As an aside, some of this is coming from the only use of canWrite() in the tree.

cloned to rdar://problem/9082880

For reference, some background for r114187 is here, though the proposed solution is not what was ultimately implemented:
http://lists.cs.uiuc.edu/pipermail/cfe-dev/2010-September/011002.html

Bob, can someone take a look at this?

I pointed Eli at this earlier but we weren't sure what to do about it. I think the next step is to talk with Argyrios to better understand the issues. We'll do that this week.

Would it be ok if we managed to reduce the number of system calls to the absolute minimum necessary to

1. create the temporary file
2. rename the temporary over the output file

?

One option would be to put the responsibility for doing the temporary file plus rename in the client which needs this. For example, in place of having the client doing the equivalent of "clang ... -o out.o", the client could wrap this in some simple logic which does the equivalent of "clang ... -o out.o.blahblah && mv out.o.blahblah out.o". That way, the client would be in control of its own safety, and all other clients would be spared the overhead.

How about temp file + rename only when writing PCH files ?

After changes to trunk, for writing PCH:
10 119 open:entry clang test.pch-0fa1da06
12 121 close:entry clang test.pch-0fa1da06
12 365 rename:entry clang test.pch-0fa1da06 test.pch

for other output files:
4 119 open:entry clang test.o
4 121 close:entry clang test.o

llvm r136310 +
clang r136315.