OSDN Git Service

android-x86/external-llvm.git
6 years agoRevert r319531 "[SLPVectorizer] Failure to beneficially vectorize 'copyable' elements...
Hans Wennborg [Fri, 1 Dec 2017 16:17:24 +0000 (16:17 +0000)]
Revert r319531 "[SLPVectorizer] Failure to beneficially vectorize 'copyable' elements in integer binary ops."

It causes builds to fail with "Instruction does not dominate all uses" (PR35497).

> Patch tries to improve vectorization of the following code:
>
> void add1(int * __restrict dst, const int * __restrict src) {
>   *dst++ = *src++;
>   *dst++ = *src++ + 1;
>   *dst++ = *src++ + 2;
>   *dst++ = *src++ + 3;
> }
> Allows to vectorize even if the very first operation is not a binary add, but just a load.
>
> Fixed issues related to previous commit.
>
> Reviewers: spatel, mzolotukhin, mkuper, hfinkel, RKSimon, filcab, ABataev
>
> Reviewed By: ABataev, RKSimon
>
> Subscribers: llvm-commits, RKSimon
>
> Differential Revision: https://reviews.llvm.org/D28907

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319550 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] and + load combine tests
Sam Parker [Fri, 1 Dec 2017 15:31:41 +0000 (15:31 +0000)]
[ARM] and + load combine tests

Add a few more tests cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319548 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM][DAG] Reenable post-legalize store merge
Nirav Dave [Fri, 1 Dec 2017 14:49:26 +0000 (14:49 +0000)]
[ARM][DAG] Reenable post-legalize store merge

Summary: Reenable post-legalize stores with constant merging computation and cofrresponding test case.

Reviewers: eastig, efriedma

Subscribers: aemerson, javed.absar, kristof.beyls, hiraditya, llvm-commits

Differential Revision: https://reviews.llvm.org/D40701

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319547 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd more triples to llc_test_checks.py
Sam Parker [Fri, 1 Dec 2017 14:27:11 +0000 (14:27 +0000)]
Add more triples to llc_test_checks.py

Added some commonly used Arm triples to the script, with and without
the -eabi suffix.

Differential Revision: https://reviews.llvm.org/D40708

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319545 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Improvement in CodeGen instruction selection for LEAs.
Jatin Bhateja [Fri, 1 Dec 2017 14:07:38 +0000 (14:07 +0000)]
[X86] Improvement in CodeGen instruction selection for LEAs.

Summary:
1/  Operand folding during complex pattern matching for LEAs has been extended, such that it promotes Scale to
     accommodate similar operand appearing in the DAG  e.g.
                 T1 = A + B
                 T2 = T1 + 10
                 T3 = T2 + A
    For above DAG rooted at T3, X86AddressMode will now look like
                Base = B , Index = A , Scale = 2 , Disp = 10

2/  During OptimizeLEAPass down the pipeline factorization is now performed over LEAs so that if there is an opportunity
     then complex LEAs (having 3 operands) could be factored out  e.g.
                 leal 1(%rax,%rcx,1), %rdx
                 leal 1(%rax,%rcx,2), %rcx
     will be factored as following
                 leal 1(%rax,%rcx,1), %rdx
                 leal (%rdx,%rcx)   , %edx

3/ Aggressive operand folding for AM based selection for LEAs is sensitive to loops, thus avoiding creation of any complex LEAs within a loop.

4/ Simplify LEA converts (lea (BASE,1,INDEX,0)  --> add (BASE, INDEX) which offers better through put.

PR32755 will be taken care of by this pathc.

Previous patch revisions : r313343 , r314886

Reviewers: lsaba, RKSimon, craig.topper, qcolombet, jmolloy, jbhateja

Reviewed By: lsaba, RKSimon, jbhateja

Subscribers: jmolloy, spatel, igorb, llvm-commits

Differential Revision: https://reviews.llvm.org/D35014

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319543 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] and + load combine tests
Sam Parker [Fri, 1 Dec 2017 13:42:39 +0000 (13:42 +0000)]
[ARM] and + load combine tests

Adding autogenerated tests for narrow load combines.

Differential Revision: https://reviews.llvm.org/D40709

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319542 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][AVX512] Tag vshift/vpermv/pshufd/pshufb instructions scheduler classes
Simon Pilgrim [Fri, 1 Dec 2017 13:25:54 +0000 (13:25 +0000)]
[X86][AVX512] Tag vshift/vpermv/pshufd/pshufb instructions scheduler classes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319540 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert r319537: Bail out of a SimplifyCFG switch table opt at undef values.
Mikael Holmen [Fri, 1 Dec 2017 13:11:39 +0000 (13:11 +0000)]
Revert r319537: Bail out of a SimplifyCFG switch table opt at undef values.

Broke build bots so reverting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319539 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstSimplify] More fcmp cases when comparing against negative constants.
Florian Hahn [Fri, 1 Dec 2017 12:34:16 +0000 (12:34 +0000)]
[InstSimplify] More fcmp cases when comparing against negative constants.

Summary:
For known positive non-zero value X:
    fcmp uge X, -C => true
    fcmp ugt X, -C => true
    fcmp une X, -C => true
    fcmp oeq X, -C => false
    fcmp ole X, -C => false
    fcmp olt X, -C => false

Patch by Paul Walker.

Reviewers: majnemer, t.p.northover, spatel, RKSimon

Reviewed By: spatel

Subscribers: fhahn, llvm-commits

Differential Revision: https://reviews.llvm.org/D40012

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319538 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoBail out of a SimplifyCFG switch table opt at undef values.
Mikael Holmen [Fri, 1 Dec 2017 12:30:49 +0000 (12:30 +0000)]
Bail out of a SimplifyCFG switch table opt at undef values.

Summary:
A true or false result is expected from a comparison, but it seems the possibility of undef was overlooked, which could lead to a failed assert. This is fixed by this patch by bailing out if we encounter undef.

