oota-llvm.git
9 years agoFix (re-)creation of unittest lit.site.cfg for clang-tools-extra.
Benjamin Kramer [Wed, 13 Aug 2014 20:41:26 +0000 (20:41 +0000)]
Fix (re-)creation of unittest lit.site.cfg for clang-tools-extra.

This has been hiding really well. Hopefully brings the builders suffering from
outdated lit.site.cfg files back to life.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215575 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoutils: Fix segfault in flattencfg
Jan Vesely [Wed, 13 Aug 2014 20:31:53 +0000 (20:31 +0000)]
utils: Fix segfault in flattencfg

v2: continue iterating through the rest of the bb
    use for loop

v3: initialize FlattenCFG pass in ScalarOps
    add test

v4: split off initializing flattencfg to a separate patch
    add comment

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215574 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInitialize FlattenCFG pass
Jan Vesely [Wed, 13 Aug 2014 20:31:52 +0000 (20:31 +0000)]
Initialize FlattenCFG pass

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215573 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSimplify memory ownership with std::unique_ptr.
Rafael Espindola [Wed, 13 Aug 2014 18:59:01 +0000 (18:59 +0000)]
Simplify memory ownership with std::unique_ptr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215567 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSimplify ownership with std::unique_ptr. NFC.
Rafael Espindola [Wed, 13 Aug 2014 18:49:01 +0000 (18:49 +0000)]
Simplify ownership with std::unique_ptr. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215566 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600: Correctly set the src value offset for scalarized kernel args
Matt Arsenault [Wed, 13 Aug 2014 18:14:11 +0000 (18:14 +0000)]
R600: Correctly set the src value offset for scalarized kernel args

This for some reason fixes v1i64 kernel arguments on pre-SI. This
currently breaks some other cases in the kernel-args.ll test for R600,
but I'm not particularly confident in the new output. VTX_READ_* are not
used for some of the scalarized cases, and the code reading from the
constant buffer doesn't make much sense to me.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215564 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCanonicalize header guards into a common format.
Benjamin Kramer [Wed, 13 Aug 2014 16:26:38 +0000 (16:26 +0000)]
Canonicalize header guards into a common format.

Add header guards to files that were missing guards. Remove #endif comments
as they don't seem common in LLVM (we can easily add them back if we decide
they're useful)

Changes made by clang-tidy with minor tweaks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215558 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTest commit, remove trailing whitespace
Benjamin Foster [Wed, 13 Aug 2014 16:11:50 +0000 (16:11 +0000)]
Test commit, remove trailing whitespace

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215556 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[DAGCombiner] Improved target independent vector shuffle combine rule.
Andrea Di Biagio [Wed, 13 Aug 2014 16:09:40 +0000 (16:09 +0000)]
[DAGCombiner] Improved target independent vector shuffle combine rule.

This patch improves the existing algorithm in DAGCombiner that
attempts to fold shuffles according to rule:
  shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(y, undef, M3)

Before this change, there were cases where the DAGCombiner conservatively
avoided folding shuffles even if the resulting mask would have been legal.
That is because the algorithm wrongly assumed that commuting
an illegal shuffle mask would always produce an illegal mask.

With this change, we now correctly compute the commuted shuffle mask before
calling method 'isShuffleMaskLegal' on it.
On X86, this improves for example the codegen for the following function:

define <4 x i32> @test(<4 x i32> %A, <4 x i32> %B) {
  %1 = shufflevector <4 x i32> %B, <4 x i32> %A, <4 x i32> <i32 1, i32 2, i32 6, i32 7>
  %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 2, i32 3>
  ret <4 x i32> %2
}

Before this change the X86 backend (-mcpu=corei7) generated
the following assembly code for function @test:
  shufps $-23, %xmm0, %xmm1  # xmm1 = xmm1[1,2],xmm0[2,3]
  movhlps %xmm1, %xmm1       # xmm1 = xmm1[1,1]
  movaps %xmm1, %xmm0

Now we produce:
  movhlps %xmm0, %xmm0       # xmm0 = xmm0[1,1]

Added extra test cases in combine-vec-shuffle-2.ll to verify that we correctly
fold according to the above-mentioned rule.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215555 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips] Refactor calls to setCanHaveModuleDir.
Toma Tabacu [Wed, 13 Aug 2014 12:48:12 +0000 (12:48 +0000)]
[mips] Refactor calls to setCanHaveModuleDir.

Summary:
Moved some calls to setCanHaveModuleDir to the MipsTargetStreamer base class and removed the resulting empty functions from the MipsTargetELFStreamer class.

Also fixed a missing call to setCanHaveModuleDir in MipsTargetELFStreamer::emitDirectiveSetMicroMips.

Reviewers: dsanders

Reviewed By: dsanders

Subscribers: tomatabacu

Differential Revision: http://reviews.llvm.org/D4781

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215542 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[shuffle] Stand back! I'm about to (try to) do math!
Chandler Carruth [Wed, 13 Aug 2014 12:27:18 +0000 (12:27 +0000)]
[shuffle] Stand back! I'm about to (try to) do math!

Especially with blends and large tree heights there was a problem with
the fuzzer where it would end up with enough undef shuffle elements in
enough parts of the tree that in a birthday-attack kind of way we ended
up regularly having large numbers of undef elements in the result. I was
seeing reasonably frequent cases of *all* results being undef which
prevents us from doing any correctness checking at all. While having
undef lanes is important, this was too much.

