oota-llvm.git
9 years agoEmit correct linkage-name attribute based on DWARF version.
Paul Robinson [Tue, 10 Mar 2015 22:44:45 +0000 (22:44 +0000)]
Emit correct linkage-name attribute based on DWARF version.

There are still 4 tests that check for DW_AT_MIPS_linkage_name,
because they specify DWARF 2 or 3 in the module metadata. So, I didn't
create an explicit version-based test for the attribute.

Differential Revision: http://reviews.llvm.org/D8227

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231880 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInfer known bits from dominating conditions
Philip Reames [Tue, 10 Mar 2015 22:43:20 +0000 (22:43 +0000)]
Infer known bits from dominating conditions

This patch adds limited support in ValueTracking for inferring known bits of a value from conditional expressions which must be true to reach the instruction we're trying to optimize. At this time, the feature is off by default. Once landed, I'm hoping for feedback from others on both profitability and compile time impact.

Forms of conditional value propagation have been tried in LLVM before and have failed due to compile time problems.  In an attempt to side step that, this patch only considers conditions where the edge leaving the branch dominates the context instruction. It does not attempt full dataflow.  Even with that restriction, it handles many interesting cases:
 * Early exits from functions
 * Early exits from loops (for context instructions in the loop and after the check)
 * Conditions which control entry into loops, including multi-version loops (such as those produced during vectorization, IRCE, loop unswitch, etc..)

Possible applications include optimizing using information provided by constructs such as: preconditions, assumptions, null checks, & range checks.

This patch implements two approaches to the problem that need further benchmarking.  Approach 1 is to directly walk the dominator tree looking for interesting conditions.  Approach 2 is to inspect other uses of the value being queried for interesting comparisons.  From initial benchmarking, it appears that Approach 2 is faster than Approach 1, but this needs to be further validated.

Differential Revision: http://reviews.llvm.org/D7708

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231879 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove the use of the subtarget in MCCodeEmitter creation and
Eric Christopher [Tue, 10 Mar 2015 22:03:14 +0000 (22:03 +0000)]
Remove the use of the subtarget in MCCodeEmitter creation and
update all ports accordingly. Required a couple of small rewrites
in handling subtarget features during creation in PPC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231861 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCreate symbols marking the start of a section earlier.
Rafael Espindola [Tue, 10 Mar 2015 22:00:25 +0000 (22:00 +0000)]
Create symbols marking the start of a section earlier.

This lets us pass the symbol to the constructor and avoid the mutable field.

This also opens the way for outputting the symbol only when needed, instead
of outputting them at the start of the file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231859 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove createAMDGPUMCCodeEmitter and instead just register the correct
Eric Christopher [Tue, 10 Mar 2015 21:57:34 +0000 (21:57 +0000)]
Remove createAMDGPUMCCodeEmitter and instead just register the correct
MCCodeEmitter creation routine based on TargetMachine since the only
64-bit R600 gpus are part of the GCN target.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231856 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[CodeGenPrepare] Refine the cost model provided by the promotion helper.
Quentin Colombet [Tue, 10 Mar 2015 21:48:15 +0000 (21:48 +0000)]
[CodeGenPrepare] Refine the cost model provided by the promotion helper.
- Use TargetLowering to check for the actual cost of each extension.
- Provide a factorized method to check for the cost of an extension:
  TargetLowering::isExtFree.
- Provide a virtual method TargetLowering::isExtFreeImpl for targets to be able
  to tune the cost of non-free extensions.

This refactoring offers a better granularity to model what really happens on
different targets.

No performance changes and very few code differences.

Part of <rdar://problem/19267165>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231855 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LoopAccesses] Add debug message to indicate the result of the analysis
Adam Nemet [Tue, 10 Mar 2015 21:47:39 +0000 (21:47 +0000)]
[LoopAccesses] Add debug message to indicate the result of the analysis

The debug message was pretty confusing here.  It only reported the
situation with memchecks without the result of the dependence analysis.

Now it prints whether the loop is safe from the POV of the dependence
analysis and if yes, whether we need memchecks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231854 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMove a non-trivial virtual function out of line.
Rafael Espindola [Tue, 10 Mar 2015 21:35:16 +0000 (21:35 +0000)]
Move a non-trivial virtual function out of line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231853 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding frame index + add load/store patterns.
Colin LeMahieu [Tue, 10 Mar 2015 21:24:13 +0000 (21:24 +0000)]
[Hexagon] Adding frame index + add load/store patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231850 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoclang-format code that is about to change.
Rafael Espindola [Tue, 10 Mar 2015 21:16:18 +0000 (21:16 +0000)]
clang-format code that is about to change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231848 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Simplifying deallocret definitions.
Colin LeMahieu [Tue, 10 Mar 2015 21:12:32 +0000 (21:12 +0000)]
[Hexagon] Simplifying deallocret definitions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231847 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoclang-format these declarations. NFC.
Rafael Espindola [Tue, 10 Mar 2015 21:05:09 +0000 (21:05 +0000)]
clang-format these declarations. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231846 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDon't repeat names in comments. NFC.
Rafael Espindola [Tue, 10 Mar 2015 21:01:50 +0000 (21:01 +0000)]
Don't repeat names in comments. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231845 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Separating InstHexagon from OpcodeHexagon.
Colin LeMahieu [Tue, 10 Mar 2015 20:56:22 +0000 (20:56 +0000)]
[Hexagon] Separating InstHexagon from OpcodeHexagon.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231844 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd support for part-word atomics for PPC
Nemanja Ivanovic [Tue, 10 Mar 2015 20:51:07 +0000 (20:51 +0000)]
Add support for part-word atomics for PPC
http://reviews.llvm.org/D8090#inline-67337

