OSDN Git Service

android-x86/external-llvm.git
7 years agoAMDGPU: Add definition for v_swap_b32
Matt Arsenault [Tue, 28 Feb 2017 21:09:04 +0000 (21:09 +0000)]
AMDGPU: Add definition for v_swap_b32

This is somewhat tricky because there are two
pairs of tied operands, and it isn't allowed to be
VOP3 encoded.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296519 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add definition for v_xad_u32
Matt Arsenault [Tue, 28 Feb 2017 20:27:30 +0000 (20:27 +0000)]
AMDGPU: Add definition for v_xad_u32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296515 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DWARFv5] Emit new unit header format.
Paul Robinson [Tue, 28 Feb 2017 20:24:55 +0000 (20:24 +0000)]
[DWARFv5] Emit new unit header format.

Requesting DWARF v5 will now get you the new compile-unit and
type-unit headers.  llvm-dwarfdump will also recognize them.

Differential Revision: http://reviews.llvm.org/D30206

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296514 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add ds_nop to assembler
Matt Arsenault [Tue, 28 Feb 2017 20:15:46 +0000 (20:15 +0000)]
AMDGPU: Add ds_nop to assembler

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296513 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add definitions for ds_{read|write}_b{96|128}
Matt Arsenault [Tue, 28 Feb 2017 20:15:43 +0000 (20:15 +0000)]
AMDGPU: Add definitions for ds_{read|write}_b{96|128}

It's not clear to me if this is always better than
doing ds_write2_b64 This adds the constraint of
a 128-bit register input instead of a pair of
64-bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296512 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AMDGPU] Add second pass of the scheduler
Stanislav Mekhanoshin [Tue, 28 Feb 2017 19:20:33 +0000 (19:20 +0000)]
[AMDGPU] Add second pass of the scheduler

If during scheduling we have identified that we cannot keep optimistic
occupancy increase critical register pressure limit and try scheduling
of the whole function again. In this case blocks with smaller pressure
will have a chance for better scheduling.

Differential Revision: https://reviews.llvm.org/D30442

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296506 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAGCombiner] use dyn_cast values in foldSelectOfConstants(); NFC
Sanjay Patel [Tue, 28 Feb 2017 18:41:49 +0000 (18:41 +0000)]
[DAGCombiner] use dyn_cast values in foldSelectOfConstants(); NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296502 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix -Wcovered-switch-default warning.
Zachary Turner [Tue, 28 Feb 2017 18:35:40 +0000 (18:35 +0000)]
Fix -Wcovered-switch-default warning.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296501 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LCG] Fix EXPENSIVE_CHECKS typo. NFC
Francis Visoiu Mistrih [Tue, 28 Feb 2017 18:34:55 +0000 (18:34 +0000)]
[LCG] Fix EXPENSIVE_CHECKS typo. NFC

Differential Revision: https://reviews.llvm.org/D30434

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296500 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd function importing info from samplepgo profile to the module summary.
Dehao Chen [Tue, 28 Feb 2017 18:09:44 +0000 (18:09 +0000)]
Add function importing info from samplepgo profile to the module summary.

Summary: For SamplePGO, the profile may contain cross-module inline stacks. As we need to make sure the profile annotation happens when all the hot inline stacks are expanded, we need to pass this info to the module importer so that it can import proper functions if necessary. This patch implemented this feature by emitting cross-module targets as part of function entry metadata. In the module-summary phase, the metadata is used to build call edges that points to functions need to be imported.

Reviewers: mehdi_amini, tejohnson

Reviewed By: tejohnson

Subscribers: davidxl, llvm-commits

Differential Revision: https://reviews.llvm.org/D30053

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296498 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoWorkaround MSVC bug when using TrailingObjects from a template.
James Y Knight [Tue, 28 Feb 2017 18:05:41 +0000 (18:05 +0000)]
Workaround MSVC bug when using TrailingObjects from a template.

MSVC appears to be getting confused as to whether OverloadToken is
supposed to be public or not.

This was discovered by code in Swift, and has been reported to
microsoft by hughbe:
https://connect.microsoft.com/VisualStudio/feedback/details/3116517

Differential Revision: https://reviews.llvm.org/D29880

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296497 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] add alternate IR tests for select of constants; NFC
Sanjay Patel [Tue, 28 Feb 2017 18:02:38 +0000 (18:02 +0000)]
[x86] add alternate IR tests for select of constants; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296496 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PDB] Add BinaryStreamError.
Zachary Turner [Tue, 28 Feb 2017 17:49:34 +0000 (17:49 +0000)]
[PDB] Add BinaryStreamError.

This migrates the stream code away from MSFError to using its
own custom Error class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296494 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoSet default CPU for OpenBSD/arm to Cortex-A8
Brad Smith [Tue, 28 Feb 2017 17:28:35 +0000 (17:28 +0000)]
Set default CPU for OpenBSD/arm to Cortex-A8

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296493 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix issue with test case. Make test x86_64 specific
David Bozier [Tue, 28 Feb 2017 17:25:38 +0000 (17:25 +0000)]
Fix issue with test case. Make test x86_64 specific

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296492 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AMDGPU] New method to estimate register pressure
Stanislav Mekhanoshin [Tue, 28 Feb 2017 17:22:39 +0000 (17:22 +0000)]
[AMDGPU] New method to estimate register pressure

This change introduces new method to estimate register pressure in
GCNScheduler. Standard RPTracker gives huge error due to the following
reasons:

1. It does not account for live-ins or live-outs if value is not used
in the region itself. That creates a huge error in a very common case
if there are a lot of live-thu registers.
2. It does not properly count subregs.
3. It assumes a register used as an input operand can be reused as an
output. This is not always possible by itself, this is not what RA
will finally do in many cases for various reasons not limited to RA's
inability to do so, and this is not so if the value is actually a
live-thu.

