[SLP] Revert r312791 and other necessary commits, except for TTI and
CostModel.
The original patch added support for horizontal min/max reductions to
the SLP vectorizer.
This patch causes LLVM to miscompile fairly simple signed min
reductions. I have attached a test progrom to http://llvm.org/PR34635
that shows the behavior change after this patch. We found this in a test
for the open source Eigen library, but also in other code.
Unfortunately, the revert is moderately challenging. It required
reverting:
r313042: [SLP] Test with multiple uses of conditional op and wrong parent.
r312853: [SLP] Fix buildbots, NFC.
r312793: [SLP] Fix the warning about paths not returning the value, NFC.
r312791: [SLP] Support for horizontal min/max reduction.
And even then, I had to completely skip reverting the changes to TTI and
CostModel because r312832 rewrote so much of this code. Plus, the cost
modeling changes aren implicated in the miscompile, so they should be
fine and will just not be used until this gets re-introduced.
Resubmit "[lit] Force site configs to run before source-tree configs"
This is a resubmission of r313270. It broke standalone builds of
compiler-rt because we were not correctly generating the llvm-lit
script in the standalone build directory.
The fixes incorporated here attempt to find llvm/utils/llvm-lit
from the source tree returned by llvm-config. If present, it
will generate llvm-lit into the output directory. Regardless,
the user can specify -DLLVM_EXTERNAL_LIT to point to a specific
lit.py on their file system. This supports the use case of
someone installing lit via a package manager. If it cannot find
a source tree, and -DLLVM_EXTERNAL_LIT is either unspecified or
invalid, then we print a warning that tests will not be able
to run.
[DebugInfo] Insert DW_OP_deref when spilling indirect DBG_VALUEs
Summary:
This comes up in optimized debug info for C++ programs that pass and
return objects indirectly by address. In these programs,
llvm.dbg.declare survives optimization, which causes us to emit indirect
DBG_VALUE instructions. The fast register allocator knows to insert
DW_OP_deref when spilling indirect DBG_VALUE instructions, but the
LiveDebugVariables did not until this change.
This fixes part of PR34513. I need to look into why this doesn't work at
-O0 and I'll send follow up patches to handle that.
[DebugInfo] Add missing DW_OP_deref when an NRVO pointer is spilled
Summary:
Fixes PR34513.
Indirect DBG_VALUEs typically come from dbg.declares of non-trivially
copyable C++ objects that must be passed by address. We were already
handling the case where the virtual register gets allocated to a
physical register and is later spilled. That's what usually happens for
normal parameters that aren't NRVO variables: they usually appear in
physical register parameters, and are spilled later in the function,
which would correctly add deref.
NRVO variables are different because the dbg.declare can come much later
after earlier instructions cause the incoming virtual register to be
spilled.
Also, clean up this code. We only need to look at the first operand of a
DBG_VALUE, which eliminates the operand loop.
Steven Wu [Fri, 15 Sep 2017 21:12:14 +0000 (21:12 +0000)]
[AutoUpgrade] Fix a compatibility issue with module flag
Summary:
After r304661, module flag to record objective-c image info section is
encoded without whitespaces after comma. The new name is equivalent to
the old one, except that when LTO a module built by old compiler and a
module built by a new compiler, it will fail with conflicting values.
Fix the issue by removing whitespaces in bitcode upgrade path.
Davide Italiano [Fri, 15 Sep 2017 20:53:05 +0000 (20:53 +0000)]
[ConstantFold] Return the correct type when folding a GEP with vector indices.
As Eli pointed out (and I got wrong in the first place), langref says: "The
getelementptr returns a vector of pointers, instead of a single address, when one
or more of its arguments is a vector. In such cases, all vector arguments should
have the same number of elements, and every scalar argument will be effectively
broadcast into a vector during address calculation."
Costantfold for gep doesn't really take in account this paragraph, returning a
pointer instead of a vector of pointer which triggers an assertion in RAUW,
as we're trying to replace values with mistmatching types.
Sam Clegg [Fri, 15 Sep 2017 20:34:47 +0000 (20:34 +0000)]
Change encodeU/SLEB128 to pad to certain number of bytes
Previously the 'Padding' argument was the number of padding
bytes to add. However most callers that use 'Padding' know
how many overall bytes they need to write. With the previous
code this would mean encoding the LEB once to find out how
many bytes it would occupy and then using this to calulate
the 'Padding' value.
