Track new virtual registers by register number, rather than by the live
interval created for them. This is the first step in separating the
creation of new virtual registers and new live intervals. Eventually
live intervals will be created and populated on demand after the virtual
registers have been created and used in instructions.
llvm-svn: 188434
The previous change to local live range allocation also suppressed
eviction of local ranges. In rare cases, this could result in more
expensive register choices. This commit actually revives a feature
that I added long ago: check if live ranges can be reassigned before
eviction. But now it only happens in rare cases of evicting a local
live range because another local live range wants a cheaper register.
The benefit is improved code size for some benchmarks on x86 and armv7.
I measured no significant compile time increase and performance
changes are noise.
llvm-svn: 187140
Also avoid locals evicting locals just because they want a cheaper register.
Problem: MI Sched knows exactly how many registers we have and assumes
they can be colored. In cases where we have large blocks, usually from
unrolled loops, greedy coloring fails. This is a source of
"regressions" from the MI Scheduler on x86. I noticed this issue on
x86 where we have long chains of two-address defs in the same live
range. It's easy to see this in matrix multiplication benchmarks like
IRSmk and even the unit test misched-matmul.ll.
A fundamental difference between the LLVM register allocator and
conventional graph coloring is that in our model a live range can't
discover its neighbors, it can only verify its neighbors. That's why
we initially went for greedy coloring and added eviction to deal with
the hard cases. However, for singly defined and two-address live
ranges, we can optimally color without visiting neighbors simply by
processing the live ranges in instruction order.
Other beneficial side effects:
It is much easier to understand and debug regalloc for large blocks
when the live ranges are allocated in order. Yes, global allocation is
still very confusing, but it's nice to be able to comprehend what
happened locally.
Heuristics could be added to bias register assignment based on
instruction locality (think late register pairing, banks...).
Intuituvely this will make some test cases that are on the threshold
of register pressure more stable.
llvm-svn: 187139
These floats all represented block frequencies anyway, so just use the
BlockFrequency class directly.
Some floating point computations remain in tryLocalSplit(). They are
estimating spill weights which are still floats.
llvm-svn: 186435
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
BitVector/SmallBitVector::reference::operator bool remain implicit since
they model more exactly a bool, rather than something else that can be
boolean tested.
The most common (non-buggy) case are where such objects are used as
return expressions in bool-returning functions or as boolean function
arguments. In those cases I've used (& added if necessary) a named
function to provide the equivalent (or sometimes negative, depending on
convenient wording) test.
One behavior change (YAMLParser) was made, though no test case is
included as I'm not sure how to reach that code path. Essentially any
comparison of llvm::yaml::document_iterators would be invalid if neither
iterator was at the end.
This helped uncover a couple of bugs in Clang - test cases provided for
those in a separate commit along with similar changes to `operator bool`
instances in Clang.
llvm-svn: 181868
Virtual registers with a known preferred register are prioritized by
RAGreedy. This function makes the condition explicit without depending
on getRegAllocPref().
llvm-svn: 169179
Sooooo many of these had incorrect or strange main module includes.
I have manually inspected all of these, and fixed the main module
include to be the nearest plausible thing I could find. If you own or
care about any of these source files, I encourage you to take some time
and check that these edits were sensible. I can't have broken anything
(I strictly added headers, and reordered them, never removed), but they
may not be the headers you'd really like to identify as containing the
API being implemented.
Many forward declarations and missing includes were added to a header
files to allow them to parse cleanly when included first. The main
module rule does in fact have its merits. =]
llvm-svn: 169131
No functional change, just moved header files.
Targets can inject custom passes between register allocation and
rewriting. This makes it possible to tweak the register allocation
before rewriting, using the full global interference checking available
from LiveRegMatrix.
llvm-svn: 168806
Based on CR feedback from r162301 and Craig Topper's refactoring in r162347
here are a few other places that could use the same API (& in one instance drop
a Function.h dependency).
llvm-svn: 162367
Even looking at the revision history I couldn't quite piece together why this
cast was ever written in the first place, but I assume it was because of some
change in the inheritance, perhaps this function was reimplemented in a
derived type & this caller was meant to get the base version (& it wasn't
virtual)?
llvm-svn: 162301
Stop depending on the LiveIntervalUnions in RegAllocBase, they are about
to be removed.
