platform/upstream/llvm.git
9 years agoSimplify.
Rui Ueyama [Wed, 21 Jan 2015 18:54:26 +0000 (18:54 +0000)]
Simplify.

What we are trying to do here is to skip object files in group if
group is repeated. This code is simpler than before.

llvm-svn: 226688

9 years agoLiveIntervalAnalysis: Factor out code to update liveness on physreg def removal
Matthias Braun [Wed, 21 Jan 2015 18:50:21 +0000 (18:50 +0000)]
LiveIntervalAnalysis: Factor out code to update liveness on physreg def removal

This cleans up code and is more in line with the general philosophy of
modifying LiveIntervals through LiveIntervalAnalysis instead of changing
them directly.

llvm-svn: 226687

9 years agoLiveIntervalAnalysis: Remove unused pruneValue() variant.
Matthias Braun [Wed, 21 Jan 2015 18:45:57 +0000 (18:45 +0000)]
LiveIntervalAnalysis: Remove unused pruneValue() variant.

llvm-svn: 226686

9 years agoclang-format: Fix crasher caused by incorrect resetting of token info.
Daniel Jasper [Wed, 21 Jan 2015 18:35:47 +0000 (18:35 +0000)]
clang-format: Fix crasher caused by incorrect resetting of token info.

llvm-svn: 226685

9 years agoLet subprograms with instructions without parent scopes fail the
Adrian Prantl [Wed, 21 Jan 2015 18:32:56 +0000 (18:32 +0000)]
Let subprograms with instructions without parent scopes fail the
verification. Tested via a unit test.

Follow-up to r226616.

llvm-svn: 226684

9 years agoRevert "Some fixes for linking Python on Windows."
Zachary Turner [Wed, 21 Jan 2015 18:30:40 +0000 (18:30 +0000)]
Revert "Some fixes for linking Python on Windows."

This reverts commit r226679.  For some reason it was
not generating the same behavior as manually specifying
the include dir, library path, and exe path, and it was
causing the test suite to fail to run.

llvm-svn: 226683

9 years agoR600/SI: Custom lower fround
Matt Arsenault [Wed, 21 Jan 2015 18:18:25 +0000 (18:18 +0000)]
R600/SI: Custom lower fround

This fixes it for SI. It also removes the pattern
used previously for Evergreen for f32. I'm not sure
if the the new R600 output is better or not, but it uses
1 fewer instructions if BFI is available.

llvm-svn: 226682

9 years ago[Hexagon] Converting multiply and accumulate with immediate intrinsics to patterns.
Colin LeMahieu [Wed, 21 Jan 2015 18:13:15 +0000 (18:13 +0000)]
[Hexagon] Converting multiply and accumulate with immediate intrinsics to patterns.

llvm-svn: 226681

9 years agoclang-format: Fix bad memory access.
Daniel Jasper [Wed, 21 Jan 2015 18:04:02 +0000 (18:04 +0000)]
clang-format: Fix bad memory access.

llvm-svn: 226680

9 years agoSome fixes for linking Python on Windows.
Zachary Turner [Wed, 21 Jan 2015 17:53:10 +0000 (17:53 +0000)]
Some fixes for linking Python on Windows.

CMake FindPythonLibs will look for multiple versions of Python
including both debug and release, and build up a list such as
(debug <debugpath> optimized <optimizedpath>).  This confuses
the logic we have in CMake to copy the correct python dll to
the output directory so that it need not be in your system's PATH.

To alleviate this, we manually split this list and extract out
the debug and release versions of the python library, and copy
only the correct one to the output directory.

llvm-svn: 226679

9 years agoclang-format: Fix use-heap-after-free bug.
Daniel Jasper [Wed, 21 Jan 2015 17:35:29 +0000 (17:35 +0000)]
clang-format: Fix use-heap-after-free bug.

Discovered by the awesome test case and ASAN.

llvm-svn: 226678

9 years agoUsed CHECK-DAG since the order is not important.
Rafael Espindola [Wed, 21 Jan 2015 17:12:04 +0000 (17:12 +0000)]
Used CHECK-DAG since the order is not important.

llvm-svn: 226677

9 years ago[X86] Declare SSE4.1/AVX2 vector extloads covered by PMOV[SZ]X legal.
Ahmed Bougacha [Wed, 21 Jan 2015 17:07:06 +0000 (17:07 +0000)]
[X86] Declare SSE4.1/AVX2 vector extloads covered by PMOV[SZ]X legal.

Now that we can fully specify extload legality, we can declare them
legal for the PMOVSX/PMOVZX instructions.  This for instance enables
a DAGCombine to fire on code such as
  (and (<zextload-equivalent> ...), <redundant mask>)
to turn it into:
  (zextload ...)
as seen in the testcase changes.

There is one regression, in widen_load-2.ll: we're no longer able
to do store-to-load forwarding with illegal extload memory types.
This will be addressed separately.

Differential Revision: http://reviews.llvm.org/D6533

llvm-svn: 226676

9 years agoUsed CHECK-DAG since the order is not important.
Rafael Espindola [Wed, 21 Jan 2015 16:56:43 +0000 (16:56 +0000)]
Used CHECK-DAG since the order is not important.

llvm-svn: 226675

9 years ago[msan] Fix origins in realloc.
Evgeniy Stepanov [Wed, 21 Jan 2015 16:48:29 +0000 (16:48 +0000)]
[msan] Fix origins in realloc.

