OSDN Git Service

android-x86/external-llvm.git
5 years ago[NVPTX] Lower instructions that expand into libcalls.
Artem Belevich [Fri, 14 Dec 2018 23:53:06 +0000 (23:53 +0000)]
[NVPTX] Lower instructions that expand into libcalls.

The change is an effort to split and refactor abandoned
D34708 into smaller parts.

Here the behaviour of unsupported instructions is changed
to match the behaviour of explicit intrinsics calls.
Currently LLVM crashes with:
> Assertion getInstruction() && "Not a call or invoke instruction!" failed.

With this patch LLVM produces a more sensible error message:
> Cannot select: ... i32 = ExternalSymbol'__foobar'

Author: Denys Zariaiev <denys.zariaiev@gmail.com>

Differential Revision: https://reviews.llvm.org/D55145

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349213 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoDebugInfo: Avoid using split DWARF when the split unit would be empty.
David Blaikie [Fri, 14 Dec 2018 22:44:46 +0000 (22:44 +0000)]
DebugInfo: Avoid using split DWARF when the split unit would be empty.

In ThinLTO many split CUs may be effectively empty because of the lack
of support for cross-unit references in split DWARF.

Using a split unit in those cases is just a waste/overhead - and turned
out to be one contributor to a significant symbolizer performance issue
when global variable debug info was being imported (see r348416 for the
primary fix) due to symbolizers seeing CUs with no ranges, assuming
there might still be addresses covered and walking into the split CU to
see if there are any ranges (when that split CU was in a DWP file, that
meant loading the DWP and its index, the index was extra large because
of all these fractured/empty CUs... and so was very expensive to load).

(the 3rd fix which will follow, is to assume that a CU with no ranges is
empty rather than merely missing its CU level range data - and to not
walk into its DIEs (split or otherwise) in search of address information
that is generally not present)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349207 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[codeview] Add begin/endSymbolRecord helpers, NFC
Reid Kleckner [Fri, 14 Dec 2018 22:40:28 +0000 (22:40 +0000)]
[codeview] Add begin/endSymbolRecord helpers, NFC

Previously beginning a symbol record was excessively verbose. Now it's a
bit simpler. This follows the same pattern as begin/endCVSubsection.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349205 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoDebugInfo: Move addAddrBase from DwarfUnit to DwarfCompileUnit
David Blaikie [Fri, 14 Dec 2018 22:34:03 +0000 (22:34 +0000)]
DebugInfo: Move addAddrBase from DwarfUnit to DwarfCompileUnit

Only CUs need an address table reference.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349203 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Hexagon] Add patterns for shifts of v2i16
Krzysztof Parzyszek [Fri, 14 Dec 2018 22:33:48 +0000 (22:33 +0000)]
[Hexagon] Add patterns for shifts of v2i16

This fixes https://llvm.org/PR39983.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349202 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[GlobalISel] LegalizerHelper: Implement fewerElementsVector for G_LOAD/G_STORE
Volkan Keles [Fri, 14 Dec 2018 22:11:20 +0000 (22:11 +0000)]
[GlobalISel] LegalizerHelper: Implement fewerElementsVector for G_LOAD/G_STORE

Reviewers: aemerson, dsanders, bogner, paquette, aditya_nandakumar

Reviewed By: dsanders

Subscribers: rovka, kristof.beyls, javed.absar, tschuett, llvm-commits

Differential Revision: https://reviews.llvm.org/D53728

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349200 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Hexagon] Use IMPLICIT_DEF to any-extend 32-bit values to 64 bits
Krzysztof Parzyszek [Fri, 14 Dec 2018 22:05:44 +0000 (22:05 +0000)]
[Hexagon] Use IMPLICIT_DEF to any-extend 32-bit values to 64 bits

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349199 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] Promote constant offset to the immediate by finding a new base with 13bit...
Farhana Aleen [Fri, 14 Dec 2018 21:13:14 +0000 (21:13 +0000)]
[AMDGPU] Promote constant offset to the immediate by finding a new base with 13bit constant offset from the nearby instructions.

Summary: Promote constant offset to immediate by recomputing the relative 13bit offset from nearby instructions.
 E.g.
  s_movk_i32 s0, 0x1800
  v_add_co_u32_e32 v0, vcc, s0, v2
  v_addc_co_u32_e32 v1, vcc, 0, v6, vcc

  s_movk_i32 s0, 0x1000
  v_add_co_u32_e32 v5, vcc, s0, v2
  v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
  global_load_dwordx2 v[5:6], v[5:6], off
  global_load_dwordx2 v[0:1], v[0:1], off
  =>
  s_movk_i32 s0, 0x1000
  v_add_co_u32_e32 v5, vcc, s0, v2
  v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
  global_load_dwordx2 v[5:6], v[5:6], off
  global_load_dwordx2 v[0:1], v[5:6], off offset:2048

Author: FarhanaAleen

Reviewed By: arsenm, rampitec

Subscribers: llvm-commits, AMDGPU

Differential Revision: https://reviews.llvm.org/D55539

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349196 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAdd missing includes and forward decls to unbreak build
Eric Fiselier [Fri, 14 Dec 2018 21:04:00 +0000 (21:04 +0000)]
Add missing includes and forward decls to unbreak build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349193 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SDAG] Ignore chain operand in REG_SEQUENCE when emitting instructions
Krzysztof Parzyszek [Fri, 14 Dec 2018 20:14:12 +0000 (20:14 +0000)]
[SDAG] Ignore chain operand in REG_SEQUENCE when emitting instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349186 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AArch64] Simplify the scheduling predicates (NFC)
Evandro Menezes [Fri, 14 Dec 2018 20:04:58 +0000 (20:04 +0000)]
[AArch64] Simplify the scheduling predicates (NFC)

The instruction encodings make it unnecessary to distinguish extended W-form
from X-form instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349185 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TransformWarning] Do not warn missed transformations in optnone functions.
Michael Kruse [Fri, 14 Dec 2018 19:45:43 +0000 (19:45 +0000)]
[TransformWarning] Do not warn missed transformations in optnone functions.

Optimization transformations are intentionally disabled by the 'optnone'
function attribute. Therefore do not warn if transformation metadata is
still present.

Using the legacy pass manager structure, the `skipFunction` method takes
care for the optnone attribute (already called before this patch). For
the new pass manager, there is no equivalent, so we check for the
'optnone' attribute manually.