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231843 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd new LLVM_OPTIMIZED_TABLEGEN build setting which configures, builds and uses a...
Chris Bieneman [Tue, 10 Mar 2015 20:48:02 +0000 (20:48 +0000)]
Add new LLVM_OPTIMIZED_TABLEGEN build setting which configures, builds and uses a release tablegen build when LLVM is configured with assertions enabled.

Summary: This change leverages the cross-compiling functionality in the build system to build a release tablegen executable for use during the build.

Reviewers: resistor, rnk

Reviewed By: rnk

Subscribers: rnk, joker.eph, llvm-commits

Differential Revision: http://reviews.llvm.org/D7349

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231842 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AArch64] Avoid going through GPRs for across-vector instructions.
Ahmed Bougacha [Tue, 10 Mar 2015 20:45:38 +0000 (20:45 +0000)]
[AArch64] Avoid going through GPRs for across-vector instructions.

This adds new node types for each intrinsic.
For instance, for addv, we have AArch64ISD::UADDV, such that:
  (v4i32 (uaddv ...))
is the same as
  (v4i32 (scalar_to_vector (i32 (int_aarch64_neon_uaddv ...))))
that is,
  (v4i32 (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)),
           (i32 (int_aarch64_neon_uaddv ...)), ssub)

In a combine, we transform all such across-vector-lanes intrinsics to:

  (i32 (extract_vector_elt (uaddv ...), 0))

This has one big advantage: by making the extract_element explicit, we
enable the existing patterns for lane-aware instructions to fire.
This lets us avoid needlessly going through the GPRs.  Consider:

    uint32x4_t test_mul(uint32x4_t a, uint32x4_t b) {
        return vmulq_n_u32(a, vaddvq_u32(b));
    }

We now generate:
    addv.4s  s1, v1
    mul.4s   v0, v0, v1[0]
instead of the previous:
    addv.4s  s1, v1
    fmov     w8, s1
    dup.4s   v1, w8
    mul.4s   v0, v1, v0

rdar://20044838

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231840 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AArch64] Remove integer INSvi*lane patterns. NFCI.
Ahmed Bougacha [Tue, 10 Mar 2015 20:37:19 +0000 (20:37 +0000)]
[AArch64] Remove integer INSvi*lane patterns.  NFCI.

Most are redundant, and they never seem to fire.

The V128 integer patterns already exist in the INS multiclass.
The duplicates only fire when the vector index type isn't i64,
because they accept "imm" instead of an explicit "i64", as the
instruction definition patterns do.

TLI::getVectorIdxTy is i64 on AArch64, so this should never happen.
Also, one of them had a typo: for i64, INSvi32lane was used.
I noticed because I mistakenly used an explicit i32 as the idx type,
and got ins.s for an i64 vector_insert.

The V64 patterns also don't seem to ever fire, as V64 vector
extract/insert are legalized to V128.

The equivalent float patterns are unique and useful, so keep them.

No functional change intended;  none exhibited on the LIT and LNT tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231838 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDon't evaluate rend() on every iteration of the loop.
Chad Rosier [Tue, 10 Mar 2015 20:29:59 +0000 (20:29 +0000)]
Don't evaluate rend() on every iteration of the loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231837 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLoopAccessAnalysis: Silence -Wreturn-type diagnostic from GCC
David Majnemer [Tue, 10 Mar 2015 20:23:29 +0000 (20:23 +0000)]
LoopAccessAnalysis: Silence -Wreturn-type diagnostic from GCC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231836 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDon't use LLVM_LIBRARY_VISIBILITY in cpp files.
Benjamin Kramer [Tue, 10 Mar 2015 20:07:44 +0000 (20:07 +0000)]
Don't use LLVM_LIBRARY_VISIBILITY in cpp files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231831 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AsmPrinter][TLOF] Reintroduce AArch64 test
Bruno Cardoso Lopes [Tue, 10 Mar 2015 20:05:23 +0000 (20:05 +0000)]
[AsmPrinter][TLOF] Reintroduce AArch64 test

Follow up from r231505.

Fix the non-determinism by using a MapVector and reintroduce the AArch64
testcase. Defer deleting the got candidates up to the end and remove
them in a bulk, avoiding linear time removal of each element.

Thanks to Renato Golin for trying it out on other platforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231830 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding nodes for PIC support.
Colin LeMahieu [Tue, 10 Mar 2015 20:04:44 +0000 (20:04 +0000)]
[Hexagon] Adding nodes for PIC support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231829 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding DuplexInst instruction format and duplex class defs.
Colin LeMahieu [Tue, 10 Mar 2015 19:53:14 +0000 (19:53 +0000)]
[Hexagon] Adding DuplexInst instruction format and duplex class defs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231828 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoChange the generation of the vmuluwm instruction to be based on the MUL opcode.
Kit Barton [Tue, 10 Mar 2015 19:49:38 +0000 (19:49 +0000)]
Change the generation of the vmuluwm instruction to be based on the MUL opcode.

