7 years ago[Orc] Add header comment to IndirectionUtils.cpp.
Lang Hames [Sun, 22 Feb 2015 01:45:31 +0000 (01:45 +0000)]
[Orc] Add header comment to IndirectionUtils.cpp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230153 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIRCE: generalize InductiveRangeCheck::computeSafeIterationSpace to
Sanjoy Das [Sat, 21 Feb 2015 22:20:22 +0000 (22:20 +0000)]
IRCE: generalize InductiveRangeCheck::computeSafeIterationSpace to
work with a non-canonical induction variable.

This is currently a non-functional change because we only ever call
computeSafeIterationSpace on a canonical induction variable; but the
generalization will be useful in a later commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230151 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIRCE: use SCEVs instead of llvm::Value's for intermediate
Sanjoy Das [Sat, 21 Feb 2015 22:07:32 +0000 (22:07 +0000)]
IRCE: use SCEVs instead of llvm::Value's for intermediate
calculations.  Semantically non-functional change.

This gets rid of some of the SCEV -> Value -> SCEV round tripping and
the Construct(SMin|SMax)Of and MaybeSimplify helper routines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230150 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Use v_madmk_f32
Matt Arsenault [Sat, 21 Feb 2015 21:29:10 +0000 (21:29 +0000)]
R600/SI: Use v_madmk_f32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230149 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Try to use v_madak_f32
Matt Arsenault [Sat, 21 Feb 2015 21:29:07 +0000 (21:29 +0000)]
R600/SI: Try to use v_madak_f32

This is a code size optimization when the constant
only has one use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230148 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Don't crash when getting immediate operand size
Matt Arsenault [Sat, 21 Feb 2015 21:29:04 +0000 (21:29 +0000)]
R600/SI: Don't crash when getting immediate operand size

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230147 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Fix mad*k definitions
Matt Arsenault [Sat, 21 Feb 2015 21:29:00 +0000 (21:29 +0000)]
R600/SI: Fix mad*k definitions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230146 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][SSE] Added shuffle based integer zero extension tests.
Simon Pilgrim [Sat, 21 Feb 2015 21:25:16 +0000 (21:25 +0000)]
[X86][SSE] Added shuffle based integer zero extension tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230145 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agodocs: Mention that assertions must be enabled to use the -stats flag
Justin Bogner [Sat, 21 Feb 2015 20:53:36 +0000 (20:53 +0000)]
docs: Mention that assertions must be enabled to use the -stats flag

Patch by Rob Stewart. Thanks!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230144 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Orc] Move Orc code into a namespace (llvm::orc), update Kaleidoscope code.
Lang Hames [Sat, 21 Feb 2015 20:44:36 +0000 (20:44 +0000)]
[Orc] Move Orc code into a namespace (llvm::orc), update Kaleidoscope code.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230143 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMachineInstr: Use range-based for loops. NFC.
Benjamin Kramer [Sat, 21 Feb 2015 17:08:08 +0000 (17:08 +0000)]
MachineInstr: Use range-based for loops. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230142 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoCalling memmove on a MachineOperand is totally safe.
Benjamin Kramer [Sat, 21 Feb 2015 16:22:48 +0000 (16:22 +0000)]
Calling memmove on a MachineOperand is totally safe.

While it's not POD due to the user-defined constructor, it's still a trivially
copyable type. No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230141 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove dead prototype.
Benjamin Kramer [Sat, 21 Feb 2015 14:35:00 +0000 (14:35 +0000)]
Remove dead prototype.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230137 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoX86: Remove custom lowering of SIGN_EXTEND_INREG
Benjamin Kramer [Sat, 21 Feb 2015 14:31:29 +0000 (14:31 +0000)]
X86: Remove custom lowering of SIGN_EXTEND_INREG

This was just replicating logic from the legalizer. Covered by existing

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230136 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUnconditionally create a new MCInstrInfo in the asm printer for
Eric Christopher [Sat, 21 Feb 2015 09:09:15 +0000 (09:09 +0000)]
Unconditionally create a new MCInstrInfo in the asm printer for
asm parsing since it's not subtarget dependent and we can't depend
upon the one hanging off the MachineFunction's subtarget still
being around.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230135 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove obsolete comment.
Eric Christopher [Sat, 21 Feb 2015 08:48:23 +0000 (08:48 +0000)]
Remove obsolete comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230134 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoHave the MipsAsmPrinter fp stub emission code take a custom
Eric Christopher [Sat, 21 Feb 2015 08:48:22 +0000 (08:48 +0000)]
Have the MipsAsmPrinter fp stub emission code take a custom
MCSubtargetInfo as the MachineFunction has gone away and we need
to emit code at the module level.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230133 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTurn an if+llvm_unreachable into an assert and reword comment.
Eric Christopher [Sat, 21 Feb 2015 08:32:38 +0000 (08:32 +0000)]
Turn an if+llvm_unreachable into an assert and reword comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230132 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoEndianness can be gotten from the DataLayout which we already
Eric Christopher [Sat, 21 Feb 2015 08:32:22 +0000 (08:32 +0000)]
Endianness can be gotten from the DataLayout which we already
have. Also, the subtarget is invalid at this point.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230131 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoX86: Call __main using the SelectionDAG
David Majnemer [Sat, 21 Feb 2015 05:49:45 +0000 (05:49 +0000)]
X86: Call __main using the SelectionDAG