So I've tried to apply some math to the probabilities of having an undef
lane and balance them against the tree height. Please be gentle, I'm
really terrible at math. I probably made a bunch of amateur mistakes
here. Fixes, etc. are quite welcome. =D At least in running it some, it
seems to be producing more interesting (for correctness testing)
results.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215540 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAsserting that the call to chdir succeeds in this test. Fixes some -Wunused-result...
Aaron Ballman [Wed, 13 Aug 2014 11:17:41 +0000 (11:17 +0000)]
Asserting that the call to chdir succeeds in this test. Fixes some -Wunused-result warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215539 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[optnone] Make the optnone attribute effective at suppressing function
Chandler Carruth [Wed, 13 Aug 2014 10:49:33 +0000 (10:49 +0000)]
[optnone] Make the optnone attribute effective at suppressing function
attribute and function argument attribute synthesizing and propagating.

As with the other uses of this attribute, the goal remains a best-effort
(no guarantees) attempt to not optimize the function or assume things
about the function when optimizing. This is particularly useful for
compiler testing, bisecting miscompiles, triaging things, etc. I was
hitting specific issues using optnone to isolate test code from a test
driver for my fuzz testing, and this is one step of fixing that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215538 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSilence a -Wparenthesis warning with these asserts. NFC.
Aaron Ballman [Wed, 13 Aug 2014 10:49:07 +0000 (10:49 +0000)]
Silence a -Wparenthesis warning with these asserts. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215537 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[SKX] Extended non-temporal load/store instructions for AVX512VL subsets.
Robert Khasanov [Wed, 13 Aug 2014 10:46:00 +0000 (10:46 +0000)]
[SKX] Extended non-temporal load/store instructions for AVX512VL subsets.

Added avx512_movnt_vl multiclass for handling 256/128-bit forms of instruction.
Added encoding and lowering tests.

Reviewed by Elena Demikhovsky <elena.demikhovsky@intel.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215536 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRe-commit: [mips] Implement .ent, .end, .frame, .mask and .fmask.
Daniel Sanders [Wed, 13 Aug 2014 10:07:34 +0000 (10:07 +0000)]
Re-commit: [mips] Implement .ent, .end, .frame, .mask and .fmask.

Patch by Matheus Almeida and Toma Tabacu

The lld test failure on the previous attempt to commit was caused by the
addition of the .pdr section causing the offsets it was checking to change.
This has been fixed by removing the .ent/.end directives from that test since
they weren't really needed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215535 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[shuffle] Make the seed an optional component and add support for
Chandler Carruth [Wed, 13 Aug 2014 10:00:46 +0000 (10:00 +0000)]
[shuffle] Make the seed an optional component and add support for
letting the python very directly compute a UUID.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215533 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert r215415 which causse MSan to crash on a great deal of C++ code.
Chandler Carruth [Wed, 13 Aug 2014 09:19:39 +0000 (09:19 +0000)]
Revert r215415 which causse MSan to crash on a great deal of C++ code.
I've followed up on the original commit as well.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215532 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[shuffle] Teach the shuffle fuzzer to fuzz blends, including forming
Chandler Carruth [Wed, 13 Aug 2014 09:05:59 +0000 (09:05 +0000)]
[shuffle] Teach the shuffle fuzzer to fuzz blends, including forming
a tree of inputs to blend iteratively together.

This required a pretty substantial rewrite of the innards. The number of
shuffle instructions is now bounded in terms of tree-height. There is
a flag to disable blends so that its still possible to test single input
shuffles. I've also improved various aspects of how the test program is
generated, primarily to simplify the test harness and allow some
optimizations to clean up how we actually check the results and build up
the inputs.

Again, apologies for my likely horrible use of Python... But hey, it
works! (Ish?)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215530 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAVX-512: Fixed a bug in shufflevector lowering.
Elena Demikhovsky [Wed, 13 Aug 2014 07:58:43 +0000 (07:58 +0000)]
AVX-512: Fixed a bug in shufflevector lowering.
PALIGNR instruction does not exist in AVX-512F set.
Added a test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215526 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInstCombine: Combine (xor (or %a, %b) (xor %a, %b)) to (add %a, %b)
Karthik Bhat [Wed, 13 Aug 2014 05:13:14 +0000 (05:13 +0000)]
InstCombine: Combine (xor (or %a, %b) (xor %a, %b)) to (add %a, %b)

Correctness proof of the transform using CVC3-

$ cat t.cvc
A, B : BITVECTOR(32);
QUERY BVXOR(A | B, BVXOR(A,B) ) = A & B;

$ cvc3 t.cvc
Valid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215524 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[NVPTX] Remove MemIntrinsicSDNode/MemSDNode duplicate checking
Hal Finkel [Wed, 13 Aug 2014 04:59:51 +0000 (04:59 +0000)]
[NVPTX] Remove MemIntrinsicSDNode/MemSDNode duplicate checking

As of r214452, isa<MemSDNode> will return true for nodes for which
isa<MemIntrinsicSDNode> will return true (classof now respects the actual class
hierarchy). So we no longer need to check for both MemIntrinsicSDNode and
MemSDNode separately.

No functionality change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215523 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix examples of "named metadata" (some of which isn't named).
Nick Lewycky [Wed, 13 Aug 2014 04:54:05 +0000 (04:54 +0000)]
Fix examples of "named metadata" (some of which isn't named).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215522 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[shuffle] Tweak the shuffle fuzzer to support bigger seeds. I'm
Chandler Carruth [Wed, 13 Aug 2014 03:21:11 +0000 (03:21 +0000)]
[shuffle] Tweak the shuffle fuzzer to support bigger seeds. I'm
currently using UUIDs to seed this in order to scan a bigger range.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215521 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[x86] Rewrite a core part of the new vector shuffle lowering to handle
Chandler Carruth [Wed, 13 Aug 2014 01:25:45 +0000 (01:25 +0000)]
[x86] Rewrite a core part of the new vector shuffle lowering to handle
one pesky test case correctly.

