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
The previous PBQP solver was very robust but consumed a lot of memory,
performed a lot of redundant computation, and contained some unnecessarily tight
coupling that prevented experimentation with novel solution techniques. This new
solver is an attempt to address these shortcomings.
Important/interesting changes:
1) The domain-independent PBQP solver class, HeuristicSolverImpl, is gone.
It is replaced by a register allocation specific solver, PBQP::RegAlloc::Solver
(see RegAllocSolver.h).
The optimal reduction rules and the backpropagation algorithm have been extracted
into stand-alone functions (see ReductionRules.h), which can be used to build
domain specific PBQP solvers. This provides many more opportunities for
domain-specific knowledge to inform the PBQP solvers' decisions. In theory this
should allow us to generate better solutions. In practice, we can at least test
out ideas now.
As a side benefit, I believe the new solver is more readable than the old one.
2) The solver type is now a template parameter of the PBQP graph.
This allows the graph to notify the solver of any modifications made (e.g. by
domain independent rules) without the overhead of a virtual call. It also allows
the solver to supply policy information to the graph (see below).
3) Significantly reduced memory overhead.
Memory management policy is now an explicit property of the PBQP graph (via
the CostAllocator typedef on the graph's solver template argument). Because PBQP
graphs for register allocation tend to contain many redundant instances of
single values (E.g. the value representing an interference constraint between
GPRs), the new RASolver class uses a uniquing scheme. This massively reduces
memory consumption for large register allocation problems. For example, looking
at the largest interference graph in each of the SPEC2006 benchmarks (the
largest graph will always set the memory consumption high-water mark for PBQP),
the average memory reduction for the PBQP costs was 400x. That's times, not
percent. The highest was 1400x. Yikes. So - this is fixed.
"PBQP: No longer feasting upon every last byte of your RAM".
Minor details:
- Fully C++11'd. Never copy-construct another vector/matrix!
- Cute tricks with cost metadata: Metadata that is derived solely from cost
matrices/vectors is attached directly to the cost instances themselves. That way
if you unique the costs you never have to recompute the metadata. 400x less
memory means 400x less cost metadata (re)computation.
Special thanks to Arnaud de Grandmaison, who has been the source of much
encouragement, and of many very useful test cases.
This new solver forms the basis for future work, of which there's plenty to do.
I will be adding TODO notes shortly.
- Lang.
llvm-svn: 202551
during the finalization for CGDebugInfo in clang we would RAUW
a type and it would result in a corrupted MDNode for an
imported declaration.
Testcase pending as reducing has been difficult.
llvm-svn: 202540
Tools that use the CommandLine library currently exit with an error
when invoked with -version or -help. This is unusual and non-standard,
so we'll fix them to exit successfully instead.
I don't expect that anyone relies on the current behaviour, so this
should be a fairly safe change.
llvm-svn: 202530
X86Operand is extracted into individual header, because it allows to create an
arbitrary memory operand and append it to MCInst. It'll be reused in X86 inline
assembly instrumentation.
Patch by Yuri Gorshenin.
llvm-svn: 202496
* Align targets of indirect jumps to instruction bundle boundaries (in MI layer).
* Add masking instructions before indirect jumps (in MC layer).
Differential Revision: http://llvm-reviews.chandlerc.com/D2847
llvm-svn: 202479
A 'remark' is information that is not an error or a warning, but rather some
additional information provided to the user. In contrast to a 'note' a 'remark'
is an independent diagnostic, whereas a 'note' always depends on another
diagnostic.
A typical use case for remark nodes is information provided to the user, e.g.
information provided by the vectorizer about loops that have been vectorized.
llvm-svn: 202474
The PPC isel instruction can fold 0 into the first operand (thus eliminating
the need to materialize a zero-containing register when the 'true' result of
the isel is 0). When the isel is fed by a bit register operation that we can
invert, do so as part of the bit-register-operation peephole routine.
llvm-svn: 202469
The CR bit tracking code broke PPC/Darwin; trying to get it working again...
(the darwin11 builder, which defaults to the darwin ABI when running PPC tests,
asserted when running test/CodeGen/PowerPC/inverted-bool-compares.ll)
llvm-svn: 202459
This change enables tracking i1 values in the PowerPC backend using the
condition register bits. These bits can be treated on PowerPC as separate
registers; individual bit operations (and, or, xor, etc.) are supported.
Tracking booleans in CR bits has several advantages:
- Reduction in register pressure (because we no longer need GPRs to store
boolean values).
