Sanjay Patel [Wed, 9 Aug 2017 18:37:41 +0000 (18:37 +0000)]
[InstCombine] narrow rotate left/right patterns to eliminate zext/trunc (PR34046)
I couldn't find any smaller folds to help the cases in:
https://bugs.llvm.org/show_bug.cgi?id=34046
after:
rL310141
The truncated rotate-by-variable patterns elude all of the existing transforms because
of multiple uses and knowledge about demanded bits and knownbits that doesn't exist
without the whole pattern. So we need an unfortunately large pattern match. But by
simplifying this pattern in IR, the backend is already able to generate
rolb/rolw/rorb/rorw for x86 using its existing rotate matching logic (although
there is a likely extraneous 'and' of the rotate amount).
Note that rotate-by-constant doesn't have this problem - smaller folds should already
produce the narrow IR ops.
Guy Blank [Wed, 9 Aug 2017 17:21:01 +0000 (17:21 +0000)]
[X86][AVX512] Choose correct registers in vpbroadcastb/w
Fixes the vpbroadcastb/w instructions which use GPRs as source operands, to use the correct registers.
The full GPR should be used, and not the subregister, as it happens before the patch.
Nuno Lopes [Wed, 9 Aug 2017 17:02:18 +0000 (17:02 +0000)]
CFLAA: return MustAlias when pointers p, q are equal, i.e.,
must-alias(p, sz_p, p, sz_q) irrespective of access sizes sz_p, sz_q
As discussed a couple of weeks ago on the ML.
This makes the behavior consistent with that of BasicAA.
AA clients already check the obj size themselves and may not require the
obj size to match exactly the access size (e.g., in case of store forwarding)
Florian Hahn [Wed, 9 Aug 2017 15:39:10 +0000 (15:39 +0000)]
[ARM] Emit error when ARM exec mode is not available.
Summary:
A similar error message has been removed from the ARMTargetMachineBase
constructor in r306939. With this patch, we generate an error message
for the example below, compiled with -mcpu=cortex-m0, which does not
have ARM execution mode.
__attribute__((target("arm"))) int foo(int a, int b)
{
return a + b % a;
}
__attribute__((target("thumb"))) int bar(int a, int b)
{
return a + b % a;
}
By adding this error message to ARMBaseTargetMachine::getSubtargetImpl,
we can deal with functions that set -thumb-mode in target-features.
At the moment it seems like Clang does not have access to target-feature
specific information, so adding the error message to the frontend will
be harder.
Coby Tayree [Wed, 9 Aug 2017 15:34:55 +0000 (15:34 +0000)]
[X86][Asm]Allow far jmp/call to be picked when using explicit FWORD size specifier
Currently, far jmp/call which utilizes a 48bit memory operand would have been invoked via the 'lcall/ljmp' mnemonic (intel style).
This patch align those variants to formal intel spec
Davide Italiano [Wed, 9 Aug 2017 15:13:50 +0000 (15:13 +0000)]
[ValueTracking] Honour recursion limit.
The recently improved support for `icmp` in ValueTracking
(r307304) exposes the fact that `isImplied` condition doesn't
really bail out if we hit the recursion limit (and calls
`computeKnownBits` which increases the depth and asserts).
Sjoerd Meijer [Wed, 9 Aug 2017 14:59:54 +0000 (14:59 +0000)]
[AArch64] Assembler support for the ARMv8.2a dot product instructions
Dot product is an optional ARMv8.2a extension, see also the public architecture
specification here:
https://developer.arm.com/products/architecture/a-profile/exploration-tools.
This patch adds AArch64 assembler support for these dot product instructions.
Florian Hahn [Wed, 9 Aug 2017 13:53:28 +0000 (13:53 +0000)]
[ARM] Remove FeatureNoARM implies ModeThumb.
Summary:
By removing FeatureNoARM implies ModeThumb, we can detect cases where a
function's target-features contain -thumb-mode (enables ARM codegen for the
function), but the architecture does not support ARM mode. Previously, the
implication caused the FeatureNoARM bit to be cleared for functions with
-thumb-mode, making the assertion in ARMSubtarget::ARMSubtarget [1]
pointless for such functions.
