Daniel Sanders [Fri, 6 Jan 2017 14:29:34 +0000 (14:29 +0000)]
[globalisel] Stop requiring -debug/-debug-only=registerbankinfo for assertions.
Summary:
I've noticed that these assertions don't trigger when the condition is false.
The problem is that the DEBUG(x) macro only executes x when the pass is
emitting debug output via the -debug and -debug-only=registerbankinfo command
line arguments.
Debug builds should always execute the assertions so use '#ifndef NDEBUG' instead.
Also removed an assertion that is only true the first time it's tested. <Target>RegisterBankInfo's constructor will re-use register banks causing them to be valid on subsequent tests. That
assertion will fail on the first test too in the near future.
Michal Gorny [Fri, 6 Jan 2017 08:23:33 +0000 (08:23 +0000)]
[llvm-config] Add --cmakedir to obtain CMake module location
Add a --cmakedir option to llvm-config that returns the correct path to
built/installed CMake modules (i.e. lib/cmake/llvm). This is mostly
intended as a convenience option for stand-alone builds of other LLVM
projects that frequently reconstruct LLVM_CMAKE_PATH after querying
llvm-config.
Lang Hames [Fri, 6 Jan 2017 06:22:31 +0000 (06:22 +0000)]
[Orc][RPC] Fix an obvious locking-order bug in RawByteChannel::startSendMessage.
The lock needs to be acquired before the data is sent, not afterwards. This
think-o slipped in during the refactor in r286620, but went unnoticed as the
resulting bug only manifests in multi-threaded clients (of which there are none
in-tree).
No unit test as the bug depends on thread scheduling.
Rui Ueyama [Fri, 6 Jan 2017 05:33:45 +0000 (05:33 +0000)]
TarWriter: Emit PAX headers only when needed.
We use PAX headers to store long filenames (>= 100 bytes).
It is not needed to emit PAX headers if filenames fit in the
Ustar header. This patch implements that optimization.
Rui Ueyama [Fri, 6 Jan 2017 02:29:48 +0000 (02:29 +0000)]
Add a class to create a tar archive file.
In LLD, we create cpio archive files for --reproduce command.
cpio was not a bad choice because it is very easy to create, but
it was sometimes hard to use because people are not familiar with
cpio command.
I noticed that creating a tar archive isn't as hard as I thought.
So I implemented it in this patch.
After making this change, I discovered that _Unwind_Backtrace is
unable to unwind past a signal handler after an assertion failure.
I filed a bug report about that issue in rdar://29866587 but even if
we get a fix soon, it will be awhile before it get released.
Bob Wilson [Fri, 6 Jan 2017 02:26:30 +0000 (02:26 +0000)]
Disable sigaltstack on Apple platforms
Using sigaltstack on Apple platforms is a bad idea. Darwin's backtrace()
function does not work with sigaltstack, and my change in r286851 was
supposed to solve that by using _Unwind_Backtrace instead. I tested that
_Unwind_Backtrace works for crashes but then discovered that it does not
work for assertion failures when using sigaltstack, at least on macOS.
The stack trace shows only the frames on the alternate stack.
I also saw some reports of this happening for crashes, but it fails
consistently for assertion failures. I tried various things to get it to
work but the problem seems to be in _Unwind_Backtrace itself. Disabling
sigaltstack is unfortunate since it would be nice to get backtraces for
stack overflows, but at least this gets us backtraces for the more common
cases. rdar://problem/29662459
LowerTypeTests: Split the pass in two: a resolution phase and a lowering phase.
This change separates how type identifiers are resolved from how intrinsic
calls are lowered. All information required to lower an intrinsic call
is stored in a new TypeIdLowering data structure. The idea is that this
data structure can either be initialized using the module itself during
regular LTO, or using the module summary in ThinLTO backends.
David Majnemer [Fri, 6 Jan 2017 00:11:46 +0000 (00:11 +0000)]
[SelectionDAG] Correctly transform range metadata to AssertZExt
We used the logBase2 of the high instead of the ceilLogBase2 resulting
in the wrong result for certain values. For example, it resulted in an
i1 AssertZExt when the exclusive portion of the range was 3.
Teresa Johnson [Thu, 5 Jan 2017 21:34:18 +0000 (21:34 +0000)]
ThinLTO: add early "dead-stripping" on the Index
Summary:
Using the linker-supplied list of "preserved" symbols, we can compute
the list of "dead" symbols, i.e. the one that are not reachable from
a "preserved" symbol transitively on the reference graph.
Right now we are using this information to mark these functions as
non-eligible for import.
The impact is two folds:
- Reduction of compile time: we don't import these functions anywhere
or import the function these symbols are calling.
- The limited number of import/export leads to better internalization.
PR 31534: When emitting both DWARF unwind tables and debug information,
do not use .cfi_sections. This requires checking if any non-declaration
function in the module needs an unwind table.
[LICM] Allow promotion of some stores that are not guaranteed to execute.
Promotion is always legal when a store within the loop is guaranteed to execute.
