Toma Tabacu [Thu, 25 Jun 2015 12:44:38 +0000 (12:44 +0000)]
[mips] [IAS] Refactor the emitDirectiveModuleFP() functions. NFC.
Summary:
Simplify emitDirectiveModuleFP() by having it just print the current information
from MipsABIFlagsSection and doing an updateABIInfo() before such calls.
This prevents us from forgetting to update the STI.FeatureBits,
because updateABIInfo() uses those to update the MipsABIFlagsSection object,
and also makes sure we use the update mechanism from MipsABIFlagsSection.
Ulrich Weigand [Thu, 25 Jun 2015 11:52:36 +0000 (11:52 +0000)]
[SystemZ] Only attempt RxSBG optimization for integer types
As pointed out by Justin Bogner (see r240520), SystemZDAGToDAGISel::Select
currently attempts to convert boolean operations into RxSBG even on some
non-integer types (in particular, vector types). This would not work in
any case, and it happened to trigger undefined behaviour in allOnes.
This patch verifies that we have a (<= 64-bit) integer type before
attempting to perform this optimization.
Toma Tabacu [Thu, 25 Jun 2015 10:56:57 +0000 (10:56 +0000)]
[mips] [IAS] Refactor the emitDirectiveModuleOddSPReg() functions. NFC.
Summary:
We can simplify emitDirectiveModuleOddSPReg() by having it print the current OddSPReg information
from MipsABIFlagsSection and doing an updateABIInfo() before such calls.
This prevents us from forgetting to update the STI.FeatureBits, because updateABIInfo() uses those to update the MipsABIFlagsSection object,
and also makes sure we use the update mechanism from MipsABIFlagsSection.
Toma Tabacu [Thu, 25 Jun 2015 09:52:02 +0000 (09:52 +0000)]
[mips] [IAS] Fix parsing of memory offset expressions with parenthesis depth >1.
Summary:
In an expression such as "(((a+b)+c)+d)", parseParenExpression() would only parse the "a+b)+c", which would result in an error later on in the parser.
This means that we can only parse one level of inner parentheses.
In order to fix this, I added a new function called parseParenExprOfDepth(), which parses a specified number of trailing parenthesis expressions
(except for the outermost parenthesis), and changed MipsAsmParser to use it in parseMemOffset instead of parseParenExpression().
r240214 fixed some UB in IndVarSimplify, and it needed a temporary
`WeakVH` to do it. Add `simplify_type<const WeakVH>` so that this
temporary isn't necessary.
Douglas Katzman [Wed, 24 Jun 2015 22:04:55 +0000 (22:04 +0000)]
[X86] Simplify some stuff in X86DisassemblerDecoder. NFC
- Deciding that insn->sibIndex is SIB_INDEX_NONE does not require another
check beyond the fully decoded bits being equal to 0x4.
The expression insn->sibIndex == SIB_INDEX_sib could not have been true unless
index were 0x4, because SIB_INDEX_sib is merely the range base (SIB_INDEX_EAX)
plus 4. Respectively SIB_INDEX_sib64.
- Don't use a switch statement to perform left-shift.
Akira Hatanaka [Wed, 24 Jun 2015 20:34:35 +0000 (20:34 +0000)]
[If Converter] Convert recursion to iteration.
This commit makes changes to IfConverter::AnalyzeBlock to use iteration instead
of recursion. Previously, this function would get called recursively a large
number of times and eventually segfault when a function with the following CFG
was compiled:
BB0:
if (condition0)
goto BB1
goto BB2
BB1:
goto BB2
BB2:
if (condition1)
goto BB3
goto BB4
BB3:
...
(repeat until BB7488)
Jingyue Wu [Wed, 24 Jun 2015 20:20:16 +0000 (20:20 +0000)]
Add NVPTXPeephole pass to reduce unnecessary address cast
Summary:
This patch first change the register that holds local address for stack
frame to %SPL. Then the new NVPTXPeephole pass will try to scan the
following pattern
Rafael Espindola [Wed, 24 Jun 2015 19:57:32 +0000 (19:57 +0000)]
Make computeSymbolSizes never fail.
On ELF that was already the case since getting the size of a symbol
never fails.