The bug is old and the assert has been there since the end of 2014, so it seems this is unusual enough to forego optimization.

Patch by: JesperAntonsson

Reviewers: spatel, eeckstein, hans

Reviewed By: hans

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D40639

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319537 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFollow-up to r319434 to turn the pass on by default
Nemanja Ivanovic [Fri, 1 Dec 2017 12:02:59 +0000 (12:02 +0000)]
Follow-up to r319434 to turn the pass on by default

Now that the patch has gone through the buildbot cycle,
turn it on by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319535 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AMDGPU] SiFixSGPRCopies should not modify non-divergent PHI
Alexander Timofeev [Fri, 1 Dec 2017 11:56:34 +0000 (11:56 +0000)]
[AMDGPU] SiFixSGPRCopies should not modify non-divergent PHI

Differential revision: https://reviews.llvm.org/D40556

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319534 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[cmake] Enable zlib support on windows
Pavel Labath [Fri, 1 Dec 2017 11:41:07 +0000 (11:41 +0000)]
[cmake] Enable zlib support on windows

Summary:
zlib support was hard-wired to off for (non-cygwin) windows targets.
This disables some features, such as reading debug info from compressed
dwarf sections.

This has been this way since zlib support was added in 2013 (r180083),
but there is no obvious reason for that. Zlib is perfectly capable of
being compiled for windows (it even has a cmake file that works out of
the box).

This enables one to turn on zlib support on windows, if one has zlib
avaliable.

Reviewers: rnk, beanz

Subscribers: mgorny, aprantl, llvm-commits

Differential Revision: https://reviews.llvm.org/D40655

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319533 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SLPVectorizer] Failure to beneficially vectorize 'copyable' elements in integer...
Dinar Temirbulatov [Fri, 1 Dec 2017 11:10:47 +0000 (11:10 +0000)]
[SLPVectorizer] Failure to beneficially vectorize 'copyable' elements in integer binary ops.

            Patch tries to improve vectorization of the following code:

            void add1(int * __restrict dst, const int * __restrict src) {
              *dst++ = *src++;
              *dst++ = *src++ + 1;
              *dst++ = *src++ + 2;
              *dst++ = *src++ + 3;
            }
            Allows to vectorize even if the very first operation is not a binary add, but just a load.

            Fixed issues related to previous commit.

            Reviewers: spatel, mzolotukhin, mkuper, hfinkel, RKSimon, filcab, ABataev

            Reviewed By: ABataev, RKSimon

            Subscribers: llvm-commits, RKSimon

            Differential Revision: https://reviews.llvm.org/D28907

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319531 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[lit] Don't enable LSan on Darwin for Apple clang 9.0.0
Jonas Devlieghere [Fri, 1 Dec 2017 10:49:47 +0000 (10:49 +0000)]
[lit] Don't enable LSan on Darwin for Apple clang 9.0.0

The latest clang that ships with Xcode (clang 900 or 9.0.0) does not
support LSan. This fixes the lit configuration to reflect that.

Differential revision: https://reviews.llvm.org/D40672

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319530 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[lit] Implement non-pipelined ‘mkdir’, ‘diff’ and ‘rm’ commands internally
Ying Yi [Fri, 1 Dec 2017 09:54:27 +0000 (09:54 +0000)]
[lit] Implement non-pipelined ‘mkdir’, ‘diff’ and ‘rm’ commands internally

Summary:
The internal shell already supports 'cd', ‘export’ and ‘echo’ commands.
This patch adds implementation of non-pipelined ‘mkdir’, ‘diff’ and ‘rm’
commands as the internal shell builtins.

Reviewed by: Zachary Turner, Reid Kleckner

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D39567

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319528 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoGlobalISel: Enable the legalization of G_MERGE_VALUES and G_UNMERGE_VALUES
Volkan Keles [Fri, 1 Dec 2017 08:19:10 +0000 (08:19 +0000)]
GlobalISel: Enable the legalization of G_MERGE_VALUES and G_UNMERGE_VALUES

Summary: LegalizerInfo assumes all G_MERGE_VALUES and G_UNMERGE_VALUES instructions are legal, so it is not possible to legalize vector operations on illegal vector types. This patch fixes the problem by removing the related check and adding default actions for G_MERGE_VALUES and G_UNMERGE_VALUES.

Reviewers: qcolombet, ab, dsanders, aditya_nandakumar, t.p.northover, kristof.beyls

Reviewed By: dsanders

Subscribers: rovka, javed.absar, igorb, llvm-commits

Differential Revision: https://reviews.llvm.org/D39823

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319524 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRecommit rL319407: [SROA] enable splitting for non-whole-alloca loads and stores
Hiroshi Inoue [Fri, 1 Dec 2017 06:05:05 +0000 (06:05 +0000)]
Recommit rL319407: [SROA] enable splitting for non-whole-alloca loads and stores

Recommiting once reverted patch rL319407 after adding a check for bit vector size to avoid failures in some build bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319522 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Custom legalize v2i32 gathers via widening rather than promoting.
Craig Topper [Fri, 1 Dec 2017 06:02:02 +0000 (06:02 +0000)]
[X86] Custom legalize v2i32 gathers via widening rather than promoting.

The default legalization for v2i32 is promotion to v2i64. This results in a gather that reads 64-bit elements rather than 32. If one of the elements is near a page boundary this can cause an illegal access that can fault.

We also miscalculate the scale for the gather which is an even worse problem, but we probably could have found a separate way to fix that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319521 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SelectionDAG] Make sure we explicitly sign extend the index when type promoting...
Craig Topper [Fri, 1 Dec 2017 06:02:00 +0000 (06:02 +0000)]
[X86][SelectionDAG] Make sure we explicitly sign extend the index when type promoting the index of scatter and gather.

