oota-llvm.git
6 years agoUse range-based for loops. NFC
Craig Topper [Wed, 14 Oct 2015 04:36:00 +0000 (04:36 +0000)]
Use range-based for loops. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250266 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert r250204 and r250240 due to bot failure. We failed to build PGO-ed clang.
Manman Ren [Wed, 14 Oct 2015 03:04:03 +0000 (03:04 +0000)]
Revert r250204 and r250240 due to bot failure. We failed to build PGO-ed clang.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250264 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[msan] Fix crash on multiplication by a non-integer constant.
Evgeniy Stepanov [Wed, 14 Oct 2015 00:21:13 +0000 (00:21 +0000)]
[msan] Fix crash on multiplication by a non-integer constant.

Fixes PR25160.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250260 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[asan] Disabling speculative loads under asan. Patch by Mike Aizatsky
Kostya Serebryany [Wed, 14 Oct 2015 00:21:05 +0000 (00:21 +0000)]
[asan] Disabling speculative loads under asan. Patch by Mike Aizatsky

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250259 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRename one of our two llvm::GCOVOptions classes to llvm::GCOV::Options. We used
Richard Smith [Wed, 14 Oct 2015 00:04:19 +0000 (00:04 +0000)]
Rename one of our two llvm::GCOVOptions classes to llvm::GCOV::Options. We used
to get away with this because llvm/Support/GCOV.h was an implementation detail
of the llvm-gcov tool, but it's now being used by FDO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250258 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSplit -pedantic and -Wno-long-long into two separate flags so libc++ can remove ...
Eric Fiselier [Tue, 13 Oct 2015 23:51:58 +0000 (23:51 +0000)]
Split -pedantic and -Wno-long-long into two separate flags so libc++ can remove -pedantic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250255 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SCEV] Use `SCEV::isAllOnesValue` directly; NFC.
Sanjoy Das [Tue, 13 Oct 2015 23:28:31 +0000 (23:28 +0000)]
[SCEV] Use `SCEV::isAllOnesValue` directly; NFC.

Instead of `dyn_cast` ing to `SCEVConstant` and checking the contained
`ConstantInteger.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250251 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake] Add compatibility and current versioning to libLTO to match autoconf support.
Chris Bieneman [Tue, 13 Oct 2015 22:54:29 +0000 (22:54 +0000)]
[CMake] Add compatibility and current versioning to libLTO to match autoconf support.

This also adds LLVM_LTO_VERSION_OFFSET to support functional equivalence to autoconf.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250245 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake] libLLVM's compatibility version should be 1.
Chris Bieneman [Tue, 13 Oct 2015 22:54:27 +0000 (22:54 +0000)]
[CMake] libLLVM's compatibility version should be 1.

This is in matching with the autoconf build system.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250244 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSample profile reader - remove dead code. NFC.
Diego Novillo [Tue, 13 Oct 2015 22:48:48 +0000 (22:48 +0000)]
Sample profile reader - remove dead code. NFC.

This removes old remnants from the gcov reader. I missed these when I
re-wrote it recently.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250242 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSample profiles - Add a name table to the binary encoding.
Diego Novillo [Tue, 13 Oct 2015 22:48:46 +0000 (22:48 +0000)]
Sample profiles - Add a name table to the binary encoding.

Binary encoded profiles used to encode all function names inline at
every reference.  This is clearly suboptimal in terms of space.  This
patch fixes this by adding a name table to the header of the file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250241 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUpdate MachineBranchProbabilityInfo::normalizeEdgeWeights to make sure there is no...
Cong Hou [Tue, 13 Oct 2015 22:27:41 +0000 (22:27 +0000)]
Update MachineBranchProbabilityInfo::normalizeEdgeWeights to make sure there is no zero weight in the output, and also add a missing test for JumpThreading.

The test is for the patch in http://reviews.llvm.org/D10979 but was missing when committing that patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250240 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd missing #include, found by modules build.
Richard Smith [Tue, 13 Oct 2015 22:24:10 +0000 (22:24 +0000)]
Add missing #include, found by modules build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250239 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InlineFunction] Correctly inline TerminatePadInst
David Majnemer [Tue, 13 Oct 2015 22:08:17 +0000 (22:08 +0000)]
[InlineFunction] Correctly inline TerminatePadInst

We forgot to append the terminatepad's arguments which resulted in us
treating the old terminatepad as an argument to the new terminatepad
causing us to crash immediately.  Instead, add the old terminatepad's
arguments to the new terminatepad.

This fixes PR25155.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250234 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Remove a TODO comment which is no longer needed. NFC.
Dan Gohman [Tue, 13 Oct 2015 22:06:40 +0000 (22:06 +0000)]
[WebAssembly] Remove a TODO comment which is no longer needed. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250233 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[uselistorder] Insert blank line, NFC
Vedant Kumar [Tue, 13 Oct 2015 21:51:02 +0000 (21:51 +0000)]
[uselistorder] Insert blank line, NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250231 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix layering violation: Analysis should not include headers from Transforms
Richard Smith [Tue, 13 Oct 2015 21:32:32 +0000 (21:32 +0000)]
Fix layering violation: Analysis should not include headers from Transforms
(and in this case isn't even using anything from those headers).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250228 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTypo.
Chad Rosier [Tue, 13 Oct 2015 20:59:16 +0000 (20:59 +0000)]
Typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250224 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTweak to r250117 and change to use ErrorOr and drop isSizeValid for
Kevin Enderby [Tue, 13 Oct 2015 20:48:04 +0000 (20:48 +0000)]
Tweak to r250117 and change to use ErrorOr and drop isSizeValid for
ArchiveMemberHeader, suggestion by Rafael Espíndola.

