oota-llvm.git
7 years ago[x86 asm] allow fwait alias in both At&t and Intel modes (PR21208)
Hans Wennborg [Tue, 14 Oct 2014 21:41:17 +0000 (21:41 +0000)]
[x86 asm] allow fwait alias in both At&t and Intel modes (PR21208)

Differential Revision: http://reviews.llvm.org/D5741

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219725 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoARM: set preferred aggregate alignment to 32 universally.
Tim Northover [Tue, 14 Oct 2014 20:57:26 +0000 (20:57 +0000)]
ARM: set preferred aggregate alignment to 32 universally.

Before, ARM and Thumb mode code had different preferred alignments, which could
lead to some rather unexpected results. There's justification for reducing it
from the default 64-bits (wasted space), but I don't think there is for going
below 32-bits.

There's no actual ABI change here, just to reassure people.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219719 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[CFL-AA] CFL-AA should not assert on an va_arg instruction
Hal Finkel [Tue, 14 Oct 2014 20:51:26 +0000 (20:51 +0000)]
[CFL-AA] CFL-AA should not assert on an va_arg instruction

The CFL-AA implementation was missing a visit* routine for va_arg instructions,
causing it to assert when run on a function that had one. For now, handle these
in a conservative way.

Fixes PR20954.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219718 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoOptimize away fabs() calls when input is squared (known positive).
Sanjay Patel [Tue, 14 Oct 2014 20:43:11 +0000 (20:43 +0000)]
Optimize away fabs() calls when input is squared (known positive).

Eliminate library calls and intrinsic calls to fabs when the input
is a squared value.

Note that no unsafe-math / fast-math assumptions are needed for
this optimization.

Differential Revision: http://reviews.llvm.org/D5777

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219717 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[FastISel][AArch64] Fix sign-/zero-extend folding when SelectionDAG is involved.
Juergen Ributzka [Tue, 14 Oct 2014 20:36:02 +0000 (20:36 +0000)]
[FastISel][AArch64] Fix sign-/zero-extend folding when SelectionDAG is involved.

Sign-/zero-extend folding depended on the load and the integer extend to be
both selected by FastISel. This cannot always be garantueed and SelectionDAG
might interfer. This commit adds additonal checks to load and integer extend
lowering to catch this.

Related to rdar://problem/18495928.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219716 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstCombine: Don't miscompile X % ((Pow2 << A) >>u B)
David Majnemer [Tue, 14 Oct 2014 20:28:40 +0000 (20:28 +0000)]
InstCombine: Don't miscompile X % ((Pow2 << A) >>u B)

We assumed that A must be greater than B because the right hand side of
a remainder operator must be nonzero.

However, it is possible for A to be less than B if Pow2 is a power of
two greater than 1.

Take for example:
i32 %A = 0
i32 %B = 31
i32 Pow2 = 2147483648

((Pow2 << 0) >>u 31) is non-zero but A is less than B.

This fixes PR21274.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219713 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoReapply "R600: Add new intrinsic to read work dimensions"
Jan Vesely [Tue, 14 Oct 2014 20:05:26 +0000 (20:05 +0000)]
Reapply "R600: Add new intrinsic to read work dimensions"

This effectively reverts revert 219707. After fixing the test to work with
new function name format and renamed intrinsic.

Reviewed-by: Tom Stellard <tom@stellard.net>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219710 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "r216914 - Revert: [APFloat] Fixed a bug in method 'fusedMultiplyAdd'"
Hal Finkel [Tue, 14 Oct 2014 19:23:07 +0000 (19:23 +0000)]
Revert "r216914 - Revert: [APFloat] Fixed a bug in method 'fusedMultiplyAdd'"

Reapply r216913, a fix for PR20832 by Andrea Di Biagio. The commit was reverted
because of buildbot failures, and credit goes to Ulrich Weigand for isolating
the underlying issue (which can be confirmed by Valgrind, which does helpfully
light up like the fourth of July). Uli explained the problem with the original
patch as:

  It seems the problem is calling multiplySignificand with an addend of category
  fcZero; that is not expected by this routine.  Note that for fcZero, the
  significand parts are simply uninitialized, but the code in (or rather, called
  from) multiplySignificand will unconditionally access them -- in effect using
  uninitialized contents.

This version avoids using a category == fcZero addend within
multiplySignificand, which avoids this problem (the Valgrind output is also now
clean).

Original commit message:

[APFloat] Fixed a bug in method 'fusedMultiplyAdd'.

When folding a fused multiply-add builtin call, make sure that we propagate the
correct result in the case where the addend is zero, and the two other operands
are finite non-zero.

Example:
  define double @test() {
    %1 = call double @llvm.fma.f64(double 7.0, double 8.0, double 0.0)
    ret double %1
  }

Before this patch, the instruction simplifier wrongly folded the builtin call
in function @test to constant 'double 7.0'.
With this patch, method 'fusedMultiplyAdd' correctly evaluates the multiply and
propagates the expected result (i.e. 56.0).

Added test fold-builtin-fma.ll with the reproducible from PR20832 plus extra
test cases to verify the behavior of method 'fusedMultiplyAdd' in the presence
of NaN/Inf operands.

