Philip Reames [Thu, 11 Dec 2014 01:47:23 +0000 (01:47 +0000)]
GCStrategy should not own GCFunctionInfo
This change moves the ownership and access of GCFunctionInfo (the object which describes the safepoints associated with a safepoint under GCRoot) to GCModuleInfo. Previously, this was owned by GCStrategy which was in turned owned by GCModuleInfo. This made GCStrategy module specific which is 'surprising' given it's name and other purposes.
There's a few more changes needed, but we're getting towards the point we can reuse GCStrategy for gc.statepoint as well.
p.s. The style of this code ends up being a mess. I was trying to move code around without otherwise changing much. Once I get the ownership structure rearranged, I will go through and fixup spacing, naming, comments etc.
Tim Northover [Wed, 10 Dec 2014 23:40:50 +0000 (23:40 +0000)]
ARM: correctly expand LDR-lit based globals.
Quite a major error here: the expansions for the Pseudos with and without
folded load were mixed up. Fortunately it only affects ARM-mode, when not using
movw/movt, on Darwin. I'm guessing no-one actually uses that combination.
DW_OP_const <const> doesn't describe a constant value, but a value at a constant address.
The proper way to describe a constant value is DW_OP_constu <const>, DW_OP_stack_value.
Added DW_OP_stack_value to the stack.
-This line, and those below, will be ignored--
M lib/CodeGen/AsmPrinter/DwarfDebug.cpp
A test/DebugInfo/incorrect-variable-debugloc1.ll
Mark Heffernan [Wed, 10 Dec 2014 22:53:52 +0000 (22:53 +0000)]
Fix PR21694. r219517 added a use of SCEV divide in HowFarToZero computation. This divide can produce incorrect results as we are using an unsigned divide for what should be a modular divide. This change reverts back to a more conservative computation using trailing zeros.
Rafael Espindola [Wed, 10 Dec 2014 20:46:55 +0000 (20:46 +0000)]
Move three methods only used by MCJIT to MCJIT.
These methods are only used by MCJIT and are very specific to it. In fact, they
are also fairly specific to the fact that we have a dynamic linker of
relocatable objects.
Juergen Ributzka [Wed, 10 Dec 2014 19:43:32 +0000 (19:43 +0000)]
[AArch64] MachO large code-model: Materialize FP constants in code.
In the large code model we have to first get the address of the GOT entry, load
the address of the constant, and then load the constant itself.
To avoid these loads and the GOT entry alltogether this commit changes the way
how FP constants are materialized in the large code model. The constats are now
materialized in a GPR and then bitconverted/moved into the FPR.
Sanjay Patel [Wed, 10 Dec 2014 16:58:54 +0000 (16:58 +0000)]
Match new shuffle codegen for MOVHPD patterns
Add patterns to match SSE (shufpd) and AVX (vpermilpd) shuffle codegen
when storing the high element of a v2f64. The existing patterns were
only checking for an unpckh type of shuffle.
[X86] Make a code path in EltsFromConsecutiveLoads work only on vectors it expects
EltsFromConsecutiveLoads was apparently only ever called for 128-bit vectors, and assumed this implicitly. r223518 started calling it for AVX-sized vectors, causing the code path that had this assumption to crash.
This adds a check to make this path fire only for 128-bit vectors.
Eric Fiselier [Wed, 10 Dec 2014 03:42:09 +0000 (03:42 +0000)]
[LIT] Add support for `UNSUPPORTED` tag to `TestRunner.parseIntegratedTestScript`
Summary:
This patch gives me just enough to leverage the existing functionality in `TestRunner` for use in `libc++` and `libc++abi` .
It does the following:
* Adds the `UNSUPPORTED` tag to `TestRunner.parseIntegratedTestScript`.
* Allows `parseIntegratedTestScript` to return an empty script if a script is not required by the caller.
David Majnemer [Wed, 10 Dec 2014 01:17:08 +0000 (01:17 +0000)]
DataLayout: Move asserts over to report_fatal_error
As indicated by the tests, it is possible to feed the AsmParser an
invalid datalayout string. We should verify the result of parsing this
string regardless of whether or not we have assertions enabled.
Matthias Braun [Wed, 10 Dec 2014 01:13:11 +0000 (01:13 +0000)]
MachineVerifier: Allow LiveInterval segments to end at a partial write.
In the subregister liveness tracking case we do not create implicit
reads on partial register writes anymore, still we need to produce a new
SSA value for partial writes so the live segment has to end.
Matthias Braun [Wed, 10 Dec 2014 01:13:04 +0000 (01:13 +0000)]
VirtRegMap: No implicit defs/uses for super registers with subreg liveness tracking.
Adding the implicit defs/uses to the superregisters is semantically questionable
but was not dangerous before as the register allocator never assigned the same
register to two overlapping LiveIntervals even when the actually live
subregisters do not overlap. With subregister liveness tracking enabled this
does actually happen and leads to subsequent bugs if we don't stop adding
the superregister defs/uses.
