Nirav Dave [Tue, 28 Feb 2017 14:24:15 +0000 (14:24 +0000)]
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled.
Recommiting after fixup of 32-bit aliasing sign offset bug in DAGCombiner.
* Simplify Consecutive Merge Store Candidate Search
Now that address aliasing is much less conservative, push through
simplified store merging search and chain alias analysis which only
checks for parallel stores through the chain subgraph. This is cleaner
as the separation of non-interfering loads/stores from the
store-merging logic.
When merging stores search up the chain through a single load, and
finds all possible stores by looking down from through a load and a
TokenFactor to all stores visited.
This improves the quality of the output SelectionDAG and the output
Codegen (save perhaps for some ARM cases where we correctly constructs
wider loads, but then promotes them to float operations which appear
but requires more expensive constant generation).
Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)
Additional Minor Changes:
1. Finishes removing unused AliasLoad code
2. Unifies the chain aggregation in the merged stores across code
paths
3. Re-add the Store node to the worklist after calling
SimplifyDemandedBits.
4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
arbitrary, but seems sufficient to not cause regressions in
tests.
5. Remove Chain dependencies of Memory operations on CopyfromReg
nodes as these are captured by data dependence
6. Forward loads-store values through tokenfactors containing
{CopyToReg,CopyFromReg} Values.
7. Peephole to convert buildvector of extract_vector_elt to
extract_subvector if possible (see
CodeGen/AArch64/store-merge.ll)
8. Store merging for the ARM target is restricted to 32-bit as
some in some contexts invalid 64-bit operations are being
generated. This can be removed once appropriate checks are
added.
This finishes the change Matt Arsenault started in r246307 and
jyknight's original patch.
Many tests required some changes as memory operations are now
reorderable, improving load-store forwarding. One test in
particular is worth noting:
CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
forwarding converts a load-store pair into a parallel store and
a memory-realized bitcast of the same value. However, because we
lose the sharing of the explicit and implicit store values we
must create another local store. A similar transformation
happens before SelectionDAG as well.
Daniel Sanders [Tue, 28 Feb 2017 14:21:31 +0000 (14:21 +0000)]
[globalisel] Change LLT constructor string into an LLT subclass that knows how to generate it.
Summary:
This will allow future patches to inspect the details of the LLT. The implementation is now split between
the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns.
Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem.
Diana Picus [Tue, 28 Feb 2017 14:17:53 +0000 (14:17 +0000)]
[ARM] GlobalISel: Lower i32 and fp call parameters on the stack
Lower i32, float and double parameters that need to live on the stack. This
boils down to creating some G_GEPs starting from the stack pointer and storing
the values there. During the process we also keep track of the stack size and
use the final value in the ADJCALLSTACKDOWN/UP instructions.
We currently assert for smaller types, since they usually require extensions.
They will be handled in a separate patch.
Sanne Wouda [Tue, 28 Feb 2017 10:34:48 +0000 (10:34 +0000)]
[Assembler] Add test for !srcloc references in assembler diags
Summary:
clang adds !srcloc metadata to inline assembly in LLVM bitcode generated
for inline assembly in C. The value of this !srcloc is passed to the
diagnostics handler if the inline assembly generates a diagnostic.
clang is able to turn this cookie back to a location in the C source
file.
To test this functionality without a dependency, make llc print the
!srcloc metadata if it is present. The added test uses this mechanism
to test that the correct !srclocs are passed to the diag handler.
Vassil Vassilev [Tue, 28 Feb 2017 07:11:59 +0000 (07:11 +0000)]
Allow externally dlopen-ed libraries to be registered as permanent libraries.
This is also useful in cases when llvm is in a shared library. First we dlopen
the llvm shared library and then we register it as a permanent library in order
to keep the JIT and other services working.
Sanjoy Das [Tue, 28 Feb 2017 07:04:49 +0000 (07:04 +0000)]
[ImplicitNullCheck] Add alias analysis usage
Summary:
With this change ImplicitNullCheck optimization uses alias analysis
and can use load/store memory access for implicit null check if there
are other load/store before but memory accesses do not alias.
Matthias Braun [Tue, 28 Feb 2017 00:33:32 +0000 (00:33 +0000)]
Add MIR-level outlining pass
This is a patch for the outliner described in the RFC at:
http://lists.llvm.org/pipermail/llvm-dev/2016-August/104170.html
The outliner is a code-size reduction pass which works by finding
repeated sequences of instructions in a program, and replacing them with
calls to functions. This is useful to people working in low-memory
environments, where sacrificing performance for space is acceptable.