This fixes PR20832.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219708 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "R600: Add new intrinsic to read work dimensions"
Rafael Espindola [Tue, 14 Oct 2014 18:58:04 +0000 (18:58 +0000)]
Revert "R600: Add new intrinsic to read work dimensions"

This reverts commit r219705.

CodeGen/R600/work-item-intrinsics.ll was failing on linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219707 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove unused member variable.
Rafael Espindola [Tue, 14 Oct 2014 18:53:16 +0000 (18:53 +0000)]
Remove unused member variable.

Fixes pr20904.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219706 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600: Add new intrinsic to read work dimensions
Jan Vesely [Tue, 14 Oct 2014 18:52:07 +0000 (18:52 +0000)]
R600: Add new intrinsic to read work dimensions

v2: Add SI lowering
    Add test

v3: Place work dimensions after the kernel arguments.
v4: Calculate offset while lowering arguments
v5: rebase
v6: change prefix to AMDGPU

Reviewed-by: Tom Stellard <tom@stellard.net>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219705 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600: FMA is VecALU only instruction
Jan Vesely [Tue, 14 Oct 2014 18:52:04 +0000 (18:52 +0000)]
R600: FMA is VecALU only instruction

Reviewed-by: Tom Stellard <tom@stellard.net>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219704 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFinish getting Mips fast-isel to match up with AArch64 fast-isel
Reed Kotler [Tue, 14 Oct 2014 18:27:58 +0000 (18:27 +0000)]
Finish getting Mips fast-isel to match up with AArch64 fast-isel

Summary:
In order to facilitate use of common code, checking by reviewers of other fast-isel ports, and hopefully to eventually move most of Mips and other fast-isel ports into target independent code, I've tried to get the two implementations to line up.

There is no functional code change. Just methods moved in the file to be in the same order as in AArch64.

Test Plan: No functional change.

Reviewers: dsanders

Reviewed By: dsanders

Subscribers: llvm-commits, aemerson, rfuhler

Differential Revision: http://reviews.llvm.org/D5692

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219703 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDebugInfo: Ensure that all debug location scope chains from instructions within a...
David Blaikie [Tue, 14 Oct 2014 18:22:52 +0000 (18:22 +0000)]
DebugInfo: Ensure that all debug location scope chains from instructions within a function, lead to the function itself.

Let me tell you a tale...

Originally committed in r211723 after discovering a nasty case of weird
scoping due to inlining, this was reverted in r211724 after it fired in
ASan/compiler-rt.

(minor diversion where I accidentally committed/reverted again in
r211871/r211873)

After further testing and fixing bugs in ArgumentPromotion (r211872) and
Inlining (r212065) it was recommitted in r212085. Reverted in r212089
after the sanitizer buildbots still showed problems.

Fixed another bug in ArgumentPromotion (r212128) found by this
assertion.

Recommitted in r212205, reverted in r212226 after it crashed some more
on sanitizer buildbots.

Fix clang some more in r212761.

Recommitted in r212776, reverted in r212793. ASan failures.
Recommitted in r213391, reverted in r213432, trying to reproduce flakey
ASan build failure.

Fixed bugs in r213805 (ArgPromo + DebugInfo), r213952
(LiveDebugVariables strips dbg_value intrinsics in functions not
described by debug info).

Recommitted in r214761, reverted in r214999, flakey failure on Windows
buildbot.

Fixed DeadArgElimination + DebugInfo bug in r219210.

Recommitted in r219215, reverted in r219512, failure on ObjC++ atomic
properties in the test-suite on Darwin.

Fixed ObjC++ atomic properties issue in Clang in r219690.

[This commit is provided 'as is' with no hope that this is the last time
I commit this change either expressed or implied]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219702 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove method that is identical to the base class one.
Rafael Espindola [Tue, 14 Oct 2014 17:38:38 +0000 (17:38 +0000)]
Remove method that is identical to the base class one.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219700 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Use DS offsets for constant addresses
Matt Arsenault [Tue, 14 Oct 2014 17:21:19 +0000 (17:21 +0000)]
R600/SI: Use DS offsets for constant addresses

Use 0 as the base address for a constant address, so if
we have a constant address we can save moves and form
read2/write2s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219698 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "Fix stuff... again."
David Blaikie [Tue, 14 Oct 2014 17:13:09 +0000 (17:13 +0000)]
Revert "Fix stuff... again."

Accidental commit.

This reverts commit r219693.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219695 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert some parts of r196288 that were confusing and untested.
David Blaikie [Tue, 14 Oct 2014 17:12:02 +0000 (17:12 +0000)]
Revert some parts of r196288 that were confusing and untested.

If we figure out why they should be here, let's add some testing of some
kind so we can better demonstrate why it's needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219694 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix stuff... again.
David Blaikie [Tue, 14 Oct 2014 17:11:59 +0000 (17:11 +0000)]
Fix stuff... again.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219693 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LVI] Check for @llvm.assume dominating the edge branch
Hal Finkel [Tue, 14 Oct 2014 16:04:49 +0000 (16:04 +0000)]
[LVI] Check for @llvm.assume dominating the edge branch

When LazyValueInfo uses @llvm.assume intrinsics to provide edge-value
constraints, we should check for intrinsics that dominate the edge's branch,
not just any potential context instructions. An assumption that dominates the
edge's branch represents a truth on that edge. This is specifically useful, for
example, if multiple predecessors assume a pointer to be nonnull, allowing us
to simplify a later null comparison.