Fixes 2 issues in origins arising from realloc() calls:
 * In the in-place grow case origin for the new memory is not set at all.
 * In the copy-realloc case __msan_memcpy is used, which unwinds stack from
   inside the MSan runtime. This does not generally work (as we may be built
   w/o frame pointers), and produces "bad" stack trace anyway, with several
   uninteresting (internal) frames on top.

This change also makes realloc() honor "zeroise" and "poison_in_malloc" flags.

See https://code.google.com/p/memory-sanitizer/issues/detail?id=73.

llvm-svn: 226674

9 years ago[msan] Refactor shadow operations.
Evgeniy Stepanov [Wed, 21 Jan 2015 16:42:30 +0000 (16:42 +0000)]
[msan] Refactor shadow operations.

Move a bunch of functions to a new source file and rename some of them for
consistency. No functional changes.

llvm-svn: 226673

9 years ago[lit] Format JSONMetricValue strings better.
Eric Fiselier [Wed, 21 Jan 2015 16:38:31 +0000 (16:38 +0000)]
[lit] Format JSONMetricValue strings better.

llvm-svn: 226672

9 years agoFixed a bug with how we determine bitset indices.
George Burgess IV [Wed, 21 Jan 2015 16:37:21 +0000 (16:37 +0000)]
Fixed a bug with how we determine bitset indices.

llvm-svn: 226671

9 years agoInitial support for C++ parameter completion
Francisco Lopes da Silva [Wed, 21 Jan 2015 16:24:11 +0000 (16:24 +0000)]
Initial support for C++ parameter completion

The improved completion in call context now works with:

 - Functions.
 - Member functions.
 - Constructors.
 - New expressions.
 - Function call expressions.
 - Template variants of the previous.

There are still rough edges to be fixed:

 - Provide support for optional parameters.         (fix known)
 - Provide support for member initializers.         (fix known)
 - Provide support for variadic template functions. (fix unknown)
 - Others?

llvm-svn: 226670

9 years agoAdd missing include guards to WindowsSupport.h.
Yaron Keren [Wed, 21 Jan 2015 16:20:38 +0000 (16:20 +0000)]
Add missing include guards to WindowsSupport.h.

llvm-svn: 226669

9 years agoDon't assume variable name.
Rafael Espindola [Wed, 21 Jan 2015 16:18:42 +0000 (16:18 +0000)]
Don't assume variable name.

Should fix the test in -Asserts builds.

llvm-svn: 226668

9 years agoMake the test a bit stricter. NFC.
Rafael Espindola [Wed, 21 Jan 2015 16:13:57 +0000 (16:13 +0000)]
Make the test a bit stricter. NFC.

llvm-svn: 226667

9 years agoMake the test a bit stricter. NFC.
Rafael Espindola [Wed, 21 Jan 2015 16:03:26 +0000 (16:03 +0000)]
Make the test a bit stricter. NFC.

llvm-svn: 226666

9 years agoRevert "DAGCombine: fold (or (and X, M), (and X, N)) -> (and X, (or M, N))"
Tim Northover [Wed, 21 Jan 2015 15:48:52 +0000 (15:48 +0000)]
Revert "DAGCombine: fold (or (and X, M), (and X, N)) -> (and X, (or M, N))"

It hadn't gone through review yet, but was still on my local copy.

This reverts commit r226663

llvm-svn: 226665

9 years agoAArch64: add backend option to reserve x18 (platform register)
Tim Northover [Wed, 21 Jan 2015 15:43:31 +0000 (15:43 +0000)]
AArch64: add backend option to reserve x18 (platform register)

AAPCS64 says that it's up to the platform to specify whether x18 is
reserved, and a first step on that way is to add a flag controlling
it.

From: Andrew Turner <andrew@fubar.geek.nz>
llvm-svn: 226664

9 years agoDAGCombine: fold (or (and X, M), (and X, N)) -> (and X, (or M, N))
Tim Northover [Wed, 21 Jan 2015 15:43:28 +0000 (15:43 +0000)]
DAGCombine: fold (or (and X, M), (and X, N)) -> (and X, (or M, N))

llvm-svn: 226663

9 years agoclang-format function. NFC.
Rafael Espindola [Wed, 21 Jan 2015 14:55:00 +0000 (14:55 +0000)]
clang-format function. NFC.

llvm-svn: 226662

9 years ago[x32] Fast ISel should use LEA64_32r instead of LEA32r to adjust addresses in x32...
Michael Kuperstein [Wed, 21 Jan 2015 14:44:05 +0000 (14:44 +0000)]
[x32] Fast ISel should use LEA64_32r instead of LEA32r to adjust addresses in x32 mode.

llvm-svn: 226661

9 years agoUse a smaller pragma unroll threshold to reduce test execution time.
Alexander Potapenko [Wed, 21 Jan 2015 13:52:02 +0000 (13:52 +0000)]
Use a smaller pragma unroll threshold to reduce test execution time.
When opt is compiled with AddressSanitizer it takes more than 30 seconds
to unroll the loop in unroll_1M().

llvm-svn: 226660

9 years agotsan: remove sleeps from tests
Dmitry Vyukov [Wed, 21 Jan 2015 13:50:02 +0000 (13:50 +0000)]
tsan: remove sleeps from tests

Even sleep(1) lead to episodical flakes on some machines.
Use an invisible by tsan barrier to enforce required execution order instead.
This makes the tests deterministic and faster.

llvm-svn: 226659

9 years ago[msan] Update origin for the entire destination range on memory store.
Evgeniy Stepanov [Wed, 21 Jan 2015 13:21:31 +0000 (13:21 +0000)]
[msan] Update origin for the entire destination range on memory store.

