with the new pass manager, and no longer relying on analysis groups.
This builds essentially a ground-up new AA infrastructure stack for
LLVM. The core ideas are the same that are used throughout the new pass
manager: type erased polymorphism and direct composition. The design is
as follows:
- FunctionAAResults is a type-erasing alias analysis results aggregation
interface to walk a single query across a range of results from
different alias analyses. Currently this is function-specific as we
always assume that aliasing queries are *within* a function.
- AAResultBase is a CRTP utility providing stub implementations of
various parts of the alias analysis result concept, notably in several
cases in terms of other more general parts of the interface. This can
be used to implement only a narrow part of the interface rather than
the entire interface. This isn't really ideal, this logic should be
hoisted into FunctionAAResults as currently it will cause
a significant amount of redundant work, but it faithfully models the
behavior of the prior infrastructure.
- All the alias analysis passes are ported to be wrapper passes for the
legacy PM and new-style analysis passes for the new PM with a shared
result object. In some cases (most notably CFL), this is an extremely
naive approach that we should revisit when we can specialize for the
new pass manager.
- BasicAA has been restructured to reflect that it is much more
fundamentally a function analysis because it uses dominator trees and
loop info that need to be constructed for each function.
All of the references to getting alias analysis results have been
updated to use the new aggregation interface. All the preservation and
other pass management code has been updated accordingly.
The way the FunctionAAResultsWrapperPass works is to detect the
available alias analyses when run, and add them to the results object.
This means that we should be able to continue to respect when various
passes are added to the pipeline, for example adding CFL or adding TBAA
passes should just cause their results to be available and to get folded
into this. The exception to this rule is BasicAA which really needs to
be a function pass due to using dominator trees and loop info. As
a consequence, the FunctionAAResultsWrapperPass directly depends on
BasicAA and always includes it in the aggregation.
This has significant implications for preserving analyses. Generally,
most passes shouldn't bother preserving FunctionAAResultsWrapperPass
because rebuilding the results just updates the set of known AA passes.
The exception to this rule are LoopPass instances which need to preserve
all the function analyses that the loop pass manager will end up
needing. This means preserving both BasicAAWrapperPass and the
aggregating FunctionAAResultsWrapperPass.
Now, when preserving an alias analysis, you do so by directly preserving
that analysis. This is only necessary for non-immutable-pass-provided
alias analyses though, and there are only three of interest: BasicAA,
GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is
preserved when needed because it (like DominatorTree and LoopInfo) is
marked as a CFG-only pass. I've expanded GlobalsAA into the preserved
set everywhere we previously were preserving all of AliasAnalysis, and
I've added SCEVAA in the intersection of that with where we preserve
SCEV itself.
One significant challenge to all of this is that the CGSCC passes were
actually using the alias analysis implementations by taking advantage of
a pretty amazing set of loop holes in the old pass manager's analysis
management code which allowed analysis groups to slide through in many
cases. Moving away from analysis groups makes this problem much more
obvious. To fix it, I've leveraged the flexibility the design of the new
PM components provides to just directly construct the relevant alias
analyses for the relevant functions in the IPO passes that need them.
This is a bit hacky, but should go away with the new pass manager, and
is already in many ways cleaner than the prior state.
Another significant challenge is that various facilities of the old
alias analysis infrastructure just don't fit any more. The most
significant of these is the alias analysis 'counter' pass. That pass
relied on the ability to snoop on AA queries at different points in the
analysis group chain. Instead, I'm planning to build printing
functionality directly into the aggregation layer. I've not included
that in this patch merely to keep it smaller.
Note that all of this needs a nearly complete rewrite of the AA
documentation. I'm planning to do that, but I'd like to make sure the
new design settles, and to flesh out a bit more of what it looks like in
the new pass manager first.
Differential Revision: http://reviews.llvm.org/D12080
llvm-svn: 247167
PR24139 contains an analysis of poor register allocation. One of the findings
was that when calculating the spill weight, a rematerializable interval once
split is no longer rematerializable. This is because the isRematerializable
check in CalcSpillWeights.cpp does not follow the copies introduced by live
range splitting (after splitting, the live interval register definition is a
copy which is not rematerializable).
