Craig Topper [Sun, 8 Oct 2017 16:57:23 +0000 (16:57 +0000)]
[X86] Prefer MOVSS/SD over BLENDI during legalization. Remove BLENDI versions of scalar arithmetic patterns
Summary:
We currently disable some converting of shuffles to MOVSS/MOVSD during legalization if SSE41 is enabled. But later during shuffle combining we go back to prefering MOVSS/MOVSD.
Additionally we have patterns that look for BLENDIs to detect scalar arithmetic operations. I believe due to the combining using MOVSS/MOVSD these are unnecessary.
Interestingly, we still codegen blend instructions even though lowering/isel emit movss/movsd instructions. Turns out machine CSE commutes them to blend, and then commuting those blends back into blends that are equivalent to the original movss/movsd.
This patch fixes the inconsistency in legalization to prefer MOVSS/MOVSD. The one test change was caused by this change. The problem is that we have integer types and are mostly selecting integer instructions except for the shufps. This shufps forced the execution domain, but the vpblendw couldn't have its domain changed with a naive instruction swap. We could fix this by special casing VPBLENDW based on the immediate to widen the element type.
The rest of the patch is removing all the excess scalar patterns.
Long term we should probably add isel patterns to make MOVSS/MOVSD emit blends directly instead of relying on the double commute. We may also want to consider emitting movss/movsd for optsize. I also wonder if we should still use the VEX encoded blendi instructions even with AVX512. Blends have better throughput, and that may outweigh the register constraint.
Gadi Haber [Sun, 8 Oct 2017 12:52:54 +0000 (12:52 +0000)]
[X86][SKX] Adding the scheduling information for the SKX target.
Adding the scheduling information for the SkylakeServer (SKX) target.
This patch adds the instruction scheduling information for the SkylakeServer (SKX) architecture target by adding the file X86SchedSkylakeServer.td located under the X86 Target.
We used the scheduling information retrieved from the Skylake architects in order to create the file.
The scheduling information includes latency, number of micro-Ops and used ports by each SKL instruction.
The patch continues the scheduling replacement and insertion effort started with the SNB target in r310792, the HSW target in r311879 and the SkylakeClient (SKL) target in rL313613.
Please expect some performance fluctuations due to code alignment effects.
Ayman Musa [Sun, 8 Oct 2017 09:20:32 +0000 (09:20 +0000)]
[X86][TableGen] Recommitting the X86 memory folding tables TableGen backend while disabling it by default.
After the original commit ([[ https://reviews.llvm.org/rL304088 | rL304088 ]]) was reverted, a discussion in llvm-dev was opened on 'how to accomplish this task'.
In the discussion we concluded that the best way to achieve our goal (which is to automate the folding tables and remove the manually maintained tables) is:
# Commit the tablegen backend disabled by default.
# Proceed with an incremental updating of the manual tables - while checking the validity of each added entry.
# Repeat previous step until we reach a state where the generated and the manual tables are identical. Then we can safely remove the manual tables and include the generated tables instead.
# Schedule periodical (1 week/2 weeks/1 month) runs of the pass:
- if changes appear (new entries):
- make sure the entries are legal
- If they are not, mark them as illegal to folding
- Commit the changes (if there are any).
CMake flag added for this purpose is "X86_GEN_FOLD_TABLES". Building with this flags will run the pass and emit the X86GenFoldTables.inc file under build/lib/Target/X86/ directory which is a good reference for any developer who wants to take part in the effort of completing the current folding tables.
Ayman Musa [Sun, 8 Oct 2017 08:32:56 +0000 (08:32 +0000)]
[X86] Add new attribute to X86 instructions to enable marking them as "not memory foldable"
This attribute will be used in a tablegen backend that generated the X86 memory folding tables which will be added in a future pass.
Instructions with this attribute unset will be excluded from the full set of X86 instructions available for the pass.
[MachineOutliner] Disable outlining from LinkOnceODRs by default
Say you have two identical linkonceodr functions, one in M1 and one in M2.
Say that the outliner outlines A,B,C from one function, and D,E,F from another
function (where letters are instructions). Now those functions are not
identical, and cannot be deduped. Locally to M1 and M2, these outlining
choices would be good-- to the whole program, however, this might not be true!
To mitigate this, this commit makes it so that the outliner sees linkonceodr
functions as unsafe to outline from. It also adds a flag,
-enable-linkonceodr-outlining, which allows the user to specify that they
want to outline from such functions when they know what they're doing.
