oota-llvm.git
6 years agoFix unaligned memory read issue exposed by ubsan
Xinliang David Li [Tue, 17 Nov 2015 03:47:21 +0000 (03:47 +0000)]
Fix unaligned memory read issue exposed by ubsan

Indexed profile data as designed today does not guarantee
counter data to be well aligned, so reading needs to use
the slower form (with memcpy). This is less than ideal and
should be improved in the future (i.e., with fixed length
function key instead of variable length name key).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253309 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Documentation] Add guidelines for grouping tests together.
Davide Italiano [Tue, 17 Nov 2015 02:17:35 +0000 (02:17 +0000)]
[Documentation] Add guidelines for grouping tests together.

This was considered a good practice but it was not documented. Now it is.

Differential Revision: http://reviews.llvm.org/D14733

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253281 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDrop prelink support.
Rafael Espindola [Tue, 17 Nov 2015 00:51:23 +0000 (00:51 +0000)]
Drop prelink support.

The way prelink used to work was

* The compiler decides if a given section only has relocations that
are know to point to the same DSO. If so, it names it
.data.rel.ro.local<something>.
* The static linker puts all of these together.
* The prelinker program assigns addresses to each library and resolves
the local relocations.

There are many problems with this:
* It is incompatible with address space randomization.
* The information passed by the compiler is redundant. The linker
knows if a given relocation is in the same DSO or not. If could sort
by that if so desired.
* There are newer ways of speeding up DSO (gnu hash for example).
* Even if we want to implement this again in the compiler, the previous
  implementation is pretty broken. It talks about relocations that are
  "resolved by the static linker". If they are resolved, there are none
  left for the prelinker. What one needs to track is if an expression
  will require only dynamic relocations that point to the same DSO.

At this point it looks like the prelinker is an historical curiosity.
For example, fedora has retired it because it failed to build for two
releases
(http://pkgs.fedoraproject.org/cgit/prelink.git/commit/?id=eb43100a8331d91c801ee3dcdb0a0bb9babfdc1f)

This patch removes support for it. That is, it stops printing the
".local" sections.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253280 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAssume lane masks are always precise
Matthias Braun [Tue, 17 Nov 2015 00:50:55 +0000 (00:50 +0000)]
Assume lane masks are always precise

Allowing imprecise lane masks in case of more than 32 sub register lanes
lead to some tricky corner cases, and I need another bugfix for another
one. Instead I rather declare lane masks as precise and let tablegen
abort if we do not have enough bits.

This does not affect any in-tree target, even AMDGPU only needs 16 lanes
at the moment. If the 32 lanes turn out to be a problem in the future,
then we can easily change the LaneBitmask typedef to uint64_t.

Differential Revision: http://reviews.llvm.org/D14557

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253279 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix indentation
David Blaikie [Tue, 17 Nov 2015 00:41:02 +0000 (00:41 +0000)]
Fix indentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253278 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agodwarfdump: Use the index to find the right abbrev offset in DWP files
David Blaikie [Tue, 17 Nov 2015 00:39:55 +0000 (00:39 +0000)]
dwarfdump: Use the index to find the right abbrev offset in DWP files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253277 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Fix printing of global operands
Derek Schuff [Tue, 17 Nov 2015 00:20:44 +0000 (00:20 +0000)]
[WebAssembly] Fix printing of global operands

This was regressed in r252656 which wasn't quite NFC. Instead of using a
custom instruction as before, use a pattern to select CONST_I32 for the
global addrs.

Differential Revision: http://reviews.llvm.org/D14587

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253276 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PRE] Preserve !invariant.load metadata
Philip Reames [Tue, 17 Nov 2015 00:15:09 +0000 (00:15 +0000)]
[PRE] Preserve !invariant.load metadata

Spoted via inspection.  Test case included.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253275 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Merged BLEND shuffle decode comments. NFC.
Simon Pilgrim [Mon, 16 Nov 2015 23:03:18 +0000 (23:03 +0000)]
[X86][SSE] Merged BLEND shuffle decode comments. NFC.

Now that we can recognise different vector sizes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253268 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDemote a single-use named function object to a lambda
David Blaikie [Mon, 16 Nov 2015 22:56:30 +0000 (22:56 +0000)]
Demote a single-use named function object to a lambda

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253267 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Merged ALIGNR/SLLDQ/SRLDQ shuffle decode comments. NFC.
Simon Pilgrim [Mon, 16 Nov 2015 22:54:41 +0000 (22:54 +0000)]
[X86][SSE] Merged ALIGNR/SLLDQ/SRLDQ shuffle decode comments. NFC.

Now that we can recognise different vector sizes - will make future AVX512 additions easier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253266 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Merged SHUF/PERM shuffle decode comments. NFC.
Simon Pilgrim [Mon, 16 Nov 2015 22:39:27 +0000 (22:39 +0000)]
[X86][SSE] Merged SHUF/PERM shuffle decode comments. NFC.

Now that we can recognise different vector sizes - will make future AVX512 additions easier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253260 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Merged UNPCK shuffle decode comments. NFC.
Simon Pilgrim [Mon, 16 Nov 2015 22:21:10 +0000 (22:21 +0000)]
[X86][SSE] Merged UNPCK shuffle decode comments. NFC.

