oota-llvm.git
11 years agoSimplify test cases for Atom preferring call register indirect over
Preston Gurd [Tue, 2 Apr 2013 14:25:06 +0000 (14:25 +0000)]
Simplify test cases for Atom preferring call register indirect over
call memory indirect (32 and 64 bit).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178541 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[NVPTX] Fix a few style issues in NVVMReflect
Justin Holewinski [Tue, 2 Apr 2013 12:37:11 +0000 (12:37 +0000)]
[NVPTX] Fix a few style issues in NVVMReflect

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178536 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse a worklist to avoid a sneaky iterator invalidation.
Bill Wendling [Tue, 2 Apr 2013 08:16:45 +0000 (08:16 +0000)]
Use a worklist to avoid a sneaky iterator invalidation.

The iterator could be invalidated when it's recursively deleting a whole bunch
of constant expressions in a constant initializer.

Note: This was only reproducible if `opt' was run on a `.bc' file. If `opt' was
run on a `.ll' file, it wouldn't crash. This is why the test first pushes the
`.ll' file through `llvm-as' before feeding it to `opt'.

PR15440

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178531 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd 64-bit load and store instructions.
Jakob Stoklund Olesen [Tue, 2 Apr 2013 04:09:28 +0000 (04:09 +0000)]
Add 64-bit load and store instructions.

There is only a few new instructions, the rest is handled with patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178528 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoBasic 64-bit ALU operations.
Jakob Stoklund Olesen [Tue, 2 Apr 2013 04:09:23 +0000 (04:09 +0000)]
Basic 64-bit ALU operations.

SPARC v9 extends all ALU instructions to 64 bits, so we simply need to
add patterns to use them for both i32 and i64 values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178527 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMaterialize 64-bit immediates.
Jakob Stoklund Olesen [Tue, 2 Apr 2013 04:09:17 +0000 (04:09 +0000)]
Materialize 64-bit immediates.

The last resort pattern produces 6 instructions, and there are still
opportunities for materializing some immediates in fewer instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178526 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd 64-bit shift instructions.
Jakob Stoklund Olesen [Tue, 2 Apr 2013 04:09:12 +0000 (04:09 +0000)]
Add 64-bit shift instructions.

SPARC v9 defines new 64-bit shift instructions. The 32-bit shift right
instructions are still usable as zero and sign extensions.

This adds new F3_Sr and F3_Si instruction formats that probably should
be used for the 32-bit shifts as well. They don't really encode an
simm13 field.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178525 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd predicates for distinguishing 32-bit and 64-bit modes.
Jakob Stoklund Olesen [Tue, 2 Apr 2013 04:09:06 +0000 (04:09 +0000)]
Add predicates for distinguishing 32-bit and 64-bit modes.

The 'sparc' architecture produces 32-bit code while 'sparcv9' produces
64-bit code.

It is also possible to run 32-bit code using SPARC v9 instructions with:

  llc -march=sparc -mattr=+v9

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178524 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support for 64-bit calling convention.
Jakob Stoklund Olesen [Tue, 2 Apr 2013 04:09:02 +0000 (04:09 +0000)]
Add support for 64-bit calling convention.

This is far from complete, but it is enough to make it possible to write
test cases using i64 arguments.

Missing features:
- Floating point arguments.
- Receiving arguments on the stack.
- Calls.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178523 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd an I64Regs register class for 64-bit registers.
Jakob Stoklund Olesen [Tue, 2 Apr 2013 04:08:54 +0000 (04:08 +0000)]
Add an I64Regs register class for 64-bit registers.

We are going to use the same registers for 32-bit and 64-bit values, but
in two different register classes. The I64Regs register class has a
larger spill size and alignment.

The addition of an i64 register class confuses TableGen's type
inference, so it is necessary to clarify the type of some immediates and
the G0 register.

In 64-bit mode, pointers are i64 and should use the I64Regs register
class. Implement getPointerRegClass() to dynamically provide the pointer
register class depending on the subtarget. Use ptr_rc and iPTR for
memory operands.