The test case, and an initial patch, were provided by Philip Reames. Thanks!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219688 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r219638, (r219640 and r219676), "Removing the static destructor from ManagedSt...
NAKAMURA Takumi [Tue, 14 Oct 2014 15:58:16 +0000 (15:58 +0000)]
Revert r219638, (r219640 and r219676), "Removing the static destructor from ManagedStatic.cpp by controlling the allocation and de-allocation of the mutex."

It caused hang-up on msc17 builder, probably deadlock.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219687 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX512] Extended avx512_binop_rm to DQ/VL subsets.
Robert Khasanov [Tue, 14 Oct 2014 15:13:56 +0000 (15:13 +0000)]
[AVX512] Extended avx512_binop_rm to DQ/VL subsets.
Added encoding tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219686 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX512] Extended avx512_binop_rm to BW/VL subsets.
Robert Khasanov [Tue, 14 Oct 2014 14:36:19 +0000 (14:36 +0000)]
[AVX512] Extended avx512_binop_rm to BW/VL subsets.
Added encoding tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219685 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64] Fix crash with empty/pseudo-only blocks in A53 erratum (835769) workaround
Bradley Smith [Tue, 14 Oct 2014 14:02:41 +0000 (14:02 +0000)]
[AArch64] Fix crash with empty/pseudo-only blocks in A53 erratum (835769) workaround

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219684 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[llvm-symbolizer] Minor typedef cleanup. NFC.
Alexander Potapenko [Tue, 14 Oct 2014 13:40:44 +0000 (13:40 +0000)]
[llvm-symbolizer] Minor typedef cleanup. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219682 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoThreading.h: Use \tparam for template parameters. [-Wdocumentation]
NAKAMURA Takumi [Tue, 14 Oct 2014 09:34:16 +0000 (09:34 +0000)]
Threading.h: Use \tparam for template parameters. [-Wdocumentation]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219676 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGrab the subtarget info off of the MachineFunction rather than
Eric Christopher [Tue, 14 Oct 2014 08:44:19 +0000 (08:44 +0000)]
Grab the subtarget info off of the MachineFunction rather than
indirecting through the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219674 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse the triple to figure out if this is a darwin target, not
Eric Christopher [Tue, 14 Oct 2014 08:25:26 +0000 (08:25 +0000)]
Use the triple to figure out if this is a darwin target, not
the subtarget.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219673 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove unnecessary TargetMachine.h includes.
Eric Christopher [Tue, 14 Oct 2014 07:22:08 +0000 (07:22 +0000)]
Remove unnecessary TargetMachine.h includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219672 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGrab the subtarget and subtarget dependent variables off of
Eric Christopher [Tue, 14 Oct 2014 07:22:00 +0000 (07:22 +0000)]
Grab the subtarget and subtarget dependent variables off of
MachineFunction rather than TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219671 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGrab the subtarget and subtarget dependent variables off of
Eric Christopher [Tue, 14 Oct 2014 07:17:23 +0000 (07:17 +0000)]
Grab the subtarget and subtarget dependent variables off of
MachineFunction rather than TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219670 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstead of the TargetMachine cache the MachineFunction
Eric Christopher [Tue, 14 Oct 2014 07:17:20 +0000 (07:17 +0000)]
Instead of the TargetMachine cache the MachineFunction
and TargetRegisterInfo in the peephole optimizer. This
makes it easier to grab subtarget dependent variables off
of the MachineFunction rather than the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219669 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAccess subtarget specific variables off of the MachineFunction's
Eric Christopher [Tue, 14 Oct 2014 07:00:33 +0000 (07:00 +0000)]
Access subtarget specific variables off of the MachineFunction's
cached subtarget and not the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219668 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd lld to the parallel set of directories since it doesn't depend
Eric Christopher [Tue, 14 Oct 2014 06:56:28 +0000 (06:56 +0000)]
Add lld to the parallel set of directories since it doesn't depend
on any of the other tools directories.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219667 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAccess the subtarget off of the MachineFunction via the DAG
Eric Christopher [Tue, 14 Oct 2014 06:56:25 +0000 (06:56 +0000)]
Access the subtarget off of the MachineFunction via the DAG
scheduler or via the SelectionDAG if available. Otherwise
grab the subtarget off of the MachineFunction by going up
the parent chain.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219666 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64]Select wide immediate offset into [Base+XReg] addressing mode
Hao Liu [Tue, 14 Oct 2014 06:50:36 +0000 (06:50 +0000)]
[AArch64]Select wide immediate offset into [Base+XReg] addressing mode
e.g Currently we'll generate following instructions if the immediate is too wide:
    MOV  X0, WideImmediate
    ADD  X1, BaseReg, X0
    LDR  X2, [X1, 0]

    Using [Base+XReg] addressing mode can save one ADD as following:
    MOV  X0, WideImmediate
    LDR  X2, [BaseReg, X0]

