platform/upstream/llvm.git
9 years agoCOFF: Fix /include. Included symbols are GC-roots.
Rui Ueyama [Mon, 1 Jun 2015 03:42:54 +0000 (03:42 +0000)]
COFF: Fix /include. Included symbols are GC-roots.

llvm-svn: 238717

9 years ago[opaque pointer type] Explicitly store the pointee type of the result of a GEP
David Blaikie [Mon, 1 Jun 2015 03:09:34 +0000 (03:09 +0000)]
[opaque pointer type] Explicitly store the pointee type of the result of a GEP

Alternatively, this type could be derived on-demand whenever
getResultElementType is called - if someone thinks that's the better
choice (simple time/space tradeoff), I'm happy to give it a go.

llvm-svn: 238716

9 years agoApply name change to CMake build system.
Jonathan Peyton [Mon, 1 Jun 2015 03:05:13 +0000 (03:05 +0000)]
Apply name change to CMake build system.

This change has the CMake build system create a dynamic library named
libomp instead of libiomp5.  Also any reference to libiomp is replaced
with libomp.  One can still use the LIBOMP_LIB_NAME variable to enforce
a different name, and everything will still work as expected.  An important
note is that libiomp5 and libgomp symlinks are created at install time when
on Unix systems.  On Windows, copies are created with the legacy names.

llvm-svn: 238715

9 years agoCOFF: Define an error category for the linker.
Rui Ueyama [Mon, 1 Jun 2015 02:58:15 +0000 (02:58 +0000)]
COFF: Define an error category for the linker.

Instead of returning non-categorized errors, return categorized errors.
All uses of make_dynamic_error_code are removed.

Because we don't have error reporting mechanism, I just chose to print out
error messages to stderr, and then return an error object. Not sure if
that's the right thing to do, but at least it seems practical.

http://reviews.llvm.org/D10129

llvm-svn: 238714

9 years agoApply name change to tools/* files.
Jonathan Peyton [Mon, 1 Jun 2015 02:41:44 +0000 (02:41 +0000)]
Apply name change to tools/* files.

These changes are all in comments.  Change libiomp references to
libomp.

llvm-svn: 238713

9 years agoApply name change to src/* files.
Jonathan Peyton [Mon, 1 Jun 2015 02:37:28 +0000 (02:37 +0000)]
Apply name change to src/* files.

These changes are mostly in comments, but there are a few
that aren't.  Change libiomp5 => libomp everywhere.  One internal
function name is changed in kmp_gsupport.c, and in kmp_i18n.c, the
static char[] variable 'name' is changed to "libomp".

llvm-svn: 238712

9 years agoApply name change to doxygen files.
Jonathan Peyton [Mon, 1 Jun 2015 02:32:03 +0000 (02:32 +0000)]
Apply name change to doxygen files.

This change only affects the doxygen file that contains a
reference to libiomp5.  change libiomp5 to libomp.

llvm-svn: 238711

9 years agoAdding libomp.rc.var for name change
Jonathan Peyton [Mon, 1 Jun 2015 02:26:08 +0000 (02:26 +0000)]
Adding libomp.rc.var for name change

Adding a new file libomp.rc.var in anticipation of more
name change commits.  This file is a copy of libiomp.rc.var.

llvm-svn: 238710

9 years agoTry to fix the build of IntelJITEventListener.
Rafael Espindola [Mon, 1 Jun 2015 02:18:14 +0000 (02:18 +0000)]
Try to fix the build of IntelJITEventListener.

llvm-svn: 238709

9 years agoRename HasData to IsRegistered.
Rafael Espindola [Mon, 1 Jun 2015 01:52:18 +0000 (01:52 +0000)]
Rename HasData to IsRegistered.

There is no MCSectionData, so the old name is now meaningless.

Also remove some asserts/checks that were there just because the information
they used was in MCSectionData.

llvm-svn: 238708

9 years agoRemove trivial forwarding function.
Rafael Espindola [Mon, 1 Jun 2015 01:39:15 +0000 (01:39 +0000)]
Remove trivial forwarding function.

llvm-svn: 238707

9 years agoStore a bit in MCSection saying if it was registered with MCAssembler.
Rafael Espindola [Mon, 1 Jun 2015 01:30:01 +0000 (01:30 +0000)]
Store a bit in MCSection saying if it was registered with MCAssembler.

With this we can replace a SetVector with a plain std::vector.

llvm-svn: 238706

9 years agoUse a bitfield. NFC.
Rafael Espindola [Mon, 1 Jun 2015 01:05:07 +0000 (01:05 +0000)]
Use a bitfield. NFC.

llvm-svn: 238705

9 years agoUse a 32 bit field for the symbol index.
Rafael Espindola [Mon, 1 Jun 2015 00:58:31 +0000 (00:58 +0000)]
Use a 32 bit field for the symbol index.