Matthias Braun [Wed, 10 Dec 2014 01:11:56 +0000 (01:11 +0000)]
Let tablegen compute maximum lanemask for regs/regclasses.
Let tablegen compute the combination of subregister lanemasks for all
subregisters in a register/register class. This is preparation for further
work subregister allocation
Rafael Espindola [Wed, 10 Dec 2014 00:09:35 +0000 (00:09 +0000)]
Simplify the handling of aliases in the gold plugin.
The complicated situation is when we have to keep an alias but drop a GV
that is part of the aliasee.
We used to clone the dropped GV and make the clone internal. This is wasteful
as we know the original will be dropped.
With this patch what is done instead is set the linkage of the original to
internal and replace all uses (but the one in the alias) with a new
declaration that takes the name of the old GV. This saves us from having
to copy the body.
We used to only combine intrinsics, and turn them into VLD1_UPD/VST1_UPD
when the base pointer is incremented after the load/store.
We can do the same thing for generic load/stores.
Note that we can only combine the first load/store+adds pair in
a sequence (as might be generated for a v16f32 load for instance),
because other combines turn the base pointer addition chain (each
computing the address of the next load, from the address of the last
load) into independent additions (common base pointer + this load's
offset).
Philip Reames [Tue, 9 Dec 2014 23:57:54 +0000 (23:57 +0000)]
Remove the Module pointer from GCStrategy and GCMetadataPrinter
In the current implementation, GCStrategy is a part of the ownership structure for the gc metadata which describes a Module. It also contains a reference to the module in question. As a result, GCStrategy instances are essentially Module specific.
I plan to transition away from this design. Instead, a GCStrategy will be owned by the LLVMContext. It will be a lightweight policy object which contains no information about the Modules or Functions involved, but can be easily reached given a Function.
The first step in this transition is to remove the direct Module reference from GCStrategy. This also requires removing the single user of this reference, the GCMetadataPrinter hierarchy. In theory, this will allow the lifetime of the printers to be scoped to the LLVMContext as well, but in practice, I'm not actually changing that. (Yet?)
An alternate design would have been to move the direct Module reference into the GCMetadataPrinter and change the keying of the owning maps to explicitly key off both GCStrategy and Module. I'm open to doing it that way instead, but didn't see much value in preserving the per Module association for GCMetadataPrinters.
The next change in this sequence will be to start unwinding the intertwined ownership between GCStrategy, GCModuleInfo, and GCFunctionInfo.
Hans Wennborg [Tue, 9 Dec 2014 20:39:15 +0000 (20:39 +0000)]
Try fixing MSVC build after r223802
LLVM_EXPLICIT is only supported by recent version of MSVC, and it seems
the not-so-recent versions get confused about the operator bool() when
tryint to resolve operator== calls.
This removed the operator bool()'s since they don't seem to be used
anyway.
[FastISel][AArch64] Fix a missing nullptr check in 'computeAddress'.
The load/store value type is currently not available when lowering the memcpy
intrinsic. Add the missing nullptr check to support this in 'computeAddress'.
Revert r223764 which taught instcombine about integer-based elment extraction
patterns.
This is causing Clang to miscompile itself for 32-bit x86 somehow, and likely
also on ARM and PPC. I really don't know how, but reverting now that I've
confirmed this is actually the culprit. I have a reproduction as well and so
should be able to restore this shortly.
This reverts commit r223764.
Original commit log follows:
Teach instcombine to canonicalize "element extraction" from a load of an
integer and "element insertion" into a store of an integer into actual
element extraction, element insertion, and vector loads and stores.
Previously various parts of LLVM (including instcombine itself) would
introduce integer loads and stores into the code as a way of opaquely
loading and storing "bits". In some cases (such as a memcpy of
std::complex<float> object) we will eventually end up using those bits
in non-integer types. In order for SROA to effectively promote the
allocas involved, it splits these "store a bag of bits" integer loads
and stores up into the constituent parts. However, for non-alloca loads
and tsores which remain, it uses integer math to recombine the values
into a large integer to load or store.
All of this would be "fine", except that it forces LLVM to go through
integer math to combine and split up values. While this makes perfect
sense for integers (and in fact is critical for bitfields to end up
lowering efficiently) it is *terrible* for non-integer types, especially
floating point types. We have a much more canonical way of representing
the act of concatenating the bits of two SSA values in LLVM: a vector
and insertelement. This patch teaching InstCombine to use this
representation.
With this patch applied, LLVM will no longer introduce integer math into
the critical path of every loop over std::complex<float> operations such
as those that make up the hot path of ... oh, most HPC code, Eigen, and
any other heavy linear algebra library.
For the record, I looked *extensively* at fixing this in other parts of
the compiler, but it just doesn't work:
- We really do want to canonicalize memcpy and other bit-motion to
integer loads and stores. SSA values are tremendously more powerful
than "copy" intrinsics. Not doing this regresses massive amounts of
LLVM's scalar optimizer.