However, this is not a necessary condition - for promotion to be memory model
semantics-preserving, it is enough to have a store that dominates every exit
block. This is because if the store dominates every exit block, the fact the
exit block was executed implies the original store was executed as well.
Matthias Braun [Thu, 5 Jan 2017 20:01:19 +0000 (20:01 +0000)]
CodeGen: Assert that liveness is up to date when reading block live-ins.
Add an assert that checks whether liveins are up to date before they are
used.
- Do not print liveins into .mir files anymore in situations where they
are out of date anyway.
- The assert in the RegisterScavenger is superseded by the new one in
livein_begin().
- Skip parts of the liveness updating logic in IfConversion.cpp when
liveness isn't tracked anymore (just enough to avoid hitting the new
assert()).
lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp:1638:12: error: could not convert ‘(const unsigned int*)(& Variants)’ from ‘const unsigned int*’ to ‘llvm::ArrayRef<unsigned int>’
return Variants;
Zvi Rackover [Thu, 5 Jan 2017 15:11:43 +0000 (15:11 +0000)]
[X86] Optimize vector shifts with variable but uniform shift amounts
Summary:
For instructions such as PSLLW/PSLLD/PSLLQ a variable shift amount may be passed in an XMM register.
The lower 64-bits of the register are evaluated to determine the shift amount.
This patch improves the construction of the vector containing the shift amount.
Teresa Johnson [Thu, 5 Jan 2017 14:59:56 +0000 (14:59 +0000)]
[ThinLTO] Use DenseSet instead of SmallPtrSet for holding GUIDs
Should fix some more bot failures from r291108.
This should have been a DenseSet, since GUID is not a pointer type.
It caused some bots to fail, but for some reason I wasnt't getting a
build failure.
Teresa Johnson [Thu, 5 Jan 2017 14:32:16 +0000 (14:32 +0000)]
[ThinLTO] Subsume all importing checks into a single flag
Summary:
This adds a new summary flag NotEligibleToImport that subsumes
several existing flags (NoRename, HasInlineAsmMaybeReferencingInternal
and IsNotViableToInline). It also subsumes the checking of references
on the summary that was being done during the thin link by
eligibleForImport() for each candidate. It is much more efficient to
do that checking once during the per-module summary build and record
it in the summary.
Currently isLikelyComplexAddressComputation tries to figure out if the given stride seems to be 'complex' and need some extra cost for address computation handling.
This code seems to be target dependent which may not be the same for all targets.
Passed the decision whether the given stride is complex or not to the target by sending stride information via SCEV to getAddressComputationCost instead of 'IsComplex'.
Specifically at X86 targets we dont see any significant address computation cost in case of the strided access in general.
Kristof Beyls [Thu, 5 Jan 2017 13:27:52 +0000 (13:27 +0000)]
[GlobalISel] Add support for address-taken basic blocks
To make this work, pointers from the MachineBasicBlock to the LLVM-IR-level
basic blocks need to be initialized, as the AsmPrinter uses this link to be
able to print out labels for the basic blocks that are address-taken.
Most of the changes in this commit are about adapting existing tests to include
the basic block name that is now printed out in the MIR format, now that the
name becomes available as the link to the LLVM-IR basic block is initialized.
The relevant test change for the functionality added in this patch are the
added "(address-taken)" strings in
test/CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll.
Kristof Beyls [Thu, 5 Jan 2017 11:28:51 +0000 (11:28 +0000)]
[GlobalISel] Add support for switch statements
This commit does this using a trivial chain of conditional branches. In the
future, we probably want to reuse the optimized switch lowering used in
SelectionDAG.
Craig Topper [Thu, 5 Jan 2017 05:47:29 +0000 (05:47 +0000)]
[X86] Change getHostCPUName to report Intel model 0x4e as "skylake" instead of "skylake-avx512". Add the proper 0x55 model for "skylake-avx512".
Summary:
Intel's i5-6300U CPU is reporting to have a model id of 78 (4e).
The Host detection assumes that to be Skylake Xeon (with AVX512 support),
instead of a normal Skylake machine.
IR: Module summary representation for type identifiers; summary test scaffolding for lowertypetests.
Set up basic YAML I/O support for module summaries, plumb the summary into
the pass and add a few command line flags to test YAML I/O support. Bitcode
support to come separately, as will the code in LowerTypeTests that actually
uses the summary. Also add a couple of tests that pass by virtue of the pass
doing nothing with the summary (which happens to be the correct thing to do
for those tests).
[gtest] Work around broken installs of libc++ where we don't have
a cxxabi.h in the include search paths.
This comes up when libc++ is installed with some other abi library. At
some points in time in history we have had CMake hackery to try and get
a cxxabi.h installed that would work, but there are lots of examples
lacking this. Also, the just-built tree with libc++ seems to not quite
get this right.
To let folks make progress, we can easily work around this by detecting
that the header is missing and disabling the relevant parts of gtest.
This should fix the last remainging build bot failures. While these
failures are typically indicative of a questionable install, I don't
think gtest should be the thing that surfaces those issues and I don't
want folks blocked on this.