Previously we always stored 4 bytes of origin at the destination address
even for 8-byte (and longer) stores.

This should fix rare missing, or incorrect, origin stacks in MSan reports.

llvm-svn: 226658

9 years ago[mips][microMIPS] MicroMIPS 16-bit unconditional branch instruction B
Jozef Kolek [Wed, 21 Jan 2015 12:39:30 +0000 (12:39 +0000)]
[mips][microMIPS] MicroMIPS 16-bit unconditional branch instruction B

Implement microMIPS 16-bit unconditional branch instruction B.

Implemented 16-bit microMIPS unconditional instruction has real name B16, and
B is an alias which expands to either B16 or BEQ according to the rules:
b 256 --> b16 256 # R_MICROMIPS_PC10_S1
b 12256 --> beq $zero, $zero, 12256 # R_MICROMIPS_PC16_S1
b label --> beq $zero, $zero, label # R_MICROMIPS_PC16_S1

Differential Revision: http://reviews.llvm.org/D3514

llvm-svn: 226657

9 years ago[mips][microMIPS] Implement ADDIUPC instruction
Jozef Kolek [Wed, 21 Jan 2015 12:10:11 +0000 (12:10 +0000)]
[mips][microMIPS] Implement ADDIUPC instruction

Differential Revision: http://reviews.llvm.org/D6582

llvm-svn: 226656

9 years ago[PM] Refactor the InstCombiner interface to use an external worklist.
Chandler Carruth [Wed, 21 Jan 2015 11:38:17 +0000 (11:38 +0000)]
[PM] Refactor the InstCombiner interface to use an external worklist.

Because in its primary function pass the combiner is run repeatedly over
the same function until doing so produces no changes, it is essentially
to not re-allocate the worklist. However, as a utility, the more common
pattern would be to put a limited set of instructions in the worklist
rather than the entire function body. That is also the more likely
pattern when used by the new pass manager.

The result is a very light weight combiner that does the visiting with
a separable worklist. This can then be wrapped up in a helper function
for users that want a combiner utility, or as I have here it can be
wrapped up in a pass which manages the iterations used when combining an
entire function's instructions.

Hopefully this removes some of the worst of the interface warts that
became apparant with the last patch here. However, there is clearly more
work. I've again left some FIXMEs for the most egregious. The ones that
stick out to me are the exposure of the worklist and IR builder as
public members, and the use of pointers rather than references. However,
fixing these is likely to be much more mechanical and less interesting
so I didn't want to touch them in this patch.

llvm-svn: 226655

9 years ago[PM] Simplify (ha! ha!) the way that instcombine calls the
Chandler Carruth [Wed, 21 Jan 2015 11:23:40 +0000 (11:23 +0000)]
[PM] Simplify (ha! ha!) the way that instcombine calls the
SimplifyLibCalls utility by sinking it into the specific call part of
the combiner.

This will avoid us needing to do any contortions to build this object in
a subsequent refactoring I'm doing and seems generally better factored.
We don't need this utility everywhere and it carries no interesting
state so we might as well build it on demand.

llvm-svn: 226654

9 years agoAST: Don't ignore alignas on EnumDecls when calculating alignment
David Majnemer [Wed, 21 Jan 2015 10:54:38 +0000 (10:54 +0000)]
AST: Don't ignore alignas on EnumDecls when calculating alignment

We didn't consider any alignment attributes on an EnumDecl when
calculating alignment.

While we are here, ignore alignment specifications on typedef types if
one is used as the underlying type.  Otherwise, weird things happen:

enum Y : int;
Y y;

typedef int __attribute__((aligned(64))) u;
enum Y : u {};

What is the alignment of 'Y'?  It would be more consistent with the
overall design of enums with fixed underlying types to consider the
underlying type's UnqualifiedDesugaredType.

This fixes PR22279.

llvm-svn: 226653

9 years ago[Mips][Disassembler]When disassembler meets load/store from coprocessor 2 instruction...
Vladimir Medic [Wed, 21 Jan 2015 10:47:36 +0000 (10:47 +0000)]
[Mips][Disassembler]When disassembler meets load/store from coprocessor 2 instructions for mips r6 it crashes as the access to operands array is out of range. This patch adds dedicated decoder method that properly handles decoding of these instructions.

llvm-svn: 226652

9 years ago[ASan] Fix the test expectations for dyld_insert_libraries_reexec.cc
Alexander Potapenko [Wed, 21 Jan 2015 10:04:41 +0000 (10:04 +0000)]
[ASan] Fix the test expectations for dyld_insert_libraries_reexec.cc

llvm-svn: 226651

9 years ago[Msan] Fix the accept unit test to pass on FreeBSD
Viktor Kutuzov [Wed, 21 Jan 2015 09:46:21 +0000 (09:46 +0000)]
[Msan] Fix the accept unit test to pass on FreeBSD
Differential Revision: http://reviews.llvm.org/D7083

llvm-svn: 226650

9 years ago[Sanitizers] Intercept confstr() on FreeBSD
Viktor Kutuzov [Wed, 21 Jan 2015 08:56:40 +0000 (08:56 +0000)]
[Sanitizers] Intercept confstr() on FreeBSD
Committed unreviewed with permission.