Now that we can recognise different vector sizes - will make future AVX512 additions easier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253258 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agouse range-based for loop; NFCI
Sanjay Patel [Mon, 16 Nov 2015 22:16:52 +0000 (22:16 +0000)]
use range-based for loop; NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253256 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd isInteger() to APFloat.
Stephen Canon [Mon, 16 Nov 2015 21:52:48 +0000 (21:52 +0000)]
Add isInteger() to APFloat.

Useful utility function; this wasn't too hard to do before, but also wasn't
obviously discoverable.  Make it explicit.  Reviewed offline by Michael
Gottesman.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253254 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PR25538]: Fix a failure caused by r253126.
Michael Zolotukhin [Mon, 16 Nov 2015 21:17:26 +0000 (21:17 +0000)]
[PR25538]: Fix a failure caused by r253126.

In r253126 we stopped to recompute LCSSA after loop unrolling in all
cases, except the unrolling is full and at least one of the loop exits
is outside the parent loop. In other cases the transformation should not
break LCSSA, but it turned out, that we also call SimplifyLoop on the
parent loop, which might break LCSSA by itself. This fix just triggers
LCSSA recomputation in this case as well.

I'm committing it without a test case for now, but I'll try to invent
one. It's a bit tricky because in an isolated test LoopSimplify would
be scheduled before LoopUnroll, and thus will change the test and hide
the problem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253253 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Fix function return type printing
Derek Schuff [Mon, 16 Nov 2015 21:12:41 +0000 (21:12 +0000)]
[WebAssembly] Fix function return type printing

Summary:
Previously return type information for a function was derived from
return dag nodes. But this didn't work for dags with != return node. So
instead compute it directly from the LLVM function as is done for imports.

Differential Revision: http://reviews.llvm.org/D14593

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253251 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Reverse the order of operands for br_if
Derek Schuff [Mon, 16 Nov 2015 21:04:51 +0000 (21:04 +0000)]
[WebAssembly] Reverse the order of operands for br_if

Summary: This is to match the new version in the spec

Reviewers: sunfish

Subscribers: jfb, llvm-commits, dschuff

Differential Revision: http://reviews.llvm.org/D14519

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253249 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[IR] Manage TheNoneToken with a std::unique_ptr
David Majnemer [Mon, 16 Nov 2015 20:55:57 +0000 (20:55 +0000)]
[IR] Manage TheNoneToken with a std::unique_ptr

Hopefully, this will make the sanitizer build bots happy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253248 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFind available scratch register to use in function prologue and epilogue as part...
Kit Barton [Mon, 16 Nov 2015 20:22:15 +0000 (20:22 +0000)]
Find available scratch register to use in function prologue and epilogue as part of shrink wrapping.

Phabricator: http://reviews.llvm.org/D13955

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253247 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WinEH] Don't let UnwindHelp alias the return address
Reid Kleckner [Mon, 16 Nov 2015 18:47:25 +0000 (18:47 +0000)]
[WinEH] Don't let UnwindHelp alias the return address

On top of that, don't bother allocating and initializing UnwindHelp if
we don't have any funclets. Currently we always use RBP as our frame
pointer when funclets are present, so this change makes it impossible to
come here without any fixed stack objects.

Fixes PR25533.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253245 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUse the subtarget reference that we already have
Reid Kleckner [Mon, 16 Nov 2015 18:47:12 +0000 (18:47 +0000)]
Use the subtarget reference that we already have

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253244 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd intermediate subtract instructions to reassociation worklist.
Owen Anderson [Mon, 16 Nov 2015 18:07:30 +0000 (18:07 +0000)]
Add intermediate subtract instructions to reassociation worklist.

We sometimes create intermediate subtract instructions during
reassociation.  Adding these to the worklist to revisit exposes many
additional reassociation opportunities.

Patch by Aditya Nandakumar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253240 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LoopStrengthReduce] Don't increment iterator past the end of the BB
David Majnemer [Mon, 16 Nov 2015 17:37:58 +0000 (17:37 +0000)]
[LoopStrengthReduce] Don't increment iterator past the end of the BB

We tried to move the insertion point beyond instructions like landingpad
and cleanuppad.
However, we *also* tried to move past catchpad.  This is problematic
because catchpad is also a terminator.

This fixes PR25541.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253238 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips] Disable code generation through FastISel for MIPS32R6.
Vasileios Kalintiris [Mon, 16 Nov 2015 17:05:01 +0000 (17:05 +0000)]
[mips] Disable code generation through FastISel for MIPS32R6.

Reviewers: dsanders

Subscribers: llvm-commits, dsanders

Differential Revision: http://reviews.llvm.org/D14708

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253225 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SimplifyLibCalls] Generalize a comment. This doesn't apply only to sqrt.
Davide Italiano [Mon, 16 Nov 2015 16:54:28 +0000 (16:54 +0000)]
[SimplifyLibCalls] Generalize a comment. This doesn't apply only to sqrt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253224 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Prevent use of a value pointed by end() iterator when placing a jump table
Petr Pavlu [Mon, 16 Nov 2015 16:41:13 +0000 (16:41 +0000)]
[ARM] Prevent use of a value pointed by end() iterator when placing a jump table

Function ARMConstantIslands::doInitialJumpTablePlacement() iterates over all
basic blocks in a machine function. It calls `MI = MBB.getLastNonDebugInstr()`
to get the last instruction in each block and then uses MI->getOpcode() to
decide what to do. If getLastNonDebugInstr() returns MBB.end() (for example,
when the block does not contain any instructions) then calling getOpcode() on
this value is incorrect. Avoid this problem by checking the result of
getLastNonDebugInstr().