Synthesizing a call directly using the MI layer would confuse the frame
lowering code.  This is problematic as frame lowering is highly
sensitive the particularities of calls, etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230129 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[obj2yaml/yaml2obj] Add SHT_GROUP support.
Shankar Easwaran [Sat, 21 Feb 2015 04:28:26 +0000 (04:28 +0000)]
[obj2yaml/yaml2obj] Add SHT_GROUP support.

This adds section group support to the tools obj2yaml and yaml2obj.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230124 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoSmall cleanup. Don't use else when not needed.
Davide Italiano [Sat, 21 Feb 2015 02:36:54 +0000 (02:36 +0000)]
Small cleanup. Don't use else when not needed.
Pointed out by David Majnemer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230122 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoCodeGen: convert CCState interface to using ArrayRefs
Tim Northover [Sat, 21 Feb 2015 02:11:17 +0000 (02:11 +0000)]
CodeGen: convert CCState interface to using ArrayRefs

Everyone except R600 was manually passing the length of a static array
at each callsite, calculated in a variety of interesting ways. Far
easier to let ArrayRef handle that.

There should be no functional change, but out of tree targets may have
to tweak their calls as with these examples.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230118 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoWin64: Stack alignment constraints aren't applied during SET_FPREG
David Majnemer [Sat, 21 Feb 2015 01:04:47 +0000 (01:04 +0000)]
Win64: Stack alignment constraints aren't applied during SET_FPREG

Stack realignment occurs after the prolog, not during, for Win64.
Because of this, don't factor in the maximum stack alignment when
establishing a frame pointer.

This fixes PR22572.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230113 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAsmParser/Writer: Handle symbolic constants in DI 'flags:'
Duncan P. N. Exon Smith [Sat, 21 Feb 2015 01:02:18 +0000 (01:02 +0000)]
AsmParser/Writer: Handle symbolic constants in DI 'flags:'

Parse (and write) symbolic constants in debug info `flags:` fields.
This prevents a readability (and CHECK-ability) regression with the new
debug info hierarchy.

Old (well, current) assembly, with pretty-printing:

    !{!"...\\0016387", ...} ; ... [public] [rvalue reference]

Flags field without this change:

   !MDDerivedType(flags: 16387, ...)

Flags field with this change:

   !MDDerivedType(flags: DIFlagPublic | DIFlagRValueReference, ...)

As discussed in the review thread, this isn't a final state.  Most of
these flags correspond to `DW_AT_` symbolic constants, and we might
eventually want to support arbitrary attributes in some form.  However,
as it stands now, some of the flags correspond to other concepts (like
`FlagStaticMember`); until things are refactored this is the simplest
way to move forward without regressing assembly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230111 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIR: Add helper to split debug info flags bitfield
Duncan P. N. Exon Smith [Sat, 21 Feb 2015 00:45:26 +0000 (00:45 +0000)]
IR: Add helper to split debug info flags bitfield

Split debug info 'flags' bitfield over a vector so the current flags can
be iterated over.  This API (in combination with r230107) will be used
for assembly support for symbolic constants.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230108 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIR: Add debug info flag string conversions
Duncan P. N. Exon Smith [Sat, 21 Feb 2015 00:43:09 +0000 (00:43 +0000)]
IR: Add debug info flag string conversions

Add `DIDescriptor::getFlag(StringRef)` and
`DIDescriptor::getFlagString(unsigned)`.  The latter only converts exact
matches; I'll add separate API for breaking the flags bitfield up into

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230107 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIR: Move DebugInfo Flag* definitions to .def file, NFC
Duncan P. N. Exon Smith [Sat, 21 Feb 2015 00:37:53 +0000 (00:37 +0000)]
IR: Move DebugInfo Flag* definitions to .def file, NFC

This prepares for adding string support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230105 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAsmParser: Use StringRef for keyword comparisons, NFC
Duncan P. N. Exon Smith [Sat, 21 Feb 2015 00:18:40 +0000 (00:18 +0000)]
AsmParser: Use StringRef for keyword comparisons, NFC

Leverage `StringRef` inside keyword comparison macros.  There's no
reason to be so low-level here, and I'm about to add another
`startswith()` use, so let's make it easy to read.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230100 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoSet the datalayout in the gold plugin.
Rafael Espindola [Sat, 21 Feb 2015 00:13:15 +0000 (00:13 +0000)]
Set the datalayout in the gold plugin.

