Daniel Berlin [Fri, 19 May 2017 20:22:20 +0000 (20:22 +0000)]
NewGVN: Fix PR32838.
This is a complicated bug involving two issues:
1. What do we do with phi nodes when we prove all arguments are not
live?
2. When is it safe to use value leaders to determine if we can ignore
an argumnet?
Zachary Turner [Fri, 19 May 2017 19:26:58 +0000 (19:26 +0000)]
Resubmit "[CodeView] Provide a common interface for type collections."
This was originally reverted because it was a breaking a bunch
of bots and the breakage was not surfacing on Windows. After much
head-scratching this was ultimately traced back to a bug in the
lit test runner related to its pipe handling. Now that the bug
in lit is fixed, Windows correctly reports these test failures,
and as such I have finally (hopefully) fixed all of them in this
patch.
Daniel Berlin [Fri, 19 May 2017 19:01:27 +0000 (19:01 +0000)]
Last of the major pieces to NewGVN - yay!
Summary:
NewGVN: Handle equivalence between phi of ops and op of phis.
This makes our GVN mostly-complete. It would be complete, modulo some
deliberate choices we make. This means it detects roughly all herband
equivalences in polynomial time, including cases notoriously hard for
other GVN's to detect. It also detects a very large swath of the
cases we currently rely on instcombine to detect that involve folding
upwards through phis.
Amaury Sechet [Fri, 19 May 2017 18:20:44 +0000 (18:20 +0000)]
[DAGCombine] (addcarry 0, 0, X) -> (ext/trunc X)
Summary:
While this makes some case better and some case worse - so it's unclear if it is a worthy combine just by itself - this is a useful canonicalisation.
Zachary Turner [Fri, 19 May 2017 18:12:07 +0000 (18:12 +0000)]
[lit] Take the last error when executing pipelines.
This seems to have been present since the beginning of time,
which is quite surprising. The symptom was this: Suppose you
have a test with a run line that looks like this:
RUN: foo | FileCheck %s
foo prints some output and then due to a bug in the program it
asserts. On Windows this results in the program returning a
negative exit code. But if enough output had been printed
already by the tool so that the FileCheck match would succeed
then FileCheck would return 0, and because of bad logic in
lit this 0 return value would overwrite the failed return
value from previous items in the pipeline. This only happened
with negative exit codes.
The most sensible behavior is to just take whatever the first
exit code is. There is no logical ordering defined on exit
codes, so comparing with < and > does not make a lot of sense.
Instead, as soon as we find the first non-successful return
value, that should be the result of the entire expression.
This fixes the issue, as now tests which fail on non-Windows
platforms also fail for me on Windows as well.
Anna Thomas [Fri, 19 May 2017 17:05:36 +0000 (17:05 +0000)]
[LoopIdiom] Refactor return value of isLegalStore [NFC]
Summary:
This NFC simply refactors the return value of LoopIdiomRecognize::isLegalStore() from bool to an enumeration, and
removes the return-through-parameter mechanism that the function was using. This function is constructed such that it will
only ever recognize a single store idiom (memset, memset_pattern, or memcpy), and never a combination of these. As such it
makes much more sense for the return value to be the single idiom that the store matches, rather than
having a separate argument-return for each idiom -- it's cleaner, and makes it clearer that
only a single idiom can be matched.
Craig Topper [Fri, 19 May 2017 16:43:54 +0000 (16:43 +0000)]
[APInt] Add support for dividing or remainder by a uint64_t or int64_t.
Summary:
This patch adds udiv/sdiv/urem/srem/udivrem/sdivrem methods that can divide by a uint64_t. This makes division consistent with all the other arithmetic operations.
This modifies the interface of the divide helper method to work on raw arrays instead of APInts. This way we can pass the uint64_t in for the RHS without wrapping it in an APInt. This required moving all the Quotient and Remainder allocation handling up to the callers. For udiv/urem this was as simple as just creating the Quotient/Remainder with the right size when they were declared. For udivrem we have to rely on reallocate not changing the contents of the variable LHS or RHS is aliased with the Quotient or Remainder APInts. We also have to zero the upper bits of Remainder and Quotient that divide doesn't write to if lhsWords/rhsWords is smaller than the width.