Changing this handles most code size regressions in the test suite caused by
competing with linker dedupe. It also doesn't have a huge impact on the code
size improvements from the outliner. There are 6 tests that regress > 5% from
outlining WITH linkonceodrs to outlining WITHOUT linkonceodrs. Overall, most
tests either improve or are not impacted.
Not outlined vs outlined without linkonceodrs:
https://hastebin.com/raw/qeguxavuda
Not outlined vs outlined with linkonceodrs:
https://hastebin.com/raw/edepoqoqic
Outlined with linkonceodrs vs outlined without linkonceodrs:
https://hastebin.com/raw/awiqifiheb
Numbers generated using compare.py with -m size.__text. Tests run for AArch64
with -Oz -mllvm -enable-machine-outliner -mno-red-zone.
Sanjay Patel [Fri, 6 Oct 2017 23:20:16 +0000 (23:20 +0000)]
[InstCombine] rename SimplifyDivRemOfSelect to be clearer, add comments, simplify code; NFCI
There's at least one bug here - this code can fail with vector types (PR34856).
It's also being called for FREM; I'm still trying to understand how that is valid.
[dwarfdump] Verify that unit type matches root DIE
This patch adds two new verifiers:
- It checks that the root DIE of a CU is actually a valid unit DIE.
(based on its tag)
- For DWARF5 which contains a unit type int he CU header, it checks that
this matches the type of the unit DIE.
Zachary Turner [Fri, 6 Oct 2017 22:05:15 +0000 (22:05 +0000)]
[llvm-rc] Implement escape sequences in .rc files.
This allows the escape sequences (\a, \n, \r, \t, \\, \x[0-9a-f]*,
\[0-7]*, "") to appear in .rc scripts. These are parsed and output in
the same way as it's done in original MS implementation.
The way these sequences are processed depends on the type of the
resource it resides in, and on whether the user declared the string to
be "wide" or "narrow".
I tried to maintain the maximum compatibility with the original tool
(and fail in some erroneous situations that are accepted by .rc).
However, there are some (extremely rare) cases where Microsoft tool
outputs nonsense. I found it infeasible to detect such casses.
Zachary Turner [Fri, 6 Oct 2017 21:30:55 +0000 (21:30 +0000)]
[llvm-rc] Serialize STRINGTABLE statements to .res file.
This allows llvm-rc to serialize STRINGTABLE resources.
These are output in an unusual way: we locate them at the end of the
file, and strings are merged into bundles of max 16 strings, depending
on their IDs, language, and characteristics.
Zachary Turner [Fri, 6 Oct 2017 21:25:44 +0000 (21:25 +0000)]
[llvm-rc] Serialize CURSOR and ICON resources to .res
This is part 6 of llvm-rc serialization.
This adds ability to output cursors and icons as resources.
Unfortunately, we can't just copy .cur or .ico files to output - as each
file might contain multiple images, each of them needs to be unpacked
and stored as a separate resource. This forces us to parse cursor and
icon contents. (Fortunately, these formats are pretty similar and can be
processed by mostly common code).
As test files are binary, here is a short explanation of .cur and .ico
files stored:
cursor.cur, cursor-8.cur, cursor-32.cur are sample correct cursor files,
differing in their bit depth.
icon-old.ico, icon-new.ico are sample correct icon files;
icon-png.ico is a sample correct icon file in PNG format (instead of
usual BMP);
cursor-eof.cur is an incorrect cursor file - this is cursor.cur with
some of its final bytes removed.
cursor-bad-offset.cur is an incorrect cursor file - image header states
that image data begins at offset 0xFFFFFFFF.
Sample correct cursors and icons were created by Nico Weber.
Patch by Marek Sokolowski
Differential Revision: https://reviews.llvm.org/D37878
Reid Kleckner [Fri, 6 Oct 2017 21:17:51 +0000 (21:17 +0000)]
Revert "Roll forward r314928"
This appears to be miscompiling Clang, as shown on two Windows bootstrap
bots:
http://lab.llvm.org:8011/builders/clang-x86-windows-msvc2015/builds/7611
http://lab.llvm.org:8011/builders/clang-x64-ninja-win7/builds/6870
Nothing else is in the blame list. Both emit errors on this valid code
in the Windows ucrt headers:
C:\...\ucrt\malloc.h:95:32: error: invalid operands to binary expression ('char *' and 'int')
_Ptr = (char*)_Ptr + _ALLOCA_S_MARKER_SIZE;
~~~~~~~~~~~ ^ ~~~~~~~~~~~~~~~~~~~~~
Zachary Turner [Fri, 6 Oct 2017 20:51:20 +0000 (20:51 +0000)]
[llvm-rc] Add optional serialization support for DIALOG(EX) resources.