Type promotion makes no guarantee about the contents of the promoted bits. Since the gather/scatter instruction will use the bits to calculate addresses, we need to ensure they aren't garbage.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319520 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add another v2i32 gather test case with v2i64 index that wasn't sign extended.
Craig Topper [Fri, 1 Dec 2017 06:01:59 +0000 (06:01 +0000)]
[X86] Add another v2i32 gather test case with v2i64 index that wasn't sign extended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319519 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Add a DAG combine to simplify masks for AVX2 gather instructions.
Craig Topper [Fri, 1 Dec 2017 02:49:07 +0000 (02:49 +0000)]
[X86] Add a DAG combine to simplify masks for AVX2 gather instructions.

AVX2 gathers only use the upper bit of the mask allowing us to simplify sign_extend_inreg to a shift left.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319514 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[cmake] Expose opt-viewer availability
Adam Nemet [Fri, 1 Dec 2017 01:44:26 +0000 (01:44 +0000)]
[cmake] Expose opt-viewer availability

This will be used in https://github.com/apple/swift/pull/12938

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319511 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Update MC tests now that hidden attr is supported
Sam Clegg [Fri, 1 Dec 2017 01:18:47 +0000 (01:18 +0000)]
[WebAssembly] Update MC tests now that hidden attr is supported

Summary:
Support was added in rL319488 but these tests were not
updated.

Subscribers: jfb, dschuff, jgravelle-google, aheejin, sunfish

Differential Revision: https://reviews.llvm.org/D40693

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319510 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd flag to ArchiveWriter to test GNU64 format more efficiently
Jake Ehrlich [Fri, 1 Dec 2017 00:54:28 +0000 (00:54 +0000)]
Add flag to ArchiveWriter to test GNU64 format more efficiently

Even with the sparse file optimizations the SYM64 test can still be painfully
slow. This unnecessarily slows down devs. It's critical that we test that the
switch to the SYM64 format occurs at 4GB but there isn't any better of a way to
fake the size of the file than sparse files. This change introduces a flag that
allows the cutoff to be arbitrarily set to whatever power of two is desired.
The flag is hidden as it really isn't meant to be used outside this one test.
This is unfortunate but appears necessary, at least until the average hard
drive is much faster.

The changes to the test require some explanation. Prior to this change we knew
that the SYM64 format was being used because the file was simply too large to
have validly handled this case if the SYM64 format were not used. To ensure
that the SYM64 format is still being used I am grepping the file for "SYM64".
Without changing the filename however this would be pointless because "SYM64"
would occur in the file either way. So the filename of the test is also changed
in order to avoid this issue.

Differential Revision: https://reviews.llvm.org/D40632

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319507 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMark all library options as hidden.
Zachary Turner [Fri, 1 Dec 2017 00:53:10 +0000 (00:53 +0000)]
Mark all library options as hidden.

These command line options are not intended for public use, and often
don't even make sense in the context of a particular tool anyway. About
90% of them are already hidden, but when people add new options they
forget to hide them, so if you were to make a brand new tool today, link
against one of LLVM's libraries, and run tool -help you would get a
bunch of junk that doesn't make sense for the tool you're writing.

This patch hides these options. The real solution is to not have
libraries defining command line options, but that's a much larger effort
and not something I'm prepared to take on.

Differential Revision: https://reviews.llvm.org/D40674

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319505 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agodocs/GettingStarted.rst: Update the list of release versions and tags
Hans Wennborg [Thu, 30 Nov 2017 23:47:30 +0000 (23:47 +0000)]
docs/GettingStarted.rst: Update the list of release versions and tags

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319502 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Use carry-less adds in FI elimination
Matt Arsenault [Thu, 30 Nov 2017 23:42:30 +0000 (23:42 +0000)]
AMDGPU: Use carry-less adds in FI elimination

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319501 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoThinLTOBitcodeWriter: Try harder to discard unused references to the merged module.
Peter Collingbourne [Thu, 30 Nov 2017 23:05:52 +0000 (23:05 +0000)]
ThinLTOBitcodeWriter: Try harder to discard unused references to the merged module.

If the thin module has no references to an internal global in the
merged module, we need to make sure to preserve that property if the
global is a member of a comdat group, as otherwise promotion can end
up adding global symbols to the comdat, which is not allowed.

This situation can arise if the external global in the thin module
has dead constant users, which would cause use_empty() to return
false and would cause us to try to promote it. To prevent this from
happening, discard the dead constant users before asking whether a
global is empty.

Differential Revision: https://reviews.llvm.org/D40593

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319494 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSimplify the DenseSet used for hashing CodeView records.
Zachary Turner [Thu, 30 Nov 2017 23:00:30 +0000 (23:00 +0000)]
Simplify the DenseSet used for hashing CodeView records.

This was storing the hash alongside the key so that the hash
doesn't need to be re-computed every time, but in doing so it
was allocating a structure to keep the key size small in the
DenseMap.  This is a noble goal, but it also leads to a pointer
indirection on every probe, and this cost of this pointer
indirection ends up being higher than the cost of having a
slightly larger entry in the hash table.  Removing this not only
simplifies the code, but yields a small but noticeable
performance improvement in the type merging algorithm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319493 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Use gfx9 carry-less add/sub instructions
Matt Arsenault [Thu, 30 Nov 2017 22:51:26 +0000 (22:51 +0000)]
AMDGPU: Use gfx9 carry-less add/sub instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319491 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoXOR the frame pointer with the stack cookie when protecting the stack
Reid Kleckner [Thu, 30 Nov 2017 22:41:21 +0000 (22:41 +0000)]
XOR the frame pointer with the stack cookie when protecting the stack

Summary: This strengthens the guard and matches MSVC.