llvm-svn: 226649

9 years ago[Sanitizers] Intercept opendir()
Viktor Kutuzov [Wed, 21 Jan 2015 08:54:01 +0000 (08:54 +0000)]
[Sanitizers] Intercept opendir()
Differential Revision: http://reviews.llvm.org/D6968

llvm-svn: 226648

9 years ago[x86] Remove some unnecessary and slightly confusing typecasts from some patterns...
Craig Topper [Wed, 21 Jan 2015 08:43:57 +0000 (08:43 +0000)]
[x86] Remove some unnecessary and slightly confusing typecasts from some patterns. I think it actually went i32->iPtr->i32 in some of these cases.

llvm-svn: 226647

9 years ago[X86] Convert all the i8imm used by AVX512 and MMX instructions to u8imm.
Craig Topper [Wed, 21 Jan 2015 08:43:49 +0000 (08:43 +0000)]
[X86] Convert all the i8imm used by AVX512 and MMX instructions to u8imm.

llvm-svn: 226646

9 years ago[X86] Convert all the i8imm used by SSE and AVX instructions to u8imm.
Craig Topper [Wed, 21 Jan 2015 08:15:54 +0000 (08:15 +0000)]
[X86] Convert all the i8imm used by SSE and AVX instructions to u8imm.

This makes the assembler check their size and removes a hack from the disassembler to avoid sign extending the immediate.

llvm-svn: 226645

9 years agotsan: don't unroll memory access loop in debug mode
Dmitry Vyukov [Wed, 21 Jan 2015 08:09:13 +0000 (08:09 +0000)]
tsan: don't unroll memory access loop in debug mode

MemoryAccess function consumes ~4K of stack in debug mode,
in significant part due to the unrolled loop.
And gtest gives only 4K of stack to death test
threads, which causes stack overflows in debug mode.

llvm-svn: 226644

9 years ago[ELF] Minimal implementation for ARM static linking
Greg Fitzgerald [Wed, 21 Jan 2015 07:35:48 +0000 (07:35 +0000)]
[ELF] Minimal implementation for ARM static linking

The code is able to statically link the simplest case of:

  int main() { return 0; }

* Only works with ARM code - no Thumb code, no interwork (-marm -mno-thumb-interwork)
* musl libc built with no interwork and no Thumb code

Differential Revision: http://reviews.llvm.org/D6716

From: Denis Protivensky <dprotivensky@accesssoftek.com>
llvm-svn: 226643

9 years ago[x86] Add assembly parser bounds checking to the immediate value for cmpss/cmpsd...
Craig Topper [Wed, 21 Jan 2015 06:07:53 +0000 (06:07 +0000)]
[x86] Add assembly parser bounds checking to the immediate value for cmpss/cmpsd/cmpps/cmppd.

llvm-svn: 226642

9 years agotuple: Make operator<() linear instead of exponential
Duncan P. N. Exon Smith [Wed, 21 Jan 2015 02:51:17 +0000 (02:51 +0000)]
tuple: Make operator<() linear instead of exponential

Patch by Matthew Dempsky!

llvm-svn: 226641

9 years ago[PM] Replace an abuse of inheritance to override a single function with
Chandler Carruth [Wed, 21 Jan 2015 02:11:59 +0000 (02:11 +0000)]
[PM] Replace an abuse of inheritance to override a single function with
a more direct approach: a type-erased glorified function pointer. Now we
can pass a function pointer into this for the easy case and we can even
pass a lambda into it in the interesting case in the instruction
combiner.

I'll be using this shortly to simplify the interfaces to InstCombiner,
but this helps pave the way and seems like a better design for the
libcall simplifier utility.

llvm-svn: 226640

9 years ago[sanitizer] First step toward supporting 42-bit AS on aarch64
Kostya Serebryany [Wed, 21 Jan 2015 02:11:05 +0000 (02:11 +0000)]
[sanitizer] First step toward supporting 42-bit AS on aarch64

aarch64-linux kernel has configurable 39, 42 or 47 bit virtual address
space.  Most distros AFAIK use 42-bit VA right now, but there are also
39-bit VA users too.  The ppc64 handling can be used for this just fine
and support all the 3 sizes.

There are other issues, like allocator32 not really being able to support
the larger addres spaces, and hardcoded 39-bit address space size in other
macros.

Patch by Jakub Jelinek.

llvm-svn: 226639

9 years agoMuch spelling. Wow.
Nico Weber [Wed, 21 Jan 2015 02:09:16 +0000 (02:09 +0000)]
Much spelling. Wow.

llvm-svn: 226638

9 years ago[sanitizer] Fix aarch64 sanitizer build with recent glibc
Kostya Serebryany [Wed, 21 Jan 2015 02:08:42 +0000 (02:08 +0000)]
[sanitizer] Fix aarch64 sanitizer build with recent glibc

glibc recently changed ABI on aarch64-linux:
https://sourceware.org/git/?p=glibc.git;a=commit;h=5c40c3bab2fddaca8cfe12d75944d1fef8adf1a4
Instead of having unsigned short mode; unsigned short __pad1; it now has
unsigned int mode; field in ipc_perm structure.