Reviewers: qcolombet
Differential Revision: http://reviews.llvm.org/D11686
llvm-svn: 244439
The idea of deferred spilling is to delay the insertion of spill code until the
very end of the allocation. A "candidate" to spill variable might not required
to be spilled because of other evictions that happened after this decision was
taken. The spirit is similar to the optimistic coloring strategy implemented in
Preston and Briggs graph coloring algorithm.
For now, this feature is highly experimental. Although correct, it would require
much more modification to properly model the effect of spilling.
Anyway, this early patch helps prototyping this feature.
Note: The test case cannot unfortunately be reduced and is probably fragile.
llvm-svn: 242585
Pass a const reference to LiveRegMatrix to getRegAllocationHints()
because some targets can prodive better hints if they can test whether a
physreg has been used for register allocation yet.
llvm-svn: 242340
Do not use MachineRegisterInfo::setPhysRegUsed()/isPhysRegUsed()
anymore. This bitset changes function-global state and is set by the
VirtRegRewriter anyway.
Simply use a bitvector private to RAGreedy.
Differential Revision: http://reviews.llvm.org/D10910
llvm-svn: 242169
Specify an allocation order with a register class. This is used by register
allocators with a greedy heuristic. This is usefull as it is sometimes
beneficial to color more constrained classes first.
Differential Revision: http://reviews.llvm.org/D8626
llvm-svn: 233743
When allocating live intervals in linear order and all of them are local
to a single basic block you get an optimal coloring. This is also true
if you reverse the order, but it is not true if you sort live ranges
beginnings in reverse order, change to sort live range endings in
reverse order. Take the following live ranges for example:
|---| |--------|
|----------| |-------|
They get colored suboptimally with 3 registers if you sort the live range
starting points in reverse order (but optimally with live range begins in order,
or live range ends in reverse order).
Apparently the previous strategy was intentional because of allocation
time considerations. I am having a hard time replicating these effects,
while I see substantial improvements in allocation quality with this
change.
No testcase as none of the (in tree) targets use reverse order mode.
Differential Revision: http://reviews.llvm.org/D8625
llvm-svn: 233742
A broken hint is a copy where both ends are assigned different colors. When a
variable gets evicted in the neighborhood of such copies, it is likely we can
reconcile some of them.
** Context **
Copies are inserted during the register allocation via splitting. These split
points are required to relax the constraints on the allocation problem. When
such a point is inserted, both ends of the copy would not share the same color
with respect to the current allocation problem. When variables get evicted,
the allocation problem becomes different and some split point may not be
required anymore. However, the related variables may already have been colored.
This usually shows up in the assembly with pattern like this:
def A
...
save A to B
def A
use A
restore A from B
...
use B
Whereas we could simply have done:
def B
...
def A
use A
...
use B
** Proposed Solution **
A variable having a broken hint is marked for late recoloring if and only if
selecting a register for it evict another variable. Indeed, if no eviction
happens this is pointless to look for recoloring opportunities as it means the
situation was the same as the initial allocation problem where we had to break
the hint.
Finally, when everything has been allocated, we look for recoloring
opportunities for all the identified candidates.
The recoloring is performed very late to rely on accurate copy cost (all
involved variables are allocated).
The recoloring is simple unlike the last change recoloring. It propagates the
color of the broken hint to all its copy-related variables. If the color is
available for them, the recoloring uses it, otherwise it gives up on that hint
even if a more complex coloring would have worked.
The recoloring happens only if it is profitable. The profitability is evaluated
using the expected frequency of the copies of the currently recolored variable
with a) its current color and b) with the target color. If a) is greater or
equal than b), then it is profitable and the recoloring happen.
** Example **
Consider the following example:
BB1:
a =
b =
BB2:
...
= b
= a
Let us assume b gets split:
BB1:
a =
b =
BB2:
c = b
...
d = c
= d
= a
Because of how the allocation work, b, c, and d may be assigned different
colors. Now, if a gets evicted to make room for c, assuming b and d were
assigned to something different than a.