Finally, add the i64 type to the IntRegs register class. This register
class is not used to hold i64 values, I64Regs is for that. The type is
required to appease TableGen's type checking in output patterns like this:

  def : Pat<(add i64:$a, i64:$b), (ADDrr $a, $b)>;

SPARC v9 uses the same ADDrr instruction for i32 and i64 additions, and
TableGen doesn't know to check the type of register sub-classes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178522 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix typo in PPCISelLowering
Hal Finkel [Tue, 2 Apr 2013 03:29:51 +0000 (03:29 +0000)]
Fix typo in PPCISelLowering

Thanks to Bill Schmidt for finding this in review of r178480.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178521 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoThe divide unit is not pipeline, but it is still buffered.
Andrew Trick [Tue, 2 Apr 2013 01:58:47 +0000 (01:58 +0000)]
The divide unit is not pipeline, but it is still buffered.

Buffered means a later divide may be executed out-of-order while a
prior divide is sitting (buffered) in a reservation station.

You can tell it's not pipelined, because operations that use it
reserve it for more than one cycle:

def : WriteRes<WriteIDiv, [HWPort0, HWDivider]> {
  let Latency = 25;
  let ResourceCycles = [1, 10];
}

We don't currently distinguish between an unpipeline operation and one
that is split into multiple micro-ops requiring the same unit. Except
that the later may have NumMicroOps > 1 if they also consume
issue/dispatch resources.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178519 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agounindent the file to follow coding standards, change class doc comment
Chris Lattner [Mon, 1 Apr 2013 23:00:01 +0000 (23:00 +0000)]
unindent the file to follow coding standards, change class doc comment
to be correct.  No functionality or behavior change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178511 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTarget/R600: Fix CMake build to add missing files.
NAKAMURA Takumi [Mon, 1 Apr 2013 22:05:58 +0000 (22:05 +0000)]
Target/R600: Fix CMake build to add missing files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178508 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMips direct object exception handling regression
Jack Carter [Mon, 1 Apr 2013 21:55:15 +0000 (21:55 +0000)]
Mips direct object exception handling regression

Revision 177141 caused a regression in all but
mips64 little endian. That is because none of the
other Mips targets had test cases checking the
contents of the .eh_frame section. This patch fixes
both the llvm code and adds an assembler test case
to include the current 4 flavors.

The test cases unfortunately rely on llvm-objdump. A
preferable method would be to use a pretty printer output
such as what readelf -wf <elf_file> would give.

I also changed the name of the test case to correct a typo.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178506 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Add support for native control flow
Vincent Lejeune [Mon, 1 Apr 2013 21:48:05 +0000 (21:48 +0000)]
R600: Add support for native control flow

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178505 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600/SI: Share code recording ShaderTypeAttribute between generations
Vincent Lejeune [Mon, 1 Apr 2013 21:47:53 +0000 (21:47 +0000)]
R600/SI: Share code recording ShaderTypeAttribute between generations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178504 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Emit CF_ALU and use true kcache register.
Vincent Lejeune [Mon, 1 Apr 2013 21:47:42 +0000 (21:47 +0000)]
R600: Emit CF_ALU and use true kcache register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178503 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix top-comment header and some indentation
Eli Bendersky [Mon, 1 Apr 2013 19:47:56 +0000 (19:47 +0000)]
Fix top-comment header and some indentation

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178492 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix a bad assert in PPCTargetLowering
Hal Finkel [Mon, 1 Apr 2013 18:42:58 +0000 (18:42 +0000)]
Fix a bad assert in PPCTargetLowering

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178489 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd triple to test/CodeGen/PowerPC/stfiwx-2
Hal Finkel [Mon, 1 Apr 2013 18:18:44 +0000 (18:18 +0000)]
Add triple to test/CodeGen/PowerPC/stfiwx-2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178486 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoCorrect assertion condition
Shuxin Yang [Mon, 1 Apr 2013 18:13:05 +0000 (18:13 +0000)]
Correct assertion condition

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178484 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMerge load/store sequences with adresses: base + index + offset
Arnold Schwaighofer [Mon, 1 Apr 2013 18:12:58 +0000 (18:12 +0000)]
Merge load/store sequences with adresses: base + index + offset

We would also like to merge sequences that involve a variable index like in the
example below.

    int index = *idx++
    int i0 = c[index+0];
    int i1 = c[index+1];
    b[0] = i0;
    b[1] = i1;

By extending the parsing of the base pointer to handle dags that contain a
base, index, and offset we can handle examples like the one above.

The dag for the code above will look something like:

 (load (i64 add (i64 copyfromreg %c)
                (i64 signextend (i8 load %index))))

 (load (i64 add (i64 copyfromreg %c)
                (i64 signextend (i32 add (i32 signextend (i8 load %index))
                                         (i32 1)))))

The code that parses the tree ignores the intermediate sign extensions. However,
if there is a sign extension it needs to be on all indexes.