Differential Revision: http://reviews.llvm.org/D14694

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253222 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM,AArch64] Store source location of asm constant pool entries
Oliver Stannard [Mon, 16 Nov 2015 16:25:47 +0000 (16:25 +0000)]
[ARM,AArch64] Store source location of asm constant pool entries

Storing the source location of the expression that created a constant pool
entry allows us to emit better error messages if we later discover that the
expression cannot be represented by a relocation.

Differential Revision: http://reviews.llvm.org/D14646

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253220 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM,AArch64] Store source location for values in assembly files
Oliver Stannard [Mon, 16 Nov 2015 16:22:47 +0000 (16:22 +0000)]
[ARM,AArch64] Store source location for values in assembly files

The MCValue class can store a SMLoc to allow better error messages to be
emitted if an error is detected after parsing. The ARM and AArch64 assembly
parsers were not setting this, so error messages did not have source
information.

Differential Revision: http://reviews.llvm.org/D14645

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253219 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Prototype passes for register coloring and register stackifying.
Dan Gohman [Mon, 16 Nov 2015 16:18:28 +0000 (16:18 +0000)]
[WebAssembly] Prototype passes for register coloring and register stackifying.

These passes are not yet enabled by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253217 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips][ias] Remove spurious ';' from inline assembly test.
Daniel Sanders [Mon, 16 Nov 2015 14:19:32 +0000 (14:19 +0000)]
[mips][ias] Remove spurious ';' from inline assembly test.

IAS will not emit it. NFC at the moment but will prevent a test failure once
IAS is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253210 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips][ias] Accept $31 or $ra in hf16call32.ll. IAS prints the latter.
Daniel Sanders [Mon, 16 Nov 2015 14:16:45 +0000 (14:16 +0000)]
[mips][ias] Accept $31 or $ra in hf16call32.ll. IAS prints the latter.

NFC at the moment, but it will prevent a test failure once IAS is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253209 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips][ias] Allow whitespace after commas in inlineasm*.ll tests.
Daniel Sanders [Mon, 16 Nov 2015 14:14:59 +0000 (14:14 +0000)]
[mips][ias] Allow whitespace after commas in inlineasm*.ll tests.

IAS always prints whitespace after a comma. NFC at the moment but this will
prevent failures when IAS is enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253208 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoHandle ARMv6KZ naming
Artyom Skrobov [Mon, 16 Nov 2015 14:05:32 +0000 (14:05 +0000)]
Handle ARMv6KZ naming

Summary:
* ARMv6KZ is the "canonical" name, given in the ARMARM
* ARMv6Z is an "official abbreviation" for it, mentioned in the ARMARM
* ARMv6ZK is a popular misspelling, which we should support as an alias.

The patch corrects the handling of the names.

Functional changes:
* ARMv6Z no longer treated as an architecture in its own right
* ARMv6ZK renamed to ARMv6KZ, accepting ARMv6ZK as an alias
* arm1176jz-s and arm1176jzf-s recognized as ARMv6ZK, instead of ARMv6K
* default ARMv6K CPU changed to arm1176j-s

Reviewers: rengolin, logan, compnerd

Subscribers: aemerson, llvm-commits, rengolin

Differential Revision: http://reviews.llvm.org/D14568

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253206 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoNFC refactorings in lib/Support/TargetParser.cpp
Artyom Skrobov [Mon, 16 Nov 2015 12:08:05 +0000 (12:08 +0000)]
NFC refactorings in lib/Support/TargetParser.cpp

Summary:
* declare FPUNames, ARCHNames, ARCHExtNames, HWDivNames, CPUNames
  as static const
* implement getDefaultExtensions with a StringSwitch, in the same
  way getDefaultFPU is implemented

Reviewers: rengolin

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14648

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253201 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Allow TargetParser to accurately target architectures
Bradley Smith [Mon, 16 Nov 2015 11:15:22 +0000 (11:15 +0000)]
[ARM] Allow TargetParser to accurately target architectures

Instead of defaulting to an empty string, we want to default to
the CPU 'generic' in the case of no valid default CPU being found,
(as long as the architecture is actually valid).

In order to do this we add a default FPU for each architecture, as
well as falling back to architecture defaults for extensions and FPU
in the case of a generic CPU is specified.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253198 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Introduce subtarget features per ARM architecture.
Bradley Smith [Mon, 16 Nov 2015 11:10:19 +0000 (11:10 +0000)]
[ARM] Introduce subtarget features per ARM architecture.

This allows for accurate architecture targeting as well as removing
duplicate information (hardcoded feature strings) from MCTargetDesc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253196 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoProperly check if a CMPZ node is in fact comparing against zero
James Molloy [Mon, 16 Nov 2015 10:49:25 +0000 (10:49 +0000)]
Properly check if a CMPZ node is in fact comparing against zero

This was left implicit and never ever checked, which means we could have a CMPZ against some non-zero value and we were carrying on with BFI conversion regardless.