Phabricator review: http://reviews.llvm.org/D8185

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231827 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoremove function names from comments; NFC
Sanjay Patel [Tue, 10 Mar 2015 19:42:57 +0000 (19:42 +0000)]
remove function names from comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231826 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding nodes for vector insert/extract lowering.
Colin LeMahieu [Tue, 10 Mar 2015 19:40:03 +0000 (19:40 +0000)]
[Hexagon] Adding nodes for vector insert/extract lowering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231825 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Renaming HexagonJT to JT and adding CP for constantpool.
Colin LeMahieu [Tue, 10 Mar 2015 19:29:53 +0000 (19:29 +0000)]
[Hexagon] Renaming HexagonJT to JT and adding CP for constantpool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231824 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoChange the datatype of DwarfExpression::Emit(Un)Signed to (u)int64_t
Adrian Prantl [Tue, 10 Mar 2015 19:23:37 +0000 (19:23 +0000)]
Change the datatype of DwarfExpression::Emit(Un)Signed to (u)int64_t
so it matches the one used by ByteStreamer::Emit(U|S)LEB128.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231823 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoNVPTX: move NVPTXAllocaHoisting into the cpp file
Benjamin Kramer [Tue, 10 Mar 2015 19:20:52 +0000 (19:20 +0000)]
NVPTX: move NVPTXAllocaHoisting into the cpp file

Also initialize without using static initialization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231822 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LAA-memchecks] Comment improvement
Adam Nemet [Tue, 10 Mar 2015 19:12:41 +0000 (19:12 +0000)]
[LAA-memchecks] Comment improvement

I forgot to roll this into r231816.  It was requested by Hal in D8122.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231821 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoEnable loop-rotate before loop-vectorize by default
Michael Zolotukhin [Tue, 10 Mar 2015 19:07:41 +0000 (19:07 +0000)]
Enable loop-rotate before loop-vectorize by default

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231820 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LAA-memchecks 3/3] Introduce pointer partitions for memchecks
Adam Nemet [Tue, 10 Mar 2015 18:54:26 +0000 (18:54 +0000)]
[LAA-memchecks 3/3] Introduce pointer partitions for memchecks

This is the final patch that actually introduces the new parameter of
partition mapping to RuntimePointerCheck::needsChecking.

Another API (LAI::getInstructionsForAccess) is also exposed that helps
to map pointers to instructions because ultimately we partition
instructions.

The WIP version of the Loop Distribution pass in D6930 has been adapted
to use all this.  See for example, how
InstrPartitionContainer::computePartitionSetForPointers sets up the
partitions using the above API and then calls to LAI::addRuntimeCheck
with the pointer partitions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231818 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LAA-memchecks 2/3] Move number of memcheck threshold checking to LV
Adam Nemet [Tue, 10 Mar 2015 18:54:23 +0000 (18:54 +0000)]
[LAA-memchecks 2/3] Move number of memcheck threshold checking to LV

Now the analysis won't "fail" if the memchecks exceed the threshold.  It
is the transform pass' responsibility to perform the check.

This allows the transform pass to further analyze/eliminate the
memchecks.  E.g. in Loop distribution we only need to check pointers
that end up in different partitions.

Note that there is a slight change of functionality here.  The logic in
analyzeLoop is that if dependence checking fails due to non-constant
distance between the pointers, another attempt is made to prove safety
of the dependences purely using run-time checks.

Before this patch we could fail the loop due to exceeding the memcheck
threshold after the first step, now we only check the threshold in the
client after the full analysis.  There is no measurable compile-time
effect but I wanted to record this here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231817 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LAA-memchecks 1/3] Split out NumComparisons checks. NFC
Adam Nemet [Tue, 10 Mar 2015 18:54:19 +0000 (18:54 +0000)]
[LAA-memchecks 1/3] Split out NumComparisons checks. NFC

The check for the number of memchecks will be moved to the client of
this analysis.  Besides allowing for transform-specific thresholds, this
also lets Loop Distribution post-process the memchecks; Loop
Distribution only needs memchecks between pointers of different
partitions.

The motivation for this first patch is to untangle the CanDoRT check
from the NumComparison check before moving the NumComparison part.
CanDoRT means that we couldn't determine the bounds for the pointer.
Note that NumComparison is set independent of this flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231816 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoremove names from comments; NFC
Sanjay Patel [Tue, 10 Mar 2015 18:41:22 +0000 (18:41 +0000)]
remove names from comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231813 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agofix typos; NFC
Sanjay Patel [Tue, 10 Mar 2015 18:37:05 +0000 (18:37 +0000)]
fix typos; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231812 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoNVPTX: Remove copy of LLVMInitializeNVPTXAsmPrinter.
Benjamin Kramer [Tue, 10 Mar 2015 18:19:24 +0000 (18:19 +0000)]
NVPTX: Remove copy of LLVMInitializeNVPTXAsmPrinter.

