[Power9]Legalize and emit code for quad-precision add/div/mul/sub

[lei]

Summary

This is the first patch of several to come to add support for quad-precision floating point operations.

Legalize and emit code for quad-precision floating point operations:

• xsaddqp
• xssubqp
• xsdivqp
• xsmulqp

[nemanjai]

LGTM. Please add the requested test case on the commit - no need for an additional review.

lib/Target/PowerPC/PPCInstrVSX.td

2898 ↗

(On Diff #138492)

Can you please add an additional test case for this. We want to make sure no future patches break it. Something along these lines should suffice:

define void @test(i8* nocapture readonly %PtrC, fp128* nocapture %PtrF) {
entry:
  %add.ptr = getelementptr inbounds i8, i8* %PtrC, i64 4
  %0 = bitcast i8* %add.ptr to fp128*
  %1 = load fp128, fp128* %0, align 16
  %2 = bitcast fp128* %PtrF to i8*
  %add.ptr1 = getelementptr inbounds i8, i8* %2, i64 8
  %3 = bitcast i8* %add.ptr1 to fp128*
  store fp128 %1, fp128* %3, align 16
  ret void
}

We need to make sure we emit LXVX and STXVX here rather than their D-Form analogs.

llvm/trunk/lib/Target/PowerPC/PPCISelLowering.cpp

if (Subtarget.hasP9Vector()) {
       setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
 
       // 128 bit shifts can be accomplished via 3 instructions for SHL and
       // SRL, but not for SRA because of the instructions available:
       // VS{RL} and VS{RL}O.
       setOperationAction(ISD::SHL, MVT::v1i128, Legal);
       setOperationAction(ISD::SRL, MVT::v1i128, Legal);
       setOperationAction(ISD::SRA, MVT::v1i128, Expand);
+
+      addRegisterClass(MVT::f128, &PPC::VRRCRegClass);
+      setOperationAction(ISD::FADD, MVT::f128, Legal);
+      setOperationAction(ISD::FSUB, MVT::f128, Legal);
+      setOperationAction(ISD::FDIV, MVT::f128, Legal);
+      setOperationAction(ISD::FMUL, MVT::f128, Legal);
+
     }
 
     if (Subtarget.hasP9Altivec()) {
       setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
       setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
     }
   }
 

llvm/trunk/lib/Target/PowerPC/PPCInstrVSX.td

let AddedComplexity = 400, Predicates = [HasP9Vector] in {
   def XSABSQP   : X_VT5_XO5_VB5<63,  0, 804, "xsabsqp" , []>;
   def XSNABSQP  : X_VT5_XO5_VB5<63,  8, 804, "xsnabsqp", []>;
   def XSNEGQP   : X_VT5_XO5_VB5<63, 16, 804, "xsnegqp" , []>;
 
   //===--------------------------------------------------------------------===//
   // Quad-Precision Scalar Floating-Point Arithmetic Instructions:
 
   // Add/Divide/Multiply/Subtract
-  def XSADDQP   : X_VT5_VA5_VB5   <63,   4, "xsaddqp" , []>;
+  let isCommutable = 1 in {
+  def XSADDQP   : X_VT5_VA5_VB5   <63,   4, "xsaddqp",
+                                   [(set f128:$vT, (fadd f128:$vA, f128:$vB))]>;
   def XSADDQPO  : X_VT5_VA5_VB5_Ro<63,   4, "xsaddqpo", []>;
-  def XSDIVQP   : X_VT5_VA5_VB5   <63, 548, "xsdivqp" , []>;
-  def XSDIVQPO  : X_VT5_VA5_VB5_Ro<63, 548, "xsdivqpo", []>;
-  def XSMULQP   : X_VT5_VA5_VB5   <63,  36, "xsmulqp" , []>;
+  def XSMULQP   : X_VT5_VA5_VB5   <63,  36, "xsmulqp",
+                                   [(set f128:$vT, (fmul f128:$vA, f128:$vB))]>;
   def XSMULQPO  : X_VT5_VA5_VB5_Ro<63,  36, "xsmulqpo", []>;
-  def XSSUBQP   : X_VT5_VA5_VB5   <63, 516, "xssubqp" , []>;
+  }
+
+  def XSSUBQP   : X_VT5_VA5_VB5   <63, 516, "xssubqp" ,
+                                   [(set f128:$vT, (fsub f128:$vA, f128:$vB))]>;
   def XSSUBQPO  : X_VT5_VA5_VB5_Ro<63, 516, "xssubqpo", []>;
+  def XSDIVQP   : X_VT5_VA5_VB5   <63, 548, "xsdivqp",
+                                   [(set f128:$vT, (fdiv f128:$vA, f128:$vB))]>;
+  def XSDIVQPO  : X_VT5_VA5_VB5_Ro<63, 548, "xsdivqpo", []>;
 