This is part 5 of llvm-rc serialization support.
This allows DIALOG and DIALOGEX to serialize if dialog-specific optional
statements are provided. These are (as of now): CAPTION, FONT, and
STYLE.
Notably, FONT statement can take more than two arguments when describing
DIALOGEX resources (as in
msdn.microsoft.com/en-us/library/windows/desktop/aa381013.aspx). I made
some changes to the parser to reflect this fact.
Patch by Marek Sokolowski
Differential Revision: https://reviews.llvm.org/D37864
Adrian Prantl [Fri, 6 Oct 2017 20:24:34 +0000 (20:24 +0000)]
llvm-dwarfdump: Add an option to collect debug info quality metrics.
At the last LLVM dev meeting we had a debug info for optimized code
BoF session. In that session I presented some graphs that showed how
the quality of the debug info produced by LLVM changed over the last
couple of years. This is a cleaned up version of the patch I used to
collect the this data. It is implemented as an extension of
llvm-dwarfdump, adding a new --statistics option. The intended
use-case is to automatically run this on the debug info produced by,
e.g., our bots, to identify eyebrow-raising changes or regressions
introduced by new transformations that we could act on.
In the current form, two kinds of data are being collected:
- The number of variables that have a debug location versus the number
of variables in total (this takes into account inlined instances of
the same function, so if a variable is completely missing form only
one instance it will be found).
- The PC range covered by variable location descriptions versus the PC
range of all variables' containing lexical scopes.
The output format is versioned and extensible, so I'm looking forward
to both bug fixes and ideas for other data that would be interesting
to track.
Amara Emerson [Fri, 6 Oct 2017 19:24:15 +0000 (19:24 +0000)]
[GlobalISel] Fix legalizer trying to process a deleted instruction.
In some cases an instruction is deleted from the block during combining,
however it can still exist in the legalizer worklist.
This change modifies the combiner routines to add the given MI to the
dead instruction list rather than trying to remove it from the block
themselves. The responsibility is then on the caller to delete the instruction
from the block and any worklists.
Reid Kleckner [Fri, 6 Oct 2017 18:21:19 +0000 (18:21 +0000)]
[PEI] Remove required properties and use 'if' instead of std::function
Summary:
After r303360, we initialize UsesCalleeSaves in runOnMachineFunction,
which runs after getRequiredProperties. UsesCalleeSaves was initialized
to 'false', so getRequiredProperties would always return an empty set.
We don't have a TargetMachine available early anymore after r303360.
Just removing the requirement of NoVRegs seems to make things work, so
let's do that.
The bitcode reader looks specifically for `__DATA, __objc_catlist` as a
section name. However, SVN r304661 removed the spaces (the two names
are functionally equivalent but do not compare equally
lexicographically). This causes compatibility issues. Add an
auto-upgrade path for removing the spaces as well as use the new name in
the LTO plugin.
Zachary Turner [Fri, 6 Oct 2017 17:54:46 +0000 (17:54 +0000)]
[lit] Improve tool substitution in lit.
This addresses two sources of inconsistency in test configuration
files.
1. Substitution boundaries. Previously you would specify a
substitution, such as 'lli', and then additionally a set
of characters that should fail to match before and after
the tool. This was used, for example, so that matches that
are parts of full paths would not be replaced. But not all
tools did this, and those that did would often re-invent
the set of characters themselves, leading to inconsistency.
Now, every tool substitution defaults to using a sane set
of reasonable defaults and you have to explicitly opt out
of it. This actually fixed a few latent bugs that were
never being surfaced, but only on accident.
2. There was no standard way for the system to decide how to
locate a tool. Sometimes you have an explicit path, sometimes
we would search for it and build up a path ourselves, and
sometimes we would build up a full command line. Furthermore,
there was no standardized way to handle missing tools. Do we
warn, fail, ignore, etc? All of this is now encapsulated in
the ToolSubst class. You either specify an exact command to
run, or an instance of FindTool('<tool-name>') and everything
else just works. Furthermore, you can specify an action to
take if the tool cannot be resolved.