Differential Revision: https://reviews.llvm.org/D55690

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349184 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] add tests for extractelement of FP binops; NFC
Sanjay Patel [Fri, 14 Dec 2018 19:15:54 +0000 (19:15 +0000)]
[x86] add tests for extractelement of FP binops; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349179 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM] make test immune to scalarization improvements; NFC
Sanjay Patel [Fri, 14 Dec 2018 18:47:04 +0000 (18:47 +0000)]
[ARM] make test immune to scalarization improvements; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349177 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] make tests immune to scalarization improvements; NFC
Sanjay Patel [Fri, 14 Dec 2018 18:44:16 +0000 (18:44 +0000)]
[x86] make tests immune to scalarization improvements; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349176 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[globalisel][combiner] Fix r349167 for release mode bots
Daniel Sanders [Fri, 14 Dec 2018 18:25:05 +0000 (18:25 +0000)]
[globalisel][combiner] Fix r349167 for release mode bots

This test relies on -debug-only which is unavailable in non-asserts builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349174 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ADT] Fix bugs in SmallBitVector.
Zachary Turner [Fri, 14 Dec 2018 18:21:20 +0000 (18:21 +0000)]
[ADT] Fix bugs in SmallBitVector.

Fixes:
  * find_last/find_last_unset - off-by-one error
  * Compound assignment ops and operator== when mixing big/small modes

Patch by Brad Moody
Differential Revision: https://reviews.llvm.org/D54933

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349173 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix Visual Studio PointerIntPair visualizer
Zachary Turner [Fri, 14 Dec 2018 18:20:21 +0000 (18:20 +0000)]
Fix Visual Studio PointerIntPair visualizer

Patch by: Trass3r

Differential Revision: https://reviews.llvm.org/D55252

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349172 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Transforms] Preserve metadata when converting invoke to call.
Michael Kruse [Fri, 14 Dec 2018 18:15:11 +0000 (18:15 +0000)]
[Transforms] Preserve metadata when converting invoke to call.

The `changeToCall` function did not preserve the invoke's metadata.
Currently, there is probably no metadata that depends on being applied
on a CallInst or InvokeInst. Therefore we can replace the instruction's
metadata.

This fixes http://llvm.org/PR39994

Suggested-by: Moritz Kreutzer <moritz.kreutzer@siemens.com>
Differential Revision: https://reviews.llvm.org/D55666

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349170 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[MS Demangler] Fail gracefully on invalid pointer types.
Zachary Turner [Fri, 14 Dec 2018 18:10:13 +0000 (18:10 +0000)]
[MS Demangler] Fail gracefully on invalid pointer types.

Once we detect a 'P', we know we a pointer type is upcoming, so
we make some assumptions about the output that follows.  If those
assumptions didn't hold, we would assert.  Instead, we should
fail gracefully and propagate the error up.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349169 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[MS Demangler] Add a regression test for an invalid mangled name.
Zachary Turner [Fri, 14 Dec 2018 17:59:27 +0000 (17:59 +0000)]
[MS Demangler] Add a regression test for an invalid mangled name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349168 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[globalisel][combiner] Make the CombinerChangeObserver a MachineFunction::Delegate
Daniel Sanders [Fri, 14 Dec 2018 17:50:14 +0000 (17:50 +0000)]
[globalisel][combiner] Make the CombinerChangeObserver a MachineFunction::Delegate

Summary:
This allows us to register it with the MachineFunction delegate and be
notified automatically about erasure and creation of instructions. However,
we still need explicit notification for modifications such as those caused
by setReg() or replaceRegWith().

There is a catch with this though. The notification for creation is
delivered before any operands can be added. While appropriate for
scheduling combiner work. This is unfortunate for debug output since an
opcode by itself doesn't provide sufficient information on what happened.
As a result, the work list remembers the instructions (when debug output is
requested) and emits a more complete dump later.

Another nit is that the MachineFunction::Delegate provides const pointers
which is inconvenient since we want to use it to schedule future
modification. To resolve this GISelWorkList now has an optional pointer to
the MachineFunction which describes the scope of the work it is permitted
to schedule. If a given MachineInstr* is in this function then it is
permitted to schedule work to be performed on the MachineInstr's. An
alternative to this would be to remove the const from the
MachineFunction::Delegate interface, however delegates are not permitted
to modify the MachineInstr's they receive.

In addition to this, the observer has three interface changes.
* erasedInstr() is now erasingInstr() to indicate it is about to be erased
  but still exists at the moment.
* changingInstr() and changedInstr() have been added to report changes
  before and after they are made. This allows us to trace the changes
  in the debug output.
* As a convenience changingAllUsesOfReg() and
  finishedChangingAllUsesOfReg() will report changingInstr() and
  changedInstr() for each use of a given register. This is primarily useful
  for changes caused by MachineRegisterInfo::replaceRegWith()

With this in place, both combine rules have been updated to report their
changes to the observer.

Finally, make some cosmetic changes to the debug output and make Combiner
and CombinerHelp

Reviewers: aditya_nandakumar, bogner, volkan, rtereshin, javed.absar

Reviewed By: aditya_nandakumar

Subscribers: mgorny, rovka, kristof.beyls, llvm-commits

Differential Revision: https://reviews.llvm.org/D52947

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349167 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AArch64] make test immune to scalarization improvements; NFC
Sanjay Patel [Fri, 14 Dec 2018 17:44:07 +0000 (17:44 +0000)]
[AArch64] make test immune to scalarization improvements; NFC

This is explicitly implementing what the comment says rather
than relying on the implicit zext of a costant operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349166 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix a crash in llvm-undname with invalid types.
Zachary Turner [Fri, 14 Dec 2018 17:43:56 +0000 (17:43 +0000)]
Fix a crash in llvm-undname with invalid types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349165 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SystemZ] make test immune to scalarization improvements; NFC
Sanjay Patel [Fri, 14 Dec 2018 17:28:52 +0000 (17:28 +0000)]
[SystemZ] make test immune to scalarization improvements; NFC

The undef operands mean this test is probably still too fragile
to accomplish what the comments suggest.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349164 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Hexagon] make test immune to scalarization improvements; NFC
Sanjay Patel [Fri, 14 Dec 2018 17:23:01 +0000 (17:23 +0000)]
[Hexagon] make test immune to scalarization improvements; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349163 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] auto-generate complete checks; NFC
Sanjay Patel [Fri, 14 Dec 2018 16:49:57 +0000 (16:49 +0000)]
[x86] auto-generate complete checks; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349162 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] regenerate test checks; NFC
Sanjay Patel [Fri, 14 Dec 2018 16:46:21 +0000 (16:46 +0000)]
[x86] regenerate test checks; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349161 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] make tests immune to scalarization improvements; NFC
Sanjay Patel [Fri, 14 Dec 2018 16:44:58 +0000 (16:44 +0000)]
[x86] make tests immune to scalarization improvements; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349160 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoNFC. Adding an empty line to test the updated commit credentials.
Ehsan Amiri [Fri, 14 Dec 2018 16:19:02 +0000 (16:19 +0000)]
NFC. Adding an empty line to test the updated commit credentials.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349158 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoImplement -frecord-command-line (-frecord-gcc-switches)
Scott Linder [Fri, 14 Dec 2018 15:38:15 +0000 (15:38 +0000)]
Implement -frecord-command-line (-frecord-gcc-switches)

Implement options in clang to enable recording the driver command-line
in an ELF section.