This adds an interprocedural outliner directly before printing assembly.
For reference on how this would work, this patch also includes X86
target hooks and an X86 test.
[CGP] Split some critical edges coming out of indirect branches
Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.
This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.
Zachary Turner [Tue, 28 Feb 2017 00:04:07 +0000 (00:04 +0000)]
[PDB] Make streams carry their own endianness.
Before the endianness was specified on each call to read
or write of the StreamReader / StreamWriter, but in practice
it's extremely rare for streams to have data encoded in
multiple different endiannesses, so we should optimize for the
99% use case.
This makes the code cleaner and more general, but otherwise
has NFC.
Dan Gohman [Mon, 27 Feb 2017 22:44:37 +0000 (22:44 +0000)]
[MC] Factor out non-COFF handling of COFF-specific directives.
Instead of requiring every non-COFF MCObjectStreamer to implement the
COFF hooks just to do an llvm_unreachable to say that they're not
supported, do the llvm_unreachable in the default implementation, as
suggested by rnk in https://reviews.llvm.org/D26722.
Zachary Turner [Mon, 27 Feb 2017 22:11:43 +0000 (22:11 +0000)]
[PDB] Partial resubmit of r296215, which improved PDB Stream Library.
This was reverted because it was breaking some builds, and
because of incorrect error code usage. Since the CL was
large and contained many different things, I'm resubmitting
it in pieces.
This portion is NFC, and consists of:
1) Renaming classes to follow a consistent naming convention.
2) Fixing the const-ness of the interface methods.
3) Adding detailed doxygen comments.
4) Fixing a few instances of passing `const BinaryStream& X`. These
are now passed as `BinaryStreamRef X`.
...starting in PerformADDCombineWithOperands(), but it wasn't actually checking for a setcc node
for the sext/zext patterns.
This is exactly the opposite of a transform I'd like to add to DAGCombiner's foldSelectOfConstants(),
so I was seeing infinite loops with my draft of a patch applied.
The changes in select_const.ll look positive (less instructions). The change in arm-and-tst-peephole.ll
is unrelated. We're changing the input IR in that test to preserve the intent of the test, but that's
not affected by this code change.
Lang Hames [Mon, 27 Feb 2017 21:09:47 +0000 (21:09 +0000)]
[Support][Error] Add a 'cantFail' utility function for known-safe calls to
fallible functions.
Some fallible functions (those returning Error or Expected<T>) may only fail
for a subset of their inputs. For example, a "safe" square root function will
succeed for all finite positive inputs:
At a safe callsite for such a function, checking the error return value is
redundant:
if (auto ValOrErr = safeSqrt(42.0)) {
// use *ValOrErr.
} else
llvm_unreachable("safeSqrt should always succeed for +ve values");
The cantFail function wraps this check and extracts the contained value,
simplifying control flow:
double Result = cantFail(safeSqrt(42.0));
This function should be used with care: it is a programmatic error to wrap a
call with cantFail if it can in fact fail. For debug builds this will
result in llvm_unreachable being called. For release builds the behavior is
undefined.
Use of this function is likely to be rare in library code, but more common
for tool and unit-test code where inputs and mock functions may be known to be
safe.
Simon Pilgrim [Mon, 27 Feb 2017 21:01:57 +0000 (21:01 +0000)]
[X86][SSE] Attempt to extract vector elements through target shuffles
DAGCombiner already supports peeking thorough shuffles to improve vector element extraction, but legalization often leaves us in situations where we need to extract vector elements after shuffles have already been lowered.
This patch adds support for VECTOR_EXTRACT_ELEMENT/PEXTRW/PEXTRB instructions to attempt to handle target shuffles as well. I've covered some basic scenarios including handling shuffle mask scaling and the implicit zero-extension of PEXTRW/PEXTRB, there is more that could be done here (that I've mentioned in TODOs) but I haven't found many cases where its worth it.
Taewook Oh [Mon, 27 Feb 2017 19:30:01 +0000 (19:30 +0000)]
[TailDuplicator] Maintain DebugLoc for branch instructions
Summary: Existing implementation of duplicateSimpleBB function drops DebugLoc metadata of branch instructions during the transformation. This patch addresses this issue by making newly created branch instructions to keep the metadata of replaced branch instructions.
Craig Topper [Mon, 27 Feb 2017 16:15:32 +0000 (16:15 +0000)]
[X86] Use APInt instead of SmallBitVector tracking undef elements from getTargetConstantBitsFromNode and getConstVector.