Also The clang-x86-win2008-selfhost bot still does not like the
malformed-machos 00000031.a test, so removing it for now.  All
the other bots are fine with it however.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250222 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WinEH] Add CoreCLR EH table emission
Joseph Tremoulet [Tue, 13 Oct 2015 20:18:27 +0000 (20:18 +0000)]
[WinEH] Add CoreCLR EH table emission

Summary:
Emit the handler and clause locations immediately after the standard
xdata.
Clauses are emitted in the same order and format used to communiate them
to the CLR Execution Engine.
Add a lit test to verify correct table generation on a small but
interesting example function.

Reviewers: majnemer, andrew.w.kaylor, rnk

Subscribers: pgavlin, AndyAyers, llvm-commits

Differential Revision: http://reviews.llvm.org/D13451

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250219 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Remove implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 20:07:10 +0000 (20:07 +0000)]
AMDGPU: Remove implicit ilist iterator conversions, NFC

One of the changes in lib/Target/AMDGPU/AMDGPUMCInstLower.cpp was a new
one.  Previously, bundle iterators and single-instruction iterators
could be compared to each other (comparing on underlying pointers).
I changed a comparison from using `MBB->end()` to using
`MBB->instr_end()`, since both end iterators should point at the some
place anyway.

I don't think the implicit conversion between the two iterator types is
a good idea since it's fairly easy to accidentally compare to the wrong
thing (they aren't always end iterators).  Otherwise I would have just
added the conversion.

Even with that, no there should be functionality change here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250218 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAArch64: Remove implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 20:02:15 +0000 (20:02 +0000)]
AArch64: Remove implicit ilist iterator conversions, NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250216 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSelectionDAG: Remove implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 19:47:46 +0000 (19:47 +0000)]
SelectionDAG: Remove implicit ilist iterator conversions, NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250214 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoScalar: Remove remaining ilist iterator implicit conversions
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 19:26:58 +0000 (19:26 +0000)]
Scalar: Remove remaining ilist iterator implicit conversions

Remove remaining `ilist_iterator` implicit conversions from
LLVMScalarOpts.

This change exposed some scary behaviour in
lib/Transforms/Scalar/SCCP.cpp around line 1770.  This patch changes a
call from `Function::begin()` to `&Function::front()`, since the return
was immediately being passed into another function that takes a
`Function*`.  `Function::front()` started to assert, since the function
was empty.  Note that `Function::end()` does not point at a legal
`Function*` -- it points at an `ilist_half_node` -- so the other
function was getting garbage before.  (I added the missing check for
`Function::isDeclaration()`.)

Otherwise, no functionality change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250211 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AArch64] Check the size of the vector before accessing its elements.
Akira Hatanaka [Tue, 13 Oct 2015 18:55:34 +0000 (18:55 +0000)]
[AArch64] Check the size of the vector before accessing its elements.

This fixes an assert in AArch64AsmParser::MatchAndEmitInstruction.

rdar://problem/23081753

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250207 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUpdate the branch weight metadata in JumpThreading pass.
Cong Hou [Tue, 13 Oct 2015 18:43:10 +0000 (18:43 +0000)]
Update the branch weight metadata in JumpThreading pass.

Currently in JumpThreading pass, the branch weight metadata is not updated after CFG modification. Consider the jump threading on PredBB, BB, and SuccBB. After jump threading, the weight on BB->SuccBB should be adjusted as some of it is contributed by the edge PredBB->BB, which doesn't exist anymore. This patch tries to update the edge weight in metadata on BB->SuccBB by scaling it by 1 - Freq(PredBB->BB) / Freq(BB->SuccBB).

Differential revision: http://reviews.llvm.org/D10979

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250204 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PGO]: Eliminate calls to __llvm_profile_register_function for Linux.
Xinliang David Li [Tue, 13 Oct 2015 18:39:48 +0000 (18:39 +0000)]
[PGO]: Eliminate calls to __llvm_profile_register_function for Linux.

On Linux, the profile runtime can use __start_SECTNAME and __stop_SECTNAME
symbols defined by the linker to locate the start and end location of
a named section (with C name). This eliminates the need for instrumented
binary to call __llvm_profile_register_function during start-up time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250199 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoScalar: Remove some implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 18:26:00 +0000 (18:26 +0000)]
Scalar: Remove some implicit ilist iterator conversions, NFC

Remove some of the implicit ilist iterator conversions in
LLVMScalarOpts.  More to go.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250197 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoExecutionEngine: Remove implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 18:11:02 +0000 (18:11 +0000)]
ExecutionEngine: Remove implicit ilist iterator conversions, NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250193 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoOrcJIT: Remove implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 18:10:59 +0000 (18:10 +0000)]
OrcJIT: Remove implicit ilist iterator conversions, NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250192 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoIPO: Remove implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 17:51:03 +0000 (17:51 +0000)]
IPO: Remove implicit ilist iterator conversions, NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250187 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoInstrumentation: Remove ilist iterator implicit conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 17:39:10 +0000 (17:39 +0000)]
Instrumentation: Remove ilist iterator implicit conversions, NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250186 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoInterpreter: Remove implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 17:33:41 +0000 (17:33 +0000)]
Interpreter: Remove implicit ilist iterator conversions, NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250185 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoThe issue with the malformed-machos 00000031.a test is that it needed ‘-arch x86_64’
Kevin Enderby [Tue, 13 Oct 2015 17:06:34 +0000 (17:06 +0000)]
The issue with the malformed-machos 00000031.a test is that it needed ‘-arch x86_64’
flag as it was a Mach-O universal file.