This test case caused the old code to infloop occilating between solving
the low-half and the high-half. The 'side balancing' part of
single-input v8 shuffle lowering didn't handle the one pattern which can
cause it to occilate. Fortunately the fuzz testing found this case.
Unfortuately it was *terrible* to handle. I'm really sorry for the
amount and density of the code here, I'd love suggestions on how to
simplify it. I feel like there *must* be a simpler form here, but after
a lot of days I've not found it. This is the only one I've found that
even works. I've added the one pesky test case along with some nice
comments explaining the core problem that we have to solve here.

So far this has survived approximately 32k test cases. More strenuous
fuzzing commencing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215519 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[PowerPC] Implement PPCTargetLowering::getTgtMemIntrinsic
Hal Finkel [Wed, 13 Aug 2014 01:15:40 +0000 (01:15 +0000)]
[PowerPC] Implement PPCTargetLowering::getTgtMemIntrinsic

This implements PPCTargetLowering::getTgtMemIntrinsic for Altivec load/store
intrinsics. As with the construction of the MachineMemOperands for the
intrinsic calls used for unaligned load/store lowering, the only slight
complication is that we need to represent a larger memory range than the
loaded/stored value-type size (because the address is rounded down to an
aligned address, and we need to conservatively represent the entire possible
range of the actual access). This required adding an extra size field to
TargetLowering::IntrinsicInfo, and this was done in a way that required no
modifications to other targets (the size defaults to the store size of the
provided memory data type).

This fixes test/CodeGen/PowerPC/unal-altivec-wint.ll (so it can be un-XFAILed).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215512 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix classof for ISD::INTRINSIC_W_CHAIN and INTRINSIC_VOID
Hal Finkel [Wed, 13 Aug 2014 01:15:37 +0000 (01:15 +0000)]
Fix classof for ISD::INTRINSIC_W_CHAIN and INTRINSIC_VOID

Unfortunately, our use of the SDNode class hierarchy for INTRINSIC_W_CHAIN and
INTRINSIC_VOID nodes is somewhat broken right now. These nodes sometimes are
used for memory intrinsics (those with MachineMemOperands), and sometimes not.
When not, the nodes are not created as instances of MemIntrinsicSDNode, but
rather created as some other subclass of SDNode using DAG::getNode. When they
are memory intrinsics, they are created using DAG::getMemIntrinsicNode as
instances of MemIntrinsicSDNode. MemIntrinsicSDNode is a subclass of
MemSDNode, but prior to r214452, we had a non-self-consistent setup whereby
MemIntrinsicSDNode::classof on INTRINSIC_W_CHAIN and INTRINSIC_VOID would
return true but MemSDNode::classof on INTRINSIC_W_CHAIN and INTRINSIC_VOID
would return false. In r214452, MemSDNode::classof was changed to return true
for INTRINSIC_W_CHAIN and INTRINSIC_VOID, which is now self-consistent. The
problem is that neither the pre-r214452 logic and the post-r214452 logic are
really right. The truth is that not all INTRINSIC_W_CHAIN and INTRINSIC_VOID
nodes are instances of MemIntrinsicSDNode (or MemSDNode for that matter), and
the return value from classof needs to reflect that. This was broken before
r214452 (because MemIntrinsicSDNode::classof always returned true), and was
broken afterward (because MemSDNode::classof also always returned true), and
will now be correct.

The minimal solution is to grab one of the SubclassData bits (there is one left
for MemIntrinsicSDNode nodes) and use it to store whether or not a particular
INTRINSIC_W_CHAIN or INTRINSIC_VOID is really an instance of
MemIntrinsicSDNode or not. Doing this allows both MemIntrinsicSDNode::classof
and MemSDNode::classof to return the correct answer for the underlying object
for both the memory-intrinsic and non-memory-intrinsic cases.

This fixes the problem that r214452 created in the SelectionDAGDumper (thanks
to Matt Arsenault for pointing it out).

Because PowerPC does not implement getTgtMemIntrinsic, this change breaks
test/CodeGen/PowerPC/unal-altivec-wint.ll. I've XFAILed it for now, and will
fix it in a follow-up commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215511 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AVX512] Verify the code generated for the intrinsic _mm512_broadcastsd_pd
Adam Nemet [Wed, 13 Aug 2014 00:30:05 +0000 (00:30 +0000)]
[AVX512] Verify the code generated for the intrinsic _mm512_broadcastsd_pd

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215487 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix -Wsign-compare warnings
David Blaikie [Tue, 12 Aug 2014 23:23:05 +0000 (23:23 +0000)]
Fix -Wsign-compare warnings

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215483 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAPInt: Make self-move-assignment a no-op to fix stage3 clang-cl
Reid Kleckner [Tue, 12 Aug 2014 22:01:39 +0000 (22:01 +0000)]
APInt: Make self-move-assignment a no-op to fix stage3 clang-cl

It's not clear what the semantics of a self-move should be.  The
consensus appears to be that a self-move should leave the object in a
moved-from state, which is what our existing move assignment operator
does.

However, the MSVC 2013 STL will perform self-moves in some cases.  In
particular, when doing a std::stable_sort of an already sorted APSInt
vector of an appropriate size, one of the merge steps will self-move
half of the elements.