In addition we can now see clear separation between live-in pressure
which we cannot change with the scheduling and tentative pressure
which we can change.

Differential Revision: https://reviews.llvm.org/D30439

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296491 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AMDGPU] Change amd_kernel_code_t's minor version to 1
Konstantin Zhuravlyov [Tue, 28 Feb 2017 17:17:52 +0000 (17:17 +0000)]
[AMDGPU] Change amd_kernel_code_t's minor version to 1
  - We do emit amd_kernel_code_t v1.1

Differential Revision: https://reviews.llvm.org/D30433

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296489 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoStrip debug info when inlining into a nodebug function.
Adrian Prantl [Tue, 28 Feb 2017 16:58:13 +0000 (16:58 +0000)]
Strip debug info when inlining into a nodebug function.

The LLVM backend cannot produce any debug info for an llvm::Function
without a DISubprogram attachment. When inlining a debug-info-carrying
function into a nodebug function, there is therefore no reason to keep
any debug info intrinsic calls or debug locations on the instructions.

This fixes a problem discovered in PR32042.

rdar://problem/30679307

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296488 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[llvm-cov] Error-out when an unsupported format is used (PR32087)
Vedant Kumar [Tue, 28 Feb 2017 16:57:28 +0000 (16:57 +0000)]
[llvm-cov] Error-out when an unsupported format is used (PR32087)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296487 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAGISel] When checking if chain node is foldable, make sure the intermediate nodes...
Craig Topper [Tue, 28 Feb 2017 16:52:05 +0000 (16:52 +0000)]
[DAGISel] When checking if chain node is foldable, make sure the intermediate nodes have a single use across all results not just the result that was used to reach the chain node.

This recovers a test case that was severely broken by r296476, my making sure we don't create ADD/ADC that loads and stores when there is also a flag dependency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296486 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AMDGPU] Fix read-undef flags when schedule is reverted
Stanislav Mekhanoshin [Tue, 28 Feb 2017 16:26:27 +0000 (16:26 +0000)]
[AMDGPU] Fix read-undef flags when schedule is reverted

If two subregs of the same register are defined and we need to revert
schedule changing def order, we will end up with both instructions
having def,read-undef flags because adjustLaneLiveness() will only set
this flag but will not remove it.

Fix this by removing read-undef flags before calling adjustLaneLiveness.

Differential Revision: https://reviews.llvm.org/D30428

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296484 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Stack Protection] Add diagnostic information for why stack protection was applied...
David Bozier [Tue, 28 Feb 2017 16:02:37 +0000 (16:02 +0000)]
[Stack Protection] Add diagnostic information for why stack protection was applied to a function

Stack Smash Protection is not completely free, so in hot code, the overhead it causes can cause performance issues. By adding diagnostic information for which functions have SSP and why, a user can quickly determine what they can do to stop SSP being applied to a specific hot function.

This change adds a remark that is reported by the stack protection code when an instruction or attribute is encountered that causes SSP to be applied.

Patch by: James Henderson

Differential Revision: https://reviews.llvm.org/D29023

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296483 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[mips] Fix 64bit slt/sltu/nor with immediates
Simon Dardis [Tue, 28 Feb 2017 15:55:23 +0000 (15:55 +0000)]
[mips] Fix 64bit slt/sltu/nor with immediates

Patch By: Alexander Richardson

Reviewers: atanasyan, theraven, sdardis

Differential Revision: https://reviews.llvm.org/D30330

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296482 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r296474 - [globalisel] Change LLT constructor string into an LLT subclass...
Daniel Sanders [Tue, 28 Feb 2017 15:00:27 +0000 (15:00 +0000)]
Revert r296474 - [globalisel] Change LLT constructor string into an LLT subclass that knows how to generate it.

There's a circular dependency that's only revealed when LLVM_ENABLE_MODULES=1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296478 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIn visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled.
Nirav Dave [Tue, 28 Feb 2017 14:24:15 +0000 (14:24 +0000)]
In visitSTORE, always use FindBetterChain, rather than only when UseAA is enabled.

    Recommiting after fixup of 32-bit aliasing sign offset bug in DAGCombiner.

    * Simplify Consecutive Merge Store Candidate Search

    Now that address aliasing is much less conservative, push through
    simplified store merging search and chain alias analysis which only
    checks for parallel stores through the chain subgraph. This is cleaner
    as the separation of non-interfering loads/stores from the
    store-merging logic.

    When merging stores search up the chain through a single load, and
    finds all possible stores by looking down from through a load and a
    TokenFactor to all stores visited.

    This improves the quality of the output SelectionDAG and the output
    Codegen (save perhaps for some ARM cases where we correctly constructs
    wider loads, but then promotes them to float operations which appear
    but requires more expensive constant generation).

    Some minor peephole optimizations to deal with improved SubDAG shapes (listed below)

    Additional Minor Changes:

      1. Finishes removing unused AliasLoad code

      2. Unifies the chain aggregation in the merged stores across code
         paths

      3. Re-add the Store node to the worklist after calling
         SimplifyDemandedBits.

      4. Increase GatherAllAliasesMaxDepth from 6 to 18. That number is
         arbitrary, but seems sufficient to not cause regressions in
         tests.