Even 64 ELF uses a 32 bit field to refer to symbols.

llvm-svn: 238704

9 years agoSimplify another function that doesn't fail.
Rafael Espindola [Mon, 1 Jun 2015 00:27:26 +0000 (00:27 +0000)]
Simplify another function that doesn't fail.

llvm-svn: 238703

9 years ago[PHITransAddr] Don't translate unreachable values
David Majnemer [Mon, 1 Jun 2015 00:15:08 +0000 (00:15 +0000)]
[PHITransAddr] Don't translate unreachable values

Unreachable values may use themselves in strange ways due to their
dominance property.  Attempting to translate through them can lead to
infinite recursion, crashing LLVM.  Instead, claim that we weren't able
to translate the value.

This fixes PR23096.

llvm-svn: 238702

9 years ago[PHITransAddr] Use std::find instead of std::count
David Majnemer [Mon, 1 Jun 2015 00:15:04 +0000 (00:15 +0000)]
[PHITransAddr] Use std::find instead of std::count

There is no need to visit all the elements if we are merely performing a
membership check.  NFCI.

llvm-svn: 238701

9 years agoSimplify interface of function that doesn't fail.
Rafael Espindola [Sun, 31 May 2015 23:52:50 +0000 (23:52 +0000)]
Simplify interface of function that doesn't fail.

llvm-svn: 238700

9 years ago[DWARF] Fix a bug in line info handling
Keno Fischer [Sun, 31 May 2015 23:37:04 +0000 (23:37 +0000)]
[DWARF] Fix a bug in line info handling

This fixes a bug in the line info handling in the dwarf code, based on a
problem I when implementing RelocVisitor support for MachO.
Since addr+size will give the first address past the end of the function,
we need to back up one line table entry. Fix this by looking up the
end_addr-1, which is the last address in the range. Note that this also
removes a duplicate output from the llvm-rtdyld line table dump. The
relevant line is the end_sequence one in the line table and has an offset
of the first address part the end of the range and hence should not be
included.
Also factor out the common functionality into a separate function.
This comes up on MachO much more than on ELF, since MachO
doesn't store the symbol size separately, hence making
said situation always occur.

Differential Revision: http://reviews.llvm.org/D9925

llvm-svn: 238699

9 years agoFor COFF and MachO, compute the gap between to symbols.
Rafael Espindola [Sun, 31 May 2015 23:15:35 +0000 (23:15 +0000)]
For COFF and MachO, compute the gap between to symbols.

Before r238028 we used to do this in O(N^2), now we do it in O(N log N).

llvm-svn: 238698

9 years agoARMConstantIslandPass.cpp: Prune an empty \brief. [-Wdocumentation]
NAKAMURA Takumi [Sun, 31 May 2015 23:05:35 +0000 (23:05 +0000)]
ARMConstantIslandPass.cpp: Prune an empty \brief.  [-Wdocumentation]

llvm-svn: 238697

9 years agoCOFF: Add /alternatename option.
Rui Ueyama [Sun, 31 May 2015 22:31:31 +0000 (22:31 +0000)]
COFF: Add /alternatename option.

Previously, this feature was implemented using a special type of
undefined symbol, in addition to an intricate way to make the resolver
read a virtual file containing that renaming symbols.

Now the feature is directly handled by the symbol table.
The symbol table has a function, rename(), to rename symbols, whose
definition is 4 lines long. Symbol renaming is naturally modeled using
Symbol and SymbolBody.

llvm-svn: 238696

9 years ago[Hexagon] Including raw_ostream for debug builds.
Colin LeMahieu [Sun, 31 May 2015 22:29:33 +0000 (22:29 +0000)]
[Hexagon] Including raw_ostream for debug builds.

llvm-svn: 238695

9 years ago[Hexagon] classes are actually structs.
Colin LeMahieu [Sun, 31 May 2015 22:18:42 +0000 (22:18 +0000)]
[Hexagon] classes are actually structs.

llvm-svn: 238694

9 years agoUse a range loop. NFC.
Rafael Espindola [Sun, 31 May 2015 22:13:51 +0000 (22:13 +0000)]
Use a range loop. NFC.

llvm-svn: 238693

9 years ago[Hexagon] Adding MC packet shuffler.
Colin LeMahieu [Sun, 31 May 2015 21:57:09 +0000 (21:57 +0000)]
[Hexagon] Adding MC packet shuffler.

llvm-svn: 238692

9 years agoCOFF: Detect file type by file magic.
Rui Ueyama [Sun, 31 May 2015 21:17:10 +0000 (21:17 +0000)]
COFF: Detect file type by file magic.

llvm-svn: 238691

9 years agoCOFF: Make the Driver own all MemoryBuffers. NFC.
Rui Ueyama [Sun, 31 May 2015 21:04:56 +0000 (21:04 +0000)]
COFF: Make the Driver own all MemoryBuffers. NFC.