We don't notice this when building with MSVC, because MSVC will not
synthesize the move assignment operator for APSInt.  Presumably MSVC
does this because APInt, the base class, has user-declared special
members that implicitly delete move special members.  Instead, MSVC
selects the copy-assign operator, which defends against self-assignment.
Clang, on the other hand, selects the move-assign operator, and we get
garbage APInts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215478 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRemove a condition that can never be true, as wittnessed by the assert
Adrian Prantl [Tue, 12 Aug 2014 21:55:58 +0000 (21:55 +0000)]
Remove a condition that can never be true, as wittnessed by the assert
above.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215477 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AVX512] Handle valign masking intrinsic via C++ lowering
Adam Nemet [Tue, 12 Aug 2014 21:13:12 +0000 (21:13 +0000)]
[AVX512] Handle valign masking intrinsic via C++ lowering

I think that this will scale better in most cases than adding a Pat<> for each
mapping from the intrinsic DAG to the intruction (i.e. rri, rrik, rrikz).  We
can just lower to the SDNode and have the resulting DAG be matches by the DAG
patterns.

Alternatively (long term), we could keep the Pat<>s but generate them via the
new AVX512_masking multiclass.  The difficulty is that in order to formulate
that we would have to concatenate DAGs.  Currently this is only supported if
the operators of the input DAGs are identical.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215473 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAllwo bitcast + struct GEP transform to work with addrspacecast
Matt Arsenault [Tue, 12 Aug 2014 19:46:13 +0000 (19:46 +0000)]
Allwo bitcast + struct GEP transform to work with addrspacecast

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215467 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600: Use optimized 24bit path in udivrem
Jan Vesely [Tue, 12 Aug 2014 17:31:20 +0000 (17:31 +0000)]
R600: Use optimized 24bit path in udivrem

v2: drop enum keyword
    use correct extension mode
    don't bother computing the sign in unsinged case

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215462 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600: Remove unused code.
Jan Vesely [Tue, 12 Aug 2014 17:31:19 +0000 (17:31 +0000)]
R600: Remove unused code.

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215461 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600: Use i24 optimized path for SREM
Jan Vesely [Tue, 12 Aug 2014 17:31:17 +0000 (17:31 +0000)]
R600: Use i24 optimized path for SREM

v2: add tests
    rename LowerSDIV24 to LowerSDIVREM24
    handle the rem part in this function

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215460 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix a parentheses warning introduced in r215394.
Quentin Colombet [Tue, 12 Aug 2014 17:11:26 +0000 (17:11 +0000)]
Fix a parentheses warning introduced in r215394.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215459 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDon't upgrade global constructors when reading bitcode
Duncan P. N. Exon Smith [Tue, 12 Aug 2014 16:46:37 +0000 (16:46 +0000)]
Don't upgrade global constructors when reading bitcode

An optional third field was added to `llvm.global_ctors` (and
`llvm.global_dtors`) in r209015.  Most of the code has been changed to
deal with both versions of the variables.  Users of the C API might
create either version, the helper functions in LLVM create the two-field
version, and clang now creates the three-field version.

However, the BitcodeReader was changed to always upgrade to the
three-field version.  This created an unnecessary inconsistency in the
IR before/after serializing to bitcode.

This commit resolves the inconsistency by making the third field truly
optional (and not upgrading in the bitcode reader).  Since `llvm-link`
was relying on this upgrade code, rather than deleting it I've moved it
into `ModuleLinker`, where it upgrades these arrays as necessary to
resolve inconsistencies between modules.

The ideal resolution would be to remove the 2-field version and make the
third field required.  I filed PR20506 to track that.

I changed `test/Bitcode/upgrade-global-ctors.ll` to a negative test and
duplicated the `llvm-link` check in `test/Linker/global_ctors.ll` to
check both upgrade directions.

Since I came across this as part of PR5680 (serializing use-list order),
I've also added the missing `verify-uselistorder` RUN line to
`test/Bitcode/metadata-2.ll`.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215457 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agofixed typos
Sanjay Patel [Tue, 12 Aug 2014 16:00:06 +0000 (16:00 +0000)]
fixed typos

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215451 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMake the test a bit more strict.
Rafael Espindola [Tue, 12 Aug 2014 15:55:27 +0000 (15:55 +0000)]
Make the test a bit more strict.

Before it would pass even if @b or @c ended up pointing to a variable named
@a123.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215450 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd a plugin testcase for merging weak variables.
Rafael Espindola [Tue, 12 Aug 2014 15:39:14 +0000 (15:39 +0000)]
Add a plugin testcase for merging weak variables.

I initially thought I could implement COMDATs with aliases by just
internalizing GVs instead of dropping them. This is a counter
example: Internalizing one of the @a would make @b and @c point
to different variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215447 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agollvm/test/TableGen/*Foreach*.td: Remove XFAIL:vg_leak. They have not been failing...
NAKAMURA Takumi [Tue, 12 Aug 2014 14:06:21 +0000 (14:06 +0000)]
llvm/test/TableGen/*Foreach*.td: Remove XFAIL:vg_leak. They have not been failing since r215176.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215445 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoReverted my "Testing commit access" commit.
Toma Tabacu [Tue, 12 Aug 2014 12:41:44 +0000 (12:41 +0000)]
Reverted my "Testing commit access" commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215441 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoTesting commit access.
Toma Tabacu [Tue, 12 Aug 2014 12:29:40 +0000 (12:29 +0000)]
Testing commit access.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215440 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agollvm-objdump: print contents of MachO __unwind_info sections
Tim Northover [Tue, 12 Aug 2014 11:52:59 +0000 (11:52 +0000)]
llvm-objdump: print contents of MachO __unwind_info sections

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215437 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoHave MachineRegisterInfo take and store the MachineFunction it
Eric Christopher [Tue, 12 Aug 2014 08:00:56 +0000 (08:00 +0000)]
Have MachineRegisterInfo take and store the MachineFunction it
was created for rather than the TargetMachine since we only
needed the TM for the subtarget and we can get that from the
MF.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215432 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MachineCombiner] Fix for ICE bug 20598
Gerolf Hoflehner [Tue, 12 Aug 2014 07:54:12 +0000 (07:54 +0000)]
[MachineCombiner] Fix for ICE bug 20598

The combiner ignored DBG nodes when checking
the uses of a virtual register.