Reviewers: hans, etienneb

Subscribers: hiraditya, JDevlieghere, vlad.tsyrklevich, llvm-commits

Differential Revision: https://reviews.llvm.org/D40622

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319490 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd visibility flag to Wasm symbol flags
Sam Clegg [Thu, 30 Nov 2017 22:34:58 +0000 (22:34 +0000)]
Add visibility flag to Wasm symbol flags

The LLVM "hidden" flag needs to be passed through the Wasm
intermediate objects in order for the linker to apply
it to the final Wasm object.

The corresponding change in LLD is here: https://github.com/WebAssembly/lld/pull/14

Patch by Nicholas Wilson

Differential Revision: https://reviews.llvm.org/D40442

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319488 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[memcpyopt] Commit file missed in r319482.
Dan Gohman [Thu, 30 Nov 2017 22:13:13 +0000 (22:13 +0000)]
[memcpyopt] Commit file missed in r319482.

This change was meant to be included with r319482 but was accidentally
omitted.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319483 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[memcpyopt] Teach memcpyopt to optimize across basic blocks
Dan Gohman [Thu, 30 Nov 2017 22:10:53 +0000 (22:10 +0000)]
[memcpyopt] Teach memcpyopt to optimize across basic blocks

This teaches memcpyopt to make a non-local memdep query when a local query
indicates that the dependency is non-local. This notably allows it to
eliminate many more llvm.memcpy calls in common Rust code, often by 20-30%.

Fixes PR28958.

Differential Revision: https://reviews.llvm.org/D38374

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319482 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InlineCost] Prefer getFunction() to two calls to getParent().
Davide Italiano [Thu, 30 Nov 2017 22:10:35 +0000 (22:10 +0000)]
[InlineCost] Prefer getFunction() to two calls to getParent().

Improves clarity, also slightly cheaper. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319481 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm] Add stripped installation targets
Shoaib Meenai [Thu, 30 Nov 2017 21:48:26 +0000 (21:48 +0000)]
[llvm] Add stripped installation targets

CMake's generated installation scripts support `CMAKE_INSTALL_DO_STRIP`
to enable stripping the installed binaries. LLVM's build system doesn't
expose this option to the `install-` targets, but it's useful in
conjunction with `install-distribution`.

Add a new function to create the install targets, which creates both the
regular install target and a second install target that strips during
installation. Change the creation of all installation targets to use
this new function. Stripping doesn't make a whole lot of sense for some
installation targets (e.g. the LLVM headers), but consistency doesn't
hurt.

I'll make other repositories (e.g. clang, compiler-rt) use this in a
follow-up, and then add an `install-distribution-stripped` target to
actually accomplish the end goal of creating a stripped distribution. I
don't want to do that step yet because the creation of that target would
depend on the presence of the `install-*-stripped` target for each
distribution component, and the distribution components from other
repositories will be missing that target right now.

Differential Revision: https://reviews.llvm.org/D40620

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319480 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Hexagon] Implement HexagonSubtarget::useAA()
Krzysztof Parzyszek [Thu, 30 Nov 2017 21:25:28 +0000 (21:25 +0000)]
[Hexagon] Implement HexagonSubtarget::useAA()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319477 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Hexagon] Fix wrong check in test/CodeGen/Hexagon/newvaluejump-solo.mir
Krzysztof Parzyszek [Thu, 30 Nov 2017 21:23:19 +0000 (21:23 +0000)]
[Hexagon] Fix wrong check in test/CodeGen/Hexagon/newvaluejump-solo.mir

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319476 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[globalisel][tablegen] Add support for relative AtomicOrderings
Daniel Sanders [Thu, 30 Nov 2017 21:05:59 +0000 (21:05 +0000)]
[globalisel][tablegen] Add support for relative AtomicOrderings

No test yet because the relevant rules are blocked on the atomic_load,
and atomic_store nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319475 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Hexagon] Fix wrong pass in testcase
Krzysztof Parzyszek [Thu, 30 Nov 2017 20:39:15 +0000 (20:39 +0000)]
[Hexagon] Fix wrong pass in testcase

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319471 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Hexagon] Solo instructions cannot be used with new value jumps
Krzysztof Parzyszek [Thu, 30 Nov 2017 20:32:54 +0000 (20:32 +0000)]
[Hexagon] Solo instructions cannot be used with new value jumps

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319470 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AMDGPU] Convert test/tools/llvm-objdump/AMDGPU/source-lines.ll to amdgiz
Yaxun Liu [Thu, 30 Nov 2017 20:27:56 +0000 (20:27 +0000)]
[AMDGPU] Convert test/tools/llvm-objdump/AMDGPU/source-lines.ll to amdgiz

Differential Revision: https://reviews.llvm.org/D40653

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319469 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Promote i8 CTPOP to i32 instead of i16 when we have the POPCNT instruction.
Craig Topper [Thu, 30 Nov 2017 20:15:31 +0000 (20:15 +0000)]
[X86] Promote i8 CTPOP to i32 instead of i16 when we have the POPCNT instruction.

The 32-bit version is shorter to encode and the zext we emit for the promotion is likely going to be a 32-bit zero extend anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319468 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-objcopy] Add support for --only-keep/-j and --keep
Jake Ehrlich [Thu, 30 Nov 2017 20:14:53 +0000 (20:14 +0000)]
[llvm-objcopy] Add support for --only-keep/-j and --keep

This change adds support for the --only-keep option and the -j alias as well.
A common use case for these being used together is to dump a specific section's
data. Additionally the --keep option is added (GNU objcopy doesn't have this)
to avoid removing a bunch of things. This allows people to err on the side of
stripping aggressively and then to keep the specific bits that they need for
their application.

Differential Revision: https://reviews.llvm.org/D39021

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319467 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[aarch64][globalisel] Legalize G_ATOMIC_CMPXCHG_WITH_SUCCESS and G_ATOMICRMW_*
Daniel Sanders [Thu, 30 Nov 2017 20:11:42 +0000 (20:11 +0000)]
[aarch64][globalisel] Legalize G_ATOMIC_CMPXCHG_WITH_SUCCESS and G_ATOMICRMW_*

G_ATOMICRMW_* is generally legal on AArch64. The exception is G_ATOMICRMW_NAND.