Previously, a MemoryBuffer of a file was owned by each InputFile object.
This patch makes the Driver own all of them. InputFiles now have only
MemoryBufferRefs. This change simplifies ownership managment
(particularly for ObjectFile -- the object owned a MemoryBuffer only when
it's not created from an archive file, because in that case a parent
archive file owned the entire buffer. Now it owns nothing unconditionally.)

llvm-svn: 238690

9 years ago[Mips] Add a couple of MipsAbiInfoHandler functions to check linked code type
Simon Atanasyan [Sun, 31 May 2015 20:37:22 +0000 (20:37 +0000)]
[Mips] Add a couple of MipsAbiInfoHandler functions to check linked code type

No functional changes.

llvm-svn: 238689

9 years ago[Mips] Do not put the .reginfo section into the separate segment
Simon Atanasyan [Sun, 31 May 2015 20:37:13 +0000 (20:37 +0000)]
[Mips] Do not put the .reginfo section into the separate segment

The .reginfo should not belong to the separate segment if there is
a .MIPS.abiflags section.

llvm-svn: 238688

9 years ago[Mips] Sort segments so PT_MIPS_ABIFLAGS goes right after the PT_INTERP
Simon Atanasyan [Sun, 31 May 2015 20:36:58 +0000 (20:36 +0000)]
[Mips] Sort segments so PT_MIPS_ABIFLAGS goes right after the PT_INTERP

llvm-svn: 238687

9 years ago[Mips] Reading, merging and writing .MIPS.abiflags section
Simon Atanasyan [Sun, 31 May 2015 20:36:43 +0000 (20:36 +0000)]
[Mips] Reading, merging and writing .MIPS.abiflags section

http://dmz-portal.mips.com/wiki/MIPS_O32_ABI_-_FR0_and_FR1_Interlinking

llvm-svn: 238686

9 years ago[Mips] Delete MipsReginfo structure. Use the Elf_Mips_RegInfo instead.
Simon Atanasyan [Sun, 31 May 2015 20:36:21 +0000 (20:36 +0000)]
[Mips] Delete MipsReginfo structure. Use the Elf_Mips_RegInfo instead.

llvm-svn: 238685

9 years ago[Mips] Collect all ABI related info in the single MipsAbiInfoHandler class
Simon Atanasyan [Sun, 31 May 2015 20:36:11 +0000 (20:36 +0000)]
[Mips] Collect all ABI related info in the single MipsAbiInfoHandler class

New MipsAbiInfoHandler merges and hold both ELF header flags
and registries usage masks. In the future commits it will manage some
additional information.

llvm-svn: 238684

9 years agoCOFF: /libpath should not take precedence over the current directory.
Rui Ueyama [Sun, 31 May 2015 20:20:37 +0000 (20:20 +0000)]
COFF: /libpath should not take precedence over the current directory.

llvm-svn: 238683

9 years agoCOFF: Add /libpath option.
Rui Ueyama [Sun, 31 May 2015 20:10:11 +0000 (20:10 +0000)]
COFF: Add /libpath option.

llvm-svn: 238682

9 years agoCOFF: Add /include option.
Rui Ueyama [Sun, 31 May 2015 19:55:40 +0000 (19:55 +0000)]
COFF: Add /include option.

It does not involve notions of virtual archives or virtual files,
nor store a list of undefined symbols somewhere else to consume them later.
We did that before. In this patch, undefined symbols are just added to
the symbol table, which now can be done in very few lines of code.

llvm-svn: 238681

9 years agoARM: recommit r237590: allow jump tables to be placed as constant islands.
Tim Northover [Sun, 31 May 2015 19:22:07 +0000 (19:22 +0000)]
ARM: recommit r237590: allow jump tables to be placed as constant islands.

The original version didn't properly account for the base register
being modified before the final jump, so caused miscompilations in
Chromium and LLVM. I've fixed this and tested with an LLVM self-host
(I don't have the means to build & test Chromium).

The general idea remains the same: in pathological cases jump tables
can be too far away from the instructions referencing them (like other
constants) so they need to be movable.

Should fix PR23627.

llvm-svn: 238680

9 years agoCOFF: Add /nodefaultlib option.
Rui Ueyama [Sun, 31 May 2015 19:17:14 +0000 (19:17 +0000)]
COFF: Add /nodefaultlib option.

llvm-svn: 238679

9 years agoCOFF: Refactor functions to find files from search paths.
Rui Ueyama [Sun, 31 May 2015 19:17:12 +0000 (19:17 +0000)]
COFF: Refactor functions to find files from search paths.

llvm-svn: 238678

9 years agoCOFF: Create LinkerDriver class.
Rui Ueyama [Sun, 31 May 2015 19:17:09 +0000 (19:17 +0000)]
COFF: Create LinkerDriver class.

