oota-llvm.git
9 years agoEnable FeatureFastUAMem for btver2
Sanjay Patel [Fri, 28 Nov 2014 18:40:18 +0000 (18:40 +0000)]
Enable FeatureFastUAMem for btver2

Allow unaligned 16-byte memop codegen for btver2. No functional changes for any other subtargets.

Replace the existing supposed small memcpy test with an actual test of a small memcpy.
The previous test wasn't using FileCheck either.

This patch should allow us to close PR21541 ( http://llvm.org/bugs/show_bug.cgi?id=21541 ).

Differential Revision: http://reviews.llvm.org/D6360

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222925 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd back r222727 with a fix.
Rafael Espindola [Fri, 28 Nov 2014 16:41:24 +0000 (16:41 +0000)]
Add back r222727 with a fix.

The original patch would fail when:

* A dst opaque type (%A) is matched with a src type (%A).
* A src opaque (%E) type is then speculatively matched with %A and the
  speculation fails afterward.
* When rolling back the speculation we would cancel the source %A to dest
  %A mapping.

The fix is to keep an explicit list of which resolutions are speculative.

Original message:

Fix overly aggressive type merging.

If we find out that two types are *not* isomorphic, we learn nothing about
opaque sub types in both the source and destination.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222923 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd an assert and use a range loop. NFC.
Rafael Espindola [Fri, 28 Nov 2014 16:26:14 +0000 (16:26 +0000)]
Add an assert and use a range loop. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222922 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd a testcase reduced from clang lto bootstrap on OS X.
Rafael Espindola [Fri, 28 Nov 2014 15:45:31 +0000 (15:45 +0000)]
Add a testcase reduced from clang lto bootstrap on OS X.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222921 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix wrong encoding of MRSBanked.
Charlie Turner [Fri, 28 Nov 2014 15:01:06 +0000 (15:01 +0000)]
Fix wrong encoding of MRSBanked.

Patch by Matthew Wahab.

Change-Id: Ia2a001ca2760028ea360fe77b56f203a219eefbc

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222920 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[msan] Fix origin propagation for select of floats.
Evgeniy Stepanov [Fri, 28 Nov 2014 11:17:58 +0000 (11:17 +0000)]
[msan] Fix origin propagation for select of floats.

MSan does not assign origin for instrumentation temps (i.e. the ones that do
not come from the application code), but "select" instrumentation erroneously
tried to use one of those.

https://code.google.com/p/memory-sanitizer/issues/detail?id=78

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222918 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTest all <build attribute, value> pairs.
Charlie Turner [Fri, 28 Nov 2014 11:14:47 +0000 (11:14 +0000)]
Test all <build attribute, value> pairs.

Add more tests to make sure the encoding/decoding of build attributes works
correctly for all permissible values of build attributes. For cases where there
are an infinite number of such values, a representative subset has been settled
for.

Change-Id: I2643c9624c211b2d56405306e16eec2d487bc5d6

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222917 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemoved extra line from a comment to test first commit. NFC.
Ankur Garg [Fri, 28 Nov 2014 10:38:18 +0000 (10:38 +0000)]
Removed extra line from a comment to test first commit. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222916 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix another memory leak in TableGen AsmMatcher by deleting CodeGenInstAliases.
Craig Topper [Fri, 28 Nov 2014 05:01:21 +0000 (05:01 +0000)]
Fix another memory leak in TableGen AsmMatcher by deleting CodeGenInstAliases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222912 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd missing 'override' keyword.
Craig Topper [Fri, 28 Nov 2014 03:58:26 +0000 (03:58 +0000)]
Add missing 'override' keyword.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222911 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse unique_ptr to fix some memory leaks in Tablegen AsmMatcherEmitter.
Craig Topper [Fri, 28 Nov 2014 03:53:02 +0000 (03:53 +0000)]
Use unique_ptr to fix some memory leaks in Tablegen AsmMatcherEmitter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222909 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse range-based for loops and const-correct a few things.
Craig Topper [Fri, 28 Nov 2014 03:53:00 +0000 (03:53 +0000)]
Use range-based for loops and const-correct a few things.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222908 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoStop using ArrayRef of a const type.
Tim Northover [Thu, 27 Nov 2014 21:29:20 +0000 (21:29 +0000)]
Stop using ArrayRef of a const type.

I *think* this is what the GCC bots are complaining about.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222905 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAArch64: treat [N x Ty] as a block during procedure calls.
Tim Northover [Thu, 27 Nov 2014 21:02:42 +0000 (21:02 +0000)]
AArch64: treat [N x Ty] as a block during procedure calls.

The AAPCS treats small structs and homogeneous floating (or vector) aggregates
specially, and guarantees they either get passed as a contiguous block of
registers, or prevent any future use of those registers and get passed on the
stack.