This fixes the gold tests after r230054.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230098 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PlaceSafepoints] Adjust enablement logic to default to off and be GC configurable...
Philip Reames [Sat, 21 Feb 2015 00:09:09 +0000 (00:09 +0000)]
[PlaceSafepoints] Adjust enablement logic to default to off and be GC configurable per GC

Previously, this pass ran over every function in the Module if added to the pass order.  With this change, it runs only over those with a GC attribute where the GC explicitly opts in.  A GC can also choose which of entry safepoint polls, backedge safepoint polls, and call safepoints it wants.  I hope to get these exposed as checks on the GCStrategy at some point, but for now, the checks are manual string comparisons.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230097 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAsmParser: Use do{}while(false) in macros, NFC
Duncan P. N. Exon Smith [Fri, 20 Feb 2015 23:49:24 +0000 (23:49 +0000)]
AsmParser: Use do{}while(false) in macros, NFC

`do { ... } while (false)` is standard macro etiquette for forcing
instantiations into a single statement and requiring a `;` afterwards,
making statement-like macros easier to reason about (and harder to use

I'm about to modify the macros in `LexIdentifier()`.  I noticed that the
`KEYWORD` macro *does* follow the rule, so I thought I'd clean up the
other macros to match (otherwise might not be worth changing, since the
benefits of this pattern are fairly irrelevant here).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230095 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove some unnecessary unreachables in favor of (sometimes implicit) assertions
David Blaikie [Fri, 20 Feb 2015 23:44:24 +0000 (23:44 +0000)]
Remove some unnecessary unreachables in favor of (sometimes implicit) assertions

Also simplify some else-after-return cases including some standard
algorithm convenience/use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230094 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLiveRangeCalc: Don't start liveranges of PHI instruction at the block begin.
Matthias Braun [Fri, 20 Feb 2015 23:43:14 +0000 (23:43 +0000)]
LiveRangeCalc: Don't start liveranges of PHI instruction at the block begin.

Letting them begin at the PHI instruction slightly simplifies the code
but more importantly avoids breaking the assumption that live ranges
starting at the block begin are also live at the end of the predecessor
blocks. The MachineVerifier checks that but was apparently never run in
the few instances where liveranges are calculated for machine-SSA

Reviewers: qcolombet

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D7779

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230093 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoHide a bunch of advanced testing options in default opt --help output
Philip Reames [Fri, 20 Feb 2015 23:32:03 +0000 (23:32 +0000)]
Hide a bunch of advanced testing options in default opt --help output

These are internal options.  I need to go through, evaluate which are worth keeping and which not.  Many of them should probably be renamed as well.  Until I have time to do that, we can at least stop poluting the standard opt -help output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230088 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse short names for jumptable sections.
Rafael Espindola [Fri, 20 Feb 2015 23:28:28 +0000 (23:28 +0000)]
Use short names for jumptable sections.

Also refactor code to remove some duplication.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230087 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "[llvm-pdbdump] Add some tests for llvm-pdbdump."
Zachary Turner [Fri, 20 Feb 2015 23:21:21 +0000 (23:21 +0000)]
Revert "[llvm-pdbdump] Add some tests for llvm-pdbdump."

It is not correctly detecting the situations where the test is
unsupported.  Reverting until we can figure it out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230085 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[RewriteStatepointsForGC] Use DenseSet in place of std::set [NFC]
Philip Reames [Fri, 20 Feb 2015 23:16:52 +0000 (23:16 +0000)]
[RewriteStatepointsForGC] Use DenseSet in place of std::set [NFC]

This should be the last cleanup on non-llvm preferred data structures.  I left one use of std::set in an assertion; DenseSet didn't seem to have a tombstone for CallSite defined.  That might be worth fixing, but wasn't worth it for a debug only use.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230084 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[llvm-pdbdump] Add some tests for llvm-pdbdump.
Zachary Turner [Fri, 20 Feb 2015 23:05:57 +0000 (23:05 +0000)]
[llvm-pdbdump] Add some tests for llvm-pdbdump.

This adds only a very basic set of tests that dump a few
functions and object files.

Differential Revision: http://reviews.llvm.org/D7656
Reviewed By: David Blaikie

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230083 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[RewriteStatepointsForGC] Replace std::map with DenseMap
Philip Reames [Fri, 20 Feb 2015 22:48:20 +0000 (22:48 +0000)]
[RewriteStatepointsForGC] Replace std::map with DenseMap

I'd done the work of extracting the typedef in a previous commit, but didn't actually change it.  Hopefully this will make any subtle changes easier to isolate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230081 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[RewriteStatepointsForGC] Cleanup - replace std::vector usage [NFC]
Philip Reames [Fri, 20 Feb 2015 22:39:41 +0000 (22:39 +0000)]
[RewriteStatepointsForGC] Cleanup - replace std::vector usage [NFC]

Migrate std::vector usage to a combination of SmallVector and ArrayRef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230079 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUsed the cached subtarget off of the MachineFunction.
Eric Christopher [Fri, 20 Feb 2015 22:36:11 +0000 (22:36 +0000)]
Used the cached subtarget off of the MachineFunction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230078 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoX86: Remove pre-2010 dead code in mergeSPUpdatesDown
Reid Kleckner [Fri, 20 Feb 2015 22:13:25 +0000 (22:13 +0000)]
X86: Remove pre-2010 dead code in mergeSPUpdatesDown