Caught by Oliver Stannard using csmith; regression test added.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253195 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDon't generate discriminators for calls to debug intrinsics
Pavel Labath [Mon, 16 Nov 2015 10:40:38 +0000 (10:40 +0000)]
Don't generate discriminators for calls to debug intrinsics

Summary:
This fails a check in Verifier.cpp, which checks for location matches between the declared
variable and the !dbg attachments.

Reviewers: dnovillo, dblaikie, danielcdh

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14657

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253194 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AArch64] ldr= pseudo-instruction silently ignored if register invalid
Oliver Stannard [Mon, 16 Nov 2015 10:25:19 +0000 (10:25 +0000)]
[AArch64] ldr= pseudo-instruction silently ignored if register invalid

The AArch64 assembler was silently ignoring instructions like this:
  ldr foo, =bar

AArch64AsmParser::parseOperand was returning true as the parse failed, but was
not calling AArch64AsmParser::Error to report this to the user, so the
instruction was ignored without printing an error message.

Differential Revision: http://reviews.llvm.org/D14651

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253193 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[GlobalOpt] Address post-commit review comments on r253168
James Molloy [Mon, 16 Nov 2015 10:16:22 +0000 (10:16 +0000)]
[GlobalOpt] Address post-commit review comments on r253168

Address Duncan Exon Smith's comments on D14148, which was added after the patch had been LGTM'd and committed:
  * clang-format one area where whitespace diffs occurred.
  * Add a threshold to limit the store/load dominance checks as they are quadratic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253192 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMove helper classes into anonymous namespaces. NFC.
Benjamin Kramer [Mon, 16 Nov 2015 09:01:28 +0000 (09:01 +0000)]
Move helper classes into anonymous namespaces. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253189 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix r253186 test case
Keno Fischer [Mon, 16 Nov 2015 08:25:14 +0000 (08:25 +0000)]
Fix r253186 test case

Referencing a DILocation whose scope is a different subprogram causes
an assertion failure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253187 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[DIBuilder] Make createReferenceType take size and align
Keno Fischer [Mon, 16 Nov 2015 07:57:32 +0000 (07:57 +0000)]
[DIBuilder] Make createReferenceType take size and align

Summary: Since we're passing references to dbg.value as pointers,
we need to have the frontend properly declare their sizes and
alignments (as it already does for regular pointers) in preparation
for my upcoming patch to have the verifer check that the sizes agree.

Also augment the backend logic that skips actually emitting this
information into DWARF such that it also handles reference types.

Reviewers: aprantl, dexonsmith, dblaikie

Subscribers: dblaikie, llvm-commits

Differential Revision: http://reviews.llvm.org/D14275

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253186 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAVX512: Implemented encoding and intrinsics for VMOVSHDUP/VMOVSLDUP instructions.
Igor Breger [Mon, 16 Nov 2015 07:22:00 +0000 (07:22 +0000)]
AVX512: Implemented encoding and intrinsics for VMOVSHDUP/VMOVSLDUP instructions.

Differential Revision: http://reviews.llvm.org/D14322

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253185 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAlso map the personality function in CloneFunctionInto
Keno Fischer [Mon, 16 Nov 2015 05:13:30 +0000 (05:13 +0000)]
Also map the personality function in CloneFunctionInto

Summary: The Old personality function gets copied over, but the
Materializer didn't  have a chance to inspect it (e.g. to fix up
references to the correct module for the target function).
Also add a verifier check that makes sure the personality routine
is in the same module as the function whose personality it is.

Reviewers: majnemer

Subscribers: jevinskie, llvm-commits

Differential Revision: http://reviews.llvm.org/D14474

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253183 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Sink] Don't move landingpads
Keno Fischer [Mon, 16 Nov 2015 04:47:58 +0000 (04:47 +0000)]
[Sink] Don't move landingpads

Summary: Moving landingpads into successor basic blocks makes the
verifier sad. Teach Sink that much like PHI nodes and terminator
instructions, landingpads (and cleanuppads, etc.) may not be moved
between basic blocks.

Reviewers: majnemer

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14475

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253182 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ADT] Make capacity_in_bytes support BitVector. NFC.
Vedant Kumar [Mon, 16 Nov 2015 04:02:36 +0000 (04:02 +0000)]
[ADT] Make capacity_in_bytes support BitVector. NFC.

This makes it a bit easier to replace instances of vector<bool> with
BitVector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253180 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Use tabs instead of spaces in assembly output.
Dan Gohman [Sun, 15 Nov 2015 15:34:19 +0000 (15:34 +0000)]
[WebAssembly] Use tabs instead of spaces in assembly output.

This seems to be the most popular convention among the other backends.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253172 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Tidyup with implicit SDValue bool check. NFC.
Simon Pilgrim [Sun, 15 Nov 2015 14:57:07 +0000 (14:57 +0000)]
[X86][SSE] Tidyup with implicit SDValue bool check. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253171 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix mapping of unmaterialized global values during metadata linking
Teresa Johnson [Sun, 15 Nov 2015 14:50:14 +0000 (14:50 +0000)]
Fix mapping of unmaterialized global values during metadata linking

Summary:
The patch to move metadata linking after global value linking didn't
correctly map unmaterialized global values to null as desired. They
were in fact mapped to the source copy. It largely worked by accident
since most module linker clients destroyed the source module which
caused the source GVs to be replaced by null, but caused a failure with
LTO linking on Windows:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20151109/312869.html

The problem is that a null return value from materializeValueFor is
handled by mapping the value to self. This is the desired behavior when
materializeValueFor is passed a non-GlobalValue. The problem is how to
distinguish that case from the case where we really do want to map to
null.

