[llvm-bugs] [Bug 47631] New: [PGO] Consider enabling pre-cleanup passes before PGO instrumentation at -Os

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=47631

            Bug ID: 47631
           Summary: [PGO] Consider enabling pre-cleanup passes before PGO
                    instrumentation at -Os
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Interprocedural Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: rnk at google.com
                CC: davidxl at google.com, llvm-bugs at lists.llvm.org,
                    xur at google.com

Rong Xu added a series of cleanup passes that run before PGO in this 2016
change: 
https://reviews.llvm.org/D21405

Chromium's build uses a mix of -O2 and -Os for different targets. We discovered
that with PGO, without the initial cleanup inliner pass, PGO-instrumented
object files are much, much larger than they should be. Switching from -Os to
-O2 everywhere reduced total object file inputs to link chrome.dll from ~20GB
to ~9GB, and greatly reduced linker memory usage.

If the inliner is not run before PGO instrumentation, every linkonce_odr
function ends up being retained. Consider this example:

$ cat t.cpp 
void foo();
void bar();
inline void inlineMe(int x) {
  if (x > 0)
    foo();
  else
    bar();
}
int getVal();
int main() { inlineMe(getVal()); }

Normally, -Os and -O2 will inline away inlineMe, and then delete it.

However, if we do PGO instrumentation before the inliner can delete inlineMe,
the __profd_ data ends up recording the function address of all linkonce_odr
functions. This command line shows that it is retained with Os but not O2:

$ clang -S -emit-llvm -O2 t.cpp  -o - -fprofile-generate |& grep
define.*inlineM
# nothing
$ clang -S -emit-llvm -Os t.cpp  -o - -fprofile-generate |& grep
define.*inlineM
define linkonce_odr dso_local void @_Z8inlineMei(i32 %x) local_unnamed_addr #1
comdat {

The retention of all linkonce_odr functions by PGO instrumentation is
controlled here:
https://github.com/llvm/llvm-project/blob/master/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp#L830
David Li added it in this change:
http://github.com/llvm/llvm-project/commit/7008ce3f9891

Almost every inline function in C++ is linkonce_odr, so this code ends up
retaining practically everything. This results in a lot of extra duplicate
sections and symbols that wouldn't normally make it to the final link in an
optimized build. In our case, the linker would run out of memory:
https://crbug.com/1058040

We fixed the problem by switching to O2, but this behavior at Os seems
undesirable, so I am reporting it.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20200923/41a7cbe9/attachment.html>