[Power9] Add __float128 builtins for Round To Odd

GCC has builtins for these round to odd instructions:

__float128 __builtin_sqrtf128_round_to_odd (__float128)
__float128 __builtin_{add,sub,mul,div}f128_round_to_odd (__float128, __float128)
__float128 __builtin_fmaf128_round_to_odd (__float128, __float128, __float128)

Differential Revision: https://reviews.llvm.org/D47550

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@336578 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/IR/Intrinsics.h

@@ -97,7 +97,7 @@ namespace Intrinsic {
   /// intrinsic. This is returned by getIntrinsicInfoTableEntries.
   struct IITDescriptor {
     enum IITDescriptorKind {
-      Void, VarArg, MMX, Token, Metadata, Half, Float, Double,
+      Void, VarArg, MMX, Token, Metadata, Half, Float, Double, Quad,
       Integer, Vector, Pointer, Struct,
       Argument, ExtendArgument, TruncArgument, HalfVecArgument,
       SameVecWidthArgument, PtrToArgument, PtrToElt, VecOfAnyPtrsToElt

include/llvm/IR/IntrinsicsPowerPC.td

@@ -61,6 +61,26 @@ let TargetPrefix = "ppc" in {  // All intrinsics start with "llvm.ppc.".
   def int_ppc_bpermd : GCCBuiltin<"__builtin_bpermd">,
                        Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty],
                                  [IntrNoMem]>;
+
+  def int_ppc_sqrtf128_round_to_odd
+      : GCCBuiltin<"__builtin_sqrtf128_round_to_odd">,
+        Intrinsic <[llvm_f128_ty], [llvm_f128_ty], [IntrNoMem]>;
+  def int_ppc_addf128_round_to_odd
+      : GCCBuiltin<"__builtin_addf128_round_to_odd">,
+        Intrinsic <[llvm_f128_ty], [llvm_f128_ty,llvm_f128_ty], [IntrNoMem]>;
+  def int_ppc_subf128_round_to_odd
+      : GCCBuiltin<"__builtin_subf128_round_to_odd">,
+        Intrinsic <[llvm_f128_ty], [llvm_f128_ty,llvm_f128_ty], [IntrNoMem]>;
+  def int_ppc_mulf128_round_to_odd
+      : GCCBuiltin<"__builtin_mulf128_round_to_odd">,
+        Intrinsic <[llvm_f128_ty], [llvm_f128_ty,llvm_f128_ty], [IntrNoMem]>;
+  def int_ppc_divf128_round_to_odd
+      : GCCBuiltin<"__builtin_divf128_round_to_odd">,
+        Intrinsic <[llvm_f128_ty], [llvm_f128_ty,llvm_f128_ty], [IntrNoMem]>;
+  def int_ppc_fmaf128_round_to_odd
+      : GCCBuiltin<"__builtin_fmaf128_round_to_odd">,
+        Intrinsic <[llvm_f128_ty], [llvm_f128_ty,llvm_f128_ty,llvm_f128_ty], [IntrNoMem]>;
+
 }
 
 

lib/IR/Function.cpp

@@ -673,7 +673,8 @@ enum IIT_Info {
   IIT_V1024 = 37,
   IIT_STRUCT6 = 38,
   IIT_STRUCT7 = 39,
-  IIT_STRUCT8 = 40
+  IIT_STRUCT8 = 40,
+  IIT_F128 = 41
 };
 
 static void DecodeIITType(unsigned &NextElt, ArrayRef<unsigned char> Infos,
@@ -708,6 +709,9 @@ static void DecodeIITType(unsigned &NextElt, ArrayRef<unsigned char> Infos,
   case IIT_F64:
     OutputTable.push_back(IITDescriptor::get(IITDescriptor::Double, 0));
     return;
+  case IIT_F128:
+    OutputTable.push_back(IITDescriptor::get(IITDescriptor::Quad, 0));
+    return;
   case IIT_I1:
     OutputTable.push_back(IITDescriptor::get(IITDescriptor::Integer, 1));
     return;
@@ -892,6 +896,7 @@ static Type *DecodeFixedType(ArrayRef<Intrinsic::IITDescriptor> &Infos,
   case IITDescriptor::Half: return Type::getHalfTy(Context);
   case IITDescriptor::Float: return Type::getFloatTy(Context);
   case IITDescriptor::Double: return Type::getDoubleTy(Context);
+  case IITDescriptor::Quad: return Type::getFP128Ty(Context);
 
