Joel Jones [Fri, 21 Apr 2017 17:31:03 +0000 (17:31 +0000)]
[AArch64] Refactor instruction selection lowering for addresses. NFCI
Factor out the common code used for generating addresses into common
templated functions that call overloaded versions of a new function,
getTargetNode.
This marks the beginning of an effort to port remaining
MSVC toolchain miscellaneous utilities to all platforms.
Currently clang-cl shells out to certain additional tools
such as the IDL compiler, resource compiler, and a few
other tools, but as these tools are Windows-only it
limits the ability of clang to target Windows on other
platforms. having a full suite of these tools directly
in LLVM should eliminate this constraint.
The current implementation provides no actual functionality,
it is just an empty skeleton executable for the purposes
of making incremental changes.
Differential Revision: https://reviews.llvm.org/D32095
Patch by Eric Beckmann (ecbeckmann@google.com)
Tim Northover [Fri, 21 Apr 2017 17:21:59 +0000 (17:21 +0000)]
ARM: don't try to create an i8 -> i32 vpaddl.
DAG combine was mistakenly assuming that the step-up it was looking at was
always a doubling, but it can sometimes be a larger extension in which case
we'd crash.
Kuba Mracek [Fri, 21 Apr 2017 16:57:37 +0000 (16:57 +0000)]
[libFuzzer] Check for target(popcnt) capability before usage
Older compilers (e.g. LLVM 3.4) do not support the attribute target("popcnt").
In order to support those, this diff check the attribute support using the preprocessor.
[APInt] Add compare/compareSigned methods that return -1, 0, 1. Reimplement slt/ult and friends using them
Currently sle and ule have to call slt/ult and eq to get the proper answer. This results in extra code for both calls and additional scans of multiword APInts.
This patch replaces slt/ult with a compareSigned/compare that can return -1, 0, or 1 so we can cover all the comparison functions with a single call.
While I was there I removed the activeBits calls and other checks at the start of the slow part of ult. Both of the activeBits calls potentially scan through each of the APInts separately. I can't imagine that's any better than just scanning them in parallel and doing the compares. Now we just share the code with tcCompare.
These changes seem to be good for about a 7-8k reduction on the size of the opt binary on my local x86-64 build.
Daniel Sanders [Fri, 21 Apr 2017 15:59:56 +0000 (15:59 +0000)]
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule.
Summary:
The SelectionDAG importer now imports rules with Predicate's attached via
Requires, PredicateControl, etc. These predicates are implemented as
bitset's to allow multiple predicates to be tested together. However,
unlike the MC layer subtarget features, each target only pays for it's own
predicates (e.g. AArch64 doesn't have 192 feature bits just because X86
needs a lot).
Both AArch64 and X86 derive at least one predicate from the MachineFunction
or Function so they must re-initialize AvailableFeatures before each
function. They also declare locals in <Target>InstructionSelector so that
computeAvailableFeatures() can use the code from SelectionDAG without
modification.
Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab
[SimplifyCFG] Fix the determination of PostBB in conditional store merging to handle the targets on the second branch being commuted
Currently we choose PostBB as the single successor of QFB, but its possible that QTB's single successor is QFB which would make QFB the correct choice.
Wei Mi [Fri, 21 Apr 2017 15:50:16 +0000 (15:50 +0000)]
[ConstHoisting] Add BFI in constanthoisting pass and select the best insertion
places based on it.
Existing constant hoisting pass will merge a group of contants in a small range
and hoist the const materialization code to the common dominator of their uses.
However, if the uses are all in cold pathes, existing implementation may hoist
the materialization code from cold pathes to a hot place. This may hurt performance.
The patch introduces BFI to the pass and selects the best insertion places based
on it.
The change is controlled by an option consthoist-with-block-frequency which is
off by default for now.
Daniel Sanders [Fri, 21 Apr 2017 12:51:43 +0000 (12:51 +0000)]
[globalisel][tablegen] Try again to fix builds on old MSVC's after r300964
This should fix llvm-clang-x86_64-expensive-checks-win
I reproduced the error using the following code:
namespace llvm {
// Moving this out of the llvm namespace fixes the error.
template<unsigned NumBits> class PredicateBitsetImpl {};
}
namespace {
const unsigned MAX_SUBTARGET_PREDICATES = 11;
// This works on Clang but is broken on MSVC
// using PredicateBitset = PredicateBitsetImpl<MAX_SUBTARGET_PREDICATES>;
// Some versions emit a syntax error here ("error C2061: syntax error: identifier
// 'PredicateBitsetImpl'") but others accept it and only emit the C3646 below.