- Logical operations on booleans can be handled more efficiently; we used to
have to move all results from comparisons into GPRs, perform promoted
logical operations in GPRs, and then move the result back into condition
register bits to be used by conditional branches. This can be very
inefficient, because the throughput of these CR <-> GPR moves have high
latency and low throughput (especially when other associated instructions
are accounted for).
- On the POWER7 and similar cores, we can increase total throughput by using
the CR bits. CR bit operations have a dedicated functional unit.
Most of this is more-or-less mechanical: Adjustments were needed in the
calling-convention code, support was added for spilling/restoring individual
condition-register bits, and conditional branch instruction definitions taking
specific CR bits were added (plus patterns and code for generating bit-level
operations).
This is enabled by default when running at -O2 and higher. For -O0 and -O1,
where the ability to debug is more important, this feature is disabled by
default. Individual CR bits do not have assigned DWARF register numbers,
and storing values in CR bits makes them invisible to the debugger.
It is critical, however, that we don't move i1 values that have been promoted
to larger values (such as those passed as function arguments) into bit
registers only to quickly turn around and move the values back into GPRs (such
as happens when values are returned by functions). A pair of target-specific
DAG combines are added to remove the trunc/extends in:
trunc(binary-ops(binary-ops(zext(x), zext(y)), ...)
and:
zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...)
In short, we only want to use CR bits where some of the i1 values come from
comparisons or are used by conditional branches or selects. To put it another
way, if we can do the entire i1 computation in GPRs, then we probably should
(on the POWER7, the GPR-operation throughput is higher, and for all cores, the
CR <-> GPR moves are expensive).
POWER7 test-suite performance results (from 10 runs in each configuration):
SingleSource/Benchmarks/Misc/mandel-2: 35% speedup
MultiSource/Benchmarks/Prolangs-C++/city/city: 21% speedup
MultiSource/Benchmarks/MiBench/automotive-susan: 23% speedup
SingleSource/Benchmarks/CoyoteBench/huffbench: 13% speedup
SingleSource/Benchmarks/Misc-C++/Large/sphereflake: 13% speedup
SingleSource/Benchmarks/Misc-C++/mandel-text: 10% speedup
SingleSource/Benchmarks/Misc-C++-EH/spirit: 10% slowdown
MultiSource/Applications/lemon/lemon: 8% slowdown
llvm-svn: 202451
This extract-and-trunc vector optimization cannot work for i1 values as
currently implemented, and so I'm disabling this for now for i1 values. In the
future, this can be fixed properly.
Soon I'll commit support for i1 CR bit tracking in the PowerPC backend, and
this will be covered by one of the existing regression tests.
llvm-svn: 202449
This is a temporary workaround for native arm linux builds:
PR18996: Changing regalloc order breaks "lencod" on native arm linux builds.
llvm-svn: 202433
scan the register file for sub- and super-registers.
No functionality change intended.
(Tests are updated because the comments in the assembler output are
different.)
llvm-svn: 202416
If a function returns a large struct by value return the first 4 words
in registers and the rest on the stack in a location reserved by the
caller. This is needed to support the xC language which supports
functions returning an arbitrary number of return values. This is
r202397 reapplied with a fix to avoid an uninitialized read of a member.
llvm-svn: 202414
Summary:
If a function returns a large struct by value return the first 4 words
in registers and the rest on the stack in a location reserved by the
caller. This is needed to support the xC language which supports
functions returning an arbitrary number of return values.
Reviewers: robertlytton
Reviewed By: robertlytton
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2889
llvm-svn: 202397
Summary:
If the src, dst and size of a memcpy are known to be 4 byte aligned we
can call __memcpy_4() instead of memcpy().
Reviewers: robertlytton
Reviewed By: robertlytton
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2871
llvm-svn: 202395
any ranges - this includes CU ranges where we were previously emitting an
end list marker even if we didn't have a list.
Testcase includes a test for line table only code emission as the problem
was noticed while writing this test.
llvm-svn: 202357
If the SI_KILL operand is constant, we can either clear the exec mask if
the operand is negative, or do nothing otherwise.
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 202337
any ranges to the list of ranges for the CU as we don't want to emit
them anyway. This ensures that we will still emit ranges if we have
a compile unit compiled with only line tables and one compiled with
full debug info requested (we'll emit for the one with full debug info).
Update testcase metadata accordingly to continue emitting ranges.
llvm-svn: 202333
and update everything accordingly. This can be used to conditionalize
the amount of output in the backend based on the amount of debug
requested/metadata emission scheme by a front end (e.g. clang).
