From 7bd840d0586c75068e813149c2b7aa3605e63c67 Mon Sep 17 00:00:00 2001 From: Chandler Carruth Date: Thu, 26 Feb 2015 21:29:06 +0000 Subject: [PATCH] [x86] Restructure the comments and the conditions for handling dynamic blends. This makes it much more clear what is going on. The case we're handling is that of dynamic conditions, and we're bailing when the nature of the vector types and subtarget preclude lowering the dynamic condition vselect as an actual blend. No functionality changed here, but this will make a subsequent bug-fix to this code much more clear. llvm-svn: 230690 --- llvm/lib/Target/X86/X86ISelLowering.cpp | 32 +++++++++++++++---------- 1 file changed, 19 insertions(+), 13 deletions(-) diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 6265e79891da..656b32ec5563 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -20759,21 +20759,12 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, return Shuffle; } - // If we know that this node is legal then we know that it is going to be - // matched by one of the SSE/AVX BLEND instructions. These instructions only - // depend on the highest bit in each word. Try to use SimplifyDemandedBits - // to simplify previous instructions. + // If this is a *dynamic* select (non-constant condition) and we can match + // this node with one of the variable blend instructions, restructure the + // condition so that the blends can use the high bit of each element and use + // SimplifyDemandedBits to simplify the condition operand. if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() && !DCI.isBeforeLegalize() && - // We explicitly check against SSE4.1, v8i16 and v16i16 because, although - // vselect nodes may be marked as Custom, they might only be legal when - // Cond is a build_vector of constants. This will be taken care in - // a later condition. - (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) && - Subtarget->hasSSE41() && VT != MVT::v16i16 && VT != MVT::v8i16) && - // Don't optimize vector of constants. Those are handled by - // the generic code and all the bits must be properly set for - // the generic optimizer. !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) { unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits(); @@ -20781,6 +20772,21 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, if (BitWidth == 1) return SDValue(); + // We can only handle the cases where VSELECT is directly legal on the + // subtarget. We custom lower VSELECT nodes with constant conditions and + // this makes it hard to see whether a dynamic VSELECT will correctly + // lower, so we both check the operation's status and explicitly handle the + // cases where a *dynamic* blend will fail even though a constant-condition + // blend could be custom lowered. + // FIXME: We should find a better way to handle this class of problems. + // Potentially, we should combine constant-condition vselect nodes + // pre-legalization into shuffles and not mark as many types as custom + // lowered. + if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT)) + return SDValue(); + if (!Subtarget->hasSSE41() || VT == MVT::v16i16 || VT == MVT::v8i16) + return SDValue(); + assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size"); APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);