    Differential Revision: http://reviews.llvm.org/D5477

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219665 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove the use and member variable of the TargetMachine from
Eric Christopher [Tue, 14 Oct 2014 06:26:57 +0000 (06:26 +0000)]
Remove the use and member variable of the TargetMachine from
MachineLICM as we can get the same data off of the MachineFunction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219663 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoHave MachineInstrBundle use the MachineFunction for subtarget
Eric Christopher [Tue, 14 Oct 2014 06:26:55 +0000 (06:26 +0000)]
Have MachineInstrBundle use the MachineFunction for subtarget
access rather than the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219662 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAccess the subtarget off of the MachineFunction rather than
Eric Christopher [Tue, 14 Oct 2014 06:26:53 +0000 (06:26 +0000)]
Access the subtarget off of the MachineFunction rather than
through the TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219661 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoSwitch to select optimization for two-case switches
Marcello Maggioni [Tue, 14 Oct 2014 01:58:26 +0000 (01:58 +0000)]
Switch to select optimization for two-case switches

This is the same optimization of r219233 with modifications to support PHIs with multiple incoming edges from the same block
and a test to check that this condition is handled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219656 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDon't include DFAPacketizer in TargetInstrInfo, there's no reason.
Eric Christopher [Tue, 14 Oct 2014 01:13:53 +0000 (01:13 +0000)]
Don't include DFAPacketizer in TargetInstrInfo, there's no reason.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219653 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInclude map into the A15SDOptimizer rather than pick it up
Eric Christopher [Tue, 14 Oct 2014 01:13:51 +0000 (01:13 +0000)]
Include map into the A15SDOptimizer rather than pick it up
transitively from the DFAPacketizer via TargetInstrInfo.h.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219652 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove the TargetMachine from DFAPacketizer since it was only
Eric Christopher [Tue, 14 Oct 2014 01:03:16 +0000 (01:03 +0000)]
Remove the TargetMachine from DFAPacketizer since it was only
being used to grab subtarget specific things that we can grab
from the MachineFunction anyhow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219650 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIntroduce Go coding standards for LLVM.
Peter Collingbourne [Tue, 14 Oct 2014 00:40:53 +0000 (00:40 +0000)]
Introduce Go coding standards for LLVM.

Rather than define our own standards, we adopt a set of best practices that
are already in use by the Go community.

Differential Revision: http://reviews.llvm.org/D5761

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219646 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agofix formatting; NFC
Sanjay Patel [Tue, 14 Oct 2014 00:33:23 +0000 (00:33 +0000)]
fix formatting; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219645 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd some optional passes around the vectorizer to both better prepare
Chandler Carruth [Tue, 14 Oct 2014 00:31:29 +0000 (00:31 +0000)]
Add some optional passes around the vectorizer to both better prepare
the IR going into it and to clean up the IR produced by the vectorizers.

Note that these are *off by default* right now while folks collect data
on whether the performance tradeoff is reasonable.

In a build of the 'opt' binary, I see about 2% compile time regression
due to this change on average. This is in my mind essentially the worst
expected case: very little of the opt binary is going to *benefit* from
these extra passes.

I've seen several benchmarks improve in performance my small amounts due
to running these passes, and there are certain (rare) cases where these
passes make a huge difference by either enabling the vectorizer at all
or by hoisting runtime checks out of the outer loop. My primary
motivation is to prevent people from seeing runtime check overhead in
benchmarks where the existing passes and optimizers would be able to
eliminate that.

I've chosen the sequence of passes based on the kinds of things that
seem likely to be relevant for the code at each stage: rotaing loops for
the vectorizer, finding correlated values, loop invariants, and
unswitching opportunities from any runtime checks, and cleaning up
commonalities exposed by the SLP vectorizer.

I'll be pinging existing threads where some of these issues have come up
and will start new threads to get folks to benchmark and collect data on
whether this is the right tradeoff or we should do something else.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219644 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoIntroduce LLVMWriteBitcodeToMemoryBuffer C API function.
Peter Collingbourne [Tue, 14 Oct 2014 00:30:59 +0000 (00:30 +0000)]
Introduce LLVMWriteBitcodeToMemoryBuffer C API function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219643 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUpdating documentation as per Chandler's feedback.
Chris Bieneman [Mon, 13 Oct 2014 23:03:45 +0000 (23:03 +0000)]
Updating documentation as per Chandler's feedback.

This goes with the earlier commit to remove the static destructor from ManagedStatic.cpp by controlling the allocation and de-allocation of the mutex.

Summary: This is part of the ongoing work to remove static constructors and destructors.

Reviewers: chandlerc, rnk

Reviewed By: rnk

Subscribers: rnk, llvm-commits

Differential Revision: http://reviews.llvm.org/D5473

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219640 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstCombine: Fix miscompile in X % -Y -> X % Y transform
David Majnemer [Mon, 13 Oct 2014 22:37:51 +0000 (22:37 +0000)]
InstCombine: Fix miscompile in X % -Y -> X % Y transform

We assumed that negation operations of the form (0 - %Z) resulted in a
negative number.  This isn't true if %Z was originally negative.
Substituting the negative number into the remainder operation may result
in undefined behavior because the dividend might be INT_MIN.