On MachO and COFF we could fail trying to get the section of a symbol. But
we don't really need the section, just the section number to know if two
symbols are in the same section or not.
Alex Lorenz [Wed, 24 Jun 2015 19:56:10 +0000 (19:56 +0000)]
MIR Serialization: Serialize simple MachineRegisterInfo attributes.
This commit serializes the 3 scalar boolean attributes from the
MachineRegisterInfo class: IsSSA, TracksRegLiveness, and
TracksSubRegLiveness. These attributes are serialized as part
of the machine function YAML mapping.
Jingyue Wu [Wed, 24 Jun 2015 19:28:40 +0000 (19:28 +0000)]
[LSR] canonicalize Prod*(1<<C) to Prod<<C
Summary:
Because LSR happens at a late stage where mul of a power of 2 is
typically canonicalized to shl, this canonicalization emits code that
can be better CSE'ed.
Test Plan:
Transforms/LoopStrengthReduce/shl.ll shows how this change makes GVN more
powerful. Fixes some existing tests due to this change.
The old pattern was to call replaceUsesOfWithOnConstant which was overridden by
subclasses. Those could then call replaceUsesOfWithOnConstantImpl on Constant
to handle deleting the current value.
To be consistent with other parts of the code, this has been changed so that we
call the method on Constant, and that dispatches to an Impl on subclasses.
As part of this, it made sense to rename the methods to be more descriptive. The
new name is Constant::handleOperandChange, and it requires that all subclasses of
Constant implement handleOperandChangeImpl, even if they just throw an error if
they shouldn't be called.
Stop taking a `dwarf::Form` in `DIEValue::EmitValue()` and
`DIEValue::SizeOf()`, since they're always passed `DIEValue::getForm()`
anyway. This is just left over from when `DIEValue` didn't know its own
form.
[CaptureTracking] Avoid long compilation time on large basic blocks
CaptureTracking becomes very expensive in large basic blocks while
calling PointerMayBeCaptured. PointerMayBeCaptured scans the BB the
number of times equal to the number of uses of 'BeforeHere', which is
currently capped at 20 and bails out with Tracker->tooManyUses().
The bottleneck here is the number of calls to PointerMayBeCaptured * the
basic block scan. In a testcase with a 82k instruction BB,
PointerMayBeCaptured is called 130k times, leading to 'shouldExplore'
taking 527k runs, this currently takes ~12min.
To fix this we locally (within PointerMayBeCaptured) number the
instructions in the basic block using a DenseMap to cache instruction
positions/numbers. We build the cache incrementally every time we need
to scan an unexplored part of the BB, improving compile time to only
take ~2min.
This triggers in the flow: DeadStoreElimination -> MepDepAnalysis ->
CaptureTracking.
Side note: after multiple runs in the test-suite I've seen no
performance nor compile time regressions, but could note a couple of
compile time improvements:
Performance Improvements - Compile Time Delta Previous Current StdDev
SingleSource/Benchmarks/Misc-C++/bigfib -4.48% 0.8547 0.8164 0.0022
MultiSource/Benchmarks/TSVC/LoopRerolling-dbl/LoopRerolling-dbl -1.47% 1.3912 1.3707 0.0056
Alex Lorenz [Wed, 24 Jun 2015 17:34:58 +0000 (17:34 +0000)]
MIR Serialization: Serialize the null register operands.
This commit serializes the null register machine operands.
It uses the '_' keyword to represent them, but the parser
also allows the '%noreg' named register syntax.
[LoopVectorizer] Fix bailing-out condition for OptForSize case.
With option OptForSize enabled, the Loop Vectorizer is not supposed to
create tail loop. The condition checking that was invalid and was not
matching to the comment above.
Pawel Bylica [Wed, 24 Jun 2015 12:49:42 +0000 (12:49 +0000)]
Fix instruction scheduling live register tracking
Summary:
This patch fixes PR23405 (https://llvm.org/bugs/show_bug.cgi?id=23405).
During a node unscheduling an entry in LiveRegGens can be replaced with a new value. That corrupts the live reg tracking and LiveReg* structure is not cleared as should be during unscheduling. Problematic condition that enforces Gen replacement is `I->getSUnit()->getHeight() < LiveRegGens[I->getReg()]->getHeight()`. This condition should be checked only if LiveRegGen was set in current node unscheduling.