The default as to which architecture slice that is dumped without an -arch flag
depends on the host architecture and the contents of the universal file.  The
malformed archive 00000031.a file has both an x86_64 and i386 slice.  So for
for x86_64 hosts only that slice is dumped, for non-x86_64 hosts, which is many
of the bots both slices are dumped.

The test is intended to only check that the malformation of the x86_64 which
has a non-decimal characters in the size field of the archive header so it no
longer crashes.

The problem turned out that the i388 slice of the malformed archive had a
different malformation which was causing the non-x86_64 bots to get this error:

llvm-objdump -macho -disassemble -arch i386 00000031.a
Archive : .00000031.a
00000031.a(c_start.o):
LLVM ERROR: Symbol name entry points before beginning or past end of file.

and causing the test as it was written to fail.  So by adding ‘-arch x86_64’ it
should correct the test and the malformation on the i388 slice will not be
dumped.

Also the removal of the malformed-machos mem-crup-0261.macho was not causing
the issue so that is put back in.

Sorry for the churn on these tests, Kev

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250184 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoInstCombine: Remove ilist iterator implicit conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 16:59:33 +0000 (16:59 +0000)]
InstCombine: Remove ilist iterator implicit conversions, NFC

Stop relying on implicit conversions of ilist iterators in
LLVMInstCombine.  No functionality change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250183 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoBitcodeReader: Remove ilist iterator implicit conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 16:48:55 +0000 (16:48 +0000)]
BitcodeReader: Remove ilist iterator implicit conversions, NFC

Get LLVMBitReader building without relying on `ilist_iterator` implicit
conversions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250181 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WinEH] Iterate state changes instead of invokes
Joseph Tremoulet [Tue, 13 Oct 2015 16:44:30 +0000 (16:44 +0000)]
[WinEH] Iterate state changes instead of invokes

Summary:
Add an iterator that can walk across blocks and which visits the state
transitions rather than state ranges, with explicit transitions to -1
indicating the presence of top-level calls that may throw and cause the
current function to unwind to caller.  This will simplify code that needs
to identify nested try regions.

Refactor SEH and C++EH table generation to use the new
InvokeStateChangeIterator, and remove the InvokeLabelIterator they were
using.

Reviewers: majnemer, andrew.w.kaylor, rnk

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D13623

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250179 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix a couple of comments; NFC
Xinliang David Li [Tue, 13 Oct 2015 16:35:59 +0000 (16:35 +0000)]
Fix a couple of comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250177 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agofunction names should start with a lower case letter; NFC
Sanjay Patel [Tue, 13 Oct 2015 16:23:00 +0000 (16:23 +0000)]
function names should start with a lower case letter; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250174 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agodon't repeat function/class/variable names in comments; NFC
Sanjay Patel [Tue, 13 Oct 2015 15:12:27 +0000 (15:12 +0000)]
don't repeat function/class/variable names in comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250162 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine][SSE4A] Remove broken INSERTQI range combining optimization
Simon Pilgrim [Tue, 13 Oct 2015 14:48:54 +0000 (14:48 +0000)]
[InstCombine][SSE4A] Remove broken INSERTQI range combining optimization

As discussed in D13348 - the INSERTQI range combining code is wrong in that it confuses the insertion bit index with an extraction bit index.

The remaining legal combines are very unlikely (especially once we've converted to shuffles in D13348) so I'm removing the optimization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250160 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[GlobalsAA] Turn GlobalsAA on again by default
James Molloy [Tue, 13 Oct 2015 10:43:57 +0000 (10:43 +0000)]
[GlobalsAA] Turn GlobalsAA on again by default

Now that all the known faults with GlobalsAA have been fixed, flip the big switch on -enable-non-lto-gmr again.

Feel free to pester me with any more bugs found, and don't hesitate to flip the switch back off.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250157 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[GlobalsAA] Don't assume anything about functions that may be overridden
James Molloy [Tue, 13 Oct 2015 10:43:33 +0000 (10:43 +0000)]
[GlobalsAA] Don't assume anything about functions that may be overridden

Weak linkage and friends allow a symbol to be overriden outside the
code generator's model, so GlobalsAA shouldn't assume that anything it
can compute about such a symbol is valid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250156 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTest commit
Christof Douma [Tue, 13 Oct 2015 09:38:21 +0000 (09:38 +0000)]
Test commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250154 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRemove unnecessary argument in ipo iterator
Michael Liao [Tue, 13 Oct 2015 07:26:04 +0000 (07:26 +0000)]
Remove unnecessary argument in ipo iterator

- Otherwise, it won't be compiled if that template is instantiated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250153 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[IndVars] NFC Cleanup.
Sanjoy Das [Tue, 13 Oct 2015 07:17:38 +0000 (07:17 +0000)]
[IndVars] NFC Cleanup.