If anyone is using this for some strange reason,
LLVMInitializeNVPTXAsmPrinter does exactly the same thing and is what
other LLVM tools are calling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231810 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoHexagon: Remove unused InstrMapping.
Benjamin Kramer [Tue, 10 Mar 2015 18:19:16 +0000 (18:19 +0000)]
Hexagon: Remove unused InstrMapping.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231809 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LoopAccesses 3/3] Print the dependences with -analyze
Adam Nemet [Tue, 10 Mar 2015 17:40:43 +0000 (17:40 +0000)]
[LoopAccesses 3/3] Print the dependences with -analyze

The dependences are now expose through the new getInterestingDependences
API so we can use that with -analyze too and fix the FIXME.

This lets us remove the test that relied on -debug to check the
dependences.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231807 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LoopAccesses 2/3] Allow querying of interesting dependences
Adam Nemet [Tue, 10 Mar 2015 17:40:37 +0000 (17:40 +0000)]
[LoopAccesses 2/3] Allow querying of interesting dependences

Gather an array of interesting dependences rather than just failing
after the first unsafe one and regarding the loop unsafe.  Loop
Distribution needs to be able to collect all dependences in order to
isolate the dependence cycles into their own partition.

Since the dependence checking algorithm is quadratic in terms of
accesses sharing the same underlying pointer, I am applying a cut-off
threshold (MaxInterestingDependence).  Exceeding that, the logic reverts
back to the original approach deeming the loop unsafe upon encountering
the first unsafe dependence.

The main idea of the patch is to split isDepedent from directly
answering the question whether the dep is safe for vectorization to
return a dependence type which then gets mapped to old boolean result
using Dependence::isSafeForVectorization.

Tested that this was compile-time neutral on SpecINT2006 LTO bitcode
inputs.  No assembly change on the testsuite including external.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231806 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LoopAccesses 1/3] Expose MemoryDepChecker to LAA users
Adam Nemet [Tue, 10 Mar 2015 17:40:34 +0000 (17:40 +0000)]
[LoopAccesses 1/3] Expose MemoryDepChecker to LAA users

LoopDistribution needs to query various results of the dependence
analysis.  This series will expose some more APIs and state of the
dependence checker.

This patch is a simple one to just expose the DepChecker instance.  The
set is compile-time neutral measured with LTO bitcode files of
SpecINT2006.  Also there is no assembly change on the testsuite.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231805 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoStore an optional section start label in MCSection.
Rafael Espindola [Tue, 10 Mar 2015 16:58:10 +0000 (16:58 +0000)]
Store an optional section start label in MCSection.

This makes code that uses section relative expressions (debug info) simpler and
less brittle.

This is still a bit awkward as the symbol is created late and has to be
stored in a mutable field.

I will move the symbol creation earlier in the next patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231802 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoremove function names from comments; NFC
Sanjay Patel [Tue, 10 Mar 2015 16:42:24 +0000 (16:42 +0000)]
remove function names from comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231801 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTeach lowering to correctly handle invoke statepoint and gc results tied to them...
Igor Laevsky [Tue, 10 Mar 2015 16:26:48 +0000 (16:26 +0000)]
Teach lowering to correctly handle invoke statepoint and gc results tied to them. Note that we still can not lower gc.relocates for invoke statepoints.
Also it extracts getCopyFromRegs helper function in SelectionDAGBuilder as we need to be able to customize type of the register exported from basic block during lowering of the gc.result.
(Resubmitting this change after not being able to reproduce buildbot failure)

Differential Revision: http://reviews.llvm.org/D7760

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231800 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[BranchFolding] Remove MMOs during tail merge to preserve dependencies.
Chad Rosier [Tue, 10 Mar 2015 16:22:52 +0000 (16:22 +0000)]
[BranchFolding] Remove MMOs during tail merge to preserve dependencies.

When tail merging it may be necessary to remove MMOs from memory operations to
ensures later passes (e.g., MI sched) conservatively compute dependencies.
Currently, we only remove the MMO from the common tail if the MMO doesn't match
with the relative instruction in the non-common tail(s).

A more robust solution would be to add multiple MMOs from the duplicate MIs to
the new MI. Currently ScheduleDAGInstrs.cpp ignores all MMOs on instructions
with multiple MMOs, so this solution is equivalent for the time being.

No test case included as this is incredibly difficult to reproduce.

Patch was a collaborative effort between Ana Pazos and myself.
Phabricator: http://reviews.llvm.org/D7769

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231799 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Add _IDXEN and _BOTHEN variants for buffer_store
Tom Stellard [Tue, 10 Mar 2015 16:16:51 +0000 (16:16 +0000)]
R600/SI: Add _IDXEN and _BOTHEN variants for buffer_store

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231798 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Re-order MUBUF operands to match asm strings.
Tom Stellard [Tue, 10 Mar 2015 16:16:49 +0000 (16:16 +0000)]
R600/SI: Re-order MUBUF operands to match asm strings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231797 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Move kill flag to second instruction when splitting SMRD
Tom Stellard [Tue, 10 Mar 2015 16:16:48 +0000 (16:16 +0000)]
R600/SI: Move kill flag to second instruction when splitting SMRD