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230075 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLowerScalarImmediateShift - Merged v16i8 and v32i8 shift lowering. NFC.
Simon Pilgrim [Fri, 20 Feb 2015 22:13:03 +0000 (22:13 +0000)]
LowerScalarImmediateShift - Merged v16i8 and v32i8 shift lowering. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230074 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Remove v_sub_f64 pseudo
Matt Arsenault [Fri, 20 Feb 2015 22:10:45 +0000 (22:10 +0000)]
R600/SI: Remove v_sub_f64 pseudo

The expansion code does the same thing. Since
the operands were not defined with the correct
types, this has the side effect of fixing operand
folding since the expanded pseudo would never use
SGPRs or inline immediates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230072 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600: Use new fmad node.
Matt Arsenault [Fri, 20 Feb 2015 22:10:41 +0000 (22:10 +0000)]
R600: Use new fmad node.

This enables a few useful combines that used to only
use fma.

Also since v_mad_f32 apparently does not support denormals,
disable the existing cases that are custom handled if they are

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230071 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd generic fmad DAG node.
Matt Arsenault [Fri, 20 Feb 2015 22:10:33 +0000 (22:10 +0000)]
Add generic fmad DAG node.

This allows sharing of FMA forming combines to work
with instructions that have the same semantics as a separate
multiply and add.

This is expand by default, and only formed post legalization
so it shouldn't have much impact on targets that do not want it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230070 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[RewriteStatepointsForGC] More style cleanup [NFC]
Philip Reames [Fri, 20 Feb 2015 22:05:18 +0000 (22:05 +0000)]
[RewriteStatepointsForGC] More style cleanup [NFC]

Use llvm_unreachable where appropriate, use SmallVector where easy to do so, introduce typedefs for planned type migrations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230068 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[RewriteStatepointsForGC] Remove notion of SafepointBounds [NFC]
Philip Reames [Fri, 20 Feb 2015 21:34:11 +0000 (21:34 +0000)]
[RewriteStatepointsForGC] Remove notion of SafepointBounds [NFC]

The notion of a range of inserted safepoint related code is no longer really applicable.  This survived over from an earlier implementation.  Just saving the inserted gc.statepoint and working from that is far clearer given the current code structure.  Particularly when invokable statepoints get involved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230063 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRaising minimum required CMake version to
Chris Bieneman [Fri, 20 Feb 2015 21:28:18 +0000 (21:28 +0000)]
Raising minimum required CMake version to

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230062 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGrab the DataLayout off of the TargetMachine since that's where
Eric Christopher [Fri, 20 Feb 2015 20:56:39 +0000 (20:56 +0000)]
Grab the DataLayout off of the TargetMachine since that's where
it's stored.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230059 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLoopRotate: When reconstructing loop simplify form don't split edges from indirectbrs.
Benjamin Kramer [Fri, 20 Feb 2015 20:49:25 +0000 (20:49 +0000)]
LoopRotate: When reconstructing loop simplify form don't split edges from indirectbrs.

Yet another chapter in the endless story. While this looks like we leave
the loop in a non-canonical state this replicates the logic in
LoopSimplify so it doesn't diverge from the canonical form in any way.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230058 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIR: Change MDFile to directly store the filename/directory
Duncan P. N. Exon Smith [Fri, 20 Feb 2015 20:35:17 +0000 (20:35 +0000)]
IR: Change MDFile to directly store the filename/directory

In the old (well, current) schema, there are two types of file
references: untagged and tagged (the latter references the former).

    !0 = !{!"filename", !"/directory"}
    !1 = !{!"0x29", !1} ; DW_TAG_file_type [filename] [/directory]

The interface to `DIBuilder` universally takes the tagged version,
described by `DIFile`.  However, most `file:` references actually use
the untagged version directly.

In the new hierarchy, I'm merging this into a single node: `MDFile`.

Originally I'd planned to keep the old schema unchanged until after I
moved the new hierarchy into place.

However, it turns out to be trivial to make `MDFile` match both nodes at
the same time.

  - Anyone referencing !1 does so through `DIFile`, whose implementation
    I need to gut anyway (as I do the rest of the `DIDescriptor`s).
  - Anyone referencing !0 just references an `MDNode`, and expects a
    node with two `MDString` operands.

This commit achieves that, and updates all the testcases for the parts
of the new hierarchy that used the two-node schema (I've replaced the
untagged nodes with `distinct !{}` to make the diff clear (otherwise the
metadata all gets renumbered); it might be worthwhile to come back and
delete those nodes and renumber the world, not sure).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230057 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIntroduce bitset metadata format and bitset lowering pass.
Peter Collingbourne [Fri, 20 Feb 2015 20:30:47 +0000 (20:30 +0000)]
Introduce bitset metadata format and bitset lowering pass.