G_ATOMIC_CMPXCHG_WITH_SUCCESS needs to be lowered to G_ATOMIC_CMPXCHG with an
external comparison.

Note that IRTranslator doesn't generate these instructions yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319466 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[GlobalISel][IRTranslator] Fix crash during translation of zero sized loads/stores...
Amara Emerson [Thu, 30 Nov 2017 20:06:02 +0000 (20:06 +0000)]
[GlobalISel][IRTranslator] Fix crash during translation of zero sized loads/stores/args/returns.

This fixes PR35358.

rdar://35619533

Differential Revision: https://reviews.llvm.org/D40604

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319465 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PGO] Skip counter promotion for infinite loops
Xinliang David Li [Thu, 30 Nov 2017 19:16:25 +0000 (19:16 +0000)]
[PGO] Skip counter promotion for infinite loops

Differential Revision: http://reviews.llvm.org/D40662

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319462 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[cmake] Include project name in Sphinx doctree dir to fix race conditions
Michal Gorny [Thu, 30 Nov 2017 19:09:22 +0000 (19:09 +0000)]
[cmake] Include project name in Sphinx doctree dir to fix race conditions

Modify add_sphinx_target() to include the project name alongside builder
in Sphinx doctree directory. This aims to avoid crashes due to race
conditions between multiple Sphinx instances running in parallel that
attempt to create or read that directory simultaneously.

This problem has originally been addressed in r283188. However, that
commit presumed that there will be only one target per builder being
run. However, r314863 introduced a second manpage target, reintroducing
the race condition.

Differential Revision: https://reviews.llvm.org/D40656

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319461 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[globalisel][tablegen] Add support for specific immediates in the match pattern
Daniel Sanders [Thu, 30 Nov 2017 18:48:35 +0000 (18:48 +0000)]
[globalisel][tablegen] Add support for specific immediates in the match pattern

This enables a few rules such as ARM's uxtb instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319457 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSplit TypeTableBuilder into two classes.
Zachary Turner [Thu, 30 Nov 2017 18:39:50 +0000 (18:39 +0000)]
Split TypeTableBuilder into two classes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319456 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-readobj] Fix mismatched line endings
Zachary Turner [Thu, 30 Nov 2017 18:33:34 +0000 (18:33 +0000)]
[llvm-readobj] Fix mismatched line endings

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319453 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Revert r319186 "Support bitcasted function addresses with varargs."
Dan Gohman [Thu, 30 Nov 2017 18:16:49 +0000 (18:16 +0000)]
[WebAssembly] Revert r319186 "Support bitcasted function addresses with varargs."

The patch broke Emscripten's EM_ASM macros, which utiltize unprototyped
functions.

See https://bugs.llvm.org/show_bug.cgi?id=35385 for details.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319452 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MIR] Fix DebugInfo tests after r319445
Francis Visoiu Mistrih [Thu, 30 Nov 2017 16:48:53 +0000 (16:48 +0000)]
[MIR] Fix DebugInfo tests after r319445

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319447 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CodeGen] Always use `printReg` to print registers in both MIR and debug
Francis Visoiu Mistrih [Thu, 30 Nov 2017 16:12:24 +0000 (16:12 +0000)]
[CodeGen] Always use `printReg` to print registers in both MIR and debug
output

As part of the unification of the debug format and the MIR format,
always use `printReg` to print all kinds of registers.

Updated the tests using '_' instead of '%noreg' until we decide which
one we want to be the default one.

Differential Revision: https://reviews.llvm.org/D40421

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319445 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[FuzzMutate] Bailout from injecting into empty basic blocks.
Igor Laevsky [Thu, 30 Nov 2017 15:41:58 +0000 (15:41 +0000)]
[FuzzMutate] Bailout from injecting into empty basic blocks.
In rare cases we can receive request to inject into completelly empty basic block. In the normal case
all basic blocks contain at least terminator instruction, but it is possible that the only instruction is
catchpad instruction which is not part of the instruction iterator. This case seems rare enough to not care
about it.
Submiting without review, since it seems almost NFC. I couldn't come up with any reasonable way to test this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319444 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[FuzzMutate] Correctly handle vector types in the insertvalue operation
Igor Laevsky [Thu, 30 Nov 2017 15:31:13 +0000 (15:31 +0000)]
[FuzzMutate] Correctly handle vector types in the insertvalue operation

Differential Revision: https://reviews.llvm.org/D40397

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319442 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[FuzzMutate] Don't use index operands as sinks
Igor Laevsky [Thu, 30 Nov 2017 15:29:16 +0000 (15:29 +0000)]
[FuzzMutate] Don't use index operands as sinks

Differential Revision: https://reviews.llvm.org/D40396

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319441 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[FuzzMutate] Pick correct index for the insertvalue instruction
Igor Laevsky [Thu, 30 Nov 2017 15:26:48 +0000 (15:26 +0000)]
[FuzzMutate] Pick correct index for the insertvalue instruction

Differential Revision: https://reviews.llvm.org/D40395

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319440 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[FuzzMutate] Don't create load as a new source if it doesn't match with the descriptor
Igor Laevsky [Thu, 30 Nov 2017 15:24:41 +0000 (15:24 +0000)]
[FuzzMutate] Don't create load as a new source if it doesn't match with the descriptor

Differential Revision: https://reviews.llvm.org/D40394

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319439 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[FuzzMutate] Don't crash when we can't remove instruction from empty function
Igor Laevsky [Thu, 30 Nov 2017 15:07:38 +0000 (15:07 +0000)]
[FuzzMutate] Don't crash when we can't remove instruction from empty function

Differential Revision: https://reviews.llvm.org/D40393

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319438 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LangRef] clarify semantics of the frem instruction
Sanjay Patel [Thu, 30 Nov 2017 14:59:03 +0000 (14:59 +0000)]
[LangRef] clarify semantics of the frem instruction

As noted in D40594, the frem instruction corresponds to fmod() except that it can't set errno.
I modified the text that we currently use for intrinsics that map to libm functions and applied
it to frem.