Justin Lebar [Thu, 5 Jan 2017 00:12:51 +0000 (00:12 +0000)]
[PM] Edit comments in PassManager.h.
Summary:
This covers most of PassManager.h, up to the introduction of inner/outer
analysis proxies.
If there's a theme to these changes, it's simplifying the language. For
example:
* PreservedAnalyses is a "set of analyses", not an "abstract set".
"Abstract" doesn't have any particular meaning here.
* "Build types for the concept types" becomes "define the concept types".
* Instead of "data structures optimized for pointer-like types using
the alignment-provided low bits", say "data structures that use the
low bits of pointers."
* "Clear the map pointing into the results list" becomes
"Delete the map entries that point into the results list."
This patch also fixes a few places where we referred to "function" and
"module" pass/analysis managers, instead of the more abstract "IRUnitT"
PM/AMs we have now.
Wolfgang Pieb [Wed, 4 Jan 2017 23:58:26 +0000 (23:58 +0000)]
[DWARF] Null out the debug locs of load instructions that have been moved by GVN
performing partial redundancy elimination (PRE). Not doing so can cause jumpy line
tables and confusing (though correct) source attributions.
[gtest] Fix the way we disable a warning for unittests.
I somehow wrote this fix and then lost it prior to commit. Really sorry
about the noise. This should fix some issues with hacking add_definition
to do things with warning flags.
Mehdi Amini [Wed, 4 Jan 2017 22:54:33 +0000 (22:54 +0000)]
Use lazy-loading of Metadata in MetadataLoader when importing is enabled (NFC)
Summary:
This is a relatively simple scheme: we use the index emitted in the
bitcode to avoid loading all the global metadata. Instead we load
the index with their position in the bitcode so that we can load each
of them individually. Materializing the global metadata block in this
condition only triggers loading the named metadata, and the ones
referenced from there (transitively). When materializing a function,
metadata from the global block are loaded lazily as they are
referenced.
Two main current limitations are:
1) Global values other than functions are not materialized on demand,
so we need to eagerly load METADATA_GLOBAL_DECL_ATTACHMENT records
(and their transitive dependencies).
2) When we load a single metadata, we don't recurse on the operands,
instead we use a placeholder or a temporary metadata. Unfortunately
tepmorary nodes are very expensive. This is why we don't have it
always enabled and only for importing.
These two limitations can be lifted in a subsequent improvement if
needed.
With this change, the total link time of opt with ThinLTO and Debug
Info enabled is going down from 282s to 224s (~20%).
David Blaikie [Wed, 4 Jan 2017 22:36:33 +0000 (22:36 +0000)]
Reapply "Make BitCodeAbbrev ownership explicit using shared_ptr rather than IntrusiveRefCntPtr""
If this is a problem for anyone (shared_ptr is two pointers in size,
whereas IntrusiveRefCntPtr is 1 - and the ref count control block that
make_shared adds is probably larger than the one int in RefCountedBase)
I'd prefer to address this by adding a lower-overhead version of
shared_ptr (possibly refactoring IntrusiveRefCntPtr into such a thing)
to avoid the intrusiveness - this allows memory ownership to remain
orthogonal to types and at least to me, seems to make code easier to
understand (since no implicit ownership acquisition can happen).
Tim Shen [Wed, 4 Jan 2017 22:11:42 +0000 (22:11 +0000)]
[Legalizer] Fix fp-to-uint to fp-tosint promotion assertion.
Summary:
When promoting fp-to-uint16 to fp-to-sint32, the result is actually zero
extended. For example, given double 65534.0, without legalization:
fp-to-uint16: 65534.0 -> 0xfffe
With the legalization:
fp-to-sint32: 65534.0 -> 0x0000fffe
Without this patch, legalization wrongly emits a signed extend assertion,
which is consumed by later icmp instruction, and cause miscompile.
Note that the floating point value must be in [0, 65535), otherwise the
behavior is undefined.
This patch reverts r279223 behavior and adds more tests and
documentations.
In PR29041's context, James Molloy mentioned that:
We don't need to mask because conversion from float->uint8_t is
undefined if the integer part of the float value is not representable in
uint8_t. Therefore we can assume this doesn't happen!
which is totally true and good, because fptoui is documented clearly to
have undefined behavior when overflow/underflow happens. We should take
the advantage of this behavior so that we can save unnecessary mask
instructions.
David Blaikie [Wed, 4 Jan 2017 21:13:35 +0000 (21:13 +0000)]
Make BitCodeAbbrev ownership explicit using shared_ptr rather than IntrusiveRefCntPtr
If this is a problem for anyone (shared_ptr is two pointers in size,
whereas IntrusiveRefCntPtr is 1 - and the ref count control block that
make_shared adds is probably larger than the one int in RefCountedBase)
I'd prefer to address this by adding a lower-overhead version of
shared_ptr (possibly refactoring IntrusiveRefCntPtr into such a thing)
to avoid the intrusiveness - this allows memory ownership to remain
orthogonal to types and at least to me, seems to make code easier to
understand (since no implicit ownership acquisition can happen).
projects / llvm / log