      5. Remove Chain dependencies of Memory operations on CopyfromReg
         nodes as these are captured by data dependence

      6. Forward loads-store values through tokenfactors containing
          {CopyToReg,CopyFromReg} Values.

      7. Peephole to convert buildvector of extract_vector_elt to
         extract_subvector if possible (see
         CodeGen/AArch64/store-merge.ll)

      8. Store merging for the ARM target is restricted to 32-bit as
         some in some contexts invalid 64-bit operations are being
         generated. This can be removed once appropriate checks are
         added.

    This finishes the change Matt Arsenault started in r246307 and
    jyknight's original patch.

    Many tests required some changes as memory operations are now
    reorderable, improving load-store forwarding. One test in
    particular is worth noting:

      CodeGen/PowerPC/ppc64-align-long-double.ll - Improved load-store
      forwarding converts a load-store pair into a parallel store and
      a memory-realized bitcast of the same value. However, because we
      lose the sharing of the explicit and implicit store values we
      must create another local store. A similar transformation
      happens before SelectionDAG as well.

    Reviewers: arsenm, hfinkel, tstellarAMD, jyknight, nhaehnle

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296476 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[globalisel] Change LLT constructor string into an LLT subclass that knows how to...
Daniel Sanders [Tue, 28 Feb 2017 14:21:31 +0000 (14:21 +0000)]
[globalisel] Change LLT constructor string into an LLT subclass that knows how to generate it.

Summary:
This will allow future patches to inspect the details of the LLT. The implementation is now split between
the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns.

Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem.

Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar

Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls

Differential Revision: https://reviews.llvm.org/D30046

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296474 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Lower i32 and fp call parameters on the stack
Diana Picus [Tue, 28 Feb 2017 14:17:53 +0000 (14:17 +0000)]
[ARM] GlobalISel: Lower i32 and fp call parameters on the stack

Lower i32, float and double parameters that need to live on the stack. This
boils down to creating some G_GEPs starting from the stack pointer and storing
the values there. During the process we also keep track of the stack size and
use the final value in the ADJCALLSTACKDOWN/UP instructions.

We currently assert for smaller types, since they usually require extensions.
They will be handled in a separate patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296473 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Select 32-bit G_CONSTANT
Diana Picus [Tue, 28 Feb 2017 13:05:42 +0000 (13:05 +0000)]
[ARM] GlobalISel: Select 32-bit G_CONSTANT

Put it into a register by means of a MOVi.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296471 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Add mapping for G_CONSTANT
Diana Picus [Tue, 28 Feb 2017 12:13:58 +0000 (12:13 +0000)]
[ARM] GlobalISel: Add mapping for G_CONSTANT

Like G_FRAME_INDEX, G_CONSTANT has one register operand and one non-register
operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296469 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Legalize 32-bit constants
Diana Picus [Tue, 28 Feb 2017 11:33:46 +0000 (11:33 +0000)]
[ARM] GlobalISel: Legalize 32-bit constants

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296468 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Assembler] Add test for !srcloc references in assembler diags
Sanne Wouda [Tue, 28 Feb 2017 10:34:48 +0000 (10:34 +0000)]
[Assembler] Add test for !srcloc references in assembler diags

Summary:
clang adds !srcloc metadata to inline assembly in LLVM bitcode generated
for inline assembly in C.  The value of this !srcloc is passed to the
diagnostics handler if the inline assembly generates a diagnostic.
clang is able to turn this cookie back to a location in the C source
file.

To test this functionality without a dependency, make llc print the
!srcloc metadata if it is present.  The added test uses this mechanism
to test that the correct !srclocs are passed to the diag handler.

Reviewers: rengolin, rnk, echristo, grosbach, mehdi_amini

Reviewed By: mehdi_amini

Subscribers: mehdi_amini, llvm-commits

Differential Revision: https://reviews.llvm.org/D30167

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296465 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoReformat a blank line.
NAKAMURA Takumi [Tue, 28 Feb 2017 10:15:25 +0000 (10:15 +0000)]
Reformat a blank line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296464 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r296442 (and r296443), "Allow externally dlopen-ed libraries to be registered...
NAKAMURA Takumi [Tue, 28 Feb 2017 10:15:18 +0000 (10:15 +0000)]
Revert r296442 (and r296443), "Allow externally dlopen-ed libraries to be registered as permanent libraries."

It broke clang/test/Analysis/checker-plugins.c

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296463 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Select G_GEP
Diana Picus [Tue, 28 Feb 2017 10:14:38 +0000 (10:14 +0000)]
[ARM] GlobalISel: Select G_GEP

At this point, G_GEP is just an add, so we treat it exactly like a G_ADD.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296462 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] Diagnose PC-writing instructions in IT blocks
Oliver Stannard [Tue, 28 Feb 2017 10:04:36 +0000 (10:04 +0000)]
[ARM] Diagnose PC-writing instructions in IT blocks

In Thumb2, instructions which write to the PC are UNPREDICTABLE if they are in
an IT block but not the last instruction in the block.

Previously, we only diagnosed this for LDM instructions, this patch extends the
diagnostic to cover all of the relevant instructions.

