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.
Alexey Bataev [Mon, 7 Aug 2017 15:25:49 +0000 (15:25 +0000)]
[SLP] General improvements of SLP vectorization process.
Patch tries to improve two-pass vectorization analysis, existing in SLP vectorizer. What it does:
1. Defines key nodes, that are the vectorization roots. Previously vectorization started if StoreInst or ReturnInst is found. For now, the vectorization started for all Instructions with no users and void types (Terminators, StoreInst) + CallInsts.
2. CmpInsts, InsertElementInsts and InsertValueInsts are stored in the
array. This array is processed only after the vectorization of the
first-after-these instructions key node is finished. Vectorization goes
in reverse order to try to vectorize as much code as possible.
Matt Arsenault [Mon, 7 Aug 2017 14:58:04 +0000 (14:58 +0000)]
AMDGPU: Cleanup subtarget features
Try to avoid mutually exclusive features. Don't use
a real default GPU, and use a fake "generic". The goal
is to make it easier to see which set of features are
incompatible between feature strings.
Most of the test changes are due to random scheduling changes
from not having a default fullspeed model.
Nirav Dave [Mon, 7 Aug 2017 14:07:49 +0000 (14:07 +0000)]
[DAG] Extend visitSCALAR_TO_VECTOR optimization to truncated vector.
Relanding after case to insert explicit truncation as necessary.
Allow SCALAR_TO_VECTOR of EXTRACT_VECTOR_ELT to reduce to
EXTRACT_SUBVECTOR of vector shuffle when output is smaller. Marginally
improves vector shuffle computations.
Alexey Bataev [Mon, 7 Aug 2017 14:03:17 +0000 (14:03 +0000)]
[SLP] General improvements of SLP vectorization process.
Summary:
Patch tries to improve two-pass vectorization analysis, existing in SLP vectorizer. What it does:
1. Defines key nodes, that are the vectorization roots. Previously vectorization started if StoreInst or ReturnInst is found. For now, the vectorization started for all Instructions with no users and void types (Terminators, StoreInst) + CallInsts.
2. CmpInsts, InsertElementInsts and InsertValueInsts are stored in the array. This array is processed only after the vectorization of the first-after-these instructions key node is finished. Vectorization goes in reverse order to try to vectorize as much code as possible.
Nirav Dave [Mon, 7 Aug 2017 13:55:27 +0000 (13:55 +0000)]
[TableGen] AsmMatcher: fix OpIdx computation when HasOptionalOperands is true
Relanding after fixing UB issue with DefaultOffsets.
Consider the following instruction: "inst.eq $dst, $src" where ".eq"
is an optional flag operand. The $src and $dst operands are
registers. If we parse the instruction "inst r0, r1", the flag is not
present and it will be marked in the "OptionalOperandsMask" variable.
After the matching is complete we call the "convertToMCInst" method.
The current implementation works only if the optional operands are at
the end of the array. The "Operands" array looks like [token:"inst",
reg:r0, reg:r1]. The first operand that must be added to the MCInst
is the destination, the r0 register. The "OpIdx" (in the Operands
array) for this register is 2. However, since the flag is not present
in the Operands, the actual index for r0 should be 1. The flag is not
present since we rely on the default value.
This patch removes the "NumDefaults" variable and replaces it with an
array (DefaultsOffset). This array contains an index for each operand
(excluding the mnemonic). At each index, the array contains the
number of optional operands that should be subtracted. For the
previous example, this array looks like this: [0, 1, 1]. When we need
to access the r0 register, we compute its index as 2 -
DefaultsOffset[1] = 1.
[X86][LLVM]Expanding Supports lowerInterleavedStore() in X86InterleavedAccess (VF16 stride 4).
This patch expands the support of lowerInterleavedStore to 16x8i stride 4.
LLVM creates suboptimal shuffle code-gen for AVX2. In overall, this patch is a specific fix for the pattern (Strid=4 VF=16) and we plan to include more patterns in the future.
The patch goal is to optimize the following sequence:
At the end of the computation, we have ymm2, ymm0, ymm12 and ymm3 holding
each 16 chars:
Guy Blank [Mon, 7 Aug 2017 05:51:14 +0000 (05:51 +0000)]
[SelectionDAG] reset NewNodesMustHaveLegalTypes flag between basic blocks
The NewNodesMustHaveLegalTypes flag is set to false at the beginning of CodeGenAndEmitDAG, and set to true after legalizing types.
But before calling CodeGenAndEmitDAG we build the DAG for the basic block.
So for the first basic block NewNodesMustHaveLegalTypes would be 'false' during the SDAG building, and for all other basic blocks it would be 'true'.
This patch sets the flag to false before SDAG building each basic block.
Sanjay Patel [Sun, 6 Aug 2017 16:27:07 +0000 (16:27 +0000)]
[x86] use more shift or LEA for select-of-constants
We can convert any select-of-constants to math ops:
http://rise4fun.com/Alive/d7d
For this patch, I'm enhancing an existing x86 transform that uses fake multiplies
(they always become shl/lea) to avoid cmov or branching. The current code misses
cases where we have a negative constant and a positive constant, so this is just
trying to plug that hole.
The DAGCombiner diff prevents us from hitting a terrible inefficiency: we can start
with a select in IR, create a select DAG node, convert it into a sext, convert it
back into a select, and then lower it to sext machine code.
Some notes about the test diffs:
1. 2010-08-04-MaskedSignedCompare.ll - We were creating control flow that didn't exist in the IR.
2. memcmp.ll - Choose -1 or 1 is the case that got me looking at this again. I
think we could avoid the push/pop in some cases if we used 'movzbl %al' instead of an xor on
a different reg? That's a post-DAG problem though.