Previously the main linker routine is just a non-member function.
We store some context information to the Config object.
This patch makes it belong to Driver.

llvm-svn: 238677

9 years ago[MC] Simplify code. No functionality change intended.
Benjamin Kramer [Sun, 31 May 2015 18:49:28 +0000 (18:49 +0000)]
[MC] Simplify code. No functionality change intended.

llvm-svn: 238676

9 years agoCOFF: Use range-based for loop.
Rui Ueyama [Sun, 31 May 2015 16:10:50 +0000 (16:10 +0000)]
COFF: Use range-based for loop.

llvm-svn: 238675

9 years agoRemove debugging code
Marshall Clow [Sun, 31 May 2015 14:01:54 +0000 (14:01 +0000)]
Remove debugging code

llvm-svn: 238674

9 years ago[Format] Move UnwrappedLines instead of copying.
Benjamin Kramer [Sun, 31 May 2015 11:18:05 +0000 (11:18 +0000)]
[Format] Move UnwrappedLines instead of copying.

No functional change intended.

llvm-svn: 238673

9 years agoclang-format: NFC. Cleanup after r237895.
Daniel Jasper [Sun, 31 May 2015 08:51:54 +0000 (08:51 +0000)]
clang-format: NFC. Cleanup after r237895.

Specifically adhere to LLVM Coding Standards (no 'else' after
return/break/continue) and remove yet another implementation of
paren counting. We already have enough of those in the
UnwrappedLineParser.

No functional changes intended.

llvm-svn: 238672

9 years agoclang-format: [JS] Fix line breaks in computed property names.
Daniel Jasper [Sun, 31 May 2015 08:40:37 +0000 (08:40 +0000)]
clang-format: [JS] Fix line breaks in computed property names.

Before:
  let foo = {
    [someLongKeyHere]: 1,
    someOtherLongKeyHere: 2, [keyLongEnoughToWrap]: 3,
    lastLongKey: 4
  };

After:
  let foo = {
    [someLongKeyHere]: 1,
    someOtherLongKeyHere: 2,
    [keyLongEnoughToWrap]: 3,
    lastLongKey: 4
  };

llvm-svn: 238671

9 years agoCOFF: Remove redundant options from tests.
Rui Ueyama [Sun, 31 May 2015 04:21:30 +0000 (04:21 +0000)]
COFF: Remove redundant options from tests.

llvm-svn: 238670

9 years agoFix unsafe memory access.
Rui Ueyama [Sun, 31 May 2015 03:57:30 +0000 (03:57 +0000)]
Fix unsafe memory access.

llvm-svn: 238669

9 years agoCOFF: Implement subsystem inference.
Rui Ueyama [Sun, 31 May 2015 03:55:46 +0000 (03:55 +0000)]
COFF: Implement subsystem inference.

llvm-svn: 238668

9 years agoCOFF: Infer entry symbol name if /entry is not given.
Rui Ueyama [Sun, 31 May 2015 03:34:08 +0000 (03:34 +0000)]
COFF: Infer entry symbol name if /entry is not given.

`main` is not the only main function in Windows. You can choose one
from these four -- {w,}{WinMain,main}. There are four different entry
point functions for them, {w,}{WinMain,main}CRTStartup, respectively.
The linker needs to choose the right one depending on which `main`
function is defined.

llvm-svn: 238667

9 years agoDon't try to memcpy zero bytes; sometimes the source pointer is NULL, and that's...
Marshall Clow [Sun, 31 May 2015 03:13:31 +0000 (03:13 +0000)]
Don't try to memcpy zero bytes; sometimes the source pointer is NULL, and that's UB. Thanks to Nuno Lopes for the catch.

llvm-svn: 238666

9 years agoClarify how the binary file checked in was generated.
Davide Italiano [Sat, 30 May 2015 22:43:36 +0000 (22:43 +0000)]
Clarify how the binary file checked in was generated.

llvm-svn: 238665

9 years ago[Hexagon] Adding override specifier and removing erroneous assertion
Colin LeMahieu [Sat, 30 May 2015 20:03:07 +0000 (20:03 +0000)]
[Hexagon] Adding override specifier and removing erroneous assertion

llvm-svn: 238664

9 years agoAdd RelocVisitor support for MachO
Keno Fischer [Sat, 30 May 2015 19:44:53 +0000 (19:44 +0000)]
Add RelocVisitor support for MachO

This commit adds partial support for MachO relocations to RelocVisitor.
A simple test case is added to show that relocations are indeed being
applied and that using llvm-dwarfdump on MachO files no longer errors.
Correctness is not yet tested, due to an unrelated bug in DebugInfo,
which will be fixed with appropriate testcase in a followup commit.

Differential Revision: http://reviews.llvm.org/D8148

llvm-svn: 238663

9 years agoUse initializer instead of memset to zero out.
Rui Ueyama [Sat, 30 May 2015 19:28:58 +0000 (19:28 +0000)]
Use initializer instead of memset to zero out.

llvm-svn: 238662

9 years agoCOFF: Support long section name.
Rui Ueyama [Sat, 30 May 2015 19:09:50 +0000 (19:09 +0000)]
COFF: Support long section name.

