Chandler Carruth [Sat, 24 Jan 2015 10:18:47 +0000 (10:18 +0000)]
[PM] Move the LowerExpectIntrinsic pass to the Scalar library.
It was already in the Scalar header and referenced extensively as being
in this library, the source file was just in the utils directory for
some reason. No actual functionality changed. I noticed as it didn't
make sense to add a pass header to the utils headers.
Chandler Carruth [Sat, 24 Jan 2015 04:19:17 +0000 (04:19 +0000)]
[PM] Port instcombine to the new pass manager!
This is exciting as this is a much more involved port. This is
a complex, existing transformation pass. All of the core logic is shared
between both old and new pass managers. Only the access to the analyses
is separate because the actual techniques are separate. This also uses
a bunch of different and interesting analyses and is the first time
where we need to use an analysis across an IR layer.
This also paves the way to expose instcombine utility functions. I've
got a static function that implements the core pass logic over
a function which might be mildly interesting, but more interesting is
likely exposing a routine which just uses instructions *already in* the
worklist and combines until empty.
I've switched one of my favorite instcombine tests to run with both as
well to make sure this keeps working.
Chandler Carruth [Sat, 24 Jan 2015 02:06:09 +0000 (02:06 +0000)]
[PM] Rework how the TargetLibraryInfo pass integrates with the new pass
manager to support the actual uses of it. =]
When I ported instcombine to the new pass manager I discover that it
didn't work because TLI wasn't available in the right places. This is
a somewhat surprising and/or subtle aspect of the new pass manager
design that came up before but I think is useful to be reminded of:
While the new pass manager *allows* a function pass to query a module
analysis, it requires that the module analysis is already run and cached
prior to the function pass manager starting up, possibly with
a 'require<foo>' style utility in the pass pipeline. This is an
intentional hurdle because using a module analysis from a function pass
*requires* that the module analysis is run prior to entering the
function pass manager. Otherwise the other functions in the module could
be in who-knows-what state, etc.
A somewhat surprising consequence of this design decision (at least to
me) is that you have to design a function pass that leverages
a module analysis to do so as an optional feature. Even if that means
your function pass does no work in the absence of the module analysis,
you have to handle that possibility and remain conservatively correct.
This is a natural consequence of things being able to invalidate the
module analysis and us being unable to re-run it. And it's a generally
good thing because it lets us reorder passes arbitrarily without
breaking correctness, etc.
This ends up causing problems in one case. What if we have a module
analysis that is *definitionally* impossible to invalidate. In the
places this might come up, the analysis is usually also definitionally
trivial to run even while other transformation passes run on the module,
regardless of the state of anything. And so, it follows that it is
natural to have a hard requirement on such analyses from a function
pass.
It turns out, that TargetLibraryInfo is just such an analysis, and
InstCombine has a hard requirement on it.
The approach I've taken here is to produce an analysis that models this
flexibility by making it both a module and a function analysis. This
exposes the fact that it is in fact safe to compute at any point. We can
even make it a valid CGSCC analysis at some point if that is useful.
However, we don't want to have a copy of the actual target library info
state for each function! This state is specific to the triple. The
somewhat direct and blunt approach here is to turn TLI into a pimpl,
with the state and mutators in the implementation class and the query
routines primarily in the wrapper. Then the analysis can lazily
construct and cache the implementations, keyed on the triple, and
on-demand produce wrappers of them for each function.
One minor annoyance is that we will end up with a wrapper for each
function in the module. While this is a bit wasteful (one pointer per
function) it seems tolerable. And it has the advantage of ensuring that
we pay the absolute minimum synchronization cost to access this
information should we end up with a nice parallel function pass manager
in the future. We could look into trying to mark when analysis results
are especially cheap to recompute and more eagerly GC-ing the cached
results, or we could look at supporting a variant of analyses whose
results are specifically *not* cached and expected to just be used and
discarded by the consumer. Either way, these seem like incremental
enhancements that should happen when we start profiling the memory and
CPU usage of the new pass manager and not before.
The other minor annoyance is that if we end up using the TLI in both
a module pass and a function pass, those will be produced by two
separate analyses, and thus will point to separate copies of the
implementation state. While a minor issue, I dislike this and would like
to find a way to cleanly allow a single analysis instance to be used
across multiple IR unit managers. But I don't have a good solution to
this today, and I don't want to hold up all of the work waiting to come
up with one. This too seems like a reasonable thing to incrementally
improve later.