Implement a new special named metadata, llvm.commandline, to support
frontends embedding their command-line options in IR/ASM/ELF.

This differs from the GCC implementation in some key ways:

* In GCC there is only one command-line possible per compilation-unit,
  in LLVM it mirrors llvm.ident and multiple are allowed.
* In GCC individual options are separated by NULL bytes, in LLVM entire
  command-lines are separated by NULL bytes. The advantage of the GCC
  approach is to clearly delineate options in the face of embedded
  spaces. The advantage of the LLVM approach is to support merging
  multiple command-lines unambiguously, while handling embedded spaces
  with escaping.

Differential Revision: https://reviews.llvm.org/D54487
Clang Differential Revision: https://reviews.llvm.org/D54489

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349155 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RegAllocGreedy] IMPLICIT_DEF values shouldn't prefer registers
John Brawn [Fri, 14 Dec 2018 14:07:57 +0000 (14:07 +0000)]
[RegAllocGreedy] IMPLICIT_DEF values shouldn't prefer registers

It costs nothing to spill an IMPLICIT_DEF value (the only spill code that's
generated is a KILL of the value), so when creating split constraints if the
live-out value is IMPLICIT_DEF the exit constraint should be DontCare instead
of PrefReg.

Differential Revision: https://reviews.llvm.org/D55652

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349151 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM GlobalISel] Thumb2: casts between int and ptr
Diana Picus [Fri, 14 Dec 2018 13:45:38 +0000 (13:45 +0000)]
[ARM GlobalISel] Thumb2: casts between int and ptr

Mark as legal and add tests. Nothing special to do.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349147 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM GlobalISel] Remove duplicate test. NFCI
Diana Picus [Fri, 14 Dec 2018 13:28:34 +0000 (13:28 +0000)]
[ARM GlobalISel] Remove duplicate test. NFCI

Fixup for r349026. I forgot to delete these test functions from the
original file when I moved them to arm-legalize-exts.mir.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349146 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM GlobalISel] Minor refactoring. NFCI
Diana Picus [Fri, 14 Dec 2018 12:37:24 +0000 (12:37 +0000)]
[ARM GlobalISel] Minor refactoring. NFCI

Refactor the ARMInstructionSelector to cache some opcodes in the
constructor instead of checking all the time if we're in ARM or Thumb
mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349143 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM GlobalISel] Allow simple binary ops in Thumb2
Diana Picus [Fri, 14 Dec 2018 11:58:14 +0000 (11:58 +0000)]
[ARM GlobalISel] Allow simple binary ops in Thumb2

Mark G_ADD, G_SUB, G_MUL, G_AND, G_OR and G_XOR as legal for both ARM
and Thumb2.

Extract the legalizer tests for these opcodes into another file.

Add tests for the instruction selector.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349142 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TableGen:AsmWriter] Cope with consecutive tied operands.
Simon Tatham [Fri, 14 Dec 2018 11:39:55 +0000 (11:39 +0000)]
[TableGen:AsmWriter] Cope with consecutive tied operands.

When you define an instruction alias as a subclass of InstAlias, you
specify all the MC operands for the instruction it expands to, except
for operands that are tied to a previous one, which you leave out in
the expectation that the Tablegen output code will fill them in
automatically.

But the code in Tablegen's AsmWriter backend that skips over a tied
operand was doing it using 'if' instead of 'while', because it wasn't
expecting to find two tied operands in sequence.

So if an instruction updates a pair of registers in place, so that its
MC representation has two input operands tied to the output ones (for
example, Arm's UMLAL instruction), then any alias which wants to
expand to a special case of that instruction is likely to fail to
match, because the indices of subsequent operands will be off by one
in the generated printAliasInstr function.

This patch re-indents some existing code, so it's clearest when
viewed as a diff with whitespace changes ignored.

Reviewers: fhahn, rengolin, sdesmalen, atanasyan, asb, jholewinski, t.p.northover, kparzysz, craig.topper, stoklund

Reviewed By: rengolin

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

Differential Revision: https://reviews.llvm.org/D53816

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349141 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert rL349136: [llvm-exegesis] Optimize ToProcess in dbScan
Simon Pilgrim [Fri, 14 Dec 2018 09:25:08 +0000 (09:25 +0000)]
Revert rL349136: [llvm-exegesis] Optimize ToProcess in dbScan

Summary:
Use `vector<char> Added + vector<size_t> ToProcess` to replace `SetVector ToProcess`

We also check `Added[P]` to enqueueing a point more than once, which
also saves us a `ClusterIdForPoint_[Q].isUndef()` check.

Reviewers: courbet, RKSimon, gchatelet, john.brawn, lebedev.ri

Subscribers: tschuett, llvm-commits

Differential Revision: https://reviews.llvm.org/D54442
........
Patch wasn't approved and breaks buildbots

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349139 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombiner][X86] Prevent visitSIGN_EXTEND from returning N when (sext (setcc))...
Craig Topper [Fri, 14 Dec 2018 08:28:24 +0000 (08:28 +0000)]
[DAGCombiner][X86] Prevent visitSIGN_EXTEND from returning N when (sext (setcc)) already has the target desired type for the setcc

Summary:
If the setcc already has the target desired type we can reach the getSetCC/getSExtOrTrunc after the MatchingVecType check with the exact same types as the nodes we started with. This causes those causes VsetCC to be CSEd to N0 and the getSExtOrTrunc will CSE to N. When we return N, the caller will think that meant we called CombineTo and did our own worklist management. But that's not what happened. This prevents target hooks from being called for the node.

To fix this, I've now returned SDValue if the setcc is already the desired type. But to avoid some regressions in X86 I've had to disable one of the target combines that wasn't being reached before in the case of a (sext (setcc)). If we get vector widening legalization enabled that entire function will be deleted anyway so hopefully this is only for the short term.

Reviewers: RKSimon, spatel

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D55459

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349137 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis] Optimize ToProcess in dbScan
Fangrui Song [Fri, 14 Dec 2018 08:27:35 +0000 (08:27 +0000)]
[llvm-exegesis] Optimize ToProcess in dbScan

Summary:
Use `vector<char> Added + vector<size_t> ToProcess` to replace `SetVector ToProcess`

We also check `Added[P]` to enqueueing a point more than once, which
also saves us a `ClusterIdForPoint_[Q].isUndef()` check.