Paired with a commit to clang.
llvm-svn: 202332
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
The current approach to lower a vsetult is to flip the sign bit of the
operands, swap the operands and then use a (signed) pcmpgt. psubus (unsigned
saturating subtract) can be used to emulate a vsetult more efficiently:
+ case ISD::SETULT: {
+ // If the comparison is against a constant we can turn this into a
+ // setule. With psubus, setule does not require a swap. This is
+ // beneficial because the constant in the register is no longer
+ // destructed as the destination so it can be hoisted out of a loop.
I also enable lowering via psubus in a few other cases where it's clearly
beneficial: setule and setuge if minu/maxu cannot be used.
rdar://problem/14338765
Patch by Adam Nemet <anemet@apple.com>.
llvm-svn: 202301
The aggressive anti-dependency breaker scans instructions, bottom-up, within the
scheduling region in order to find opportunities where register renaming can
be used to break anti-dependencies.
Unfortunately, the aggressive anti-dep breaker was treating a register definition
as defining all of that register's aliases (including super registers). This behavior
is incorrect when the super register is live and there are other definitions of
subregisters of the super register.
For example, given the following sequence:
%CR2EQ<def> = CROR %CR3UN, %CR3UN<kill>
%CR2GT<def> = IMPLICIT_DEF
%X4<def> = MFOCRF8 %CR2
the analysis of the first subregister definition would work as expected:
Anti: %CR2GT<def> = IMPLICIT_DEF
Def Groups: CR2GT=g194->g0(via CR2)
Antidep reg: CR2GT (zero group)
Use Groups:
but the analysis of the second one would not:
Anti: %CR2EQ<def> = CROR %CR3UN, %CR3UN<kill>
Def Groups: CR2EQ=g195
Antidep reg: CR2EQ
Rename Candidates for Group g195: ...
because, when processing the %CR2GT<def>, we'd mark all super registers of
%CR2GT (%CR2 in this case) as defined. As a result, when processing
%CR2EQ<def>, %CR2 no longer appears to be live, and %CR2EQ<def>'s group is not
%unioned with the %CR2 group.
I don't have an in-tree test case for this yet (and even if I did, I don't have
a small one).
llvm-svn: 202294
We should apply fastcc whenever profitable. We can expand this list,
but there are lots of conventions with performance implications that we
don't want to change.
Differential Revision: http://llvm-reviews.chandlerc.com/D2705
llvm-svn: 202293
COFF object files with 0 as string table size are currently rejected. This
prevents us from reading object files written by tools like cvtres that
violate the PECOFF spec and write 0 instead of 4 for the size of an empty
string table.
llvm-svn: 202292
We don't have any test with more than 6 address spaces, so a DenseMap is
probably not the correct answer.
An unsorted array would also be OK, but we have to sort it for printing anyway.
llvm-svn: 202275
Summary:
This should fix the MCJIT unit tests that were broken by r201792 on the MIPS buildbot.
MIPS currently uses the default implementation of sys::getHostCPUName() which
always returns "generic". For now, we will accept "generic" and coerce it to
"mips32" or "mips64" depending on the target architecture like we do for empty
CPU names.
Reviewers: jacksprat, matheusalmeida
Reviewed By: jacksprat
Differential Revision: http://llvm-reviews.chandlerc.com/D2878
llvm-svn: 202253
the default.
Based on the patch by Matt Arsenault, D1764!
I switched one place to use the more direct pointer type to compute the
desired address space, and I reworked the memcpy rewriting section to
reflect significant refactorings that this patch helped inspire.
Thanks to several of the folks who helped review and improve the patch
as well.
llvm-svn: 202247
to work independently for the slice side and the other side.
This allows us to only compute the minimum of the two when we actually
rewrite to a memcpy that needs to take the minimum, and preserve higher
alignment for one side or the other when rewriting to loads and stores.
This fix was inspired by seeing the result of some refactoring that
makes addrspace handling better.
llvm-svn: 202242
target_link_libraries(INTERFACE) doesn't bring inter-target dependencies in add_library,
although final targets have dependencies to whole dependent libraries.
It makes most libraries can be built in parallel.
target_link_libraries(PRIVATE) is used to shaared library.
Each dependent library is linked to the target.so, and its user will not see its grandchildren.
For example,
- libclang.so has sufficient libclang*.a(s).
- c-index-test requires just only libclang.so.
FIXME: lld is tweaked minimally. Adding INTERFACE in each library would be better thing.
llvm-svn: 202241
D1764, which in turn set off the other refactorings to make
'getSliceAlign()' a sensible thing.
There are two possible inputs to the required alignment of a memory
transfer intrinsic: the alignment constraints of the source and the
destination. If we are *only* introducing a (potentially new) offset
onto one side of the transfer, we don't need to consider the alignment
constraints of the other side. Use this to simplify the logic feeding
into alignment computation for unsplit transfers.