This fixes a machine verifier error in the salu-to-valu.ll, which
would have been exposed by a future commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231796 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Add 32-bit encoding of v_cndmask_b32
Tom Stellard [Tue, 10 Mar 2015 16:16:44 +0000 (16:16 +0000)]
R600/SI: Add 32-bit encoding of v_cndmask_b32

This was done by refactoring the v_cndmask_b32 tablegen definition
to use inherit from VOP2Inst.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231795 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[X86, AVX] replace vinsertf128 intrinsics with generic shuffles
Sanjay Patel [Tue, 10 Mar 2015 16:08:36 +0000 (16:08 +0000)]
[X86, AVX] replace vinsertf128 intrinsics with generic shuffles

We want to replace as much custom x86 shuffling via intrinsics
as possible because pushing the code down the generic shuffle
optimization path allows for better codegen and less complexity
in LLVM.

This is the sibling patch for the Clang half of this change:
http://reviews.llvm.org/D8088

Differential Revision: http://reviews.llvm.org/D8086

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231794 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoHexagon: Remove pass that does nothing at all
Benjamin Kramer [Tue, 10 Mar 2015 15:06:38 +0000 (15:06 +0000)]
Hexagon: Remove pass that does nothing at all

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231791 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove effectively dead code.
Rafael Espindola [Tue, 10 Mar 2015 14:48:01 +0000 (14:48 +0000)]
Remove effectively dead code.

Switching back and forth between sections does nothing (other than producing
larger .s files).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231790 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a memory corruption in Dependency Analysis.
Karthik Bhat [Tue, 10 Mar 2015 14:32:02 +0000 (14:32 +0000)]
Fix a memory corruption in Dependency Analysis.
This crash occurs due to memory corruption when trying to update dependency
direction based on Constraints.

This crash was observed during lnt regression of Polybench benchmark test case dynprog.

Review: http://reviews.llvm.org/D8059

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231788 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDon't repeat names and clang-format this file.
Rafael Espindola [Tue, 10 Mar 2015 13:56:44 +0000 (13:56 +0000)]
Don't repeat names and clang-format this file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231786 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemoving dead code to silence warning C4060: switch statement contains no 'case'...
Aaron Ballman [Tue, 10 Mar 2015 13:56:28 +0000 (13:56 +0000)]
Removing dead code to silence warning C4060: switch statement contains no 'case' or 'default' labels; NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231785 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a crash in Dependency Analysis.
Karthik Bhat [Tue, 10 Mar 2015 13:31:03 +0000 (13:31 +0000)]
Fix a crash in Dependency Analysis.
This crash in Dependency analysis is because we assume here that in case of UsefulGEP
both source and destination have the same number of operands which may not be true.
This incorrect assumption results in crash while populating Pairs. Fix the same.

This crash was observed during lnt regression for code such as-
  struct s{
    int A[10][10];
    int C[10][10][10];
  } S;
  void dep_constraint_crash_test(int k,int N)  {
     for( int i=0;i<N;i++)
       for( int j=0;j<N;j++)
         S.A[0][0] = S.C[0][0][k];
  }
Review: http://reviews.llvm.org/D8162

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231784 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoThe operand flag word used in ISD::INLINEASM is an i32 not a pointer. NFC.
Daniel Sanders [Tue, 10 Mar 2015 10:42:59 +0000 (10:42 +0000)]
The operand flag word used in ISD::INLINEASM is an i32 not a pointer. NFC.

Summary:
This is part of the work to support memory constraints that behave
differently to 'm'. The subsequent patches will expand on the existing
encoding (which is a 32-bit int) and as a result in some flag words will no
longer fit into an i16. This problem only affected the MSP430 target which
appears to have 16-bit pointers.

Reviewers: hfinkel

Reviewed By: hfinkel

Subscribers: hfinkel, llvm-commits

Differential Revision: http://reviews.llvm.org/D8168

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231783 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTeach raw_ostream to accept SmallString.
Yaron Keren [Tue, 10 Mar 2015 07:33:23 +0000 (07:33 +0000)]
Teach raw_ostream to accept SmallString.

Saves adding .str() call to any raw_ostream << SmallString usage
and a small step towards making .str() consistent in the ADTs by
removing one of the SmallString::str() use cases, discussion at

http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20141013/240026.html

I'll update the Phabricator patch http://reviews.llvm.org/D6372
for review of the Twine SmallString support, it's more complex
than this one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231763 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a crash in InstCombine where we could try to truncate a switch comparison to...
Owen Anderson [Tue, 10 Mar 2015 06:51:39 +0000 (06:51 +0000)]
Fix a crash in InstCombine where we could try to truncate a switch comparison to zero width.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231761 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a stack overflow in the assembler when checking that GEPs must be over sized...
Owen Anderson [Tue, 10 Mar 2015 06:34:57 +0000 (06:34 +0000)]
Fix a stack overflow in the assembler when checking that GEPs must be over sized types.

We failed to use a marking set to properly handle recursive types, which caused use
to recurse infinitely and eventually overflow the stack.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231760 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix an issue in the verifier where we could try to read information out of a malforme...
Owen Anderson [Tue, 10 Mar 2015 05:58:21 +0000 (05:58 +0000)]
Fix an issue in the verifier where we could try to read information out of a malformed statepoint intrinsic.

