Using Chandler's words from r265331:
This commit was greatly exacerbating PR17409 and effectively regressed
build time for lot of (very large) code when compiled with ASan or MSan.
PR17409 is fixed by r269249, so this is fine to reapply r263460.
Original commit message:
The bad behavior happens when we have a function with a long linear
chain of basic blocks, and have a live range spanning most of this
chain, but with very few uses.
Let say we have only 2 uses.
The Hopfield network is only seeded with two active blocks where the
uses are, and each iteration of the outer loop in
`RAGreedy::growRegion()` only adds two new nodes to the network due to
the completely linear shape of the CFG. Meanwhile,
`SpillPlacer->iterate()` visits the whole set of discovered nodes, which
adds up to a quadratic algorithm.
This is an historical accident effect from r129188.
When the Hopfield network is expanding, most of the action is happening
on the frontier where new nodes are being added. The internal nodes in
the network are not likely to be flip-flopping much, or they will at
least settle down very quickly. This means that while
`SpillPlacer->iterate()` is recomputing all the nodes in the network, it
is probably only the two frontier nodes that are changing their output.
Instead of recomputing the whole network on each iteration, we can
maintain a SparseSet of nodes that need to be updated:
- `SpillPlacement::activate()` adds the node to the todo list.
- When a node changes value (i.e., `update()` returns true), its
neighbors are added to the todo list.
- `SpillPlacement::iterate()` only updates the nodes in the list.
The result of Hopfield iterations is not necessarily exact. It should
converge to a local minimum, but there is no guarantee that it will find
a global minimum. It is possible that updating nodes in a different
order will cause us to switch to a different local minimum. In other
words, this is not NFC, but although I saw a few runtime improvements
and regressions when I benchmarked this change, those were side effects
and actually the performance change is in the noise as expected.
Huge thanks to Jakob Stoklund Olesen <stoklund@2pi.dk> for his
feedbacks, guidance and time for the review.
llvm-svn: 270149
That commit looks wonderful and awesome. Sadly, it greatly exacerbates
PR17409 and effectively regresses build time for a lot of (very large)
code when compiled with ASan or MSan.
We thought this could be fixed forward by landing D15302 which at last
fixes that PR, but some issues were discovered and it looks like that
got reverted, so reverting this as well temporarily. As soon as the fix
for PR17409 lands and sticks, we should re-land this patch as it won't
trigger more significant test cases hitting that bug.
Many thanks to Quentin and Wei here as they're doing all the awesome
hard work!!!
llvm-svn: 265331
The bad behavior happens when we have a function with a long linear chain of
basic blocks, and have a live range spanning most of this chain, but with very
few uses.
Let say we have only 2 uses.
The Hopfield network is only seeded with two active blocks where the uses are,
and each iteration of the outer loop in `RAGreedy::growRegion()` only adds two
new nodes to the network due to the completely linear shape of the CFG.
Meanwhile, `SpillPlacer->iterate()` visits the whole set of discovered nodes,
which adds up to a quadratic algorithm.
This is an historical accident effect from r129188.
When the Hopfield network is expanding, most of the action is happening on the
frontier where new nodes are being added. The internal nodes in the network are
not likely to be flip-flopping much, or they will at least settle down very
quickly. This means that while `SpillPlacer->iterate()` is recomputing all the
nodes in the network, it is probably only the two frontier nodes that are
changing their output.
Instead of recomputing the whole network on each iteration, we can maintain a
SparseSet of nodes that need to be updated:
- `SpillPlacement::activate()` adds the node to the todo list.
- When a node changes value (i.e., `update()` returns true), its neighbors are
added to the todo list.
- `SpillPlacement::iterate()` only updates the nodes in the list.
The result of Hopfield iterations is not necessarily exact. It should converge
to a local minimum, but there is no guarantee that it will find a global
minimum. It is possible that updating nodes in a different order will cause us
to switch to a different local minimum. In other words, this is not NFC, but
although I saw a few runtime improvements and regressions when I benchmarked
this change, those were side effects and actually the performance change is in
the noise as expected.
Huge thanks to Jakob Stoklund Olesen <stoklund@2pi.dk> for his feedbacks,
guidance and time for the review.
llvm-svn: 263460
to be a ManagedStatic in r218163 to not be a global variable written and
read to from within the innards of SpillPlacement.
This will fix a really scary race condition for anyone that has two
copies of LLVM running spill placement concurrently. Yikes!
This will also fix a really significant compile time hit that r218163
caused because the spill placement threshold read is actually in the
*very* hot path of this code. The memory fence on each read was showing
up as huge compile time regressions when spilling is responsible for
most of the compile time. For example, optimizing sanitized code showed
over 50% compile time regressions here. =/
llvm-svn: 218921
define below all header includes in the lib/CodeGen/... tree. While the
current modules implementation doesn't check for this kind of ODR
violation yet, it is likely to grow support for it in the future. It
also removes one layer of macro pollution across all the included
headers.
Other sub-trees will follow.
llvm-svn: 206837
Until r197284, the entry frequency was constant -- i.e., set to 2^14.
Although current ToT still has a constant entry frequency, since r197284
that has been an implementation detail (which is soon going to change).