The changes are mostly replacing register alias iterators with regunit
iterators, and querying LiveRegMatrix instrad of RegAllocBase.
InterferenceCache is converted to work with per-regunit
LiveIntervalUnions, and it checks fixed regunit interference separately,
using the fixed live intervals provided by LiveIntervalAnalysis.
The local splitting helper calcGapWeights() is also considering fixed
regunit interference which is kept on the side now.
llvm-svn: 158867
This deduplicates some code from the optimizing register allocators, and
it means that it is now possible to change the register allocators'
solutions simply by editing the VirtRegMap between the register
allocator pass and the rewriter.
llvm-svn: 158249
OK, not really. We don't want to reintroduce the old rewriter hacks.
This patch extracts virtual register rewriting as a separate pass that
runs after the register allocator. This is possible now that
CodeGen/Passes.cpp can configure the full optimizing register allocator
pipeline.
The rewriter pass uses register assignments in VirtRegMap to rewrite
virtual registers to physical registers, and it inserts kill flags based
on live intervals.
These finalization steps are the same for the optimizing register
allocators: RABasic, RAGreedy, and PBQP.
llvm-svn: 158244
LLVM is now -Wunused-private-field clean except for
- lib/MC/MCDisassembler/Disassembler.h. Not sure why it keeps all those unaccessible fields.
- gtest.
llvm-svn: 158096
No functional change intended.
Sorry for the churn. The iterator classes are supposed to help avoid
giant commits like this one in the future. The TableGen-produced
register lists are getting quite large, and it may be necessary to
change the table representation.
This makes it possible to do so without changing all clients (again).
llvm-svn: 157854
It helps compile exotic inline asm. In the test case, normal GR32
virtual registers use up eax-edx so the final GR32_ABCD live range has
no registers left. Since all the live ranges were tiny, we had no way of
prioritizing the smaller register class.
This patch allows tiny unspillable live ranges to be evicted by tiny
unspillable live ranges from a smaller register class.
<rdar://problem/11542429>
llvm-svn: 157715
Live ranges with a constrained register class may benefit from splitting
around individual uses. It allows the remaining live range to use a
larger register class where it may allocate. This is like spilling to a
different register class.
This is only attempted on constrained register classes.
<rdar://problem/11438902>
llvm-svn: 157354
Passes after RegAlloc should be able to rely on MRI->getNumVirtRegs() == 0.
This makes sharing code for pre/postRA passes more robust.
Now, to check if a pass is running before the RA pipeline begins, use MRI->isSSA().
To check if a pass is running after the RA pipeline ends, use !MRI->getNumVirtRegs().
PEI resets virtual regs when it's done scavenging.
PTX will either have to provide its own PEI pass or assign physregs.
llvm-svn: 151032
This makes global live range splitting behave identically with and
without register mask operands.
This is not necessarily the best way of using register masks for live
range splitting. It would be more efficient to first split global live
ranges around calls (i.e., register masks), and reserve the fine grained
per-physreg interference guidance for global live ranges that do not
cross calls.
For now the goal is to produce identical assembly when enabling register
masks.
llvm-svn: 150259
Creates a configurable regalloc pipeline.
Ensure specific llc options do what they say and nothing more: -reglloc=... has no effect other than selecting the allocator pass itself. This patch introduces a new umbrella flag, "-optimize-regalloc", to enable/disable the optimizing regalloc "superpass". This allows for example testing coalscing and scheduling under -O0 or vice-versa.
When a CodeGen pass requires the MachineFunction to have a particular property, we need to explicitly define that property so it can be directly queried rather than naming a specific Pass. For example, to check for SSA, use MRI->isSSA, not addRequired<PHIElimination>.