 - Rename methods according to the LLVM Coding Style
 - Merge adjacent anonymous namespace block
 - Use `auto` in two places

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250152 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix line-ending issue. NFC.
Michael Kuperstein [Tue, 13 Oct 2015 06:22:30 +0000 (06:22 +0000)]
Fix line-ending issue. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250151 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake] LLVM_PROFDATA_FILE only works if you're using clang, so we should error out...
Chris Bieneman [Tue, 13 Oct 2015 05:35:12 +0000 (05:35 +0000)]
[CMake] LLVM_PROFDATA_FILE only works if you're using clang, so we should error out if it is specified when not using clang.

Also updated the CMake docs.

Based on post-commit review of r250108 from Sean Silvas.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250150 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Mark the AAD and AAM aliases as not valid in 64-bit mode.
Craig Topper [Tue, 13 Oct 2015 05:12:07 +0000 (05:12 +0000)]
[X86] Mark the AAD and AAM aliases as not valid in 64-bit mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250148 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Change all the i8imm operands in XOP instructions to u8imm so the parser will...
Craig Topper [Tue, 13 Oct 2015 05:06:25 +0000 (05:06 +0000)]
[X86] Change all the i8imm operands in XOP instructions to u8imm so the parser will check the size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250147 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert 250089 due to bot failure. It failed when building clang itself with PGO.
Manman Ren [Tue, 13 Oct 2015 03:38:02 +0000 (03:38 +0000)]
Revert 250089 due to bot failure. It failed when building clang itself with PGO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250145 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoBitcodeWriter: Stop using implicit ilist iterator conversion, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 03:26:19 +0000 (03:26 +0000)]
BitcodeWriter: Stop using implicit ilist iterator conversion, NFC

Now LLVMBitWriter compiles without implicit ilist iterator conversions.

In these cases, the cleanest thing was to switch to range-based for
loops.  Since there wasn't much noise I converted sub-loops and parent
loops as a drive-by.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250144 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SCEV] Put some utilites in the ScalarEvolution class
Sanjoy Das [Tue, 13 Oct 2015 02:53:27 +0000 (02:53 +0000)]
[SCEV] Put some utilites in the ScalarEvolution class

In a later commit, `SplitBinaryAdd` will be used outside `IsConstDiff`,
so lift that out.  And lift out `IsConstDiff` as
`computeConstantDifference` to keep things clean and to avoid playing
C++ access specifier games.

NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250143 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTransformUtils: Remove implicit ilist iterator conversions, NFC
Duncan P. N. Exon Smith [Tue, 13 Oct 2015 02:39:05 +0000 (02:39 +0000)]
TransformUtils: Remove implicit ilist iterator conversions, NFC

Continuing the work from last week to remove implicit ilist iterator
conversions.  First related commit was probably r249767, with some more
motivation in r249925.  This edition gets LLVMTransformUtils compiling
without the implicit conversions.

No functional change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250142 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoLooks like malformed-machos 00000031.a test is just getting a different error
Kevin Enderby [Tue, 13 Oct 2015 01:27:28 +0000 (01:27 +0000)]
Looks like malformed-machos 00000031.a test is just getting a different error
on some of the bots.  I’ll remove this test for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250141 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDAGCombiner: Don't stop finding better chain on 2 aliases
Matt Arsenault [Tue, 13 Oct 2015 00:49:00 +0000 (00:49 +0000)]
DAGCombiner: Don't stop finding better chain on 2 aliases

The comment says this was stopped because it was unlikely to be
profitable. This is not true if you want to combine vector loads
with multiple components.

For a simple case that looks like

t0 = load t0 ...
t1 = load t0 ...
t2 = load t0 ...
t3 = load t0 ...

t4 = store t0:1, t0:1
  t5 = store t4, t1:0
    t6 = store t5, t2:0
  t7 = store t6, t3:0

We want to get all of these stores onto a chain
that is a TokenFactor of these N loads. This mostly
solves the AMDGPU merge-stores.ll regressions
with -combiner-alias-analysis for merging vector
stores of vector loads.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250138 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agox86: preserve flags when folding atomic operations
JF Bastien [Tue, 13 Oct 2015 00:28:47 +0000 (00:28 +0000)]
x86: preserve flags when folding atomic operations

Summary:
D4796 taught LLVM to fold some atomic integer operations into a single
instruction. The pattern was unaware that the instructions clobbered
flags.

This patch adds the missing EFLAGS definition.

Floating point operations don't set flags, the subsequent fadd
optimization is therefore correct. The same applies for surrounding
load/store optimizations.