- r204690 made the wrong assumption for the CSRCost metric. Adjust
callee-saved register cost based on entry frequency.
- r185393 made the wrong assumption (although it was valid at the
time). Update SpillPlacement.cpp::Threshold to be relative to the
entry frequency.
Since ToT still has 2^14 entry frequency, this should have no observable
functionality change.
<rdar://problem/14292693>
llvm-svn: 205789
Inside iterate, we scan backwards then scan forwards in a loop. When iteration
is not zero, the last node was just updated so we can skip it. But when
iteration is zero, we can't skip the last node.
For the testing case, fixing this will save a spill and move register copies
from hot path to cold path.
llvm-svn: 202557
Original commit message:
Remove floating point computations from SpillPlacement.cpp.
Patch by Benjamin Kramer!
Use the BlockFrequency class instead of floats in the Hopfield network
computations. This rescales the node Bias field from a [-2;2] float
range to two block frequencies BiasN and BiasP pulling in opposite
directions. This construct has a more predictable behavior when block
frequencies saturate.
The per-node scaling factors are no longer necessary, assuming the block
frequencies around a bundle are consistent.
This patch can cause the register allocator to make different spilling
decisions. The differences should be small.
llvm-svn: 186434
"Remove floating point computations form SpillPlacement.cpp."
These commits caused test failures in lencod on clang-native-arm-lnt.
I suspect these changes are only exposing an existing issue, but
reverting anyway to keep the bots passing while we investigate.
llvm-svn: 185447
Patch by Benjamin Kramer!
Use the BlockFrequency class instead of floats in the Hopfield network
computations. This rescales the node Bias field from a [-2;2] float
range to two block frequencies BiasN and BiasP pulling in opposite
directions. This construct has a more predictable behavior when block
frequencies saturate.
The per-node scaling factors are no longer necessary, assuming the block
frequencies around a bundle are consistent.
This patch can cause the register allocator to make different spilling
decisions. The differences should be small.
llvm-svn: 185393
The main advantages here are way better heuristics, taking into account not
just loop depth but also __builtin_expect and other static heuristics and will
eventually learn how to use profile info. Most of the work in this patch is
pushing the MachineBlockFrequencyInfo analysis into the right places.
This is good for a 5% speedup on zlib's deflate (x86_64), there were some very
unfortunate spilling decisions in its hottest loop in longest_match(). Other
benchmarks I tried were mostly neutral.
This changes register allocation in subtle ways, update the tests for it.
2012-02-20-MachineCPBug.ll was deleted as it's very fragile and the instruction
it looked for was gone already (but the FileCheck pattern picked up unrelated
stuff).
llvm-svn: 184105
This helps compile time when the greedy register allocator splits live
ranges in giant functions. Without the bias, we would try to grow
regions through the giant edge bundles, usually to find out that the
region became too big and expensive.
If a live range has many uses in blocks near the giant bundle, the small
negative bias doesn't make a big difference, and we still consider
regions including the giant edge bundle.
Giant edge bundles are usually connected to landing pads or indirect
branches.
llvm-svn: 157174
Apply twice the negative bias on transparent blocks when computing the
compact regions. This excludes loop backedges from the region when only
one of the loop blocks uses the register.
Previously, we would include the backedge in the region if the loop
preheader and the loop latch both used the register, but the loop header
didn't.
When both the header and latch blocks use the register, we still keep it
live on the backedge.
llvm-svn: 136832
The PrefBoth constraint is used for blocks that ideally want a live-in
value both on the stack and in a register. This would be used by a block
that has a use before interference forces a spill.
Secondly, add the ChangesValue flag to BlockConstraint. This tells
SpillPlacement if a live-in value on the stack can be reused as a
live-out stack value for free. If the block redefines the virtual
register, a spill would be required for that.
This extra information will be used by SpillPlacement to more accurately
calculate spill costs when a value can exist both on the stack and in a
register.
The simplest example is a basic block that reads the virtual register,
but doesn't change its value. Spilling around such a block requires a
reload, but no spill in the block.
The spiller already knows this, but the spill placer doesn't. That can
sometimes lead to suboptimal regions.
llvm-svn: 136731
This method matches addLinks - All the listed blocks are considered to
have interference, so they add a negative bias to their bundles.
This could also be done by addConstraints, but that requires building a
separate BlockConstraint array.
llvm-svn: 135844
It is common for large live ranges to have few basic blocks with register uses
and many live-through blocks without any uses. This approach grows the Hopfield
network incrementally around the use blocks, completely avoiding checking
interference for some through blocks.
llvm-svn: 129188
Analyze the live range's behavior entering and leaving basic blocks. Compute an
interference pattern for each allocation candidate, and use SpillPlacement to
find an optimal region where that register can be live.
This code is still not enabled.
llvm-svn: 123774
This pass precomputes CFG block frequency information that can be used by the
register allocator to find optimal spill code placement.
Given an interference pattern, placeSpills() will compute which basic blocks
should have the current variable enter or exit in a register, and which blocks
prefer the stack.
The algorithm is ready to consume block frequencies from profiling data, but for
now it gets by with the static estimates used for spill weights.
This is a work in progress and still not hooked up to RegAllocGreedy.
llvm-svn: 122938
Quentin Colombet