This patch addresses this by passing in a new flag to the value mapper
indicating that unmapped global values should be mapped to null. Other
Value types are handled as before.

Note that the documented behavior of asserting on unmapped values when
the flag RF_IgnoreMissingValues isn't set is currently disabled with
FIXME notes due to bootstrap failures. I modified these disabled asserts
so when they are eventually enabled again it won't assert for the
unmapped values when the new RF_NullMapMissingGlobalValues flag is set.

I also considered using a callback into the value materializer, but a
flag seemed cleaner given that there are already existing flags.
I also considered modifying materializeValueFor to return the input
value when we want to map to source and then treat a null return
to mean map to null. However, there are other value materializer
subclasses that implement materializeValueFor, and they would all need
to be audited and the return values possibly changed, which seemed
error-prone.

Reviewers: dexonsmith, joker.eph

Subscribers: pcc, llvm-commits

Differential Revision: http://reviews.llvm.org/D14682

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253170 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[GlobalOpt] Demote globals to locals more aggressively
James Molloy [Sun, 15 Nov 2015 14:21:37 +0000 (14:21 +0000)]
[GlobalOpt] Demote globals to locals more aggressively

Global to local demotion can speed up programs that use globals a lot. It is particularly useful with LTO, when the entire call graph is known and most functions have been internalized.

For a global to be demoted, it must only be accessed by one function and that function:
  1. Must never recurse directly or indirectly, else the GV would be clobbered.
  2. Must never rely on the value in GV at the start of the function (apart from the initializer).

GlobalOpt can already do this, but it is hamstrung and only ever tries to demote globals inside "main", because C++ gives extra guarantees about how main is called - once and only once.

In LTO mode, we can often prove the first property (if the function is internal by this point, we know enough about the callgraph to determine if it could possibly recurse). FunctionAttrs now infers the "norecurse" attribute for this reason.

The second property can be proven for a subset of functions by proving that all loads from GV are dominated by a store to GV. This is conservative in the name of compile time - this only requires a DominatorTree which is fairly cheap in the grand scheme of things. We could do more fancy stuff with MemoryDependenceAnalysis too to catch more cases but this appears to catch most of the useful ones in my testing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253168 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Docs] Fix typo
Alex Denisov [Sun, 15 Nov 2015 14:13:24 +0000 (14:13 +0000)]
[Docs] Fix typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253167 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert r253160.
Igor Breger [Sun, 15 Nov 2015 12:19:11 +0000 (12:19 +0000)]
Revert r253160.

It broke layering violation. Reproducible with BUILD_SHARED_LIBS=ON.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253163 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFixed GEP visitor in the InstCombine pass.
Elena Demikhovsky [Sun, 15 Nov 2015 08:19:35 +0000 (08:19 +0000)]
Fixed GEP visitor in the InstCombine pass.
The current implementation of GEP visitor in InstCombine fails with assertion on Vector GEP with mix of scalar and vector types, like this:

getelementptr double, double* %a, <8 x i32> %i
(It fails to create a "sext" from <8 x i32> to <8 x i64>)

I fixed it and added some tests.

Differential Revision: http://reviews.llvm.org/D14485

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253162 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAVX512: Implemented encoding and intrinsics for VMOVSHDUP/VMOVSLDUP instructions.
Igor Breger [Sun, 15 Nov 2015 07:23:13 +0000 (07:23 +0000)]
AVX512: Implemented encoding and intrinsics for VMOVSHDUP/VMOVSLDUP instructions.

Differential Revision: http://reviews.llvm.org/D14322

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253160 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoNFC: Document MSVC getters on Triple
Dylan McKay [Sun, 15 Nov 2015 05:43:11 +0000 (05:43 +0000)]
NFC: Document MSVC getters on Triple

Summary:
Document the differences between the isKnownWindowsMSVC() and isWindowsMSVC() methods on Triple.

Also removed the '\brief' Doxygen annotations - now that 'AUTOBRIEF' is set to on, they are unnecessary.

Subscribers: jfb, tberghammer, danalbert, srhines, dschuff

Differential Revision: http://reviews.llvm.org/D14110

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253159 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUse a different block id for block of metadata kind records
Teresa Johnson [Sun, 15 Nov 2015 02:00:09 +0000 (02:00 +0000)]
Use a different block id for block of metadata kind records

Summary:
There are currently two blocks with the METADATA_BLOCK id at module
scope. The first has the module-level metadata values (consisting of
some combination of METADATA_* record codes except for METADATA_KIND).
The second consists only of METADATA_KIND records. The latter is used
only in the METADATA_ATTACHMENT block within function blocks (for
metadata attached to instructions).

For ThinLTO we want to delay the parsing of module level metadata
until all functions have been imported from that module (there is some
bookkeeping used to suture it up when we read it during a post-pass).
However, we do need the METADATA_KIND records when parsing the function
body during importing, since those kinds are used as described above.