This patch introduces a new mechanism that allows IR modules to co-operatively
build pointer sets corresponding to addresses within a given set of
globals. One particular use case for this is to allow a C++ program to
efficiently verify (at each call site) that a vtable pointer is in the set
of valid vtable pointers for the class or its derived classes. One way of
doing this is for a toolchain component to build, for each class, a bit set
that maps to the memory region allocated for the vtables, such that each 1
bit in the bit set maps to a valid vtable for that class, and lay out the
vtables next to each other, to minimize the total size of the bit sets.

The patch introduces a metadata format for representing pointer sets, an
'@llvm.bitset.test' intrinsic and an LTO lowering pass that lays out the globals
and builds the bitsets, and documents the new feature.

Differential Revision: http://reviews.llvm.org/D7288

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230054 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoReversed revision 229706. The reason is regression, which is caused by the
Jozef Kolek [Fri, 20 Feb 2015 20:26:52 +0000 (20:26 +0000)]
Reversed revision 229706. The reason is regression, which is caused by the
usage of instruction ADDU16 by CodeGen. For this instruction an improper
register is allocated, i.e. the register that is not from register set defined
for the instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230053 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoVerifier: Unused comdats might not have a corresponding GV
David Majnemer [Fri, 20 Feb 2015 19:58:48 +0000 (19:58 +0000)]
Verifier: Unused comdats might not have a corresponding GV

This fixes PR22646.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230051 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix an asan use-after-free bug introduced by the asm printer
Eric Christopher [Fri, 20 Feb 2015 19:54:07 +0000 (19:54 +0000)]
Fix an asan use-after-free bug introduced by the asm printer
changes to remove non-Function based subtargets out of the asm
printer. For module level emission we'll need to construct up
an MCSubtargetInfo so that we can encode instructions for

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230050 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[GC, RewriteStatepointsForGC] Style cleanup and bug fix
Philip Reames [Fri, 20 Feb 2015 19:51:56 +0000 (19:51 +0000)]
[GC, RewriteStatepointsForGC] Style cleanup and bug fix

When doing style cleanup, I noticed a minor bug in this code.  If we have a pointer that we think is unused after a statepoint and thus doesn't need relocation, we store a null pointer into the alloca we're about to promote.  This helps turn a mistake in liveness analysis into an easily debuggable crash.  It turned out this code had never been updated to handle invoke statepoints.

There's no test for this.  Without a bug in liveness, it appears impossible to make this trigger in a way which is visible in the resulting IR.  We might store the null, but when promoting the alloca, there will be no uses and thus nothing to test against.  Suggestions on how to test are very welcome.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230047 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse unreachable instead of assert(false) to silence MSVC warning
Reid Kleckner [Fri, 20 Feb 2015 19:46:02 +0000 (19:46 +0000)]
Use unreachable instead of assert(false) to silence MSVC warning

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230045 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][FastIsel] Teach how to select float-half conversion intrinsics.
Andrea Di Biagio [Fri, 20 Feb 2015 19:37:14 +0000 (19:37 +0000)]
[X86][FastIsel] Teach how to select float-half conversion intrinsics.

This patch teaches X86FastISel how to select intrinsic 'convert_from_fp16' and
intrinsic 'convert_to_fp16'.
If the target has F16C, we can select VCVTPS2PHrr for a float-half conversion,
and VCVTPH2PSrr for a half-float conversion.

Differential Revision: http://reviews.llvm.org/D7673

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230043 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[GC] Style cleanup for RewriteStatepointForGC (1 of many) [NFC]
Philip Reames [Fri, 20 Feb 2015 19:26:04 +0000 (19:26 +0000)]
[GC] Style cleanup for RewriteStatepointForGC (1 of many) [NFC]

Starting to update variable naming and types to match LLVM style.  This will be an incremental process to minimize the chance of breakage as I work.  Step one, rename member variables to LLVM CamelCase and use llvm's ADT.  Much more to come.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230042 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoSetting up CMake to default to Debug when no build type is specified.
Chris Bieneman [Fri, 20 Feb 2015 19:02:59 +0000 (19:02 +0000)]
Setting up CMake to default to Debug when no build type is specified.

Summary: Turns out if you don't set CMAKE_BUILD_TYPE the default is an empty string. This results in some of the behaviors of debug builds, but not all of them. For example ENABLE_ASSERTIONS is false.

Reviewers: rnk

Reviewed By: rnk

Subscribers: chapuni, llvm-commits

Differential Revision: http://reviews.llvm.org/D7360

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230041 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoBugfix for 229954
Philip Reames [Fri, 20 Feb 2015 18:56:14 +0000 (18:56 +0000)]
Bugfix for 229954

Before calling Function::getGC to test for enablement, we need to make sure there's actually a GC at all via Function::hasGC.  Otherwise, we'd crash on functions without a GC.  Thankfully, this only mattered if you manually scheduled the pass, but still, oops. :(