Differential Revision: https://reviews.llvm.org/D30398

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296459 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Add reg bank mapping for G_GEP
Diana Picus [Tue, 28 Feb 2017 09:35:10 +0000 (09:35 +0000)]
[ARM] GlobalISel: Add reg bank mapping for G_GEP

This should be the same as the mapping for G_ADD etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296455 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Legalize G_GEP with 32-bit offsets
Diana Picus [Tue, 28 Feb 2017 09:02:42 +0000 (09:02 +0000)]
[ARM] GlobalISel: Legalize G_GEP with 32-bit offsets

At the moment we're only interested in GEPs for putting call parameters on the
stack, so we'll stick to 32-bit offsets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296452 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRelate the CHECK: lines to the functions that they're checking [NFC]
Artyom Skrobov [Tue, 28 Feb 2017 08:58:40 +0000 (08:58 +0000)]
Relate the CHECK: lines to the functions that they're checking [NFC]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296450 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTest commit, fix typo, NFC.
Vadzim Dambrouski [Tue, 28 Feb 2017 08:27:43 +0000 (08:27 +0000)]
Test commit, fix typo, NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296447 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[IR] Add range accessors for the indices of a GEP instruction.
Chandler Carruth [Tue, 28 Feb 2017 08:04:20 +0000 (08:04 +0000)]
[IR] Add range accessors for the indices of a GEP instruction.

These were noticed as missing in a code review. Add them and the boring
unit test to make sure they compile and DTRT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296444 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix Win bots.
Vassil Vassilev [Tue, 28 Feb 2017 07:26:21 +0000 (07:26 +0000)]
Fix Win bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296443 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAllow externally dlopen-ed libraries to be registered as permanent libraries.
Vassil Vassilev [Tue, 28 Feb 2017 07:11:59 +0000 (07:11 +0000)]
Allow externally dlopen-ed libraries to be registered as permanent libraries.

This is also useful in cases when llvm is in a shared library. First we dlopen
the llvm shared library and then we register it as a permanent library in order
to keep the JIT and other services working.

Patch reviewed by Vedant Kumar (D29955)!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296442 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ImplicitNullCheck] Add alias analysis usage
Sanjoy Das [Tue, 28 Feb 2017 07:04:49 +0000 (07:04 +0000)]
[ImplicitNullCheck] Add alias analysis usage

Summary:
With this change ImplicitNullCheck optimization uses alias analysis
and can use load/store memory access for implicit null check if there
are other load/store before but memory accesses do not alias.

Patch by Serguei Katkov!

Reviewers: sanjoy

Reviewed By: sanjoy

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30331

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296440 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoEmpty line. NFCI
Xin Tong [Tue, 28 Feb 2017 05:30:48 +0000 (05:30 +0000)]
Empty line. NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296438 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LoopUnswitch] Common pushing LIC's user to worklist.
Xin Tong [Tue, 28 Feb 2017 03:32:41 +0000 (03:32 +0000)]
[LoopUnswitch] Common pushing LIC's user to worklist.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296432 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "Add MIR-level outlining pass"
Matthias Braun [Tue, 28 Feb 2017 02:24:30 +0000 (02:24 +0000)]
Revert "Add MIR-level outlining pass"

Revert Machine Outliner for now, as it breaks the asan bot.

This reverts commit r296418.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296426 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoThis script was meant to be committed with the DebugCounter changes.
Daniel Berlin [Tue, 28 Feb 2017 02:19:11 +0000 (02:19 +0000)]
This script was meant to be committed with the DebugCounter changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296425 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[docs] Fix a think-o in the Programmer's Manual.
Lang Hames [Tue, 28 Feb 2017 01:35:31 +0000 (01:35 +0000)]
[docs] Fix a think-o in the Programmer's Manual.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296421 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd test case for usubo combine. NFC.
Amaury Sechet [Tue, 28 Feb 2017 01:16:39 +0000 (01:16 +0000)]
Add test case for usubo combine. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296420 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd MIR-level outlining pass
Matthias Braun [Tue, 28 Feb 2017 00:33:32 +0000 (00:33 +0000)]
Add MIR-level outlining pass

This is a patch for the outliner described in the RFC at:
http://lists.llvm.org/pipermail/llvm-dev/2016-August/104170.html

The outliner is a code-size reduction pass which works by finding
repeated sequences of instructions in a program, and replacing them with
calls to functions. This is useful to people working in low-memory
environments, where sacrificing performance for space is acceptable.

This adds an interprocedural outliner directly before printing assembly.
For reference on how this would work, this patch also includes X86
target hooks and an X86 test.

The outliner is run like so:

clang -mno-red-zone -mllvm -enable-machine-outliner file.c

Patch by Jessica Paquette<jpaquette@apple.com>!

rdar://29166825

Differential Revision: https://reviews.llvm.org/D26872

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296418 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd test case for computing known bits of substraction operations. NFC
Amaury Sechet [Tue, 28 Feb 2017 00:15:13 +0000 (00:15 +0000)]
Add test case for computing known bits of substraction operations. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296417 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[CGP] Split some critical edges coming out of indirect branches
Michael Kuperstein [Tue, 28 Feb 2017 00:11:34 +0000 (00:11 +0000)]
[CGP] Split some critical edges coming out of indirect branches

Splitting critical edges when one of the source edges is an indirectbr
is hard in general (because it requires changing the memory the indirectbr
reads). But if a block only has a single indirectbr predecessor (which is
the common case), we can simulate splitting that edge by splitting
the destination block, and retargeting the *direct* branches.