Differential Revision: https://reviews.llvm.org/D40629

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319437 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] Additional test for PR35354, NFC.
Alexey Bataev [Thu, 30 Nov 2017 14:33:58 +0000 (14:33 +0000)]
[InstCombine] Additional test for PR35354, NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319436 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PowerPC] Recommit r314244 with refactoring and off by default
Nemanja Ivanovic [Thu, 30 Nov 2017 13:39:10 +0000 (13:39 +0000)]
[PowerPC] Recommit r314244 with refactoring and off by default

This re-commits everything that was pulled in r314244. The transformation
is off by default (patch to enable it to follow). The code is refactored
to have a single entry-point and provide fine-grained control over patterns
that it selects. This patch also fixes the bugs in the original code.

Everything that failed with the original patch has been re-tested with this
patch (with the transformation turned on). So the patch to turn this on is
soon to follow.

Differential Revision: https://reviews.llvm.org/D38575

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319434 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][AVX512] Tag fcmp/ptest/ternlog instructions scheduler classes
Simon Pilgrim [Thu, 30 Nov 2017 13:18:06 +0000 (13:18 +0000)]
[X86][AVX512] Tag fcmp/ptest/ternlog instructions scheduler classes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319433 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][AVX512] Regenerate avx512 schedule tests
Simon Pilgrim [Thu, 30 Nov 2017 13:09:21 +0000 (13:09 +0000)]
[X86][AVX512] Regenerate avx512 schedule tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319432 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MC] Function stack size section.
Sean Eveson [Thu, 30 Nov 2017 13:05:14 +0000 (13:05 +0000)]
[MC] Function stack size section.

Re applying after fixing issues in the diff, sorry for any painful conflicts/merges!

Original RFC: http://lists.llvm.org/pipermail/llvm-dev/2017-August/117028.html

This change adds a '.stack-size' section containing metadata on function stack sizes to output ELF files behind the new -stack-size-section flag. The section contains pairs of function symbol references (8 byte) and stack sizes (unsigned LEB128).

The contents of this section can be used to measure changes to stack sizes between different versions of the compiler or a source base. The advantage of having a section is that we can extract this information when examining binaries that we didn't build, and it allows users and tools easy access to that information just by referencing the binary.

There is a follow up change to add an option to clang.

Thanks.

Reviewers: hfinkel, MatzeB

Reviewed By: MatzeB

Subscribers: thegameg, asb, llvm-commits

Differential Revision: https://reviews.llvm.org/D39788

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319430 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert r319423: [MC] Function stack size section.
Sean Eveson [Thu, 30 Nov 2017 12:43:25 +0000 (12:43 +0000)]
Revert r319423: [MC] Function stack size section.

I messed up the diff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319429 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM GlobalISel] Bail out for byval
Diana Picus [Thu, 30 Nov 2017 12:23:44 +0000 (12:23 +0000)]
[ARM GlobalISel] Bail out for byval

Fallback if we have a byval parameter or argument since we don't support
them yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319428 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CodeGen] Print "%vreg0" as "%0" in both MIR and debug output
Francis Visoiu Mistrih [Thu, 30 Nov 2017 12:12:19 +0000 (12:12 +0000)]
[CodeGen] Print "%vreg0" as "%0" in both MIR and debug output

As part of the unification of the debug format and the MIR format, avoid
printing "vreg" for virtual registers (which is one of the current MIR
possibilities).

Basically:

* find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E "s/%vreg([0-9]+)/%\1/g"
* grep -nr '%vreg' . and fix if needed
* find . \( -name "*.mir" -o -name "*.cpp" -o -name "*.h" -o -name "*.ll" \) -type f -print0 | xargs -0 sed -i '' -E "s/ vreg([0-9]+)/ %\1/g"
* grep -nr 'vreg[0-9]\+' . and fix if needed

Differential Revision: https://reviews.llvm.org/D40420

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319427 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][AVX512] Tag binop/rounding/sae instructions scheduler classes
Simon Pilgrim [Thu, 30 Nov 2017 12:01:52 +0000 (12:01 +0000)]
[X86][AVX512] Tag binop/rounding/sae instructions scheduler classes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319424 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MC] Function stack size section.
Sean Eveson [Thu, 30 Nov 2017 12:01:16 +0000 (12:01 +0000)]
[MC] Function stack size section.

Summary:
Original RFC: http://lists.llvm.org/pipermail/llvm-dev/2017-August/117028.html

I wasn't sure who to put as reviewers, so please add/remove people as appropriate.

This change adds a '.stack-size' section containing metadata on function stack sizes to output ELF files behind the new -stack-size-section flag. The section contains pairs of function symbol references (8 byte) and stack sizes (unsigned LEB128).

The contents of this section can be used to measure changes to stack sizes between different versions of the compiler or a source base. The advantage of having a section is that we can extract this information when examining binaries that we didn't build, and it allows users and tools easy access to that information just by referencing the binary.

There is a follow up change to add an option to clang.

Thanks.

Reviewers: hfinkel, MatzeB

Reviewed By: MatzeB

Subscribers: thegameg, asb, llvm-commits

Differential Revision: https://reviews.llvm.org/D39788

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319423 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[DAGCombine] Refactor ReduceLoadWidth
Sam Parker [Thu, 30 Nov 2017 11:49:11 +0000 (11:49 +0000)]
[DAGCombine] Refactor ReduceLoadWidth

visitAND attempts to narrow the width of extending loads that are
then masked off. ReduceLoadWidth already exists for a similar purpose
and handles shifts, so I've moved the code to handle AND nodes there.