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230040 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove a use of the Subtarget in the darwin ppc asm printer.
Eric Christopher [Fri, 20 Feb 2015 18:53:42 +0000 (18:53 +0000)]
Remove a use of the Subtarget in the darwin ppc asm printer.
EmitFunctionStubs is called from doFinalization and so can't
depend on the Subtarget existing. It's also irrelevant as
we know we're darwin since we're in the darwin asm printer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230039 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGet the function specific subtarget.
Eric Christopher [Fri, 20 Feb 2015 18:44:17 +0000 (18:44 +0000)]
Get the function specific subtarget.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230038 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGet the cached subtarget off the MachineFunction rather than
Eric Christopher [Fri, 20 Feb 2015 18:44:15 +0000 (18:44 +0000)]
Get the cached subtarget off the MachineFunction rather than
inquiring for a new one from the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230037 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agocanonicalize a v2f64 blendi of 2 registers
Sanjay Patel [Fri, 20 Feb 2015 16:55:27 +0000 (16:55 +0000)]
canonicalize a v2f64 blendi of 2 registers

This canonicalization step saves us 3 pattern matching possibilities * 4 math ops
for scalar FP math that uses xmm regs. The backend can re-commute the operands
post-instruction-selection if that makes register allocation better.

The tests in llvm/test/CodeGen/X86/sse-scalar-fp-arith.ll cover this scenario already,
so there are no new tests with this patch.

Differential Revision: http://reviews.llvm.org/D7777

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230024 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoPut MSVC back into the dumb compiler's corner.
Benjamin Kramer [Fri, 20 Feb 2015 16:35:42 +0000 (16:35 +0000)]
Put MSVC back into the dumb compiler's corner.

It fails to compile std::trivially_copyable for forward-declared enums.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230023 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoBase isPodLike on is_trivially_copyable for GCC 5 and MSVC
Benjamin Kramer [Fri, 20 Feb 2015 16:19:28 +0000 (16:19 +0000)]
Base isPodLike on is_trivially_copyable for GCC 5 and MSVC

It would be nice to get rid of the version checks here, but that will
have to wait until libstdc++ is upgraded to 5.0 everywhere ...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230021 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoI incorrectly marked the VORC instruction as isCommutable when I added it.
Kit Barton [Fri, 20 Feb 2015 15:54:58 +0000 (15:54 +0000)]
I incorrectly marked the VORC instruction as isCommutable when I added it.
This fix removes the VORC instruction definition from the isCommutable block.

Phabricator review: http://reviews.llvm.org/D7772

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230020 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGeneralize statepoint lowering to use ImmutableStatepoint. Move statepoint lowering...
Igor Laevsky [Fri, 20 Feb 2015 15:28:35 +0000 (15:28 +0000)]
Generalize statepoint lowering to use ImmutableStatepoint. Move statepoint lowering into a separate function 'LowerStatepoint' which uses ImmutableStatepoint instead of a CallInst. Also related utility functions are changed to receive ImmutableCallSite.

Differential Revision: http://reviews.llvm.org/D7756

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230017 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoConstants.cpp: Only read 32 bits for float.
Benjamin Kramer [Fri, 20 Feb 2015 15:11:55 +0000 (15:11 +0000)]
Constants.cpp: Only read 32 bits for float.

Otherwise we'll discard the wrong half of a uint64_t on big-endian systems.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230016 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoConstants.cpp: getElementAsAPFloat(): Don't handle constant value via host's float...
NAKAMURA Takumi [Fri, 20 Feb 2015 14:24:49 +0000 (14:24 +0000)]
Constants.cpp: getElementAsAPFloat(): Don't handle constant value via host's float/double, just handle with APInt/APFloat.

x87 FPU didn't keep SNAN, but demoted to QNAN.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230013 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRewriteStatepointsForGC: Move details into anonymous namespaces. NFC.
Benjamin Kramer [Fri, 20 Feb 2015 14:00:58 +0000 (14:00 +0000)]
RewriteStatepointsForGC: Move details into anonymous namespaces. NFC.

While there reduce the number of duplicated std::map lookups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230012 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMake the static instance of None just const.
Benjamin Kramer [Fri, 20 Feb 2015 13:16:05 +0000 (13:16 +0000)]
Make the static instance of None just const.

This way there shouldn't be any unused variable warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230010 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoWrap recursive function only used in assert in #ifndef NDEBUG.
Benjamin Kramer [Fri, 20 Feb 2015 13:15:49 +0000 (13:15 +0000)]
Wrap recursive function only used in assert in #ifndef NDEBUG.