 (load (i64 add (i64 copyfromreg %c)
                (i64 signextend (add (i8 load %index)
                                     (i8 1))))
 vs

 (load (i64 add (i64 copyfromreg %c)
                (i64 signextend (i32 add (i32 signextend (i8 load %index))
                                         (i32 1)))))
radar://13536387

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178483 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd more PPC floating-point conversion instructions
Hal Finkel [Mon, 1 Apr 2013 17:52:07 +0000 (17:52 +0000)]
Add more PPC floating-point conversion instructions

The P7 and A2 have additional floating-point conversion instructions which
allow a direct two-instruction sequence (plus load/store) to convert from all
combinations (signed/unsigned i32/i64) <--> (float/double) (on previous cores,
only some combinations were directly available).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178480 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse ImmToIdxMap.count in PPCRegisterInfo
Hal Finkel [Mon, 1 Apr 2013 17:02:06 +0000 (17:02 +0000)]
Use ImmToIdxMap.count in PPCRegisterInfo

Code improvement suggested by Jakob (in review of r178450). No functionality
change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178473 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix PowerPC/cttz.ll to specify a cpu (and use FileCheck)
Hal Finkel [Mon, 1 Apr 2013 16:31:56 +0000 (16:31 +0000)]
Fix PowerPC/cttz.ll to specify a cpu (and use FileCheck)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178472 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd the PPC popcntw instruction
Hal Finkel [Mon, 1 Apr 2013 15:58:15 +0000 (15:58 +0000)]
Add the PPC popcntw instruction

The popcntw instruction is available whenever the popcntd instruction is
available, and performs a separate popcnt on the lower and upper 32-bits.
Ignoring the high-order count, this can be used for the 32-bit input case
(saving on the explicit zero extension otherwise required to use popcntd).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178470 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support for vector data types in the LLVM interpreter.
Nadav Rotem [Mon, 1 Apr 2013 15:53:30 +0000 (15:53 +0000)]
Add support for vector data types in the LLVM interpreter.

Patch by:
Veselov, Yuri <Yuri.Veselov@intel.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178469 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTreat PPCISD::STFIWX like the memory opcode that it is
Hal Finkel [Mon, 1 Apr 2013 15:37:53 +0000 (15:37 +0000)]
Treat PPCISD::STFIWX like the memory opcode that it is

PPCISD::STFIWX is really a memory opcode, and so it should come after
FIRST_TARGET_MEMORY_OPCODE, and we should use DAG.getMemIntrinsicNode to create
nodes using it.

No functionality change intended (although there could be optimization benefits
from preserving the MMO information).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178468 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove unused typedef.
Duncan Sands [Mon, 1 Apr 2013 13:46:15 +0000 (13:46 +0000)]
Remove unused typedef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178462 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoARM Scheduler Model: Add resources instructions, map resources in subtargets
Arnold Schwaighofer [Mon, 1 Apr 2013 13:07:05 +0000 (13:07 +0000)]
ARM Scheduler Model: Add resources instructions, map resources in subtargets

Reapply r177968:
After commit 178074 we can now have undefined scheduler variants.

Move the CortexA9 resources into the CortexA9 SchedModel namespace. Define
resource mappings under the CortexA9 SchedModel. Define resources and mappings
for the SwiftModel.

Incooperate Andrew's feedback.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178460 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoX86TTI: Add accurate costs for itofp operations, based on the actual instruction...
Benjamin Kramer [Mon, 1 Apr 2013 10:23:49 +0000 (10:23 +0000)]
X86TTI: Add accurate costs for itofp operations, based on the actual instruction counts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178459 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoWhitespace cleanup
Joe Abbey [Mon, 1 Apr 2013 02:28:07 +0000 (02:28 +0000)]
Whitespace cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178454 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Emit native instructions for tex
Vincent Lejeune [Sun, 31 Mar 2013 19:33:04 +0000 (19:33 +0000)]
R600: Emit native instructions for tex

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178452 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoThere is no longer any need to silence this compiler warning as the warning has
Duncan Sands [Sun, 31 Mar 2013 17:44:09 +0000 (17:44 +0000)]
There is no longer any need to silence this compiler warning as the warning has
been turned off globally.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178451 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoCleanup ImmToIdxMap and noImmForm in PPCRegisterInfo
Hal Finkel [Sun, 31 Mar 2013 14:43:31 +0000 (14:43 +0000)]
Cleanup ImmToIdxMap and noImmForm in PPCRegisterInfo

ImmToIdxMap should be a DenseMap (not a std::map) because there
is no ordering requirement. Also, we don't need a separate list
of instructions for noImmForm in eliminateFrameIndex, because this
list is essentially the complement of the keys in ImmToIdxMap.

