From 04e5518783fa3eb679cecc14c9ea3c3c60fa7cb4 Mon Sep 17 00:00:00 2001 From: Evan Cheng Date: Wed, 12 Dec 2012 00:42:09 +0000 Subject: Avoid using lossy load / stores for memcpy / memset expansion. e.g. f64 load / store on non-SSE2 x86 targets. llvm-svn: 169944 --- llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 36 +++++++++++++++----------- 1 file changed, 21 insertions(+), 15 deletions(-) (limited to 'llvm/lib/CodeGen') diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp index 4cb63ce9a66..36592e54f8f 100644 --- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp @@ -3474,27 +3474,33 @@ static bool FindOptimalMemOpLowering(std::vector &MemOps, unsigned VTSize = VT.getSizeInBits() / 8; while (VTSize > Size) { // For now, only use non-vector load / store's for the left-over pieces. - EVT NewVT; + EVT NewVT = VT; unsigned NewVTSize; + + bool Found = false; if (VT.isVector() || VT.isFloatingPoint()) { NewVT = (VT.getSizeInBits() > 64) ? MVT::i64 : MVT::i32; - while (!TLI.isOperationLegalOrCustom(ISD::STORE, NewVT)) { - if (NewVT == MVT::i64 && - TLI.isOperationLegalOrCustom(ISD::STORE, MVT::f64)) { - // i64 is usually not legal on 32-bit targets, but f64 may be. - NewVT = MVT::f64; - break; - } - NewVT = (MVT::SimpleValueType)(NewVT.getSimpleVT().SimpleTy - 1); + if (TLI.isOperationLegalOrCustom(ISD::STORE, NewVT) && + TLI.isLegalMemOpType(NewVT.getSimpleVT())) + Found = true; + else if (NewVT == MVT::i64 && + TLI.isOperationLegalOrCustom(ISD::STORE, MVT::f64) && + TLI.isLegalMemOpType(MVT::f64)) { + // i64 is usually not legal on 32-bit targets, but f64 may be. + NewVT = MVT::f64; + Found = true; } - NewVTSize = NewVT.getSizeInBits() / 8; - } else { - // This can result in a type that is not legal on the target, e.g. - // 1 or 2 bytes on PPC. - NewVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); - NewVTSize = VTSize >> 1; } + if (!Found) { + do { + NewVT = (MVT::SimpleValueType)(NewVT.getSimpleVT().SimpleTy - 1); + if (NewVT == MVT::i8) + break; + } while (!TLI.isLegalMemOpType(NewVT.getSimpleVT())); + } + NewVTSize = NewVT.getSizeInBits() / 8; + // If the new VT cannot cover all of the remaining bits, then consider // issuing a (or a pair of) unaligned and overlapping load / store. // FIXME: Only does this for 64-bit or more since we don't have proper -- cgit v1.2.3