[llvm] r194542 - SelectionDAG: Teach the legalizer to split SETCC if VSELECT needs splitting too.
Owen Anderson
resistor at mac.com
Wed Nov 20 23:30:06 PST 2013
I'm fine with both of these patches.
--Owen
On Nov 20, 2013, at 7:35 PM, Tom Stellard <tom at stellard.net> wrote:
> On Tue, Nov 19, 2013 at 06:40:26PM -0800, Juergen Ributzka wrote:
>> Hi Tom
>>
>> Great, that makes things much easier. I rewrote the patch and it shouldn’t require any changes in the backend.
>> Please let me know if this patch works for you.
>
> cc'ing Owen, because I would like to get these patches into the 3.4
> branch.
>
> Juergen,
>
> Your patch does fix the crash, but there is still a performance
> regression in the test case I provided. I have attached a patch to fix
> this performance regression.
>
> Owen,
>
> There are two patches attached to this email that I would like to merge
> into the 3.4 branch. The first is a patch from Juergen which fixes a
> crash in the R600 backend. The crash is a regression caused by r194542.
>
> Even with the crash fixed there is still a performance regression which
> was uncovered by r194542. This performance issue is caused by the
> DAGLegalizer falling back to stack loads and stores when trying to
> expand certain kinds of bitcasts. The second patch improves the
> handling of bitcasts in the legalizer and fixes this regression.
>
> Thanks,
> Tom
>
>
>>
>> Thanks
>>
>> Cheers,
>> Juergen
>>
>>
>
>
>>
>> On Nov 19, 2013, at 3:31 PM, Tom Stellard <tom at stellard.net> wrote:
>>
>>> Hi Juergen,
>>>
>>> Sorry the test case is only meant for VLIW chips, it should work with
>>> these arguments to llc:
>>>
>>> llc -march=r600 -mcpu=redwood vselect.ll
>>>
>>> -Tom
>>>
>>>
>>> On Tue, Nov 19, 2013 at 03:16:08PM -0800, Juergen Ributzka wrote:
>>>> Hi Tom,
>>>>
>>>> even after disabling my changes I get the following error from your new test case during instruction selection:
>>>>
>>>> LLVM ERROR: Cannot select: 0x7ffc34825a10: i64 = select 0x7ffc34825910, 0x7ffc3480f510, 0x7ffc3480e210 [ORD=26] [ID=49]
>>>> 0x7ffc34825910: i1 = setcc 0x7ffc34825810, 0x7ffc34825710, 0x7ffc3480db10 [ORD=25] [ID=45]
>>>> 0x7ffc34825810: i32 = extract_vector_elt 0x7ffc34049810, 0x7ffc34049510 [ORD=25] [ID=41]
>>>> 0x7ffc34049810: v2i32 = BUILD_VECTOR 0x7ffc34810c10, 0x7ffc3480e610 [ORD=25] [ID=37]
>>>> 0x7ffc34810c10: i32 = extract_vector_elt 0x7ffc3480f110, 0x7ffc34049510 [ORD=24] [ID=32]
>>>> 0x7ffc3480f110: v4i32,ch = load 0x7ffc33c160d8, 0x7ffc34046910, 0x7ffc34810210<LD16[undef]> [ORD=24] [ID=27]
>>>> 0x7ffc34046910: i64 = add 0x7ffc34047310, 0x7ffc3480f410 [ORD=24] [ID=23]
>>>> 0x7ffc34047310: i64,ch = CopyFromReg 0x7ffc33c160d8, 0x7ffc34049910 [ORD=24] [ID=21]
>>>> 0x7ffc34049910: i64 = Register %vreg0 [ID=1]
>>>> 0x7ffc3480f410: i64 = Constant<52> [ID=4]
>>>> 0x7ffc34810210: i64 = undef [ID=3]
>>>> 0x7ffc34049510: i32 = Constant<0> [ID=5]
>>>> 0x7ffc3480e610: i32 = extract_vector_elt 0x7ffc3480f110, 0x7ffc34049a10 [ORD=24] [ID=31]
>>>> 0x7ffc3480f110: v4i32,ch = load 0x7ffc33c160d8, 0x7ffc34046910, 0x7ffc34810210<LD16[undef]> [ORD=24] [ID=27]
>>>> 0x7ffc34046910: i64 = add 0x7ffc34047310, 0x7ffc3480f410 [ORD=24] [ID=23]
>>>> 0x7ffc34047310: i64,ch = CopyFromReg 0x7ffc33c160d8, 0x7ffc34049910 [ORD=24] [ID=21]
>>>> 0x7ffc34049910: i64 = Register %vreg0 [ID=1]
>>>> 0x7ffc3480f410: i64 = Constant<52> [ID=4]
>>>> 0x7ffc34810210: i64 = undef [ID=3]
>>>> 0x7ffc34049a10: i32 = Constant<1> [ID=17]
>>>> 0x7ffc34049510: i32 = Constant<0> [ID=5]
>>>> 0x7ffc34825710: i32 = extract_vector_elt 0x7ffc3480f310, 0x7ffc34049510 [ORD=25] [ID=26]
>>>> 0x7ffc3480f310: v2i32 = BUILD_VECTOR 0x7ffc34049510, 0x7ffc34049510 [ORD=25] [ID=22]
>>>> 0x7ffc34049510: i32 = Constant<0> [ID=5]
>>>> 0x7ffc34049510: i32 = Constant<0> [ID=5]
>>>> 0x7ffc34049510: i32 = Constant<0> [ID=5]
>>>> 0x7ffc3480f510: i64 = Constant<0> [ID=7]
>>>> 0x7ffc3480e210: i64 = Constant<4> [ID=11]
>>>> In function: test_select_v4i64
>>>>
>>>> Do you get the same result?