Reviewers: rsmith, rtrieu

Subscribers: llvm-commits, reames, morisset

Differential Revision: http://reviews.llvm.org/D13680

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250135 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Refactor isVGPRToSGPRCopy
Matt Arsenault [Tue, 13 Oct 2015 00:07:54 +0000 (00:07 +0000)]
AMDGPU: Refactor isVGPRToSGPRCopy

It should now correctly handle physical registers and make
it easier to identify the other direction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250132 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRemove the correct unstable malformed-machos test mem-crup-0261.macho and
Kevin Enderby [Tue, 13 Oct 2015 00:05:17 +0000 (00:05 +0000)]
Remove the correct unstable malformed-machos test mem-crup-0261.macho and
restore the malformed-machos 00000031.a test.  Hopefully this will get all the
build bots happy again.  I’ll again keep an eye on them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250130 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDAGCombiner: Combine extract_vector_elt from build_vector
Matt Arsenault [Mon, 12 Oct 2015 23:59:50 +0000 (23:59 +0000)]
DAGCombiner: Combine extract_vector_elt from build_vector

This basic combine was surprisingly missing.
AMDGPU legalizes many operations in terms of 32-bit vector components,
so not doing this results in many extra copies and subregister extracts
that need to be cleaned up later.

InstCombine already does this for the hasOneUse case. The target hook
is to fix a handful of tests which break (e.g. ARM/vmov.ll) which turn
from a vector materialize repeated immediate instruction to a constant
vector load with more scalar copies from it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250129 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] Tidied up SSE4A tests.
Simon Pilgrim [Mon, 12 Oct 2015 23:07:06 +0000 (23:07 +0000)]
[InstCombine] Tidied up SSE4A tests.

First stage of bugfix discussed in D13348

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250121 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTemporarily remove the test added in r250117 while I investigate why two
Kevin Enderby [Mon, 12 Oct 2015 23:03:43 +0000 (23:03 +0000)]
Temporarily remove the test added in r250117 while I investigate why two
of the build bots get a different error on that malformed file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250120 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAssign correct edge weights to unwind destinations when lowering invoke statement.
Cong Hou [Mon, 12 Oct 2015 23:02:58 +0000 (23:02 +0000)]
Assign correct edge weights to unwind destinations when lowering invoke statement.

When lowering invoke statement, all unwind destinations are directly added as successors of call site block, and the weight of those new edges are not assigned properly. Actually, default weight 16 are used for those edges. This patch calculates the proper edge weights for those edges when collecting all unwind destinations.

Differential revision: http://reviews.llvm.org/D13354

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250119 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SelectionDAG] Add common vector constant folding helper function
Simon Pilgrim [Mon, 12 Oct 2015 23:00:11 +0000 (23:00 +0000)]
[SelectionDAG] Add common vector constant folding helper function

We have a number of functions that implement constant folding of vectors (unary and binary ops) in near identical manners (and the differences don't appear to be critical).

This patch introduces a common implementation (SelectionDAG::FoldConstantVectorArithmetic) and calls this in both the unary and binary op cases.

After this initial patch I intend to begin enabling vector constant folding for a wider number of opcodes in SelectionDAG::getNode().

Differential Revision: http://reviews.llvm.org/D13665

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250118 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFixed bugs in llvm-obdump while parsing Mach-O files from malformed archives
Kevin Enderby [Mon, 12 Oct 2015 22:04:54 +0000 (22:04 +0000)]
Fixed bugs in llvm-obdump while parsing Mach-O files from malformed archives
that caused aborts.  This was because of the characters of the ‘Size’ field in
the archive header did not contain decimal characters.

rdar://22983603

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250117 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake] Adding support for passing in profiling data.
Chris Bieneman [Mon, 12 Oct 2015 21:13:20 +0000 (21:13 +0000)]
[CMake] Adding support for passing in profiling data.

Adds LLVM_PROFDATA_FILE option to allow specifying a profile data file to be used during compilation of LLVM and subprojects.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250108 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUpdate the branch weight metadata in JumpThreading pass.
Cong Hou [Mon, 12 Oct 2015 19:44:08 +0000 (19:44 +0000)]
Update the branch weight metadata in JumpThreading pass.

In JumpThreading pass, the branch weight metadata is not updated after CFG modification. Consider the jump threading on PredBB, BB, and SuccBB. After jump threading, the weight on BB->SuccBB should be adjusted as some of it is contributed by the edge PredBB->BB, which doesn't exist anymore. This patch tries to update the edge weight in metadata on BB->SuccBB by scaling it by 1 - Freq(PredBB->BB) / Freq(BB->SuccBB).

Differential revision: http://reviews.llvm.org/D10979

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250089 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMake Win64 localescape offsets FP relative instead of SP relative
Reid Kleckner [Mon, 12 Oct 2015 19:43:34 +0000 (19:43 +0000)]
Make Win64 localescape offsets FP relative instead of SP relative

We made them SP relative back in March (r233137) because that's the
value the runtime passes to EH functions. With the new cleanuppad IR,
funclets adjust their frame argument from SP to FP, so our offsets
should now be FP-relative.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250088 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-symbolizer] Add -print-address option
Hemant Kulkarni [Mon, 12 Oct 2015 19:26:44 +0000 (19:26 +0000)]
[llvm-symbolizer] Add -print-address option

Differential Revision: http://reviews.llvm.org/D13518

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250086 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[x86] Fix wrong lowering of vsetcc nodes (PR25080).
Andrea Di Biagio [Mon, 12 Oct 2015 19:22:30 +0000 (19:22 +0000)]
[x86] Fix wrong lowering of vsetcc nodes (PR25080).