Reviewers: courbet, RKSimon, gchatelet, john.brawn, lebedev.ri

Subscribers: tschuett, llvm-commits

Differential Revision: https://reviews.llvm.org/D54442

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349136 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ThinLTO] Fix test added in rL349076
Fangrui Song [Fri, 14 Dec 2018 08:21:08 +0000 (08:21 +0000)]
[ThinLTO] Fix test added in rL349076

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349135 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Object] Rename getRelrRelocationType to getRelativeRelocationType
Fangrui Song [Fri, 14 Dec 2018 07:46:58 +0000 (07:46 +0000)]
[Object] Rename getRelrRelocationType to getRelativeRelocationType

Summary:
The two utility functions were added in D47919 to support SHT_RELR.
However, these are just relative relocations types and are't
necessarily be named Relr.

Reviewers: phosek, dberris

Reviewed By: dberris

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D55691

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349133 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-xray] Use correct variable name
Petr Hosek [Fri, 14 Dec 2018 06:06:19 +0000 (06:06 +0000)]
[llvm-xray] Use correct variable name

This fixes the compiler error introduced in r349129.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349130 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-xray] Store offset pointers in temporaries
Petr Hosek [Fri, 14 Dec 2018 05:56:20 +0000 (05:56 +0000)]
[llvm-xray] Store offset pointers in temporaries

DataExtractor::getU64 modifies the OffsetPtr which also pass to
RelocateOrElse which breaks on Windows. This addresses the issue
introduced in r349120.

Differential Revision: https://reviews.llvm.org/D55689

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349129 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[gn build] Merge r348963 and r349076
Nico Weber [Fri, 14 Dec 2018 03:20:46 +0000 (03:20 +0000)]
[gn build] Merge r348963 and r349076

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349124 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-xray] Support for PIE
Petr Hosek [Fri, 14 Dec 2018 01:37:56 +0000 (01:37 +0000)]
[llvm-xray] Support for PIE

When the instrumented binary is linked as PIE, we need to apply the
relative relocations to sleds. This is handled by the dynamic linker
at runtime, but when processing the file we have to do it ourselves.

Differential Revision: https://reviews.llvm.org/D55542

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349120 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[macho] save the SDK version stored in module metadata into the version min and
Alex Lorenz [Fri, 14 Dec 2018 01:14:10 +0000 (01:14 +0000)]
[macho] save the SDK version stored in module metadata into the version min and
build version load commands in the object file

This commit introduces a new metadata node called "SDK Version". It will be set
by the frontend to mark the platform SDK (macOS/iOS/etc) version which was used
during that particular compilation.
This node is used when machine code is emitted, by either saving the SDK version
into the appropriate macho load command (version min/build version), or by
emitting the assembly for these load commands with the SDK version specified as
well.
The assembly for both load commands is extended by allowing it to contain the
sdk_version X, Y [, Z] trailing directive to represent the SDK version
respectively.

rdar://45774000

Differential Revision: https://reviews.llvm.org/D55612

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349119 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoSilence CMP0048 warning in the benchmark utility library
Reid Kleckner [Fri, 14 Dec 2018 00:17:12 +0000 (00:17 +0000)]
Silence CMP0048 warning in the benchmark utility library

I'm testing this in LLVM before sending it upstream.

Part of PR38874

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349097 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[gn build] Add infrastructure to create symlinks and use it to create lld's symlinks
Nico Weber [Fri, 14 Dec 2018 00:16:33 +0000 (00:16 +0000)]
[gn build] Add infrastructure to create symlinks and use it to create lld's symlinks

This is slightly involved, see the comments in the code.

The GN build now builds a functional lld!

Differential Revision: https://reviews.llvm.org/D55606

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349096 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombiner] clean up visitEXTRACT_VECTOR_ELT
Sanjay Patel [Fri, 14 Dec 2018 00:09:08 +0000 (00:09 +0000)]
[DAGCombiner] clean up visitEXTRACT_VECTOR_ELT

This isn't quite NFC, but I don't know how to expose
any outward diffs from these changes. Mostly, this
was confusing because it used 'VT' to refer to the
operand type rather the usual type of the input node.

There's also a large block at the end that is dedicated
solely to matching loads, but that wasn't obvious. This
could probably be split up into separate functions to
make it easier to see.

It's still not clear to me when we make certain transforms
because the legality and constant conditions are
intertwined in a way that might be improved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349095 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Demote EmitTest to a helper function of EmitCmp. Route all callers except EmitC...
Craig Topper [Thu, 13 Dec 2018 23:55:30 +0000 (23:55 +0000)]
[X86] Demote EmitTest to a helper function of EmitCmp. Route all callers except EmitCmp through EmitCmp.

This requires the two callers to manifest a 0 to make EmitCmp call EmitTest.

I'm looking into changing how we combine TEST and flag setting instructions to not be part of lowering. And instead be part of DAG combine or isel. Which will mean EmitTest will probably become gutted and maybe disappear entirely.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349094 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert "[hwasan] Android: Switch from TLS_SLOT_TSAN(8) to TLS_SLOT_SANITIZER(6)"
Evgeniy Stepanov [Thu, 13 Dec 2018 23:47:50 +0000 (23:47 +0000)]
Revert "[hwasan] Android: Switch from TLS_SLOT_TSAN(8) to TLS_SLOT_SANITIZER(6)"

Breaks sanitizer-android buildbot.

This reverts commit af8443a984c3b491c9ca2996b8d126ea31e5ecbe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349092 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AArch64] Fix Exynos predicates (NFC)
Evandro Menezes [Thu, 13 Dec 2018 23:19:46 +0000 (23:19 +0000)]
[AArch64] Fix Exynos predicates (NFC)

Fix the logic in the definition of the `ExynosShiftExPred` as a more
specific version of `ExynosShiftPred`.  But, since `ExynosShiftExPred` is
not used yet, this change has NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349091 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SampleFDO] handle ProfileSampleAccurate when initializing function entry count
Wei Mi [Thu, 13 Dec 2018 21:51:42 +0000 (21:51 +0000)]
[SampleFDO] handle ProfileSampleAccurate when initializing function entry count

ProfileSampleAccurate is used to indicate the profile has exact match to the
code to be optimized.

Previously ProfileSampleAccurate is handled in ProfileSummaryInfo::isColdCallSite
and ProfileSummaryInfo::isColdBlock. A better solution is to initialize function
entry count to 0 when ProfileSampleAccurate is true, so we don't have to handle
ProfileSampleAccurate in multiple places.