In this situation we would always have already flagged an error on the statepoint intrinsic,
but then we carry on to parse other, related GC intrinsics, and could end up crashing during that
verification when they try to access data from the malformed statepoint.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231759 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix an infinite loop in InstCombine when an instruction with no users and side effect...
Owen Anderson [Tue, 10 Mar 2015 05:13:47 +0000 (05:13 +0000)]
Fix an infinite loop in InstCombine when an instruction with no users and side effects can be constant folded.

ReplaceInstUsesWith needs to return nullptr when the input has no users,
because in that case it does not mutate the program.  Otherwise, we can
get stuck in an infinite loop of repeatedly attempting to constant fold
and instruction with no users.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231755 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMove variable into assert to fix -Asserts builds.
Rafael Espindola [Tue, 10 Mar 2015 04:28:09 +0000 (04:28 +0000)]
Move variable into assert to fix -Asserts builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231753 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove incredibly confusing isBaseAddressKnownZero.
Rafael Espindola [Tue, 10 Mar 2015 04:11:52 +0000 (04:11 +0000)]
Remove incredibly confusing isBaseAddressKnownZero.

When referring to a symbol in a dwarf section on ELF we should use

.long foo

instead of

.long foo - .debug_something

because ELF is unaware of the content of the sections and therefore needs
relocations. This has nothing to do with optimizing a -0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231751 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse a better name for compile unit labels.
Rafael Espindola [Tue, 10 Mar 2015 03:58:36 +0000 (03:58 +0000)]
Use a better name for compile unit labels.

They mark the start of a compile unit, so name them .Lcu_*. Using
Section->getLabelBeginName() makes it looks like they mark the start of the
section.

While at it, switch to createTempSymbol to avoid collisions with labels
created in inline assembly. Not sure if a "don't crash" test is worth it.

With this getLabelBeginName is dead, delete it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231750 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoremoved function names from comments; NFC
Sanjay Patel [Tue, 10 Mar 2015 03:48:14 +0000 (03:48 +0000)]
removed function names from comments; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231749 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDwarfAccelTable: remove unneeded bucket terminators.
Frederic Riss [Tue, 10 Mar 2015 03:47:55 +0000 (03:47 +0000)]
DwarfAccelTable: remove unneeded bucket terminators.

Last commit fixed the handling of hash collisions, but it introdcuced
unneeded bucket terminators in some places. The generated table was
correct, it can just be a tiny bit smaller. As the previous table was
correct, the test doesn't need updating. If we really wanted to test
this, I could add the section size to the dwarf dump and test for a
precise value there. IMO the correctness test is sufficient.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231748 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agouse range-based for loops; NFC
Sanjay Patel [Tue, 10 Mar 2015 03:26:39 +0000 (03:26 +0000)]
use range-based for loops; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231747 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoImprove and simplify EnforceSmallerThan for vector types.
Craig Topper [Tue, 10 Mar 2015 03:25:07 +0000 (03:25 +0000)]
Improve and simplify EnforceSmallerThan for vector types.

Explicitly compare the size of the scalar types and the whole vector size rather than just comparing enum encodings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231746 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove extra indentation of entire function body. NFC.
Craig Topper [Tue, 10 Mar 2015 03:25:04 +0000 (03:25 +0000)]
Remove extra indentation of entire function body. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231745 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMove label creation close to emission. NFC.
Rafael Espindola [Tue, 10 Mar 2015 03:11:11 +0000 (03:11 +0000)]
Move label creation close to emission. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231744 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdded ConstantExpr support to CFLAA.
George Burgess IV [Tue, 10 Mar 2015 02:58:15 +0000 (02:58 +0000)]
Added ConstantExpr support to CFLAA.

CFLAA didn't know how to properly handle ConstantExprs; it would silently
ignore them. This was a problem if the ConstantExpr is, say, a GEP of a global,
because CFLAA wouldn't realize that there's a global there. :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231743 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdded special handling for inttoptr in CFLAA.
George Burgess IV [Tue, 10 Mar 2015 02:40:06 +0000 (02:40 +0000)]
Added special handling for inttoptr in CFLAA.

We now treat pointers given to ptrtoint and pointers retrieved from
inttoptr as similar to arguments or globals (can alias anything, etc.)

This solves some of the problems we were having with giving incorrect
results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231741 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDataLayout is mandatory, update the API to reflect it with references.
Mehdi Amini [Tue, 10 Mar 2015 02:37:25 +0000 (02:37 +0000)]
DataLayout is mandatory, update the API to reflect it with references.

Summary:
Now that the DataLayout is a mandatory part of the module, let's start
cleaning the codebase. This patch is a first attempt at doing that.

This patch is not exactly NFC as for instance some places were passing
a nullptr instead of the DataLayout, possibly just because there was a
default value on the DataLayout argument to many functions in the API.
Even though it is not purely NFC, there is no change in the
validation.

I turned as many pointer to DataLayout to references, this helped
figuring out all the places where a nullptr could come up.

I had initially a local version of this patch broken into over 30
independant, commits but some later commit were cleaning the API and
touching part of the code modified in the previous commits, so it
seemed cleaner without the intermediate state.