This concept can fit quite neatly into LLVM's own type system, mapping an HFA
to [N x float] and so on, and small structs to [N x i64]. Doing so allows
front-ends to emit AAPCS compliant code without having to duplicate the
register counting logic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222903 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips][microMIPS] Implement SWM16 and LWM16 instructions
Zoran Jovanovic [Thu, 27 Nov 2014 18:28:59 +0000 (18:28 +0000)]
[mips][microMIPS] Implement SWM16 and LWM16 instructions
Differential Revision: http://reviews.llvm.org/D5579

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222901 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips][microMIPS] Implement BREAK16 and SDBBP16 instructions
Jozef Kolek [Thu, 27 Nov 2014 18:18:42 +0000 (18:18 +0000)]
[mips][microMIPS] Implement BREAK16 and SDBBP16 instructions

Patch by Radovan Obradovic.

Differential Revision: http://reviews.llvm.org/D5048

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222900 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips] Add synci instruction.
Daniel Sanders [Thu, 27 Nov 2014 17:28:10 +0000 (17:28 +0000)]
[mips] Add synci instruction.

Patch by Amaury Pouly

Reviewers: dsanders

Reviewed By: dsanders

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D6421

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222899 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoWiden ELFYAML relocation type to 32 bits
Will Newton [Thu, 27 Nov 2014 17:20:48 +0000 (17:20 +0000)]
Widen ELFYAML relocation type to 32 bits

The current 8 bits is sufficient for ELF32 targets but ELF64 requires
32 bits. Add a test for AArch64 that exposes the issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222898 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCommit back the correct bits of r222760 (was r222538).
Rafael Espindola [Thu, 27 Nov 2014 17:13:56 +0000 (17:13 +0000)]
Commit back the correct bits of r222760 (was r222538).

I also added a test.

Original message:

Allow FDE references outside the +/-2GB range supported by PC relative
offsets for code models other than small/medium. For JIT application,
memory layout is less controlled and can result in truncations
otherwise.

Patch from Akos Kiss.

Differential Revision: http://reviews.llvm.org/D6079

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222897 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "Reapply 222538 and update tests to explicitly request small code model and...
Rafael Espindola [Thu, 27 Nov 2014 17:13:51 +0000 (17:13 +0000)]
Revert "Reapply 222538 and update tests to explicitly request small code model and PIC:"

This reverts commit r222760.

It changed our behaviour on PIC so we don't match gas anymore. It also
included lots of unnecessary changes to tests.

If those changes are desirable, there should be an independent discussion
as they are out of scope for that patch.

I will recommit the other bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222896 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "Fix overly aggressive type merging."
Duncan P. N. Exon Smith [Thu, 27 Nov 2014 17:01:10 +0000 (17:01 +0000)]
Revert "Fix overly aggressive type merging."

This reverts commit r222727, which causes LTO bootstrap failures.

Last passing @ r222698:
http://lab.llvm.org:8080/green/job/clang-Rlto_master_build/532/

First failing @ r222843:
http://lab.llvm.org:8080/green/job/clang-Rlto_master_build/533/

Internal bootstraps pointed at a much narrower range: r222725 is
passing, and r222731 is failing.

LTO crashes while handling libclang.dylib:
http://lab.llvm.org:8080/green/job/clang-Rlto_master_build/533/consoleFull#-158682280549ba4694-19c4-4d7e-bec5-911270d8a58c

    GEP is not of right type for indices!
      %InfoObj.i.i = getelementptr inbounds %"class.llvm::OnDiskIterableChainedHashTable"* %.lcssa, i64 0, i32 0, i32 4, !dbg !123627
     %"class.clang::serialization::reader::ASTIdentifierLookupTrait" = type { %"class.clang::ASTReader.31859"*, %"class.clang::serialization::ModuleFile.31870"*, %"class.clang::IdentifierInfo"* }LLVM ERROR: Broken function found, compilation aborted!
    clang: error: linker command failed with exit code 1 (use -v to see invocation)

Looks like the new algorithm doesn't merge types aggressively enough.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222895 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoreinstate r222872: Peephole optimization in switch table lookup: reuse the guarding...
Erik Eckstein [Thu, 27 Nov 2014 15:13:14 +0000 (15:13 +0000)]
reinstate r222872: Peephole optimization in switch table lookup: reuse the guarding table comparison if possible.

Fixed missing dominance check.
Original commit message:

This optimization tries to reuse the generated compare instruction, if there is a comparison against the default value after the switch.
Example:
   if (idx < tablesize)
      r = table[idx]; // table does not contain default_value
   else
      r = default_value;
   if (r != default_value)
      ...
Is optimized to:
   cond = idx < tablesize;
   if (cond)
      r = table[idx];
   else
      r = default_value;
   if (cond)
      ...
Jump threading will then eliminate the second if(cond).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222891 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[msan] Remove indirect call wrapping code.
Evgeniy Stepanov [Thu, 27 Nov 2014 14:54:02 +0000 (14:54 +0000)]
[msan] Remove indirect call wrapping code.

