Craig Topper [Mon, 18 Feb 2019 06:46:17 +0000 (06:46 +0000)]
[MC] Make SubtargetFeatureKV only store one FeatureBitset and use an 'unsigned' to hold the value.
This class is used for two difference tablegen generated tables. For one of the tables the Value FeatureBitset only has one bit set. For the other usage the Implies field was unused.
This patch changes the Value field to just be an unsigned. For the usage that put a real vector in bitset, we now use the previously unused Implies field and leave the Value field unused instead.
This is good for a 16K reduction in the size of llc on my local build with all targets enabled.
Craig Topper [Sun, 17 Feb 2019 19:23:49 +0000 (19:23 +0000)]
[X86] In FP_TO_INTHelper, when moving data from SSE register to X87 register file via the stack, use the same stack slot we use for the integer conversion.
No need for a separate stack slot. The lifetimes don't overlap.
Also fix the MachinePointerInfo for the final load after the integer conversion to indicate it came from the stack slot.
Nikita Popov [Sun, 17 Feb 2019 17:40:47 +0000 (17:40 +0000)]
[SelectionDAG] Extract [US]MULO expansion into TL method; NFC
In preparation for supporting vector expansion.
Add an isPostTypeLegalization flag to makeLibCall(), because this
expansion relies on the legalized form using MERGE_VALUES. Drop
the corresponding variant of ExpandLibCall, which is no longer used.
Sanjay Patel [Sun, 17 Feb 2019 16:48:50 +0000 (16:48 +0000)]
[InstCombine] reduce more unsigned saturated add with 'not' op
We want to use the sum in the icmp to allow matching with
m_UAddWithOverflow and eliminate the 'not'. This is discussed
in D51929 and is another step towards solving PR14613:
https://bugs.llvm.org/show_bug.cgi?id=14613
Sanjay Patel [Sun, 17 Feb 2019 15:58:48 +0000 (15:58 +0000)]
[InstCombine] reduce unsigned saturated add with 'not' op
We want to use the sum in the icmp to allow matching with
m_UAddWithOverflow and eliminate the 'not'. This is discussed
in D51929 and is another step towards solving PR14613:
https://bugs.llvm.org/show_bug.cgi?id=14613
(The matching here is incomplete. Trying to take minimal steps
to make sure we don't induce infinite looping from existing
canonicalizations of the 'select'.)
Kristina Brooks [Sun, 17 Feb 2019 01:52:18 +0000 (01:52 +0000)]
[LLVMSupport]: Remove a severely outdated README.
The LLVM Support library implementation has resided in
//llvm/lib/Support for a significant amount of time now,
with documentation having been updated with all references
to the "System library" being replaced with "Support library".
Since this file mirrors already existing documentation available
for Support library, includes dead links to documentation and
still refers to it as "System library", having it there is
confusing and updating it has very little point as it duplicates
information in documentation, except documentation is a lot more
up to date while this file has not been maintained.
Up to date documentation concerning this can be found here:
http://llvm.org/docs/SupportLibrary.html
whitequark [Sat, 16 Feb 2019 22:33:10 +0000 (22:33 +0000)]
[bindings/go] Fix building on 32-bit systems (ARM etc.)
Summary:
The patch in https://reviews.llvm.org/D53883 (by me) fails to build on 32-bit systems like ARM. Fix the array size to be less ridiculously large. 2<<20 should still be enough for all practical purposes.
Craig Topper [Sat, 16 Feb 2019 08:25:42 +0000 (08:25 +0000)]
[X86] When type legalizing the result of a i64 fp_to_uint on 32-bit targets. Generate all of the ops as i64 and let them be legalized.
No need to manually split everything. We can let the type legalizer work for us.
The test change seems to be caused by some DAG ordering issue that was previously circumventing a one use check in LowerSELECT where FP selects are turned into blends if the setcc has one use. But it was running after an integer select and the same setcc had been legalized to cmov and X86SISD::CMP. This dropped the use count of the setcc, but wasn't what was intended.
Dave Lee [Sat, 16 Feb 2019 06:59:49 +0000 (06:59 +0000)]
llvm-nm: Observe -no-llvm-bc for archive members
Summary:
This change fixes the `-no-llvm-bc` flag to work with object files within
archives. Currently the `-no-llvm-bc` flag works for regular object files, but
not static libraries, where it continues to show bitcode symbol info.