3. mul-constant-result.ll - The trade-off between sbb+not vs. setne+neg could be addressed if
that's a regression, but I think those would always be nearly equivalent.
4. pr22338.ll and sext-i1.ll - These tests have undef operands, so I don't think we actually care about these diffs.
5. sbb.ll - This shows a win for what I think is a common case: choose -1 or 0.
6. select.ll - There's another borderline case here: cmp+sbb+or vs. test+set+lea? Also, sbb+not vs. setae+neg shows up again.
7. select_const.ll - These are motivating cases for the enhancement; replace cmov with cheaper ops.
Assembly differences between movzbl and xor to avoid a partial reg stall are caused later by the X86 Fixup SetCC pass.
Meador Inge [Sun, 6 Aug 2017 12:02:17 +0000 (12:02 +0000)]
[AVR] Compute code model if one is not provided
The patch from r310028 fixed things to work with the new
`LLVMTargetMachine` constructor that came in on r309911.
However, the fix was partial since an object of type
`CodeModel::Model` must be passed to `LLVMTargetMachine`
(not one of `Optional<CodeModel::Model>`).
This patch fixes the problem in the same fashion that r309911
did for other machines: by checking if the passed optional
code model has a value and using `CodeModel::Small` if not.
Craig Topper [Sat, 5 Aug 2017 23:34:44 +0000 (23:34 +0000)]
[X86] Enable isel to use the PAUSE instruction even when SSE2 is disabled
Summary:
On older processors this instruction encoding is treated as a NOP.
MSVC doesn't disable intrinsics based on features the way clang/gcc does. Because the PAUSE instruction encoding doesn't crash older processors, some software out there uses these intrinsics without checking for SSE2.
This change also seems to also be consistent with gcc behavior.
[ADT] Add a much simpler loop to DenseMap::clear when the types are
POD-like and we can just splat the empty key across memory.
Sadly we can't optimize the normal loop well enough because we can't
turn the conditional store into an unconditional store according to the
memory model.
This loop actually showed up in a profile of code that was calling clear
as a serious source of time. =[
Craig Topper [Sat, 5 Aug 2017 20:00:41 +0000 (20:00 +0000)]
[InstCombine] Support vector splats in foldSelectICmpAnd.
Unfortunately, it looks like there's some other missed optimizations in the generated code for some of these cases. I'll try to look at some of those next.
Florian Hahn [Sat, 5 Aug 2017 12:13:13 +0000 (12:13 +0000)]
[ARM] Add registers to debuginfo MIR test cases.
Summary:
MIRParserImpl::computeFunctionProperties uses MRI.getNumVirtRegs() to
set the NoVReg property. By adding a bunch of registers to the MIR test
cases, the NoVReg property is not set when importing the MIR. Otherwise
NoVReg is set after instruction selection while the machine instructions
still contain virtual registers, causing expensive checks to fail.
[LCG] Completely remove the parent set and leaf tracking for RefSCCs.
After the previous series of patches, this is now trivial and deletes
a pretty astonishing amount of complexity. This has been a long time
coming, as the move toward a PO sequence of RefSCCs started eroding the
underlying use cases for this half of the data structure.
Among the biggest advantages here is that now there aren't two
independent data structures that need to stay in sync.
Some of my profiling has also indicated that updating the parent sets
was among the most expensive parts of the lazy call graph. Eliminating
it whole sale is likely to be a nice win in terms of compile time.
Last but not least, I had discussed with some folks previously keeping
it around for asserts and other correctness checking, but once the
fundamentals of the parent and child checking were implemented without
the parent sets their value in correctness checking was tiny and no
where near worth the cost of the complexity required to keep everything
up-to-date.
[LCG] Re-implement the basic isParentOf, isAncestorOf, isChildOf, and
isDescendantOf methods on RefSCCs in terms of the forward edges rather
than the parent sets.
This is technically slower, but probably not interestingly slower, and
all of these routines were already so expensive that they're guarded
behind both !NDEBUG and EXPENSIVE_CHECKS.
This removes another non-critical usage of parent sets.
I've also added some comments to try and help clarify to any potential
users the costs of these routines. They're mostly useful for debugging,
asserts, or other queries.
[LCG] Add the concept of a "dead" node and use it to avoid a complex
walk over the parent set.
When removing a single function from the call graph, we previously would
walk the entire RefSCC's parent set and then walk every outgoing edge
just to find the ones to remove. In addition to this being quite high
complexity in theory, it is also the last fundamental use of the parent
sets.
With this change, when we remove a function we transform the node
containing it to be recognizably "dead" and then teach the edge
iterators to recognize edges to such nodes and skip them the same way
they skip null edges.
We can't move fully to using "dead" nodes -- when disconnecting two live
nodes we need to null out the edge. But the complexity this adds to the
edge sequence isn't too bad and the simplification of lazily handling
this seems like a significant win.
[LCG] Replace an implicit bool operator with a named function. (NFC)
The definition of 'false' here was already pretty vague and debatable,
and I'm about to add another potential 'false' that would actually make
much more sense in a bool operator. Especially given how rarely this is
used, a nicely named method seems better.
[LCG] When removing a dead function and clearing out the data
structures, actually null out the graph pointers as well. We won't ever
update these, and we certainly shouldn't be calling any methods on them,
so it seems good to defensively nuke them.
[LCG] Remove the use of the parent sets to compute connectivity when
merging RefSCCs.
The logic to directly use the reference edges is simpler and not
substantially slower (despite the comments to the contrary) because this
is not actually an especially hot part of LCG in practice.
projects / llvm / log