Rafael Espindola [Wed, 24 Jun 2015 10:20:30 +0000 (10:20 +0000)]
Change how symbol sizes are handled in lib/Object.
COFF and MachO only define symbol sizes for common symbols. Reflect that
in the class hierarchy by having a method for common symbols only in the base
and a general one in ELF.
This avoids the need of using a magic value for the size, which had a few
problems
* Most callers didn't check for it.
* The ones that did could not tell the magic value from a file actually having
that value.
Chandler Carruth [Wed, 24 Jun 2015 10:06:29 +0000 (10:06 +0000)]
[ADT] Teach DenseMap to support StringRef keys.
While often you want to use something specialized like StringMap, when
the strings already have persistent storage a normal densemap over them
can be more efficient.
This can't go into StringRef.h because of really obnoxious header chains
from the hashing code to the endian detection code to CPU feature
detection code to StringMap.
Justin Bogner [Wed, 24 Jun 2015 07:03:07 +0000 (07:03 +0000)]
Hexagon: Paper over the undefined behaviour introduced by r238692
This stops shifting a 32-bit value by such absurd amounts as 96 and
120. We do this by dropping a call to the function that was doing this
entirely, which rather surprisingly doesn't break *any* tests.
I've also added an assert in the misbehaving function to prove that
it's no longer being called with completely invalid arguments.
This change looks pretty bogus and we should probably be reverting
r238692 instead, but this is hard to do with the number of follow ups
that have happened since. It can't be any worse than the undefined
behaviour that was happening before though.
Craig Topper [Wed, 24 Jun 2015 06:19:19 +0000 (06:19 +0000)]
[TableGen] Restore the use of the TheInit field in Record to cache the Record's DefInit. I broke this when I fixed memory leaks recently. Remove the DenseMap that mapped Record's to DefInit.
Justin Bogner [Wed, 24 Jun 2015 05:59:19 +0000 (05:59 +0000)]
SystemZ: Rephrase this allOnes calculation to avoid UB
This allOnes function hits undefined behaviour if Count is greater
than 64, but we can avoid that and simplify the calculation by just
saturating if such a value is passed in.
This comes up under ubsan becauseRxSBGOperands is sometimes created
with values that are 128 bits wide. Somebody more familiar with this
code should probably look into whether that's expected, as a 64 bit
mask may or may not be appropriate for such types.
Ahmed Bougacha [Wed, 24 Jun 2015 00:07:16 +0000 (00:07 +0000)]
[X86] Don't generate vbroadcasti128 for v4i64 splats from memory.
We used to erroneously match:
(v4i64 shuffle (v2i64 load), <0,0,0,0>)
Whereas vbroadcasti128 is more like:
(v4i64 shuffle (v2i64 load), <0,1,0,1>)
This problem doesn't exist for vbroadcastf128, which kept matching
the intrinsic after r231182. We should perhaps re-introduce the
intrinsic here as well, but that's a separate issue still being
discussed.
While there, add some proper vbroadcastf128 tests. We don't currently
match those, like for loading vbroadcastsd/ss on AVX (the reg-reg
broadcasts where added in AVX2).
The only caller of this method is Value::replaceAllUsesWith which
explicitly checks that we are not a GlobalValue. So replace the
body with an unreachable to ensure that we never call it.
The unreachable itself is moved to GlobalValue not GlobalVariable
as that is the base class of all the globals we don't want to call
this method on.
Note, this patch is short lived as i'll soon refactor all callers
of this method.
Alex Lorenz [Tue, 23 Jun 2015 22:39:23 +0000 (22:39 +0000)]
MIR Parser: Use correct source locations for machine instruction diagnostics.
This commit translates the source locations for MIParser diagnostics from
the locations in the machine instruction string to the locations in the
MIR file.
Alexey Samsonov [Tue, 23 Jun 2015 21:00:08 +0000 (21:00 +0000)]
Let llvm::ReplaceInstWithInst copy debug location from old to new instruction.
Currently some users of this function do this explicitly, and all the
rest forget to do this.