This is motivated by the use of computed gotos in python 2.7: PyEval_EvalFrame()
ends up using an indirect branch with ~100 successors, and passing a constant to
each of those. Since MachineSink can't break indirect critical edges on demand
(and doing this in MIR doesn't look feasible), this causes us to emit about ~100
defs of registers containing constants, which we in the predecessor block, where
only one of those constants is used in each successor. So, at each computed goto,
we needlessly spill about a 100 constants to stack. The end result is that a
clang-compiled python interpreter can be about ~2.5x slower on a simple python
reduction loop than a gcc-compiled interpreter.

Differential Revision: https://reviews.llvm.org/D29916

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296416 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PDB] Make streams carry their own endianness.
Zachary Turner [Tue, 28 Feb 2017 00:04:07 +0000 (00:04 +0000)]
[PDB] Make streams carry their own endianness.

Before the endianness was specified on each call to read
or write of the StreamReader / StreamWriter, but in practice
it's extremely rare for streams to have data encoded in
multiple different endiannesses, so we should optimize for the
99% use case.

This makes the code cleaner and more general, but otherwise
has NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296415 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DebugInfo] Fix some Clang-tidy modernize and Include What You Use warnings; other...
Eugene Zelenko [Mon, 27 Feb 2017 23:43:14 +0000 (23:43 +0000)]
[DebugInfo] Fix some Clang-tidy modernize and Include What You Use warnings; other minor fixes (NFC).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296413 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Load sorting should not try to sort things that aren't loads.
Michael Kuperstein [Mon, 27 Feb 2017 23:18:11 +0000 (23:18 +0000)]
[SLP] Load sorting should not try to sort things that aren't loads.

We may get a VL where the first element is a load, but the others
aren't. Trying to sort such VLs can only lead to sorrow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296411 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MC] Implement the COFF directives in MCNullStreamer.
Dan Gohman [Mon, 27 Feb 2017 23:10:18 +0000 (23:10 +0000)]
[MC] Implement the COFF directives in MCNullStreamer.

This fixes -filetype=null errors introduced in r296403.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296410 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Basic folds for fmed3 intrinsic
Matt Arsenault [Mon, 27 Feb 2017 23:08:49 +0000 (23:08 +0000)]
AMDGPU: Basic folds for fmed3 intrinsic

Constant fold, canonicalize constants to RHS,
reduce to minnum/maxnum when inputs are nan/undef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296409 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove some code accidentally left in.
Zachary Turner [Mon, 27 Feb 2017 22:57:32 +0000 (22:57 +0000)]
Remove some code accidentally left in.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296407 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AddressSanitizer] Put shadow at 0 for Fuchsia
Petr Hosek [Mon, 27 Feb 2017 22:49:37 +0000 (22:49 +0000)]
[AddressSanitizer] Put shadow at 0 for Fuchsia

The Fuchsia ASan runtime reserves the low part of the address space.

Patch by Roland McGrath

Differential Revision: https://reviews.llvm.org/D30426

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296405 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[CodeGen] Fix some Clang-tidy modernize and Include What You Use warnings; other...
Eugene Zelenko [Mon, 27 Feb 2017 22:45:06 +0000 (22:45 +0000)]
[CodeGen] Fix some Clang-tidy modernize and Include What You Use warnings; other minor fixes (NFC).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296404 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MC] Factor out non-COFF handling of COFF-specific directives.
Dan Gohman [Mon, 27 Feb 2017 22:44:37 +0000 (22:44 +0000)]
[MC] Factor out non-COFF handling of COFF-specific directives.

Instead of requiring every non-COFF MCObjectStreamer to implement the
COFF hooks just to do an llvm_unreachable to say that they're not
supported, do the llvm_unreachable in the default implementation, as
suggested by rnk in https://reviews.llvm.org/D26722.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296403 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WebAssembly] Add some comments and tidy up whitespace.
Dan Gohman [Mon, 27 Feb 2017 22:41:39 +0000 (22:41 +0000)]
[WebAssembly] Add some comments and tidy up whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296402 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Use v_med3_{f16|i16|u16}
Matt Arsenault [Mon, 27 Feb 2017 22:40:39 +0000 (22:40 +0000)]
AMDGPU: Use v_med3_{f16|i16|u16}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296401 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WebAssembly] Split CFG-sorting into its own pass. NFC.
Dan Gohman [Mon, 27 Feb 2017 22:38:58 +0000 (22:38 +0000)]
[WebAssembly] Split CFG-sorting into its own pass. NFC.

CFG sorting was already an independent algorithm from block/loop insertion;
this change makes it more convenient to debug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296399 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r296366 "[InlineFunction] add nonnull assumptions based on argument attributes"
Hans Wennborg [Mon, 27 Feb 2017 22:33:02 +0000 (22:33 +0000)]
Revert r296366 "[InlineFunction] add nonnull assumptions based on argument attributes"

It causes miscompiles e.g. during self-host of Clang (PR32082).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296398 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd missing namespace qualifier.
Zachary Turner [Mon, 27 Feb 2017 22:17:50 +0000 (22:17 +0000)]
Add missing namespace qualifier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296397 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Support v2i16/v2f16 packed operations
Matt Arsenault [Mon, 27 Feb 2017 22:15:25 +0000 (22:15 +0000)]
AMDGPU: Support v2i16/v2f16 packed operations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296396 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoISel: We need to notify FastIS of the IMPLICIT_DEF we created in createSwiftErrorEntr...
Arnold Schwaighofer [Mon, 27 Feb 2017 22:12:06 +0000 (22:12 +0000)]
ISel: We need to notify FastIS of the IMPLICIT_DEF we created in createSwiftErrorEntriesInEntryBlock

Otherwise, it will insert instructions before it.

rdar://30536186

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296395 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PDB] Partial resubmit of r296215, which improved PDB Stream Library.
Zachary Turner [Mon, 27 Feb 2017 22:11:43 +0000 (22:11 +0000)]
[PDB] Partial resubmit of r296215, which improved PDB Stream Library.