CodeGen transformation passes are never "required" as an analysis
ProcessImplicitDefs does not require LiveVariables.
We have a plan to massively simplify some of the early passes within the regalloc superpass.
llvm-svn: 150226
This only adds the interference checks required for correctness.
We still need to take advantage of register masks for the
interference driven live range splitting.
llvm-svn: 150191
THe LRE_DidCloneVirtReg callback may be called with vitual registers
that RAGreedy doesn't even know about yet. In that case, there are no
data structures to update.
llvm-svn: 139702
SplitKit always computes a complement live range to cover the places
where the original live range was live, but no explicit region has been
allocated.
Currently, the complement live range is created to be as small as
possible - it never overlaps any of the regions. This minimizes
register pressure, but if the complement is going to be spilled anyway,
that is not very important. The spiller will eliminate redundant
spills, and hoist others by making the spill slot live range overlap
some of the regions created by splitting. Stack slots are cheap.
This patch adds the interface to enable spill modes in SplitKit. In
spill mode, SplitKit will assume that the complement is going to spill,
so it will allow it to overlap regions in order to avoid back-copies.
By doing some of the spiller's work early, the complement live range
becomes simpler. In some cases, it can become much simpler because no
extra PHI-defs are required. This will speed up both splitting and
spilling.
This is only the interface to enable spill modes, no implementation yet.
llvm-svn: 139500
The local ranges created get to stay in the RS_New stage, just like for
local and region splitting.
This gives tryLocalSplit a bit more freedom the first time it sees one
of these new local ranges.
llvm-svn: 137001
Normally, we don't create a live range for a single instruction in a
basic block, the spiller does that anyway. However, when splitting a
live range that belongs to a proper register sub-class, inserting these
extra COPY instructions completely remove the constraints from the
remainder interval, and it may be allocated from the larger super-class.
The spiller will mop up these small live ranges if we end up spilling
anyway. It calls them snippets.
llvm-svn: 136989
This helps generate better code in functions with high register
pressure.
The previous version of compact region splitting caused regressions
because the regions were a bit too large. A stronger negative bias
applied in r136832 fixed this problem.
llvm-svn: 136836
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
There are two conflicting strategies in play:
- Under high register pressure, we want to assign large live ranges
first. Smaller live ranges are easier to place afterwards.
- Live range splitting is guided by interference, so splitting should be
deferred until interference is as realistic as possible.
With the recent changes to the live range stages, and with compact
regions enabled, it is less traumatic to split a live range too early.
If some of the split products were too big, they can often be split
again.
By reversing the RS_Split order, we get this queue order:
1. Normal live ranges, large to small.
2. RS_Split live ranges, large to small.
The large-to-small order improves RAGreedy's puzzle solving skills under
high register pressure. It may cause a bit more iterated splitting, but
we handle that better now.
With this change, -compact-regions is mostly an improvement on SPEC.
llvm-svn: 136388
When splitting global live ranges, it is now possible to split for
multiple destination intervals at once. Previously, we only had the main
and stack intervals.
Each edge bundle is assigned to a split candidate, and splitAroundRegion
will insert copies between the candidate intervals and the stack
interval as needed.
The multi-way splitting is used to split around compact regions when
enabled with -compact-regions. The best candidate register still gets
all the bundles it wants, but everything outside the main interval is
first split around compact regions before we create single-block
intervals.
Compact region splitting still causes some regressions, so it is not
enabled by default.
llvm-svn: 136186
When dead code elimination deletes a PHI value, the virtual register may
split into multiple connected components. In that case, revert each
component to the RS_Assign stage.
The new components are guaranteed to be smaller (the original value
numbers are distributed among the components), so this will always be
making progress. The components are now allowed to evict other live
ranges or be split again.
llvm-svn: 136034
This mechanism already exists, but the RS_Split2 stage makes it clearer.
When live range splitting creates ranges that may not be making
progress, they are marked RS_Split2 instead of RS_New. These ranges may
be split again, but only in a way that can be proven to make progress.