   // Square-Root
   def XSSQRTQP  : X_VT5_XO5_VB5   <63, 27, 804, "xssqrtqp" , []>;
   def XSSQRTQPO : X_VT5_XO5_VB5_Ro<63, 27, 804, "xssqrtqpo", []>;
 
   // (Negative) Multiply-{Add/Subtract}
   def XSMADDQP  : X_VT5_VA5_VB5   <63, 388, "xsmaddqp"  , []>;
   def XSMADDQPO : X_VT5_VA5_VB5_Ro<63, 388, "xsmaddqpo" , []>;
def : Pat<(v4f32 (insertelt v4f32:$A, f32:$B, 3)),
             (v4f32 (XXINSERTW v4f32:$A, AlignValues.F32_TO_BE_WORD1, 12))>;
   } // IsLittleEndian, HasP9Vector
 
   // D-Form Load/Store
   def : Pat<(v4i32 (quadwOffsetLoad iqaddr:$src)), (LXV memrix16:$src)>;
   def : Pat<(v4f32 (quadwOffsetLoad iqaddr:$src)), (LXV memrix16:$src)>;
   def : Pat<(v2i64 (quadwOffsetLoad iqaddr:$src)), (LXV memrix16:$src)>;
   def : Pat<(v2f64 (quadwOffsetLoad iqaddr:$src)), (LXV memrix16:$src)>;
+  def : Pat<(f128  (quadwOffsetLoad iqaddr:$src)),
+            (COPY_TO_REGCLASS (LXV memrix16:$src), VRRC)>;
   def : Pat<(v4i32 (int_ppc_vsx_lxvw4x iqaddr:$src)), (LXV memrix16:$src)>;
   def : Pat<(v2f64 (int_ppc_vsx_lxvd2x iqaddr:$src)), (LXV memrix16:$src)>;
 
   def : Pat<(quadwOffsetStore v4f32:$rS, iqaddr:$dst), (STXV $rS, memrix16:$dst)>;
   def : Pat<(quadwOffsetStore v4i32:$rS, iqaddr:$dst), (STXV $rS, memrix16:$dst)>;
   def : Pat<(quadwOffsetStore v2f64:$rS, iqaddr:$dst), (STXV $rS, memrix16:$dst)>;
+  def : Pat<(quadwOffsetStore  f128:$rS, iqaddr:$dst),
+            (STXV (COPY_TO_REGCLASS $rS, VSRC), memrix16:$dst)>;
   def : Pat<(quadwOffsetStore v2i64:$rS, iqaddr:$dst), (STXV $rS, memrix16:$dst)>;
   def : Pat<(int_ppc_vsx_stxvw4x v4i32:$rS, iqaddr:$dst),
             (STXV $rS, memrix16:$dst)>;
   def : Pat<(int_ppc_vsx_stxvd2x v2f64:$rS, iqaddr:$dst),
             (STXV $rS, memrix16:$dst)>;
 