This functionality was only used in MSanDR, which is deprecated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222889 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips][microMIPS] Implement disassembler support for 16-bit instructions LI16, ADDIUR...
Jozef Kolek [Thu, 27 Nov 2014 14:41:44 +0000 (14:41 +0000)]
[mips][microMIPS] Implement disassembler support for 16-bit instructions LI16, ADDIUR1SP, ADDIUR2 and ADDIUS5

Differential Revision: http://reviews.llvm.org/D6419

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222887 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoStop uppercasing build attribute data.
Charlie Turner [Thu, 27 Nov 2014 12:13:56 +0000 (12:13 +0000)]
Stop uppercasing build attribute data.

The string data for string-valued build attributes were being unconditionally
uppercased. There is no mention in the ARM ABI addenda about case conventions,
so it's technically implementation defined as to whether the data are
capitialised in some way or not. However, there are good reasons not to
captialise the data.

  * It's less work.
  * Some vendors may legitimately have case-sensitive checks for these
    attributes which would fail on LLVM generated object files.
  * There could be locale issues with uppercasing.

The original reasons for uppercasing appear to have stemmed from an
old codesourcery toolchain behaviour, see

http://comments.gmane.org/gmane.comp.compilers.llvm.cvs/87133

This patch makes the object file emitted no longer captialise string
data, it encodes as seen in the assembly source.

Change-Id: Ibe20dd6e60d2773d57ff72a78470839033aa5538

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222882 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse FileCheck instead of grep. Change by Ankur Garg.
Suyog Sarda [Thu, 27 Nov 2014 11:22:49 +0000 (11:22 +0000)]
Use FileCheck instead of grep. Change by Ankur Garg.

Differential Revision: http://reviews.llvm.org/D6430

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222879 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "Peephole optimization in switch table lookup: reuse the guarding table compar...
Erik Eckstein [Thu, 27 Nov 2014 10:59:08 +0000 (10:59 +0000)]
Revert "Peephole optimization in switch table lookup: reuse the guarding table comparison if possible."

It is breaking the clang bootstrag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222877 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse FileCheck instead of grep. Change by Sonam.
Suyog Sarda [Thu, 27 Nov 2014 10:57:24 +0000 (10:57 +0000)]
Use FileCheck instead of grep. Change by Sonam.

Differential Revision: http://reviews.llvm.org/D6432

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222876 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoPeephole optimization in switch table lookup: reuse the guarding table comparison...
Erik Eckstein [Thu, 27 Nov 2014 08:33:51 +0000 (08:33 +0000)]
Peephole optimization in switch table lookup: reuse the guarding table comparison if possible.

This optimization tries to reuse the generated compare instruction, if there is a comparison against the default value after the switch.
Example:
    if (idx < tablesize)
       r = table[idx]; // table does not contain default_value
    else
       r = default_value;
    if (r != default_value)
       ...
Is optimized to:
    cond = idx < tablesize;
    if (cond)
       r = table[idx];
    else
       r = default_value;
    if (cond)
       ...
\endcode
Jump threading will then eliminate the second if(cond).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222872 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInstCombine: Restore optimizations lost in r210006
David Majnemer [Thu, 27 Nov 2014 07:25:21 +0000 (07:25 +0000)]
InstCombine: Restore optimizations lost in r210006

This restores our ability to optimize:
(X & C) == 0 ? X ^ C : X  into  X | C
(X & C) != 0 ? X ^ C : X  into  X & ~C

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222871 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd LLVMObject to LLVMExecutionEngine.
NAKAMURA Takumi [Thu, 27 Nov 2014 06:36:22 +0000 (06:36 +0000)]
Add LLVMObject to LLVMExecutionEngine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222869 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInstSimplify: Restore optimizations lost in r210006
David Majnemer [Thu, 27 Nov 2014 06:32:46 +0000 (06:32 +0000)]
InstSimplify: Restore optimizations lost in r210006

This restores our ability to optimize:
(X & C) ? X & ~C : X  into  X & ~C
(X & C) ? X : X & ~C  into  X
(X & C) ? X | C : X  into  X
(X & C) ? X : X | C  into  X | C

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222868 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Remove the local symbol table from RuntimeDlyd - it's not needed.
Lang Hames [Thu, 27 Nov 2014 05:40:13 +0000 (05:40 +0000)]
[MCJIT] Remove the local symbol table from RuntimeDlyd - it's not needed.

All symbols have to be stored in the global symbol to enable
cross-rtdyld-instance linking, so the local symbol table content is
redundant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222867 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Update CMakeLists.txt for llvm-rtdyld to add Object as a requirement.
Lang Hames [Thu, 27 Nov 2014 04:18:50 +0000 (04:18 +0000)]
[MCJIT] Update CMakeLists.txt for llvm-rtdyld to add Object as a requirement.

