2003-06-25-Reoptimizer1.txt [plain text]
Wed Jun 25 15:13:51 CDT 2003
We use opt to do Bytecode-to-bytecode instrumentation. Look at
back-edges and insert llvm_first_trigger() function call which takes
no arguments and no return value. This instrumentation is designed to
be easy to remove, for instance by writing a NOP over the function
Keep count of every call to llvm_first_trigger(), and maintain
counters in a map indexed by return address. If the trigger count
exceeds a threshold, we identify a hot loop and perform second-level
instrumentation on the hot loop region (the instructions between the
target of the back-edge and the branch that causes the back-edge). We
do not move code across basic-block boundaries.
We remove the first-level instrumentation by overwriting the CALL to
llvm_first_trigger() with a NOP.
The reoptimizer maintains a map between machine-code basic blocks and
LLVM BasicBlock*s. We only keep track of paths that start at the
first machine-code basic block of the hot loop region.
How do we keep track of which edges to instrument, and which edges are
exits from the hot region? 3 step process.
1) Do a DFS from the first machine-code basic block of the hot loop
region and mark reachable edges.
2) Do a DFS from the last machine-code basic block of the hot loop
region IGNORING back edges, and mark the edges which are reachable in
1) and also in 2) (i.e., must be reachable from both the start BB and
the end BB of the hot region).
3) Mark BBs which end in edges that exit the hot region; we need to
instrument these differently.
Assume that there is 1 free register. On SPARC we use %g1, which LLC
has agreed not to use. Shift a 1 into it at the beginning. At every
edge which corresponds to a conditional branch, we shift 0 for not
taken and 1 for taken into a register. This uniquely numbers the paths
through the hot region. Silently fail if we need more than 64 bits.
At the end BB we call countPath and increment the counter based on %g1
and the return address of the countPath call. We keep track of the
number of iterations and the number of paths. We only run this
version 30 or 40 times.
Find the BBs that total 90% or more of execution, and aggregate them
together to form our trace. But we do not allow more than 5 paths; if
we have more than 5 we take the ones that are executed the most. We
verify our assumption that we picked a hot back-edge in first-level
instrumentation, by making sure that the number of times we took an
exit edge from the hot trace is less than 10% of the number of
LLC has been taught to recognize llvm_first_trigger() calls and NOT
generate saves and restores of caller-saved registers around these
We turn off llvm_first_trigger() calls with NOPs, but this would hide
phase behavior from us (when some funcs/traces stop being hot and
others become hot.)
We have a SIGALRM timer that counts time for us. Every time we get a
SIGALRM we look at our priority queue of locations where we have
removed llvm_first_trigger() calls. Each location is inserted along
with a time when we will next turn instrumentation back on for that
call site. If the time has arrived for a particular call site, we pop
that off the prio. queue and turn instrumentation back on for that
When we finally generate an optimized trace we first copy the code
into the trace cache. This leaves us with 3 copies of the code: the
original code, the instrumented code, and the optimized trace. The
optimized trace does not have instrumentation. The original code and
the instrumented code are modified to have a branch to the trace
cache, where the optimized traces are kept.
We copy the code from the original to the instrumentation version
by tracing the LLVM-to-Machine code basic block map and then copying
each machine code basic block we think is in the hot region into the
trace cache. Then we instrument that code. The process is similar for
generating the final optimized trace; we copy the same basic blocks
because we might need to put in fixup code for exit BBs.
LLVM basic blocks are not typically used in the Reoptimizer except
for the mapping information.
We are restricted to using single instructions to branch between the
original code, trace, and instrumented code. So we have to keep the
code copies in memory near the original code (they can't be far enough
away that a single pc-relative branch would not work.) Malloc() or
data region space is too far away. this impacts the design of the
We use a dummy function that is full of a bunch of for loops which we
overwrite with trace-cache code. The trace manager keeps track of
whether or not we have enough space in the trace cache, etc.
The trace insertion routine takes an original start address, a vector
of machine instructions representing the trace, index of branches and
their corresponding absolute targets, and index of calls and their
corresponding absolute targets.
The trace insertion routine is responsible for inserting branches from
the beginning of the original code to the beginning of the optimized
trace. This is because at some point the trace cache may run out of
space and it may have to evict a trace, at which point the branch to
the trace would also have to be removed. It uses a round-robin
replacement policy; we have found that this is almost as good as LRU
and better than random (especially because of problems fitting the new
We cannot deal with discontiguous trace cache areas. The trace cache
is supposed to be cache-line-aligned, but it is not page-aligned.
We generate instrumentation traces and optimized traces into separate
trace caches. We keep the instrumented code around because you don't
want to delete a trace when you still might have to return to it
(i.e., return from a llvm_first_trigger() or countPath() call.)