From 57760d96e2dfb485dc53fe2799df24bd18157abb Mon Sep 17 00:00:00 2001 From: Duncan Sands Date: Tue, 28 Oct 2008 15:00:32 +0000 Subject: [PATCH] Fix darwin ppc llvm-gcc build breakage: intercept ppcf128 to i32 conversion and expand it into a code sequence like in LegalizeDAG. This needs custom ppc lowering of FP_ROUND_INREG, so turn that on and make it work with LegalizeTypes. Probably PPC should simply custom lower the original conversion. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58329 91177308-0d34-0410-b5e6-96231b3b80d8 --- .../SelectionDAG/LegalizeFloatTypes.cpp | 35 +++++++++++++++++++ lib/Target/PowerPC/PPCISelLowering.cpp | 12 ++++--- test/CodeGen/PowerPC/2008-10-28-f128-i32.ll | 33 +++++++++++++++++ 3 files changed, 75 insertions(+), 5 deletions(-) create mode 100644 test/CodeGen/PowerPC/2008-10-28-f128-i32.ll diff --git a/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp index f1343886afe..c3b2d4082cb 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp @@ -902,6 +902,18 @@ SDValue DAGTypeLegalizer::ExpandFloatOp_FP_ROUND(SDNode *N) { SDValue DAGTypeLegalizer::ExpandFloatOp_FP_TO_SINT(SDNode *N) { MVT RVT = N->getValueType(0); + + // Expand ppcf128 to i32 by hand for the benefit of llvm-gcc bootstrap on + // PPC (the libcall is not available). FIXME: Do this in a less hacky way. + if (RVT == MVT::i32) { + assert(N->getOperand(0).getValueType() == MVT::ppcf128 && + "Logic only correct for ppcf128!"); + SDValue Res = DAG.getNode(ISD::FP_ROUND_INREG, MVT::ppcf128, + N->getOperand(0), DAG.getValueType(MVT::f64)); + Res = DAG.getNode(ISD::FP_ROUND, MVT::f64, Res, DAG.getIntPtrConstant(1)); + return DAG.getNode(ISD::FP_TO_SINT, MVT::i32, Res); + } + RTLIB::Libcall LC = RTLIB::getFPTOSINT(N->getOperand(0).getValueType(), RVT); assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported FP_TO_SINT!"); return MakeLibCall(LC, RVT, &N->getOperand(0), 1, false); @@ -909,6 +921,29 @@ SDValue DAGTypeLegalizer::ExpandFloatOp_FP_TO_SINT(SDNode *N) { SDValue DAGTypeLegalizer::ExpandFloatOp_FP_TO_UINT(SDNode *N) { MVT RVT = N->getValueType(0); + + // Expand ppcf128 to i32 by hand for the benefit of llvm-gcc bootstrap on + // PPC (the libcall is not available). FIXME: Do this in a less hacky way. + if (RVT == MVT::i32) { + assert(N->getOperand(0).getValueType() == MVT::ppcf128 && + "Logic only correct for ppcf128!"); + const uint64_t TwoE31[] = {0x41e0000000000000LL, 0}; + APFloat APF = APFloat(APInt(128, 2, TwoE31)); + SDValue Tmp = DAG.getConstantFP(APF, MVT::ppcf128); + // X>=2^31 ? (int)(X-2^31)+0x80000000 : (int)X + // FIXME: generated code sucks. + return DAG.getNode(ISD::SELECT_CC, MVT::i32, N->getOperand(0), Tmp, + DAG.getNode(ISD::ADD, MVT::i32, + DAG.getNode(ISD::FP_TO_SINT, MVT::i32, + DAG.getNode(ISD::FSUB, + MVT::ppcf128, + N->getOperand(0), + Tmp)), + DAG.getConstant(0x80000000, MVT::i32)), + DAG.getNode(ISD::FP_TO_SINT, MVT::i32, N->getOperand(0)), + DAG.getCondCode(ISD::SETGE)); + } + RTLIB::Libcall LC = RTLIB::getFPTOUINT(N->getOperand(0).getValueType(), RVT); assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported FP_TO_UINT!"); return MakeLibCall(LC, N->getValueType(0), &N->getOperand(0), 