Reid Kleckner [Fri, 23 Jan 2015 22:55:31 +0000 (22:55 +0000)]
mips: Fix "XPASS" test results by removing 'not' commands
These tests are asserting and crashing for me, and 'not' sees that as a
non-zero exit code instead of a signal code for obscure Windows reasons.
This causes the test to pass, giving me an unclean 'ninja check'.
The test is already XFAILd, so just run the test without 'not' and let
lit handle the failure.
[x86] Combine x86mmx/i64 to v2i64 conversion to use scalar_to_vector
Handle the poor codegen for i64/x86xmm->v2i64 (%mm -> %xmm) moves. Instead of
using stack store/load pair to do the job, use scalar_to_vector directly, which
in the MMX case can use movq2dq. This was the current behavior prior to
improvements for vector legalization of extloads in r213897.
This commit fixes the regression and as a side-effect also remove some
unnecessary shuffles.
Tom Stellard [Fri, 23 Jan 2015 22:05:45 +0000 (22:05 +0000)]
R600/SI: Move i64 -> v2i32 load promotion into AMDGPUDAGToDAGISel::Select()
We used to do this promotion during DAG legalization, but this
caused an infinite loop in ExpandUnalignedLoad() because it assumed
that i64 loads were legal if i64 was a legal type.
It also seems better to report i64 loads as legal, since they actually
are and we were just promoting them to simplify our tablegen files.
Lang Hames [Fri, 23 Jan 2015 21:25:00 +0000 (21:25 +0000)]
[Orc] New JIT APIs.
This patch adds a new set of JIT APIs to LLVM. The aim of these new APIs is to
cleanly support a wider range of JIT use cases in LLVM, and encourage the
development and contribution of re-usable infrastructure for LLVM JIT use-cases.
These APIs are intended to live alongside the MCJIT APIs, and should not affect
existing clients.
Included in this patch:
1) New headers in include/llvm/ExecutionEngine/Orc that provide a set of
components for building JIT infrastructure.
Implementation code for these headers lives in lib/ExecutionEngine/Orc.
2) A prototype re-implementation of MCJIT (OrcMCJITReplacement) built out of the
new components.
3) Minor changes to RTDyldMemoryManager needed to support the new components.
These changes should not impact existing clients.
4) A new flag for lli, -use-orcmcjit, which will cause lli to use the
OrcMCJITReplacement class as its underlying execution engine, rather than
MCJIT itself.
Tests to follow shortly.
Special thanks to Michael Ilseman, Pete Cooper, David Blaikie, Eric Christopher,
Justin Bogner, and Jim Grosbach for extensive feedback and discussion.
Kevin Enderby [Fri, 23 Jan 2015 21:02:44 +0000 (21:02 +0000)]
Fix the problem with llvm-objdump and -archive-headers in printing the archive header size field.
This problem showed up with the clang-cmake-armv7-a15-full bot. Thanks to Renato Golin for his help.
Hans Wennborg [Fri, 23 Jan 2015 20:43:51 +0000 (20:43 +0000)]
LowerSwitch: replace unreachable default with popular case destination
SimplifyCFG currently does this transformation, but I'm planning to remove that
to allow other passes, such as this one, to exploit the unreachable default.
This patch takes care to keep track of what case values are unreachable even
after the transformation, allowing for more efficient lowering.
In llvm-mode, with electric-pair-mode turned on, typing a literal '['
would print out '[[', and '(' would print a '(('. This was a very
annoying bug caused by overzealous syntax-table entries: the parens are
already part of the '(' and ')' class by default. Fix this.
While at it, notice that i32, i64, i1 etc. are not font-locked despite a
clear intent to do so. The issue is that regexp-opt doesn't accept
regular expressions. So, spell out the common literal integers with
different widths.
Toma Tabacu [Fri, 23 Jan 2015 10:40:19 +0000 (10:40 +0000)]
[mips] Add new error message and improve testing for parsing the .module directive.
Summary:
We used to silently ignore any empty .module's and we used to give an error saying that we found
an "unexpected token at start of statement" when the value of the option wasn't an identifier (e.g. if it was a number).
We now give an error saying that we "expected .module option identifier" in both of those cases.