ThreadSanitizer was one of such users, and had missing debug
locations for calls into TSan runtime handling atomic operations,
eventually leading to poorly symbolized stack traces and malfunctioning
suppressions.
Evgeniy Stepanov [Tue, 23 Jun 2015 20:57:26 +0000 (20:57 +0000)]
Pass 2 more variables to lit tests.
Pass ADB and ADB_SERIAL environment variables to lit tests.
This would allow running Android tests in compiler-rt when
there is more than one device attached to the host.
Mark Heffernan [Tue, 23 Jun 2015 18:26:50 +0000 (18:26 +0000)]
This change fixes three bugs in loop unswitching. This change causes an 81% speed-up on a benchmark that is based on EigenConvolutionKernel2D from Eigen3, where the lack of loop unswitching blocks hoisting of loads out of a nested loop (see bug 23816 for how loop unswitching and load hoisting are related).
Change 1: Unswitching on trivial conditions should always happen regardless of the computed unswitching cost, as really the cost is zero. While there is code to make that happen, the logic that checks the unswitching cost against a threshold was moved to an earlier point (revision 147935) than the point where trivial unswitching is detected, so trivial unswitching is currently blocked by the cost threshold. This change fixes that.
Change 2: Before revision 147935 (from 2012-01-11), the threshold parameter was a per-loop threshold. So an unswitching happened only if the cost of the unswitching was less than the threshold. In an indirect way (and I believe unintentionally), the logic for this since then has been that the threshold is an over-all budget across all loops for all loop unswitching done by a given LoopUnswitch loop pass object. So if an unswitching with cost 100 happens in one function, that in effect reduces the threshold from 100 to 0 for the loops even in another function. This persists for the lifetime of that loop pass object. This makes no difference for most small examples but it is important for large examples. This revision fixes that.
Change 3: The cost is currently calculated as std::min(NumInstructions, 5 * NumBlocks). So a loop with 2 blocks and a million instructions will have an unswitching cost of 10. I changed this to just NumInstructions, as it were before revision 147935, though I'm open to e.g. instead replacing std::min with std::max.
I've tried to make the change minimally invasive while staying with what I think was the original intent of the code.
Submitted on behalf of broune@.
Alex Lorenz [Tue, 23 Jun 2015 18:22:10 +0000 (18:22 +0000)]
ADT: Add a string APSInt constructor.
This commit moves the APSInt initialization code that's used by
the LLLexer class into a new APSInt constructor that constructs
APSInts from strings.
This change is useful for MIR Serialization, as it would allow
the MILexer class to use the same APSInt initialization as
LLexer when parsing immediate machine operands.
Alex Lorenz [Tue, 23 Jun 2015 17:10:10 +0000 (17:10 +0000)]
AsmParser: Extend the API to make the global value and metadata node slot mappings publicly accessible.
This commit creates a new structure called 'SlotMapping' in the AsmParser library.
This structure can be passed into the public parsing APIs from the AsmParser library
in order to extract the data structures that map from slot numbers to unnamed global
values and metadata nodes.
This change is useful for MIR Serialization, as the MIR Parser has to lookup the
unnamed global values and metadata nodes by their slot numbers.
John Brawn [Tue, 23 Jun 2015 16:02:11 +0000 (16:02 +0000)]
[ARM] ARMLoadStoreOpt::UpdateBaseRegUses should stop on def
When UpdateBaseRegUses sees an instruction that defines the base
register it must stop, as the base register value it is updating is no
longer live. Ideally we would already have seen the register be killed
(which is already checked for), but the kill flags may be inaccurate
and we have to account for this.
Benjamin Kramer [Tue, 23 Jun 2015 15:28:10 +0000 (15:28 +0000)]
[Option] Plug a leak when move-assigning an InputArgList.
The class has a non-trivial dtor so we have to clean up before we move
in new members. Remove misleading comment as a default move assignment
operator will never be synthesized for this class.
Toma Tabacu [Tue, 23 Jun 2015 14:39:42 +0000 (14:39 +0000)]
[mips] [IAS] Add partial support for the ULHU pseudo-instruction.
Summary:
This only adds support for ULHU of an immediate address with/without a source register.
It does not include support for ULHU of the address of a symbol.
projects / llvm / log