To simplify identification and parsing of just the block containing
the metadata kinds, use a different block id (METADATA_KIND_BLOCK_ID).
Support older bitcode without the new block id as well.

Reviewers: dexonsmith, joker.eph

Subscribers: davidxl, llvm-commits

Differential Revision: http://reviews.llvm.org/D14654

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253154 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Minor code simplification. NFC.
Dan Gohman [Sat, 14 Nov 2015 23:28:15 +0000 (23:28 +0000)]
[WebAssembly] Minor code simplification. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253150 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Make indentation consistent with the other testcases. NFC.
Dan Gohman [Sat, 14 Nov 2015 23:17:07 +0000 (23:17 +0000)]
[WebAssembly] Make indentation consistent with the other testcases. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253149 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Support signext, zeroext, and several other function attributes.
Dan Gohman [Sat, 14 Nov 2015 23:15:41 +0000 (23:15 +0000)]
[WebAssembly] Support signext, zeroext, and several other function attributes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253148 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Change int_wasm_memory_size from IntrNoMem to IntrReadMem.
Dan Gohman [Sat, 14 Nov 2015 23:02:31 +0000 (23:02 +0000)]
[WebAssembly] Change int_wasm_memory_size from IntrNoMem to IntrReadMem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253147 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Fixed arch/triple and regenerated results.
Simon Pilgrim [Sat, 14 Nov 2015 20:42:01 +0000 (20:42 +0000)]
[X86][SSE] Fixed arch/triple and regenerated results.

Tidyup before diffs from new patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253144 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-ar] Use failIfError/fail helpers.
Davide Italiano [Sat, 14 Nov 2015 19:00:33 +0000 (19:00 +0000)]
[llvm-ar] Use failIfError/fail helpers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253141 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-ar] Use fail() helper to reduce duplication.
Davide Italiano [Sat, 14 Nov 2015 18:33:47 +0000 (18:33 +0000)]
[llvm-ar] Use fail() helper to reduce duplication.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253139 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-ar] Simplify the code.
Davide Italiano [Sat, 14 Nov 2015 18:25:18 +0000 (18:25 +0000)]
[llvm-ar] Simplify the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253138 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Added extra vector truncation tests
Simon Pilgrim [Sat, 14 Nov 2015 15:23:59 +0000 (15:23 +0000)]
[X86][SSE] Added extra vector truncation tests

Baseline comparison to D14588

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253132 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoReduce the size of MCRelaxableFragment.
Akira Hatanaka [Sat, 14 Nov 2015 06:35:56 +0000 (06:35 +0000)]
Reduce the size of MCRelaxableFragment.

MCRelaxableFragment previously kept a copy of MCSubtargetInfo and
MCInst to enable re-encoding the MCInst later during relaxation. A copy
of MCSubtargetInfo (instead of a reference or pointer) was needed
because the feature bits could be modified by the parser.

This commit replaces the MCSubtargetInfo copy in MCRelaxableFragment
with a constant reference to MCSubtargetInfo. The copies of
MCSubtargetInfo are kept in MCContext, and the target parsers are now
responsible for asking MCContext to provide a copy whenever the feature
bits of MCSubtargetInfo have to be toggled.

With this patch, I saw a 4% reduction in peak memory usage when I
compiled verify-uselistorder.lto.bc using llc.

rdar://problem/21736951

Differential Revision: http://reviews.llvm.org/D14346

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253127 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDon't recompute LCSSA after loop-unrolling when possible.
Michael Zolotukhin [Sat, 14 Nov 2015 05:51:41 +0000 (05:51 +0000)]
Don't recompute LCSSA after loop-unrolling when possible.

Summary:
Currently we always recompute LCSSA for outer loops after unrolling an
inner loop. That leads to compile time problem when we have big loop
nests, and we can solve it by avoiding unnecessary work. For instance,
if w eonly do partial unrolling, we don't break LCSSA, so we don't need
to rebuild it. Also, if all exits from the inner loop are inside the
enclosing loop, then complete unrolling won't break LCSSA either.

I replaced unconditional LCSSA recomputation with conditional recomputation +
unconditional assert and added several tests, which were failing when I
experimented with it.

Soon I plan to follow up with a similar patch for recalculation of dominators
tree.

Reviewers: hfinkel, dexonsmith, bogner, joker.eph, chandlerc

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14526

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253126 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MCTargetAsmParser] Move the member varialbes that reference
Akira Hatanaka [Sat, 14 Nov 2015 05:20:05 +0000 (05:20 +0000)]
[MCTargetAsmParser] Move the member varialbes that reference
MCSubtargetInfo in the subclasses into MCTargetAsmParser and define a
member function getSTI.

This is done in preparation for making changes to shrink the size of
MCRelaxableFragment. (see http://reviews.llvm.org/D14346).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253124 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd MMX to the 3dnow enum and propagate changes around. This makes
Eric Christopher [Sat, 14 Nov 2015 03:04:00 +0000 (03:04 +0000)]
Add MMX to the 3dnow enum and propagate changes around. This makes
it somewhat more consistent with how the feature is used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253122 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ShrinkWrapping] Disable the optimization for functions with sanitize like
Quentin Colombet [Sat, 14 Nov 2015 01:55:17 +0000 (01:55 +0000)]
[ShrinkWrapping] Disable the optimization for functions with sanitize like
attribute.