No functionality change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178450 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoX86: Promote sitofp <8 x i16> to <8 x i32> when AVX is available.
Benjamin Kramer [Sun, 31 Mar 2013 12:49:15 +0000 (12:49 +0000)]
X86: Promote sitofp <8 x i16> to <8 x i32> when AVX is available.

A vector sext + sitofp is a lot cheaper than 8 scalar conversions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178448 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd the PPC lfiwax instruction
Hal Finkel [Sun, 31 Mar 2013 10:12:51 +0000 (10:12 +0000)]
Add the PPC lfiwax instruction

This instruction is available on modern PPC64 CPUs, and is now used
to improve the SINT_TO_FP lowering (by eliminating the need for the
separate sign extension instruction and decreasing the amount of
needed stack space).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178446 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoCleanup PPC(64) i32 -> float/double conversion
Hal Finkel [Sun, 31 Mar 2013 01:58:02 +0000 (01:58 +0000)]
Cleanup PPC(64) i32 -> float/double conversion

The existing SINT_TO_FP code for i32 -> float/double conversion was disabled
because it relied on broken EXTSW_32/STD_32 instruction definitions. The
original intent had been to enable these 64-bit instructions to be used on CPUs
that support them even in 32-bit mode.  Unfortunately, this form of lying to
the infrastructure was buggy (as explained in the FIXME comment) and had
therefore been disabled.

This re-enables this functionality, using regular DAG nodes, but only when
compiling in 64-bit mode. The old STD_32/EXTSW_32 definitions (which were dead)
are removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178438 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDAGCombine: visitXOR can replace a node without returning it, bail out in that case.
Benjamin Kramer [Sat, 30 Mar 2013 21:28:18 +0000 (21:28 +0000)]
DAGCombine: visitXOR can replace a node without returning it, bail out in that case.

Fixes the crash reported in PR15608.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178429 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd start of user documentation for NVPTX
Justin Holewinski [Sat, 30 Mar 2013 16:41:14 +0000 (16:41 +0000)]
Add start of user documentation for NVPTX

Summary: This is the beginning of user documentation for the NVPTX back-end.  I want to ensure I am integrating this properly into the rest of the LLVM documentation.

Differential Revision: http://llvm-reviews.chandlerc.com/D600

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178428 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoChange '@SECREL' suffix to GAS-compatible '@SECREL32'.
Benjamin Kramer [Sat, 30 Mar 2013 16:21:50 +0000 (16:21 +0000)]
Change '@SECREL' suffix to GAS-compatible '@SECREL32'.

'@SECREL' is what is used by the Microsoft assembler, but GNU as expects '@SECREL32'.
With the patch, the MC-generated code works fine in combination with a recent GNU as (2.23.51.20120920 here).

Patch by David Nadlinger!
Differential Revision: http://llvm-reviews.chandlerc.com/D429

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178427 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[docs] llvmbugs is not the place for patches.
Sean Silva [Sat, 30 Mar 2013 15:33:02 +0000 (15:33 +0000)]
[docs] llvmbugs is not the place for patches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178426 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[docs] Annotate mailing lists with their "name".
Sean Silva [Sat, 30 Mar 2013 15:33:01 +0000 (15:33 +0000)]
[docs] Annotate mailing lists with their "name".

Nobody says "the developer's list" or "commits archive"; they always say
"llvmdev" or "llvm-commits". It makes sense for our documentation to
at least make that association explicitly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178425 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[docs] Reorganize mailing lists.
Sean Silva [Sat, 30 Mar 2013 15:32:54 +0000 (15:32 +0000)]
[docs] Reorganize mailing lists.