Function LowerVSETCC (in X86ISelLowering.cpp) worked under the wrong
assumption that for non-AVX512 targets, the source type and destination type
of a type-legalized setcc node were always the same type.

This assumption was unfortunately incorrect; the type legalizer is not always
able to promote the return type of a setcc to the same type as the first
operand of a setcc.

In the case of a vsetcc node, the legalizer firstly checks if the first input
operand has a legal type. If so, then it promotes the return type of the vsetcc
to that same type. Otherwise, the return type is promoted to the 'next legal
type', which, for vectors of MVT::i1 is always a 128-bit integer vector type.

Example (-mattr=+avx):

  %0 = trunc <8 x i32> %a to <8 x i23>
  %1 = icmp eq <8 x i23> %0, zeroinitializer

The initial selection dag for the code above is:

v8i1 = setcc t5, t7, seteq:ch
  t5: v8i23 = truncate t2
    t2: v8i32,ch = CopyFromReg t0, Register:v8i32 %vreg1
    t7: v8i32 = build_vector of all zeroes.

The type legalizer would firstly check if 't5' has a legal type. If so, then it
would reuse that same type to promote the return type of the setcc node.
Unfortunately 't5' is of illegal type v8i23, and therefore it cannot be used to
promote the return type of the setcc node. Consequently, the setcc return type
is promoted to v8i16. Later on, 't5' is promoted to v8i32 thus leading to the
following dag node:
  v8i16 = setcc t32, t25, seteq:ch

  where t32 and t25 are now values of type v8i32.

Before this patch, function LowerVSETCC would have wrongly expanded the setcc
to a single X86ISD::PCMPEQ. Surprisingly, ISel was still able to match an
instruction. In our case, ISel would have matched a VPCMPEQWrr:
  t37: v8i16 = X86ISD::VPCMPEQWrr t36, t25

However, t36 and t25 are both VR256, while the result type is instead of class
VR128. This inconsistency ended up causing the insertion of COPY instructions
like this:
  %vreg7<def> = COPY %vreg3; VR128:%vreg7 VR256:%vreg3

Which is an invalid full copy (not a sub register copy).
Eventually, the backend would have hit an UNREACHABLE "Cannot emit physreg copy
instruction" in the attempt to expand the malformed pseudo COPY instructions.

This patch fixes the problem adding the missing logic in LowerVSETCC to handle
the corner case of a setcc with 128-bit return type and 256-bit operand type.

This problem was originally reported by Dimitry as PR25080. It has been latent
for a very long time. I have added the minimal reproducible from that bugzilla
as test setcc-lowering.ll.

Differential Revision: http://reviews.llvm.org/D13660

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250085 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd - and -= operators to BlockFrequency using saturating arithmetic.
Cong Hou [Mon, 12 Oct 2015 18:34:00 +0000 (18:34 +0000)]
Add - and -= operators to BlockFrequency using saturating arithmetic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250077 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[libFuzzer] mention more trophies and improve the link formatting
Kostya Serebryany [Mon, 12 Oct 2015 18:15:42 +0000 (18:15 +0000)]
[libFuzzer] mention more trophies and improve the link formatting

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250076 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agocombine predicates; NFCI
Sanjay Patel [Mon, 12 Oct 2015 18:15:08 +0000 (18:15 +0000)]
combine predicates; NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250075 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTurn const/const& into value type for BlockFrequency in functions of this class....
Cong Hou [Mon, 12 Oct 2015 18:14:15 +0000 (18:14 +0000)]
Turn const/const& into value type for BlockFrequency in functions of this class. Also fix a naming issue. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250074 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-symbolizer] Reverting r250067
Colin LeMahieu [Mon, 12 Oct 2015 17:57:02 +0000 (17:57 +0000)]
[llvm-symbolizer] Reverting r250067

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250072 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Register some more passes so -print-before works
Matt Arsenault [Mon, 12 Oct 2015 17:43:59 +0000 (17:43 +0000)]
AMDGPU: Register some more passes so -print-before works

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250071 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoEnable verifier after PeepholeOptimizer
Matt Arsenault [Mon, 12 Oct 2015 17:43:56 +0000 (17:43 +0000)]
Enable verifier after PeepholeOptimizer

No tests fail with this enabled so I assume it was an accident
that it isn't enabled now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250070 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDon't call PrepareEHLandingPad on non EH pads
Reid Kleckner [Mon, 12 Oct 2015 17:42:32 +0000 (17:42 +0000)]
Don't call PrepareEHLandingPad on non EH pads

This was a minor bug in r249492. Calling PrepareEHLandingPad on a
non-landingpad was a no-op, but it attempted to get the generic pointer
register class, which apparently doesn't exist for some targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250068 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-symbolizer] Add -print-address option
Hemant Kulkarni [Mon, 12 Oct 2015 17:31:22 +0000 (17:31 +0000)]
[llvm-symbolizer] Add -print-address option

Differential Revision  http://reviews.llvm.org/D13518

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250067 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WinEH] Remove CatchObjRecoverIdx
David Majnemer [Mon, 12 Oct 2015 16:44:22 +0000 (16:44 +0000)]
[WinEH] Remove CatchObjRecoverIdx