Section names were truncated to 8 bytes because the section table's
name field is 8 byte long. This patch creates the string table to
store long names.

llvm-svn: 238661

9 years ago[Hexagon] Adding basic relaxation functionality.
Colin LeMahieu [Sat, 30 May 2015 18:55:47 +0000 (18:55 +0000)]
[Hexagon] Adding basic relaxation functionality.

llvm-svn: 238660

9 years ago[MC] Allow backends to decide relaxation for unresolved fixups.
Colin LeMahieu [Sat, 30 May 2015 18:42:22 +0000 (18:42 +0000)]
[MC] Allow backends to decide relaxation for unresolved fixups.

Differential Revision: http://reviews.llvm.org/D8217

llvm-svn: 238659

9 years ago[lib/Fuzzer] make assertions more informative and update comments for the user-suppli...
Kostya Serebryany [Sat, 30 May 2015 17:33:13 +0000 (17:33 +0000)]
[lib/Fuzzer] make assertions more informative and update comments for the user-supplied mutator

llvm-svn: 238658

9 years agoubsan: Check for null pointers given to certain builtins, such
Nuno Lopes [Sat, 30 May 2015 16:11:40 +0000 (16:11 +0000)]
ubsan: Check for null pointers given to certain builtins, such
as memcpy, memset, memmove, and bzero.

Reviewed by: Richard Smith

Differential Revision: http://reviews.llvm.org/D9673

llvm-svn: 238657

9 years agoCode cleanup: Reindent statements.
Logan Chien [Sat, 30 May 2015 14:00:39 +0000 (14:00 +0000)]
Code cleanup: Reindent statements.

llvm-svn: 238656

9 years ago[MC] Reorder MCSymbol members to reduce padding.
Benjamin Kramer [Sat, 30 May 2015 13:52:30 +0000 (13:52 +0000)]
[MC] Reorder MCSymbol members to reduce padding.

sizeof(MCSymbol) goes from 72 to 64 bytes on x86_64.

llvm-svn: 238655

9 years agoStripped trailing whitespace. NFC.
Simon Pilgrim [Sat, 30 May 2015 13:01:42 +0000 (13:01 +0000)]
Stripped trailing whitespace. NFC.

llvm-svn: 238654

9 years agoComment change. NFC
Renato Golin [Sat, 30 May 2015 10:44:07 +0000 (10:44 +0000)]
Comment change. NFC

That comment misleads the current discussions in mentioned bug. Leave
the discussions to the bug. Also, adding a future change FIXME.

llvm-svn: 238653

9 years ago[x86] Unify the horizontal adding used for popcount lowering taking the
Chandler Carruth [Sat, 30 May 2015 10:35:03 +0000 (10:35 +0000)]
[x86] Unify the horizontal adding used for popcount lowering taking the
best approach of each.

For vNi16, we use SHL + ADD + SRL pattern that seem easily the best.

For vNi32, we use the PUNPCK + PSADBW + PACKUSWB pattern. In some cases
there is a huge improvement with this in IACA's estimated throughput --
over 2x higher throughput!!!! -- but the measurements are too good to be
true. In one narrow case, the SHL + ADD + SHL + ADD + SRL pattern looks
slightly faster, but I'm not sure I believe any of the measurements at
this point. Both are the exact same uops though. Hard to be confident of
anything past that.

If anyone wants to collect very detailed (Agner-level) timings with the
result of this patch, or with the i32 case replaced with SHL + ADD + SHl
+ ADD + SRL, I'd be very interested. Note that you'll need to test it on
both Ivybridge and Haswell, with both SSE3, SSSE3, and AVX selected as
I saw unique behavior in each of these buckets with IACA all of which
should be checked against measured performance.

But this patch is still a useful improvement by dropping duplicate work
and getting the much nicer PSADBW lowering for v2i64.

I'd still like to rephrase this in terms of generic horizontal sum. It's
a bit lame to have a special case of that just for popcount.

llvm-svn: 238652

9 years ago[ARMTargetParser] Move IAS arch ext parser. NFC
Renato Golin [Sat, 30 May 2015 10:30:02 +0000 (10:30 +0000)]
[ARMTargetParser] Move IAS arch ext parser. NFC

The plan was to move the whole table into the already existing ArchExtNames
but some fields depend on a table-generated file, and we don't yet have this
feature in the generic lib/Support side.

Once the minimum target-specific table-generated files are available in a
generic fashion to these libraries, we'll have to keep it in the ASM parser.

llvm-svn: 238651

9 years ago[x86] Split out the horizontal byte sum lowering component of the LUT
Chandler Carruth [Sat, 30 May 2015 09:46:16 +0000 (09:46 +0000)]
[x86] Split out the horizontal byte sum lowering component of the LUT
lowering into a helper function.