Order them roughly by "which one should a newbie join first".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178424 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[docs] Pull IRC and Mailing Lists under a new "Community" heading.
Sean Silva [Sat, 30 Mar 2013 15:32:51 +0000 (15:32 +0000)]
[docs] Pull IRC and Mailing Lists under a new "Community" heading.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178423 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[docs] The GEP FAQ is not "design and overview"
Sean Silva [Sat, 30 Mar 2013 15:32:50 +0000 (15:32 +0000)]
[docs] The GEP FAQ is not "design and overview"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178422 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[docs] Put DeveloperPolicy under "Development Process Documentation"
Sean Silva [Sat, 30 Mar 2013 15:32:47 +0000 (15:32 +0000)]
[docs] Put DeveloperPolicy under "Development Process Documentation"

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178421 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoPut private class into an anonmyous namespace.
Benjamin Kramer [Sat, 30 Mar 2013 15:23:08 +0000 (15:23 +0000)]
Put private class into an anonmyous namespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178420 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[NVPTX] Remove support for SM < 2.0. This was never fully supported anyway.
Justin Holewinski [Sat, 30 Mar 2013 14:29:30 +0000 (14:29 +0000)]
[NVPTX] Remove support for SM < 2.0.  This was never fully supported anyway.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178417 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[NVPTX] Add NVVMReflect pass to allow compile-time selection of
Justin Holewinski [Sat, 30 Mar 2013 14:29:25 +0000 (14:29 +0000)]
[NVPTX] Add NVVMReflect pass to allow compile-time selection of
specific code paths.

This allows us to write code like:

  if (__nvvm_reflect("FOO"))
    // Do something
  else
    // Do something else

and compile into a library, then give "FOO" a value at kernel
compile-time so the check becomes a no-op.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178416 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[NVPTX] Run clang-format on all NVPTX sources.
Justin Holewinski [Sat, 30 Mar 2013 14:29:21 +0000 (14:29 +0000)]
[NVPTX] Run clang-format on all NVPTX sources.

Hopefully this resolves any outstanding style issues and gives us
an automated way of ensuring we conform to the style guidelines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178415 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoObject: Turn a couple of degenerate for loops into while loops.
Benjamin Kramer [Sat, 30 Mar 2013 13:07:51 +0000 (13:07 +0000)]
Object: Turn a couple of degenerate for loops into while loops.

No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178413 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoImplement XOR reassociation. It is based on following rules:
Shuxin Yang [Sat, 30 Mar 2013 02:15:01 +0000 (02:15 +0000)]
Implement XOR reassociation. It is based on following rules:

  rule 1: (x | c1) ^ c2 => (x & ~c1) ^ (c1^c2),
     only useful when c1=c2
  rule 2: (x & c1) ^ (x & c2) = (x & (c1^c2))
  rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2
  rule 4: (x | c1) ^ (x & c2) => (x & c3) ^ c1, where c3 = ~c1 ^ c2

 It reduces an application's size (in terms of # of instructions) by 8.9%.
 Reviwed by Pete Cooper. Thanks a lot!

 rdar://13212115

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178409 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Add patterns for DSP indexed load instructions.
Akira Hatanaka [Sat, 30 Mar 2013 02:14:45 +0000 (02:14 +0000)]
[mips] Add patterns for DSP indexed load instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178408 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Define reg+imm load/store pattern templates.
Akira Hatanaka [Sat, 30 Mar 2013 02:01:48 +0000 (02:01 +0000)]
[mips] Define reg+imm load/store pattern templates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178407 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Fix DSP instructions to have explicit accumulator register operands.
Akira Hatanaka [Sat, 30 Mar 2013 01:58:00 +0000 (01:58 +0000)]
[mips] Fix DSP instructions to have explicit accumulator register operands.

Check that instruction selection can select multiply-add/sub DSP instructions
from a pattern that doesn't have intrinsics.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178406 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove unused variables.
Akira Hatanaka [Sat, 30 Mar 2013 01:46:28 +0000 (01:46 +0000)]
Remove unused variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178405 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Move the code which does dag-combine for multiply-add/sub nodes to
Akira Hatanaka [Sat, 30 Mar 2013 01:42:24 +0000 (01:42 +0000)]
[mips] Move the code which does dag-combine for multiply-add/sub nodes to
derived class MipsSETargetLowering.

We shouldn't be generating madd/msub nodes if target is Mips16, since Mips16
doesn't have support for multipy-add/sub instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178404 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Fix definitions of multiply, multiply-add/sub and divide instructions.
Akira Hatanaka [Sat, 30 Mar 2013 01:36:35 +0000 (01:36 +0000)]
[mips] Fix definitions of multiply, multiply-add/sub and divide instructions.

The new instructions have explicit register output operands and use table-gen
patterns instead of C++ code to do instruction selection.