Hopefully this will fix
http://bb.pgr.jp/builders/clang-3stage-x86_64-linux/builds/8271/steps/stage1_build/logs/stdio

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222866 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Replace JITEventListener::anchor (temporarily removed in r222861), and
Lang Hames [Thu, 27 Nov 2014 01:41:16 +0000 (01:41 +0000)]
[MCJIT] Replace JITEventListener::anchor (temporarily removed in r222861), and
move GDBRegistrationListener into ExecutionEngine to avoid layering violation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222864 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Remove JITEventListener's anchor until I can determine the right place
Lang Hames [Thu, 27 Nov 2014 00:15:28 +0000 (00:15 +0000)]
[MCJIT] Remove JITEventListener's anchor until I can determine the right place
to put it. This should unbreak the Mips bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222861 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTeach LLVM about llgo subproject.
Peter Collingbourne [Thu, 27 Nov 2014 00:15:21 +0000 (00:15 +0000)]
Teach LLVM about llgo subproject.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222860 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Move get-any-symbol-load-address logic out of RuntimeDyld and into
Lang Hames [Thu, 27 Nov 2014 00:12:28 +0000 (00:12 +0000)]
[MCJIT] Move get-any-symbol-load-address logic out of RuntimeDyld and into
RuntimeDyldChecker.

RuntimeDyld instances should only provide lookup for locally defined
symbols.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222859 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "Added inst combine transforms for single bit tests from Chris's note"
David Majnemer [Wed, 26 Nov 2014 23:00:38 +0000 (23:00 +0000)]
Revert "Added inst combine transforms for single bit tests from Chris's note"

This reverts commit r210006, it miscompiled libapr which is used in who
knows how many projects.

A test has been added to ensure that we don't regress again.

I'll work on a rewrite of what the optimization was trying to do later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222856 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd a small "usage:" comment at the top of not.cpp
Sean Silva [Wed, 26 Nov 2014 22:53:46 +0000 (22:53 +0000)]
Add a small "usage:" comment at the top of not.cpp

Mostly pulled from Rafael's r185678 commit message.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222855 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoObject/COFF: Fix off-by-one error for object having lots of relocations
Rui Ueyama [Wed, 26 Nov 2014 22:17:25 +0000 (22:17 +0000)]
Object/COFF: Fix off-by-one error for object having lots of relocations

llvm-objdump printed out an error message for this off-by-one error,
but because it always exits with 0 whether or not it found an error,
the test (llvm-objdump/coff-many-relocs.test) succeeded.
I made llvm-objdump exit with EXIT_FAILURE when an error is found.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222852 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Use ZeroOrNegativeOneBooleanContent
Matt Arsenault [Wed, 26 Nov 2014 21:23:15 +0000 (21:23 +0000)]
R600/SI: Use ZeroOrNegativeOneBooleanContent

This sort of doesn't matter since the setcc type is i1, but
this previously was using the default UndefinedBooleanContent. This
makes it more consistent with R600. This enables more optimizations
which typically give up on UndefinedBooleanContent. For example,
there is already a special case target DAG combine for
setcc + sext which can be eliminated in favor of what the generic
DAG combiner can do if it assumes boolean values are sign extended.
Since -1 is an inline immediate, using it is basically free and the
backend already uses it when a boolean value is needed in a wider type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222850 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding cmp* immediate form instructions.
Colin LeMahieu [Wed, 26 Nov 2014 19:43:12 +0000 (19:43 +0000)]
[Hexagon] Adding cmp* immediate form instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222849 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips][microMIPS] Implement disassembler support for 16-bit instructions LBU16, LHU16...
Jozef Kolek [Wed, 26 Nov 2014 18:56:38 +0000 (18:56 +0000)]
[mips][microMIPS] Implement disassembler support for 16-bit instructions LBU16, LHU16, LW16, SB16, SH16 and SW16

Differential Revision: http://reviews.llvm.org/D6405

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222847 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding and64, or64, and xor64 instructions.
Colin LeMahieu [Wed, 26 Nov 2014 18:55:59 +0000 (18:55 +0000)]
[Hexagon] Adding and64, or64, and xor64 instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222846 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Create e64 versions of and/or/xor in SILowerI1Copies
Matt Arsenault [Wed, 26 Nov 2014 18:18:28 +0000 (18:18 +0000)]
R600/SI: Create e64 versions of and/or/xor in SILowerI1Copies

This fixes moving boolean constants into registers before operating
on them. They get permuted and shrunk down to e32 anyway later. This
is a temporary fix until the patch that removes these pseudos is
committed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222844 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Fix missing return statement.
Lang Hames [Wed, 26 Nov 2014 17:21:41 +0000 (17:21 +0000)]
[MCJIT] Fix missing return statement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222841 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Reapply r222828 and r222810-r222812 with fix for MSVC move-op issues.
Lang Hames [Wed, 26 Nov 2014 16:54:40 +0000 (16:54 +0000)]
[MCJIT] Reapply r222828 and r222810-r222812 with fix for MSVC move-op issues.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222840 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoReverting r222828 and r222810-r222812 as they broke the build on Windows.
Aaron Ballman [Wed, 26 Nov 2014 15:27:39 +0000 (15:27 +0000)]
Reverting r222828 and r222810-r222812 as they broke the build on Windows.

http://bb.pgr.jp/builders/ninja-clang-i686-msc17-R/builds/11753

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222833 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemoving a spurious semicolon; NFC
Aaron Ballman [Wed, 26 Nov 2014 13:55:55 +0000 (13:55 +0000)]
Removing a spurious semicolon; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222830 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd missing "override".
Evgeniy Stepanov [Wed, 26 Nov 2014 12:26:03 +0000 (12:26 +0000)]
Add missing "override".