This patch fixes https://bugs.llvm.org/show_bug.cgi?id=32352
It enables OptimizationRemarkEmitter::allowExtraAnalysis and MachineOptimizationRemarkEmitter::allowExtraAnalysis to return true not only for -fsave-optimization-record but when specific remarks are requested with
command line options.
The diagnostic handler used to be callback now this patch adds a class
DiagnosticHandler. It has virtual method to provide custom diagnostic handler
and methods to control which particular remarks are enabled.
However LLVM-C API users can still provide callback function for diagnostic handler.
This patch fixes https://bugs.llvm.org/show_bug.cgi?id=32352
It enables OptimizationRemarkEmitter::allowExtraAnalysis and MachineOptimizationRemarkEmitter::allowExtraAnalysis to return true not only for -fsave-optimization-record but when specific remarks are requested with
command line options.
The diagnostic handler used to be callback now this patch adds a class
DiagnosticHandler. It has virtual method to provide custom diagnostic handler
and methods to control which particular remarks are enabled.
However LLVM-C API users can still provide callback function for diagnostic handler.
Hans Wennborg [Fri, 15 Sep 2017 18:40:26 +0000 (18:40 +0000)]
Revert r313343 "[X86] PR32755 : Improvement in CodeGen instruction selection for LEAs."
This caused PR34629: asserts firing when building Chromium. It also broke some
buildbots building test-suite as reported on the commit thread.
> Summary:
> 1/ Operand folding during complex pattern matching for LEAs has been
> extended, such that it promotes Scale to accommodate similar operand
> appearing in the DAG.
> e.g.
> T1 = A + B
> T2 = T1 + 10
> T3 = T2 + A
> For above DAG rooted at T3, X86AddressMode will no look like
> Base = B , Index = A , Scale = 2 , Disp = 10
>
> 2/ During OptimizeLEAPass down the pipeline factorization is now performed over LEAs
> so that if there is an opportunity then complex LEAs (having 3 operands)
> could be factored out.
> e.g.
> leal 1(%rax,%rcx,1), %rdx
> leal 1(%rax,%rcx,2), %rcx
> will be factored as following
> leal 1(%rax,%rcx,1), %rdx
> leal (%rdx,%rcx) , %edx
>
> 3/ Aggressive operand folding for AM based selection for LEAs is sensitive to loops,
> thus avoiding creation of any complex LEAs within a loop.
>
> Reviewers: lsaba, RKSimon, craig.topper, qcolombet
>
> Reviewed By: lsaba
>
> Subscribers: spatel, igorb, llvm-commits
>
> Differential Revision: https://reviews.llvm.org/D35014
[X86] Don't create i64 constants on 32-bit targets when lowering v64i1 constant build vectors
When handling a v64i1 build vector of constants on 32-bit targets we were creating an illegal i64 constant that we then bitcasted back to v64i1. We need to instead create two 32-bit constants, bitcast them to v32i1 and concat the result. We should also take care to handle the halves being all zeros/ones after the split.
This patch splits the build vector and then recursively lowers the two pieces. This allows us to handle the all ones and all zeros cases with minimal effort. Ideally we'd just do the split and concat, and let lowering get called again on the new nodes, but getNode has special handling for CONCAT_VECTORS that reassembles the pieces back into a single BUILD_VECTOR. Hopefully the two temporary BUILD_VECTORS we had to create to do this that don't get returned don't cause any issues.
[X86] Add isel pattern infrastructure to begin recognizing when we're inserting 0s into the upper portions of a vector register and the producing instruction as already produced the zeros.
Currently if we're inserting 0s into the upper elements of a vector register we insert an explicit move of the smaller register to implicitly zero the upper bits. But if we can prove that they are already zero we can skip that. This is based on a similar idea of what we do to avoid emitting explicit zero extends for GR32->GR64.
Unfortunately, this is harder for vector registers because there are several opcodes that don't have VEX equivalent instructions, but can write to XMM registers. Among these are SHA instructions and a MMX->XMM move. Bitcasts can also get in the way.
So for now I'm starting with explicitly allowing only VPMADDWD because we emit zeros in combineLoopMAddPattern. So that is placing extra instruction into the reduction loop.