Craig Topper [Sat, 16 Feb 2019 03:34:54 +0000 (03:34 +0000)]
[X86] Don't prevent load folding for cvtsi2ss/cvtsi2sd based on hasPartialRegUpdate.
Preventing the load fold won't fix the partial register update since the
input we can fold is a GPR. So it will do nothing to prevent a false dependency
on an XMM register.
Julian Lettner [Sat, 16 Feb 2019 00:44:00 +0000 (00:44 +0000)]
[lit] Remove LitTestCase
From the docs: `class LitTestCase(unittest.TestCase)`
LitTestCase is an adaptor for providing a 'unittest' compatible
interface to 'lit' tests so that we can run lit tests with standard
python test runners.
Julian Lettner [Sat, 16 Feb 2019 00:40:40 +0000 (00:40 +0000)]
[lit][NFC] Cleanup lit worker process handling
Move code that is executed on worker process to separate file. This
makes the use of the pickled arguments stored in global variables in the
worker a bit clearer. (Still not pretty though.)
Extract handling of parallelism groups to it's own function.
Use BoundedSemaphore instead of Semaphore. BoundedSemaphore raises for
unmatched release() calls.
Alina Sbirlea [Fri, 15 Feb 2019 22:47:54 +0000 (22:47 +0000)]
[EarlyCSE & MSSA] Cap the clobbering calls in EarlyCSE.
Summary:
Unlimitted number of calls to getClobberingAccess can lead to high
compile times in pathological cases.
Limitting getClobberingAccess to a fairly high number. Can be adjusted
based on users/need.
Note: this is the only user of MemorySSA currently enabled by default.
The same handling exists in LICM (disabled atm). As MemorySSA gains more
users, this logic of capping will need to move inside MemorySSA.
Craig Topper [Fri, 15 Feb 2019 21:59:33 +0000 (21:59 +0000)]
[X86] Don't set exception mask bits when modifying FPCW to change rounding mode for fp->int conversion
When we need to do an fp->int conversion using x87 instructions, we need to temporarily change the rounding mode to 0b11 and perform a store. To do this we save the old value of the fpcw to the stack, then set the fpcw to 0xc7f, do the store, then restore fpcw. But the 0xc7f value forces the exception mask bits 1. While this is what they would be in the default FP environment, as we move to support changing the FP environments, we shouldn't make this assumption.
This patch changes the code to explicitly OR 0xc00 with the old value so that only the rounding mode is changed. Unfortunately, this requires two stack temporaries instead of one. One to hold the old value and one to hold the new value. Without two stack temporaries we would need an additional GPR. We already need one to do the OR operation in. This is similar to what gcc and icc do for this operation. Though they are both better at reusing the stack temporaries when there are multiple truncates in a function(or at least in a basic block)
Philip Reames [Fri, 15 Feb 2019 21:23:51 +0000 (21:23 +0000)]
[InstCombine] Convert atomicrmws to xchg or store where legal
Implement two more transforms of atomicrmw:
1) We can convert an atomicrmw which produces a known value in memory into an xchg instead.
2) We can convert an atomicrmw xchg w/o users into a store for some orderings.
Craig Topper [Fri, 15 Feb 2019 19:21:39 +0000 (19:21 +0000)]
[X86] Move all the SSE legality checks out of FP_TO_INTHelper and up to LowerFP_TO_INT. NFCI
These checks aren't needed on the call to FP_TO_INTHelper from the type legalizer for splitting i64. We always want to use X87 FIST/FISTT to memory there.
Moving up the SSE checks will allow this routine to focus on what it cares about and makes its return semantics cleaner.
Jonas Paulsson [Fri, 15 Feb 2019 19:13:55 +0000 (19:13 +0000)]
Recommit "[SystemZ] Do not emit VEXTEND or VROUND nodes without vector support."
It seems there were some problem with using a .mir test. For some reason
doing '-stop-before=codegenprepare' and then '-start-before=codegenprepare'
on the output .mir file results in the NoVRegs Property after instruction
selection.
Recommitting the same test as an .ll file instead.