Also, hoist the clamp of the magical zero alignment for these intrinsics
to the more customary one alignment early. This lets several other
conditions melt away.
No functionality changed. There is a further improvement this exposes
which *will* change functionality, but that's arriving in a separate
patch.
llvm-svn: 202232
rewriting logic: don't pass custom offsets for the adjusted pointer to
the new alloca.
We always passed NewBeginOffset here. Sometimes we spelled it
BeginOffset, but only when they were in fact equal. Whats worse, the API
is set up so that you can't reasonably call it with anything else -- it
assumes that you're passing it an offset relative to the *original*
alloca that happens to fall within the new one. That's the whole point
of NewBeginOffset, it's the clamped beginning offset.
No functionality changed.
llvm-svn: 202231
alignment of the slice being rewritten, not any arbitrary offset.
Every caller is really just trying to compute the alignment for the
whole slice, never for some arbitrary alignment. They are also just
passing a type when they have one to see if we can skip an explicit
alignment in the IR by using the type's alignment. This makes for a much
simpler interface.
Another refactoring inspired by the addrspace patch for SROA, although
only loosely related.
llvm-svn: 202230
consistency with memcpy rewriting, and fix a latent bug in the alignment
management for memset.
The alignment issue is that getAdjustedAllocaPtr is computing the
*relative* offset into the new alloca, but the alignment isn't being set
to the relative offset, it was using the the absolute offset which is
into the old alloca.
I don't think its possible to write a test case that actually reaches
this code where the resulting alignment would be observably different,
but the intent was clearly to use the relative offset within the new
alloca.
llvm-svn: 202229
rather than passing them as arguments.
While I generally prefer actual arguments, in this case the readability
loss is substantial. By using members we avoid repeatedly calculating
the offsets, and once we're using members it is useful to ensure that
those names *always* refer to the original-alloca-relative new offset
for a rewritten slice.
No functionality changed. Follow-up refactoring, all toward getting the
address space patch merged.
llvm-svn: 202228
slice being rewritten.
We had the same code scattered across most of the visits. Instead,
compute the new offsets and the slice size once when we start to visit
a particular slice, and use the member variables from then on. This
reduces quite a bit of code duplication.
No functionality changed. Refactoring inspired to make it easier to
apply the address space patch to SROA.
llvm-svn: 202227
checking in SROA.
The primary change is to just rely on uge for checking that the offset
is within the allocation size. This removes the explicit checks against
isNegative which were terribly error prone (including the reversed logic
that led to PR18615) and prevented us from supporting stack allocations
larger than half the address space.... Ok, so maybe the latter isn't
*common* but it's a silly restriction to have.
Also, we used to try to support a PHI node which loaded from before the
start of the allocation if any of the loaded bytes were within the
allocation. This doesn't make any sense, we have never really supported
loading or storing *before* the allocation starts. The simplified logic
just doesn't care.
We continue to allow loading past the end of the allocation in part to
support cases where there is a PHI and some loads are larger than others
and the larger ones reach past the end of the allocation. We could solve
this a different and more conservative way, but I'm still somewhat
paranoid about this.
llvm-svn: 202224
Eventually DataLayoutPass should go away, but for now that is the only easy
way to get a DataLayout in some APIs. This patch only changes the ones that
have easy access to a Module.
One interesting issue with sometimes using DataLayoutPass and sometimes
fetching it from the Module is that we have to make sure they are equivalent.
We can get most of the way there by always constructing the pass with a Module.
In fact, the pass could be changed to point to an external DataLayout instead
of owning one to make this stricter.
Unfortunately, the C api passes a DataLayout, so it has to be up to the caller
to make sure the pass and the module are in sync.
llvm-svn: 202204
No tool does this currently, but as everything else in a module we should be
able to change its DataLayout.
Most of the fix is in DataLayout to make sure it can be reset properly.
The test uses Module::setDataLayout since the fact that we mutate a DataLayout
is an implementation detail. The module could hold a OwningPtr<DataLayout> and
the DataLayout itself could be immutable.
Thanks to Philip Reames for pushing me in the right direction.
llvm-svn: 202198
their inputs come from std::stable_sort and they are not total orders.
I'm not a huge fan of this, but the really bad std::stable_sort is right
at the beginning of Reassociate. After we commit to stable-sort based
consistent respect of source order, the downstream sorts shouldn't undo
that unless they have a total order or they are used in an
order-insensitive way. Neither appears to be true for these cases.
I don't have particularly good test cases, but this jumped out by
inspection when looking for output instability in this pass due to
changes in the ordering of std::sort.
llvm-svn: 202196