>>>>
>>>> -Juergen
>>>>
>>>> On Nov 18, 2013, at 5:51 PM, Juergen Ributzka <juergen at apple.com> wrote:
>>>>
>>>>> Hi Tom,
>>>>>
>>>>> sorry for not coming back to you earlier. I was working on a patch for this problem this weekend, but then I stumbled across another issue that was exposed while working on the patch. The two pending patches for SelectionDAG that I posted to the mailing list are in preparation for fixing this issue. I looked at your second patch and using SplitVecRes_SETCC would fix the problem, but it also splits the SETCC result vector again, instead of using the already split vectors provided by GetSplitVector. GetSplitVector is just a cache of already split vectors for that given operation. I would prefer not to duplicate split vector nodes if they already exist. The issue for your backend is that split never occurred and that is why the assertion is failing. I hope to get this fixed for you soon.
>>>>>
>>>>> Cheers,
>>>>> Juergen
>>>>>
>>>>> On Nov 18, 2013, at 4:49 PM, Tom Stellard <tom at stellard.net> wrote:
>>>>>
>>>>>> Hi Juergen,
>>>>>>
>>>>>> These two patches fix the crash for me. The first one is an R600 fix
>>>>>> and the second is a fix for illegal SELECT with legal SETCC as its
>>>>>> condition. I'm not so sure patch number 2 is the right way to
>>>>>> fix the bug, but it was the most simple fix I could come up with.
>>>>>>
>>>>>> -Tom
>>>>>>
>>>>>>
>>>>>> On Fri, Nov 15, 2013 at 10:45:11AM -0800, Juergen Ributzka wrote:
>>>>>>> Hi Tom,
>>>>>>>
>>>>>>> I took a quick look yesterday night and I didn?t see a quick fix either. I will look into it today.
>>>>>>>
>>>>>>> Cheers,
>>>>>>> Juergen
>>>>>>>
>>>>>>>
>>>>>>> On Nov 14, 2013, at 8:45 PM, Tom Stellard <tom at stellard.net> wrote:
>>>>>>>
>>>>>>>> On Wed, Nov 13, 2013 at 01:57:54AM -0000, Juergen Ributzka wrote:
>>>>>>>>> Author: ributzka
>>>>>>>>> Date: Tue Nov 12 19:57:54 2013
>>>>>>>>> New Revision: 194542
>>>>>>>>>
>>>>>>>>> URL: http://llvm.org/viewvc/llvm-project?rev=194542&view=rev
>>>>>>>>> Log:
>>>>>>>>> SelectionDAG: Teach the legalizer to split SETCC if VSELECT needs splitting too.
>>>>>>>>>
>>>>>>>>> This patch reapplies r193676 with an additional fix for the Hexagon backend. The
>>>>>>>>> SystemZ backend has already been fixed by r194148.
>>>>>>>>>
>>>>>>>>> The Type Legalizer recognizes that VSELECT needs to be split, because the type
>>>>>>>>> is to wide for the given target. The same does not always apply to SETCC,
>>>>>>>>> because less space is required to encode the result of a comparison. As a result
>>>>>>>>> VSELECT is split and SETCC is unrolled into scalar comparisons.
