[X86] Use xorps to create fp128 +0.0 constants.

This matches what we do for f32/f64. gcc also does this for fp128.

llvm-svn: 371357

5 files changed, 24 insertions, 4 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 2fa368764a5..0a8219214f4 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -628,6 +628,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
       setOperationAction(ISD::FABS , MVT::f128, Custom);
       setOperationAction(ISD::FNEG , MVT::f128, Custom);
       setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom);
+
+      addLegalFPImmediate(APFloat::getZero(APFloat::IEEEquad())); // xorps
     }
 
     addRegisterClass(MVT::f80, &X86::RFP80RegClass);
diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 46a948bc28d..c17aa29f66f 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -464,7 +464,9 @@ let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
   def AVX512_FsFLD0SS : I<0, Pseudo, (outs FR32X:$dst), (ins), "",
                           [(set FR32X:$dst, fp32imm0)]>;
   def AVX512_FsFLD0SD : I<0, Pseudo, (outs FR64X:$dst), (ins), "",
-                          [(set FR64X:$dst, fpimm0)]>;
+                          [(set FR64X:$dst, fp64imm0)]>;
+  def AVX512_FsFLD0F128 : I<0, Pseudo, (outs VR128X:$dst), (ins), "",
+                            [(set VR128X:$dst, fp128imm0)]>;
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td b/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
index 096cc27861c..d75b492594b 100644
--- a/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ b/llvm/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -963,6 +963,10 @@ def fp64imm0 : PatLeaf<(f64 fpimm), [{
   return N->isExactlyValue(+0.0);
 }]>;
 
+def fp128imm0 : PatLeaf<(f128 fpimm), [{
+  return N->isExactlyValue(+0.0);
+}]>;
+
 // EXTRACT_get_vextract128_imm xform function: convert extract_subvector index
 // to VEXTRACTF128/VEXTRACTI128 imm.
 def EXTRACT_get_vextract128_imm : SDNodeXForm<extract_subvector, [{
diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index 809f93ed602..1831986e2ab 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -497,9 +497,11 @@ bool X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
   case X86::AVX512_512_SETALLONES:
   case X86::AVX512_FsFLD0SD:
   case X86::AVX512_FsFLD0SS:
+  case X86::AVX512_FsFLD0F128:
   case X86::AVX_SET0:
   case X86::FsFLD0SD:
   case X86::FsFLD0SS:
+  case X86::FsFLD0F128:
   case X86::KSET0D:
   case X86::KSET0Q:
   case X86::KSET0W:
@@ -4026,6 +4028,7 @@ bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
   case X86::V_SET0:
   case X86::FsFLD0SS:
   case X86::FsFLD0SD:
+  case X86::FsFLD0F128:
     return Expand2AddrUndef(MIB, get(HasAVX ? X86::VXORPSrr : X86::XORPSrr));
   case X86::AVX_SET0: {
     assert(HasAVX && "AVX not supported");
@@ -4039,7 +4042,8 @@ bool X86InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
   }
   case X86::AVX512_128_SET0:
   case X86::AVX512_FsFLD0SS:
-  case X86::AVX512_FsFLD0SD: {
+  case X86::AVX512_FsFLD0SD:
+  case X86::AVX512_FsFLD0F128: {
     bool HasVLX = Subtarget.hasVLX();
     Register SrcReg = MIB->getOperand(0).getReg();
     const TargetRegisterInfo *TRI = &getRegisterInfo();
@@ -5152,6 +5156,8 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
     case X86::V_SET0:
     case X86::V_SETALLONES:
     case X86::AVX512_128_SET0:
+    case X86::FsFLD0F128:
+    case X86::AVX512_FsFLD0F128:
       Alignment = 16;
       break;
     case X86::MMX_SET0:
@@ -5201,7 +5207,9 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
   case X86::FsFLD0SD:
   case X86::AVX512_FsFLD0SD:
   case X86::FsFLD0SS:
-  case X86::AVX512_FsFLD0SS: {
+  case X86::AVX512_FsFLD0SS:
+  case X86::FsFLD0F128:
+  case X86::AVX512_FsFLD0F128: {
     // Folding a V_SET0 or V_SETALLONES as a load, to ease register pressure.
     // Create a constant-pool entry and operands to load from it.
 
@@ -5231,6 +5239,8 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl(
       Ty = Type::getFloatTy(MF.getFunction().getContext());
     else if (Opc == X86::FsFLD0SD || Opc == X86::AVX512_FsFLD0SD)
       Ty = Type::getDoubleTy(MF.getFunction().getContext());
+    else if (Opc == X86::FsFLD0F128 || Opc == X86::AVX512_FsFLD0F128)
+      Ty = Type::getFP128Ty(MF.getFunction().getContext());
     else if (Opc == X86::AVX512_512_SET0 || Opc == X86::AVX512_512_SETALLONES)
       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()),16);
     else if (Opc == X86::AVX2_SETALLONES || Opc == X86::AVX_SET0 ||
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index db1dbf9be5f..5ed4674cef2 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -115,7 +115,9 @@ let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
   def FsFLD0SS : I<0, Pseudo, (outs FR32:$dst), (ins), "",
                    [(set FR32:$dst, fp32imm0)]>, Requires<[HasSSE1, NoAVX512]>;
   def FsFLD0SD : I<0, Pseudo, (outs FR64:$dst), (ins), "",
-                   [(set FR64:$dst, fpimm0)]>, Requires<[HasSSE2, NoAVX512]>;
+                   [(set FR64:$dst, fp64imm0)]>, Requires<[HasSSE2, NoAVX512]>;
+  def FsFLD0F128 : I<0, Pseudo, (outs VR128:$dst), (ins), "",
+                     [(set VR128:$dst, fp128imm0)]>, Requires<[HasSSE1, NoAVX512]>;
 }
 
 //===----------------------------------------------------------------------===//