It combined a sequence like
   %vreg1 = madd %vreg2, %vreg3,...
   DBG_VALUE (%vreg1 ...)
   %vreg4 = add %vreg1,...
to
  %vreg4 = madd %vreg2, %vreg3

leaving behind a dangling DBG_VALUE with
a definition. This triggered an assertion
in the MachineTraceMetrics.cpp module.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215431 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoIR: Print a newline when dumping Types
Justin Bogner [Tue, 12 Aug 2014 03:24:59 +0000 (03:24 +0000)]
IR: Print a newline when dumping Types

Type::dump() doesn't print a newline, which makes for a poor
experience in a debugger. This looks like it was an ommission
considering Value::dump() two lines above, so I've changed Type to add
a newline as well.

Of the two in-tree callers, one added a newline anyway, and I've
updated the other one to use Type::print instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215421 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[OCaml] Expose Llvm.get_operand_use.
Peter Zotov [Tue, 12 Aug 2014 02:55:45 +0000 (02:55 +0000)]
[OCaml] Expose Llvm.get_operand_use.

Patch by Gabriel Radanne <drupyog@zoho.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215420 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[LLVM-C] Expose User::getOperandUse as LLVMGetOperandUse.
Peter Zotov [Tue, 12 Aug 2014 02:55:40 +0000 (02:55 +0000)]
[LLVM-C] Expose User::getOperandUse as LLVMGetOperandUse.

Patch by Gabriel Radanne <drupyog@zoho.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215419 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDebugLocEntry: Restore the comparison predicate from before the
Adrian Prantl [Tue, 12 Aug 2014 01:07:53 +0000 (01:07 +0000)]
DebugLocEntry: Restore the comparison predicate from before the
refactoring in 215384. This way it can unique multiple entries describing
the same piece even if they don't have the exact same location.
(The same piece may get merged in and be added from OpenRanges).
There ought to be a more elegant solution for this, though.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215418 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agomsan: Handle musttail calls
Reid Kleckner [Tue, 12 Aug 2014 00:12:43 +0000 (00:12 +0000)]
msan: Handle musttail calls

First, avoid calling setTailCall(false) on musttail calls.  The funciton
prototypes should be "congruent", so the shadow layout should be exactly
the same.

Second, avoid inserting instrumentation after a musttail call to
propagate the return value shadow.  We don't need to propagate the
result of a tail call, it should already be in the right place.

Reviewed By: eugenis

Differential Revision: http://reviews.llvm.org/D4331

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215415 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMove helper for getting a terminating musttail call to BasicBlock
Reid Kleckner [Tue, 12 Aug 2014 00:05:15 +0000 (00:05 +0000)]
Move helper for getting a terminating musttail call to BasicBlock

No functional change.  To be used in future commits that need to look
for such instructions.

Reviewed By: rafael

Differential Revision: http://reviews.llvm.org/D4504

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215413 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "Partially revert r214761 that asserted that all concrete debug info variables...
David Blaikie [Tue, 12 Aug 2014 00:00:31 +0000 (00:00 +0000)]
Revert "Partially revert r214761 that asserted that all concrete debug info variables had DIEs, due to a failure on Darwin."

I believe this was addressed by r215157 and r215227, so let's have
another go at the bots, etc.

This reverts commit r214880.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215412 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[MachineSink] Improve the compile time by preserving the dominance information
Quentin Colombet [Mon, 11 Aug 2014 23:52:01 +0000 (23:52 +0000)]
[MachineSink] Improve the compile time by preserving the dominance information
as long as possible.

** Context **

Each time the dominance information is modified, the dominator tree analysis
switches in a slow query mode. After a few queries without any modification on
the dominator tree, it performs an expensive update of its internal structure to
provide fast queries again.

** Problem **

Prior to this patch, the MachineSink pass was splitting the critical edges on
demand while relying heavy on the dominator tree information. In some cases,
this leads to pathological behavior where:
- We end up in the slow query mode right after splitting an edge.
- We update the dominance information.
- We break the dominance information again, thus ending up in the slow query
  mode and so on.

** Proposed Solution **

To mitigate this effect, this patch postpones all the splitting of the edges at
the end of each iteration of the main loop.
The benefits are:
- The dominance information is valid for the life time of an iteration.
- This simplifies the code as we do not have to special treat instructions that
  are sunk on critical edges. Indeed, the related block will be available
  through the next iteration.

The downside is that when edges splitting is required, this incurs an additional
iteration of the main loop compared to the previous scheme.

** Performance **

Thanks to this patch, the motivating example compiles in 6+ minutes instead of
10+ minutes. No test case added as the motivating example as nothing special but
being huge!

I have measured only noise for both the compile time and the runtime on the llvm
test-suite + SPECs with Os and O3.