>>>>>>>>>
>>>>>>>>> This commit fixes the issue by checking for VSELECT-SETCC patterns in the DAG
>>>>>>>>> Combiner. If a matching pattern is found, then the result mask of SETCC is
>>>>>>>>> promoted to the expected vector mask type for the given target. Now the type
>>>>>>>>> legalizer will split both VSELECT and SETCC.
>>>>>>>>>
>>>>>>>>> This allows the following X86 DAG Combine code to sucessfully detect the MIN/MAX
>>>>>>>>> pattern. This fixes PR16695, PR17002, and <rdar://problem/14594431>.
>>>>>>>>>
>>>>>>>>> Reviewed by Nadav
>>>>>>>>>
>>>>>>>>
>>>>>>>> Hi Juergen,
>>>>>>>>
>>>>>>>> This caused a regression on R600. I have attached a patch to this email
>>>>>>>> with a testcase. See my comments below.
>>>>>>>>
>>>>>>>>> Added:
>>>>>>>>> llvm/trunk/test/CodeGen/X86/vec_split.ll
>>>>>>>>> Modified:
>>>>>>>>> llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
>>>>>>>>> llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp
>>>>>>>>> llvm/trunk/lib/Target/Hexagon/HexagonISelLowering.h
>>>>>>>>> llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
>>>>>>>>>
>>>>>>>>> Modified: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
>>>>>>>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp?rev=194542&r1=194541&r2=194542&view=diff
>>>>>>>>> ==============================================================================
>>>>>>>>> --- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (original)
>>>>>>>>> +++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp Tue Nov 12 19:57:54 2013
>>>>>>>>> @@ -4364,6 +4364,29 @@ SDValue DAGCombiner::visitVSELECT(SDNode
>>>>>>>>> }
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> + // Treat SETCC as a vector mask and promote the result type based on the
>>>>>>>>> + // targets expected SETCC result type. This will ensure that SETCC and VSELECT
>>>>>>>>> + // are both split by the type legalizer. This is done to prevent the type
>>>>>>>>> + // legalizer from unrolling SETCC into scalar comparions.
>>>>>>>>> + EVT SelectVT = N->getValueType(0);
>>>>>>>>> + EVT MaskVT = getSetCCResultType(SelectVT);
>>>>>>>>> + assert(MaskVT.isVector() && "Expected a vector type.");
>>>>>>>>> + if (N0.getOpcode() == ISD::SETCC && N0.getValueType() != MaskVT) {
>>>>>>>>> + SDLoc MaskDL(N0);
>>>>>>>>> +
>>>>>>>>> + // Extend the mask to the desired value type.
>>>>>>>>> + ISD::NodeType ExtendCode =
>>>>>>>>> + TargetLowering::getExtendForContent(TLI.getBooleanContents(true));
>>>>>>>>> + SDValue Mask = DAG.getNode(ExtendCode, MaskDL, MaskVT, N0);
>>>>>>>>> +
>>>>>>>>> + AddToWorkList(Mask.getNode());
>>>>>>>>> +
>>>>>>>>> + SDValue LHS = N->getOperand(1);
>>>>>>>>> + SDValue RHS = N->getOperand(2);
>>>>>>>>> +
>>>>>>>>> + return DAG.getNode(ISD::VSELECT, DL, SelectVT, Mask, LHS, RHS);
>>>>>>>>> + }
>>>>>>>>> +
>>>>>>>>> return SDValue();
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Modified: llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp
>>>>>>>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp?rev=194542&r1=194541&r2=194542&view=diff
>>>>>>>>> ==============================================================================
>>>>>>>>> --- llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp (original)
>>>>>>>>> +++ llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp Tue Nov 12 19:57:54 2013
>>>>>>>>> @@ -492,14 +492,19 @@ void DAGTypeLegalizer::SplitRes_SELECT(S
>>>>>>>>> SDValue Cond = N->getOperand(0);
>>>>>>>>> CL = CH = Cond;
>>>>>>>>> if (Cond.getValueType().isVector()) {
>>>>>>>>> - assert(Cond.getValueType().getVectorElementType() == MVT::i1 &&
>>>>>>>>> - "Condition legalized before result?");
>>>>>>>>> - unsigned NumElements = Cond.getValueType().getVectorNumElements();
>>>>>>>>> - EVT VCondTy = EVT::getVectorVT(*DAG.getContext(), MVT::i1, NumElements / 2);
>>>>>>>>> - CL = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VCondTy, Cond,
>>>>>>>>> - DAG.getConstant(0, TLI.getVectorIdxTy()));
>>>>>>>>> - CH = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VCondTy, Cond,
>>>>>>>>> - DAG.getConstant(NumElements / 2, TLI.getVectorIdxTy()));
>>>>>>>>> + if (Cond.getOpcode() == ISD::SETCC) {
>>>>>>>>> + assert(Cond.getValueType() == getSetCCResultType(N->getValueType(0)) &&
>>>>>>>>> + "Condition has not been prepared for split!");
>>>>>>>>> + GetSplitVector(Cond, CL, CH);
>>>>>>>>
>>>>>>>> Even with the adjustment to getSetCCResultType in the attached patch,
>>>>>>>> when GetSplitVector() is called here, I am getting an assertion failure:
>>>>>>>>
>>>>>>>> llc: LegalizeTypes.cpp:828: void llvm::DAGTypeLegalizer::GetSplitVector(llvm::SDValue, llvm::SDValue &, llvm::SDValue &): Assertion `Entry.first.getNode() && "Operand isn't split"' failed.