Summary:
SmallBitVector uses a malloc for more than 58 bits on a 64-bit target and more than 27 bits on a 32-bit target. Some of the vector types we deal with here use more than those number of elements and therefore cause a malloc.
APInt on the other hand supports up to 64 bits without a malloc. That's the maximum number of bits we need here so we can avoid a malloc for all cases by using APInt.
Craig Topper [Mon, 27 Feb 2017 16:15:30 +0000 (16:15 +0000)]
[X86] Use APInt instead of SmallBitVector for tracking Zeroable elements in shuffle lowering
Summary:
SmallBitVector uses a malloc for more than 58 bits on a 64-bit target and more than 27 bits on a 32-bit target. Some of the vector types we deal with here use more than those number of elements and therefore cause a malloc.
APInt on the other hand supports up to 64 bits without a malloc. That's the maximum number of bits we need here so we can avoid a malloc for all cases by using APInt.
Craig Topper [Mon, 27 Feb 2017 16:15:25 +0000 (16:15 +0000)]
[X86] Use APInt instead of SmallBitVector for tracking undef elements in constant pool shuffle decoding
Summary:
SmallBitVector uses a malloc for more than 58 bits on a 64-bit target and more than 27 bits on a 32-bit target. Some of the vector types we deal with here use more than those number of elements and therefore cause a malloc.
APInt on the other hand supports up to 64 bits without a malloc. That's the maximum number of bits we need here so we can avoid a malloc for all cases by using APInt. This will incur a minor increase in stack usage due to APInt storing the bit count separately from the data bits unlike SmallBitVector, but that should be ok.
Artur Pilipenko [Mon, 27 Feb 2017 15:44:49 +0000 (15:44 +0000)]
Loop predication expand both sides of the widened condition
This is a fix for a loop predication bug which resulted in malformed IR generation.
Loop invariant side of the widened condition is not guaranteed to be available in the preheader as is, so we need to expand it as well. See added unsigned_loop_0_to_n_hoist_length test for example.
Sjoerd Meijer [Mon, 27 Feb 2017 14:45:34 +0000 (14:45 +0000)]
AArch64InstPrinter: rewrite of printSysAlias
This is a cleanup/rewrite of the printSysAlias function. This was not using the
tablegen instruction descriptions, but was "manually" decoding the
instructions. This has been replaced with calls to lookup_XYZ_ByEncoding
tablegen calls.
This revealed several problems. First, instruction IVAU had the wrong encoding.
This was cancelled out by the parser that incorrectly matched the wrong
encoding. Second, instruction CVAP was missing from the SystemOperands tablegen
descriptions, so this has been added. And third, the required target features
were not captured in the tablegen descriptions, so support for this has also
been added.
John Brawn [Mon, 27 Feb 2017 14:40:51 +0000 (14:40 +0000)]
[ARM] LSL #0 is an alias of MOV
Currently we handle this correctly in arm, but in thumb we don't which leads to
an unpredictable instruction being emitted for LSL #0 in an IT block and SP not
being permitted in some cases when it should be.
For the thumb2 LSL we can handle this by making LSL #0 an alias of MOV in the
.td file, but for thumb1 we need to handle it in checkTargetMatchPredicate to
get the IT handling right. We also need to adjust the handling of
MOV rd, rn, LSL #0 to avoid generating the 16-bit encoding in an IT block. We
should also adjust it to allow SP in the same way that it is allowed in
MOV rd, rn, but I haven't done that here because it looks like it would take
quite a lot of work to get right.
Additionally correct the selection of the 16-bit shift instructions in
processInstruction, where it was checking if the two registers were equal when
it should have been checking if they were low. It appears that previously this
code was never executed and the 16-bit encoding was selected by default, but
the other changes I've done here have somehow made it start being used.
Current implementation would identify %v1 load as the first byte load and would mistakenly emit a i16 load from %p1.i16 address. This patch adds a check that the first byte is loaded from a non-zero offset of the first load address. This way this address can be used as the base address for the combined value. Otherwise just give up combining.
[AMDGPU] Runtime metadata fixes:
- Verify that runtime metadata is actually valid runtime metadata when assembling, otherwise we could accept the following when assembling, but ocl runtime will reject it:
.amdgpu_runtime_metadata
{ amd.MDVersion: [ 2, 1 ], amd.RandomUnknownKey, amd.IsaInfo: ...
- Make IsaInfo optional, and always emit it.
projects / llvm / log