Test Plan:

Reviewers: echristo

Subscribers: llvm-commits

From: Mehdi Amini <mehdi.amini@apple.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231740 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[sanitizer] fix instrumentation with -mllvm -sanitizer-coverage-block-threshold=0...
Kostya Serebryany [Tue, 10 Mar 2015 01:58:27 +0000 (01:58 +0000)]
[sanitizer] fix instrumentation with -mllvm -sanitizer-coverage-block-threshold=0 to actually do something useful.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231736 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[sanitizer] decrease sanitizer-coverage-block-threshold from 1000 to 500 as another...
Kostya Serebryany [Tue, 10 Mar 2015 01:11:53 +0000 (01:11 +0000)]
[sanitizer] decrease sanitizer-coverage-block-threshold from 1000 to 500 as another horrible workaround for PR17409

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231733 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDwarfAccelTable: Fix handling of hash collisions.
Frederic Riss [Tue, 10 Mar 2015 00:46:31 +0000 (00:46 +0000)]
DwarfAccelTable: Fix handling of hash collisions.

It turns out accelerator tables where totally broken if they contained
entries with colliding hashes. The failure mode is pretty bad, as it not
only impacted the colliding entries, but would basically make all the
entries after the first hash collision pointing in the wrong place.

The testcase uses the symbol names that where found to collide during a
clang build.

From a performance point of view, the patch adds a sort and a linear
walk over each bucket contents. While it has a measurable impact on the
accelerator table emission, it's not showing up significantly in clang
profiles (and I'd argue that correctness is priceless :-)).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231732 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTemporarily revert r231726 and r231724 as they're breaking the build.:
Eric Christopher [Tue, 10 Mar 2015 00:33:27 +0000 (00:33 +0000)]
Temporarily revert r231726 and r231724 as they're breaking the build.:

Author: Lang Hames <lhames@gmail.com>
Date:   Mon Mar 9 23:51:09 2015 +0000

    [Orc][MCJIT][RuntimeDyld] Add header that was accidentally left out of r231724.

Author: Lang Hames <lhames@gmail.com>
Date:   Mon Mar 9 23:44:13 2015 +0000

    [Orc][MCJIT][RuntimeDyld] Add symbol flags to symbols in RuntimeDyld. Thread the
    new types through MCJIT and Orc.

    In particular, add a 'weak' flag. When plumbed through RTDyldMemoryManager, this
    will allow us to distinguish between weak and strong definitions and find the
    right ones during symbol resolution.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231731 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove an unused variable.
Eric Christopher [Tue, 10 Mar 2015 00:33:22 +0000 (00:33 +0000)]
Remove an unused variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231730 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDon't treat .foo as two path components in path::iterators
Ben Langmuir [Tue, 10 Mar 2015 00:04:29 +0000 (00:04 +0000)]
Don't treat .foo as two path components in path::iterators

We were treating '/.foo' as ['/', '.', 'foo'] instead of ['/', '.foo'],
which lead to insanity.  Same for '..'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231727 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Orc][MCJIT][RuntimeDyld] Add header that was accidentally left out of r231724.
Lang Hames [Mon, 9 Mar 2015 23:51:09 +0000 (23:51 +0000)]
[Orc][MCJIT][RuntimeDyld] Add header that was accidentally left out of r231724.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231726 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Orc][MCJIT][RuntimeDyld] Add symbol flags to symbols in RuntimeDyld. Thread the
Lang Hames [Mon, 9 Mar 2015 23:44:13 +0000 (23:44 +0000)]
[Orc][MCJIT][RuntimeDyld] Add symbol flags to symbols in RuntimeDyld. Thread the
new types through MCJIT and Orc.

In particular, add a 'weak' flag. When plumbed through RTDyldMemoryManager, this
will allow us to distinguish between weak and strong definitions and find the
right ones during symbol resolution.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231724 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Removing unused patterns.
Colin LeMahieu [Mon, 9 Mar 2015 23:08:46 +0000 (23:08 +0000)]
[Hexagon] Removing unused patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231723 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLLParser: gep: Simplify parsing error handling
David Blaikie [Mon, 9 Mar 2015 23:08:44 +0000 (23:08 +0000)]
LLParser: gep: Simplify parsing error handling

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231722 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[CodeGen] Replace the reused stores' chain for extractelt expansion.
Ahmed Bougacha [Mon, 9 Mar 2015 22:51:05 +0000 (22:51 +0000)]
[CodeGen] Replace the reused stores' chain for extractelt expansion.

This fixes a subtle issue that was introduced in r205153.

When reusing a store for the extractelement expansion (to load directly
from it, inserting of going through the stack), later stores to the
same location might have overwritten the data we were expecting to
extract from.

To fix that, we need to explicitly replace the chain going out of the
reused store, so that later stores also have an explicit dependency on
the generated element-extracting loads, and can't clobber them.

rdar://20066785
Differential Revision: http://reviews.llvm.org/D8180

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231721 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[X86] Add nounwind to vector-idiv.ll testcases. NFC.
Ahmed Bougacha [Mon, 9 Mar 2015 22:46:02 +0000 (22:46 +0000)]
[X86] Add nounwind to vector-idiv.ll testcases.  NFC.