//
// This works on Clang and MSVC
using PredicateBitset = llvm::PredicateBitsetImpl<MAX_SUBTARGET_PREDICATES>;
[ARM] GlobalISel: Make struct arguments fail elegantly
The condition in isSupportedType didn't handle struct/array arguments
properly. Fix the check and add a test to make sure we use the fallback
path in this kind of situation. The test deals with some common cases
where the call lowering should error out. There are still some issues
here that need to be addressed (tail calls come to mind), but they can
be addressed in other patches.
Daniel Sanders [Fri, 21 Apr 2017 10:27:20 +0000 (10:27 +0000)]
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule.
Summary:
The SelectionDAG importer now imports rules with Predicate's attached via
Requires, PredicateControl, etc. These predicates are implemented as
bitset's to allow multiple predicates to be tested together. However,
unlike the MC layer subtarget features, each target only pays for it's own
predicates (e.g. AArch64 doesn't have 192 feature bits just because X86
needs a lot).
Both AArch64 and X86 derive at least one predicate from the MachineFunction
or Function so they must re-initialize AvailableFeatures before each
function. They also declare locals in <Target>InstructionSelector so that
computeAvailableFeatures() can use the code from SelectionDAG without
modification.
Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab
X86 memcpy: use REPMOVSB instead of REPMOVS{Q,D,W} for inline copies
when the subtarget has fast strings.
This has two advantages:
- Speed is improved. For example, on Haswell thoughput improvements increase
linearly with size from 256 to 512 bytes, after which they plateau:
(e.g. 1% for 260 bytes, 25% for 400 bytes, 40% for 508 bytes).
- Code is much smaller (no need to handle boundaries).
George Rimar [Fri, 21 Apr 2017 09:12:18 +0000 (09:12 +0000)]
[DWARF] - Refactoring: localize handling of relocations in a single place.
This is splitted from D32228,
currently DWARF parsers code has few places that applied relocations values manually.
These places has similar duplicated code. Patch introduces separate method that can be
used to obtain relocated value. That helps to reduce code and simplifies things.
Davide Italiano [Fri, 21 Apr 2017 04:25:00 +0000 (04:25 +0000)]
[PartialInliner] Fix crash when inlining functions with unreachable blocks.
CodeExtractor looks up the dominator node corresponding to return blocks
when splitting them. If one of these blocks is unreachable, there's no
node in the Dom and CodeExtractor crashes because it doesn't check
for domtree node validity.
In theory, we could add just a check for skipping null DTNodes in
`splitReturnBlock` but the fix I propose here is slightly different. To the
best of my knowledge, unreachable blocks are irrelevant for the algorithm,
therefore we can just skip them when building the candidate set in the
constructor.
[AsmWriter/APFloat] FP constant printing: Avoid usage of locale dependent snprinf
This should fix the bug https://bugs.llvm.org/show_bug.cgi?id=12906
To print the FP constant AsmWriter does the following:
1) convert FP value to String (actually using snprintf function which is locale dependent).
2) Convert String back to FP Value
3) Compare original and got FP values. If they are not equal just dump as hex.
The problem happens on the 2nd step when APFloat does not expect group delimiter or
fraction delimiter other than period symbol and so on, which can be produced on the
first step if LLVM library is used in an environment with corresponding locale set.
To fix this issue the locale independent APFloat:toString function is used.
However it prints FP values slightly differently than snprintf does. Specifically
it suppress trailing zeros in significant, use capital E and so on.
It results in 117 test failures during make check.
To avoid this I've also updated APFloat.toString a bit to pass make check at least.
[AArch64] Improve code generation for logical instructions taking
immediate operands.
This commit adds an AArch64 dag-combine that optimizes code generation
for logical instructions taking immediate operands. The optimization
uses demanded bits to change a logical instruction's immediate operand
so that the immediate can be folded into the immediate field of the
instruction.
This recommits r300913, which broke bots because I didn't fix a call to
ShrinkDemandedConstant in SIISelLowering.cpp after changing the APIs of
TargetLoweringOpt and TargetLowering.
[InstCombine] Remove the zextOrTrunc from ShrinkDemandedConstant.
The demanded mask and the constant should always be the same width for all callers today.
Also stop copying the demanded mask as its passed in. We should avoid allocating memory unless we are going to do something. The final AND to create the new constant will take care of it.
X86RegisterInfo::eliminateFrameIndex() and
X86FrameLowering::getFrameIndexReference() both had logic to compute the
base register. This consolidates the code.
Also use MachineInstr::isReturn instead of manually enumerating tail
call instructions (return instructions were not included in the previous
list because they never reference frame indexes).
X86RegisterInfo: eliminateFrameIndex: Force SP for AfterFPPop; NFC
AfterFPPop is used for tailcall/tailjump instructions. We shouldn't ever
have frame-pointer/base-pointer relative addressing for those. After all
the frame/base pointer should already be restored to their previous
values at the return.
Make this fact explicit in preparation for an upcoming refactoring.