This assertion is the only guard against generating ARM code for
architectures without ARM codegen support. Is there a place where we
could easily generate error messages for the user? At the moment, we
would generate ARM code for Thumb-only architectures. X86 has the same
behavior as ARM, as in it only has an assertion and no error message,
but I think for ARM an error message would be helpful. What do you
think?
For the example below, `llc -mtriple=armv7m-eabi test.ll -o -` will
generate ARM assembler (or fail with an assertion error with this patch).
Note that if we run the resulting assembler through llvm-mc, we get
an appropriate error message, but not when codegen is handled
through clang.
formatv_object currently uses the implicitly defined move constructor,
but it is buggy. In typical use-cases, the problem doesn't show-up
because all calls to the move constructor are elided. Thus, the buggy
constructors are never invoked.
The issue especially shows-up when code is compiled using the
-fno-elide-constructors compiler flag. For instance, this is useful when
attempting to collect accurate code coverage statistics.
The exact issue is the following:
The Parameters data member is correctly moved, thus making the
parameters occupy a new memory location in the target
object. Unfortunately, the default copying of the Adapters blindly
copies the vector of pointers, leaving each of these pointers
referencing the parameters in the original object instead of the copied
one. These pointers quickly become dangling when the original object is
deleted. This quickly leads to crashes.
The solution is to update the Adapters pointers when performing a move.
The copy constructor isn't useful for format objects and can thus be
deleted.
isLegalAddressingMode() has recently gained the extra optional Instruction*
parameter, and therefore it can now do the job that previously only
isFoldableMemAccess() could do.
The SystemZ implementation of isLegalAddressingMode() has gained the
functionality of checking for offsets, which used to be done with
isFoldableMemAccess().
The isFoldableMemAccess() hook has been removed everywhere.
[LCG] Completely remove the map-based association of post-order numbers
to Nodes when removing ref edges from a RefSCC.
This map based association turns out to be pretty expensive for large
RefSCCs and pointless as we already have embedded data members inside
nodes that we use to track the DFS state. We can reuse one of those and
the map becomes unnecessary.
This also fuses the update of those numbers into the scan across the
pending stack of nodes so that we don't walk the nodes twice during the
DFS.
With this I expect the new PM to be faster than the old PM for the test
case I have been optimizing. That said, it also seems simpler and more
direct in many ways. The side storage was always pretty awkward.
The last remaining hot-spot in the profile of the LCG once this is done
will be the edge iterator walk in the DFS. I'll take a look at improving
that next.
[LCG] Special case when removing a ref edge from a RefSCC leaves
that RefSCC still connected.
This is common and can be handled much more efficiently. As soon as we
know we've covered every node in the RefSCC with the DFS, we can simply
reset our state and return. This avoids numerous data structure updates
and other complexity.
On top of other changes, this appears to get new PM back to parity with
the old PM for a large protocol buffer message source code. The dense
map updates are very hot in this function.
[LCG] Switch one of the update methods for the LazyCallGraph to support
limited batch updates.
Specifically, allow removing multiple reference edges starting from
a common source node. There are a few constraints that play into
supporting this form of batching:
1) The way updates occur during the CGSCC walk, about the most we can
functionally batch together are those with a common source node. This
also makes the batching simpler to implement, so it seems
a worthwhile restriction.
2) The far and away hottest function for large C++ files I measured
(generated code for protocol buffers) showed a huge amount of time
was spent removing ref edges specifically, so it seems worth focusing
there.
3) The algorithm for removing ref edges is very amenable to this
restricted batching. There are just both API and implementation
special casing for the non-batch case that gets in the way. Once
removed, supporting batches is nearly trivial.
This does modify the API in an interesting way -- now, we only preserve
the target RefSCC when the RefSCC structure is unchanged. In the face of
any splits, we create brand new RefSCC objects. However, all of the
users were OK with it that I could find. Only the unittest needed
interesting updates here.
How much does batching these updates help? I instrumented the compiler
when run over a very large generated source file for a protocol buffer
and found that the majority of updates are intrinsically updating one
function at a time. However, nearly 40% of the total ref edges removed
are removed as part of a batch of removals greater than one, so these
are the cases batching can help with.
When compiling the IR for this file with 'opt' and 'O3', this patch
reduces the total time by 8-9%.