I've update the toString method to use the new udivrem.
Guy Blank [Fri, 19 May 2017 12:35:15 +0000 (12:35 +0000)]
[X86][AVX512] Make i1 illegal in the CodeGen
This patch defines the i1 type as illegal in the X86 backend for AVX512.
For DAG operations on <N x i1> types (build vector, extract vector element, ...) i8 is used, and should be truncated/extended.
This should produce better scalar code for i1 types since GPRs will be used instead of mask registers.
Zachary Turner [Fri, 19 May 2017 05:57:45 +0000 (05:57 +0000)]
Revert "[CodeView] Provide a common interface for type collections."
This is a squash of ~5 reverts of, well, pretty much everything
I did today. Something is seriously broken with lit on Windows
right now, and as a result assertions that fire in tests are
triggering failures. I've been breaking non-Windows bots all
day which has seriously confused me because all my tests have
been passing, and after running lit with -a to view the output
even on successful runs, I find out that the tool is crashing
and yet lit is still reporting it as a success!
At this point I don't even know where to start, so rather than
leave the tree broken for who knows how long, I will get this
back to green, and then once lit is fixed on Windows, hopefully
hopefully fix the remaining set of problems for real.
Zachary Turner [Fri, 19 May 2017 04:56:48 +0000 (04:56 +0000)]
[CodeView] Reduce memory usage in TypeSerializer.
We were using a BumpPtrAllocator to allocate stable storage for
a record, then trying to insert that into a hash table. If a
collision occurred, the bytes were never inserted and the
allocation was unnecessary. At the cost of an extra hash
computation, check first if it exists, and only if it does do
we allocate and insert.
Zachary Turner [Fri, 19 May 2017 03:04:08 +0000 (03:04 +0000)]
Fix a broken test.
Similar to my previous fix, it turns out llvm-pdbdump has been
printing an incorrect value since the beginning of time, but
we didn't know it was incorrect. Specifically, we were interpreting
a TypeIndex as referencing a type from the TPI stream when it
actually should come from the IPI stream. So we were printing a
string that looked like a valid string, but was just from the
wrong place.
Zachary Turner [Fri, 19 May 2017 00:56:39 +0000 (00:56 +0000)]
Fix crasher in CodeView test.
Apparently this was always broken, but previously we were more
graceful about it and we would print "unknown udt" if we couldn't
find the type index, whereas now we just segfault because we
assume it's valid. But this exposed a real bug, which is that
we weren't looking in the right place. So fix that, and also
fix this crash at the same time.
Matthias Braun [Fri, 19 May 2017 00:18:03 +0000 (00:18 +0000)]
LiveIntervalAnalysis: Fix missing case in pruneSubRegValues()
pruneSubRegValues() needs to remove subregister ranges starting at
instructions that later get removed by eraseInstrs(). It missed to check
one case in which eraseInstrs() would remove an instruction.
Davide Italiano [Thu, 18 May 2017 23:22:44 +0000 (23:22 +0000)]
[NewGVN] Break infinite recursion in singleReachablePHIPath().
We can have cycles between PHIs and this causes singleReachablePhi()
to call itself indefintely (until we run out of stack). The proper
solution would be that of computing SCCs, but it's not worth for
now, so just keep a visited set and give up when we find a cycle.
Thanks to Dan for the discussion/help with this.
Zachary Turner [Thu, 18 May 2017 23:03:41 +0000 (23:03 +0000)]
[llvm-pdbdump] Add the ability to merge PDBs.
Merging PDBs is a feature that will be used heavily by
the linker. The functionality already exists but does not
have deep test coverage because it's not easily exposed through
any tools. This patch aims to address that by adding the
ability to merge PDBs via llvm-pdbdump. It takes arbitrarily
many PDBs and outputs a single PDB.
Using this new functionality, a test is added for merging
type records. Future patches will add the ability to merge
symbol records, module information, etc.
Zachary Turner [Thu, 18 May 2017 23:03:06 +0000 (23:03 +0000)]
[CodeView] Provide a common interface for type collections.
