For a reference: there are total of 1301 instances this optimization for a bootstrap build of llvm:

grep "Found one" bootstrap_buildlog |sort |uniq -c

1094 Found one path (len=1):
  176 Found one path (len=2):
   30 Found one path (len=3):
    1 Found one path (len=4):

integrated David's review comments to use an available api.

Fixed type cast warnings.

using __llvm_profile_get_size_for_buffer_internal directly to get the profile size.

integrated David's review comments

move inline attribute into InstrProfilingPort.h

using wrappers to always_inline function to avoid function call overhead.

My previous refers to using gcc as the compiler.

I checked the assembly in
clang_rt.profile-i386.dir/ and
clang_rt.profile-x86_64.dir/

This is the updated patch. It fixed some issues in previous patch. It also adds a new API for updating value-count in batch mode. Without this, PGO instrumented Clang incurs 20x slow down in building itself (with %9m prefix -- with %m is even greater slowdown ).

lgtm

Ping.

ping

Integrated David's review comments: (1) add vararg test (2) simplify function attributes.

ping.

Updated the patch to integrate comments from ctopper and dberlin.

Look good to me.

This is a good improvement.

Looks good to me.

Added more switch cases to make the lowering generate binary search compares.
Thanks to Hans's suggestion.

Add a test case that tests merging adjacent cases.

Here is the patch that Integrated with the review comments from Hans.

Integrated with Hans's review comments.

Hans, Thanks for the detailed reviews. Please check if the updated patch addresses your reviews.

Integrated the review comments and suggestion from Hans.

Updated the patch with review comments.

Thanks for the review. Here is the updated patch.

Integrated the review comments from vsk.

This patch is wrong. Please ignore.

LGTM

Thanks for the fix.

Looks good to me. Thanks for fixing.

Lgtm.

Looks good to me.

integrated David's review comments

This new patch integrated David's review suggestion to make this optimization a separated pass.
This pass will be called after inlining to avoid the potential interference to in-lining decisions. It will use both the metadata annotation for memop size profiling and the enclosing BB's edge count to find the actual count.

Integrated David's suggestion to make the optimization a stand along pass (a function pass).
It currently resides in IndirectCallPromotion.cpp. I will have a follow-up patch to change IndirectCallPromotion.cpp to a more appropriate name.

randomize the order the VP profile entry in proftext, as suggested by David.

Fix the testcase.

Integrate David's comments. Using a separated option to control the number of annotation for memop_size.