Francis Ricci [Fri, 6 Oct 2017 17:43:37 +0000 (17:43 +0000)]
[dsymutil] Emit valid debug locations when no symbol flags are set
Summary:
swiftc emits symbols without flags set, which led dsymutil to ignore
them when searching for global symbols, causing dwarf location data
to be omitted. Xcode's dsymutil handles this case correctly, and emits
valid location data. Add this functionality to llvm-dsymutil by
allowing parsing of symbols with no flags set.
Reid Kleckner [Fri, 6 Oct 2017 17:21:49 +0000 (17:21 +0000)]
[MC] Use unique_ptr to manage WinFrameInfos, NFC
The FrameInfo cannot be stored directly in the vector because chained
frames may refer to parent frames, so we need pointers that are stable
across a vector resize.
Support: Rewrite Windows implementation of sys::fs::rename to be more POSIXy.
The current implementation of rename uses ReplaceFile if the
destination file already exists. According to the documentation for
ReplaceFile, the source file is opened without a sharing mode. This
means that there is a short interval of time between when ReplaceFile
renames the file and when it closes the file during which the
destination file cannot be opened.
This behaviour is not POSIX compliant because rename is supposed
to be atomic. It was also causing intermittent link failures when
linking with a ThinLTO cache; the ThinLTO cache implementation expects
all cache files to be openable.
This patch addresses that problem by re-implementing rename
using CreateFile and SetFileInformationByHandle. It is roughly a
reimplementation of ReplaceFile with a better sharing policy as well
as support for renaming in the case where the destination file does
not exist.
This implementation is still not fully POSIX. Specifically in the case
where the destination file is open at the point when rename is called,
there will be a short interval of time during which the destination
file will not exist. It isn't clear whether it is possible to avoid
this using the Windows API.
Dehao Chen [Fri, 6 Oct 2017 17:04:55 +0000 (17:04 +0000)]
Directly return promoted direct call instead of rely on stripPointerCast.
Summary: stripPointerCast is not reliably returning the value that's being type-casted. Instead it may look further at function attributes to further propagate the value. Instead of relying on stripPOintercast, the more reliable solution is to directly use the pointer to the promoted direct call.
Francis Ricci [Fri, 6 Oct 2017 15:33:28 +0000 (15:33 +0000)]
[llvm-objdump] Add RAII for xar apis
Summary:
xar_open and xar_iter_new require manual calls to close/free functions
to deallocate resources. This makes it easy to introduce memory leaks,
so add RAII struct wrappers for these resources.
Francis Ricci [Fri, 6 Oct 2017 14:49:20 +0000 (14:49 +0000)]
[llvm-dsymutil] Add support for __swift_ast MachO DWARF section
Summary:
Xcode's dsymutil emits a __swift_ast DWARF section, which is required for debugging,
and which contains a byte-for-byte dump of the swiftmodule file.
Add this feature to llvm-dsymutil.
Tested with `gobjdump --dwarf=info -s`, by verifying that the contents of
`__DWARF.__swift_ast` match between Xcode's dsymutil and llvm-dsymutil
(Xcode's dwarfdump and llvm-dwarfdump don't currently recognize the
__swift_ast section).
Davide Italiano [Fri, 6 Oct 2017 04:39:40 +0000 (04:39 +0000)]
[PM] Retire disable unit-at-a-time switch.
This is a vestige from the GCC-3 days, which disables IPO passes
when set. I don't think anybody actually uses it as there are
several IPO passes which still run with this flag set and
nobody complained/noticed. This reduces the delta between
current and new pass manager and allows us to easily review
the difference when we decide to flip the switch (or audit
which passes should run, FWIW).
Jakub Kuderski [Fri, 6 Oct 2017 03:37:06 +0000 (03:37 +0000)]
[CodeExtractor] Fix multiple bugs under certain shape of extracted region
Summary:
If the extracted region has multiple exported data flows toward the same BB which is not included in the region, correct resotre instructions and PHI nodes won't be generated inside the exitStub. The solution is simply put the restore instructions right after the definition of output values instead of putting in exitStub.
Unittest for this bug is included.
ModuleUtils: Stop using comdat members to generate unique module ids.
It is possible for two modules to define the same set of external
symbols without causing a duplicate symbol error at link time,
as long as each of the symbols is a comdat member. So we cannot
use them as part of a unique id for the module.
Derek Schuff [Thu, 5 Oct 2017 21:18:42 +0000 (21:18 +0000)]
[WebAssembly] Add the rest of the atomic loads
Add extending loads and constant offset patterns
A bit more refactoring of the tablegen to make the patterns fairly nice and
uniform between the regular and atomic loads.