Vedant Kumar [Fri, 15 Feb 2019 18:46:58 +0000 (18:46 +0000)]
[CodeExtractor] Do not lift lifetime.end markers for region inputs
If a lifetime.end marker occurs along one path through the extraction
region, but not another, then it's still incorrect to lift the marker,
because there is some path through the extracted function which would
ordinarily not reach the marker. If the call to the extracted function
is in a loop, unrolling can cause inputs to the function to become
optimized out as undef after the first iteration.
To prevent incorrect stack slot merging in the calling function, it
should be sufficient to lift lifetime.start markers for region inputs.
I've tested this theory out by doing a stage2 check-all with randomized
splitting enabled.
This is a follow-up to r353973, and there's additional context for this
change in https://reviews.llvm.org/D57834.
Vedant Kumar [Fri, 15 Feb 2019 18:46:44 +0000 (18:46 +0000)]
[HotColdSplit] Schedule splitting late to fix perf regression
With or without PGO data applied, splitting early in the pipeline
(either before the inliner or shortly after it) regresses performance
across SPEC variants. The cause appears to be that splitting hides
context for subsequent optimizations.
Schedule splitting late again, in effect reversing r352080, which
scheduled the splitting pass early for code size benefits (documented in
https://reviews.llvm.org/D57082).
Matt Arsenault [Fri, 15 Feb 2019 15:24:34 +0000 (15:24 +0000)]
GlobalISel: Fix inadequate verification of g_build_vector
Testing based on the total size of the elements failed to catch a few
invalid scenarios, so explicitly check the number of elements/operands
and types.
This failed to catch situations like
<4 x s16> = G_BUILD_VECTOR s32, s32 since the total size added
up. This also would fail to catch an implicit conversion between
pointers and scalars.
Matt Arsenault [Fri, 15 Feb 2019 15:24:31 +0000 (15:24 +0000)]
Try to organize MachineVerifier tests
The Verifier is separate from the MachineVerifier, so move it to a
different directory. Some other verifier tests were scattered in
target codegen tests as well (although I'm sure I missed some). Work
towards using a more consistent naming scheme to make it clearer where
the gaps still are for generic instructions.
Serge Guelton [Fri, 15 Feb 2019 15:17:29 +0000 (15:17 +0000)]
OptionalStorage implementation for trivial type, take III
This is another attempt at implementating optional storage
for trivially copyable type, using an union instead of a
raw buffer to hold the actual storage. This make it possible
to get rid of the reinterpret_cast, and hopefully to fix the UB
of the previous attempts.
This validates fine on my laptop for gcc 8.2 and gcc 4.8, I'll
revert if it breaks the validation.
Clement Courbet [Fri, 15 Feb 2019 14:17:17 +0000 (14:17 +0000)]
[MergeICmps] Make base ordering really deterministic.
Summary:
The idea is that we now manipulate bases through a `unsigned BaseID` based on
order of appearance in the comparison chain rather than through the `Value*`.
Fixes 40714.
Reviewers: gchatelet
Subscribers: mgrang, jfb, jdoerfert, llvm-commits, hans
Hans Wennborg [Fri, 15 Feb 2019 12:20:33 +0000 (12:20 +0000)]
Speculatively revert r354051 "Recommit Optional specialization for trivially copyable types"
and
r354055 "Optional specialization for trivially copyable types, part2"
These are suspected to cause Clang to get miscompiled on Ubuntu 14.04
(Trusty) which uses GCC 4.8.4. Reverting for an hour to see if this
helps. See llvm-commits thread.
> Recommit Optional specialization for trivially copyable types
>
> Unfortunately the original code gets misscompiled by GCC (at least 8.1),
> this is a tentative workaround using std::memcpy instead of inplace new
> for trivially copyable types. I'll revert if it breaks.
>
> Original revision: https://reviews.llvm.org/D57097
Sam Parker [Fri, 15 Feb 2019 11:50:21 +0000 (11:50 +0000)]
[BPI] Look through bitcasts in calcZeroHeuristic
Constant hoisting may have hidden a constant behind a bitcast so that
it isn't folded into its users. However, this prevents BPI from
calculating some of its heuristics that are based upon constant
values. So, I've added a simple helper function to look through these
casts.