1, false); diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp index eb327dcdf23..5cf7eb28ef9 100644 --- a/lib/Target/PowerPC/PPCISelLowering.cpp +++ b/lib/Target/PowerPC/PPCISelLowering.cpp @@ -2865,9 +2865,10 @@ SDValue PPCTargetLowering::LowerFP_ROUND_INREG(SDValue Op, assert(Op.getValueType() == MVT::ppcf128); SDNode *Node = Op.getNode(); assert(Node->getOperand(0).getValueType() == MVT::ppcf128); - assert(Node->getOperand(0).getNode()->getOpcode() == ISD::BUILD_PAIR); - SDValue Lo = Node->getOperand(0).getNode()->getOperand(0); - SDValue Hi = Node->getOperand(0).getNode()->getOperand(1); + SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::f64, Node->getOperand(0), + DAG.getIntPtrConstant(0)); + SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::f64, Node->getOperand(0), + DAG.getIntPtrConstant(1)); // This sequence changes FPSCR to do round-to-zero, adds the two halves // of the long double, and puts FPSCR back the way it was. We do not @@ -2916,7 +2917,7 @@ SDValue PPCTargetLowering::LowerFP_ROUND_INREG(SDValue Op, // We know the low half is about to be thrown away, so just use something // convenient. - return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg); + return DAG.getNode(ISD::BUILD_PAIR, MVT::ppcf128, FPreg, FPreg); } SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) { @@ -3883,7 +3884,8 @@ SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) { SDNode *PPCTargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) { switch (N->getOpcode()) { - default: assert(0 && "Wasn't expecting to be able to lower this!"); + default: + return PPCTargetLowering::LowerOperation(SDValue (N, 0), DAG).getNode(); case ISD::FP_TO_SINT: { SDValue Res = LowerFP_TO_SINT(SDValue(N, 0), DAG); // Use MERGE_VALUES to drop the chain result value and get a node with one diff --git a/test/CodeGen/PowerPC/2008-10-28-f128-i32.ll b/test/CodeGen/PowerPC/2008-10-28-f128-i32.ll new file mode 100644 index 00000000000..4db577357da --- /dev/null +++ b/test/CodeGen/PowerPC/2008-10-28-f128-i32.ll @@ -0,0 +1,33 @@ +; RUN: llvm-as < %s | llc -march=ppc32 -o - | not grep fixunstfsi + +define i64 @__fixunstfdi(ppc_fp128 %a) nounwind readnone { +entry: + %0 = fcmp olt ppc_fp128 %a, 0xM00000000000000000000000000000000 ; [#uses=1] + br i1 %0, label %bb5, label %bb1 + +bb1: ; preds = %entry + %1 = mul ppc_fp128 %a, 0xM3DF00000000000000000000000000000 ; [#uses=1] + %2 = fptoui ppc_fp128 %1 to i32 ; [#uses=1] + %3 = zext i32 %2 to i64 ; [#uses=1] + %4 = shl i64 %3, 32 ; [#uses=3] + %5 = uitofp i64 %4 to ppc_fp128 ; [#uses=1] + %6 = sub ppc_fp128 %a, %5 ; [#uses=3] + %7 = fcmp olt ppc_fp128 %6, 0xM00000000000000000000000000000000 ; [#uses=1] + br i1 %7, label %bb2, label %bb3 + +bb2: ; preds = %bb1 + %8 = sub ppc_fp128 0xM80000000000000000000000000000000, %6 ; [#uses=1] + %9 = fptoui ppc_fp128 %8 to i32 ; [#uses=1] + %10 = zext i32 %9 to i64 ; [#uses=1] + %11 = sub i64 %4, %10 ; [#uses=1] + ret i64 %11 + +bb3: ; preds = %bb1 + %12 = fptoui ppc_fp128 %6 to i32 ; [#uses=1] + %13 = zext i32 %12 to i64 ; [#uses=1] + %14 = or i64 %13, %4 ; [#uses=1] + ret i64 %14 + +bb5: ; preds = %entry + ret i64 0 +}