   case IITDescriptor::Integer:
     return IntegerType::get(Context, D.Integer_Width);
@@ -1034,6 +1039,7 @@ bool Intrinsic::matchIntrinsicType(Type *Ty, ArrayRef<Intrinsic::IITDescriptor>
     case IITDescriptor::Half: return !Ty->isHalfTy();
     case IITDescriptor::Float: return !Ty->isFloatTy();
     case IITDescriptor::Double: return !Ty->isDoubleTy();
+    case IITDescriptor::Quad: return !Ty->isFP128Ty();
     case IITDescriptor::Integer: return !Ty->isIntegerTy(D.Integer_Width);
     case IITDescriptor::Vector: {
       VectorType *VT = dyn_cast<VectorType>(Ty);

lib/Target/PowerPC/PPCInstrVSX.td

@@ -2451,30 +2451,49 @@ let AddedComplexity = 400, Predicates = [HasP9Vector] in {
   let isCommutable = 1 in {
   def XSADDQP   : X_VT5_VA5_VB5   <63,   4, "xsaddqp",
                                    [(set f128:$vT, (fadd f128:$vA, f128:$vB))]>;
-  def XSADDQPO  : X_VT5_VA5_VB5_Ro<63,   4, "xsaddqpo", []>;
+  def XSADDQPO : X_VT5_VA5_VB5_Ro<63, 4, "xsaddqpo",
+                                  [(set f128:$vT,
+                                  (int_ppc_addf128_round_to_odd
+                                  f128:$vA, f128:$vB))]>;
   def XSMULQP   : X_VT5_VA5_VB5   <63,  36, "xsmulqp",
                                    [(set f128:$vT, (fmul f128:$vA, f128:$vB))]>;
-  def XSMULQPO  : X_VT5_VA5_VB5_Ro<63,  36, "xsmulqpo", []>;
+  def XSMULQPO : X_VT5_VA5_VB5_Ro<63, 36, "xsmulqpo",
+                                  [(set f128:$vT,
+                                  (int_ppc_mulf128_round_to_odd
+                                  f128:$vA, f128:$vB))]>;
   }
 
   def XSSUBQP   : X_VT5_VA5_VB5   <63, 516, "xssubqp" ,
                                    [(set f128:$vT, (fsub f128:$vA, f128:$vB))]>;
-  def XSSUBQPO  : X_VT5_VA5_VB5_Ro<63, 516, "xssubqpo", []>;
+  def XSSUBQPO : X_VT5_VA5_VB5_Ro<63, 516, "xssubqpo",
+                                  [(set f128:$vT,
+                                  (int_ppc_subf128_round_to_odd
+                                  f128:$vA, f128:$vB))]>;
   def XSDIVQP   : X_VT5_VA5_VB5   <63, 548, "xsdivqp",
                                    [(set f128:$vT, (fdiv f128:$vA, f128:$vB))]>;
-  def XSDIVQPO  : X_VT5_VA5_VB5_Ro<63, 548, "xsdivqpo", []>;
+  def XSDIVQPO : X_VT5_VA5_VB5_Ro<63, 548, "xsdivqpo",
+                                  [(set f128:$vT,
+                                  (int_ppc_divf128_round_to_odd
+                                  f128:$vA, f128:$vB))]>;
 
   // Square-Root
   def XSSQRTQP  : X_VT5_XO5_VB5   <63, 27, 804, "xssqrtqp",
                                    [(set f128:$vT, (fsqrt f128:$vB))]>;
-  def XSSQRTQPO : X_VT5_XO5_VB5_Ro<63, 27, 804, "xssqrtqpo", []>;
+  def XSSQRTQPO : X_VT5_XO5_VB5_Ro<63, 27, 804, "xssqrtqpo",
+                                  [(set f128:$vT,
+                                  (int_ppc_sqrtf128_round_to_odd f128:$vB))]>;
 
   // (Negative) Multiply-{Add/Subtract}
   def XSMADDQP : X_VT5_VA5_VB5_FMA <63, 388, "xsmaddqp",
                                     [(set f128:$vT,
                                           (fma f128:$vA, f128:$vB,
                                                f128:$vTi))]>;
-  def XSMADDQPO : X_VT5_VA5_VB5_FMA_Ro<63, 388, "xsmaddqpo" , []>;
+
+  def XSMADDQPO : X_VT5_VA5_VB5_FMA_Ro<63, 388, "xsmaddqpo",
+                                      [(set f128:$vT,
+                                      (int_ppc_fmaf128_round_to_odd
+                                      f128:$vA,f128:$vB,f128:$vTi))]>;
+
   def XSMSUBQP  : X_VT5_VA5_VB5_FMA   <63, 420, "xsmsubqp"  ,
                                        [(set f128:$vT,
                                              (fma f128:$vA, f128:$vB,