I'd like to allow PSADBW as well after D37453, but that's currently blocked by a bitcast. We either need to peek through bitcasts or canonicalize insert_subvectors with zeros to remove bitcasts on the value being inserted.
Longer term we should probably have a cleanup pass that removes superfluous zeroing moves even when the producer is in another basic block which is something these isel tricks can't do. See PR32544.
Anna Thomas [Fri, 15 Sep 2017 15:56:05 +0000 (15:56 +0000)]
[RuntimeUnroll] Add heuristic for unrolling multi-exit loop
Add a profitability heuristic to enable runtime unrolling of multi-exit
loop: There can be atmost two unique exit blocks for the loop and the
second exit block should be a deoptimizing block. Also, there can be one
other exiting block other than the latch exiting block. The reason for
the latter is so that we limit the number of branches in the unrolled
code to being at most the unroll factor. Deoptimizing blocks are rarely
taken so these additional number of branches created due to the
unrolling are predictable, since one of their target is the deopt block.
Anna Thomas [Fri, 15 Sep 2017 13:29:33 +0000 (13:29 +0000)]
[RuntimeUnrolling] Populate the VMap entry correctly when default generated through lookup
During runtime unrolling on loops with multiple exits, we update the
exit blocks with the correct phi values from both original and remainder
loop.
In this process, we lookup the VMap for the mapped incoming phi values,
but did not update the VMap if a default entry was generated in the VMap
during the lookup. This default value is generated when constants or
values outside the current loop are looked up.
This patch fixes the assertion failure when null entries are present in
the VMap because of this lookup. Added a testcase that showcases the
problem.
[AArch64] allow v8f16 types when FullFP16 is supported
This adds support for allowing v8f16 vector types, thus avoiding conversions
from/to single precision for these types. This is a follow up patch of
commits r311154 and r312104, which added support for scalars and v4f16
types, respectively.
[SLPVectorizer] Failure to beneficially vectorize 'copyable' elements in integer binary ops.
Patch tries to improve vectorization of the following code:
void add1(int * __restrict dst, const int * __restrict src) {
*dst++ = *src++;
*dst++ = *src++ + 1;
*dst++ = *src++ + 2;
*dst++ = *src++ + 3;
}
Allows to vectorize even if the very first operation is not a binary add, but just a load.
[X86] PR32755 : Improvement in CodeGen instruction selection for LEAs.
Summary:
1/ Operand folding during complex pattern matching for LEAs has been
extended, such that it promotes Scale to accommodate similar operand
appearing in the DAG.
e.g.
T1 = A + B
T2 = T1 + 10
T3 = T2 + A
For above DAG rooted at T3, X86AddressMode will no look like
Base = B , Index = A , Scale = 2 , Disp = 10
2/ During OptimizeLEAPass down the pipeline factorization is now performed over LEAs
so that if there is an opportunity then complex LEAs (having 3 operands)
could be factored out.
e.g.
leal 1(%rax,%rcx,1), %rdx
leal 1(%rax,%rcx,2), %rcx
will be factored as following
leal 1(%rax,%rcx,1), %rdx
leal (%rdx,%rcx) , %edx
3/ Aggressive operand folding for AM based selection for LEAs is sensitive to loops,
thus avoiding creation of any complex LEAs within a loop.
Martin Pelikan [Fri, 15 Sep 2017 04:22:16 +0000 (04:22 +0000)]
[XRay] fix and clarify comments in the log file decoder
Summary:
For readers unfamiliar with the XRay code base, reference the compiler-rt
implementation even though we're not allowed to share any code and explain
our little-endian views more clearly.
For code clarity either get rid of obvious comments or explain their
intentions, fix typos, correct coding style according to LLVM's standards
and manually CSE long expressions to point out it is the same expression.
Petr Hosek [Fri, 15 Sep 2017 02:59:55 +0000 (02:59 +0000)]
[Object] Fix missing arguments to getType and getSymbol in Elf_Rel_Impl
Somehow this was compiling without these methods having their arguments
passed to them. I used these methods in some code I wrote and it raised
an error on me. It appears no one else has used these methods let (LLD
uses setSymbolAndType however). This change resolves the issue.
Revert "[lit] Force site configs to run before source-tree configs"
This patch is still breaking several multi-stage compiler-rt bots.