We end up with:
BB1:
a =
st a, SpillSlot
b =
BB2:
c = b
...
d = c
= d
e = ld SpillSlot
= e
This is likely that we can assign the same register for b, c, and d,
getting rid of 2 copies.
** Performances **
Both ARM64 and x86_64 show performance improvements of up to 3% for the
llvm-testsuite + externals with Os and O3. There are a few regressions too that
comes from the (in)accuracy of the block frequency estimate.
<rdar://problem/18312047>
llvm-svn: 225422
Indices into the table are stored in each MCRegisterClass instead of a pointer. A new method, getRegClassName, is added to MCRegisterInfo and TargetRegisterInfo to lookup the string in the table.
llvm-svn: 222118
This patch improves how the different costs (register, interference, spill
and coalescing) relates together. The assumption is now that:
- coalescing (or any other "side effect" of reg alloc) is negative, and
instead of being derived from a spill cost, they use the block
frequency info.
- spill costs are in the [MinSpillCost:+inf( range
- register or interference costs are in [0.0:MinSpillCost( or +inf
The current MinSpillCost is set to 10.0, which is a random value high
enough that the current constraint builders do not need to worry about
when settings costs. It would however be worth adding a normalization
step for register and interference costs as the last step in the
constraint builder chain to ensure they are not greater than SpillMinCost
(unless this has some sense for some architectures). This would work well
with the current builder pipeline, where all costs are tweaked relatively
to each others, but could grow above MinSpillCost if the pipeline is
deep enough.
The current heuristic is tuned to depend rather on the number of uses of
a live interval rather than a density of uses, as used by the greedy
allocator. This heuristic provides a few percent improvement on a number
of benchmarks (eembc, spec, ...) and will definitely need to change once
spill placement is implemented: the current spill placement is really
ineficient, so making the cost proportionnal to the number of use is a
clear win.
llvm-svn: 221292
That commit was introduced in order to help investigate a problem in ARM
codegen breaking from commit 202304 (Add a limit to the heuristic that register
allocates instructions in local order). Recent analisys indicated that the
problem no longer exists, so I'm reverting this change.
See PR18996.
llvm-svn: 218981
It's also possible to just write "= nullptr", but there's some question
of whether that's as readable, so I leave it up to authors to pick which
they prefer for now. If we want to discuss standardizing on one or the
other, we can do that at some point in the future.
llvm-svn: 213438
heuristic.
By default, no functionality change.
This is a follow-up of r212099.
This hook provides a finer grain to control the optimization.
<rdar://problem/17444599>
llvm-svn: 212204
By default, no functionality change.
Before evicting a local variable, this heuristic tries to find another (set of)
local(s) that can be reassigned to a free color.
In some extreme cases (large basic blocks with tons of local variables), the
compilation time is dominated by the local interference checks that this
heuristic must perform, with no code gen gain.
E.g., the motivating example takes 4 minutes to compile with this heuristic, 12
seconds without.
Improving the situation will likely require to make drastic changes to the
register allocator and/or the interference check framework.
For now, provide this flag to better understand the impact of that heuristic.
<rdar://problem/17444599>
llvm-svn: 212099
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
-fexhaustive-register-search option to allow an exhaustive search during last
chance recoloring.
This is related to PR18747
Patch by MAYUR PANDEY <mayur.p@samsung.com>.
llvm-svn: 206072
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
recoloring cut-offs are encountered and register allocation failed.
This is related to PR18747
Patch by MAYUR PANDEY <mayur.p@samsung.com>.
llvm-svn: 205601
When register allocator's stage is RS_Spill, we choose spill over using the CSR
for the first time, if the spill cost is lower than CSRCost.
When register allocator's stage is < RS_Split, we choose pre-splitting over
using the CSR for the first time, if the cost of splitting is lower than
CSRCost.
CSRCost is set with command-line option "regalloc-csr-first-time-cost". The
default value is 0 to generate the same codes as before this commit.