This fixes PR21256.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219639 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemoving the static destructor from ManagedStatic.cpp by controlling the allocation...
Chris Bieneman [Mon, 13 Oct 2014 22:37:25 +0000 (22:37 +0000)]
Removing the static destructor from ManagedStatic.cpp by controlling the allocation and de-allocation of the mutex.

This patch adds a new llvm_call_once function which is used by the ManagedStatic implementation to safely initialize a global to avoid static construction and destruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219638 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMigrate another set of getSubtargetImpl away.
Eric Christopher [Mon, 13 Oct 2014 21:57:44 +0000 (21:57 +0000)]
Migrate another set of getSubtargetImpl away.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219636 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRemove unused debug info constants.
Peter Collingbourne [Mon, 13 Oct 2014 21:50:30 +0000 (21:50 +0000)]
Remove unused debug info constants.

These became unused in r219010.

Differential Revision: http://reviews.llvm.org/D5760

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219635 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstCombine: Don't miscompile (x lshr C1) udiv C2
David Majnemer [Mon, 13 Oct 2014 21:48:30 +0000 (21:48 +0000)]
InstCombine: Don't miscompile (x lshr C1) udiv C2

We have a transform that changes:
  (x lshr C1) udiv C2
into:
  x udiv (C2 << C1)

However, it is unsafe to do so if C2 << C1 discards any of C2's bits.

This fixes PR21255.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219634 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMake first of several changes to bring up to AArch64 fast-isel style
Reed Kotler [Mon, 13 Oct 2014 21:46:41 +0000 (21:46 +0000)]
Make first of several changes to bring up to AArch64 fast-isel style

Summary:
Make Mips fast-isel track the form of AArch64 where practical.
This makes it easier for people to review the code, to borrow similar code, and to see how to eventually move a lot of this
 target code for fast-isels into target independent code.

These are just cosmetic changes. Should be no functional difference.

Test Plan:
make check
test-suite for 4 flavors mips32 r1/r2 , -O0/-O2

Reviewers: dsanders

Reviewed By: dsanders

Subscribers: aemerson, llvm-commits, rfuhler

Differential Revision: http://reviews.llvm.org/D5595

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219633 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUpdate the example of using a command-line option custom parser to
Paul Robinson [Mon, 13 Oct 2014 21:11:22 +0000 (21:11 +0000)]
Update the example of using a command-line option custom parser to
match the current implementation.

Patch by Douglas Yung!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219631 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd an assertion about the integrity of the iterator.
Adrian Prantl [Mon, 13 Oct 2014 20:44:58 +0000 (20:44 +0000)]
Add an assertion about the integrity of the iterator.
Broken parent scope pointers in inlined DIVariables can cause
ensureAbstractVariableIsCreated to insert new abstract scopes, thus
invalidating the iterator in this loop and leading to hard-to-debug
crashes. Useful when manually reducing IR for testcases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219628 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoconstify the getters in SDNodeDbgValue.
Adrian Prantl [Mon, 13 Oct 2014 20:43:47 +0000 (20:43 +0000)]
constify the getters in SDNodeDbgValue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219627 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRefactor debug statement and remove dead argument. NFC.
Chad Rosier [Mon, 13 Oct 2014 19:46:39 +0000 (19:46 +0000)]
Refactor debug statement and remove dead argument. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219626 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd VS2012-generated test inputs for test/tools/llvm-readobj/codeview-linetables...
Timur Iskhodzhanov [Mon, 13 Oct 2014 17:03:13 +0000 (17:03 +0000)]
Add VS2012-generated test inputs for test/tools/llvm-readobj/codeview-linetables.test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219621 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix a broadcast related regression on the vector shuffle lowering.
Filipe Cabecinhas [Mon, 13 Oct 2014 16:16:16 +0000 (16:16 +0000)]
Fix a broadcast related regression on the vector shuffle lowering.

Summary: Test by Robert Lougher!

Reviewers: chandlerc

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D5745

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219617 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Minor cleanup of function
Matt Arsenault [Mon, 13 Oct 2014 15:47:59 +0000 (15:47 +0000)]
R600/SI: Minor cleanup of function

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219616 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[asan-asm-instrumentation] Follow-up fixes to r219602: asserts are moved into
Yuri Gorshenin [Mon, 13 Oct 2014 11:44:06 +0000 (11:44 +0000)]
[asan-asm-instrumentation] Follow-up fixes to r219602: asserts are moved into
function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219610 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdds support for the Cortex-A17 to the ARM backend
Renato Golin [Mon, 13 Oct 2014 10:22:19 +0000 (10:22 +0000)]
Adds support for the Cortex-A17 to the ARM backend

Patch by Matthew Wahab.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219606 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[mips] Mark redundant instructions with a comment in test/CodeGen/Mips/Fast-ISel...
Daniel Sanders [Mon, 13 Oct 2014 10:18:02 +0000 (10:18 +0000)]
[mips] Mark redundant instructions with a comment in test/CodeGen/Mips/Fast-ISel/icmpa.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219605 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64] Add workaround for Cortex-A53 erratum (835769)
Bradley Smith [Mon, 13 Oct 2014 10:12:35 +0000 (10:12 +0000)]
[AArch64] Add workaround for Cortex-A53 erratum (835769)

Some early revisions of the Cortex-A53 have an erratum (835769) whereby it is
possible for a 64-bit multiply-accumulate instruction in AArch64 state to
generate an incorrect result.  The details are quite complex and hard to
determine statically, since branches in the code may exist in some
 circumstances, but all cases end with a memory (load, store, or prefetch)
instruction followed immediately by the multiply-accumulate operation.