I also fixed the other tests in mips-abi-bad.s, which all seemed to be broken.
Mehdi Amini [Fri, 23 Jan 2015 07:07:20 +0000 (07:07 +0000)]
DAGCombine: always constant fold FMA when target disable FP exceptions
Summary: When trying to constant fold an FMA in the DAG, getNode()
fails to fold the FMA if an operand is not finite. In this case this
patch allows the constant folding if !TLI->hasFloatingPointExceptions()
These things are potentially used for non-DWARF data (see the discussion
in PR22235), so take the `Dwarf` out of the name. Since the new name
gives fewer clues, update the doxygen to properly describe what they
are.
Chandler Carruth [Thu, 22 Jan 2015 21:53:09 +0000 (21:53 +0000)]
[PM] Actually add the new pass manager support for the assumption cache.
I had already factored this analysis specifically to enable doing this,
but hadn't actually committed the necessary wiring to get at this from
the new pass manager. This also nicely shows how the separate cache
object can be directly managed by the new pass manager.
This analysis didn't have any direct tests and so I've added a printer
pass and a boring test case. I chose to print the i1 value which is
being assumed rather than the call to llvm.assume as that seems much
more useful for testing... but suggestions on an even better printing
strategy welcome. My main goal was to make sure things actually work. =]
IR: Update references to temporaries before deleting
During `MDNode::deleteTemporary()`, call `replaceAllUsesWith(nullptr)`
to update all tracking references to `nullptr`.
This fixes PR22280, where inverted destruction order between tracking
references and the temporaries themselves caused a use-after-free in
`LLParser`.
An alternative fix would be to add an assertion that there are no users,
and continue to fix inverted destruction order in clients (like
`LLParser`), but instead I decided to make getting-teardown-right easy.
(If someone disagrees let me know.)
Chris Bieneman [Thu, 22 Jan 2015 21:01:12 +0000 (21:01 +0000)]
Refactoring cl::parser construction and initialization.
Summary:
Some parsers need references back to the option they are members of. This is used for handling the argument string as well as by the various pass name parsers for making pass names into flags.
Making parsers that need to refer back to the option have a reference to the option eliminates some of the members of various parsers, and enables further code cleanup.
Sanjay Patel [Thu, 22 Jan 2015 18:21:26 +0000 (18:21 +0000)]
merge consecutive stores of extracted vector elements (PR21711)
This is a 2nd try at the same optimization as http://reviews.llvm.org/D6698.
That patch was checked in at r224611, but reverted at r225031 because it
caused a failure outside of the regression tests.
The cause of the crash was not recognizing consecutive stores that have mixed
source values (loads and vector element extracts), so this patch adds a check
to bail out if any store value is not coming from a vector element extract.
This patch also refactors the shared logic of the constant source and vector
extracted elements source cases into a helper function.
David Blaikie [Thu, 22 Jan 2015 17:49:59 +0000 (17:49 +0000)]
Revert "PR21408: Workaround the appearance of duplicate variables due to problems when inlining two calls to the same function from the same call site."
The underlying bug has been fixed in r226736 so there's no need to
workaround this anymore.
Tim Northover [Thu, 22 Jan 2015 17:23:04 +0000 (17:23 +0000)]
AArch64: decode all MRS/MSR forms early to avoid saving FeatureBits.
Currently, we're adding a uint64_t describing the current subtarget so
that matching can check whether the specified register is valid.
However, we want to move to a bitset for those bits (x86 has more than
64 of them).
This can't live in a union so it's probably better to do the checks
early (especially as there are only 3 of them).
Rafael Espindola [Thu, 22 Jan 2015 14:20:45 +0000 (14:20 +0000)]
[pr21886] Change MCJIT/ELF to support MSVC C++ mangled symbol.
The ELF format is used on Windows by the MCJIT engine. Thus, on Windows, the
ELFObjectWriter can encounter symbols mangled using the MS Visual Studio C++
name mangling. Symbols mangled using the MSVC C++ name mangling can legally
have "@@@" as a substring. The EFLObjectWriter should not interpret the "@@@"
substring as specifying GNU-style symbol versioning. The ELFObjectWriter
therefore check for the MSVC C++ name mangling prefix which is either "?", "@?",
"imp_?" or "imp_?@".