NFC.

llvm-svn: 238650

9 years ago[CodeGen] Indirect fields can initialize a union
David Majnemer [Sat, 30 May 2015 09:12:07 +0000 (09:12 +0000)]
[CodeGen] Indirect fields can initialize a union

The first named data member is the field used to default initialize the
union.  An IndirectFieldDecl can introduce the first named data member
of a union.

llvm-svn: 238649

9 years ago[TableGen] Merge RecTy::typeIsConvertibleTo and RecTy::baseClassOf. NFC
Craig Topper [Sat, 30 May 2015 07:36:01 +0000 (07:36 +0000)]
[TableGen] Merge RecTy::typeIsConvertibleTo and RecTy::baseClassOf. NFC

typeIsConvertibleTo was just calling baseClassOf(this) on the argument passed to it, but there weren't different signatures for baseClassOf so passing 'this' didn't really do anything interesting. typeIsConvertibleTo could have just been a non-virtual method in RecTy. But since that would be kind of a silly method, I instead re-distributed the logic from baseClassOf into typeIsConvertibleTo.

llvm-svn: 238648

9 years agoFix indentation. NFC.
Craig Topper [Sat, 30 May 2015 07:35:21 +0000 (07:35 +0000)]
Fix indentation. NFC.

llvm-svn: 238647

9 years ago[TableGen] Remove all the variations of RecTy::convertValue and just handle the conve...
Craig Topper [Sat, 30 May 2015 07:34:51 +0000 (07:34 +0000)]
[TableGen] Remove all the variations of RecTy::convertValue and just handle the conversions in convertInitializerTo directly. This saves a bunch of vtable entries. NFC

llvm-svn: 238646

9 years agoDump YAML schedule tree as properly indented tree in DEBUG output
Tobias Grosser [Sat, 30 May 2015 06:46:59 +0000 (06:46 +0000)]
Dump YAML schedule tree as properly indented tree in DEBUG output

llvm-svn: 238645

9 years agoAdd DEBUG output to -polly-scops pass
Tobias Grosser [Sat, 30 May 2015 06:26:21 +0000 (06:26 +0000)]
Add DEBUG output to -polly-scops pass

llvm-svn: 238644

9 years agoAdd indvar pass to canonicalization sequence
Tobias Grosser [Sat, 30 May 2015 06:16:41 +0000 (06:16 +0000)]
Add indvar pass to canonicalization sequence

Running indvar before Polly is useful as this eliminates zexts as they commonly
appear when a 32 bit induction variable (type int) was used on a 64 bit system.
These zexts confuse our delinearization and prevent for example the successful
delinearization of the nussinov kernel in polybench-c-4.1.

This fixes http://llvm.org/PR23426

Suggested-by: Xing Su <xsu.llvm@outlook.com>
llvm-svn: 238643

9 years ago[x86] Update the order of instructions after I switched to a bitcast
Chandler Carruth [Sat, 30 May 2015 06:02:37 +0000 (06:02 +0000)]
[x86] Update the order of instructions after I switched to a bitcast
helper that skips creating a cast when it isn't necessary.

It's really somewhat concerning that this was caused by the the presence
of a no-op bitcast, but...

llvm-svn: 238642

9 years ago[WinCOFF] Add support for the .safeseh directive
David Majnemer [Sat, 30 May 2015 04:56:02 +0000 (04:56 +0000)]
[WinCOFF] Add support for the .safeseh directive

.safeseh adds an entry to the .sxdata section to register all the
appropriate functions which may handle an exception.  This entry is not
a relocation to the symbol but instead the symbol table index of the
function.

llvm-svn: 238641

9 years ago[x86] Replace the long spelling of getting a bitcast with the *much*
Chandler Carruth [Sat, 30 May 2015 04:23:13 +0000 (04:23 +0000)]
[x86] Replace the long spelling of getting a bitcast with the *much*
shorter one. NFC.

In addition to being much shorter to type and requiring fewer arguments,
this change saves over 30 lines from this one file, all wasted on total
boilerplate...

llvm-svn: 238640

9 years ago[x86] Replace the long spelling of getting a bitcast with the new short
Chandler Carruth [Sat, 30 May 2015 04:19:57 +0000 (04:19 +0000)]
[x86] Replace the long spelling of getting a bitcast with the new short
spelling. NFC.

llvm-svn: 238639

9 years ago[sdag] Add the helper I most want to the DAG -- building a bitcast
Chandler Carruth [Sat, 30 May 2015 04:14:10 +0000 (04:14 +0000)]
[sdag] Add the helper I most want to the DAG -- building a bitcast
around a value using its existing SDLoc.

Start using this in just one function to save omg lines of code.

llvm-svn: 238638

9 years ago[x86] Restore the bitcasts I removed when refactoring this to avoid
Chandler Carruth [Sat, 30 May 2015 04:05:11 +0000 (04:05 +0000)]
[x86] Restore the bitcasts I removed when refactoring this to avoid
shifting vectors of bytes as x86 doesn't have direct support for that.

This removes a bunch of redundant masking in the generated code for SSE2
and SSE3.