Note: The current implementation of MachineBasicBlock::SplitCriticalEdge also
uses the dominance information and therefore, hits this problem. A subsequent
patch will address that.

<rdar://problem/17894619>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215410 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[x86] Fold extract_vector_elt of a load into the Load's address computation.
Michael J. Spencer [Mon, 11 Aug 2014 23:49:33 +0000 (23:49 +0000)]
[x86] Fold extract_vector_elt of a load into the Load's address computation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215409 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd a couple of convenience accessors to DebugLocEntry::Value to further
Adrian Prantl [Mon, 11 Aug 2014 23:22:59 +0000 (23:22 +0000)]
Add a couple of convenience accessors to DebugLocEntry::Value to further
simplify common usage patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215407 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SIInstrInfo.cpp: Suppress an warning. [-Wunused-variable]
NAKAMURA Takumi [Mon, 11 Aug 2014 23:03:38 +0000 (23:03 +0000)]
R600/SIInstrInfo.cpp: Suppress an warning. [-Wunused-variable]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215406 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[ARM] Mark VMOVDRR with the RegSequence property and implement the related
Quentin Colombet [Mon, 11 Aug 2014 22:56:22 +0000 (22:56 +0000)]
[ARM] Mark VMOVDRR with the RegSequence property and implement the related
target hook.

This patch teaches the compiler that:
dX = VMOVDRR rY, rZ
is the same as:
dX = REG_SEQUENCE rY, ssub_0, rZ, ssub_1

<rdar://problem/12702965>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215404 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMake these DebugLocEntry::Value comparison operators friend functions
Adrian Prantl [Mon, 11 Aug 2014 22:52:56 +0000 (22:52 +0000)]
Make these DebugLocEntry::Value comparison operators friend functions
as suggested by dblaikie in a comment on r215384.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215403 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd missing closing namespace comment.
Jim Grosbach [Mon, 11 Aug 2014 22:42:31 +0000 (22:42 +0000)]
Add missing closing namespace comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215402 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAArch64: Tidy up a few comments.
Jim Grosbach [Mon, 11 Aug 2014 22:42:28 +0000 (22:42 +0000)]
AArch64: Tidy up a few comments.

Have the comments match the actual parameter names. Found via clang-tidy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215401 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoInstCombine: Combine (add (and %a, %b) (or %a, %b)) to (add %a, %b)
David Majnemer [Mon, 11 Aug 2014 22:32:02 +0000 (22:32 +0000)]
InstCombine: Combine (add (and %a, %b) (or %a, %b)) to (add %a, %b)

What follows bellow is a correctness proof of the transform using CVC3.

$ < t.cvc
A, B : BITVECTOR(32);

QUERY BVPLUS(32, A & B, A | B) = BVPLUS(32, A, B);

$ cvc3 < t.cvc
Valid.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215400 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Add a ComplexPattern for selecting MUBUF _OFFSET variant
Tom Stellard [Mon, 11 Aug 2014 22:18:17 +0000 (22:18 +0000)]
R600/SI: Add a ComplexPattern for selecting MUBUF _OFFSET variant

This saves us from having to copy a 64-bit 0 value into VGPRs for
BUFFER_* instruction which only have a 12-bit immediate offset.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215399 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Add an _OFFEN variant MUBUF_STORE_* and use it for scratch writes
Tom Stellard [Mon, 11 Aug 2014 22:18:14 +0000 (22:18 +0000)]
R600/SI: Add an _OFFEN variant MUBUF_STORE_* and use it for scratch writes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215398 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Add check for low 32 bits of encoding to mubuf tests
Tom Stellard [Mon, 11 Aug 2014 22:18:11 +0000 (22:18 +0000)]
R600/SI: Add check for low 32 bits of encoding to mubuf tests

There are no variable values like registers encoded in the low 32 bits of MUBUF
instructions, so it is relatively easy to check these bits, and it will
help prevent us from introducing encoding bugs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215397 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Clear lds bit on MUBUF instructions used for private stores
Tom Stellard [Mon, 11 Aug 2014 22:18:09 +0000 (22:18 +0000)]
R600/SI: Clear lds bit on MUBUF instructions used for private stores

This bit was left uninitialized, which was causing some random failures
of piglit tests.

NOTE: This is a candidate for the 3.5 branch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215396 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoR600/SI: Fix broken test
Tom Stellard [Mon, 11 Aug 2014 22:18:05 +0000 (22:18 +0000)]
R600/SI: Fix broken test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215395 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd isRegSequence property.
Quentin Colombet [Mon, 11 Aug 2014 22:17:14 +0000 (22:17 +0000)]
Add isRegSequence property.

This patch adds a new property: isRegSequence and the related target hooks:
TargetIntrInfo::getRegSequenceInputs and
TargetInstrInfo::getRegSequenceLikeInputs to specify that a target specific
instruction is a (kind of) REG_SEQUENCE.

<rdar://problem/12702965>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215394 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[AArch64] Fix registerAllocator assigns same register for base and wback in
Quentin Colombet [Mon, 11 Aug 2014 21:39:53 +0000 (21:39 +0000)]
[AArch64] Fix registerAllocator assigns same register for base and wback in
pre/post-index load and store.