This patch allows to build against the recent glibc and disables the
ipc_perm.mode verification for older versions of glibc.

I think it shouldn't be a big deal even for older glibcs, I couldn't find
any place which would actually care about the exact mode field, rather than
the whole structure, appart from the CHECK_SIZE_AND_OFFSET macro.

Patch by Jakub Jelinek

llvm-svn: 226637

9 years ago[asan] use MADV_NOHUGEPAGE for shadow to reduce the actual memory usage
Kostya Serebryany [Wed, 21 Jan 2015 02:05:31 +0000 (02:05 +0000)]
[asan] use MADV_NOHUGEPAGE for shadow to reduce the actual memory usage

llvm-svn: 226636

9 years agoAllow individual ValueObjects to pick their preferred display language
Enrico Granata [Wed, 21 Jan 2015 01:47:13 +0000 (01:47 +0000)]
Allow individual ValueObjects to pick their preferred display language

Most of the time, we can use context information just fine to choose a language (i.e. the language of the frame that the root object was defined in, if any); but in some cases, synthetic children may be fabricated as root frame-less entities, and then we wouldn't know any better

This patch allows (internal) synthetic child providers to set a display language on the children they generate, should they so choose

llvm-svn: 226634

9 years agoRevert "Sema: err_after_alias is unreachable, remove it"
David Majnemer [Wed, 21 Jan 2015 01:30:40 +0000 (01:30 +0000)]
Revert "Sema: err_after_alias is unreachable, remove it"

This reverts commit r226626.  err_after_alias is, in fact, reachable.

llvm-svn: 226633

9 years agoFix --start-group/end-group.
Rui Ueyama [Wed, 21 Jan 2015 01:26:43 +0000 (01:26 +0000)]
Fix --start-group/end-group.

We used to manage the state whether we are in a group or not
using a counter. The counter is incremented by one if we jump from
end-group to start-group, and decremented by one if we don't.
The counter was assumed to be either zero or one, but obviously it
could be negative (if there's a group which is not repeated at all).
This is a fix for that issue.

llvm-svn: 226632

9 years agoAdding compact unwind as a source of unwind information
Jason Molenda [Wed, 21 Jan 2015 01:26:28 +0000 (01:26 +0000)]
Adding compact unwind as a source of unwind information
introduced subtle bugs in two places in
RegisterContextLLDB::GetFullUnwindPlanForFrame where
it specifically wanted to get an eh_frame unwind plan
and was using "Get CallSite UnwindPlan" as synonymous
with that.  But now we have two different types of
unwind plan that can be returned in that case, and
compact unwind won't behaves as needed.

<rdar://problem/19528559>

llvm-svn: 226631

9 years agoMS ABI: Virtual member pointer thunks should be in COMDAT groups
David Majnemer [Wed, 21 Jan 2015 01:21:31 +0000 (01:21 +0000)]
MS ABI: Virtual member pointer thunks should be in COMDAT groups

They can be emitted by multiple translation units and thus belong in a
COMDAT group.

llvm-svn: 226630

9 years agoMS ABI: Let guard variables be present in COMDATs
David Majnemer [Wed, 21 Jan 2015 01:04:30 +0000 (01:04 +0000)]
MS ABI: Let guard variables be present in COMDATs

A guard variable in a COMDAT'd function should also be in a COMDAT.

llvm-svn: 226629

9 years agoCodeGen: Compiler generated __declspec(uuid) objects should be COMDAT'd
David Majnemer [Wed, 21 Jan 2015 01:04:28 +0000 (01:04 +0000)]
CodeGen: Compiler generated __declspec(uuid) objects should be COMDAT'd

llvm-svn: 226628

9 years agoMake DIExpression::Verify() stricter by checking that the number of
Adrian Prantl [Wed, 21 Jan 2015 00:59:20 +0000 (00:59 +0000)]
Make DIExpression::Verify() stricter by checking that the number of
elements and the ordering is sane and cleanup the accessors.

llvm-svn: 226627

9 years agoSema: err_after_alias is unreachable, remove it
David Majnemer [Wed, 21 Jan 2015 00:52:17 +0000 (00:52 +0000)]
Sema: err_after_alias is unreachable, remove it

Examples this would have catched are now handled by the attribute
verification code.

llvm-svn: 226626

9 years agoFix some platform specific paths in TestSettings.py.
Zachary Turner [Wed, 21 Jan 2015 00:40:27 +0000 (00:40 +0000)]
Fix some platform specific paths in TestSettings.py.

We were referring to hardcoded paths /bin/ls and /bin/cat.  For
the purposes of this test, the actual value it's set to doesn't
matter, and it might as well be a non-existent path.  All that
matters is that the before and after values have to match, and
that trailing whitespace is stripped.  On Windows FileSpec
(correctly) converts /bin/ls to D:\bin\ls though, so the before
and after values won't match.  So this patch just correctly builds
up a valid path in a platform-agnostic manner, and verifies that
it matches before and after the set.

llvm-svn: 226625

9 years agoCorrect all typos in the initialization arguments, even if one could not
Kaelyn Takata [Wed, 21 Jan 2015 00:04:19 +0000 (00:04 +0000)]
Correct all typos in the initialization arguments, even if one could not
be corrected.