In order to avoid the really significant code size growth this would
have triggered, I also factored the completely repeatative logic for
shifting and masking into two lambdas which in turn makes all of this
much easier to read IMO.

llvm-svn: 238637

9 years ago[x86] Implement a faster vector population count based on the PSHUFB
Chandler Carruth [Sat, 30 May 2015 03:20:59 +0000 (03:20 +0000)]
[x86] Implement a faster vector population count based on the PSHUFB
in-register LUT technique.

Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html

The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.

On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.

The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.

I have extended Bruno's patch in the following ways:

0) I committed the new tests with baseline sequences so this shows
   a diff, and regenerated the tests using the update scripts.

1) Bruno had noticed and mentioned in IRC a redundant mask that
   I removed.

2) I introduced a particular optimization for the i32 vector cases where
   we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
   + PSADBW to compute doubled high i32 popcounts. This takes advantage
   of the fact that to line up the high i32 popcounts we have to shift
   them anyways, and we can shift them by one fewer bit to effectively
   divide the count by two. While the PSHUFD based horizontal add is no
   faster, it doesn't require registers or load traffic the way a mask
   would, and provides more ILP as it happens on different ports with
   high throughput.

3) I did some code cleanups throughout to simplify the implementation
   logic.

4) I refactored it to continue to use the parallel bitmath lowering when
   SSSE3 is not available to preserve the performance of that version on
   SSE2 targets where it is still much better than scalarizing as we'll
   still do a bitmath implementation of popcount even in scalar code
   there.

With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.

With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.

Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.

Differential Revision: http://reviews.llvm.org/D10084

llvm-svn: 238636

9 years ago[x86] Restructure the parallel bitmath lowering of popcount into
Chandler Carruth [Sat, 30 May 2015 03:20:55 +0000 (03:20 +0000)]
[x86] Restructure the parallel bitmath lowering of popcount into
a separate routine, generalize it to work for all the integer vector
sizes, and do general code cleanups.

This dramatically improves lowerings of byte and short element vector
popcount, but more importantly it will make the introduction of the
LUT-approach much cleaner.

The biggest cleanup I've done is to just force the legalizer to do the
bitcasting we need. We run these iteratively now and it makes the code
much simpler IMO. Other changes were minor, and mostly naming and
splitting things up in a way that makes it more clear what is going on.

The other significant change is to use a different final horizontal sum
approach. This is the same number of instructions as the old method, but
shifts left instead of right so that we can clear everything but the
final sum with a single shift right. This seems likely better than
a mask which will usually have to read the mask from memory. It is
certaily fewer u-ops. Also, this will be temporary. This and the LUT
approach share the need of horizontal adds to finish the computation,
and we have more clever approaches than this one that I'll switch over
to.

llvm-svn: 238635

9 years agoMC: Clean up MCExpr naming. NFC.
Jim Grosbach [Sat, 30 May 2015 01:25:56 +0000 (01:25 +0000)]
MC: Clean up MCExpr naming. NFC.

llvm-svn: 238634

9 years ago[BitcodeReader] Change an assert to a call to a call to Error()
Filipe Cabecinhas [Sat, 30 May 2015 00:17:20 +0000 (00:17 +0000)]
[BitcodeReader] Change an assert to a call to a call to Error()

It's reachable from user input.

Bug found with AFL fuzz.

llvm-svn: 238633

9 years agoSelectionDAG: fix logic for promoting shift types
Fiona Glaser [Fri, 29 May 2015 23:37:22 +0000 (23:37 +0000)]
SelectionDAG: fix logic for promoting shift types

r238503 fixed the problem of too-small shift types by promoting them
during legalization, but the correct solution is to promote only the
operands that actually demand promotion.

This fixes a crash on an out-of-tree target caused by trying to
promote an operand that can't be promoted.

llvm-svn: 238632

9 years agoAdd TODO note about switching to __decltype
Eric Fiselier [Fri, 29 May 2015 23:21:03 +0000 (23:21 +0000)]
Add TODO note about switching to __decltype

llvm-svn: 238631

9 years agoFix 80-column violations.
Eric Christopher [Fri, 29 May 2015 23:09:49 +0000 (23:09 +0000)]
Fix 80-column violations.

llvm-svn: 238630

9 years agoFix inferior's i/o connections to its console window on Windows 7.
Adrian McCarthy [Fri, 29 May 2015 23:01:25 +0000 (23:01 +0000)]
Fix inferior's i/o connections to its console window on Windows 7.

llvm-svn: 238629

9 years agoclang/CMakeLists.txt: s/LLVM_INSTALL_PACKAGE_DIR/CLANG_INSTALL_PACKAGE_DIR/ for the...
NAKAMURA Takumi [Fri, 29 May 2015 22:58:05 +0000 (22:58 +0000)]
clang/CMakeLists.txt: s/LLVM_INSTALL_PACKAGE_DIR/CLANG_INSTALL_PACKAGE_DIR/ for the standalone configuration.

llvm-svn: 238628

9 years ago[WinEH] Adjust the 32-bit SEH prologue to better match reality
Reid Kleckner [Fri, 29 May 2015 22:57:46 +0000 (22:57 +0000)]
[WinEH] Adjust the 32-bit SEH prologue to better match reality

It turns out that _except_handler3 and _except_handler4 really use the
same stack allocation layout, at least today. They just make different
choices about encoding the LSDA.