In preparation for a patch where cfi directives get in the way.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231720 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoReland r229944: EH: Prune unreachable resume instructions during Dwarf EH preparation
Reid Kleckner [Mon, 9 Mar 2015 22:45:16 +0000 (22:45 +0000)]
Reland r229944: EH: Prune unreachable resume instructions during Dwarf EH preparation

Fix the double-deletion of AnalysisResolver when delegating through to
Dwarf EH preparation by creating one from scratch. Hopefully the new
pass manager simplifies this.

This reverts commit r229952.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231719 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse a MapVector instead of an extra sort.
Rafael Espindola [Mon, 9 Mar 2015 22:08:37 +0000 (22:08 +0000)]
Use a MapVector instead of an extra sort.

This also has the advantage of not depending on the brittle getLabelBeginName.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231714 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Use single tailcall pseudoinst and fix checking for label jumping versus...
Colin LeMahieu [Mon, 9 Mar 2015 22:05:21 +0000 (22:05 +0000)]
[Hexagon] Use single tailcall pseudoinst and fix checking for label jumping versus tail calling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231713 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Reapply r231699. Remove assumption that second operand is an immediate...
Colin LeMahieu [Mon, 9 Mar 2015 21:48:13 +0000 (21:48 +0000)]
[Hexagon] Reapply r231699.  Remove assumption that second operand is an immediate when checking if A2_tfrsi is combinable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231710 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[SCEV] Unify getUnsignedRange and getSignedRange
Sanjoy Das [Mon, 9 Mar 2015 21:43:43 +0000 (21:43 +0000)]
[SCEV] Unify getUnsignedRange and getSignedRange

Summary:
This removes some duplicated code, and also helps optimization: e.g. in
the test case added, `%idx ULT 128` in `@x` is not currently optimized
to `true` by `-indvars` but will be, after this change.

The only functional change in ths commit is that for add recurrences,
ScalarEvolution::getRange will be more aggressive -- computing the
unsigned (resp. signed) range for a SCEVAddRecExpr will now look at the
NSW (resp. NUW) bits and check for signed (resp. unsigned) overflow.
This can be a strict improvement in some cases (such as the attached
test case), and should be no worse in other cases.

Reviewers: atrick, nlewycky

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D8142

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231709 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[SCEV] Add a `scalar-evolution-print-constant-ranges' option
Sanjoy Das [Mon, 9 Mar 2015 21:43:39 +0000 (21:43 +0000)]
[SCEV] Add a `scalar-evolution-print-constant-ranges' option

Summary:
Unused in this commit, but will be used in a subsequent change (D8142)
by a FileCheck test.

Reviewers: atrick

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D8143

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231708 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Reverting r231699
Colin LeMahieu [Mon, 9 Mar 2015 21:19:02 +0000 (21:19 +0000)]
[Hexagon] Reverting r231699

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231703 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDwarfAccelTable: fix obvious typo.
Frederic Riss [Mon, 9 Mar 2015 21:09:50 +0000 (21:09 +0000)]
DwarfAccelTable: fix obvious typo.

I have a test for that issue, but I didn't include it in the commit as it's
a 200KB file for a pretty minor issue. (The reason the file is so big is
that it needs > 1024 variables/functions to trigger and that with debug
information.

The issue/fix on the other side is totally trivial. If poeple want the test
commited, I can do that. It just didn't seem worth it to me.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231701 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Updating constant set to simpler versions.
Colin LeMahieu [Mon, 9 Mar 2015 20:33:12 +0000 (20:33 +0000)]
[Hexagon] Updating constant set to simpler versions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231699 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTableGen: Use 'enum : uint64_t' for feature flags to fix -Wmicrosoft
Reid Kleckner [Mon, 9 Mar 2015 20:23:14 +0000 (20:23 +0000)]
TableGen: Use 'enum : uint64_t' for feature flags to fix -Wmicrosoft

clang-cl would warn that this value is not representable in 'int':
  enum { FeatureX = 1ULL << 31 };
All MS enums are 'ints' unless otherwise specified, so we have to use an
explicit type.  The AMDGPU target just hit 32 features, triggering this
warning.

Now that we have C++11 strong enum types, we can also eliminate the
'const uint64_t' codepath from tablegen and just use 'enum : uint64_t'.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231697 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove the remaining uses of abs64 and nuke it.
Benjamin Kramer [Mon, 9 Mar 2015 20:20:16 +0000 (20:20 +0000)]
Remove the remaining uses of abs64 and nuke it.

std::abs works just fine and we're already using it in many places. NFC intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231696 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Removing old halfword codegen instructions and updating const32/64 splitting.
Colin LeMahieu [Mon, 9 Mar 2015 20:11:02 +0000 (20:11 +0000)]
[Hexagon] Removing old halfword codegen instructions and updating const32/64 splitting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231695 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDon't prime the section map.
Rafael Espindola [Mon, 9 Mar 2015 20:09:58 +0000 (20:09 +0000)]
Don't prime the section map.

This was just creating unused labels for .text when the module had no
functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@231694 91177308-0d34-0410-b5e6-96231b3b80d8