Differential Revision: https://reviews.llvm.org/D55660

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349088 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert r348971: [AMDGPU] Support for "uniform-work-group-size" attribute
Aakanksha Patil [Thu, 13 Dec 2018 21:23:12 +0000 (21:23 +0000)]
Revert r348971: [AMDGPU] Support for "uniform-work-group-size" attribute

This patch breaks RADV (and probably RadeonSI as well)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349084 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAMDGPU/GlobalISel: Legalize/regbankselect block_addr
Matt Arsenault [Thu, 13 Dec 2018 20:34:15 +0000 (20:34 +0000)]
AMDGPU/GlobalISel: Legalize/regbankselect block_addr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349081 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoReapply "[MemCpyOpt] memset->memcpy forwarding with undef tail"
Nikita Popov [Thu, 13 Dec 2018 20:04:27 +0000 (20:04 +0000)]
Reapply "[MemCpyOpt] memset->memcpy forwarding with undef tail"

Currently memcpyopt optimizes cases like

    memset(a, byte, N);
    memcpy(b, a, M);

to

    memset(a, byte, N);
    memset(b, byte, M);

if M <= N. Often this allows further simplifications down the line,
which drop the first memset entirely.

This patch extends this optimization for the case where M > N, but we
know that the bytes a[N..M] are undef due to alloca/lifetime.start.

This situation arises relatively often for Rust code, because Rust does
not initialize trailing structure padding and loves to insert redundant
memcpys. This also fixes https://bugs.llvm.org/show_bug.cgi?id=39844.

The previous version of this patch did not perform dependency checking
properly: While the dependency is checked at the position of the memset,
the used size must be that of the memcpy. Previously the size of the
memset was used, which missed modification in the region
MemSetSize..CopySize, resulting in miscompiles. The added tests cover
variations of this issue.

Differential Revision: https://reviews.llvm.org/D55120

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349078 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ThinLTO] Compute synthetic function entry count
Easwaran Raman [Thu, 13 Dec 2018 19:54:27 +0000 (19:54 +0000)]
[ThinLTO] Compute synthetic function entry count

Summary:
This patch computes the synthetic function entry count on the whole
program callgraph (based on module summary) and writes the entry counts
to the summary. After function importing, this count gets attached to
the IR as metadata. Since it adds a new field to the summary, this bumps
up the version.

Reviewers: tejohnson

Subscribers: mehdi_amini, inglorion, llvm-commits

Differential Revision: https://reviews.llvm.org/D43521

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349076 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm] Address base discriminator overflow in X86DiscriminateMemOps
Mircea Trofin [Thu, 13 Dec 2018 19:40:59 +0000 (19:40 +0000)]
[llvm] Address base discriminator overflow in X86DiscriminateMemOps

Summary:
Macros are expanded on a single line. In case of large expansions,
with sufficiently many instructions with memory operands (and when
-fdebug-info-for-profiling is requested), we may be unable to generate
new base discriminator values - new values overflow (base
discriminators may not be larger than 2^12).

This CL warns instead of asserting in such a case. A subsequent CL
will add APIs to check for overflow before creating new debug info.

See https://bugs.llvm.org/show_bug.cgi?id=39890

Reviewers: davidxl, wmi, gbedwell

Reviewed By: davidxl

Subscribers: aprantl, llvm-commits

Differential Revision: https://reviews.llvm.org/D55643

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349075 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-size][libobject] Add explicit "inTextSegment" methods similar to "isText" secti...
Jordan Rupprecht [Thu, 13 Dec 2018 19:40:12 +0000 (19:40 +0000)]
[llvm-size][libobject] Add explicit "inTextSegment" methods similar to "isText" section methods to calculate size correctly.

Summary:
llvm-size uses "isText()" etc. which seem to indicate whether the section contains code-like things, not whether or not it will actually go in the text segment when in a fully linked executable.

The unit test added (elf-sizes.test) shows some types of sections that cause discrepencies versus the GNU size tool. llvm-size is not correctly reporting sizes of things mapping to text/data segments, at least for ELF files.

This fixes pr38723.

Reviewers: echristo, Bigcheese, MaskRay

Reviewed By: MaskRay

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D54369

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349074 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[CostModel][X86] Don't count 2 shuffles on the last level of a pairwise arithmetic...
Craig Topper [Thu, 13 Dec 2018 19:08:10 +0000 (19:08 +0000)]
[CostModel][X86] Don't count 2 shuffles on the last level of a pairwise arithmetic or min/max reduction

This is split from D55452 with the correct patch this time.

Pairwise reductions require two shuffles on every level but the last. On the last level the two shuffles are <1, u, u, u...> and <0, u, u, u...>, but <0, u, u, u...> will be dropped by InstCombine/DAGCombine as being an identity shuffle.

Differential Revision: https://reviews.llvm.org/D55615

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349072 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[CMake] llvm_codesign workaround for Xcode double-signing errors
Stefan Granitz [Thu, 13 Dec 2018 18:51:19 +0000 (18:51 +0000)]
[CMake] llvm_codesign workaround for Xcode double-signing errors

Summary:
When using Xcode to build LLVM with code signing, the post-build rule is executed even if the actual build-step was skipped. This causes double-signing errors. We can currently only avoid it by passing the `--force` flag.

Plus some polishing for my previous patch D54443.

Reviewers: beanz, kubamracek

Reviewed By: kubamracek

Subscribers: #lldb, mgorny, llvm-commits

Differential Revision: https://reviews.llvm.org/D55116

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349070 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[LoopUtils] Use i32 instead of `void`.
Davide Italiano [Thu, 13 Dec 2018 18:37:23 +0000 (18:37 +0000)]
[LoopUtils] Use i32 instead of `void`.

The actual type of the first argument of the @dbg intrinsic
doesn't really matter as we're setting it to `undef`, but the
bitcode reader is picky about `void` types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349069 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoDon't add unnecessary compiler flags to llvm-config output
Tom Stellard [Thu, 13 Dec 2018 18:21:23 +0000 (18:21 +0000)]
Don't add unnecessary compiler flags to llvm-config output

Summary:
llvm-config --cxxflags --cflags, should only output the minimal flags
required to link against the llvm libraries.  They currently contain
all flags used to compile llvm including flags like -g, -pedantic,
-Wall, etc, which users may not always want.