This is in preparation for lowering the llvm.eh.exceptioninfo().

llvm-svn: 238627

9 years ago[sanitizer] Add -lrt to dyndd link line.
Evgeniy Stepanov [Fri, 29 May 2015 22:57:15 +0000 (22:57 +0000)]
[sanitizer] Add -lrt to dyndd link line.

llvm-svn: 238626

9 years ago[Sema] Promote compound assignment exprs. with fp16 LHS and int. RHS.
Ahmed Bougacha [Fri, 29 May 2015 22:54:57 +0000 (22:54 +0000)]
[Sema] Promote compound assignment exprs. with fp16 LHS and int. RHS.

We catch most of the various other __fp16 implicit conversions to
float, but not this one:

  __fp16 a;
  int i;
  ...
  a += i;

For which we used to generate something 'fun' like:

  %conv = sitofp i32 %i to float
  %1 = tail call i16 @llvm.convert.to.fp16.f32(float %conv)
  %add = add i16 %0, %1

Instead, when we have an __fp16 LHS and an integer RHS, we should
use float as the result type.

While there, add a bunch of missing tests for mixed
__fp16/integer expressions.

llvm-svn: 238625

9 years agoUse FileSpec::IsRelativeToCurrentWorkingDirectory instead of llvm::sys::path::is_rela...
Oleksiy Vyalov [Fri, 29 May 2015 22:54:45 +0000 (22:54 +0000)]
Use FileSpec::IsRelativeToCurrentWorkingDirectory instead of llvm::sys::path::is_relative in PlatformAndroid::GetFile.

http://reviews.llvm.org/D10141

llvm-svn: 238624

9 years agoLeave OS type and vendor as unspecified unknowns.
Oleksiy Vyalov [Fri, 29 May 2015 22:45:47 +0000 (22:45 +0000)]
Leave OS type and vendor as unspecified unknowns.

http://reviews.llvm.org/D10080

llvm-svn: 238623

9 years agoprofile: Fix link errors in this test on darwin
Justin Bogner [Fri, 29 May 2015 22:41:25 +0000 (22:41 +0000)]
profile: Fix link errors in this test on darwin

llvm-svn: 238622

9 years agoAdd descriptive names to sanitizer entries in /proc/self/maps. Helps debugging.
Evgeniy Stepanov [Fri, 29 May 2015 22:31:28 +0000 (22:31 +0000)]
Add descriptive names to sanitizer entries in /proc/self/maps. Helps debugging.

This is done by creating a named shared memory region, unlinking it
and setting up a private (i.e. copy-on-write) mapping of that instead
of a regular anonymous mapping. I've experimented with regular
(sparse) files, but they can not be scaled to the size of MSan shadow
mapping, at least on Linux/X86_64 and ext3 fs.

Controlled by a common flag, decorate_proc_maps, disabled by default.

This patch has a few shortcomings:
* not all mappings are annotated, especially in TSan.
* our handling of memset() of shadow via mmap() puts small anonymous
  mappings inside larger named mappings, which looks ugly and can, in
  theory, hit the mapping number limit.

llvm-svn: 238621

9 years ago[docs] fix the declarations of the llvm.nvvm.ptr.gen.to.* intrinsics
Jingyue Wu [Fri, 29 May 2015 22:18:03 +0000 (22:18 +0000)]
[docs] fix the declarations of the llvm.nvvm.ptr.gen.to.* intrinsics

Summary:
These intrinsics should take a generic input address space and outputs a
non-generic address space.

Test Plan: no

Reviewers: jholewinski, eliben

Reviewed By: eliben

Subscribers: eliben, jholewinski, llvm-commits

Differential Revision: http://reviews.llvm.org/D10132

llvm-svn: 238620

9 years agoDisable FP elimination in funcs using 32-bit MSVC EH personalities
Reid Kleckner [Fri, 29 May 2015 21:58:11 +0000 (21:58 +0000)]
Disable FP elimination in funcs using 32-bit MSVC EH personalities

The value in 'ebp' acts as an implicit argument to the outlined
handlers, and is recovered with frameaddress(1).

llvm-svn: 238619

9 years agoCOFF: Move machine type auto-detection to SymbolTable.
Peter Collingbourne [Fri, 29 May 2015 21:47:36 +0000 (21:47 +0000)]
COFF: Move machine type auto-detection to SymbolTable.

The new mechanism is less code, and fixes the case where all inputs
are archives.

Differential Revision: http://reviews.llvm.org/D10136

llvm-svn: 238618