I already know what the fix is, but I want to get the bots green
for now and then try re-applying in the morning.
[lit] Add a lit.llvm module that all test suites can use.
To further reduce duplicate code, this patch introduces a module
that configs can simply import and get access to a lot of useful
functionality such as setting up paths, adding features that are
useful across all projects, and other utility-type functions.
For now this only updates llvm's suite to use this new library,
but subsequent patches will update other projects.
Benjamin Kramer [Thu, 14 Sep 2017 18:33:25 +0000 (18:33 +0000)]
Remove usages of deprecated std::unary_function and std::binary_function.
These are removed in C++17. We still have some users of
unary_function::argument_type, so just spell that typedef out. No
functionality change intended.
Note that many of the argument types are actually wrong :)
Matt Arsenault [Thu, 14 Sep 2017 17:37:40 +0000 (17:37 +0000)]
AMDGPU: Stop modifying SP in call sequences
Because the stack growth direction and addressing is done
in the same direction, modifying SP at the beginning of the
call sequence was incorrect. If we had a stack passed argument,
we would end up skipping that number of bytes before pushing
arguments, leaving unused/inconsistent space.
The callee creates fixed stack objects in its frame, so
the space necessary for these is already logically allocated
in the callee, so we just let the callee increment SP if
it really requires it.
Invoke GetInlineCost for legality check before inline functions in SampleProfileLoader.
Summary: SampleProfileLoader inlines hot functions if it is inlined in the profiled binary. However, the inline needs to be guarded by legality check, otherwise it could lead to correctness issues.
Simon Dardis [Thu, 14 Sep 2017 17:27:53 +0000 (17:27 +0000)]
[mips] Implement the 'dext' aliases and it's disassembly alias.
The other members of the dext family of instructions (dextm, dextu) are
traditionally handled by the assembler selecting the right variant of
'dext' depending on the values of the position and size operands.
When these instructions are disassembled, rather than reporting the
actual instruction, an equivalent aliased form of 'dext' is generated
and is reported. This is to mimic the behaviour of binutils.
Matt Arsenault [Thu, 14 Sep 2017 17:14:57 +0000 (17:14 +0000)]
AMDGPU: Make frame register caller preserved
Using SplitCSR for the frame register was very broken. Often
the copies in the prolog and epilog were optimized out, in addition
to them being inserted after the true prolog where the FP
was clobbered.
I have a hacky solution which works that continues to use
split CSR, but for now this is simpler and will get to working
programs.
TableGen support for parameterized register class information
This replaces TableGen's type inference to operate on parameterized
types instead of MVTs, and as a consequence, some interfaces have
changed:
- Uses of MVTs are replaced by ValueTypeByHwMode.
- EEVT::TypeSet is replaced by TypeSetByHwMode.
This affects the way that types and type sets are printed, and the
tests relying on that have been updated.
There are certain users of the inferred types outside of TableGen
itself, namely FastISel and GlobalISel. For those users, the way
that the types are accessed have changed. For typical scenarios,
these replacements can be used:
- TreePatternNode::getType(ResNo) -> getSimpleType(ResNo)
- TreePatternNode::hasTypeSet(ResNo) -> hasConcreteType(ResNo)
- TypeSet::isConcrete -> TypeSetByHwMode::isValueTypeByHwMode(false)
For more information, please refer to the review page.
[lit] Force site configs to be run before source-tree configs
This patch simplifies LLVM's lit infrastructure by enforcing an ordering
that a site config is always run before a source-tree config.
A significant amount of the complexity from lit config files arises from
the fact that inside of a source-tree config file, we don't yet know if
the site config has been run. However it is *always* required to run
a site config first, because it passes various variables down through
CMake that the main config depends on. As a result, every config
file has to do a bunch of magic to try to reverse-engineer the location
of the site config file if they detect (heuristically) that the site
config file has not yet been run.
This patch solves the problem by emitting a mapping from source tree
config file to binary tree site config file in llvm-lit.py. Then, during
discovery when we find a config file, we check to see if we have a
target mapping for it, and if so we use that instead.
This mechanism is generic enough that it does not affect external users
of lit. They will just not have a config mapping defined, and everything
will work as normal.
On the other hand, for us it allows us to make many simplifications:
* We are guaranteed that a site config will be executed first
* Inside of a main config, we no longer have to assume that attributes
might not be present and use getattr everywhere.