Mips16's instructions are unaffected by this change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178403 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Remove function getFPBranchCodeFromCond. Rename invertFPCondCodeAdd.
Akira Hatanaka [Sat, 30 Mar 2013 01:16:38 +0000 (01:16 +0000)]
[mips] Remove function getFPBranchCodeFromCond. Rename invertFPCondCodeAdd.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178396 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix indentation.
Akira Hatanaka [Sat, 30 Mar 2013 01:15:17 +0000 (01:15 +0000)]
Fix indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178395 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Add mips-specific nodes which will be used to select multiply and divide
Akira Hatanaka [Sat, 30 Mar 2013 01:14:04 +0000 (01:14 +0000)]
[mips] Add mips-specific nodes which will be used to select multiply and divide
instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178394 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Implement getRepRegClassFor in MipsSETargetLowering. This function is
Akira Hatanaka [Sat, 30 Mar 2013 01:12:05 +0000 (01:12 +0000)]
[mips] Implement getRepRegClassFor in MipsSETargetLowering. This function is
called in several places in ScheduleDAGRRList.cpp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178393 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Fix MipsSEInstrInfo::copyPhysReg, loadRegFromStack and storeRegToStack
Akira Hatanaka [Sat, 30 Mar 2013 01:08:05 +0000 (01:08 +0000)]
[mips] Fix MipsSEInstrInfo::copyPhysReg, loadRegFromStack and storeRegToStack
to handle accumulator registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178392 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Expand pseudo load, store and copy instructions right before
Akira Hatanaka [Sat, 30 Mar 2013 01:04:11 +0000 (01:04 +0000)]
[mips] Expand pseudo load, store and copy instructions right before
callee-saved scan.

The code makes use of register's scavenger's capability to spill multiple
registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178391 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Define pseudo instructions for spilling and copying accumulator
Akira Hatanaka [Sat, 30 Mar 2013 00:54:52 +0000 (00:54 +0000)]
[mips] Define pseudo instructions for spilling and copying accumulator
registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178390 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse SmallVectorImpl instead of SmallVector at the uses.
Eric Christopher [Fri, 29 Mar 2013 23:34:06 +0000 (23:34 +0000)]
Use SmallVectorImpl instead of SmallVector at the uses.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178386 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRun the ObjCARCContract pass for LTO. <rdar://problem/13538084>
Bob Wilson [Fri, 29 Mar 2013 23:28:55 +0000 (23:28 +0000)]
Run the ObjCARCContract pass for LTO. <rdar://problem/13538084>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178385 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUpdated test0 of retain-not-declared.ll to reflect the fact that objc-arc-expand...
Michael Gottesman [Fri, 29 Mar 2013 22:44:59 +0000 (22:44 +0000)]
Updated test0 of retain-not-declared.ll to reflect the fact that objc-arc-expand runs before objc-arc/objc-arc-contract.

Specifically, objc-arc-expand will make sure that the
objc_retainAutoreleasedReturnValue, objc_autoreleaseReturnValue, and ret
will all have %call as an argument.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178382 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSmallVector and SmallPtrSet allocations now power-of-two aligned.
Jean-Luc Duprat [Fri, 29 Mar 2013 22:07:12 +0000 (22:07 +0000)]
SmallVector and SmallPtrSet allocations now power-of-two aligned.
This time tested on both OSX and Linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178377 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[docs] The STL "binary search" has a non-obvious name.
Sean Silva [Fri, 29 Mar 2013 21:57:47 +0000 (21:57 +0000)]
[docs] The STL "binary search" has a non-obvious name.

std::lower_bound is the canonical "binary search" in the STL
(std::binary_search generally is not what you want). The name actually
makes a lot of sense (and also has a beautiful symmetry with the
std::upper_bound algorithm). The name is nonetheless non-obvious.

Also, remove mention of "radix search". It's not even clear how that
would work in the context of a sorted vector. AFAIK "radix search" only
makes sense when you have a trie-like data structure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178376 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoExclude the X86/complex-fca.ll test at it probably wasn't supposed to work on Windows
Timur Iskhodzhanov [Fri, 29 Mar 2013 21:54:00 +0000 (21:54 +0000)]
Exclude the X86/complex-fca.ll test at it probably wasn't supposed to work on Windows

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178375 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd clang.arc.used to ModuleHasARC so ARC always runs if said call is present in...
Michael Gottesman [Fri, 29 Mar 2013 21:15:23 +0000 (21:15 +0000)]
Add clang.arc.used to ModuleHasARC so ARC always runs if said call is present in a module.