The safest work-around for this issue is to make the compiler avoid emitting
multiply-accumulate instructions immediately after memory instructions and the
simplest way to do this is to insert a NOP.

This patch implements such work-around in the backend, enabled via the option
-aarch64-fix-cortex-a53-835769.

The work-around code generation is not enabled by default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219603 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[asan-asm-instrumentation] Fixed memory references which includes %rsp as a base...
Yuri Gorshenin [Mon, 13 Oct 2014 09:37:47 +0000 (09:37 +0000)]
[asan-asm-instrumentation] Fixed memory references which includes %rsp as a base or an index register.

Summary: [asan-asm-instrumentation] Fixed memory references which includes %rsp as a base or an index register.

Reviewers: eugenis

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D5599

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219602 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUnix/Signals.inc: Let findModulesAndOffsets() built conditionally regarding to (defin...
NAKAMURA Takumi [Mon, 13 Oct 2014 04:32:43 +0000 (04:32 +0000)]
Unix/Signals.inc: Let findModulesAndOffsets() built conditionally regarding to (defined(HAVE_BACKTRACE) && defined(ENABLE_BACKTRACES)). [-Wunused-function]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219596 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r219584, "[X86] Memory folding for commutative instructions."
NAKAMURA Takumi [Mon, 13 Oct 2014 04:17:34 +0000 (04:17 +0000)]
Revert r219584, "[X86] Memory folding for commutative instructions."

It broke i686 selfhosting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219595 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[modules] Stop excluding Support/Debug.h from the Support module. This header
Richard Smith [Mon, 13 Oct 2014 00:41:03 +0000 (00:41 +0000)]
[modules] Stop excluding Support/Debug.h from the Support module. This header
has been modular since r206822, and excluding it was leading to workarounds
such as the one in r219592, which this change removes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219593 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Modules] Add some missing includes to make files compile stand-alone.
Benjamin Kramer [Sun, 12 Oct 2014 22:49:26 +0000 (22:49 +0000)]
[Modules] Add some missing includes to make files compile stand-alone.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219592 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoModernize old-style static asserts. NFC.
Benjamin Kramer [Sun, 12 Oct 2014 17:56:40 +0000 (17:56 +0000)]
Modernize old-style static asserts. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219588 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r219223, it creates invalid PHI nodes.
Joerg Sonnenberger [Sun, 12 Oct 2014 17:16:04 +0000 (17:16 +0000)]
Revert r219223, it creates invalid PHI nodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219587 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAPSInt: Simplify code to reduce the number of copies. No functionality change.
Benjamin Kramer [Sun, 12 Oct 2014 15:36:31 +0000 (15:36 +0000)]
APSInt: Simplify code to reduce the number of copies. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219586 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstCombine: Turn (x != 0 & x <u C) into the canonical range check form (x-1 <u C-1)
Benjamin Kramer [Sun, 12 Oct 2014 14:02:34 +0000 (14:02 +0000)]
InstCombine: Turn (x != 0 & x <u C) into the canonical range check form (x-1 <u C-1)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219585 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86] Memory folding for commutative instructions.
Simon Pilgrim [Sun, 12 Oct 2014 10:52:55 +0000 (10:52 +0000)]
[X86] Memory folding for commutative instructions.

This patch improves support for commutative instructions in the x86 memory folding implementation by attempting to fold a commuted version of the instruction if the original folding fails - if that folding fails as well the instruction is 're-commuted' back to its original order before returning.

This mainly helps the stack inliner better fold reloads of 3 (or more) operand instructions (VEX encoded SSE etc.) but by performing this in the lowest foldMemoryOperandImpl implementation it also replaces the X86InstrInfo::optimizeLoadInstr version and is now used by FastISel too.

Differential Revision: http://reviews.llvm.org/D5701

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219584 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstCombine: Simplify commonIDivTransforms
David Majnemer [Sun, 12 Oct 2014 08:34:24 +0000 (08:34 +0000)]
InstCombine: Simplify commonIDivTransforms

A helper routine, MultiplyOverflows, was a less efficient
reimplementation of APInt's smul_ov and umul_ov.  While we are here,
clean up the code so it's more uniform.

No functionality change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219583 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agollvm/test/CodeGen: Some tests don't REQUIRE asserts any more. Remove them.
NAKAMURA Takumi [Sun, 12 Oct 2014 06:47:47 +0000 (06:47 +0000)]
llvm/test/CodeGen: Some tests don't REQUIRE asserts any more. Remove them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219581 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoSuppress llvm-ar's MRI tests for now on win32, since line_iterator is incompatible...
NAKAMURA Takumi [Sat, 11 Oct 2014 22:21:27 +0000 (22:21 +0000)]
Suppress llvm-ar's MRI tests for now on win32, since line_iterator is incompatible to CRLF.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219579 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTest commit access (email fix)
Simon Pilgrim [Sat, 11 Oct 2014 20:28:56 +0000 (20:28 +0000)]
Test commit access (email fix)

Indentation tidyup.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219577 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAssumptionTracker: Don't create temporary CallbackVHs.
Benjamin Kramer [Sat, 11 Oct 2014 19:13:01 +0000 (19:13 +0000)]
AssumptionTracker: Don't create temporary CallbackVHs.