Fixes compilation failure in r222810.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222828 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUpdate AArch64 ELF relocations to ABI 1.0
Will Newton [Wed, 26 Nov 2014 10:49:18 +0000 (10:49 +0000)]
Update AArch64 ELF relocations to ABI 1.0

This mostly entails adding relocations, however there are a couple of
changes to existing relocations:

1. R_AARCH64_NONE is defined to be zero rather than 256

R_AARCH64_NONE has been defined to be zero for a long time elsewhere
e.g. binutils and glibc since the submission of the AArch64 port in
2012 so this is required for compatibility.

2. R_AARCH64_TLSDESC_ADR_PAGE renamed to R_AARCH64_TLSDESC_ADR_PAGE21

I don't think there is any way for relocation names to leak out of LLVM
so this should not break anything.

Tested with check-all with no regressions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222821 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAVX-512: Scalar ERI intrinsics
Elena Demikhovsky [Wed, 26 Nov 2014 10:46:49 +0000 (10:46 +0000)]
AVX-512: Scalar ERI intrinsics
including SAE mode and memory operand.
Added AVX512_maskable_scalar template, that should cover all scalar instructions in the future.

The main difference between AVX512_maskable_scalar<> and AVX512_maskable<> is using X86select instead of vselect.
I need it, because I can't create vselect node for MVT::i1 mask for scalar instruction.

http://reviews.llvm.org/D6378

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222820 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUpdate ARM ELF relocations to ABI 2.09
Will Newton [Wed, 26 Nov 2014 10:36:03 +0000 (10:36 +0000)]
Update ARM ELF relocations to ABI 2.09

Add R_ARM_IRELATIVE.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222817 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Re-enable GDB registration (temporarily disabled in r222811), but check
Lang Hames [Wed, 26 Nov 2014 07:39:03 +0000 (07:39 +0000)]
[MCJIT] Re-enable GDB registration (temporarily disabled in r222811), but check
that we actually have an object to register first.

For MachO objects, RuntimeDyld::LoadedObjectInfo::getObjectForDebug returns an
empty OwningBinary<ObjectFile> which was causing crashes in the GDB registration
code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222812 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Temporarily disable automatic JIT debugger registration.
Lang Hames [Wed, 26 Nov 2014 07:25:26 +0000 (07:25 +0000)]
[MCJIT] Temporarily disable automatic JIT debugger registration.

The RuntimeDyld cleanup patch r222810 turned on GDB registration for MachO
objects. I expected this to be harmless, but it seems to have broken on
MacsOS. Temporarily disabling debugger registration while I dig in to what's
gone wrong.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222811 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MCJIT] Clean up RuntimeDyld's quirky object-ownership/modification scheme.
Lang Hames [Wed, 26 Nov 2014 06:53:26 +0000 (06:53 +0000)]
[MCJIT] Clean up RuntimeDyld's quirky object-ownership/modification scheme.

Previously, when loading an object file, RuntimeDyld (1) took ownership of the
ObjectFile instance (and associated MemoryBuffer), (2) potentially modified the
object in-place, and (3) returned an ObjectImage that managed ownership of the
now-modified object and provided some convenience methods. This scheme accreted
over several years as features were tacked on to RuntimeDyld, and was both
unintuitive and unsafe (See e.g. http://llvm.org/PR20722).

This patch fixes the issue by removing all ownership and in-place modification
of object files from RuntimeDyld. Existing behavior, including debugger
registration, is preserved.

Noteworthy changes include:

(1) ObjectFile instances are now passed to RuntimeDyld by const-ref.
(2) The ObjectImage and ObjectBuffer classes have been removed entirely, they
    existed to model ownership within RuntimeDyld, and so are no longer needed.
(3) RuntimeDyld::loadObject now returns an instance of a new class,
    RuntimeDyld::LoadedObjectInfo, which can be used to construct a modified
    object suitable for registration with the debugger, following the existing
    debugger registration scheme.
(4) The JITRegistrar class has been removed, and the GDBRegistrar class has been
    re-written as a JITEventListener.

This should fix http://llvm.org/PR20722 .