clang.arc.used is an interesting call for ARC since ObjCARCContract
needs to run to remove said intrinsic to avoid a linker error (since the
call does not exist).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178369 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoHexagon: Add emitFrameIndexDebugValue function to emit debug information.
Jyotsna Verma [Fri, 29 Mar 2013 21:09:53 +0000 (21:09 +0000)]
Hexagon: Add emitFrameIndexDebugValue function to emit debug information.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178368 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse 12 as the magic number for our abbreviation data and our
Eric Christopher [Fri, 29 Mar 2013 20:23:06 +0000 (20:23 +0000)]
Use 12 as the magic number for our abbreviation data and our
die values. A lot of DIEs have 10 attributes in C++ code (example
clang), none had more than 12. Seems like a good default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178366 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMove the construction of the skeleton compile unit after the
Eric Christopher [Fri, 29 Mar 2013 20:23:02 +0000 (20:23 +0000)]
Move the construction of the skeleton compile unit after the
entire original compile unit has been constructed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178365 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agomove testcase into appropriate X86 subdirectory.
Adrian Prantl [Fri, 29 Mar 2013 20:14:08 +0000 (20:14 +0000)]
move testcase into appropriate X86 subdirectory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178364 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoImplement FRINT lowering on PPC using frin
Hal Finkel [Fri, 29 Mar 2013 19:41:55 +0000 (19:41 +0000)]
Implement FRINT lowering on PPC using frin

Like nearbyint, rint can be implemented on PPC using the frin instruction. The
complication comes from the fact that rint needs to set the FE_INEXACT flag
when the result does not equal the input value (and frin does not do that). As
a result, we use a custom inserter which, after the rounding, compares the
rounded value with the original, and if they differ, explicitly sets the XX bit
in the FPSCR register (which corresponds to FE_INEXACT).

Once LLVM has better modeling of the floating-point environment we should be
able to (often) eliminate this extra complexity.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178362 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Define a function which returns the GPR register class.
Akira Hatanaka [Fri, 29 Mar 2013 19:17:42 +0000 (19:17 +0000)]
[mips] Define a function which returns the GPR register class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178359 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix TableGen subtarget-emitter to handle A9/Swift.
Andrew Trick [Fri, 29 Mar 2013 19:08:31 +0000 (19:08 +0000)]
Fix TableGen subtarget-emitter to handle A9/Swift.

A9 uses itinerary classes, Swift uses RW lists. This tripped some
verification when we're expanding variants. I had to refine the
verification a bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178357 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoBuild fixes for STLPort + GCC
Matt Arsenault [Fri, 29 Mar 2013 18:48:45 +0000 (18:48 +0000)]
Build fixes for STLPort + GCC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178356 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix loop style
Matt Arsenault [Fri, 29 Mar 2013 18:48:42 +0000 (18:48 +0000)]
Fix loop style

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178355 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSplit the llvm/tools/clang/test/CodeGenObjC/debug-info-blocks.m testcase into a CFE...
Adrian Prantl [Fri, 29 Mar 2013 18:08:14 +0000 (18:08 +0000)]
Split the llvm/tools/clang/test/CodeGenObjC/debug-info-blocks.m testcase into a CFE and LLVM part.
rdar://problem/12767564

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178353 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove the old CodePlacementOpt pass.
Benjamin Kramer [Fri, 29 Mar 2013 17:14:24 +0000 (17:14 +0000)]
Remove the old CodePlacementOpt pass.

It was superseded by MachineBlockPlacement and disabled by default since LLVM 3.1.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178349 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix a typo
Nadav Rotem [Fri, 29 Mar 2013 16:34:23 +0000 (16:34 +0000)]
Fix a typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178346 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoHexagon: Disable DwarfUsesInlineInfoSection flag.
Jyotsna Verma [Fri, 29 Mar 2013 15:46:12 +0000 (15:46 +0000)]
Hexagon: Disable DwarfUsesInlineInfoSection flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178345 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd PPC FP rounding instructions fri[mnpz]
Hal Finkel [Fri, 29 Mar 2013 08:57:48 +0000 (08:57 +0000)]
Add PPC FP rounding instructions fri[mnpz]

These instructions are available on the P5x (and later) and on the A2. They
implement the standard floating-point rounding operations (floor, trunc, etc.).
One caveat: frin (round to nearest) does not implement "ties to even", and so
is only enabled in fast-math mode.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178337 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert "Fix allocations of SmallVector and SmallPtrSet so they are more prone to"
Rafael Espindola [Fri, 29 Mar 2013 07:11:21 +0000 (07:11 +0000)]
Revert "Fix allocations of SmallVector and SmallPtrSet so they are more prone to"

This reverts commit 617330909f0c26a3f2ab8601a029b9bdca48aa61.