Differential Revision: https://reviews.llvm.org/D39595

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319421 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSupport generic lowering of vector bswap
Serge Guelton [Thu, 30 Nov 2017 11:06:22 +0000 (11:06 +0000)]
Support generic lowering of vector bswap

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319419 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][AVX512] Tag RCP/RSQRT/GETEXP instructions scheduler classes
Simon Pilgrim [Thu, 30 Nov 2017 10:48:47 +0000 (10:48 +0000)]
[X86][AVX512] Tag RCP/RSQRT/GETEXP instructions scheduler classes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319418 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[dsymutil] Exclude namespace from ifdef in CFBundle
Jonas Devlieghere [Thu, 30 Nov 2017 10:41:31 +0000 (10:41 +0000)]
[dsymutil] Exclude namespace from ifdef in CFBundle

Should fix build failure introduced by r319416 on non-darwin hosts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319417 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[dsymutil] Upstream getBundleInfo implementation
Jonas Devlieghere [Thu, 30 Nov 2017 10:25:28 +0000 (10:25 +0000)]
[dsymutil] Upstream getBundleInfo implementation

This patch implements `getBundleInfo`, which uses CoreFoundation to
obtain information about the CFBundle. This information is needed to
populate the Plist in the dSYM bundle.

This change only applies to darwin and is an NFC as far as other
platforms are concerned.

Differential revision: https://reviews.llvm.org/D40244

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319416 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert rL319407: [SROA] enable splitting for non-whole-alloca loads and stores
Hiroshi Inoue [Thu, 30 Nov 2017 08:29:51 +0000 (08:29 +0000)]
Revert rL319407: [SROA] enable splitting for non-whole-alloca loads and stores

This reverts commit rL319407 due to failures in some buildbot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319410 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ] Bugfix in adjustSubwordCmp.
Jonas Paulsson [Thu, 30 Nov 2017 08:18:50 +0000 (08:18 +0000)]
[SystemZ]  Bugfix in adjustSubwordCmp.

Csmith generated a program where a store after load to the same address did
not get chained after the new load created during DAG legalizing, and so
performed an illegal overwrite of the expected value.

When the new zero-extending load is created, the chain users of the original
load must be updated, which was not done previously.

A similar case was also found and handled in lowerBITCAST.

Review: Ulrich Weigand
https://reviews.llvm.org/D40542

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319409 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SROA] enable splitting for non-whole-alloca loads and stores
Hiroshi Inoue [Thu, 30 Nov 2017 07:44:46 +0000 (07:44 +0000)]
[SROA] enable splitting for non-whole-alloca loads and stores

Currently, SROA splits loads and stores only when they are accessing the whole alloca.
This patch relaxes this limitation to allow splitting a load/store if all other loads and stores to the alloca are disjoint to or fully included in the current load/store. If there is no other load or store that crosses the boundary of the current load/store, the current splitting implementation works as is.
The whole-alloca loads and stores meet this new condition and so they are still splittable.

Here is a simplified motivating example.

struct record {
    long long a;
    int b;
    int c;
};

int func(struct record r) {
    for (int i = 0; i < r.c; i++)
        r.b++;
    return r.b;
}

When updating r.b (or r.c as well), LLVM generates redundant instructions on some platforms (such as x86_64, ppc64); here, r.b and r.c are packed into one 64-bit GPR when the struct is passed as a method argument.

With this patch, the above example is compiled into only few instructions without loop.
Without the patch, unnecessary loop-carried dependency is introduced by SROA and the loop cannot be eliminated by the later optimizers.

Differential Revision: https://reviews.llvm.org/D32998

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319407 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Optimize avx2 vgatherqps for v2f32 with v2i64 index type.
Craig Topper [Thu, 30 Nov 2017 07:01:40 +0000 (07:01 +0000)]
[X86] Optimize avx2 vgatherqps for v2f32 with v2i64 index type.

Normal type legalization will widen everything. This requires forcing 0s into the mask register. We can instead choose the form that only reads 2 elements without zeroing the mask.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319406 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Make sure we don't remove sign extends of masks with AVX2 masked gathers.
Craig Topper [Thu, 30 Nov 2017 06:31:31 +0000 (06:31 +0000)]
[X86] Make sure we don't remove sign extends of masks with AVX2 masked gathers.

We don't use k-registers and instead use the MSB so we need to make sure we sign extend the mask to the msb.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319405 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[XRay][docs] Update documentation on new default for xray_naive_log=
Dean Michael Berris [Thu, 30 Nov 2017 05:35:51 +0000 (05:35 +0000)]
[XRay][docs] Update documentation on new default for xray_naive_log=

We've recently changed the default for `xray_naive_log=` to be `false`
instead of `true` to make it more consistent with the FDR mode logging
implementation. This means we will now ask users to explicitly choose
which version of the XRay logging is being used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319400 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago- Removed unused lamba (IsReturnBlock) causing build bots to fail for r319398
Graham Yiu [Thu, 30 Nov 2017 03:36:57 +0000 (03:36 +0000)]
- Removed unused lamba (IsReturnBlock) causing build bots to fail for r319398
- Added lit testcases that were supposed to be part of r319398

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319399 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoWith PGO information, we can do more aggressive outlining of cold regions in the...
Graham Yiu [Thu, 30 Nov 2017 02:41:36 +0000 (02:41 +0000)]
With PGO information, we can do more aggressive outlining of cold regions in the inline candidate function. This contrasts with the scheme of keeping only the 'early return' portion of the inline candidate and outlining the rest of the function as a single function call.

Support for outlining multiple regions of each function is added, as well as some basic heuristics to determine which regions are good to outline. Outline candidates limited to regions that are single-entry & single-exit. We also avoid outlining regions that produce live-exit variables, which may inhibit some forms of code motion (like commoning).