For local ranges, that means they must be split into ranges used by
strictly fewer instructions.
For global ranges, region splitting is bypassed and the RS_Split2
ranges go straight to per-block splitting.
llvm-svn: 135912
The stage is used to control where a live range is going, not where it
is coming from. Live ranges created by splitting will usually be marked
RS_New, but some are marked RS_Spill to avoid wasting time trying to
split them again.
The old RS_Global and RS_Local stages are merged - they are really the
same thing for local and global live ranges.
llvm-svn: 135911
This method computes the edge bundles that should be live when splitting
around a compact region. This is independent of interference.
The function returns false if the live range was already a compact
region, or the compact region doesn't have any live bundles - it would
be the same as splitting around basic blocks.
Compact regions are computed using the normal spill placement code. We
pretend there is interference in all live-through blocks that don't use
the live range. This removes all edges from the Hopfield network used
for spill placement, so it converges instantly.
llvm-svn: 135847
A split candidate can have a null PhysReg which means that it doesn't
map to a real interference pattern. Instead, pretend that all through
blocks have interference.
This makes it possible to generate compact regions where the live range
doesn't go through blocks that don't use it. The live range will still
be live between directly connected blocks with uses.
Splitting around a compact region tends to produce a live range with a
high spill weight, so it may evict a less dense live range.
llvm-svn: 135845
This gets rid of some of the gory splitting details in RAGreedy and
makes them available to future SplitKit clients.
Slightly generalize the functionality to support multi-way splitting.
Specifically, SplitEditor::splitLiveThroughBlock() supports switching
between different register intervals in a block.
llvm-svn: 135307
Original commit message:
Count references to interference cache entries.
Each InterferenceCache::Cursor instance references a cache entry. A
non-zero reference count guarantees that the entry won't be reused for a
new register.
This makes it possible to have multiple live cursors examining
interference for different physregs.
The total number of live cursors into a cache must be kept below
InterferenceCache::getMaxCursors().
Code generation should be unaffected by this change, and it doesn't seem
to affect the cache replacement strategy either.
llvm-svn: 135130
Each InterferenceCache::Cursor instance references a cache entry. A
non-zero reference count guarantees that the entry won't be reused for a
new register.
This makes it possible to have multiple live cursors examining
interference for different physregs.
The total number of live cursors into a cache must be kept below
InterferenceCache::getMaxCursors().
Code generation should be unaffected by this change, and it doesn't seem
to affect the cache replacement strategy either.
llvm-svn: 135121
The cache entry referenced by the best split candidate could become
clobbered by an unsuccessful candidate.
The correct fix here is to use reference counts on the cache entries.
Coming up.
llvm-svn: 135113
Some pysical registers create split solutions that would spill anywhere.
They should not even be considered in future multi-way global splits.
This does not affect code generation (yet).
llvm-svn: 135080
RAGreedy::tryAssign will now evict interference from the preferred
register even when another register is free.
To support this, add the EvictionCost struct that counts how many hints
are broken by an eviction. We don't want to break one hint just to
satisfy another.
Rename canEvict to shouldEvict, and add the first bit of eviction policy
that doesn't depend on spill weights: Always make room in the preferred
register as long as the evictees can be split and aren't already
assigned to their preferred register.
Also make the CSR avoidance more accurate. When looking for a cheaper
register it is OK to use a new volatile register. Only CSR aliases that
have never been used before should be avoided.
llvm-svn: 134735
This is impossible in theory, I can prove it. In practice, our near-zero
threshold can cause the network to oscillate between equally good
solutions.
<rdar://problem/9720596>
llvm-svn: 134428
A split point inserted in a block with a landing pad successor may be
hoisted above the call to ensure that it dominates all successors. The
code that handles the rest of the basic block must take this into
account.
I am not including a test case, it would be very fragile. PR10244 comes
from building clang with exceptions enabled.
llvm-svn: 134369
Every live range is assigned a cascade number the first time it is
involved in an eviction. As the evictor, it gets a new cascade number.