Even if the target supports shrink-wrapping, the prologue and epilogue
must not move because a crash can happen anywhere and sanitizers need
to be able to unwind from the PC of the crash.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253116 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRemove some unused includes
David Blaikie [Sat, 14 Nov 2015 01:32:27 +0000 (01:32 +0000)]
Remove some unused includes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253115 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[RuntimeDyld] Fix indentation and whitespace; NFC
Sanjoy Das [Sat, 14 Nov 2015 00:16:15 +0000 (00:16 +0000)]
[RuntimeDyld] Fix indentation and whitespace; NFC

Whitespace-only change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253105 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAArch64: Default AArch64Subtarget::ReserveX18 to true on darwin
Justin Bogner [Fri, 13 Nov 2015 23:05:46 +0000 (23:05 +0000)]
AArch64: Default AArch64Subtarget::ReserveX18 to true on darwin

Darwin reserves x18, so it's never ABI compliant to generate code that
uses it. Set the default value based on the OS part of the triple
rather than forcing front-ends to set the +reserve-x18 target feature
in order to build correct code for Darwin.

This will make r243310 redundant, so I'll revert that shortly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253102 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMachineScheduler: Print initial pressure in debug dump
Matthias Braun [Fri, 13 Nov 2015 22:30:31 +0000 (22:30 +0000)]
MachineScheduler: Print initial pressure in debug dump

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253097 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMachineScheduler: Improve debug output for "only one node in readyset"
Matthias Braun [Fri, 13 Nov 2015 22:30:29 +0000 (22:30 +0000)]
MachineScheduler: Improve debug output for "only one node in readyset"

When there is only 1 node left in the ready queue and it is picked call
the reason "ONLY1" instead of "NOCAND".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253096 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agotablegen: Add a simple heuristic to get better names for pressure sets
Matthias Braun [Fri, 13 Nov 2015 22:30:27 +0000 (22:30 +0000)]
tablegen: Add a simple heuristic to get better names for pressure sets

Differential Revision: http://reviews.llvm.org/D14597

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253095 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LIR] Add support for creating memcpys from loops with a negative stride.
Chad Rosier [Fri, 13 Nov 2015 21:51:02 +0000 (21:51 +0000)]
[LIR] Add support for creating memcpys from loops with a negative stride.

This allows us to transform the below loop into a memcpy.

void test(unsigned *__restrict__ a, unsigned *__restrict__ b) {
  for (int i = 2047; i >= 0; --i) {
    a[i] = b[i];
  }
}

This is the memcpy version of r251518, which added support for memset with
negative strided loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253091 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Hexagon] Fixing memory leak during relaxation by allocating MCInst in MCContext.
Colin LeMahieu [Fri, 13 Nov 2015 21:45:50 +0000 (21:45 +0000)]
[Hexagon] Fixing memory leak during relaxation by allocating MCInst in MCContext.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253090 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WinEH] Fix ESP management with 32-bit __CxxFrameHandler3
Reid Kleckner [Fri, 13 Nov 2015 21:27:00 +0000 (21:27 +0000)]
[WinEH] Fix ESP management with 32-bit __CxxFrameHandler3

The C++ EH personality automatically restores ESP from the C++ EH
registration node after a catchret. I mistakenly thought it was like
SEH, which does not restore ESP.

It makes sense for C++ EH to differ from SEH here because SEH does not
use funclets for catches, and does not allow catching inside of finally.
C++ EH may need to unwind through multiple catch funclets and eventually
catchret to some outer funclet. Therefore, the runtime has to keep track
of which ESP to use with catchret, rather than having the compiler
reload it manually.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253084 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[safestack] Rewrite isAllocaSafe using SCEV.
Evgeniy Stepanov [Fri, 13 Nov 2015 21:21:42 +0000 (21:21 +0000)]
[safestack] Rewrite isAllocaSafe using SCEV.

Use ScalarEvolution to calculate memory access bounds.
Handle function calls based on readnone/nocapture attributes.
Handle memory intrinsics with constant size.

This change improves both recall and precision of IsAllocaSafe.
See the new tests (ex. BitCastWide) for the kind of code that was wrongly
classified as safe.

SCEV efficiency seems to be limited by the fact the SafeStack runs late
(in CodeGenPrepare), and many loops are unrolled or otherwise not in LCSSA.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253083 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Docs] Fix warning "Title underline too short."
Akira Hatanaka [Fri, 13 Nov 2015 21:09:57 +0000 (21:09 +0000)]
[Docs] Fix warning "Title underline too short."

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253082 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Rename the Const instructions to be upper-case too.
Dan Gohman [Fri, 13 Nov 2015 20:27:45 +0000 (20:27 +0000)]
[WebAssembly] Rename the Const instructions to be upper-case too.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253072 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSamplePGO - Add dump routines for LineLocation, SampleRecord and FunctionSamples
Diego Novillo [Fri, 13 Nov 2015 20:24:28 +0000 (20:24 +0000)]
SamplePGO - Add dump routines for LineLocation, SampleRecord and FunctionSamples

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253071 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Rename memory intrinsics to be upper-case, following convention. NFC.
Dan Gohman [Fri, 13 Nov 2015 20:19:11 +0000 (20:19 +0000)]
[WebAssembly] Rename memory intrinsics to be upper-case, following convention. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253070 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Combine UNPCKL with vector_shuffle into UNPCKH to save one instruction...
Cong Hou [Fri, 13 Nov 2015 19:47:43 +0000 (19:47 +0000)]
[X86][SSE] Combine UNPCKL with vector_shuffle into UNPCKH to save one instruction for sext from v16i8 to v16i16 and v8i16 to v8i32.