This changes the llvm-config output to only include flags that have been
explictly added to the COMPILE_FLAGS property of the llvm-config target
by the llvm build system.

llvm.org/PR8220

Output from llvm-config when running cmake with:
cmake -G Ninja .. -DCMAKE_CXX_FLAGS=-funroll-loops

Before:

--cppflags: -I$HEADERS_DIR/llvm/include -I$HEADERS_DIR/llvm/build/include
            -D_GNU_SOURCE -D_DEBUG -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
--cflags:   -I$HEADERS_DIR/llvm/include -I$HEADERS_DIR/llvm/build/include
            -fPIC -Werror=date-time -Wall -Wextra -Wno-unused-parameter -Wwrite-strings \
            -Wno-missing-field-initializers -pedantic -Wno-long-long -Wimplicit-fallthrough \
            -Wno-comment -fdiagnostics-color -g -D_GNU_SOURCE -D_DEBUG -D__STDC_CONSTANT_MACROS \
            -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
--cxxflags: -I$HEADERS_DIR/llvm/include -I$HEADERS_DIR/llvm/build/include\
            -funroll-loops -fPIC -fvisibility-inlines-hidden -Werror=date-time -std=c++11 -Wall \
            -Wextra -Wno-unused-parameter -Wwrite-strings -Wcast-qual -Wno-missing-field-initializers \
            -pedantic -Wno-long-long -Wimplicit-fallthrough -Wno-maybe-uninitialized \
            -Wno-class-memaccess -Wno-noexcept-type -Wdelete-non-virtual-dtor -Wno-comment \
            -fdiagnostics-color -g  -fno-exceptions -fno-rtti -D_GNU_SOURCE -D_DEBUG \
            -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS"

After:

--cppflags: -I$HEADERS_DIR/llvm/include -I$HEADERS_DIR/llvm/build/include \
            -D_GNU_SOURCE -D_DEBUG -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
--cflags:   -I$HEADERS_DIR/llvm/include -I$HEADERS_DIR/llvm/build/include \
            -D_GNU_SOURCE -D_DEBUG -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS
--cxxflags: -I$HEADERS_DIR/llvm/include -I$HEADERS_DIR/llvm/build/include \
             -std=c++11   -fno-exceptions -fno-rtti \
             -D_GNU_SOURCE -D_DEBUG -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS

Reviewers: sylvestre.ledru, infinity0, mgorny

Reviewed By: sylvestre.ledru, mgorny

Subscribers: mgorny, dmgreen, llvm-commits

Differential Revision: https://reviews.llvm.org/D55391

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349068 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoCorrectly handle skewed streams in drop_front() method.
Zachary Turner [Thu, 13 Dec 2018 18:11:33 +0000 (18:11 +0000)]
Correctly handle skewed streams in drop_front() method.

When calling BinaryStreamArray::drop_front(), if the stream
is skewed it means we must never drop the first bytes of the
stream since offsets which occur in records assume the existence
of those bytes.  So if we want to skip the first record in a
stream, then what we really want to do is just set the begin
pointer to the next record.  But we shouldn't actually remove
those bytes from the underlying view of the data.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349066 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[MachO][TLOF] Add support for local symbols in the indirect symbol table
Francis Visoiu Mistrih [Thu, 13 Dec 2018 17:23:30 +0000 (17:23 +0000)]
[MachO][TLOF] Add support for local symbols in the indirect symbol table

On 32-bit archs, before, we would assume that an indirect symbol will
never have local linkage. This can lead to miscompiles where the
symbol's value would be 0 and the linker would use that value, because
the indirect symbol table would contain the value
`INDIRECT_SYMBOL_LOCAL` for that specific symbol.

Differential Revision: https://reviews.llvm.org/D55573

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349060 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombiner] after simplifying demanded elements of vector operand of extract, revis...
Sanjay Patel [Thu, 13 Dec 2018 17:05:01 +0000 (17:05 +0000)]
[DAGCombiner] after simplifying demanded elements of vector operand of extract, revisit the extract; 2nd try

This is a retry of rL349051 (reverted at rL349056). I changed the check for dead-ness from
number of uses to an opcode test for DELETED_NODE based on existing similar code.

Differential Revision: https://reviews.llvm.org/D55655

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349058 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][SSE] Add SSE vector imm/var shift support to SimplifyDemandedVectorEltsForTarge...
Simon Pilgrim [Thu, 13 Dec 2018 16:39:29 +0000 (16:39 +0000)]
[X86][SSE] Add SSE vector imm/var shift support to SimplifyDemandedVectorEltsForTargetNode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349057 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agorevert rL349051: [DAGCombiner] after simplifying demanded elements of vector operand...
Sanjay Patel [Thu, 13 Dec 2018 16:32:44 +0000 (16:32 +0000)]
revert rL349051: [DAGCombiner] after simplifying demanded elements of vector operand of extract, revisit the extract

This causes an address sanitizer bot failure:
http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-fast/builds/27187/steps/check-llvm%20asan/logs/stdio

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349056 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRecommit r349041: [tblgen][disasm] Separate encodings from instructions
Daniel Sanders [Thu, 13 Dec 2018 16:17:54 +0000 (16:17 +0000)]
Recommit r349041: [tblgen][disasm] Separate encodings from instructions

Removed const from the ArrayRef<const EncodingAndInst> to avoid the
std::vector<const EncodingAndInst> that G++ saw

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349055 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][SSE] Fix all remaining modulo vector rotation amounts (PR38243)
Simon Pilgrim [Thu, 13 Dec 2018 15:50:31 +0000 (15:50 +0000)]
[X86][SSE] Fix all remaining modulo vector rotation amounts (PR38243)

There's still a couple of minor SimplifyDemandedElts regressions in some of the shift amount splats that will be fixed in future patches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349052 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombiner] after simplifying demanded elements of vector operand of extract, revis...
Sanjay Patel [Thu, 13 Dec 2018 15:44:26 +0000 (15:44 +0000)]
[DAGCombiner] after simplifying demanded elements of vector operand of extract, revisit the extract

Differential Revision: https://reviews.llvm.org/D55655

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349051 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Sparc] Add membar assembler tags
Daniel Cederman [Thu, 13 Dec 2018 15:29:12 +0000 (15:29 +0000)]
[Sparc] Add membar assembler tags

Summary: The Sparc V9 membar instruction can enforce different types of
memory orderings depending on the value in its immediate field.  In the
architectural manual the type is selected by combining different assembler
tags into a mask. This patch adds support for these tags.

Reviewers: jyknight, venkatra, brad

Reviewed By: jyknight

Subscribers: fedor.sergeev, jrtc27, jfb, llvm-commits

Differential Revision: https://reviews.llvm.org/D53491

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349048 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][SSE] Fix modulo rotation amounts for v8i16/v16i16/v4i32 (PR38243)
Simon Pilgrim [Thu, 13 Dec 2018 15:23:09 +0000 (15:23 +0000)]
[X86][SSE] Fix modulo rotation amounts for v8i16/v16i16/v4i32 (PR38243)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349047 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert r349041: [tblgen][disasm] Separate encodings from instructions
Daniel Sanders [Thu, 13 Dec 2018 15:14:21 +0000 (15:14 +0000)]
Revert r349041: [tblgen][disasm] Separate encodings from instructions