Avoids unused function warnings in Release builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230009 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] Switching the shuffle equivalence test to a variadic template was
Chandler Carruth [Fri, 20 Feb 2015 10:47:28 +0000 (10:47 +0000)]
[x86] Switching the shuffle equivalence test to a variadic template was
the wrong answer. We also got initializer lists which are *way* cleaner
for this kind of thing. Let's use those and make this a normal, boring
functionn accepting ArrayRef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230004 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix wording and grammar in Mips subtarget options.
Eric Christopher [Fri, 20 Feb 2015 08:42:34 +0000 (08:42 +0000)]
Fix wording and grammar in Mips subtarget options.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230001 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGet the cached subtarget off the MachineFunction rather than
Eric Christopher [Fri, 20 Feb 2015 08:39:06 +0000 (08:39 +0000)]
Get the cached subtarget off the MachineFunction rather than
inquiring for a new one from the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230000 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGet the cached subtarget off the MachineFunction rather than
Eric Christopher [Fri, 20 Feb 2015 08:24:37 +0000 (08:24 +0000)]
Get the cached subtarget off the MachineFunction rather than
inquiring for a new one from the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229999 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGet the cached subtarget off the MachineFunction rather than
Eric Christopher [Fri, 20 Feb 2015 08:24:34 +0000 (08:24 +0000)]
Get the cached subtarget off the MachineFunction rather than
inquiring for a new one from the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229998 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoSave the MachineFunction in startFunction so that we can use it for
Eric Christopher [Fri, 20 Feb 2015 08:01:55 +0000 (08:01 +0000)]
Save the MachineFunction in startFunction so that we can use it for
lookups of the subtarget later.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229996 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse the cached subtarget from the MachineFunction rather than
Eric Christopher [Fri, 20 Feb 2015 08:01:52 +0000 (08:01 +0000)]
Use the cached subtarget from the MachineFunction rather than
doing a lookup on the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229995 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMake the TargetMachine::getSubtarget that takes a Function argument
Eric Christopher [Fri, 20 Feb 2015 07:32:59 +0000 (07:32 +0000)]
Make the TargetMachine::getSubtarget that takes a Function argument
take a reference to match the getSubtargetImpl that takes a Function

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229994 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDisallow implicit conversions from None to integer types
Justin Bogner [Fri, 20 Feb 2015 07:28:28 +0000 (07:28 +0000)]
Disallow implicit conversions from None to integer types

This fixes an error introduced in r228934 where None was converted to
an int instead of the int being converted to an Optional as intended.
We make that sort of mistake a compile error by changing NoneType into
a scoped enum.

Finally, provide a static NoneType called None to avoid forcing all
users to spell it NoneType::None.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229980 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix build with gcc. This has a -Wsequence-point error on 'MII', which is a good point.
Nick Lewycky [Fri, 20 Feb 2015 07:17:40 +0000 (07:17 +0000)]
Fix build with gcc. This has a -Wsequence-point error on 'MII', which is a good point.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229979 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove more uses of TargetMachine::getSubtargetImpl from the
Eric Christopher [Fri, 20 Feb 2015 07:16:19 +0000 (07:16 +0000)]
Remove more uses of TargetMachine::getSubtargetImpl from the

getSubtargetInfo now asserts that the MachineFunction exists.
Debug printing of register naming now uses the register info
from MCAsmInfo as that's unchanging.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229978 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix build in release mode, -Wunused-variable on this lambda function used only in...
Nick Lewycky [Fri, 20 Feb 2015 07:16:17 +0000 (07:16 +0000)]
Fix build in release mode, -Wunused-variable on this lambda function used only in an assert.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229977 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix build in release mode, four cases of -Wunused-variable.
Nick Lewycky [Fri, 20 Feb 2015 07:14:02 +0000 (07:14 +0000)]
Fix build in release mode, four cases of -Wunused-variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229976 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAsmPrinter::doFinalization is at the module level and so doesn't
Eric Christopher [Fri, 20 Feb 2015 06:59:48 +0000 (06:59 +0000)]
AsmPrinter::doFinalization is at the module level and so doesn't
have access to a target specific subtarget info. Grab the module
level MCSubtargetInfo for the JumpInstrTable output stubs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229974 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Orc] Add a new JITSymbol constructor to build a symbol from an existing address.
Lang Hames [Fri, 20 Feb 2015 06:48:29 +0000 (06:48 +0000)]
[Orc] Add a new JITSymbol constructor to build a symbol from an existing address.

This constructor is more efficient for symbols that have already been emitted,
since it avoids the construction/execution of a std::function.

Update the ObjectLinkingLayer to use this new constructor where possible.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229973 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove the MCInstrInfo cached variable as it was only used in a
Eric Christopher [Fri, 20 Feb 2015 06:35:21 +0000 (06:35 +0000)]
Remove the MCInstrInfo cached variable as it was only used in a
single place and replace calls to getSubtargetImpl with calls
to get the subtarget from the MachineFunction where valid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229971 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix -Wunused-variable warning in non-asserts build, and optimize a little bit while...
David Blaikie [Fri, 20 Feb 2015 06:28:38 +0000 (06:28 +0000)]
Fix -Wunused-variable warning in non-asserts build, and optimize a little bit while I'm here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229970 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PowerPC] Loop Data Prefetching for the BG/Q
Hal Finkel [Fri, 20 Feb 2015 05:08:21 +0000 (05:08 +0000)]
[PowerPC] Loop Data Prefetching for the BG/Q

The IBM BG/Q supercomputer's A2 cores have a hardware prefetching unit, the
L1P, but it does not prefetch directly into the A2's L1 cache. Instead, it
prefetches into its own L1P buffer, and the latency to access that buffer is
significantly higher than that to the L1 cache (although smaller than the
latency to the L2 cache). As a result, especially when multiple hardware
threads are not actively busy, explicitly prefetching data into the L1 cache is

I've been using this pass out-of-tree for data prefetching on the BG/Q for well
over a year, and it has worked quite well. It is enabled by default only for
the BG/Q, but can be enabled for other cores as well via a command-line option.

