRBuilder to strip names in SROA
Diff Detail
Diff Detail
Event Timeline
Comment Actions
FYI i've been reviewing these and I'll LGTM them as soon as we're clearly on the same page with Philip in D17946's thread.
Comment Actions
LGTM patch is fine, but also see comment below.
lib/Transforms/Scalar/SROA.cpp | ||
---|---|---|
93–104 | When you measured, did you measure getting rid of this complexity too? If it doesn't slow SROA down when the context has the flag set, I would be all in favor of just completely removing the NDEBUG hacking here. My expectation is that the overhead of this while not zero (it uses an actual std::string) is *very* close to zero. That said, if you'd rather submit as-is, I can take care of cleaning this up. |
When you measured, did you measure getting rid of this complexity too?
If it doesn't slow SROA down when the context has the flag set, I would be all in favor of just completely removing the NDEBUG hacking here. My expectation is that the overhead of this while not zero (it uses an actual std::string) is *very* close to zero.
That said, if you'd rather submit as-is, I can take care of cleaning this up.