Simon Pilgrim [Fri, 15 Feb 2019 11:39:21 +0000 (11:39 +0000)]
[X86][AVX] lowerShuffleAsLanePermuteAndPermute - fully populate the lane shuffle mask (PR40730)
As detailed on PR40730, we are not correctly filling in the lane shuffle mask (D53148/rL344446) - we fill in for the correct src lane but don't add it to the correct mask element, so any reference to the correct element is likely to see an UNDEF mask index.
This allows constant folding to propagate UNDEFs prior to the lane mask being (correctly) lowered to vperm2f128.
This patch fixes the issue by fully populating the lane shuffle mask - this is more than is necessary (if we only filled in the required mask elements we might be able to match other shuffle instructions - broadcasts etc.), but its the most cautious approach as this needs to be cherrypicked into the 8.0.0 release branch.
Diana Picus [Fri, 15 Feb 2019 10:50:02 +0000 (10:50 +0000)]
[ARM GlobalISel] Style fix. NFCI
Add the opcode for ADDrr / t2ADDrr to the Opcode cache, as we did for
all other opcodes where the handling is otherwise the same between arm
mode and thumb2.
Sam Parker [Fri, 15 Feb 2019 09:04:39 +0000 (09:04 +0000)]
[ARM CGP] Fix ConvertTruncs
ConvertTruncs is used to replace a trunc for an AND mask, however
this function wasn't working as expected. By performing the change
later, we can create a wide type integer mask instead of a narrow -1
value, which could then be simply removed (incorrectly). Because we
now perform this action later, it's necessary to cache the trunc type
before we perform the promotion.
[GISel][NFC]: Add methods to speed up insertion into GISelWorklist
https://reviews.llvm.org/D58073
Speed up insertion during the initial populating phase into the
GISelWorkList by deferring repeatedly resizing the DenseMap.
This results in ~10% improvement in the combiner passes, and
~3% speedup in the Legalizer.
Matt Davis [Thu, 14 Feb 2019 23:50:35 +0000 (23:50 +0000)]
[symbolizer] Avoid collecting symbols belonging to invalid sections.
Summary:
llvm-symbolizer would originally report symbols that belonged to an invalid object file section.
Specifically the case where: `*Symbol.getSection() == ObjFile.section_end()`
This patch prevents the Symbolizer from collecting symbols that belong to invalid sections.
The test (from PR40591) introduces a case where two symbols have address 0,
one symbol is defined, 'foo', and the other is not defined, 'bar'. This patch will cause
the Symbolizer to keep 'foo' and ignore 'bar'.
As a side note, the logic for adding symbols to the Symbolizer's store
(`SymbolizableObjectFile::addSymbol`) replaces symbols with the
same <address, size> pair. At some point that logic should be revisited as in the
aforementioned case, 'bar' was overwriting 'foo' in the Symbolizer's store,
and 'foo' was forgotten.
Nick Desaulniers [Thu, 14 Feb 2019 23:35:53 +0000 (23:35 +0000)]
[INLINER] allow inlining of address taken blocks
as long as their uses does not contain calls to functions that capture
the argument (potentially allowing the blockaddress to "escape" the
lifetime of the caller).
TODO:
- add more tests
- fix crash in llvm::updateCGAndAnalysisManagerForFunctionPass when
invoking Transforms/Inline/blockaddress.ll
Matt Arsenault [Thu, 14 Feb 2019 22:41:01 +0000 (22:41 +0000)]
Replace gcroot verifier tests
These haven't been checking anything useful and have been testing the
wrong failure reason for many years. Replace them with something which
stresses what is actually implemented in the verifier now.
[AMDGPU] Ressociate 'add (add x, y), z' to use SALU
Reassociate adds to collect scalar operands in a single
instruction when possible. That will result in a scalar
add followed by vector instead of two vector adds, thus
better utilizing SALU.
Teresa Johnson [Thu, 14 Feb 2019 21:22:50 +0000 (21:22 +0000)]
[ThinLTO] Detect partially split modules during the thin link
Summary:
The changes to disable LTO unit splitting by default (r350949) and
detect inconsistently split LTO units (r350948) are causing some crashes
when the inconsistency is detected in multiple threads simultaneously.