Right now we have multiple notions of things that represent collections of
types. Most commonly used are TypeDatabase, which is supposed to keep
mappings from TypeIndex to type name when reading a type stream, which
happens when reading PDBs. And also TypeTableBuilder, which is used to
build up a collection of types dynamically which we will later serialize
(i.e. when writing PDBs).
But often you just want to do some operation on a collection of types, and
you may want to do the same operation on any kind of collection. For
example, you might want to merge two TypeTableBuilders or you might want
to merge two type streams that you loaded from various files.
This dichotomy between reading and writing is responsible for a lot of the
existing code duplication and overlapping responsibilities in the existing
CodeView library classes. For example, after building up a
TypeTableBuilder with a bunch of type records, if we want to dump it we
have to re-invent a bunch of extra glue because our dumper takes a
TypeDatabase or a CVTypeArray, which are both incompatible with
TypeTableBuilder.
This patch introduces an abstract base class called TypeCollection which
is shared between the various type collection like things. Wherever we
previously stored a TypeDatabase& in some common class, we now store a
TypeCollection&.
The advantage of this is that all the details of how the collection are
implemented, such as lazy deserialization of partial type streams, is
completely transparent and you can just treat any collection of types the
same regardless of where it came from.
Sanjay Patel [Thu, 18 May 2017 20:53:16 +0000 (20:53 +0000)]
[InstCombine] add helper to foldXorOfICmps(); NFCI
Also, fix the old-style capitalization of the related functions
and move them to the 'private' section of the class since they
are just helpers of the visit* functions.
As shown in the post-commit comments for D32143, we are missing
folds for xor-of-icmps.
Reid Kleckner [Thu, 18 May 2017 17:24:10 +0000 (17:24 +0000)]
[IR] De-virtualize ~Value to save a vptr
Summary:
Implements PR889
Removing the virtual table pointer from Value saves 1% of RSS when doing
LTO of llc on Linux. The impact on time was positive, but too noisy to
conclusively say that performance improved. Here is a link to the
spreadsheet with the original data:
This change makes it invalid to directly delete a Value, User, or
Instruction pointer. Instead, such code can be rewritten to a null check
and a call Value::deleteValue(). Value objects tend to have their
lifetimes managed through iplist, so for the most part, this isn't a big
deal. However, there are some places where LLVM deletes values, and
those places had to be migrated to deleteValue. I have also created
llvm::unique_value, which has a custom deleter, so it can be used in
place of std::unique_ptr<Value>.
I had to add the "DerivedUser" Deleter escape hatch for MemorySSA, which
derives from User outside of lib/IR. Code in IR cannot include MemorySSA
headers or call the MemoryAccess object destructors without introducing
a circular dependency, so we need some level of indirection.
Unfortunately, no class derived from User may have any virtual methods,
because adding a virtual method would break User::getHungOffOperands(),
which assumes that it can find the use list immediately prior to the
User object. I've added a static_assert to the appropriate OperandTraits
templates to help people avoid this trap.
Wei Mi [Thu, 18 May 2017 17:21:22 +0000 (17:21 +0000)]
[LSR] Call canonicalize after we generate a new Formula in GenerateTruncates. Fix PR33077.
The testcase in PR33077 generates a LSR Use Formula with two SCEVAddRecExprs for the same
loop. Such uncommon formula will become non-canonical after GenerateTruncates adds sign
extension to the ScaledReg of the Formula, and it will break the assertion that every
Formula to be inserted is canonical.
The fix is to call canonicalize for the raw Formula generated by GenerateTruncates
before inserting it.
This provides a new way to access the TargetMachine through
TargetPassConfig, as a dependency.
The patterns replaced here are:
* Passes handling a null TargetMachine call
`getAnalysisIfAvailable<TargetPassConfig>`.
* Passes not handling a null TargetMachine
`addRequired<TargetPassConfig>` and call
`getAnalysis<TargetPassConfig>`.
* MachineFunctionPasses now use MF.getTarget().
* Remove all the TargetMachine constructors.
* Remove INITIALIZE_TM_PASS.
This fixes a crash when running `llc -start-before prologepilog`.
PEI needs StackProtector, which gets constructed without a TargetMachine
by the pass manager. The StackProtector pass doesn't handle the case
where there is no TargetMachine, so it segfaults.