CatchObjRecoverIdx was used for the old scheme, it is no longer
relevant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250065 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agofix typos; NFC
Sanjay Patel [Mon, 12 Oct 2015 16:09:59 +0000 (16:09 +0000)]
fix typos; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250059 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips][micromips] Initial support for micrmomips DSP instructions and addu.qb impleme...
Zoran Jovanovic [Mon, 12 Oct 2015 16:07:25 +0000 (16:07 +0000)]
[mips][micromips] Initial support for micrmomips DSP instructions and addu.qb implementation
Differential Revision: http://reviews.llvm.org/D12798

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250058 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Debug] Look through bitcasts to find argument registers
Oliver Stannard [Mon, 12 Oct 2015 15:52:36 +0000 (15:52 +0000)]
[Debug] Look through bitcasts to find argument registers

On targets where f32 is not legal, we have to look through a BITCAST SDNode to
find the register that an argument is stored in when emitting debug info, or we
will not be able to emit a DW_AT_location for it.

Differential Revision: http://reviews.llvm.org/D13005

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250056 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips][FastISel] Clang-format switch statement. NFC.
Vasileios Kalintiris [Mon, 12 Oct 2015 15:39:41 +0000 (15:39 +0000)]
[mips][FastISel] Clang-format switch statement. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250053 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AArch64]Fix bug in function names in test case
Jun Bum Lim [Mon, 12 Oct 2015 15:34:52 +0000 (15:34 +0000)]
[AArch64]Fix bug in function names in test case

Functions in this test case need to be renamed as its names are the same
as the instructions we are comparing with.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250052 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agofix capitalization; NFC
Sanjay Patel [Mon, 12 Oct 2015 15:24:01 +0000 (15:24 +0000)]
fix capitalization; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250049 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix rename() sometimes failing if another process uses openFileForRead()
Greg Bedwell [Mon, 12 Oct 2015 15:11:47 +0000 (15:11 +0000)]
Fix rename() sometimes failing if another process uses openFileForRead()

On Windows, fs::rename() could fail is another process was reading the
file at the same time using fs::openFileForRead().  In most cases the user
wouldn't notice as fs::rename() will continue to retry for 2000ms.  Typically
this is enough for the read to complete and a retry to succeed, but if the
disk is being it too hard then the response time might be longer than the
retry time and the rename would fail with a permission error.

Add FILE_SHARE_DELETE to the sharing flags for CreateFileW() in
fs::openFileForRead() and try ReplaceFileW() prior to MoveFileExW()
in fs::rename().

Based on an initial patch by Edd Dawson!

Differential Revision: http://reviews.llvm.org/D13647

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250046 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips][ias] Implement macro expansion when bcc has an immediate where a register...
Daniel Sanders [Mon, 12 Oct 2015 14:24:05 +0000 (14:24 +0000)]
[mips][ias] Implement macro expansion when bcc has an immediate where a register belongs.

Summary: Fixes PR24915.

Reviewers: vkalintiris

Subscribers: emaste, seanbruno, llvm-commits

Differential Revision: http://reviews.llvm.org/D13533

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250042 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips] Whitespace cleanup in MIPS16 tests to reduce noise in following changes. NFC.
Daniel Sanders [Mon, 12 Oct 2015 14:16:52 +0000 (14:16 +0000)]
[mips] Whitespace cleanup in MIPS16 tests to reduce noise in following changes. NFC.

Mostly tabs -> spaces and double spacing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250041 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips] Clean up most macro expansions to use the emit*() functions.
Daniel Sanders [Mon, 12 Oct 2015 14:09:12 +0000 (14:09 +0000)]
[mips] Clean up most macro expansions to use the emit*() functions.

Reviewers: vkalintiris

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D13591

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250040 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips] Handle undef when extracting subregs from FP64 registers.
Daniel Sanders [Mon, 12 Oct 2015 13:55:44 +0000 (13:55 +0000)]
[mips] Handle undef when extracting subregs from FP64 registers.

Summary:
This removes unnecessary instructions when extracting from an undefined register
and also fixes a crash for O32 when passing undef to a double argument in
held in integer registers.

Reviewers: vkalintiris

Subscribers: llvm-commits, zoran.jovanovic, petarj

Differential Revision: http://reviews.llvm.org/D13467

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250039 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoGlobalOpt does not treat externally_initialized globals correctly
Oliver Stannard [Mon, 12 Oct 2015 13:20:52 +0000 (13:20 +0000)]
GlobalOpt does not treat externally_initialized globals correctly

GlobalOpt currently merges stores into the initialisers of internal,
externally_initialized globals, but should not do so as the value of the global
may change between the initialiser and any code in the module being run.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250035 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Mark Swift MISched model as incomplete
James Molloy [Mon, 12 Oct 2015 12:49:59 +0000 (12:49 +0000)]
[ARM] Mark Swift MISched model as incomplete

The Swift Machine Scheduler Model is incomplete. There are instructions
missing which can trigger the "incomplete machine model" abort. This was
observed when a downstream SchedMachineModel was added to the ARM
target.

Patch by Christof Douma!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250033 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LoopVectorize] Shrink integer operations into the smallest type possible
James Molloy [Mon, 12 Oct 2015 12:34:45 +0000 (12:34 +0000)]
[LoopVectorize] Shrink integer operations into the smallest type possible

C semantics force sub-int-sized values (e.g. i8, i16) to be promoted to int
type (e.g. i32) whenever arithmetic is performed on them.