One of the GCC based bots is objecting to a vector of const EncodingAndInst's:
In file included from /usr/include/c++/8/vector:64,
                 from /export/users/atombot/llvm/clang-atom-d525-fedora-rel/llvm/utils/TableGen/CodeGenInstruction.h:22,
                 from /export/users/atombot/llvm/clang-atom-d525-fedora-rel/llvm/utils/TableGen/FixedLenDecoderEmitter.cpp:15:
/usr/include/c++/8/bits/stl_vector.h: In instantiation of 'class std::vector<const {anonymous}::EncodingAndInst, std::allocator<const {anonymous}::EncodingAndInst> >':
/export/users/atombot/llvm/clang-atom-d525-fedora-rel/llvm/utils/TableGen/FixedLenDecoderEmitter.cpp:375:32:   required from here
/usr/include/c++/8/bits/stl_vector.h:351:21: error: static assertion failed: std::vector must have a non-const, non-volatile value_type
       static_assert(is_same<typename remove_cv<_Tp>::type, _Tp>::value,
                     ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/include/c++/8/bits/stl_vector.h:354:21: error: static assertion failed: std::vector must have the same value_type as its allocator
       static_assert(is_same<typename _Alloc::value_type, _Tp>::value,
                     ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349046 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Sparc] Use float register for integer constrained with "f" in inline asm
Daniel Cederman [Thu, 13 Dec 2018 15:13:29 +0000 (15:13 +0000)]
[Sparc] Use float register for integer constrained with "f"  in inline asm

Summary:
Constraining an integer value to a floating point register using "f"
causes an llvm_unreachable to trigger. This patch allows i32 integers
to be placed in a single precision float register and i64 integers to
be placed in a double precision float register. This matches the behavior
of GCC.

For other types the llvm_unreachable is removed to instead trigger an
error message that points out the offending line.

Reviewers: jyknight, venkatra

Reviewed By: jyknight

Subscribers: eraman, fedor.sergeev, jrtc27, llvm-commits

Differential Revision: https://reviews.llvm.org/D51614

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349045 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PowerPC][NFC] Sorting out Pseudo related classes to avoid confusion
Jinsong Ji [Thu, 13 Dec 2018 15:12:57 +0000 (15:12 +0000)]
[PowerPC][NFC] Sorting out Pseudo related classes to avoid confusion

There are several Pseudo in PowerPC backend.
eg:

* ISel Pseudo-instructions , which has let usesCustomInserter=1 in td
ExpandISelPseudos -> EmitInstrWithCustomInserter will deal with them.
* Post-RA pseudo instruction, which has let isPseudo = 1 in td, or Standard pseudo (SUBREG_TO_REG,COPY etc.)
ExpandPostRAPseudos -> expandPostRAPseudo will expand them
* Multi-instruction pseudo operations will expand them PPCAsmPrinter::EmitInstruction
* Pseudo instruction in CodeEmitter, which has encoding of 0.

Currently, in td files, especially PPCInstrVSX.td,
we did not distinguish Post-RA pseudo instruction and Pseudo instruction in CodeEmitter very clearly.

This patch is to

* Rename Pseudo<> class to PPCEmitTimePseudo, which means encoding of 0 in CodeEmitter
* Introduce new class PPCPostRAExpPseudo <> for previous PostRA Pseudo
* Introduce new class PPCCustomInserterPseudo <> for previous Isel Pseudo

Differential Revision: https://reviews.llvm.org/D55143

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349044 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[mir] Fix uninitialized variable in r349035 noticed by clang-atom-d525-fedora-rel...
Daniel Sanders [Thu, 13 Dec 2018 15:05:27 +0000 (15:05 +0000)]
[mir] Fix uninitialized variable in r349035 noticed by clang-atom-d525-fedora-rel and 3 other bots

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349043 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[tblgen][disasm] Separate encodings from instructions
Daniel Sanders [Thu, 13 Dec 2018 14:55:57 +0000 (14:55 +0000)]
[tblgen][disasm] Separate encodings from instructions

Summary:
Separate the concept of an encoding from an instruction. This will enable
the definition of additional encodings for the same instruction which can
be used to support variable length instruction sets in the disassembler
(and potentially assembler but I'm not working towards that right now)
without causing an explosion in the number of Instruction records that
CodeGen then has to pick between.

Reviewers: bogner, charukcs

Reviewed By: bogner

Subscribers: kparzysz, llvm-commits

Differential Revision: https://reviews.llvm.org/D52366

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349041 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][SSE] Merge the vXi16/vXi32 vector rotation expansion cases. NFCI.
Simon Pilgrim [Thu, 13 Dec 2018 14:51:28 +0000 (14:51 +0000)]
[X86][SSE] Merge the vXi16/vXi32 vector rotation expansion cases. NFCI.

Merged the repeated code into a single if().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349040 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SystemZ] Pass copy-hinted regs first from getRegAllocationHints().
Jonas Paulsson [Thu, 13 Dec 2018 14:37:05 +0000 (14:37 +0000)]
[SystemZ]  Pass copy-hinted regs first from getRegAllocationHints().

When computing register allocation hints for a GRX32Bit register, make sure
that any of the hinted registers that are also copy hints are returned first
in the list.

Review: Ulrich Weigand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349037 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[mir] Serialize DILocation inline when not possible to use a metadata reference
Daniel Sanders [Thu, 13 Dec 2018 14:25:27 +0000 (14:25 +0000)]
[mir] Serialize DILocation inline when not possible to use a metadata reference

Summary:
Sometimes MIR-level passes create DILocations that were not present in the
LLVM-IR. For example, it may merge two DILocations together to produce a
DILocation that points to line 0.

Previously, the address of these DILocations were printed which prevented the
MIR from being read back into LLVM. With this patch, DILocations will use
metadata references where possible and fall back on serializing them inline like so:
    MOV32mr %stack.0.x.addr, 1, _, 0, _, %0, debug-location !DILocation(line: 1, scope: !15)

Reviewers: aprantl, vsk, arphaman

Reviewed By: aprantl

Subscribers: probinson, llvm-commits

Tags: #debug-info

Differential Revision: https://reviews.llvm.org/D55243

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349035 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][BWI] Don't custom lower vXi8 rotations.
Simon Pilgrim [Thu, 13 Dec 2018 13:44:33 +0000 (13:44 +0000)]
[X86][BWI] Don't custom lower vXi8 rotations.

We always expand to shifts anyhow - test changes are just different scheduling only.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349034 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC][PowerPC] add verify-machineinstrs check
Chen Zheng [Thu, 13 Dec 2018 12:55:42 +0000 (12:55 +0000)]
[NFC][PowerPC] add verify-machineinstrs check

After rL349029 and rL348566, sj-ctr-loop.ll is ok for verify-machineinstrs check.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349030 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PowerPC] intrinsic llvm.eh.sjlj.setjmp should not have flag isBarrier.
Chen Zheng [Thu, 13 Dec 2018 12:25:20 +0000 (12:25 +0000)]
[PowerPC] intrinsic llvm.eh.sjlj.setjmp should not have flag isBarrier.