This was reverted because it was breaking some builds, and
because of incorrect error code usage.  Since the CL was
large and contained many different things, I'm resubmitting
it in pieces.

This portion is NFC, and consists of:

1) Renaming classes to follow a consistent naming convention.
2) Fixing the const-ness of the interface methods.
3) Adding detailed doxygen comments.
4) Fixing a few instances of passing `const BinaryStream& X`.  These
   are now passed as `BinaryStreamRef X`.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296394 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "DAG: Check if extract_vector_elt is legal or custom"
Matt Arsenault [Mon, 27 Feb 2017 21:59:07 +0000 (21:59 +0000)]
Revert "DAG: Check if extract_vector_elt is legal or custom"

This reverts r295782. This could potentially result in some
legalization loops and I avoided the need for this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296393 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoEmpty line. NFCI
Xin Tong [Mon, 27 Feb 2017 21:51:48 +0000 (21:51 +0000)]
Empty line. NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296392 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PGO] Fix a bug in reading text format value profile.
Rong Xu [Mon, 27 Feb 2017 21:42:39 +0000 (21:42 +0000)]
[PGO] Fix a bug in reading text format value profile.

Summary: Should use the Valuekind read from the profile.

Reviewers: davidxl

Reviewed By: davidxl

Subscribers: llvm-commits, xur

Differential Revision: https://reviews.llvm.org/D30420

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296391 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] don't transform an add(ext Cond), C to select unless there's a setcc of the...
Sanjay Patel [Mon, 27 Feb 2017 21:30:54 +0000 (21:30 +0000)]
[ARM] don't transform an add(ext Cond), C to select unless there's a setcc of the condition

The transform in question claims to be doing:

// fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))

...starting in PerformADDCombineWithOperands(), but it wasn't actually checking for a setcc node
for the sext/zext patterns.

This is exactly the opposite of a transform I'd like to add to DAGCombiner's foldSelectOfConstants(),
so I was seeing infinite loops with my draft of a patch applied.

The changes in select_const.ll look positive (less instructions). The change in arm-and-tst-peephole.ll
is unrelated. We're changing the input IR in that test to preserve the intent of the test, but that's
not affected by this code change.

Differential Revision:
https://reviews.llvm.org/D30355

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296389 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Support][Error] Add a 'cantFail' utility function for known-safe calls to
Lang Hames [Mon, 27 Feb 2017 21:09:47 +0000 (21:09 +0000)]
[Support][Error] Add a 'cantFail' utility function for known-safe calls to
fallible functions.

Some fallible functions (those returning Error or Expected<T>) may only fail
for a subset of their inputs. For example, a "safe" square root function will
succeed for all finite positive inputs:

  Expected<double> safeSqrt(double d) {
    if (d < 0 && !isnan(d) && !isinf(d))
      return make_error<...>("Cannot sqrt -ve values, nans or infs");
    return sqrt(d);
  }

At a safe callsite for such a function, checking the error return value is
redundant:

  if (auto ValOrErr = safeSqrt(42.0)) {
    // use *ValOrErr.
  } else
    llvm_unreachable("safeSqrt should always succeed for +ve values");

The cantFail function wraps this check and extracts the contained value,
simplifying control flow:

  double Result = cantFail(safeSqrt(42.0));

This function should be used with care: it is a programmatic error to wrap a
call with cantFail if it can in fact fail. For debug builds this will
result in llvm_unreachable being called. For release builds the behavior is
undefined.

Use of this function is likely to be rare in library code, but more common
for tool and unit-test code where inputs and mock functions may be known to be
safe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296384 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add some of the new gfx9 VOP3 instructions
Matt Arsenault [Mon, 27 Feb 2017 21:04:41 +0000 (21:04 +0000)]
AMDGPU: Add some of the new gfx9 VOP3 instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296382 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][SSE] Attempt to extract vector elements through target shuffles
Simon Pilgrim [Mon, 27 Feb 2017 21:01:57 +0000 (21:01 +0000)]
[X86][SSE] Attempt to extract vector elements through target shuffles

DAGCombiner already supports peeking thorough shuffles to improve vector element extraction, but legalization often leaves us in situations where we need to extract vector elements after shuffles have already been lowered.

This patch adds support for VECTOR_EXTRACT_ELEMENT/PEXTRW/PEXTRB instructions to attempt to handle target shuffles as well. I've covered some basic scenarios including handling shuffle mask scaling and the implicit zero-extension of PEXTRW/PEXTRB, there is more that could be done here (that I've mentioned in TODOs) but I haven't found many cases where its worth it.