With a value of 15 (1 << 14 is the entry frequency), I measured performance
gain of 3% on 253.perlbmk and 1.7% on 197.parser, with instrumented PGO,
on an arm device.
rdar://16162005
llvm-svn: 204690
Factor out two functions calculateRegionSplitCost and doRegionSplit
from tryRegionSplit. These two functions will be used in coming patches.
rdar://16162005
llvm-svn: 204684
This compiles with no changes to clang/lld/lldb with MSVC and includes
overloads to various functions which are used by those projects and llvm
which have OwningPtr's as parameters. This should allow out of tree
projects some time to move. There are also no changes to libs/Target,
which should help out of tree targets have time to move, if necessary.
llvm-svn: 203083
This is a temporary workaround for native arm linux builds:
PR18996: Changing regalloc order breaks "lencod" on native arm linux builds.
llvm-svn: 202433
This handles pathological cases in which we see 2x increase in spill
code for large blocks (~50k instructions). I don't have a unit test
for this behavior.
Fixes rdar://16072279.
llvm-svn: 202304
find a register.
The idea is to choose a color for the variable that cannot be allocated and
recolor its interferences around. Unlike the current register allocation scheme,
it is allowed to change the color of an already assigned (but maybe not
splittable or spillable) live interval while propagating this change to its
neighbors.
In other word, there are two things that may help finding an available color:
- Already assigned variables (RS_Done) can be recolored to different color.
- The recoloring allows to catch solutions that needs to touch more that just
the neighbors of the current allocated variable.
E.g.,
vA can use {R1, R2 }
vB can use { R2, R3}
vC can use {R1 }
Where vA, vB, and vC cannot be split anymore (they are reloads for instance) and
they all interfere.
vA is assigned R1
vB is assigned R2
vC tries to evict vA but vA is already done.
=> Regular register allocation heuristic fails.
Last chance recoloring kicks in:
vC does as if vA was evicted => vC uses R1.
vC is marked as fixed.
vA needs to find a color.
None are available.
vA cannot evict vC: vC is a fixed virtual register now.
vA does as if vB was evicted => vA uses R2.
vB needs to find a color.
R3 is available.
Recoloring => vC = R1, vA = R2, vB = R3.
<rdar://problem/15947839>
llvm-svn: 200883
The greedy register allocator tries to split a live-range around each
instruction where it is used or defined to relax the constraints on the entire
live-range (this is a last chance split before falling back to spill).
The goal is to have a big live-range that is unconstrained (i.e., that can use
the largest legal register class) and several small local live-range that carry
the constraints implied by each instruction.
E.g.,
Let csti be the constraints on operation i.
V1=
op1 V1(cst1)
op2 V1(cst2)
V1 live-range is constrained on the intersection of cst1 and cst2.
tryInstructionSplit relaxes those constraints by aggressively splitting each
def/use point:
V1=
V2 = V1
V3 = V2
op1 V3(cst1)
V4 = V2
op2 V4(cst2)
Because of how the coalescer infrastructure works, each new variable (V3, V4)
that is alive at the same time as V1 (or its copy, here V2) interfere with V1.
Thus, we end up with an uncoalescable copy for each split point.
To make tryInstructionSplit less aggressive, we check if the split point
actually relaxes the constraints on the whole live-range. If it does not, we do
not insert it.
Indeed, it will not help the global allocation problem:
- V1 will have the same constraints.
- V1 will have the same interference + possibly the newly added split variable
VS.
- VS will produce an uncoalesceable copy if alive at the same time as V1.
<rdar://problem/15570057>
llvm-svn: 198369
This hook reverses the order of assignment for local live ranges. This
will generally allocate shorter local live ranges first. For targets with
many registers, this could reduce regalloc compile time by a large
factor. It should still achieve optimal coloring; however, it can change
register eviction decisions. It is disabled by default for two reasons:
(1) Top-down allocation is simpler and easier to debug for targets that
don't benefit from reversing the order.
(2) Bottom-up allocation could result in poor evicition decisions on some
targets affecting the performance of compiled code.
llvm-svn: 197001
Chandler Carruth