nlewycky added inline comments.
================
Comment at: lib/Transforms/Instrumentation/GCOVProfiling.cpp:447
@@ -432,3 +446,3 @@
return Filename.str();
}
}
----------------
compnerd wrote:
> It really feels like these two cases can be collapsed.
I don't see a great way to do that, so I picked the one which minimizes the
cost of the common case. Having an !llvm.gcov node but no CU for the specific
.o is uncommon: to get here you're building a program with some .bc's built
with a frontend that emits an !llvm.gcov and some .bc's without, and then you
llvm-linked them and are running the insert-gcov-profiling pass afterwards.
Also, I assume that copying SmallString<128>'s around is slow, while looking up
Value*'s is faster.
If you have a specific idea, I'm happy to try it. I'm not thrilled with the way
it's written now.
================
Comment at: test/Transforms/GCOVProfiling/three-element-mdnode.ll:27
@@ +26,2 @@
+
+!10 = !{i32 1, !"Debug Info Version", i32 3}
----------------
The way this test was written, !10 would be parsed before !9. Swizzled it
around to make !9 the !llvm.module.flags and !10 the !llvm.gcov node.
https://reviews.llvm.org/D23734
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