Sanjay Patel [Thu, 5 Oct 2017 21:11:49 +0000 (21:11 +0000)]
[InstCombine] improve folds for icmp gt/lt (shr X, C1), C2
We can always eliminate the shift in: icmp gt/lt (shr X, C1), C2 --> icmp gt/lt X, C'
This patch was supposed to just be an efficiency improvement because we were doing this 3-step process to fold:
Dehao Chen [Thu, 5 Oct 2017 20:15:29 +0000 (20:15 +0000)]
Annotate VP prof on indirect call if it is ICPed in the profiled binary.
Summary: In SamplePGO, when an indirect call is promoted in the profiled binary, before profile annotation, it will be promoted and inlined. For the original indirect call, the current implementation will not mark VP profile on it. This is an issue when profile becomes stale. This patch annotates VP prof on indirect calls during annotation.
Francis Ricci [Thu, 5 Oct 2017 20:03:01 +0000 (20:03 +0000)]
[llvm-dsymutil] Add support for __swift_ast MachO DWARF section
Summary:
Xcode's dsymutil emits a __swift_ast DWARF section, which is required for debugging,
and which contains a byte-for-byte dump of the swiftmodule file.
Add this feature to llvm-dsymutil.
Tested with `gobjdump --dwarf=info -s`, by verifying that the contents of
`__DWARF.__swift_ast` match between Xcode's dsymutil and llvm-dsymutil
(Xcode's dwarfdump and llvm-dwarfdump don't currently recognize the
__swift_ast section).
Francis Ricci [Thu, 5 Oct 2017 19:17:28 +0000 (19:17 +0000)]
[llvm-dsymutil] Add support for __swift_ast MachO DWARF section
Summary:
Xcode's dsymutil emits a __swift_ast DWARF section, which is required for debugging,
and which contains a byte-for-byte dump of the swiftmodule file.
Add this feature to llvm-dsymutil.
Tested with `gobjdump --dwarf=info -s`, by verifying that the contents of
`__DWARF.__swift_ast` match between Xcode's dsymutil and llvm-dsymutil
(Xcode's dwarfdump and llvm-dwarfdump don't currently recognize the
__swift_ast section).
Davide Italiano [Thu, 5 Oct 2017 18:36:01 +0000 (18:36 +0000)]
[NewPassManager] Run global dead code elimination after the inliner.
This is the same exact change we did for the current pass manager
in rL314997, but the new pass manager pipeline already happened
to run GlobalOpt after the inliner, so we just insert a run of
GDCE here.
Davide Italiano [Thu, 5 Oct 2017 18:06:37 +0000 (18:06 +0000)]
[PassManager] Run global optimizations after the inliner.
The inliner performs some kind of dead code elimination as it goes,
but there are cases that are not really caught by it. We might
at some point consider teaching the inliner about them, but it
is OK for now to run GlobalOpt + GlobalDCE in tandem as their
benefits generally outweight the cost, making the whole pipeline
faster.
Matthew Simpson [Thu, 5 Oct 2017 18:03:30 +0000 (18:03 +0000)]
[SparsePropagation] Move member definitions to header (NFC)
AbstractLatticeFunction and SparseSolver are class templates parameterized by a
lattice value, so we need to move these member functions over to the header.
Petar Jovanovic [Thu, 5 Oct 2017 17:40:32 +0000 (17:40 +0000)]
[mips] implement .set dspr2 directive
Implement .set dspr2 directive with appropriate feature bits. This
directive is a counterpart of -mattr=dspr2 command line option with the
exception that it does not influence elf header flags.
The old algoritm was not correct, although it worked most of the time.
Avoid the complex reachability analysis and simply calculate the maximal
registers out of the set of all referenced registers.
Rong Xu [Thu, 5 Oct 2017 17:05:20 +0000 (17:05 +0000)]
[ProfileData] Fix data racing in merging indexed profiles
There is data racing to the static variable RecordIndex in index profile reader
when merging in multiple threads. Make it a member variable in
IndexedInstrProfReader to fix this.
Artur Pilipenko [Thu, 5 Oct 2017 16:28:21 +0000 (16:28 +0000)]
[X86] Fix chains update when lowering BUILD_VECTOR to a vector load
The code which lowers BUILD_VECTOR of consecutive loads into a single vector
load doesn't update chains properly. As a result the vector load can be
reordered with the store to the same location.
The current code in EltsFromConsecutiveLoads only updates the chain following
the first load. The fix is to update the chains following all the loads
comprising the vector.
projects / llvm / log