For targets with native i8 or i16 operations, usually InstCombine can shrink
the arithmetic type down again. However InstCombine refuses to create illegal
types, so for targets without i8 or i16 registers, the lengthening and
shrinking remains.

Most SIMD ISAs (e.g. NEON) however support vectors of i8 or i16 even when
their scalar equivalents do not, so during vectorization it is important to
remove these lengthens and truncates when deciding the profitability of
vectorization.

The algorithm this uses starts at truncs and icmps, trawling their use-def
chains until they terminate or instructions outside the loop are found (or
unsafe instructions like inttoptr casts are found). If the use-def chains
starting from different root instructions (truncs/icmps) meet, they are
unioned. The demanded bits of each node in the graph are ORed together to form
an overall mask of the demanded bits in the entire graph. The minimum bitwidth
that graph can be truncated to is the bitwidth minus the number of leading
zeroes in the overall mask.

The intention is that this algorithm should "first do no harm", so it will
never insert extra cast instructions. This is why the use-def graphs are
unioned, so that subgraphs with different minimum bitwidths do not need casts
inserted between them.

This algorithm works hard to reduce compile time impact. DemandedBits are only
queried if there are extends of illegal types and if a truncate to an illegal
type is seen. In the general case, this results in a simple linear scan of the
instructions in the loop.

No non-noise compile time impact was seen on a clang bootstrap build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250032 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add XSAVE intrinsic family
Amjad Aboud [Mon, 12 Oct 2015 11:47:46 +0000 (11:47 +0000)]
[X86] Add XSAVE intrinsic family

Add intrinsics for the
  XSAVE instructions (XSAVE/XSAVE64/XRSTOR/XRSTOR64)
  XSAVEOPT instructions (XSAVEOPT/XSAVEOPT64)
  XSAVEC instructions (XSAVEC/XSAVEC64)
  XSAVES instructions (XSAVES/XSAVES64/XRSTORS/XRSTORS64)

Differential Revision: http://reviews.llvm.org/D13012

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250029 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[x86] PR24562: fix incorrect folding of PSHUFB nodes with a mask where all indices...
Andrea Di Biagio [Mon, 12 Oct 2015 11:25:41 +0000 (11:25 +0000)]
[x86] PR24562: fix incorrect folding of PSHUFB nodes with a mask where all indices have the most significant bit set.

This patch fixes a problem in function 'combineX86ShuffleChain' that causes a
chain of shuffles to be wrongly folded away when the combined shuffle mask has
only one element.

We may end up with a combined shuffle mask of one element as a result of
multiple calls to function 'canWidenShuffleElements()'.
Function canWidenShuffleElements attempts to simplify a shuffle mask by widening
the size of the elements being shuffled.
For every pair of shuffle indices, function canWidenShuffleElements checks if
indices refer to adjacent elements. If all pairs refer to "adjacent" elements
then the shuffle mask is safely widened. As a consequence of widening, we end up
with a new shuffle mask which is half the size of the original shuffle mask.

The byte shuffle (pshufb) from test pr24562.ll has a mask of all SM_SentinelZero
indices. Function canWidenShuffleElements would combine each pair of
SM_SentinelZero indices into a single SM_SentinelZero index. So, in a
logarithmic number of steps (4 in this case), the pshufb mask is simplified to
a mask with only one index which is equal to SM_SentinelZero.

Before this patch, function combineX86ShuffleChain wrongly assumed that a mask
of size one is always equivalent to an identity mask. So, the entire shuffle
chain was just folded away as the combined shuffle mask was treated as a no-op
mask.

With this patch we know check if the only element of a combined shuffle mask is
SM_SentinelZero. In case, we propagate a zero vector.

Differential Revision: http://reviews.llvm.org/D13364

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250027 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTest commit
Zlatko Buljan [Mon, 12 Oct 2015 11:19:40 +0000 (11:19 +0000)]
Test commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250026 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agocmake: Avoid leading space in LLVM_DEFINITIONS.
Pawel Bylica [Mon, 12 Oct 2015 10:42:03 +0000 (10:42 +0000)]
cmake: Avoid leading space in LLVM_DEFINITIONS.

Summary: Unnecessary space at the beginning of LLVM_DEFINITIONS in cmake shared files can break projects that use the variable.

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D13432

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250025 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ] testcase MC/SystemZ/insn-good-z13.s extended.
Jonas Paulsson [Mon, 12 Oct 2015 10:13:57 +0000 (10:13 +0000)]
[SystemZ] testcase MC/SystemZ/insn-good-z13.s extended.

New instructions using floating point registers have been added, to check
that AsmParser can deal with fp regs in vector instructions.

This tests r249810.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250023 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MISched] Python script to check coverage of misched info
James Molloy [Mon, 12 Oct 2015 08:50:47 +0000 (08:50 +0000)]
[MISched] Python script to check coverage of misched info

This script prints a CSV of all misched models of a target when given the output of the debug output of subtarget using:

    llvm-tblgen --gen-subtarget --debug-only=subtarget-emitter ...

With thanks to Dave Estes for mentioning the idea at the 2014 LLVM Developers' Meeting.

Patch by Christof Douma!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250020 91177308-0d34-0410-b5e6-96231b3b80d8