* We no longer have to pass parameters such as --param llvm_site_config=<path>
on the command line.
* It is future-proof, meaning you don't have to edit llvm-lit.in to add
support for new projects.
* All of the duplicated logic of trying various fallback mechanisms of
finding a site config from the main config are now gone.
One potentially noteworthy thing that was required to implement this
change is that whereas the ninja check targets previously used the first
method to spawn lit, they now use the second. In particular, you can no
longer run lit.py against the source tree while specifying the various
`foo_site_config=<path>` parameters. Instead, you need to run
llvm-lit.py.
Simon Dardis [Thu, 14 Sep 2017 15:17:50 +0000 (15:17 +0000)]
[mips] Implement the 'dins' aliases.
Traditionally GAS has provided automatic selection between dins, dinsm and
dinsu. Binutils also disassembles all instructions in that family as 'dins'
rather than the actual instruction.
[InstSimplify] fold sdiv/srem based on compare of dividend and divisor
This should bring signed div/rem analysis up to the same level as unsigned.
We use icmp simplification to determine when the divisor is known greater than the dividend.
Each positive test is followed by a negative test to show that we're not overstepping the boundaries of the known bits.
There are extra tests for the signed-min-value special cases.
The idea to make an 'isDivZero' helper was suggested for the signed case in D37713:
https://reviews.llvm.org/D37713
This clean-up makes it clear that D37713 is just filling the gap for signed div/rem,
removes unnecessary code, and allows us to remove a bit of duplicated code from the
planned improvement in D37713.
[Hexagon] Make getMemAccessSize return size in bytes
It used to return the actual field value from the instruction descriptor.
There is no reason for that, that value is not interesting in any way and
the specifics of its encoding in the descriptor should not be exposed.
[dwarfdump] Add DWARF verifiers for address ranges
This patch started as an attempt to rebase Greg's differential (D32821).
The result is both quite similar and different at the same time. It adds
the following checks:
- Verify that all address ranges in a DIE are valid.
- Verify that no ranges within the DIE overlap.
- Verify that no ranges overlap with the ranges of a sibling.
- Verify that children are completely contained in its (direct)
parent's address range. (unless both are subprograms)
Simon Dardis [Thu, 14 Sep 2017 10:58:00 +0000 (10:58 +0000)]
[mips] Pick the right variant of DINS upfront and enable target instruction verification
This patch complements D16810 "[mips] Make isel select the correct DEXT variant
up front.". Now ISel picks the right variant of DINS, so now there is no need
to replace DINS with the appropriate variant during
MipsMCCodeEmitter::encodeInstruction().
This patch also enables target specific instruction verification for ins, dins,
dinsm, dinsu, ext, dext, dextm, dextu. These instructions have constraints that
are checked when generating MipsISD::Ins and MipsISD::Ext nodes, but these
constraints are not checked during instruction selection. Adding machine
verification should catch outstanding cases.
Finally, correct a bug that instruction verification uncovered, where the
position operand of a DINSU generated during lowering was being silently
and accidently corrected to the correct value.
[dwarfdump] Add DWARF verifiers for address ranges
This patch started as an attempt to rebase Greg's differential (D32821).
The result is both quite similar and different at the same time. It adds
the following checks:
- Verify that all address ranges in a DIE are valid.
- Verify that no ranges within the DIE overlap.
- Verify that no ranges overlap with the ranges of a sibling.
- Verify that children are completely contained in its (direct)
parent's address range. (unless both are subprograms)
[PM/CGSCC] Teach the CGSCC pass manager components to gracefully handle
invalidated SCCs even when we do not have an updated SCC to redirect
towards.
This comes up in a fairly subtle and surprising circumstance: we need to
have a connected but internal node in the call graph which later becomes
a disconnected island, and then gets deleted. All of this needs to
happen mid-CGSCC walk. Because it is disconnected, we have no way of
computing a new "current" SCC when it gets deleted. Instead, we need to
explicitly check for a deleted "current" SCC and bail out of the current
CGSCC step. This will bubble all the way up to the post-order walk and
then resume correctly.
I've included minimal tests for this bug. The specific behavior
matches something we've seen in the wild with the new PM combined with
ThinLTO and sample PGO, but I've not yet confirmed whether this is the
only issue there.
projects / llvm / log