Eventually, we might want to add some TTI interfaces and move this into
Transforms/Scalar (there is nothing particularly target dependent about it,
although only machines like the BG/Q will benefit from its simplistic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229966 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] Remove the old vector shuffle lowering code and its flag.
Chandler Carruth [Fri, 20 Feb 2015 04:25:04 +0000 (04:25 +0000)]
[x86] Remove the old vector shuffle lowering code and its flag.

The new shuffle lowering has been the default for some time. I've
enabled the new legality testing by default with no really blocking
regressions. I've fuzz tested this very heavily (many millions of fuzz
test cases have passed at this point). And this cleans up a ton of code.

Thanks again to the many folks that helped with this transition. There
was a lot of work by others that went into the new shuffle lowering to
make it really excellent.

In case you aren't using a diff algorithm that can handle this:
  X86ISelLowering.cpp: 22 insertions(+), 2940 deletions(-)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229964 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] Now that the new vector shuffle legality is enabled and everything
Chandler Carruth [Fri, 20 Feb 2015 03:59:35 +0000 (03:59 +0000)]
[x86] Now that the new vector shuffle legality is enabled and everything
is going well, remove the flag and the code for the old legality tests.

This is the first step toward removing the entire old vector shuffle
lowering. *Much* more code to delete coming up next.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229963 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoBitcode: Stop assuming non-null fields
Duncan P. N. Exon Smith [Fri, 20 Feb 2015 03:17:58 +0000 (03:17 +0000)]
Bitcode: Stop assuming non-null fields

When writing the bitcode serialization for the new debug info hierarchy,
I assumed two fields would never be null.

Drop that assumption, since it's brittle (and crashes the
`BitcodeWriter` if wrong), and is a check better left for the verifier
anyway.  (No need for a bitcode upgrade here, since the new hierarchy is
still not in place.)

The fields in question are `MDCompileUnit::getFile()` and
`MDDerivedType::getBaseType()`, the latter of which isn't null in
test/Transforms/Mem2Reg/ConvertDebugInfo2.ll (see !14, a pointer to
nothing).  While the testcase might have bitrotted, there's no reason
for the bitcode format to rely on non-null for metadata operands.

This also fixes a bug in `AsmWriter` where if the `file:` is null it
isn't emitted (caught by the double-round trip in the testcase I'm
adding) -- this is a required field in `LLParser`.

I'll circle back to ConvertDebugInfo2.  Once the specialized nodes are
in place, I'll be trying to turn the debug info verifier back on by
default (in the newer module pass form committed r206300) and throwing
more logic in there.  If the testcase has bitrotted (as opposed to me
not understanding the schema correctly) I'll fix it then.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229960 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] Remove unnecessary variable indexing into single-element arrays
Hal Finkel [Fri, 20 Feb 2015 03:05:53 +0000 (03:05 +0000)]
[InstCombine] Remove unnecessary variable indexing into single-element arrays

This change addresses a deficiency pointed out in PR22629. To copy from the bug

[from the bug report]

Consider this code:

int f(int x) {
  int a[] = {12};
  return a[x];

GCC knows to optimize this to

movl     $12, %eax

The code generated by recent Clang at -O3 is:

movslq   %edi, %rax
movl     .L_ZZ1fiE1a(,%rax,4), %eax

  .long    12                      # 0xc

[end from the bug report]

This definitely seems worth fixing. I've also seen this kind of code before (as
the base case of generic vector wrapper templates with one element).

The general idea is to look at the GEP feeding a load or a store, which has
some variable as its first non-zero index, and determine if that index must be
zero (or else an out-of-bounds access would occur). We can do this for allocas
and globals with constant initializers where we know the maximum size of the
underlying object. When we find such a GEP, we create a new one for the memory
access with that first variable index replaced with a constant zero.

Even if we can't eliminate the memory access (and sometimes we can't), it is
still useful because it removes unnecessary indexing calculations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229959 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[x86] Make the new vector shuffle legality test on by default, which
Chandler Carruth [Fri, 20 Feb 2015 03:05:47 +0000 (03:05 +0000)]
[x86] Make the new vector shuffle legality test on by default, which
reflects the fact that the x86 backend can in fact lower any shuffle you
want it to with reasonably high code quality.

My recent work on the new vector shuffle has made this regress *very*
little. The diff in the test cases makes me very, very happy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229958 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[fuzzer] one more experimental search mode: -use_coverage_pairs=1
Kostya Serebryany [Fri, 20 Feb 2015 03:02:37 +0000 (03:02 +0000)]
[fuzzer] one more experimental search mode: -use_coverage_pairs=1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229957 91177308-0d34-0410-b5e6-96231b3b80d8