It broke the bots:

/home/clangbuild2/clang-ppc64-2/llvm.src/unittests/ADT/SmallVectorTest.cpp:150: PushPopTest
/home/clangbuild2/clang-ppc64-2/llvm.src/unittests/ADT/SmallVectorTest.cpp:118: Failure
Value of: v[i].getValue()
  Actual: 0
Expected: value
Which is: 2

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178334 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix allocations of SmallVector and SmallPtrSet so they are more prone to
Jean-Luc Duprat [Fri, 29 Mar 2013 05:45:22 +0000 (05:45 +0000)]
Fix allocations of SmallVector and SmallPtrSet so they are more prone to
being power-of-two sized.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178332 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemoved trailing whitespace.
Michael Gottesman [Fri, 29 Mar 2013 05:13:07 +0000 (05:13 +0000)]
Removed trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178329 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Change type of accumulator registers to Untyped. Add two more accumulator
Akira Hatanaka [Fri, 29 Mar 2013 03:27:21 +0000 (03:27 +0000)]
[mips] Change type of accumulator registers to Untyped. Add two more accumulator
register classes for Mips64 and DSP-ASE.

No functionality changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178328 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Define overloaded versions of storeRegToStack and loadRegFromStack.
Akira Hatanaka [Fri, 29 Mar 2013 02:14:12 +0000 (02:14 +0000)]
[mips] Define overloaded versions of storeRegToStack and loadRegFromStack.

No functionality changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178327 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Add parameter Alignment to MipsFrameLowering's constructor.
Akira Hatanaka [Fri, 29 Mar 2013 01:51:04 +0000 (01:51 +0000)]
[mips] Add parameter Alignment to MipsFrameLowering's constructor.

No functionality changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178326 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert r178166. According to Howard, this code is actually ok.
Dan Gohman [Fri, 29 Mar 2013 00:13:08 +0000 (00:13 +0000)]
Revert r178166. According to Howard, this code is actually ok.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178319 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[Mips Assembler] Add support for OR macro with imediate opperand
Jack Carter [Thu, 28 Mar 2013 23:45:13 +0000 (23:45 +0000)]
[Mips Assembler] Add support for OR macro with imediate opperand

Mips assembler supports macros that allows the OR instruction
to have an immediate parameter. This patch adds an instruction
alias that converts this macro into a Mips ORI instruction.

Contributer: Vladimir Medic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178316 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support of RDSEED defined in AVX2 extension
Michael Liao [Thu, 28 Mar 2013 23:41:26 +0000 (23:41 +0000)]
Add support of RDSEED defined in AVX2 extension

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178314 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoEnhance boolean simplification to handle 16-/64-bit RDRAND
Michael Liao [Thu, 28 Mar 2013 23:38:52 +0000 (23:38 +0000)]
Enhance boolean simplification to handle 16-/64-bit RDRAND

- RDRAND always clears the destination value when a random value is not
  available (i.e. CF == 0). This value is truncated or zero-extended as
  the false boolean value to be returned. Boolean simplification needs
  to skip this 'zext' or 'trunc' node.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178312 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSkip moving call address loading into callseq when targets prefer register indirect...
Michael Liao [Thu, 28 Mar 2013 23:13:21 +0000 (23:13 +0000)]
Skip moving call address loading into callseq when targets prefer register indirect call.

To enable a load of a call address to be folded with that call, this
load is moved from outside of callseq into callseq. Such a moving
adds a non-glued node (that load) into a glued sequence. This non-glue
load is only removed when DAG selection folds them into a memory form
call instruction. When such instruction selection is disabled, it breaks
DAG schedule.

To prevent that, such moving is disabled when target favors register
indirect call.

Previous workaround disabling CALL32m/CALL64m insn selection is removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178308 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemoved dead code from ObjCARCOpts relating to tracking objc_retainBlocks through...
Michael Gottesman [Thu, 28 Mar 2013 23:08:44 +0000 (23:08 +0000)]
Removed dead code from ObjCARCOpts relating to tracking objc_retainBlocks through the ARC Dataflow analysis. By the time we get to the ARC dataflow analysis, any objc_retainBlock calls are not optimizable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178306 91177308-0d34-0410-b5e6-96231b3b80d8