Every evictee is assigned the same cascade number as the evictor.
Eviction is prohibited if the evictor has a lower assigned cascade
number than the evictee.
This means that assigned cascade numbers are monotonically increasing
with every eviction, yet they are bounded by NextCascade which can only
be incremented by new live ranges. Thus, infinite loops cannot happen,
but eviction cascades can still be triggered by new live ranges as we
want.
Thanks to Andy for explaining this to me.
llvm-svn: 134303
This patch will sometimes choose live range split points next to
interference instead of always splitting next to a register point. That
means spill code can now appear almost anywhere, and it was necessary
to fix code that didn't expect that.
The difficult places were:
- Between a CALL returning a value on the x87 stack and the
corresponding FpPOP_RETVAL (was FpGET_ST0). Probably also near x87
inline assembly, but that didn't actually show up in testing.
- Between a CALL popping arguments off the stack and the corresponding
ADJCALLSTACKUP.
Both are fixed now. The only place spill code can't appear is after
terminators, see SplitAnalysis::getLastSplitPoint.
Original commit message:
Rewrite RAGreedy::splitAroundRegion, now with cool ASCII art.
This function has to deal with a lot of special cases, and the old
version got it wrong sometimes. In particular, it would sometimes leave
multiple uses in the stack interval in a single block. That causes bad
code with multiple reloads in the same basic block.
The new version handles block entry and exit in a single pass. It first
eliminates all the easy cases, and then goes on to create a local
interval for the blocks with difficult interference. Previously, we
would only create the local interval for completely isolated blocks.
It can happen that the stack interval becomes completely empty because
we could allocate a register in all edge bundles, and the new local
intervals deal with the interference. The empty stack interval is
harmless, but we need to remove a SplitKit assertion that checks for
empty intervals.
llvm-svn: 134125
This function has to deal with a lot of special cases, and the old
version got it wrong sometimes. In particular, it would sometimes leave
multiple uses in the stack interval in a single block. That causes bad
code with multiple reloads in the same basic block.
The new version handles block entry and exit in a single pass. It first
eliminates all the easy cases, and then goes on to create a local
interval for the blocks with difficult interference. Previously, we
would only create the local interval for completely isolated blocks.
It can happen that the stack interval becomes completely empty because
we could allocate a register in all edge bundles, and the new local
intervals deal with the interference. The empty stack interval is
harmless, but we need to remove a SplitKit assertion that checks for
empty intervals.
llvm-svn: 134047
When local live range splitting creates a live range with the same
number of instructions as the old range, mark it as RS_Local. When such
a range is seen again, require that it be split in a way that reduces
the number of instructions. That guarantees we are making progress while
still being able to perform 3 -> 2+3 splits as required by PR10070.
This also means that the PrevSlot map is no longer needed. This was also
used to estimate new spill weights, but that is no longer necessary
after slotIndexes::insertMachineInstrInMaps() got the extra Late
insertion argument.
llvm-svn: 132697
of reserved registers.
Use RegisterClassInfo in RABasic as well. This slightly changes som
allocation orders because RegisterClassInfo puts CSR aliases last.
llvm-svn: 132581
When assigned ranges are evicted, they are put in the RS_Evicted stage and are
not allowed to evict anything else. That prevents looping automatically.
When evicting ranges just to get a cheaper register, use only spill weights to
find the possible candidates. Avoid breaking hints for this purpose, it is not
worth it.
Start implementing more complex eviction heuristics, guarded by the temporary
-complex-eviction flag. The initial version permits a heavier range to be
evicted if it doesn't have any uses where the evicting range is live. This makes
it a good candidate for live ranfge splitting.
llvm-svn: 132358
Delete the Kill and Def markers in BlockInfo. They are no longer
necessary when BlockInfo describes a continuous live range.