Differential Revision: https://reviews.llvm.org/D30176

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296381 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Support inlineasm for packed instructions
Matt Arsenault [Mon, 27 Feb 2017 20:52:10 +0000 (20:52 +0000)]
AMDGPU: Support inlineasm for packed instructions

Add packed types as legal so they may be used with inlineasm.
Keep all operations expanded for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296379 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Use different flags in tests for reduction ops and extra args.
Alexey Bataev [Mon, 27 Feb 2017 20:22:44 +0000 (20:22 +0000)]
[SLP] Use different flags in tests for reduction ops and extra args.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296376 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Don't fold immediate if clamp/omod are set
Matt Arsenault [Mon, 27 Feb 2017 20:21:31 +0000 (20:21 +0000)]
AMDGPU: Don't fold immediate if clamp/omod are set

Doesn't fix any practical problems because clamp/omod
are currently folded after peephole optimizer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296375 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Fold omod into instructions
Matt Arsenault [Mon, 27 Feb 2017 19:35:42 +0000 (19:35 +0000)]
AMDGPU: Fold omod into instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296372 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[TailDuplicator] Maintain DebugLoc for branch instructions
Taewook Oh [Mon, 27 Feb 2017 19:30:01 +0000 (19:30 +0000)]
[TailDuplicator] Maintain DebugLoc for branch instructions

Summary: Existing implementation of duplicateSimpleBB function drops DebugLoc metadata of branch instructions during the transformation. This patch addresses this issue by making newly created branch instructions to keep the metadata of replaced branch instructions.

Reviewers: qcolombet, craig.topper, aprantl, MatzeB, sanjoy, dblaikie

Reviewed By: dblaikie

Subscribers: dblaikie, llvm-commits

Differential Revision: https://reviews.llvm.org/D30026

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296371 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add f16 to shader calling conventions
Matt Arsenault [Mon, 27 Feb 2017 19:24:47 +0000 (19:24 +0000)]
AMDGPU: Add f16 to shader calling conventions

Mostly useful for writing tests for f16 features.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296370 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Modify test to check IR flags propagation for extra args.
Alexey Bataev [Mon, 27 Feb 2017 19:16:09 +0000 (19:16 +0000)]
[SLP] Modify test to check IR flags propagation for extra args.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296369 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add VOP3P instruction format
Matt Arsenault [Mon, 27 Feb 2017 18:49:11 +0000 (18:49 +0000)]
AMDGPU: Add VOP3P instruction format

Add a few non-VOP3P but instructions related to packed.

Includes hack with dummy operands for the benefit of the assembler

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296368 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRefactor xaluo.ll and xmulo.ll tests. NFC
Amaury Sechet [Mon, 27 Feb 2017 18:32:54 +0000 (18:32 +0000)]
Refactor xaluo.ll and xmulo.ll tests. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296367 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InlineFunction] add nonnull assumptions based on argument attributes
Sanjay Patel [Mon, 27 Feb 2017 18:13:48 +0000 (18:13 +0000)]
[InlineFunction] add nonnull assumptions based on argument attributes

This was suggested in D27855: have the inliner add assumptions, so we don't
lose nonnull info provided by argument attributes.

This still doesn't solve PR28430 (dyn_cast), but this gets us closer.

https://reviews.llvm.org/D29999

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296366 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Defs and clobbers can overlap
Krzysztof Parzyszek [Mon, 27 Feb 2017 18:03:35 +0000 (18:03 +0000)]
[Hexagon] Defs and clobbers can overlap

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296365 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix a bug when unswitching on partial LIV for SwitchInst
Xin Tong [Mon, 27 Feb 2017 18:00:13 +0000 (18:00 +0000)]
Fix a bug when unswitching on partial LIV for SwitchInst

Summary: Fix a bug when unswitching on partial LIV for SwitchInst.

Reviewers: hfinkel, efriedma, sanjoy

Reviewed By: sanjoy

Subscribers: david2050, mzolotukhin, llvm-commits

Differential Revision: https://reviews.llvm.org/D29107

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296363 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix comments. NFC.
Rong Xu [Mon, 27 Feb 2017 17:59:01 +0000 (17:59 +0000)]
Fix comments. NFC.

Change "Thin-LTO" to "ThinLTO" in the comments for consistency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296362 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix LLVM module build
Steven Wu [Mon, 27 Feb 2017 16:56:37 +0000 (16:56 +0000)]
Fix LLVM module build

Add WasmRelocs/WebAssembly.def to textual include header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296356 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86] Use APInt instead of SmallBitVector tracking undef elements from getTargetConst...
Craig Topper [Mon, 27 Feb 2017 16:15:32 +0000 (16:15 +0000)]
[X86] Use APInt instead of SmallBitVector tracking undef elements from getTargetConstantBitsFromNode and getConstVector.

Summary:
SmallBitVector uses a malloc for more than 58 bits on a 64-bit target and more than 27 bits on a 32-bit target. Some of the vector types we deal with here use more than those number of elements and therefore cause a malloc.

APInt on the other hand supports up to 64 bits without a malloc. That's the maximum number of bits we need here so we can avoid a malloc for all cases by using APInt.

Reviewers: RKSimon

Reviewed By: RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30392

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296355 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86] Use APInt instead of SmallBitVector for tracking Zeroable elements in shuffle...
Craig Topper [Mon, 27 Feb 2017 16:15:30 +0000 (16:15 +0000)]
[X86] Use APInt instead of SmallBitVector for tracking Zeroable elements in shuffle lowering

Summary:
SmallBitVector uses a malloc for more than 58 bits on a 64-bit target and more than 27 bits on a 32-bit target. Some of the vector types we deal with here use more than those number of elements and therefore cause a malloc.

APInt on the other hand supports up to 64 bits without a malloc. That's the maximum number of bits we need here so we can avoid a malloc for all cases by using APInt.