This fixes PR22250, which exposed the bug where if there's more than one
TypoExpr in the arguments, once one failed to be corrected none of the
TypoExprs after it would be handled at all thanks to an early return.

llvm-svn: 226624

9 years ago[X86][AVX] Simplified diff between AVX1 and SSE42 fp stack folding tests. NFC.
Simon Pilgrim [Wed, 21 Jan 2015 00:02:13 +0000 (00:02 +0000)]
[X86][AVX] Simplified diff between AVX1 and SSE42 fp stack folding tests. NFC.

Changed the AVX1 tests register spill tail call to return a xmm like the SSE42 version - makes doing diffs between them a lot easier without affecting the spills themselves.

llvm-svn: 226623

9 years ago[X86][SSE] Added SSE/AVX1 integer stack folding tests.
Simon Pilgrim [Tue, 20 Jan 2015 23:54:17 +0000 (23:54 +0000)]
[X86][SSE] Added SSE/AVX1 integer stack folding tests.

Some folding patterns + tests are missing (marked as TODO) - these will be added in a future patch for review.

llvm-svn: 226622

9 years ago[X86][SSE] Added SSE fp stack folding tests.
Simon Pilgrim [Tue, 20 Jan 2015 23:50:18 +0000 (23:50 +0000)]
[X86][SSE] Added SSE fp stack folding tests.

Some folding patterns + tests are missing (marked as TODO) - these will be added in a future patch for review.

llvm-svn: 226621

9 years ago[X86][AVX] Renamed AVX1 fp stack folding tests. NFC.
Simon Pilgrim [Tue, 20 Jan 2015 23:45:50 +0000 (23:45 +0000)]
[X86][AVX] Renamed AVX1 fp stack folding tests. NFC.

The SSE42 version of the AVX1 float stack folding tests will be added shortly, this renames the AVX1 file so that the files will be near each other in a directory listing to help ensure they are kept in sync.

llvm-svn: 226620

9 years ago[PM] Separate the InstCombiner from its pass.
Chandler Carruth [Tue, 20 Jan 2015 22:44:35 +0000 (22:44 +0000)]
[PM] Separate the InstCombiner from its pass.

This creates a small internal pass which runs the InstCombiner over
a function. This is the hard part of porting InstCombine to the new pass
manager, as at this point none of the code in InstCombine has access to
a Pass object any longer.

The resulting interface for the InstCombiner is pretty terrible. I'm not
planning on leaving it that way. The key thing missing is that we need
to separate the worklist from the combiner a touch more. Once that's
done, it should be possible for *any* part of LLVM to just create
a worklist with instructions, populate it, and then combine it until
empty. The pass will just be the (obvious and important) special case of
doing that for an entire function body.

For now, this is the first increment of factoring to make all of this
work.

llvm-svn: 226618

9 years agoAdd missing case statement to quiet a compiler warning.
Greg Clayton [Tue, 20 Jan 2015 22:41:14 +0000 (22:41 +0000)]
Add missing case statement to quiet a compiler warning.

llvm-svn: 226617

9 years agoDebugLocs without a scope should fail the verification.
Adrian Prantl [Tue, 20 Jan 2015 22:37:25 +0000 (22:37 +0000)]
DebugLocs without a scope should fail the verification.
Follow-up to r226588.

llvm-svn: 226616

9 years agoOnly use -target flag when target_triple is manually specified
Eric Fiselier [Tue, 20 Jan 2015 22:37:25 +0000 (22:37 +0000)]
Only use -target flag when target_triple is manually specified

llvm-svn: 226615

9 years agoXFAIL pexpect tests on Windows.
Zachary Turner [Tue, 20 Jan 2015 22:36:03 +0000 (22:36 +0000)]
XFAIL pexpect tests on Windows.

At some point we will need to either provide a pexpect equivalent
on Windows, or provide some other method of doing out-of-process
tests.

Even with a pexpect replacement, it may be worth re-evaluating
some of these tests to see if they would be better served as
in-process tests.  The larger issue of coming up with a pexpect
replacement on Windows is tracked in http://llvm.org/pr22274.

llvm-svn: 226614

9 years agoDon't pass -Wl,z,defs for now.
Rafael Espindola [Tue, 20 Jan 2015 22:08:20 +0000 (22:08 +0000)]
Don't pass -Wl,z,defs for now.

It broke the msan build.

llvm-svn: 226613

9 years agoFor llvm-objdump, hook up existing options to work when using -macho (the Mach-O...
Kevin Enderby [Tue, 20 Jan 2015 21:47:46 +0000 (21:47 +0000)]
For llvm-objdump, hook up existing options to work when using -macho (the Mach-O parser).

llvm-svn: 226612

9 years agoUse -Wl,defs when linking.
Rafael Espindola [Tue, 20 Jan 2015 21:23:15 +0000 (21:23 +0000)]
Use -Wl,defs when linking.

ELF linkers by default allow shared libraries to contain undefined references
and it is up to the dynamic linker to look for them.

On COFF and MachO, that is not the case.

This creates a situation where a .so might build on an ELF system, but the build
of the corresponding .dylib or .dll will fail.