Serguei Katkov [Wed, 9 Aug 2017 05:17:02 +0000 (05:17 +0000)]
[ImplicitNullCheck] Fix the bug when dependent instruction accesses memory
It is possible that dependent instruction may access memory.
In this case we must reject optimization because the memory change will
be visible in null handler basic block. So we will execute an instruction which
we must not execute if check fails.
Zachary Turner [Wed, 9 Aug 2017 04:23:59 +0000 (04:23 +0000)]
[PDB] Fix an issue writing the publics stream.
In the refactor to merge the publics and globals stream, a bug
was introduced that wrote the wrong value for one of the fields
of the PublicsStreamHeader. This caused debugging in WinDbg
to break.
We had no way of dumping any of these fields, so in addition to
fixing the bug I've added dumping support for them along with a
test that verifies the correct value is written.
Zachary Turner [Wed, 9 Aug 2017 04:23:25 +0000 (04:23 +0000)]
[PDB] Merge Global and Publics Builders.
The publics stream and globals stream are very similar. They both
contain a list of hash buckets that refer into a single shared stream,
the symbol record stream. Because of the need for each builder to manage
both an independent hash stream as well as a single shared record
stream, making the two builders be independent entities is not the right
design. This patch merges them into a single class, of which only a
single instance is needed to create all 3 streams. PublicsStreamBuilder
and GlobalsStreamBuilder are now merged into the single GSIStreamBuilder
class, which writes all 3 streams at once.
Note that this patch does not contain any functionality change. So we're
still not yet writing any records to the globals stream. All we're doing
is making it so that when we do start writing records to the globals,
this refactor won't have to be part of that patch.
Revert "[GlobalISel] Remove the GISelAccessor API."
This reverts commit r310115.
It causes a linker failure for the one of the unittests of AArch64 on one
of the linux bot:
http://lab.llvm.org:8011/builders/clang-ppc64le-linux-multistage/builds/3429
[MachineOutliner] Ensure AArch64 outliner doesn't mess with W30 or LR
Before, the outliner would mark all instructions that read from/modify LR as
illegal. This doesn't handle W30, which overlaps with LR. This shouldn't be
outlined.
This commit fixes that by making modifiesRegister() and readsRegister() look at
W30 + take in a TRI argument. This makes sure that modifiesRegister() and
readsRegister() won't outline either of W30 and LR.
Wei Mi [Tue, 8 Aug 2017 21:40:14 +0000 (21:40 +0000)]
[GVN] Remove stale entries in phitranslate cache when new phi is generated for PRE
When a new phi is generated for scalarpre of an expression, the phiTranslate cache
will become stale: Before PRE, the candidate expression must not be available in a
predecessor block, and phitranslate will cache the information. After PRE, the
expression will become available in all predecessor blocks, so the related entries
in phiTranslate cache becomes stale. The patch will simply remove the stale entries
so phiTranslate can be recomputed next time.
The stale entries in phitranslate cache will not affect correctness but will cause
missing PRE opportunity for later instructions.
Dehao Chen [Tue, 8 Aug 2017 20:57:33 +0000 (20:57 +0000)]
Make ICP uses PSI to check for hotness.
Summary: Currently, ICP checks the count against a fixed value to see if it is hot enough to be promoted. This does not work for SamplePGO because sampled count may be much smaller. This patch uses PSI to check if the count is hot enough to be promoted.
Connor Abbott [Tue, 8 Aug 2017 18:52:22 +0000 (18:52 +0000)]
[AMDGPU] Add llvm.amdgpu.update.dpp intrinsic
Summary:
Now that we've made all the necessary backend changes, we can add a new
intrinsic which exposes the new capabilities to IR producers. Since
llvm.amdgpu.update.dpp is a strict superset of llvm.amdgpu.mov.dpp, we
should deprecate the former. We also add tests for all the functionality
that was added in previous changes, now that we can access it via an IR
construct.
Zachary Turner [Tue, 8 Aug 2017 18:34:44 +0000 (18:34 +0000)]
[PDB] Fix linking of function symbols and local variables.
The compiler outputs PROC32_ID symbols into the object files
for functions, and these symbols have an embedded type index
which, when copied to the PDB, refer to the IPI stream. However,
the symbols themselves are also converted into regular symbols
(e.g. S_GPROC32_ID -> S_GPROC32), and type indices in the regular
symbol records refer to the TPI stream. So this patch applies
two fixes to function records.