Fallback to the regular partial inlining scheme is retained when either i) no regions are identified for outlining in the function, or ii) the outlined function could not be inlined in any of its callers.

Differential Revision: https://reviews.llvm.org/D38190

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319398 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[libFuzzer] mention one more trophie in the Linux Kernel
Kostya Serebryany [Thu, 30 Nov 2017 02:26:47 +0000 (02:26 +0000)]
[libFuzzer] mention one more trophie in the Linux Kernel

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319397 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU: Allow negative MUBUF vaddr for gfx9
Matt Arsenault [Thu, 30 Nov 2017 00:52:40 +0000 (00:52 +0000)]
AMDGPU: Allow negative MUBUF vaddr for gfx9

GFX9 does not enable bounds checking for the resource descriptors
used for private access, so it should be OK to use vaddr with
a potentially negative value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319393 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoCheck alignment in getSectionContentsAsArray.
Rafael Espindola [Thu, 30 Nov 2017 00:44:22 +0000 (00:44 +0000)]
Check alignment in getSectionContentsAsArray.

While the ArrayRef can technically have unaligned data, it would be
extremely surprising if iterating over it caused undefined behavior
when a reference to the underlying type was bound.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319392 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Coverage] Use the most-recent completed region count (PR35437)
Vedant Kumar [Thu, 30 Nov 2017 00:28:23 +0000 (00:28 +0000)]
[Coverage] Use the most-recent completed region count (PR35437)

This is a fix for the coverage segment builder.

If multiple regions must be popped off the active stack at once, and
more than one of them end at the same location, emit a segment using the
count from the most-recent completed region.

Fixes PR35437, rdar://35760630

Testing: invoked llvm-cov on a stage2 build of clang, additional unit
tests, check-profile

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319391 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoLowerTypeTests: Deduplicate code. NFC.
Peter Collingbourne [Thu, 30 Nov 2017 00:27:08 +0000 (00:27 +0000)]
LowerTypeTests: Deduplicate code. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319390 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoLowerTypeTests: Remove unnecessary cast. NFC.
Peter Collingbourne [Thu, 30 Nov 2017 00:02:55 +0000 (00:02 +0000)]
LowerTypeTests: Remove unnecessary cast. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319387 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Remove some questionable looking code that seems to be looking through a VZEXT...
Craig Topper [Wed, 29 Nov 2017 23:08:25 +0000 (23:08 +0000)]
[X86] Remove some questionable looking code that seems to be looking through a VZEXT to create a larger VSEXT.

If the input the vzext was signed this would do the wrong thing.

Not sure how to test this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319382 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFirst step towards more human-friendly PPC assembler output:
Joerg Sonnenberger [Wed, 29 Nov 2017 23:05:56 +0000 (23:05 +0000)]
First step towards more human-friendly PPC assembler output:
- add -ppc-reg-with-percent-prefix option to use %r3 etc as register
  names
- split off logic for Darwinish verbose conditional codes into a helper
  function
- be explicit about Darwin vs AIX vs GNUish assembler flavors

Based on the patch from Alexandre Yukio Yamashita

Differential Revision: https://reviews.llvm.org/D39016

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319381 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Update test expectations for gcc torture tests
Sam Clegg [Wed, 29 Nov 2017 23:05:50 +0000 (23:05 +0000)]
[WebAssembly] Update test expectations for gcc torture tests

I believe these were recently fixed by:
https://reviews.llvm.org/rL319186

Differential Revision: https://reviews.llvm.org/D40619

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319380 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CodeView] Factor some code out of TypeTableBuilder.
Zachary Turner [Wed, 29 Nov 2017 22:41:56 +0000 (22:41 +0000)]
[CodeView] Factor some code out of TypeTableBuilder.

This class had some code that would automatically remap type
indices before hashing and serializing.  The only caller of
this method was the TypeStreamMerger anyway, and the method
doesn't make general sense, and prevents making certain future
improvements to the class.  So, factoring this up one level
into the TypeStreamMerger where it belongs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319377 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SelectionDAG][X86] Teach promotion legalization for fp_to_sint/fp_to_uint to insert...
Craig Topper [Wed, 29 Nov 2017 22:15:43 +0000 (22:15 +0000)]
[SelectionDAG][X86] Teach promotion legalization for fp_to_sint/fp_to_uint to insert an assertsext/assertzext based on the original type

If we put in an assertsext/zext here, we're able to generate better truncate code using pack on pre-avx512 targets.

Similar is already done during type legalization. This is the equivalent for op legalization

Differential Revision: https://reviews.llvm.org/D40591

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319368 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Fix fptoui lowering bounds
Dan Gohman [Wed, 29 Nov 2017 20:20:11 +0000 (20:20 +0000)]
[WebAssembly] Fix fptoui lowering bounds

To fully avoid trapping on wasm, fptoui needs a second check to ensure that
the operand isn't below the supported range.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319354 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd libstd++-4.8 exceptions to ubsan_blacklist.txt
Sam Clegg [Wed, 29 Nov 2017 20:10:14 +0000 (20:10 +0000)]
Add libstd++-4.8 exceptions to ubsan_blacklist.txt

Differential Revision: https://reviews.llvm.org/D40589

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319353 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Hexagon] Remove HexagonISD::PACKHL
Krzysztof Parzyszek [Wed, 29 Nov 2017 19:59:29 +0000 (19:59 +0000)]
[Hexagon] Remove HexagonISD::PACKHL

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319352 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Hexagon] Create helpers extractVector and insertVector in lowering
Krzysztof Parzyszek [Wed, 29 Nov 2017 19:58:10 +0000 (19:58 +0000)]
[Hexagon] Create helpers extractVector and insertVector in lowering

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@319351 91177308-0d34-0410-b5e6-96231b3b80d8