[DAGCombine] Produce better code for constant splats
This solves PR22276.
Splats of constants would sometimes produce redundant shuffles, sometimes ridiculously so (see the PR for details). Fold these shuffles into BUILD_VECTORs early on instead.
[DAGCombine] Produce better code for constant splats
This solves PR22276.
Splats of constants would sometimes produce redundant shuffles, sometimes ridiculously so (see the PR for details). Fold these shuffles into BUILD_VECTORs early on instead.
Fixed a bug in type legalizer for masked load/store intrinsics.
The problem occurs when after vectorization we have type
<2 x i32>. This type is promoted to <2 x i64> and then requires
additional efforts for expanding loads and truncating stores.
I added EXPAND / TRUNCATE attributes to the masked load/store
SDNodes. The code now contains additional shuffles.
I've prepared changes in the cost estimation for masked memory
operations, it will be submitted separately.
Fixed a bug in narrowing store operation.
Type MVT::i1 became legal in KNL, but store operation can't be narrowed to this type,
since the size of VT (1 bit) is not equal to its actual store size(8 bits).
Sanjoy Das [Thu, 22 Jan 2015 09:32:02 +0000 (09:32 +0000)]
[NFC] Introduce a 'struct Range' for IRCE
Use the struct instead of a std::pair<Value *, Value *>. This makes a
Range an obviously immutable object, and we can now assert that a
range is well-typed (Begin->getType() == End->getType()) on its
construction.
Sanjoy Das [Thu, 22 Jan 2015 08:29:18 +0000 (08:29 +0000)]
Fix crashes in IRCE caused by mismatched types
There are places where the inductive range check elimination pass
depends on two llvm::Values or llvm::SCEVs to be of the same
llvm::Type when they do not need to be. This patch relaxes those
restrictions (by bailing out of the optimization if the types
mismatch), and adds test cases to trigger those paths.
These issues were found by bootstrapping clang with IRCE running in
the -O3 pass ordering.
Erik Eckstein [Thu, 22 Jan 2015 08:20:51 +0000 (08:20 +0000)]
SLPVectorizer: add a second limit for the number of alias checks.
Even with the current limit on the number of alias checks, the containing loop has quadratic complexity.
This begins to hurt for blocks containing > 1K load/store instructions.
This commit introduces a limit for the loop count. It reduces the runtime for such very large blocks.
Chandler Carruth [Thu, 22 Jan 2015 05:25:13 +0000 (05:25 +0000)]
[PM] Rename InstCombine.h to InstCombineInternal.h in preparation for
creating a non-internal header file for the InstCombine pass.
I thought about calling this InstCombiner.h or in some way more clearly
associating it with the InstCombiner clas that it is primarily defining,
but there are several other utility interfaces defined within this for
InstCombine. If, in the course of refactoring, those end up moving
elsewhere or going away, it might make more sense to make this the
combiner's header alone.
Naturally, this is a bikeshed to a certain degree, so feel free to lobby
for a different shade of paint if this name just doesn't suit you.
Chandler Carruth [Thu, 22 Jan 2015 05:08:12 +0000 (05:08 +0000)]
[canonicalize] Teach InstCombine to canonicalize loads which are only
ever stored to always use a legal integer type if one is available.
Regardless of whether this particular type is good or bad, it ensures we
don't get weird differences in generated code (and resulting
performance) from "equivalent" patterns that happen to end up using
a slightly different type.
After some discussion on llvmdev it seems everyone generally likes this
canonicalization. However, there may be some parts of LLVM that handle
it poorly and need to be fixed. I have at least verified that this
doesn't impede GVN and instcombine's store-to-load forwarding powers in
any obvious cases. Subtle cases are exactly what we need te flush out if
they remain.
Also note that this IR pattern should already be hitting LLVM from Clang
at least because it is exactly the IR which would be produced if you
used memcpy to copy a pointer or floating point between memory instead
of a variable.
ARM: fail less catastrophically on invalid Windows input
Windows supports a restricted set of relocations (compared to ARM ELF). In some
cases, we may end up generating an unsupported relocation. This can occur with
bad input to the assembler in particular (the frontend should never generate
code that cannot be compiled). Generate an error rather than just aborting.
The change in the API is driven by the desire to provide a slightly more helpful
message for debugging purposes.
projects / llvm / log