Those are expensive to create in cold cache scenarios. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219575 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMC: Shrink MCSymbolRefExpr by only storing the bits we need.
Benjamin Kramer [Sat, 11 Oct 2014 17:57:27 +0000 (17:57 +0000)]
MC: Shrink MCSymbolRefExpr by only storing the bits we need.

32 -> 16 bytes on x86_64. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219574 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMC: Bit pack MCSymbolData.
Benjamin Kramer [Sat, 11 Oct 2014 15:07:21 +0000 (15:07 +0000)]
MC: Bit pack MCSymbolData.

On x86_64 this brings it from 80 bytes to 64 bytes. Also make any member
variables private and clean up uses to go through the existing accessors.

NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219573 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTest commit access
Simon Pilgrim [Sat, 11 Oct 2014 14:23:36 +0000 (14:23 +0000)]
Test commit access

Fix comment typo + spelling.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219572 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstCombine: Don't fold (X <<s log(INT_MIN)) /s INT_MIN to X
David Majnemer [Sat, 11 Oct 2014 10:20:04 +0000 (10:20 +0000)]
InstCombine: Don't fold (X <<s log(INT_MIN)) /s INT_MIN to X

Consider the case where X is 2.  (2 <<s 31)/s-2147483648 is zero but we
would fold to X.  Note that this is valid when we are in the unsigned
domain because we require NUW: 2 <<u 31 results in poison.

This fixes PR21245.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219568 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstCombine, InstSimplify: (%X /s C1) /s C2 isn't always 0 when C1 * C2 overflow
David Majnemer [Sat, 11 Oct 2014 10:20:01 +0000 (10:20 +0000)]
InstCombine, InstSimplify: (%X /s C1) /s C2 isn't always 0 when C1 * C2 overflow

consider:
C1 = INT_MIN
C2 = -1

C1 * C2 overflows without a doubt but consider the following:
%x = i32 INT_MIN

This means that (%X /s C1) is 1 and (%X /s C1) /s C2 is -1.

N. B.  Move the unsigned version of this transform to InstSimplify, it
doesn't create any new instructions.

This fixes PR21243.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219567 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoInstCombine: mul to shl shouldn't preserve nsw
David Majnemer [Sat, 11 Oct 2014 10:19:52 +0000 (10:19 +0000)]
InstCombine: mul to shl shouldn't preserve nsw

consider:
mul i32 nsw %x, -2147483648

this instruction will not result in poison if %x is 1

however, if we transform this into:
shl i32 nsw %x, 31

then we will be generating poison because we just shifted into the sign
bit.

This fixes PR21242.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219566 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SCEV] Fix one more caller blindly passing the latch to SCEV's
Chandler Carruth [Sat, 11 Oct 2014 05:28:30 +0000 (05:28 +0000)]
[SCEV] Fix one more caller blindly passing the latch to SCEV's
getSmallConstantTripCount even when it isn't the exiting block.

I missed this in my first audit, very sorry. This was found in LNT and
elsewhere. I don't have a test case, but it was completely obvious from
inspection that this was the problem. I'll see if I can reduce a test
case, but I'm not really hopeful, and the value seems quite low.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219562 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoGuard the definition of the stack tracing function with the same macros
Chandler Carruth [Sat, 11 Oct 2014 01:04:40 +0000 (01:04 +0000)]
Guard the definition of the stack tracing function with the same macros
that guard its usage. Without this, we can get unused function warnings
when backtraces are disabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219558 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd basic conditional branches in mips fast-isel
Reed Kotler [Sat, 11 Oct 2014 00:55:18 +0000 (00:55 +0000)]
Add basic conditional branches in mips fast-isel

Summary: Implement the most basic form of conditional branches in Mips fast-isel.

Test Plan:
br1.ll
run 4 flavors of test-suite. mips32 r1/r2 and at -O0/O2

Reviewers: dsanders

Reviewed By: dsanders

Subscribers: llvm-commits, rfuhler

Differential Revision: http://reviews.llvm.org/D5583

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219556 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SCEV] Add some asserts to the recently improved trip count computation
Chandler Carruth [Sat, 11 Oct 2014 00:12:11 +0000 (00:12 +0000)]
[SCEV] Add some asserts to the recently improved trip count computation
routines and fix all of the bugs they expose.

I hit a test case that crashed even without these asserts due to passing
a non-exiting latch to the ExitingBlock parameter of the trip count
computation machinery. However, when I add the nice asserts, it turns
out we have plenty of coverage of these bugs, they just didn't manifest
in crashers.

The core problem seems to stem from an assumption that the latch *is*
the exiting block. While this is often true, and somewhat the "normal"
way to think about loops, it isn't necessarily true. The correct way to
call the trip count routines in a *generic* fashion (that is, without
a particular exit in mind) is to just use the loop's single exiting
block if it has one. The trip count can't be computed generically unless
it does. This works great for the loop vectorizer. The loop unroller
actually *wants* to select the latch when it has to chose between
multiple exits because for unrolling it is the latch trips that matter.
But if this is the desire, it needs to explicitly guard for non-exiting
latches and check for the generic trip count in that case.