- We really do need to split up integer loads and stores of this form in
SROA or every memcpy of a trivially copyable struct will prevent SSA
formation of the members of that struct. It essentially turns off
SROA.
- The closest alternative is to actually split the loads and stores when
partitioning with SROA, but this has all of the downsides historically
discussed of splitting up loads and stores -- the wide-store
information is fundamentally lost. We would also see performance
regressions for bitfield-heavy code and other places where the
integers aren't really intended to be split without seemingly
arbitrary logic to treat integers totally differently.
- We *can* effectively fix this in instcombine, so it isn't that hard of
a choice to make IMO.
Chris Bieneman [Tue, 9 Dec 2014 18:49:55 +0000 (18:49 +0000)]
Adding a new option to CMake to disable C++ atexit on llvm-shlib.
Summary:
This is desirable for WebKit and other clients of the llvm-shlib because C++ exit time destructors have a tendency to crash when invoked from multi-threaded applications.
Ideally this option will be temporary, because the ideal fix is to just not have exit time destructors.
Robert Khasanov [Tue, 9 Dec 2014 18:45:30 +0000 (18:45 +0000)]
[AVX512] Added lowering for VBROADCASTSS/SD instructions.
Lowering patterns were written through avx512_broadcast_pat multiclass as pattern generates VBROADCAST and COPY_TO_REGCLASS nodes.
Added lowering tests.
Split `Metadata` away from the `Value` class hierarchy, as part of
PR21532. Assembly and bitcode changes are in the wings, but this is the
bulk of the change for the IR C++ API.
I have a follow-up patch prepared for `clang`. If this breaks other
sub-projects, I apologize in advance :(. Help me compile it on Darwin
I'll try to fix it. FWIW, the errors should be easy to fix, so it may
be simpler to just fix it yourself.
This breaks the build for all metadata-related code that's out-of-tree.
Rest assured the transition is mechanical and the compiler should catch
almost all of the problems.
Here's a quick guide for updating your code:
- `Metadata` is the root of a class hierarchy with three main classes:
`MDNode`, `MDString`, and `ValueAsMetadata`. It is distinct from
the `Value` class hierarchy. It is typeless -- i.e., instances do
*not* have a `Type`.
- `MDNode`'s operands are all `Metadata *` (instead of `Value *`).
- `TrackingVH<MDNode>` and `WeakVH` referring to metadata can be
replaced with `TrackingMDNodeRef` and `TrackingMDRef`, respectively.
If you're referring solely to resolved `MDNode`s -- post graph
construction -- just use `MDNode*`.
- `MDNode` (and the rest of `Metadata`) have only limited support for
`replaceAllUsesWith()`.
As long as an `MDNode` is pointing at a forward declaration -- the
result of `MDNode::getTemporary()` -- it maintains a side map of its
uses and can RAUW itself. Once the forward declarations are fully
resolved RAUW support is dropped on the ground. This means that
uniquing collisions on changing operands cause nodes to become
"distinct". (This already happened fairly commonly, whenever an
operand went to null.)
If you're constructing complex (non self-reference) `MDNode` cycles,
you need to call `MDNode::resolveCycles()` on each node (or on a
top-level node that somehow references all of the nodes). Also,
don't do that. Metadata cycles (and the RAUW machinery needed to
construct them) are expensive.
- An `MDNode` can only refer to a `Constant` through a bridge called
`ConstantAsMetadata` (one of the subclasses of `ValueAsMetadata`).
As a side effect, accessing an operand of an `MDNode` that is known
to be, e.g., `ConstantInt`, takes three steps: first, cast from
`Metadata` to `ConstantAsMetadata`; second, extract the `Constant`;
third, cast down to `ConstantInt`.
The eventual goal is to introduce `MDInt`/`MDFloat`/etc. and have
metadata schema owners transition away from using `Constant`s when
the type isn't important (and they don't care about referring to
`GlobalValue`s).
In the meantime, I've added transitional API to the `mdconst`
namespace that matches semantics with the old code, in order to
avoid adding the error-prone three-step equivalent to every call
site. If your old code was:
- A `CallInst` -- specifically, intrinsic instructions -- can refer to
metadata through a bridge called `MetadataAsValue`. This is a
subclass of `Value` where `getType()->isMetadataTy()`.
`MetadataAsValue` is the *only* class that can legally refer to a
`LocalAsMetadata`, which is a bridged form of non-`Constant` values
like `Argument` and `Instruction`. It can also refer to any other
`Metadata` subclass.
(I'll break all your testcases in a follow-up commit, when I propagate
this change to assembly.)
Colin LeMahieu [Tue, 9 Dec 2014 18:16:49 +0000 (18:16 +0000)]
[Hexagon] Updating predicate register transfers and adding tstbit to allow select selection. Updating ll tests with predicate transfers that previously had nop encodings.
projects / llvm / log