>>>>>>>>
>>>>>>>> From what I can tell the problem here is that 'Cond' has a value type
>>>>>>>> of v4i32 which is legal on R600, so the operation was not split, and
>>>>>>>> therefore there is no entry in the SplitVectors map for it (This is what
>>>>>>>> is causing the assertion failure).
>>>>>>>>
>>>>>>>> The SELECT in this example has a type of v4i64 which is not a legal type
>>>>>>>> on R600, so the problem occurs when a SELECT node has an illegal type,
>>>>>>>> but Operand 0 is a SETCC node with a legal type. I'm not sure the best
>>>>>>>> way to fix this issue. Do you have any suggestions?
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>> Tom
>>>>>>>>
>>>>>>>>
>>>>>>>>> + } else {
>>>>>>>>> + EVT ETy = Cond.getValueType().getVectorElementType();
>>>>>>>>> + unsigned NumElements = Cond.getValueType().getVectorNumElements();
>>>>>>>>> + EVT VCondTy = EVT::getVectorVT(*DAG.getContext(), ETy, NumElements / 2);
>>>>>>>>> + CL = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VCondTy, Cond,
>>>>>>>>> + DAG.getConstant(0, TLI.getVectorIdxTy()));
>>>>>>>>> + CH = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VCondTy, Cond,
>>>>>>>>> + DAG.getConstant(NumElements / 2, TLI.getVectorIdxTy()));
>>>>>>>>> + }
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> Lo = DAG.getNode(N->getOpcode(), dl, LL.getValueType(), CL, LL, RL);
>>>>>>>>>
>>>>>>>>> Modified: llvm/trunk/lib/Target/Hexagon/HexagonISelLowering.h
>>>>>>>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/Hexagon/HexagonISelLowering.h?rev=194542&r1=194541&r2=194542&view=diff
>>>>>>>>> ==============================================================================
>>>>>>>>> --- llvm/trunk/lib/Target/Hexagon/HexagonISelLowering.h (original)
>>>>>>>>> +++ llvm/trunk/lib/Target/Hexagon/HexagonISelLowering.h Tue Nov 12 19:57:54 2013
>>>>>>>>> @@ -141,8 +141,11 @@ namespace llvm {
>>>>>>>>>
>>>>>>>>> SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
>>>>>>>>> SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
>>>>>>>>> - virtual EVT getSetCCResultType(LLVMContext &, EVT) const {
>>>>>>>>> - return MVT::i1;
>>>>>>>>> + virtual EVT getSetCCResultType(LLVMContext &C, EVT VT) const {
>>>>>>>>> + if (!VT.isVector())
>>>>>>>>> + return MVT::i1;
>>>>>>>>> + else
>>>>>>>>> + return EVT::getVectorVT(C, MVT::i1, VT.getVectorNumElements());
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
>>>>>>>>>
>>>>>>>>> Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
>>>>>>>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=194542&r1=194541&r2=194542&view=diff
>>>>>>>>> ==============================================================================
>>>>>>>>> --- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
>>>>>>>>> +++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Tue Nov 12 19:57:54 2013
>>>>>>>>> @@ -1547,7 +1547,16 @@ void X86TargetLowering::resetOperationAc
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
>>>>>>>>> - if (!VT.isVector()) return MVT::i8;
>>>>>>>>> + if (!VT.isVector())
>>>>>>>>> + return MVT::i8;
>>>>>>>>> +
>>>>>>>>> + const TargetMachine &TM = getTargetMachine();
