This patch implements the infrastructure for PGO late (i.e. IR-level) instrumentation. (Refer to RFC: PGO Late instrumentation for LLVM http://lists.llvm.org/pipermail/llvm-dev/2015-August/089058.html)
The main part of code is in a newly added file: lib/Transforms/Instrumentation/PGOLateInstr.cpp
This file implements a module pass PGOLateInstrumeatnion. It applies the instrumentation to each function by class PGOLateInstrumentationFunc. For each function, perform the following steps:
(1) Collect all the CFG edges. Assign an estimated weight to each edge. Critical edges and back-edges are assigned to high value of weights. One fake node and a few fake edges (from the fake node to the entry node, and from the exit nodes to the fake node) are also added to the worklist.
(2) Construct the MST. The edges with the higher weight will be put to MST first, unless it forms a cycle.
(3) Traverse the CFG and compute the CFG hash.
The above three steps are the same for profile-generate and profile-use compilation.
In the next step, for profile-generation compilation:
(4-gen) Instrument all the edges that not in the MST. If this is a critical edge, split the edge first. The actual instrumentation is to generate Intrinsic::instrprof_increment() in the instrumented BB. This intrinsic will be lowed by pass createInstrProfilingPass().
For profile-use compilation,
(4-use) Read in the counters and the CFG hash from the profile file.
(5-use) If there is no error, populate the counters to all the edges in reverse topological order of the MST.
(6-use) Once having all the edge counts, set the branch weights metadata for the IR having multiple branches. Also apply the cold/hot function attributes based on function level counts.
This pass is added to PassManagerBuilder when populate module passes. see lib/Transforms/IPO/PassManagerBuilder.cpp.
Nit: it is a little strange to change some of the formatting.