Patch by Steven Wu <stevenwu@apple.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215390 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDebug info: Remove an obsolete constructor from DebugLocEntry.
Adrian Prantl [Mon, 11 Aug 2014 21:06:03 +0000 (21:06 +0000)]
Debug info: Remove an obsolete constructor from DebugLocEntry.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215387 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDebug info: Modify DebugLocEntry::addValue to take multiple values so it
Adrian Prantl [Mon, 11 Aug 2014 21:06:00 +0000 (21:06 +0000)]
Debug info: Modify DebugLocEntry::addValue to take multiple values so it
only has to sort/unique values once per batch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215386 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDebug info: Further simplify the implementation of buildLocationList by
Adrian Prantl [Mon, 11 Aug 2014 21:05:57 +0000 (21:05 +0000)]
Debug info: Further simplify the implementation of buildLocationList by
getting rid of the redundant DIVariable in the OpenRanges pair.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215385 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDebug Info: Move the sorting and uniqueing of pieces from emitLocPieces()
Adrian Prantl [Mon, 11 Aug 2014 21:05:55 +0000 (21:05 +0000)]
Debug Info: Move the sorting and uniqueing of pieces from emitLocPieces()
into buildLocationList(). By keeping the list of Values sorted,
DebugLocEntry::Merge can also merge multi-piece entries.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215384 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoDebug info: Refactor DebugLocEntry's Merge function to make
Adrian Prantl [Mon, 11 Aug 2014 20:59:28 +0000 (20:59 +0000)]
Debug info: Refactor DebugLocEntry's Merge function to make
buildLocationLists easier to read.

The previous implementation conflated the merging of individual pieces
and the merging of entire DebugLocEntries.

By splitting this functionality into two separate functions the intention
of the code should be clearer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215383 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoARM: try harder to detect non-IT eligible instructions
Saleem Abdulrasool [Mon, 11 Aug 2014 20:13:25 +0000 (20:13 +0000)]
ARM: try harder to detect non-IT eligible instructions

For many Thumb-1 register register instructions, setting the CPSR is not
permitted inside an IT block.  We would not correctly flag those instructions.
The previous change to identify this scenario was insufficient as it did not
actually catch all the instances.  The current list is formed by manual
inspection of the ARMv6M ARM.

The change to the Thumb2 IT block test is due to the fact that the new more
stringent checking of the MIs results in the If Conversion pass being prevented
from executing (since not all the instructions in the BB are predicable).  This
results in code gen changes.

Thanks to Tim Northover for pointing out that the previous patch was
insufficient and hinting that the use of the v6M ARM would be much easier to use
than the v7 or v8!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215382 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix using -plugin-opt=apiflie when also using -plugin-opt=emit-llvm.
Rafael Espindola [Mon, 11 Aug 2014 19:06:54 +0000 (19:06 +0000)]
Fix using -plugin-opt=apiflie when also using -plugin-opt=emit-llvm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215378 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoCorrect a missing RUN line in the ARM codegen test for fneg ops. We should also expli...
Sanjay Patel [Mon, 11 Aug 2014 19:04:28 +0000 (19:04 +0000)]
Correct a missing RUN line in the ARM codegen test for fneg ops. We should also explicitly specify +/-neonfp.

The bug was introduced at r99570 when use of "-arm-use-neon-fp" was removed.

Differential Revision: http://reviews.llvm.org/D4846

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215377 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAdd missing test for r215031
Reid Kleckner [Mon, 11 Aug 2014 18:34:54 +0000 (18:34 +0000)]
Add missing test for r215031

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215374 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMC: Diagnose an unexpected token in COFF .section instead of asserting
Reid Kleckner [Mon, 11 Aug 2014 18:34:43 +0000 (18:34 +0000)]
MC: Diagnose an unexpected token in COFF .section instead of asserting

This can easily arise when trying to assemble and ELF style .section
directive for a COFF object file.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215373 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix typos:
Sylvestre Ledru [Mon, 11 Aug 2014 18:04:46 +0000 (18:04 +0000)]
Fix typos:
* libaries => libraries
* avaiable => available

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215366 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoFix use of uninitialized variable.
Rafael Espindola [Mon, 11 Aug 2014 17:07:34 +0000 (17:07 +0000)]
Fix use of uninitialized variable.

Fixes linking bitcode files that use the new style comdats for constructors
with ones that don't.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215364 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoUse an early return. NFC.
Rafael Espindola [Mon, 11 Aug 2014 16:55:42 +0000 (16:55 +0000)]
Use an early return. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215363 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert r215359 - [mips] Implement .ent, .end, .frame, .mask and .fmask assembler...
Daniel Sanders [Mon, 11 Aug 2014 16:10:19 +0000 (16:10 +0000)]
Revert r215359 - [mips] Implement .ent, .end, .frame, .mask and .fmask assembler directives

It seems to cause an lld test (elf/Mips/hilo16-3.test) to fail. Reverted while we investigate.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215361 91177308-0d34-0410-b5e6-96231b3b80d8

9 years ago[mips] Implement .ent, .end, .frame, .mask and .fmask assembler directives
Daniel Sanders [Mon, 11 Aug 2014 15:28:56 +0000 (15:28 +0000)]
[mips] Implement .ent, .end, .frame, .mask and .fmask assembler directives

Patch by Matheus Almeida and Toma Tabacu

Differential Revision: http://reviews.llvm.org/D4179

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215359 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoPeepholeOptimizer: make parameter ref to SmallPtrSetImpl
Hans Wennborg [Mon, 11 Aug 2014 13:52:46 +0000 (13:52 +0000)]
PeepholeOptimizer: make parameter ref to SmallPtrSetImpl

This makes the function type independent of the in-line size
of LocalMIs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215356 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoMake this SmallVector size a power of two as suggested by Chandler
Hans Wennborg [Mon, 11 Aug 2014 13:47:57 +0000 (13:47 +0000)]
Make this SmallVector size a power of two as suggested by Chandler

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215355 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAArch64: add support for dynamic-loader relocations
Tim Northover [Mon, 11 Aug 2014 10:10:27 +0000 (10:10 +0000)]
AArch64: add support for dynamic-loader relocations

LLD needs them, and it's good to be able to print them properly when
our object dumpers encounter them.