>>>>>>>>> + if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512())
>>>>>>>>> + switch(VT.getVectorNumElements()) {
>>>>>>>>> + case 8: return MVT::v8i1;
>>>>>>>>> + case 16: return MVT::v16i1;
>>>>>>>>> + }
>>>>>>>>> +
>>>>>>>>> return VT.changeVectorElementTypeToInteger();
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Added: llvm/trunk/test/CodeGen/X86/vec_split.ll
>>>>>>>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vec_split.ll?rev=194542&view=auto
>>>>>>>>> ==============================================================================
>>>>>>>>> --- llvm/trunk/test/CodeGen/X86/vec_split.ll (added)
>>>>>>>>> +++ llvm/trunk/test/CodeGen/X86/vec_split.ll Tue Nov 12 19:57:54 2013
>>>>>>>>> @@ -0,0 +1,42 @@
>>>>>>>>> +; RUN: llc -march=x86-64 -mcpu=corei7 < %s | FileCheck %s -check-prefix=SSE4
>>>>>>>>> +; RUN: llc -march=x86-64 -mcpu=corei7-avx < %s | FileCheck %s -check-prefix=AVX1
>>>>>>>>> +; RUN: llc -march=x86-64 -mcpu=core-avx2 < %s | FileCheck %s -check-prefix=AVX2
>>>>>>>>> +
>>>>>>>>> +define <16 x i16> @split16(<16 x i16> %a, <16 x i16> %b, <16 x i8> %__mask) {
>>>>>>>>> +; SSE4-LABEL: split16:
>>>>>>>>> +; SSE4: pminuw
>>>>>>>>> +; SSE4: pminuw
>>>>>>>>> +; SSE4: ret
>>>>>>>>> +; AVX1-LABEL: split16:
>>>>>>>>> +; AVX1: vpminuw
>>>>>>>>> +; AVX1: vpminuw
>>>>>>>>> +; AVX1: ret
>>>>>>>>> +; AVX2-LABEL: split16:
>>>>>>>>> +; AVX2: vpminuw
>>>>>>>>> +; AVX2: ret
>>>>>>>>> + %1 = icmp ult <16 x i16> %a, %b
>>>>>>>>> + %2 = select <16 x i1> %1, <16 x i16> %a, <16 x i16> %b
>>>>>>>>> + ret <16 x i16> %2
>>>>>>>>> +}
>>>>>>>>> +
>>>>>>>>> +define <32 x i16> @split32(<32 x i16> %a, <32 x i16> %b, <32 x i8> %__mask) {
>>>>>>>>> +; SSE4-LABEL: split32:
>>>>>>>>> +; SSE4: pminuw
>>>>>>>>> +; SSE4: pminuw
>>>>>>>>> +; SSE4: pminuw
>>>>>>>>> +; SSE4: pminuw
>>>>>>>>> +; SSE4: ret
>>>>>>>>> +; AVX1-LABEL: split32:
>>>>>>>>> +; AVX1: vpminuw
>>>>>>>>> +; AVX1: vpminuw
>>>>>>>>> +; AVX1: vpminuw
>>>>>>>>> +; AVX1: vpminuw
>>>>>>>>> +; AVX1: ret
>>>>>>>>> +; AVX2-LABEL: split32:
>>>>>>>>> +; AVX2: vpminuw
>>>>>>>>> +; AVX2: vpminuw
>>>>>>>>> +; AVX2: ret
>>>>>>>>> + %1 = icmp ult <32 x i16> %a, %b
>>>>>>>>> + %2 = select <32 x i1> %1, <32 x i16> %a, <32 x i16> %b
>>>>>>>>> + ret <32 x i16> %2
>>>>>>>>> +}
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> _______________________________________________
>>>>>>>>> llvm-commits mailing list
>>>>>>>>> llvm-commits at cs.uiuc.edu
>>>>>>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>>>>>> <0001-R600-Fix-getSetCCResultType-to-handle-new-VSELECT-DA.patch>
>>>>>>>
>>>>>> <0001-R600-Fix-getSetCCResultType-to-handle-new-VSELECT-DA.patch><0002-SelectionDAG-Fix-assertion-failure-when-splitting-ve.patch>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> llvm-commits mailing list
>>>>> llvm-commits at cs.uiuc.edu
>>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>>
>>
>
> <0001-Split-SETCC-if-VSELECT-requires-splitting-too.patch><0002-SelectionDAG-Optimize-expansion-of-vec_type-BITCAST-.patch>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20131120/580d28ae/attachment.html>
More information about the llvm-commits
mailing list