This patch changes the cmake build to use -Wl,-z,defs when linking and updates
the dependencies so that -DBUILD_SHARED_LIBS=ON build still works.

llvm-svn: 226611

9 years ago[PM] Reformat this code with clang-format so that subsequent changes
Chandler Carruth [Tue, 20 Jan 2015 21:10:35 +0000 (21:10 +0000)]
[PM] Reformat this code with clang-format so that subsequent changes
don't get muddied up by formatting changes.

Some of these don't really seem like improvements to me, but they also
don't seem any worse and I care much more about not formatting them
manually than I do about the particular formatting. =]

llvm-svn: 226610

9 years agoLink libclang with dl if available.
Rafael Espindola [Tue, 20 Jan 2015 21:10:35 +0000 (21:10 +0000)]
Link libclang with dl if available.

This is in preparation for changing the link to use -Wl,-z,defs.

llvm-svn: 226609

9 years agoRemove unused function:
Greg Clayton [Tue, 20 Jan 2015 21:05:21 +0000 (21:05 +0000)]
Remove unused function:

lldb::pid_t
Host::LaunchApplication (const FileSpec &app_file_spec);

This had use of a function FSPathMakeRef(const UInt8*, FSRef *, ...) that was deprecated in 10.8.

Removing this fucntion since it wasn't used and was causing warnings.

llvm-svn: 226608

9 years agoUpdate the deployment target on MacOSX to be 10.8. The change is mainly to get the...
Greg Clayton [Tue, 20 Jan 2015 20:45:49 +0000 (20:45 +0000)]
Update the deployment target on MacOSX to be 10.8. The change is mainly to get the REG_ENHANCED stuff in regex.h and shouldn't affect other targets.

llvm-svn: 226607

9 years ago[Hexagon] Adding intrinsics for doubleword ALU operations.
Colin LeMahieu [Tue, 20 Jan 2015 20:45:05 +0000 (20:45 +0000)]
[Hexagon] Adding intrinsics for doubleword ALU operations.

llvm-svn: 226606

9 years agoFix crashes on missing @interface for category
Ben Langmuir [Tue, 20 Jan 2015 20:41:36 +0000 (20:41 +0000)]
Fix crashes on missing @interface for category

In a few places we didn't check that Category->getClassInterface() was
not null before using it.

llvm-svn: 226605

9 years agoAdd the test that was supposed to be included with r223162.
Kaelyn Takata [Tue, 20 Jan 2015 20:15:29 +0000 (20:15 +0000)]
Add the test that was supposed to be included with r223162.

The test case is based on the reduction from PR21679 and has to be
freestanding to work correctly, since some of the expected errors (and
some of the problems that were fixed) only occur when the end of the
file is reached.

llvm-svn: 226603

9 years ago[Hexagon] Removing unnecessary clutter in intrinsic tests.
Colin LeMahieu [Tue, 20 Jan 2015 19:46:07 +0000 (19:46 +0000)]
[Hexagon] Removing unnecessary clutter in intrinsic tests.

llvm-svn: 226602

9 years agoImplement command line options for stack probe space
Hans Wennborg [Tue, 20 Jan 2015 19:45:50 +0000 (19:45 +0000)]
Implement command line options for stack probe space

This code adds the -mstack-probe-size command line option and implements the /Gs
compiler switch for clang-cl.

This should fix http://llvm.org/bugs/show_bug.cgi?id=21896

Patch by Andrew H!

Differential Revision: http://reviews.llvm.org/D6685

llvm-svn: 226601

9 years agoPrevent binary-tree deterioration in sparse switch statements.
Daniel Jasper [Tue, 20 Jan 2015 19:43:33 +0000 (19:43 +0000)]
Prevent binary-tree deterioration in sparse switch statements.

This addresses part of llvm.org/PR22262. Specifically, it prevents
considering the densities of sub-ranges that have fewer than
TLI.getMinimumJumpTableEntries() elements. Those densities won't help
jump tables.

This is not a complete solution but works around the most pressing
issue.

Review: http://reviews.llvm.org/D7070
llvm-svn: 226600

9 years ago[GC] Verify-pass void vararg functions in gc.statepoint
Ramkumar Ramachandra [Tue, 20 Jan 2015 19:42:46 +0000 (19:42 +0000)]
[GC] Verify-pass void vararg functions in gc.statepoint

With the appropriate Verifier changes, exactracting the result out of a
statepoint wrapping a vararg function crashes. However, a void vararg
function works fine: commit this first step.

Differential Revision: http://reviews.llvm.org/D7071

llvm-svn: 226599

9 years agoReapply: Teach SROA how to update debug info for fragmented variables.
Adrian Prantl [Tue, 20 Jan 2015 19:42:22 +0000 (19:42 +0000)]
Reapply: Teach SROA how to update debug info for fragmented variables.
This reapplies r225379.

ChangeLog:
- The assertion that this commit previously ran into about the inability
  to handle indirect variables has since been removed and the backend
  can handle this now.
- Testcases were upgrade to the new MDLocation format.
- Instead of keeping a DebugDeclares map, we now use
  llvm::FindAllocaDbgDeclare().

Original commit message follows.