Differential Revision: https://reviews.llvm.org/D55499

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349029 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombine] Moved X86 rotate_amount % bitwidth == 0 early out to DAGCombiner
Simon Pilgrim [Thu, 13 Dec 2018 12:23:32 +0000 (12:23 +0000)]
[DAGCombine] Moved X86 rotate_amount % bitwidth == 0 early out to DAGCombiner

Remove common code from custom lowering (code is still safe if somehow a zero value gets used).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349028 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM GlobalISel] Support exts and truncs for Thumb2
Diana Picus [Thu, 13 Dec 2018 12:06:54 +0000 (12:06 +0000)]
[ARM GlobalISel] Support exts and truncs for Thumb2

Mark G_SEXT, G_ZEXT and G_ANYEXT to 32 bits as legal and add support for
them in the instruction selector. This uses handwritten code again
because the patterns that are generated with TableGen are tuned for what
the DAG combiner would produce and not for simple sext/zext nodes.
Luckily, we only need to update the opcodes to use the Thumb2 variants,
everything else can be reused from ARM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349026 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TargetLowering] Add ISD::ROTL/ROTR vector expansion
Simon Pilgrim [Thu, 13 Dec 2018 11:20:48 +0000 (11:20 +0000)]
[TargetLowering] Add ISD::ROTL/ROTR vector expansion

Move existing rotation expansion code into TargetLowering and set it up for vectors as well.

Ideally this would share more of the funnel shift expansion, but we handle the shift amount modulo quite differently at the moment.

Begun removing x86 vector rotate custom lowering to use the expansion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349025 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RISCV] Add support for the various RISC-V FMA instruction variants
Alex Bradbury [Thu, 13 Dec 2018 10:49:05 +0000 (10:49 +0000)]
[RISCV] Add support for the various RISC-V FMA instruction variants

Adds support for the various RISC-V FMA instructions (fmadd, fmsub, fnmsub, fnmadd).

The criteria for choosing whether a fused add or subtract is used, as well as
whether the product is negated or not, is whether some of the arguments to the
llvm.fma.* intrinsic are negated or not. In the tests, extraneous fadd
instructions were added to avoid the negation being performed using a xor
trick, which prevented the proper FMA forms from being selected and thus
tested.

The FMA instruction patterns might seem incorrect (e.g., fnmadd: -rs1 * rs2 -
rs3), but they should be correct. The misleading names were inherited from
MIPS, where the negation happens after computing the sum.

The llvm.fmuladd.* intrinsics still do not generate RISC-V FMA instructions,
as that depends on TargetLowering::isFMAFasterthanFMulAndFAdd.

Some comments in the test files about what type of instructions are there
tested were updated, to better reflect the current content of those test
files.

Differential Revision: https://reviews.llvm.org/D54205
Patch by Luís Marques.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349023 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AArch64] Catch some more CMN opportunities.
Arnaud A. de Grandmaison [Thu, 13 Dec 2018 10:31:32 +0000 (10:31 +0000)]
[AArch64] Catch some more CMN opportunities.

Fixes https://bugs.llvm.org/show_bug.cgi?id=33486

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349022 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[CodeGen] Allow mempcy/memset to generate small overlapping stores.
Clement Courbet [Thu, 13 Dec 2018 09:56:19 +0000 (09:56 +0000)]
[CodeGen] Allow mempcy/memset to generate small overlapping stores.

Summary:
All targets either just return false here or properly model `Fast`, so I
don't think there is any reason to prevent CodeGen from doing the right
thing here.

Subscribers: nemanjai, javed.absar, eraman, jsji, llvm-commits

Differential Revision: https://reviews.llvm.org/D55365

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349016 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[asan] Don't check ODR violations for particular types of globals
Vitaly Buka [Thu, 13 Dec 2018 09:47:39 +0000 (09:47 +0000)]
[asan] Don't check ODR violations for particular types of globals

Summary:
private and internal: should not trigger ODR at all.
unnamed_addr: current ODR checking approach fail and rereport false violation if
a linker merges such globals
linkonce_odr, weak_odr: could cause similar problems and they are already not
instrumented for ELF.

Reviewers: eugenis, kcc

Subscribers: kubamracek, hiraditya, llvm-commits

Differential Revision: https://reviews.llvm.org/D55621

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349015 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAMDGPU/GlobalISel: Legalize f64 fadd/fmul
Matt Arsenault [Thu, 13 Dec 2018 08:27:48 +0000 (08:27 +0000)]
AMDGPU/GlobalISel: Legalize f64 fadd/fmul

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349014 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix missing C++ mode comment in header
Matt Arsenault [Thu, 13 Dec 2018 08:23:56 +0000 (08:23 +0000)]
Fix missing C++ mode comment in header

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349013 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAMDGPU/GlobalISel: RegBankSelect some simple operations
Matt Arsenault [Thu, 13 Dec 2018 08:23:51 +0000 (08:23 +0000)]
AMDGPU/GlobalISel: RegBankSelect some simple operations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349012 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAMDGPU/GlobalISel: Test cleanups
Matt Arsenault [Thu, 13 Dec 2018 08:11:45 +0000 (08:11 +0000)]
AMDGPU/GlobalISel: Test cleanups

Remove IR and registers sections

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349011 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Remove assert leftover from when i1 was a legal type. Add more accurate assert...
Craig Topper [Thu, 13 Dec 2018 06:14:25 +0000 (06:14 +0000)]
[X86] Remove assert leftover from when i1 was a legal type. Add more accurate assert. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349007 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] Fix build failure, second attempt
Stanislav Mekhanoshin [Thu, 13 Dec 2018 05:52:11 +0000 (05:52 +0000)]
[AMDGPU] Fix build failure, second attempt

Some compilers complain that variable is captured and some
complain when it is not. Switch to [&].

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349006 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] Fix build failure
Stanislav Mekhanoshin [Thu, 13 Dec 2018 05:21:25 +0000 (05:21 +0000)]
[AMDGPU] Fix build failure

Fixed error 'lambda capture 'CondReg' is not required to be captured
for this use'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349005 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] Simplify negated condition
Stanislav Mekhanoshin [Thu, 13 Dec 2018 03:17:40 +0000 (03:17 +0000)]
[AMDGPU] Simplify negated condition

Optimize sequence:

  %sel = V_CNDMASK_B32_e64 0, 1, %cc
  %cmp = V_CMP_NE_U32 1, %1
  $vcc = S_AND_B64 $exec, %cmp
  S_CBRANCH_VCC[N]Z
=>
  $vcc = S_ANDN2_B64 $exec, %cc
  S_CBRANCH_VCC[N]Z

It is the negation pattern inserted by DAGCombiner::visitBRCOND() in the
rebuildSetCC().

Differential Revision: https://reviews.llvm.org/D55402

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@349003 91177308-0d34-0410-b5e6-96231b3b80d8