[Power9] Add __float128 builtins for Rounding Operations

[stefanp]

Summary

Added __float128 support for a number of rounding operations:

trunc
rint
nearbyint
round
floor
ceil

[nemanjai]

LGTM.

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

if (Subtarget.hasP9Vector()) {
           setLoadExtAction(ISD::EXTLOAD, MVT::f128, FPT, Expand);
         setOperationAction(ISD::FMA, MVT::f128, Legal);
         setCondCodeAction(ISD::SETULT, MVT::f128, Expand);
         setCondCodeAction(ISD::SETUGT, MVT::f128, Expand);
         setCondCodeAction(ISD::SETUEQ, MVT::f128, Expand);
         setCondCodeAction(ISD::SETOGE, MVT::f128, Expand);
         setCondCodeAction(ISD::SETOLE, MVT::f128, Expand);
         setCondCodeAction(ISD::SETONE, MVT::f128, Expand);
+
+        setOperationAction(ISD::FTRUNC, MVT::f128, Legal);
+        setOperationAction(ISD::FRINT, MVT::f128, Legal);
+        setOperationAction(ISD::FFLOOR, MVT::f128, Legal);
+        setOperationAction(ISD::FCEIL, MVT::f128, Legal);
+        setOperationAction(ISD::FNEARBYINT, MVT::f128, Legal);
+        setOperationAction(ISD::FROUND, MVT::f128, Legal);
+
         setOperationAction(ISD::SELECT, MVT::f128, Expand);
         setOperationAction(ISD::FP_ROUND, MVT::f64, Legal);
         setOperationAction(ISD::FP_ROUND, MVT::f32, Legal);
         setTruncStoreAction(MVT::f128, MVT::f64, Expand);
         setTruncStoreAction(MVT::f128, MVT::f32, Expand);
         setOperationAction(ISD::BITCAST, MVT::i128, Custom);
         // No implementation for these ops for PowerPC.
         setOperationAction(ISD::FSIN , MVT::f128, Expand);

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

: Z23Form_8<opcode, xo,
                 !strconcat(opc, " $r, $vT, $vB, $rmc"), IIC_VecFP, pattern> {
     let RC = ex;
   }
 
   // Round to Quad-Precision Integer [with Inexact]
   def XSRQPI   : Z23_VT5_R1_VB5_RMC2_EX1<63,  5, 0, "xsrqpi" , []>;
   def XSRQPIX  : Z23_VT5_R1_VB5_RMC2_EX1<63,  5, 1, "xsrqpix", []>;
 
+  // Use current rounding mode
+  def : Pat<(f128 (fnearbyint f128:$vB)), (f128 (XSRQPI 0, $vB, 3))>;
+  // Round to nearest, ties away from zero
+  def : Pat<(f128 (fround f128:$vB)), (f128 (XSRQPI 0, $vB, 0))>;
+  // Round towards Zero
+  def : Pat<(f128 (ftrunc f128:$vB)), (f128 (XSRQPI 1, $vB, 1))>;
+  // Round towards +Inf
+  def : Pat<(f128 (fceil f128:$vB)), (f128 (XSRQPI 1, $vB, 2))>;
+  // Round towards -Inf
+  def : Pat<(f128 (ffloor f128:$vB)), (f128 (XSRQPI 1, $vB, 3))>;
+
+  // Use current rounding mode, [with Inexact]
+  def : Pat<(f128 (frint f128:$vB)), (f128 (XSRQPIX 0, $vB, 3))>;
+
   // Round Quad-Precision to Double-Extended Precision (fp80)
   def XSRQPXP  : Z23_VT5_R1_VB5_RMC2_EX1<63, 37, 0, "xsrqpxp", []>;
 
   //===--------------------------------------------------------------------===//
   // Insert/Extract Instructions
 
   // Insert Exponent DP/QP
   // XT NOTE: XT.dword[1] = 0xUUUU_UUUU_UUUU_UUUU

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

+; RUN: llc -mcpu=pwr9 -mtriple=powerpc64le-unknown-unknown \
+; RUN:   -enable-ppc-quad-precision -verify-machineinstrs < %s | FileCheck %s
+
+
+define void @qp_trunc(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %1 = tail call fp128 @llvm.trunc.f128(fp128 %0)
+  store fp128 %1, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qp_trunc
+; CHECK: xsrqpi 1, {{[0-9]+}}, {{[0-9]+}}, 1
+; CHECK: blr
+}
+declare fp128     @llvm.trunc.f128(fp128 %Val)
+
+define void @qp_rint(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %1 = tail call fp128 @llvm.rint.f128(fp128 %0)
+  store fp128 %1, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qp_rint
+; CHECK: xsrqpix 0, {{[0-9]+}}, {{[0-9]+}}, 3
+; CHECK: blr
+}
+declare fp128     @llvm.rint.f128(fp128 %Val)
+
+define void @qp_nearbyint(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %1 = tail call fp128 @llvm.nearbyint.f128(fp128 %0)
+  store fp128 %1, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qp_nearbyint
+; CHECK: xsrqpi 0, {{[0-9]+}}, {{[0-9]+}}, 3
+; CHECK: blr
+}
+declare fp128     @llvm.nearbyint.f128(fp128 %Val)
+
+define void @qp_round(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %1 = tail call fp128 @llvm.round.f128(fp128 %0)
+  store fp128 %1, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qp_round
+; CHECK: xsrqpi 0, {{[0-9]+}}, {{[0-9]+}}, 0
+; CHECK: blr
+}
+declare fp128     @llvm.round.f128(fp128 %Val)
+
+define void @qp_floor(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %1 = tail call fp128 @llvm.floor.f128(fp128 %0)
+  store fp128 %1, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qp_floor
+; CHECK: xsrqpi 1, {{[0-9]+}}, {{[0-9]+}}, 3
+; CHECK: blr
+}
+declare fp128     @llvm.floor.f128(fp128 %Val)
+
+define void @qp_ceil(fp128* nocapture readonly %a, fp128* nocapture %res) {
+entry:
+  %0 = load fp128, fp128* %a, align 16
+  %1 = tail call fp128 @llvm.ceil.f128(fp128 %0)
+  store fp128 %1, fp128* %res, align 16
+  ret void
+; CHECK-LABEL: qp_ceil
+; CHECK: xsrqpi 1, {{[0-9]+}}, {{[0-9]+}}, 2
+; CHECK: blr
+}
+declare fp128     @llvm.ceil.f128(fp128 %Val)
+