This only affects the relatively rare kind of basic block where a live
range looks like this:
|---x o---|
Now live range splitting can pretend that it is looking at two blocks:
|---x
o---|
This allows the code to be simplified a bit.
llvm-svn: 132245
It is important that this function returns the same number of live blocks as
countLiveBlocks(CurLI) because live range splitting uses the number of live
blocks to ensure it is making progress.
This is in preparation of supporting duplicate UseBlock entries for basic blocks
that have a virtual register live-in and live-out, but not live-though.
llvm-svn: 132244
This doesn't change functionality (much), but it allows for a more fine-grained
eviction policy. The current policy only compares spill weights, and that is not
always the best thing to do. Spill weights are designed to serve linear scan,
and they don't consider live range splitting.
Add a mechanism so canEvict() can request that a live range be evicted and
split/spilled. This is to avoid infinite eviction loops.
llvm-svn: 132101
This can't be just an assertion, users can always write impossible inline
assembly. Such an assembly statement should be included in the error message.
llvm-svn: 131024
After a virtual register is split, update any debug user variables that resided
in the old register. This ensures that the LiveDebugVariables are still correct
after register allocation.
This may create DBG_VALUE instructions that place a user variable in a register
in parts of the function and in a stack slot in other parts. DwarfDebug
currently doesn't support that.
llvm-svn: 130998
Register coalescing can sometimes create live ranges that end in the middle of a
basic block without any killing instruction. When SplitKit detects this, it will
repair the live range by shrinking it to its uses.
Live range splitting also needs to know about this. When the range shrinks so
much that it becomes allocatable, live range splitting fails because it can't
find a good split point. It is paranoid about making progress, so an allocatable
range is considered an error.
The coalescer should really not be creating these bad live ranges. They appear
when coalescing dead copies.
llvm-svn: 130787
These intervals are allocatable immediately after splitting, but they may be
evicted because of later splitting. This is rare, but when it happens they
should be split again.
The remainder intervals that cannot be allocated after splitting still move
directly to spilling.
SplitEditor::finish can optionally provide a mapping from new live intervals
back to the original interval indexes returned by openIntv().
Each original interval index can map to multiple new intervals after connected
components have been separated. Dead code elimination may also add existing
intervals to the list.
The reverse mapping allows the SplitEditor client to treat the new intervals
differently depending on the split region they came from.
llvm-svn: 129925
On the x86-64 and thumb2 targets, some registers are more expensive to encode
than others in the same register class.
Add a CostPerUse field to the TableGen register description, and make it
available from TRI->getCostPerUse. This represents the cost of a REX prefix or a
32-bit instruction encoding required by choosing a high register.
Teach the greedy register allocator to prefer cheap registers for busy live
ranges (as indicated by spill weight).
llvm-svn: 129864
This merges the behavior of splitSingleBlocks into splitAroundRegion, so the
RS_Region and RS_Block register stages can be coalesced. That means the leftover
intervals after region splitting go directly to spilling instead of a second
pass of per-block splitting.
llvm-svn: 129379
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
About 90% of the relevant blocks are live-through without uses, and the only
information required about them is their number. This saves memory and enables
later optimizations that need to look at only the use-blocks.
llvm-svn: 128985
When the greedy register allocator is splitting multiple global live ranges, it
tends to look at the same interference data many times. The InterferenceCache
class caches queries for unaltered LiveIntervalUnions.
llvm-svn: 128764
When DCE clones a live range because it separates into connected components,
make sure that the clones enter the same register allocator stage as the
register they were cloned from.
For instance, clones may be split even when they where created during spilling.
Other registers created during spilling are not candidates for splitting or even
(re-)spilling.
llvm-svn: 128524
The reassignment phase was able to move interference with a higher spill weight,
but it didn't happen very often and it was fairly expensive.
The existing interference eviction picks up the slack.
llvm-svn: 128397
This simplifies the code and makes it faster too.
The interference patterns are saved for each candidate register. It will be
reused for actually executing the split. Work in progress.
llvm-svn: 127054
Mark Lacey