This patch is enabling combining UNPCKL with vector_shuffle that moves the upper
half of a vector into the lower half, into a UNPCKH instruction. For example:

t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1, undef:v16i8
t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2

will be combined to:

t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1

Differential revision: http://reviews.llvm.org/D14399

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253067 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agodwarfdump: Add support for dumping the table contents of DWP indexes
David Blaikie [Fri, 13 Nov 2015 19:18:49 +0000 (19:18 +0000)]
dwarfdump: Add support for dumping the table contents of DWP indexes

This is a recommit of 252842 which was reverted in 252859. The issue was
using %s format specifier for a StringRef - used Format's
left_justify(StringRef, int) instead.

It'd be nice to have __attribute__((format(..))) on llvm::format, but
apparently it's only implemented for c-style variadics, not C++ variadic
templates. Perhaps we could fix that & conditionalize the attribute on
such...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253065 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd a comment that should have made my last commit.
Chad Rosier [Fri, 13 Nov 2015 19:13:40 +0000 (19:13 +0000)]
Add a comment that should have made my last commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253063 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd missing triple to WinEH test case
Reid Kleckner [Fri, 13 Nov 2015 19:11:12 +0000 (19:11 +0000)]
Add missing triple to WinEH test case

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253062 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LIR] Factor out the code to compute base ptr for negative strided loops.
Chad Rosier [Fri, 13 Nov 2015 19:11:07 +0000 (19:11 +0000)]
[LIR] Factor out the code to compute base ptr for negative strided loops.

This will allow for the code to be reused in the memcpy optimization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253061 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WinEH] Make UnwindHelp a fixed stack object allocated after XMM CSRs
Reid Kleckner [Fri, 13 Nov 2015 19:06:01 +0000 (19:06 +0000)]
[WinEH] Make UnwindHelp a fixed stack object allocated after XMM CSRs

Now the offset of UnwindHelp in our EH tables and the offset that we
store to in the prologue agree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253059 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Hexagon] Factoring bundle creation in to a utility function.
Colin LeMahieu [Fri, 13 Nov 2015 17:42:46 +0000 (17:42 +0000)]
[Hexagon] Factoring bundle creation in to a utility function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253056 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Add stony support
Tom Stellard [Fri, 13 Nov 2015 17:06:32 +0000 (17:06 +0000)]
AMDGPU: Add stony support

Patch by: Alex Deucher

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253053 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoELFYAML: Add support for parsing AMDGPU section attribute flags
Tom Stellard [Fri, 13 Nov 2015 17:06:29 +0000 (17:06 +0000)]
ELFYAML: Add support for parsing AMDGPU section attribute flags

Reviewers: silvas

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14444

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253052 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Symbolizer] Don't use PE symbol tables to override PDB symbols
Reid Kleckner [Fri, 13 Nov 2015 17:00:36 +0000 (17:00 +0000)]
[Symbolizer] Don't use PE symbol tables to override PDB symbols

Summary:
PE files are stripped by default, and only contain the names of exported
symbols.

The actual reason that we bother to do this override by default is
actually due to a quirk of the way -gline-tables-only is implemented, so
I phrased the check as "if we are symbolizing from dwarf, do the symtab
override".

This fixes lots of Windows ASan tests that I broke in r250582.

Reviewers: samsonov

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D14594

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253051 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agouse range-based for loop; NFCI
Sanjay Patel [Fri, 13 Nov 2015 16:21:23 +0000 (16:21 +0000)]
use range-based for loop; NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253048 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Replace ARMISD::RBIT with ISD::BITREVERSE
James Molloy [Fri, 13 Nov 2015 16:05:22 +0000 (16:05 +0000)]
[ARM] Replace ARMISD::RBIT with ISD::BITREVERSE

ISD::BITREVERSE matches "rbit" completely, so remove ARMISD::RBIT and mark ISD::BITREVERSE as legal, adding a test for lowering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253047 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFixing a typo in docs/CodeGenerator.rst
Artyom Skrobov [Fri, 13 Nov 2015 15:14:04 +0000 (15:14 +0000)]
Fixing a typo in docs/CodeGenerator.rst

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253045 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips][microMIPS] Implement SHRA[_R].PH, SHRAV[_R].PH, SHRAV[_R].QB, SHRAV_R.W, SHRA_...
Zlatko Buljan [Fri, 13 Nov 2015 13:14:25 +0000 (13:14 +0000)]
[mips][microMIPS] Implement SHRA[_R].PH, SHRAV[_R].PH, SHRAV[_R].QB, SHRAV_R.W, SHRA_R.W, SHRL.PH, SHRL.QB, SHRLV.PH and SHRLV.QB instructions
Differential Revision: http://reviews.llvm.org/D14010

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@253041 91177308-0d34-0410-b5e6-96231b3b80d8