test/CodeGen/PowerPC/builtins-ppc-p9-f128.ll

@@ -0,0 +1,82 @@
+; RUN: llc -verify-machineinstrs -mcpu=pwr9 -enable-ppc-quad-precision \
+; RUN:   -mtriple=powerpc64le-unknown-unknown < %s | FileCheck %s
+
+@A = common global fp128 0xL00000000000000000000000000000000, align 16
+@B = common global fp128 0xL00000000000000000000000000000000, align 16
+@C = common global fp128 0xL00000000000000000000000000000000, align 16
+
+define fp128 @testSqrtOdd() {
+entry:
+  %0 = load fp128, fp128* @A, align 16
+  %1 = call fp128 @llvm.ppc.sqrtf128.round.to.odd(fp128 %0)
+  ret fp128 %1
+; CHECK-LABEL: testSqrtOdd
+; CHECK: xssqrtqpo
+}
+
+declare fp128 @llvm.ppc.sqrtf128.round.to.odd(fp128)
+
+define fp128 @testFMAOdd() {
+entry:
+  %0 = load fp128, fp128* @A, align 16
+  %1 = load fp128, fp128* @B, align 16
+  %2 = load fp128, fp128* @C, align 16
+  %3 = call fp128 @llvm.ppc.fmaf128.round.to.odd(fp128 %0, fp128 %1, fp128 %2)
+  ret fp128 %3
+; CHECK-LABEL: testFMAOdd
+; CHECK: xsmaddqpo
+}
+
+declare fp128 @llvm.ppc.fmaf128.round.to.odd(fp128, fp128, fp128)
+
+define fp128 @testAddOdd() {
+entry:
+  %0 = load fp128, fp128* @A, align 16
+  %1 = load fp128, fp128* @B, align 16
+  %2 = call fp128 @llvm.ppc.addf128.round.to.odd(fp128 %0, fp128 %1)
+  ret fp128 %2
+; CHECK-LABEL: testAddOdd
+; CHECK: xsaddqpo
+}
+
+declare fp128 @llvm.ppc.addf128.round.to.odd(fp128, fp128)
+
+define fp128 @testSubOdd() {
+entry:
+  %0 = load fp128, fp128* @A, align 16
+  %1 = load fp128, fp128* @B, align 16
+  %2 = call fp128 @llvm.ppc.subf128.round.to.odd(fp128 %0, fp128 %1)
+  ret fp128 %2
+; CHECK-LABEL: testSubOdd
+; CHECK: xssubqpo
+}
+
+; Function Attrs: nounwind readnone
+declare fp128 @llvm.ppc.subf128.round.to.odd(fp128, fp128)
+
+; Function Attrs: noinline nounwind optnone
+define fp128 @testMulOdd() {
+entry:
+  %0 = load fp128, fp128* @A, align 16
+  %1 = load fp128, fp128* @B, align 16
+  %2 = call fp128 @llvm.ppc.mulf128.round.to.odd(fp128 %0, fp128 %1)
+  ret fp128 %2
+; CHECK-LABEL: testMulOdd
+; CHECK: xsmulqpo
+}
+
+; Function Attrs: nounwind readnone
+declare fp128 @llvm.ppc.mulf128.round.to.odd(fp128, fp128)
+
+define fp128 @testDivOdd() {
+entry:
+  %0 = load fp128, fp128* @A, align 16
+  %1 = load fp128, fp128* @B, align 16
+  %2 = call fp128 @llvm.ppc.divf128.round.to.odd(fp128 %0, fp128 %1)
+  ret fp128 %2
+; CHECK-LABEL: testDivOdd
+; CHECK: xsdivqpo
+}
+
+declare fp128 @llvm.ppc.divf128.round.to.odd(fp128, fp128)
+

utils/TableGen/IntrinsicEmitter.cpp

@@ -219,7 +219,8 @@ enum IIT_Info {
   IIT_V1024 = 37,
   IIT_STRUCT6 = 38,
   IIT_STRUCT7 = 39,
-  IIT_STRUCT8 = 40
+  IIT_STRUCT8 = 40,
+  IIT_F128 = 41
 };
 
 static void EncodeFixedValueType(MVT::SimpleValueType VT,
@@ -242,6 +243,7 @@ static void EncodeFixedValueType(MVT::SimpleValueType VT,
   case MVT::f16: return Sig.push_back(IIT_F16);
   case MVT::f32: return Sig.push_back(IIT_F32);
   case MVT::f64: return Sig.push_back(IIT_F64);
+  case MVT::f128: return Sig.push_back(IIT_F128);
   case MVT::token: return Sig.push_back(IIT_TOKEN);
   case MVT::Metadata: return Sig.push_back(IIT_METADATA);
   case MVT::x86mmx: return Sig.push_back(IIT_MMX);