Reviewers: RKSimon

Reviewed By: RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30390

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296354 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86] Fix SmallVector sizes in constant pool shuffle decoding to avoid heap allocation
Craig Topper [Mon, 27 Feb 2017 16:15:27 +0000 (16:15 +0000)]
[X86] Fix SmallVector sizes in constant pool shuffle decoding to avoid heap allocation

Some of the vectors are under sized to avoid heap allocation. In one case the vector was oversized.

Differential Revision: https://reviews.llvm.org/D30387

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296353 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86] Use APInt instead of SmallBitVector for tracking undef elements in constant...
Craig Topper [Mon, 27 Feb 2017 16:15:25 +0000 (16:15 +0000)]
[X86] Use APInt instead of SmallBitVector for tracking undef elements in constant pool shuffle decoding

Summary:
SmallBitVector uses a malloc for more than 58 bits on a 64-bit target and more than 27 bits on a 32-bit target. Some of the vector types we deal with here use more than those number of elements and therefore cause a malloc.

APInt on the other hand supports up to 64 bits without a malloc. That's the maximum number of bits we need here so we can avoid a malloc for all cases by using APInt. This will incur a minor increase in stack usage due to APInt storing the bit count separately from the data bits unlike SmallBitVector, but that should be ok.

Reviewers: RKSimon

Reviewed By: RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30386

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296352 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove an empty line in icmp-illegal.ll . NFC
Amaury Sechet [Mon, 27 Feb 2017 16:09:44 +0000 (16:09 +0000)]
Remove an empty line in icmp-illegal.ll . NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296350 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] A test for a fix of PR32038.
Alexey Bataev [Mon, 27 Feb 2017 16:07:10 +0000 (16:07 +0000)]
[SLP] A test for a fix of PR32038.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296349 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLoop predication expand both sides of the widened condition
Artur Pilipenko [Mon, 27 Feb 2017 15:44:49 +0000 (15:44 +0000)]
Loop predication expand both sides of the widened condition

This is a fix for a loop predication bug which resulted in malformed IR generation.

Loop invariant side of the widened condition is not guaranteed to be available in the preheader as is, so we need to expand it as well. See added unsigned_loop_0_to_n_hoist_length test for example.

Reviewed By: sanjoy, mkazantsev

Differential Revision: https://reviews.llvm.org/D30099

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296345 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAArch64InstPrinter: rewrite of printSysAlias
Sjoerd Meijer [Mon, 27 Feb 2017 14:45:34 +0000 (14:45 +0000)]
AArch64InstPrinter: rewrite of printSysAlias

This is a cleanup/rewrite of the printSysAlias function. This was not using the
tablegen instruction descriptions, but was "manually" decoding the
instructions. This has been replaced with calls to lookup_XYZ_ByEncoding
tablegen calls.

This revealed several problems. First, instruction IVAU had the wrong encoding.
This was cancelled out by the parser that incorrectly matched the wrong
encoding. Second, instruction CVAP was missing from the SystemOperands tablegen
descriptions, so this has been added. And third, the required target features
were not captured in the tablegen descriptions, so support for this has also
been added.

Differential Revision: https://reviews.llvm.org/D30329

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296343 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] LSL #0 is an alias of MOV
John Brawn [Mon, 27 Feb 2017 14:40:51 +0000 (14:40 +0000)]
[ARM] LSL #0 is an alias of MOV

Currently we handle this correctly in arm, but in thumb we don't which leads to
an unpredictable instruction being emitted for LSL #0 in an IT block and SP not
being permitted in some cases when it should be.

For the thumb2 LSL we can handle this by making LSL #0 an alias of MOV in the
.td file, but for thumb1 we need to handle it in checkTargetMatchPredicate to
get the IT handling right. We also need to adjust the handling of
MOV rd, rn, LSL #0 to avoid generating the 16-bit encoding in an IT block. We
should also adjust it to allow SP in the same way that it is allowed in
MOV rd, rn, but I haven't done that here because it looks like it would take
quite a lot of work to get right.

Additionally correct the selection of the 16-bit shift instructions in
processInstruction, where it was checking if the two registers were equal when
it should have been checking if they were low. It appears that previously this
code was never executed and the 16-bit encoding was selected by default, but
the other changes I've done here have somehow made it start being used.

Differential Revision: https://reviews.llvm.org/D30294

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296342 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAGCombine] Fix for a load combine bug with non-zero offset patterns on BE targets
Artur Pilipenko [Mon, 27 Feb 2017 13:04:23 +0000 (13:04 +0000)]
[DAGCombine] Fix for a load combine bug with non-zero offset patterns on BE targets

This pattern is essentially a i16 load from p+1 address:

  %p1.i16 = bitcast i8* %p to i16*
  %p2.i8 = getelementptr i8, i8* %p, i64 2
  %v1 = load i16, i16* %p1.i16
  %v2.i8 = load i8, i8* %p2.i8
  %v2 = zext i8 %v2.i8 to i16
  %v1.shl = shl i16 %v1, 8
  %res = or i16 %v1.shl, %v2

Current implementation would identify %v1 load as the first byte load and would mistakenly emit a i16 load from %p1.i16 address. This patch adds a check that the first byte is loaded from a non-zero offset of the first load address. This way this address can be used as the base address for the combined value. Otherwise just give up combining.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296336 91177308-0d34-0410-b5e6-96231b3b80d8