Patch by Daniel Stewart.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215352 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoAVX-512: added a missing bitcast from v16f32 to v16i32
Elena Demikhovsky [Mon, 11 Aug 2014 09:59:08 +0000 (09:59 +0000)]
AVX-512: added a missing bitcast from v16f32 to v16i32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215351 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agollvm-readobj: zero out timestamp in COFF auto-generated test files.
Tim Northover [Mon, 11 Aug 2014 09:53:07 +0000 (09:53 +0000)]
llvm-readobj: zero out timestamp in COFF auto-generated test files.

The timestamp meant these files changed with each invocation of
relocs.py, confusing matters when we add relocations and need to
update the tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215350 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agollvm-objdump: fix remaining use of %x format specifier for 64-bit values
Tim Northover [Mon, 11 Aug 2014 09:14:06 +0000 (09:14 +0000)]
llvm-objdump: fix remaining use of %x format specifier for 64-bit values

Third time lucky. This should finally fix the ARM (& MIPS, I think) bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215349 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoARM: __gnu_h2f_ieee and __gnu_f2h_ieee always use the soft-float calling convention
Oliver Stannard [Mon, 11 Aug 2014 09:12:32 +0000 (09:12 +0000)]
ARM: __gnu_h2f_ieee and __gnu_f2h_ieee always use the soft-float calling convention

By default, LLVM uses the "C" calling convention for all runtime
library functions. The half-precision FP conversion functions use the
soft-float calling convention, and are needed for some targets which
use the hard-float convention by default, so must have their calling
convention explicitly set.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215348 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoIn Machine CSE pass, the source register of a COPY machine instruction can
Jiangning Liu [Mon, 11 Aug 2014 05:17:19 +0000 (05:17 +0000)]
In Machine CSE pass, the source register of a COPY machine instruction can
be propagated to all its users, and this propagation could increase the
probability of finding common subexpressions. If the COPY has only one user,
the COPY itself can be removed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215344 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoIn LVI(Lazy Value Info), originally value on a BB can only be caculated once,
Jiangning Liu [Mon, 11 Aug 2014 05:02:04 +0000 (05:02 +0000)]
In LVI(Lazy Value Info), originally value on a BB can only be caculated once,
and the lattice will be updated to be a state other than "undefined". This
limiation could miss some opportunities of lowering "overdefined" to be an
even accurate value. So this patch ask the algorithm to try to lower the
lattice value again even if the value has been lowered to be "overdefined".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215343 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRe-commit "Increase the size of this SmallVector in PeepholeOptimizer." (r215340)
Hans Wennborg [Mon, 11 Aug 2014 02:50:43 +0000 (02:50 +0000)]
Re-commit "Increase the size of this SmallVector in PeepholeOptimizer." (r215340)

This time, also update the function that receives a reference to the SmallPtrSet as
a parameter.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215342 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoRevert "Increase the size of this SmallVector in PeepholeOptimizer." (r215340)
Hans Wennborg [Mon, 11 Aug 2014 02:34:52 +0000 (02:34 +0000)]
Revert "Increase the size of this SmallVector in PeepholeOptimizer." (r215340)

That broke the build:

/data/buildslave/clang-amd64-freebsd/src-llvm/lib/CodeGen/PeepholeOptimizer.cpp:729:46: error: non-const lvalue reference to type 'SmallPtrSet<[...], 8>' cannot bind to a value of unrelated type 'SmallPtrSet<[...], 16>'
        Changed |= optimizeExtInstr(MI, MBB, LocalMIs);
                                             ^~~~~~~~
/data/buildslave/clang-amd64-freebsd/src-llvm/lib/CodeGen/PeepholeOptimizer.cpp:265:49: note: passing argument to parameter 'LocalMIs' here
                 SmallPtrSet<MachineInstr*, 8> &LocalMIs) {
                                                ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215341 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoIncrease the size of this SmallVector in PeepholeOptimizer.
Hans Wennborg [Mon, 11 Aug 2014 02:21:34 +0000 (02:21 +0000)]
Increase the size of this SmallVector in PeepholeOptimizer.

During a Clang build, the median size of this was 9

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215340 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoSubTargetFeature.cpp: it seems the size of this SmallVector should be 3
Hans Wennborg [Mon, 11 Aug 2014 02:21:32 +0000 (02:21 +0000)]
SubTargetFeature.cpp: it seems the size of this SmallVector should be 3

because some subtarget feature strings have three components.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215339 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoIncrease the size of SpillPlacement::BlockFrequencies.
Hans Wennborg [Mon, 11 Aug 2014 02:21:30 +0000 (02:21 +0000)]
Increase the size of SpillPlacement::BlockFrequencies.

This SmallVector's median size during a Clang build was 7.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215338 91177308-0d34-0410-b5e6-96231b3b80d8

9 years agoLegacyPassManagers.h: increase the size of PMDataManager::HigherLevelAnalysis.
Hans Wennborg [Mon, 11 Aug 2014 02:21:25 +0000 (02:21 +0000)]
LegacyPassManagers.h: increase the size of PMDataManager::HigherLevelAnalysis.

In a Clang bootstrap, the median and max size was 9.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@215337 91177308-0d34-0410-b5e6-96231b3b80d8