1. It converts ID symbols to the proper non-ID record type.
2. After remapping the type index from the object file's index
space to the PDB file/IPI stream's index space, it then
remaps that index to the TPI stream's index space by.
Besides functions, during the remapping process we were also
discarding symbol record types which we did not recognize.
In particular, we were discarding S_BPREL32 records, which is
what MSVC uses to describe local variables on the stack. So
this patch fixes that as well by copying them to the PDB.
Adrian Prantl [Tue, 8 Aug 2017 18:26:12 +0000 (18:26 +0000)]
dsymutil: support dwarf version mismatches between object and clang module
This adds a missing call to maybeUpdateMaxDwarfVersion when visitng a
clang module. Failing to do so will cause a failure when emitting
DWARF 4 forms into a CU that AsmPrinter believes to be DWARF 2.
Anna Thomas [Tue, 8 Aug 2017 18:07:44 +0000 (18:07 +0000)]
[LoopVectorize] Fix assertion failure in Fcmp vectorization
Summary:
When vectorizing fcmps we can trip on incorrect cast assertion when setting the
FastMathFlags after generating the vectorized FCmp.
This can happen if the FCmp can be folded to true or false directly. The fix
here is to set the FastMathFlag using the FastMathFlagBuilder *before* creating
the FCmp Instruction. This is what's done by other optimizations such as
InstCombine.
Added a test case which trips on cast assertion without this patch.
Tim Northover [Tue, 8 Aug 2017 17:16:46 +0000 (17:16 +0000)]
Revert "[ARM] Fix assembly and disassembly for VMRS/VMSR"
This reverts r310243. Only MVFR2 is actually restricted to v8 and it'll be a
little while before we can get a proper fix together. Better that we allow
incorrect code than reject correct in the meantime.
[PowerPC] Don't crash on larger splats achieved through 1-byte splats
We've implemented a 1-byte splat using XXSPLTISB on P9. However, LLVM will
produce a 1-byte splat even for wider element BUILD_VECTOR nodes. This patch
prevents crashing in that situation.
Amjad Aboud [Tue, 8 Aug 2017 12:17:56 +0000 (12:17 +0000)]
[X86] Improved X86::CMOV to Branch heuristic.
Resolved PR33954.
This patch contains two more constraints that aim to reduce the noise cases where we convert CMOV into branch for small gain, and end up spending more cycles due to overhead.
Daniel Sanders [Tue, 8 Aug 2017 10:44:31 +0000 (10:44 +0000)]
[globalisel][tablegen] Add support for importing 'imm' operands.
Summary:
This patch enables the import of rules containing 'imm' operands that do not
constrain the acceptable values using predicates. Support for ImmLeaf will
arrive in a later patch.
[PM] Fix a likely more critical infloop bug in the CGSCC pass manager.
This was just a bad oversight on my part. The code in question should
never have worked without this fix. But it turns out, there are
relatively few places that involve libfunctions that participate in
a single SCC, and unless they do, this happens to not matter.
The effect of not having this correct is that each time through this
routine, the edge from write_wrapper to write was toggled between a call
edge and a ref edge. First time through, it becomes a demoted call edge
and is turned into a ref edge. Next time it is a promoted call edge from
a ref edge. On, and on it goes forever.
I've added the asserts which should have always been here to catch silly
mistakes like this in the future as well as a test case that will
actually infloop without the fix.
The other (much scarier) infinite-inlining issue I think didn't actually
occur in practice, and I simply misdiagnosed this minor issue as that
much more scary issue. The other issue *is* still a real issue, but I'm
somewhat relieved that so far it hasn't happened in real-world code
yet...
[PM] Fix new LoopUnroll function pass by invalidating loop analysis
results when a loop is completely removed.
This is very hard to manifest as a visible bug. You need to arrange for
there to be a subsequent allocation of a 'Loop' object which gets the
exact same address as the one which the unroll deleted, and you need the
LoopAccessAnalysis results to be significant in the way that they're
stale. And you need a million other things to align.