I've added the asserts, and added convenience APIs for querying the trip
count generically that check for a single exit block. I've kept the APIs
consistent between computing trip count and trip multiples.

Thansk to Mark for the help debugging and tracking down the *right* fix
here!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219550 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MCJIT] Replace memcpy with readBytesUnaligned in RuntimeDyldMachOI386.
Lang Hames [Fri, 10 Oct 2014 23:07:09 +0000 (23:07 +0000)]
[MCJIT] Replace memcpy with readBytesUnaligned in RuntimeDyldMachOI386.

This should fix the failures of the MachO_i386_DynNoPIC_relocations.s test case
on MIPS hosts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219543 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoReturn undef on FP <-> Int conversions that overflow (PR21330).
Sanjay Patel [Fri, 10 Oct 2014 23:00:21 +0000 (23:00 +0000)]
Return undef on FP <-> Int conversions that overflow (PR21330).

The LLVM Lang Ref states for signed/unsigned int to float conversions:
"If the value cannot fit in the floating point value, the results are undefined."

And for FP to signed/unsigned int:
"If the value cannot fit in ty2, the results are undefined."

This matches the C definitions.

The existing behavior pins to infinity or a max int value, but that may just
lead to more confusion as seen in:
http://llvm.org/bugs/show_bug.cgi?id=21130

Returning undef will hopefully lead to a less silent failure.

Differential Revision: http://reviews.llvm.org/D5603

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219542 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFollow-up to r219534 to make symbolization more robust.
Alexey Samsonov [Fri, 10 Oct 2014 22:58:26 +0000 (22:58 +0000)]
Follow-up to r219534 to make symbolization more robust.

1) Explicitly provide important arguments to llvm-symbolizer,
not relying on defaults.
2) Be more defensive about symbolizer output.

This might fix weird failures on ninja-x64-msvc-RA-centos6 buildbot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219541 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Change how DS offsets are printed
Matt Arsenault [Fri, 10 Oct 2014 22:16:07 +0000 (22:16 +0000)]
R600/SI: Change how DS offsets are printed

Match SC by using offset/offset0/offset1 and printing
in decimal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219537 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Match read2/write2 stride 64 versions
Matt Arsenault [Fri, 10 Oct 2014 22:12:32 +0000 (22:12 +0000)]
R600/SI: Match read2/write2 stride 64 versions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219536 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRe-land r219354: Use llvm-symbolizer to symbolize LLVM/Clang crash dumps.
Alexey Samsonov [Fri, 10 Oct 2014 22:06:59 +0000 (22:06 +0000)]
Re-land r219354: Use llvm-symbolizer to symbolize LLVM/Clang crash dumps.

In fact, symbolization is now expected to work only on Linux and
FreeBSD/NetBSD, where we have dl_iterate_phdr and can learn the
main executable name without argv0 (it will be possible on BSD systems
after http://reviews.llvm.org/D5693 lands). #ifdef-out the code for
all the rest Unix systems.

Reviewed in http://reviews.llvm.org/D5610

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219534 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoR600/SI: Add load / store machine optimizer pass.
Matt Arsenault [Fri, 10 Oct 2014 22:01:59 +0000 (22:01 +0000)]
R600/SI: Add load / store machine optimizer pass.

Currently this only functions to match simple cases
where ds_read2_* / ds_write2_* instructions can be used.

In the future it might match some of the other weird
load patterns, such as direct to LDS loads.

Currently enabled only with a subtarget feature to enable
easier testing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219533 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoThis patch teaches ScalarEvolution to pick and use !range metadata.
Sanjoy Das [Fri, 10 Oct 2014 21:22:34 +0000 (21:22 +0000)]
This patch teaches ScalarEvolution to pick and use !range metadata.
It also makes it more aggressive in querying range information by
adding a call to isKnownPredicateWithRanges to
isLoopBackedgeGuardedByCond and isLoopEntryGuardedByCond.

phabricator: http://reviews.llvm.org/D5638

Reviewed by: atrick, hfinkel

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219532 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[mips] Actually mark that the default case is unreachable as this switch
Chandler Carruth [Fri, 10 Oct 2014 21:07:03 +0000 (21:07 +0000)]
[mips] Actually mark that the default case is unreachable as this switch
is over a subset of condition codes.

This fixes the -Werror build which warns about use of uninitialized
variables in the default case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219531 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoImplement floating point compare for mips fast-isel
Reed Kotler [Fri, 10 Oct 2014 20:46:28 +0000 (20:46 +0000)]
Implement floating point compare for mips fast-isel

Summary: Expand SelectCmp to handle floating point compare

Test Plan:
fpcmpa.ll
run 4 flavors of test-suite, mips32 r1/r2 O0/O2

Reviewers: dsanders

Reviewed By: dsanders

Subscribers: llvm-commits, rfuhler

Differential Revision: http://reviews.llvm.org/D5567

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219530 91177308-0d34-0410-b5e6-96231b3b80d8