Fix that by having the code always look for the inconsistently split
LTO units during the thin link, by checking for the presence of type
tests recorded in the summaries.
Modify test added in r350948 to remove single threading required to fix
a bot failure due to this issue (and some debugging options added in the
process of diagnosing it).
Chris Bieneman [Thu, 14 Feb 2019 20:57:17 +0000 (20:57 +0000)]
[CMake] Fix ability to use LLVM_ENABLE_PROJECTS with LLVM_EXTERNAL_PROJECTS
LLVM r353148, changed the circumstances in which the project source directory variables are created to only create them for LLVM projects. This patch initializes the directory variables for projects specified in `LLVM_EXTERNAL_PROJECTS` as well.
Philip Reames [Thu, 14 Feb 2019 20:41:17 +0000 (20:41 +0000)]
Canonicalize all integer "idempotent" atomicrmw ops
For "idempotent" atomicrmw instructions which we can't simply turn into load, canonicalize the operation and constant. This reduces the matching needed elsewhere in the optimizer, but doesn't directly impact codegen.
For any architecture where OR/Zero is not a good default choice, you can extend the AtomicExpand lowerIdempotentRMWIntoFencedLoad mechanism. I reviewed X86 to make sure this works well, haven't audited other backends.
Serge Guelton [Thu, 14 Feb 2019 19:17:00 +0000 (19:17 +0000)]
Recommit Optional specialization for trivially copyable types
Unfortunately the original code gets misscompiled by GCC (at least 8.1),
this is a tentative workaround using std::memcpy instead of inplace new
for trivially copyable types. I'll revert if it breaks.
Original revision: https://reviews.llvm.org/D57097
Philip Reames [Thu, 14 Feb 2019 18:39:14 +0000 (18:39 +0000)]
Teach instcombine about remaining "idempotent" atomicrmw types
Expand on Quentin's r353471 patch which converts some atomicrmws into loads. Handle remaining operation types, and fix a slight bug. Atomic loads are required to have alignment. Since this was within the InstCombine fixed point, somewhere else in InstCombine was adding alignment before the verifier saw it, but still, we should fix.
Terminology wise, I'm using the "idempotent" naming that is used for the same operations in AtomicExpand and X86ISelLoweringInfo. Once this lands, I'll add similar tests for AtomicExpand, and move the pattern match function to a common location. In the review, there was seemingly consensus that "idempotent" was slightly incorrect for this context. Once we setle on a better name, I'll update all uses at once.
Jordan Rupprecht [Thu, 14 Feb 2019 18:35:13 +0000 (18:35 +0000)]
[llvm-ar] Implement the P modifier.
Summary:
GNU ar has a `P` modifier that changes filename comparisons to use full paths instead of the basename. As noted in the GNU docs, regular archives are not created with full path names, so P is used to deal with archives created by other archive programs (e.g. see the updated `absolute-paths.test` test case).
Since thin archives use full path names -- paths are relative to the archive -- it seems very error prone to not imply P when dealing with thin archives, so P is implied in those cases. (I think this is a deviation from GNU ar that makes sense).
This fixes PR37436 via https://github.com/ClangBuiltLinux/linux/issues/33.
Teresa Johnson [Thu, 14 Feb 2019 14:14:24 +0000 (14:14 +0000)]
Refine ArgPromotion metadata handling
Summary:
In r353537 we now copy all metadata to the new function, with the old
being removed when the old function is eliminated. In some cases the old
function is dropped to a declaration (seems to only occur with the old
PM). Go ahead and clear all metadata from the old function to handle that
case, since verification will complain otherwise. This is consistent
with what was being done for debug metadata before r353537.
Florian Hahn [Thu, 14 Feb 2019 13:59:39 +0000 (13:59 +0000)]
[LoopUnrollPeel] Add case where we should forget the peeled loop from SCEV.
The test case requires the peeled loop to be forgotten after peeling,
even though it does not have a parent. When called via the unroller,
SE->forgetTopmostLoop is also called, so the test case would also pass
without any SCEV invalidation, but peelLoop is exposed as utility
function. Also, in the test case, simplifyLoop will make changes,
removing the loop from SCEV, but it is better to not rely on this
behavior.
projects / llvm / log