But when it does, you get a deeply mysterious crash due to actually
finding a stale analysis result. This fixes the issue and tests for it
by directly checking we successfully invalidate things. I have not been
able to get *any* test case to reliably trigger this. Changes to LLVM
itself caused the only test case I ever had to cease to crash.
I've looked pretty extensively at less brittle ways of fixing this and
they are actually very, very hard to do. This is a somewhat strange and
unusual case as we have a pass which is deleting an IR unit, but is not
running within that IR unit's pass framework (which is what handles this
cleanly for the normal loop unroll). And where there isn't a definitive
way to clear *all* of the stale cache entries. And where the pass *is*
updating the core analysis that provides the IR units!
For example, we don't have any of these problems with Function analyses
because it is easy to clear out function analyses when the functions
themselves may have been deleted -- we clear an entire module's worth!
But that is too heavy of a hammer down here in the LoopAnalysisManager
layer.
A better long-term solution IMO is to require that AnalysisManager's
make their keys durable to this kind of thing. Specifically, when
caching an analysis for one IR unit that is conceptually "owned" by
a higher level IR unit, the AnalysisManager should incorporate this into
its data structures so that we can reliably clear these results without
having to teach each and every pass to do so manually as we do here. But
that is a change for another day as it will be a fairly invasive change
to the AnalysisManager infrastructure. Until then, this fortunately
seems to be quite rare.
Reid Kleckner [Mon, 7 Aug 2017 21:23:38 +0000 (21:23 +0000)]
[Object] Initialize LoadConfig member to null
Executables may not contain a load config, and clients should be able to
test for nullability. Previously we'd return uninitialized memory. Now
getLoadConfig32/64 return valid pointers or null.
Dehao Chen [Mon, 7 Aug 2017 20:23:20 +0000 (20:23 +0000)]
Move the SampleProfileLoader right after EarlyFPM.
Summary: SampleProfileLoader pass do need to happen after some early cleanup passes so that inlining can happen correctly inside the SampleProfileLoader pass.
Connor Abbott [Mon, 7 Aug 2017 19:10:56 +0000 (19:10 +0000)]
[AMDGPU] Add pseudo "old" source to all DPP instructions
Summary:
All instructions with the DPP modifier may not write to certain lanes of
the output if bound_ctrl=1 is set or any bits in bank_mask or row_mask
aren't set, so the destination register may be both defined and modified.
The right way to handle this is to add a constraint that the destination
register is the same as one of the inputs. We could tie the destination
to the first source, but that would be too restrictive for some use-cases
where we want the destination to be some other value before the
instruction executes. Instead, add a fake "old" source and tie it to the
destination. Effectively, the "old" source defines what value unwritten
lanes will get. We'll expose this functionality to users with a new
intrinsic later.
Also, we want to use DPP instructions for computing derivatives, which
means we need to set WQM for them. We also need to enable the entire
wavefront when using DPP intrinsics to implement nonuniform subgroup
reductions, since otherwise we'll get incorrect results in some cases.
To accomodate this, add a new operand to all DPP instructions which will
be interpreted by the SI WQM pass. This will be exposed with a new
intrinsic later. We'll also add support for Whole Wavefront Mode later.
I also fixed llvm.amdgcn.mov.dpp to overwrite the source and fixed up
the test. However, I could also keep the old behavior (where lanes that
aren't written are undefined) if people want it.
Matt Arsenault [Mon, 7 Aug 2017 18:12:47 +0000 (18:12 +0000)]
AMDGPU: Remove FixControlFlowLiveIntervals pass
This hasn't done anything in a long time. This was
running after the the control flow pseudos were expanded,
so this would never find them. The control flow pseudo
expansion was moved to solve the problem this pass was
supposed to solve in the first place, except handling
it earlier also fixes it for fast regalloc which doesn't
use LiveIntervals.
Craig Topper [Mon, 7 Aug 2017 18:10:39 +0000 (18:10 +0000)]
[InstCombine] Support (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2) for vector splats
Note the original code I deleted incorrectly listed this as (X | C1) & C2 --> (X & C2^(C1&C2)) | C1 Which is only valid if C1 is a subset of C2. This relied on SimplifyDemandedBits to remove any extra bits from C1 before we got to that code.
My new implementation avoids relying on that behavior so that it can be naively verified with alive.
projects / llvm / log