Zachary Turner [Thu, 18 May 2017 15:14:44 +0000 (15:14 +0000)]
Fix some minor issues in PDB parsing library.
1) Until now I'd never seen a valid PDB where the DBI stream and
the PDB Stream disagreed on the "Age" field. Because of that,
we had code to assert that they matched. Recently though I was
given a PDB where they disagreed, so this assumption has proven
to be incorrect. Remove this check.
2) We were walking the entire list of hash values for types up front
and then throwing away the values. For large PDBs this was a
significant slow down. Remove this.
With this patch, I can dump the list of all compilands from a
1.5GB PDB file in just a few seconds.
Anna Thomas [Thu, 18 May 2017 13:12:18 +0000 (13:12 +0000)]
[JumpThreading] Dont RAUW condition incorrectly
Summary:
We have a bug when RAUWing the condition if experimental.guard or assumes is a use of that
condition. This is because LazyValueInfo may have used the guards/assumes to identify the
value of the condition at the end of the block. RAUW replaces the uses
at the guard/assume as well as uses before the guard/assume. Both of
these are incorrect.
For now, disable RAUW for conditions and fix the logic as a next
step: https://reviews.llvm.org/D33257
Sam Kolton [Thu, 18 May 2017 12:12:03 +0000 (12:12 +0000)]
[AMDGPU] SDWA operands should not intersect with potential MIs
Summary:
There should be no intesection between SDWA operands and potential MIs. E.g.:
```
v_and_b32 v0, 0xff, v1 -> src:v1 sel:BYTE_0
v_and_b32 v2, 0xff, v0 -> src:v0 sel:BYTE_0
v_add_u32 v3, v4, v2
```
In that example it is possible that we would fold 2nd instruction into 3rd (v_add_u32_sdwa) and then try to fold 1st instruction into 2nd (that was already destroyed). So if SDWAOperand is also a potential MI then do not apply it.
The previous commit failed on test-suite/Bitcode/simd_ops/AArch64_halide_runtime.bc
because isImmOperandEqual() assumed MO was a register operand and that's not
always true.
Lama Saba [Thu, 18 May 2017 08:11:50 +0000 (08:11 +0000)]
[X86] Replace slow LEA instructions in X86
According to Intel's Optimization Reference Manual for SNB+:
" For LEA instructions with three source operands and some specific situations, instruction latency has increased to 3 cycles, and must
dispatch via port 1:
- LEA that has all three source operands: base, index, and offset
- LEA that uses base and index registers where the base is EBP, RBP,or R13
- LEA that uses RIP relative addressing mode
- LEA that uses 16-bit addressing mode "
This patch currently handles the first 2 cases only.
George Rimar [Thu, 18 May 2017 08:00:01 +0000 (08:00 +0000)]
[lib/Object] - Minor API update for llvm::Decompressor.
I revisited Decompressor API (issue with it was triggered during D32865 review)
and found it is probably provides more then we really need.
Issue was about next method's signature:
Error decompress(SmallString<32> &Out);
It is too strict. At first I wanted to change it to decompress(SmallVectorImpl<char> &Out),
but then found it is still not flexible because sticks to SmallVector.
During reviews was suggested to use templating to simplify code. Patch do that.
Serguei Katkov [Thu, 18 May 2017 06:11:56 +0000 (06:11 +0000)]
[BPI] Reduce the probability of unreachable edge to minimal value greater than 0
The probability of edge coming to unreachable block should be as low as possible.
The change reduces the probability to minimal value greater than zero.
The bug https://bugs.llvm.org/show_bug.cgi?id=32214 show the example when
the probability of edge coming to unreachable block is greater than for edge
coming to out of the loop and it causes incorrect loop rotation.
Please note that with this change the behavior of unreachable heuristic is a bit different
than others. Specifically, before this change the sum of probabilities
coming to unreachable blocks have the same weight for all branches
(it was just split over all edges of this block coming to unreachable blocks).
With this change it might be slightly different but not to much due to probability of
taken branch to unreachable block is really small.