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222810 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove neverHasSideEffects support from TableGen CodeGenInstruction. Everyone should...
Craig Topper [Wed, 26 Nov 2014 04:11:14 +0000 (04:11 +0000)]
Remove neverHasSideEffects support from TableGen CodeGenInstruction. Everyone should use hasSideEffects now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222809 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTry to make the modules buildbot happy again.
Richard Smith [Wed, 26 Nov 2014 03:44:47 +0000 (03:44 +0000)]
Try to make the modules buildbot happy again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222808 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoReformat with gofmt.
Eric Christopher [Wed, 26 Nov 2014 02:57:33 +0000 (02:57 +0000)]
Reformat with gofmt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222806 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMake SetMCJITOptimizationLevel more of a method and pass options
Eric Christopher [Wed, 26 Nov 2014 02:54:24 +0000 (02:54 +0000)]
Make SetMCJITOptimizationLevel more of a method and pass options
as a reference. Move closer to the type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222805 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMake sure that the go bindings call LLVMInitializeMCJITCompilerOptions
Eric Christopher [Wed, 26 Nov 2014 02:27:46 +0000 (02:27 +0000)]
Make sure that the go bindings call LLVMInitializeMCJITCompilerOptions
so that they initialize the code generation model to the correct
(non-zero) default model.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222804 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoReplace neverHasSideEffects=1 with hasSideEffects=0 in all .td files.
Craig Topper [Wed, 26 Nov 2014 00:46:26 +0000 (00:46 +0000)]
Replace neverHasSideEffects=1 with hasSideEffects=0 in all .td files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222801 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[X86][SSE] Improvements to byte shift shuffle matching
Simon Pilgrim [Tue, 25 Nov 2014 22:34:59 +0000 (22:34 +0000)]
[X86][SSE] Improvements to byte shift shuffle matching

Since (v)pslldq / (v)psrldq instructions resolve to a single input argument it is useful to match it much earlier than we currently do - this prevents more complicated shuffles (notably insertion into a zero vector) matching before it.

Differential Revision: http://reviews.llvm.org/D6409

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222796 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding add64 and sub64 instructions.
Colin LeMahieu [Tue, 25 Nov 2014 22:15:44 +0000 (22:15 +0000)]
[Hexagon] Adding add64 and sub64 instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222795 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoReverting 222792
Colin LeMahieu [Tue, 25 Nov 2014 21:39:57 +0000 (21:39 +0000)]
Reverting 222792

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222793 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding compare with immediate instructions.
Colin LeMahieu [Tue, 25 Nov 2014 21:30:28 +0000 (21:30 +0000)]
[Hexagon] Adding compare with immediate instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222792 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding NOP encoding bits.
Colin LeMahieu [Tue, 25 Nov 2014 21:23:07 +0000 (21:23 +0000)]
[Hexagon] Adding NOP encoding bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222791 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoGo bindings: add DIBuilder.InsertValueAtEnd
Peter Collingbourne [Tue, 25 Nov 2014 21:05:04 +0000 (21:05 +0000)]
Go bindings: add DIBuilder.InsertValueAtEnd

Expose llvm::DIBuilder::insertDbgValueIntrinsic as
DIBuilder.InsertValueAtEnd in the Go bindings, to support attaching
debug metadata to register values.

Patch by Andrew Wilkins!

Differential Revision: http://reviews.llvm.org/D6374

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222790 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Only use one DEBUG()
Matt Arsenault [Tue, 25 Nov 2014 21:03:22 +0000 (21:03 +0000)]
R600/SI: Only use one DEBUG()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222789 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoThis test requires asserts because of -stats.
Rafael Espindola [Tue, 25 Nov 2014 20:56:56 +0000 (20:56 +0000)]
This test requires asserts because of -stats.