 
   def : Pat<(v2f64 (nonQuadwOffsetLoad xoaddr:$src)), (LXVX xoaddr:$src)>;
   def : Pat<(v2i64 (nonQuadwOffsetLoad xoaddr:$src)), (LXVX xoaddr:$src)>;
   def : Pat<(v4f32 (nonQuadwOffsetLoad xoaddr:$src)), (LXVX xoaddr:$src)>;
   def : Pat<(v4i32 (nonQuadwOffsetLoad xoaddr:$src)), (LXVX xoaddr:$src)>;
   def : Pat<(v4i32 (int_ppc_vsx_lxvw4x xoaddr:$src)), (LXVX xoaddr:$src)>;
   def : Pat<(v2f64 (int_ppc_vsx_lxvd2x xoaddr:$src)), (LXVX xoaddr:$src)>;
+  def : Pat<(f128  (nonQuadwOffsetLoad xoaddr:$src)),
+            (COPY_TO_REGCLASS (LXVX xoaddr:$src), VRRC)>;
+  def : Pat<(nonQuadwOffsetStore f128:$rS, xoaddr:$dst),
+            (STXVX (COPY_TO_REGCLASS $rS, VSRC), xoaddr:$dst)>;
   def : Pat<(nonQuadwOffsetStore v2f64:$rS, xoaddr:$dst),
             (STXVX $rS, xoaddr:$dst)>;
   def : Pat<(nonQuadwOffsetStore v2i64:$rS, xoaddr:$dst),
             (STXVX $rS, xoaddr:$dst)>;
   def : Pat<(nonQuadwOffsetStore v4f32:$rS, xoaddr:$dst),
             (STXVX $rS, xoaddr:$dst)>;
   def : Pat<(nonQuadwOffsetStore v4i32:$rS, xoaddr:$dst),
             (STXVX $rS, xoaddr:$dst)>;

llvm/trunk/test/CodeGen/PowerPC/f128-arith.ll

+; RUN: llc -mcpu=pwr9 -mtriple=powerpc64le-unknown-unknown < %s | FileCheck %s
+
+; Function Attrs: norecurse nounwind
+define void @qpAdd(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %add = fadd fp128 %0, %0
+  store fp128 %add, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qpAdd
+; CHECK-NOT bl __addtf3
+; CHECK xsaddqp
+; CHECK stxv
+; CHECK blr
+}
+
+; Function Attrs: norecurse nounwind
+define void @qpSub(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %sub = fsub fp128 %0, %0
+  store fp128 %sub, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qpSub
+; CHECK-NOT bl __subtf3
+; CHECK xssubqp
+; CHECK stxv
+; CHECK blr
+}
+
+; Function Attrs: norecurse nounwind
+define void @qpMul(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %mul = fmul fp128 %0, %0
+  store fp128 %mul, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qpMul
+; CHECK-NOT bl __multf3
+; CHECK xsmulqp
+; CHECK stxv
+; CHECK blr
+}
+
+; Function Attrs: norecurse nounwind
+define void @qpDiv(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %div = fdiv fp128 %0, %0
+  store fp128 %div, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qpDiv
+; CHECK-NOT bl __divtf3
+; CHECK xsdivqp
+; CHECK stxv
+; CHECK blr
+}
+
+define void @testLdNSt(i8* nocapture readonly %PtrC, fp128* nocapture %PtrF) {
+entry:
+  %add.ptr = getelementptr inbounds i8, i8* %PtrC, i64 4
+  %0 = bitcast i8* %add.ptr to fp128*
+  %1 = load fp128, fp128* %0, align 16
+  %2 = bitcast fp128* %PtrF to i8*
+  %add.ptr1 = getelementptr inbounds i8, i8* %2, i64 8
+  %3 = bitcast i8* %add.ptr1 to fp128*
+  store fp128 %1, fp128* %3, align 16
+  ret void
+; CHECK-LABEL: testLdNSt
+; CHECK lxvx
+; CHECK stxvx
+; CHECK-NEXT blr
+}