Akira Hatanaka [Thu, 18 May 2017 03:52:29 +0000 (03:52 +0000)]
[ThinLTO] Do not assert when adding a module with a different but
compatible target triple
Currently, an assertion fails in ThinLTOCodeGenerator::addModule when
the target triple of the module being added doesn't match that of the
one stored in TMBuilder. This patch relaxes the constraint and makes
changes to allow target triples that only differ in their version
numbers on Apple platforms, similarly to what r228999 did.
Craig Topper [Thu, 18 May 2017 00:51:39 +0000 (00:51 +0000)]
[Statistics] Add a method to atomically update a statistic that contains a maximum
Summary:
There are several places in the codebase that try to calculate a maximum value in a Statistic object. We currently do this in one of two ways:
The first version reads from MaxNumFoo one time and uncontionally rwrites to it. The second version possibly reads it twice depending on the result of the first compare. But we have no way of knowing if the value was changed by another thread between the reads and the writes.
This patch adds a method to the Statistic object that can ensure that we only store if our value is the max and the previous max didn't change after we read it. If it changed we'll recheck if our value should still be the max or not and try again.
This spawned from an audit I'm trying to do of all places we uses the implicit conversion to unsigned on the Statistics objects. See my previous thread on llvm-dev https://groups.google.com/forum/#!topic/llvm-dev/yfvxiorKrDQ
Craig Topper [Wed, 17 May 2017 23:22:10 +0000 (23:22 +0000)]
[Statistics] Use Statistic::operator+= instead of adding and assigning separately.
I believe this technically fixes a multithreaded race condition in this code. But my primary concern was as part of looking at removing the ability to treat Statistics like a plain unsigned. There are many weird operations on Statistics in the codebase.
This breaks the GISel bot:
http://lab.llvm.org:8080/green/job/Compiler_Verifiers_GlobalISEL/5163/consoleFull#-134276167849ba4694-19c4-4d7e-bec5-911270d8a58c
Sanjay Patel [Wed, 17 May 2017 22:29:40 +0000 (22:29 +0000)]
[InstCombine] handle icmp i1 X, C early to avoid creating an unknown pattern
The missing optimization for xor-of-icmps still needs to be added, but by
being more efficient (not generating unnecessary logic ops with constants)
we avoid the bug.
See discussion in post-commit comments:
https://reviews.llvm.org/D32143
Sanjay Patel [Wed, 17 May 2017 22:15:07 +0000 (22:15 +0000)]
[InstCombine] move icmp bool canonicalizations to helper; NFC
As noted in the post-commit comments in D32143, we should be
catching the constant operand cases sooner to be more efficient
and less likely to expose a missing fold.
Matt Arsenault [Wed, 17 May 2017 21:56:25 +0000 (21:56 +0000)]
AMDGPU: Start defining a calling convention
Partially implement callee-side for arguments and return values.
byval doesn't work properly, and most likely sret or other on-stack
return values most as well.
Kyle Butt [Wed, 17 May 2017 21:54:41 +0000 (21:54 +0000)]
CodeGen: Power: Add lowering for shifts of v1i128.
When legalizing vector operations on vNi128, they will be split to v1i128
because that is a legal type on ppc64, but then the compiler will crash in
selection dag because it fails to select for these operations. This patch fixes
shift operations. Logical shift right and left shift can be performed in the
vector unit, but algebraic shift right requires being split.
Michael Liao [Wed, 17 May 2017 21:48:00 +0000 (21:48 +0000)]
Fix PR33028
- '-verify-mahcineinstrs' starts to complain allocatable live-in physical
registers on non-entry or non-landing-pad basic blocks.
- Refactor the XBEGIN translation to define EAX on a dedicated fallback code
path due to XABORT. Add a pseudo instruction to define EAX explicitly to
avoid add physical register live-in.
Matt Arsenault [Wed, 17 May 2017 21:23:14 +0000 (21:23 +0000)]
AMDGPU: Expand frame indexes to be relative to scratch wave offset
In order for an arbitrary callee to access an object
in a caller's stack frame, the 32-bit offset used as
the private pointer needs to be relative to the kernel's
scratch wave offset register.
Convert to this by finding the difference from the current
stack frame and scaling by the wavefront size.
projects / llvm / log