Sorry about that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222788 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agogold plugin: call llvm_shutdown so that -stats works.
Rafael Espindola [Tue, 25 Nov 2014 20:52:49 +0000 (20:52 +0000)]
gold plugin: call llvm_shutdown so that -stats works.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222787 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AVX512] Add 512b integer shift by variable intrinsics and patterns.
Cameron McInally [Tue, 25 Nov 2014 20:41:51 +0000 (20:41 +0000)]
[AVX512] Add 512b integer shift by variable intrinsics and patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222786 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] [NFC] Adding trailing whitespace to test files.
Colin LeMahieu [Tue, 25 Nov 2014 20:22:24 +0000 (20:22 +0000)]
[Hexagon] [NFC] Adding trailing whitespace to test files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222785 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Adding C2_mux instruction.
Colin LeMahieu [Tue, 25 Nov 2014 20:20:09 +0000 (20:20 +0000)]
[Hexagon] Adding C2_mux instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222784 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove unncessary check for Int_* and *_Int in AsmMatcherEmitter. These are all marke...
Craig Topper [Tue, 25 Nov 2014 20:11:34 +0000 (20:11 +0000)]
Remove unncessary check for Int_* and *_Int in AsmMatcherEmitter. These are all marked isCodeGenOnly these days.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222783 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse range-based for loops.
Craig Topper [Tue, 25 Nov 2014 20:11:31 +0000 (20:11 +0000)]
Use range-based for loops.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222782 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove dead code.
Craig Topper [Tue, 25 Nov 2014 20:11:29 +0000 (20:11 +0000)]
Remove dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222781 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove unused MaxSize variable.
Craig Topper [Tue, 25 Nov 2014 20:11:27 +0000 (20:11 +0000)]
Remove unused MaxSize variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222780 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMove a vector instead of copying it.
Craig Topper [Tue, 25 Nov 2014 20:11:25 +0000 (20:11 +0000)]
Move a vector instead of copying it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222779 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove space before tab in all AVX512 mnemonic strings.
Craig Topper [Tue, 25 Nov 2014 20:11:23 +0000 (20:11 +0000)]
Remove space before tab in all AVX512 mnemonic strings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222778 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove useless rdar:// comment from switch_to_lookup_table.ll test.
Hans Wennborg [Tue, 25 Nov 2014 18:45:23 +0000 (18:45 +0000)]
Remove useless rdar:// comment from switch_to_lookup_table.ll test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222772 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[Hexagon] Replacing cmp* instructions with ones that contain encoding bits.
Colin LeMahieu [Tue, 25 Nov 2014 18:20:52 +0000 (18:20 +0000)]
[Hexagon] Replacing cmp* instructions with ones that contain encoding bits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222771 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCleaning out google tests from MC.
Colin LeMahieu [Tue, 25 Nov 2014 18:03:08 +0000 (18:03 +0000)]
Cleaning out google tests from MC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222770 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLazyValueInfo: Actually re-visit partially solved block-values in solveBlockValue()
Hans Wennborg [Tue, 25 Nov 2014 17:23:05 +0000 (17:23 +0000)]
LazyValueInfo: Actually re-visit partially solved block-values in solveBlockValue()

If solveBlockValue() needs results from predecessors that are not already
computed, it returns false with the intention of resuming when the dependencies
have been resolved. However, the computation would never be resumed since an
'overdefined' result had been placed in the cache, preventing any further
computation.

The point of placing the 'overdefined' result in the cache seems to have been
to break cycles, but we can check for that when inserting work items in the
BlockValue stack instead. This makes the "stop and resume" mechanism of
solveBlockValue() work as intended, unlocking more analysis.

Using this patch shaves 120 KB off a 64-bit Chromium build on Linux.

I benchmarked compiling bzip2.c at -O2 but couldn't measure any difference in
compile time.

Tests by Jiangning Liu from r215343 / PR21238, Pete Cooper, and me.

Differential Revision: http://reviews.llvm.org/D6397

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222768 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSmall model and JIT generally don't go well with each other.
Joerg Sonnenberger [Tue, 25 Nov 2014 17:14:22 +0000 (17:14 +0000)]
Small model and JIT generally don't go well with each other.
On LP64 platforms, it will work or not depending on the choosen memory
layout, so neither PASS nor XFAIL is appropiate.
As UNSUPPORTED as per-test target doesn't exist (yet), remove the test
instead to unbreak the builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222767 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSet the body of a new struct as soon as it is created.
Rafael Espindola [Tue, 25 Nov 2014 15:33:40 +0000 (15:33 +0000)]
Set the body of a new struct as soon as it is created.

This changes the order in which different types are passed to get, but
one order is not inherently better than the other.

The main motivation is that this simplifies linkDefinedTypeBodies now that
it is only linking "real" opaque types. It is also means that we only have to
call it once and that we don't need getImpl.

A small change in behavior is that we don't copy type names when resolving
opaque types. This is an improvement IMHO, but it can be added back if
desired. A test is included with the new behavior.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222764 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[msan] Annotate zlib functions for MemorySanitizer.
Evgeniy Stepanov [Tue, 25 Nov 2014 15:24:07 +0000 (15:24 +0000)]
[msan] Annotate zlib functions for MemorySanitizer.

Mark destination buffer in zlib::compress and zlib::decompress as fully
initialized.

When building LLVM with system zlib and MemorySanitizer instrumentation,
MSan does not observe memory writes in zlib code and erroneously considers
zlib output buffers as uninitialized, resulting in false use-of-uninitialized
memory reports. This change helps MSan understand the state of that memory
and prevents such reports.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222763 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMisc style fixes. NFC.
Rafael Espindola [Tue, 25 Nov 2014 14:35:53 +0000 (14:35 +0000)]
Misc style fixes. NFC.

This just reduces the noise in the next patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222761 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoReapply 222538 and update tests to explicitly request small code model
Joerg Sonnenberger [Tue, 25 Nov 2014 13:37:55 +0000 (13:37 +0000)]
Reapply 222538 and update tests to explicitly request small code model
and PIC:

Allow FDE references outside the +/-2GB range supported by PC relative
offsets for code models other than small/medium. For JIT application,
memory layout is less controlled and can result in truncations
otherwise.

