Implement codegen for flt_rounds on x86

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44183 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 75 insertions, 3 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 27decaa..9a86561 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -205,7 +205,8 @@ X86TargetLowering::X86TargetLowering(TargetMachine &TM)
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1   , Expand);
   setOperationAction(ISD::FP_ROUND_INREG   , MVT::f32  , Expand);
   setOperationAction(ISD::FREM             , MVT::f64  , Expand);
-
+  setOperationAction(ISD::FLT_ROUNDS       , MVT::i32  , Custom);
+  
   setOperationAction(ISD::CTPOP            , MVT::i8   , Expand);
   setOperationAction(ISD::CTTZ             , MVT::i8   , Expand);
   setOperationAction(ISD::CTLZ             , MVT::i8   , Expand);
@@ -4940,6 +4941,66 @@ SDOperand X86TargetLowering::LowerTRAMPOLINE(SDOperand Op,
   }
 }
 
+SDOperand X86TargetLowering::LowerFLT_ROUNDS(SDOperand Op, SelectionDAG &DAG) {
+  /*
+   The rounding mode is in bits 11:10 of FPSR, and has the following
+   settings:
+     00 Round to nearest
+     01 Round to -inf
+     10 Round to +inf
+     11 Round to 0
+
+  FLT_ROUNDS, on the other hand, expects the following:
+    -1 Undefined
+     0 Round to 0
+     1 Round to nearest
+     2 Round to +inf
+     3 Round to -inf
+
+  To perform the conversion, we do:
+    (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
+  */
+
+  MachineFunction &MF = DAG.getMachineFunction();
+  const TargetMachine &TM = MF.getTarget();
+  const TargetFrameInfo &TFI = *TM.getFrameInfo();
+  unsigned StackAlignment = TFI.getStackAlignment();
+  MVT::ValueType VT = Op.getValueType();
+
+  // Save FP Control Word to stack slot
+  int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
+  SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
+
+  SDOperand Chain = DAG.getNode(X86ISD::FNSTCW16m, MVT::Other,
+                                DAG.getEntryNode(), StackSlot);
+
+  // Load FP Control Word from stack slot
+  SDOperand CWD = DAG.getLoad(MVT::i16, Chain, StackSlot, NULL, 0);
+
+  // Transform as necessary
+  SDOperand CWD1 =
+    DAG.getNode(ISD::SRL, MVT::i16,
+                DAG.getNode(ISD::AND, MVT::i16,
+                            CWD, DAG.getConstant(0x800, MVT::i16)),
+                DAG.getConstant(11, MVT::i8));
+  SDOperand CWD2 =
+    DAG.getNode(ISD::SRL, MVT::i16,
+                DAG.getNode(ISD::AND, MVT::i16,
+                            CWD, DAG.getConstant(0x400, MVT::i16)),
+                DAG.getConstant(9, MVT::i8));
+
+  SDOperand RetVal =
+    DAG.getNode(ISD::AND, MVT::i16,
+                DAG.getNode(ISD::ADD, MVT::i16,
+                            DAG.getNode(ISD::OR, MVT::i16, CWD1, CWD2),
+                            DAG.getConstant(1, MVT::i16)),
+                DAG.getConstant(3, MVT::i16));
+
+
+  return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
+                      ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
+}
+
 /// LowerOperation - Provide custom lowering hooks for some operations.
 ///
 SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
@@ -4982,6 +5043,7 @@ SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
   case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
   case ISD::EH_RETURN:          return LowerEH_RETURN(Op, DAG);
   case ISD::TRAMPOLINE:         return LowerTRAMPOLINE(Op, DAG);
+  case ISD::FLT_ROUNDS:         return LowerFLT_ROUNDS(Op, DAG);
   }
   return SDOperand();
 }
@@ -5029,6 +5091,7 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case X86ISD::THREAD_POINTER:     return "X86ISD::THREAD_POINTER";
   case X86ISD::EH_RETURN:          return "X86ISD::EH_RETURN";
   case X86ISD::TC_RETURN:          return "X86ISD::TC_RETURN";
+  case X86ISD::FNSTCW16m:          return "X86ISD::FNSTCW16m";
   }
 }
 
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 2921f7b..ac16ff4 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -188,7 +188,10 @@ namespace llvm {
       //   operand #1 callee (register or absolute)
       //   operand #2 stack adjustment
       //   operand #3 optional in flag
-      TC_RETURN
+      TC_RETURN,
+
+      // Store FP control world into i16 memory
+      FNSTCW16m
     };
   }
 
@@ -477,6 +480,7 @@ namespace llvm {
     SDOperand LowerFRAME_TO_ARGS_OFFSET(SDOperand Op, SelectionDAG &DAG);
     SDOperand LowerEH_RETURN(SDOperand Op, SelectionDAG &DAG);
     SDOperand LowerTRAMPOLINE(SDOperand Op, SelectionDAG &DAG);
+    SDOperand LowerFLT_ROUNDS(SDOperand Op, SelectionDAG &DAG);
   };
 }
 
diff --git a/lib/Target/X86/X86InstrFPStack.td b/lib/Target/X86/X86InstrFPStack.td
index 8ce8cdf..d2feb42 100644
--- a/lib/Target/X86/X86InstrFPStack.td
+++ b/lib/Target/X86/X86InstrFPStack.td
@@ -29,6 +29,8 @@ def SDTX86Fild      : SDTypeProfile<1, 2, [SDTCisFP<0>, SDTCisPtrTy<1>,
                                            SDTCisVT<2, OtherVT>]>;
 def SDTX86FpToIMem  : SDTypeProfile<0, 2, [SDTCisFP<0>, SDTCisPtrTy<1>]>;
 
+def SDTX86CwdStore  : SDTypeProfile<0, 1, [SDTCisPtrTy<0>]>;
+
 def X86fpget        : SDNode<"X86ISD::FP_GET_RESULT", SDTX86FpGet,
                         [SDNPHasChain, SDNPInFlag, SDNPOutFlag]>;
 def X86fpset        : SDNode<"X86ISD::FP_SET_RESULT", SDTX86FpSet,
@@ -47,6 +49,8 @@ def X86fp_to_i32mem : SDNode<"X86ISD::FP_TO_INT32_IN_MEM", SDTX86FpToIMem,
                         [SDNPHasChain]>;
 def X86fp_to_i64mem : SDNode<"X86ISD::FP_TO_INT64_IN_MEM", SDTX86FpToIMem,
                         [SDNPHasChain]>;
+def X86fp_cwd_get16 : SDNode<"X86ISD::FNSTCW16m",          SDTX86CwdStore,
+                        [SDNPHasChain]>;
 
 //===----------------------------------------------------------------------===//
 // FPStack pattern fragments
@@ -524,7 +528,8 @@ def FNSTSW8r  : I<0xE0, RawFrm,                  // AX = fp flags
                   (outs), (ins), "fnstsw", []>, DF;
 
 def FNSTCW16m : I<0xD9, MRM7m,                   // [mem16] = X87 control world
-                  (outs), (ins i16mem:$dst), "fnstcw\t$dst", []>;
+                  (outs), (ins i16mem:$dst), "fnstcw\t$dst",
+                  [(X86fp_cwd_get16 addr:$dst)]>;
 def FLDCW16m  : I<0xD9, MRM5m,                   // X87 control world = [mem16]
                   (outs), (ins i16mem:$dst), "fldcw\t$dst", []>;