[Simplify] Add testcase to show that merging conditional stores for triangles is sensitive to the order of the branch targets on the conditional branches. NFC
[AArch64] Improve code generation for logical instructions taking
immediate operands.
This commit adds an AArch64 dag-combine that optimizes code generation
for logical instructions taking immediate operands. The optimization
uses demanded bits to change a logical instruction's immediate operand
so that the immediate can be folded into the immediate field of the
instruction.
Tim Northover [Thu, 20 Apr 2017 21:57:45 +0000 (21:57 +0000)]
AArch64: lower "fence singlethread" to a pure compiler barrier.
Single-threaded fences aren't required to provide any synchronization with
other processing elements so there's no need for a DMB. They should still be a
barrier for compiler optimizations though.
Tim Northover [Thu, 20 Apr 2017 21:56:52 +0000 (21:56 +0000)]
ARM: lower "fence singlethread" to a pure compiler barrier.
Single-threaded fences aren't required to provide any synchronization with
other processing elements so there's no need for a DMB. They should still be a
barrier for compiler optimizations though.
Tim Northover [Thu, 20 Apr 2017 19:54:02 +0000 (19:54 +0000)]
ARM: handle post-indexed NEON ops where the offset isn't the access width.
Before, we assumed that any ConstantInt offset was precisely the access width,
so we could use the "[rN]!" form. ISelLowering only ever created that kind, but
further simplification during combining could lead to unexpected constants and
incorrect codegen.
Paul Robinson [Thu, 20 Apr 2017 19:16:51 +0000 (19:16 +0000)]
[DWARF] Versioning for DWARF constants; verify FORMs
Associate the version-when-defined with definitions of standard DWARF
constants. Identify the "vendor" for DWARF extensions.
Use this information to verify FORMs in .debug_abbrev are defined as
of the DWARF version specified in the associated unit.
Removed two tests that had specified DWARF v1 (which essentially does
not exist).
Benjamin Kramer [Thu, 20 Apr 2017 18:29:37 +0000 (18:29 +0000)]
[Recycler] Add asan/msan annotations.
This enables use after free and uninit memory checking for memory
returned by a recycler. SelectionDAG currently relies on the opcode of a
free'd node being ISD::DELETED_NODE, so poke a hole in the asan poison
for SDNode opcodes. This means that we won't find some issues, but only
in SDag.
Benjamin Kramer [Thu, 20 Apr 2017 18:29:14 +0000 (18:29 +0000)]
Fix use-after-frees on memory allocated in a Recycler.
This will become asan errors once the patch lands that poisons the
memory after free. The x86 change is a hack, but I don't see how to
solve this properly at the moment.
Yaxun Liu [Thu, 20 Apr 2017 18:15:34 +0000 (18:15 +0000)]
CodeGen: Let frame index value type match alloca addr space
Recently alloca address space has been added to data layout. Due to this
change, pointer returned by alloca may have different size as pointer in
address space 0.
However, currently the value type of frame index is assumed to be of the
same size as pointer in address space 0.
This patch fixes that.
Most targets assume alloca returning pointer in address space 0, which
is the default alloca address space. Therefore it is NFC for them.
AMDGCN target with amdgiz environment requires this change since it
assumes alloca returning pointer to addr space 5 and its size is 32,
which is different from the size of pointer in addr space 0 which is 64.
Have the AttributeList overload delegate to the AttrBuilder one.
Simplify the AttrBuilder overload by avoiding getSlotAttributes, which
creates temporary AttributeLists.
Simplify `AttrBuilder::removeAttributes(AttributeList, unsigned)` by
using getAttributes instead of manually iterating over slots.
Resubmit "[BitVector] Add operator<<= and operator>>=."
This was failing due to the use of assigning a Mask to an
unsigned, rather than to a BitWord. But most systems do not
have sizeof(unsigned) == sizeof(unsigned long), so the mask
was getting truncated.
getSignBit is a static function that creates an APInt with only the sign bit set. getSignMask seems like a better name to convey its functionality. In fact several places use it and then store in an APInt named SignMask.
[APInt] Add isSubsetOf method that can check if one APInt is a subset of another without creating temporary APInts
This question comes up in many places in SimplifyDemandedBits. This makes it easy to ask without allocating additional temporary APInts.
The BitVector class provides a similar functionality through its (IMHO badly named) test(const BitVector&) method. Though its output polarity is reversed.
I've provided one example use case in this patch. I plan to do more as a follow up.
In SimplifyDemandedUseBits, use computeKnownBits directly to handle Constants
Currently we don't explicitly process ConstantDataSequential, ConstantAggregateZero, or ConstantVector, or Undef before applying the Depth limit. Instead they occur after the depth check in the non-instruction path.
For the constant types that we do handle, the code is replicated from computeKnownBits.
This patch fixes the missing constant handling and the reduces the amount of code by just using computeKnownBits directly for any type of Constant.
projects / llvm / log