Debug info: Teach SROA how to update debug info for fragmented variables.
This allows us to generate debug info for extremely advanced code such as

 typedef struct { long int a; int b;} S;

 int foo(S s) {
   return s.b;
 }

which at -O1 on x86_64 is codegen'd into

 define i32 @foo(i64 %s.coerce0, i32 %s.coerce1) #0 {
   ret i32 %s.coerce1, !dbg !24
 }

with this patch we emit the following debug info for this

 TAG_formal_parameter [3]
   AT_location( 0x00000000
                0x0000000000000000 - 0x0000000000000006: rdi, piece 0x00000008, rsi, piece 0x00000004
                0x0000000000000006 - 0x0000000000000008: rdi, piece 0x00000008, rax, piece 0x00000004 )
                AT_name( "s" )
                AT_decl_file( "/Volumes/Data/llvm/_build.ninja.release/test.c" )

Thanks to chandlerc, dblaikie, and echristo for their feedback on all
previous iterations of this patch!

llvm-svn: 226598

9 years agoR600/SI: Add subtarget feature to enable VGPR spilling for all shader types
Tom Stellard [Tue, 20 Jan 2015 19:33:04 +0000 (19:33 +0000)]
R600/SI: Add subtarget feature to enable VGPR spilling for all shader types

This is disabled by default, but can be enabled with the subtarget
feature: 'vgpr-spilling'

llvm-svn: 226597

9 years agoR600/SI: Fix simple-loop.ll test
Tom Stellard [Tue, 20 Jan 2015 19:33:02 +0000 (19:33 +0000)]
R600/SI: Fix simple-loop.ll test

llvm-svn: 226596

9 years agoReverted revision 226577.
Jozef Kolek [Tue, 20 Jan 2015 19:29:28 +0000 (19:29 +0000)]
Reverted revision 226577.

llvm-svn: 226595

9 years ago[PM] Clean up a bunch of the doxygen / API docs on the InstCombiner pass
Chandler Carruth [Tue, 20 Jan 2015 19:27:58 +0000 (19:27 +0000)]
[PM] Clean up a bunch of the doxygen / API docs on the InstCombiner pass
prior to refactoring it.

llvm-svn: 226594

9 years ago[llvm link] Destroy ConstantArrays in LLVMContext if they are not used.
Manman Ren [Tue, 20 Jan 2015 19:24:59 +0000 (19:24 +0000)]
[llvm link] Destroy ConstantArrays in LLVMContext if they are not used.

ConstantArrays constructed during linking can cause quadratic memory
explosion. An example is the ConstantArrays constructed when linking in
GlobalVariables with appending linkage.

Releasing all unused constants can cause a 20% LTO compile-time
slowdown for a large application. So this commit releases unused ConstantArrays
only.

rdar://19040716. It reduces memory footprint from 20+G to 6+G.

llvm-svn: 226592

9 years agoR600/SI: Remove stray debugging code from r226586
Tom Stellard [Tue, 20 Jan 2015 19:24:31 +0000 (19:24 +0000)]
R600/SI: Remove stray debugging code from r226586

llvm-svn: 226591

9 years agoDisable C4062 on Windows.
Zachary Turner [Tue, 20 Jan 2015 19:23:44 +0000 (19:23 +0000)]
Disable C4062 on Windows.

The use of std::future introduces an implicit dependency on the PPL
subcomponent of ConcRT.  ConcRT in general is pretty noisy with
warnings, so this patch just disables one of the noisy warnings.

llvm-svn: 226590

9 years ago[PM] Don't spend time making self moves no-ops. They're allowed to leave
Chandler Carruth [Tue, 20 Jan 2015 18:54:16 +0000 (18:54 +0000)]
[PM] Don't spend time making self moves no-ops. They're allowed to leave
the object in a moved-from state, and its simpler to write the code that
way.

llvm-svn: 226589

9 years agoAdd an assertion and prefer a crash over an infinite loop.
Adrian Prantl [Tue, 20 Jan 2015 18:03:37 +0000 (18:03 +0000)]
Add an assertion and prefer a crash over an infinite loop.

llvm-svn: 226588

9 years agoDon't remove backslashes from arguments unless the following char is recognized.
Zachary Turner [Tue, 20 Jan 2015 17:53:36 +0000 (17:53 +0000)]
Don't remove backslashes from arguments unless the following char is recognized.

This fixes file paths on Windows, as you can now write, for example,
file d:\foo\bar.txt, but does not break the case that this tokenization
logic was originally designed for, which is to allow escaping of things
like quotes and double quotes, so that all of the escapable characters
can appear in the same string together.

Reviewed by: Jim Ingham, Greg Clayton
Differential Revision: http://reviews.llvm.org/D7018

llvm-svn: 226587

9 years agoR600/SI: Use external symbols for scratch buffer
Tom Stellard [Tue, 20 Jan 2015 17:49:47 +0000 (17:49 +0000)]
R600/SI: Use external symbols for scratch buffer

We were passing the scratch buffer address to the shaders via user sgprs,
but now we use external symbols and have the driver patch the shader
using reloc information.

llvm-svn: 226586

9 years agoR600/SI: Add kill flag when copying scratch offset to a register
Tom Stellard [Tue, 20 Jan 2015 17:49:45 +0000 (17:49 +0000)]
R600/SI: Add kill flag when copying scratch offset to a register

This allows us to re-use the same register for the scratch offset
when accessing large private arrays.

llvm-svn: 226585