Patch from Akos Kiss.

Differential Revision: http://reviews.llvm.org/D6079

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222760 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMark as explicit failing on x86-64 -- small memory model doesn't agree
Joerg Sonnenberger [Tue, 25 Nov 2014 13:28:56 +0000 (13:28 +0000)]
Mark as explicit failing on x86-64 -- small memory model doesn't agree
with default address selections.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222759 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove a bit of duplicated code.
Rafael Espindola [Tue, 25 Nov 2014 13:19:46 +0000 (13:19 +0000)]
Remove a bit of duplicated code.

Exactly the same checks are present in areTypesIsomorphic.

This might have been a premature performance optimization. I cannot reproduce
any slowdown with this patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222758 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert r222746: That commit did not update any tests and caused two R600
Chandler Carruth [Tue, 25 Nov 2014 10:50:41 +0000 (10:50 +0000)]
Revert r222746: That commit did not update any tests and caused two R600
tests to start failing.

Original commit log: R600/SI: Disable commutativity for MIN/MAX_LEGACY

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222753 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips][micromips] Use call instructions with short delay slots
Zoran Jovanovic [Tue, 25 Nov 2014 10:50:00 +0000 (10:50 +0000)]
[mips][micromips] Use call instructions with short delay slots
Differential Revision: http://reviews.llvm.org/D6338

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222752 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[InstCombine] Change LLVM To canonicalize toward the value type being
Chandler Carruth [Tue, 25 Nov 2014 10:09:51 +0000 (10:09 +0000)]
[InstCombine] Change LLVM To canonicalize toward the value type being
stored rather than the pointer type.

This change is analogous to r220138 which changed the canonicalization
for loads. The rationale is the same: memory does not have a type,
operations (and thus the values they produce) have a type. We should
match that type as closely as possible rather than reading some form of
semantics into the pointer type.

With this change, loads and stores should no longer be made with
nonsensical types for the values that tehy load and store. This is
particularly important when trying to match specific loaded and stored
types in the process of doing other instcombines, which is what led me
down this twisty maze of miscanonicalization.

I've put quite some effort into looking through IR to find places where
LLVM's optimizer was being unreasonably conservative in the face of
mismatched load and store types, however it is possible (let's say,
likely!) I have missed some. If you see regressions here, or from
r220138, the likely cause is some part of LLVM failing to cope with load
and store types differing. Test cases appreciated, it is important that
we root all of these out of LLVM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222748 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Disable commutativity for MIN/MAX_LEGACY
Marek Olsak [Tue, 25 Nov 2014 09:49:23 +0000 (09:49 +0000)]
R600/SI: Disable commutativity for MIN/MAX_LEGACY

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222746 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCorrectly handle Tag_CPU_arch_profile.
Charlie Turner [Tue, 25 Nov 2014 09:30:09 +0000 (09:30 +0000)]
Correctly handle Tag_CPU_arch_profile.

Fix ARMAttributeParser::CPU_arch_profile so that it doesn't switch on the value
'0' as a legal value of this build attribute.

Change-Id: Ie05a08900a82bb10b78c841b437df747ce3bb38e

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222743 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoChange the test case file to use FileCheck instead of grep. NFC.
Suyog Sarda [Tue, 25 Nov 2014 08:44:56 +0000 (08:44 +0000)]
Change the test case file to use FileCheck instead of grep. NFC.
Change by Ankur Garg.

Differential Revision: http://reviews.llvm.org/D6382

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222740 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert r220349 to re-instate r220277 with a fix for PR21330 -- quite
Chandler Carruth [Tue, 25 Nov 2014 08:20:27 +0000 (08:20 +0000)]
Revert r220349 to re-instate r220277 with a fix for PR21330 -- quite
clearly only exactly equal width ptrtoint and inttoptr casts are no-op
casts, it says so right there in the langref. Make the code agree.

Original log from r220277:
Teach the load analysis to allow finding available values which require
inttoptr or ptrtoint cast provided there is datalayout available.
Eventually, the datalayout can just be required but in practice it will
always be there today.

To go with the ability to expose available values requiring a ptrtoint
or inttoptr cast, helpers are added to perform one of these three casts.

These smarts are necessary to finish canonicalizing loads and stores to
the operational type requirements without regressing fundamental
combines.

I've added some test cases. These should actually improve as the load
combining and store combining improves, but they may fundamentally be
highlighting some missing combines for select in addition to exercising
the specific added logic to load analysis.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222739 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Fix allocating flat_scr_lo / flat_scr_hi
Matt Arsenault [Tue, 25 Nov 2014 07:53:06 +0000 (07:53 +0000)]
R600/SI: Fix allocating flat_scr_lo / flat_scr_hi

Only the super register flat_scr was marked as reserved,
so in some cases with high register usage it would still
try to allocate the subregisters.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@222737 91177308-0d34-0410-b5e6-96231b3b80d8