[SystemZ] Implement ISD::BITCAST for fp128 -> i128.

[jonpa]

Summary

The type legalizer has by default no method of doing this bitcast other than storing and reloading the value from stack.

This patch implements a custom lowering of this operation using extractions of subregs (z13 and earlier using FP128 register pairs), or of vector elements (with vector enhancements-1 using VR128 FP registers).

(Unfortunately there is no "low" subreg for a VR128 reg, so that case had to get its own separate handling.)

[jonpa]

llvm/lib/Target/SystemZ/SystemZISelLowering.cpp

296

hmm.. should the check for soft-float be done here instead?

[uweigand]

LGTM, thanks!

llvm/lib/Target/SystemZ/SystemZISelLowering.cpp

296

No, I think it's better to have it in the implementation -- just in case we need to handle some other cases in BITCAST later.

llvm/lib/Target/SystemZ/SystemZISelLowering.cpp

SystemZTargetLowering::SystemZTargetLowering(const TargetMachine &TM,
   // Expand 128 bit shifts without using a libcall.
   setOperationAction(ISD::SRL_PARTS, MVT::i64, Expand);
   setOperationAction(ISD::SHL_PARTS, MVT::i64, Expand);
   setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
   setLibcallName(RTLIB::SRL_I128, nullptr);
   setLibcallName(RTLIB::SHL_I128, nullptr);
   setLibcallName(RTLIB::SRA_I128, nullptr);
 
+  // Handle bitcast from fp128 to i128.
+  setOperationAction(ISD::BITCAST, MVT::i128, Custom);
+
   // We have native instructions for i8, i16 and i32 extensions, but not i1.
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
   for (MVT VT : MVT::integer_valuetypes()) {
     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
     setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
     setLoadExtAction(ISD::EXTLOAD,  VT, MVT::i1, Promote);
   }
 
case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
     SDValue Success = emitSETCC(DAG, DL, Res.getValue(1),
                                 SystemZ::CCMASK_CS, SystemZ::CCMASK_CS_EQ);
     Success = DAG.getZExtOrTrunc(Success, DL, N->getValueType(1));
     Results.push_back(lowerGR128ToI128(DAG, Res));
     Results.push_back(Success);
     Results.push_back(Res.getValue(2));
     break;
   }
+  case ISD::BITCAST: {
+    SDValue Src = N->getOperand(0);
+    if (N->getValueType(0) == MVT::i128 && Src.getValueType() == MVT::f128 &&
+        !useSoftFloat()) {
+      SDLoc DL(N);
+      SDValue Lo, Hi;
+      if (getRepRegClassFor(MVT::f128) == &SystemZ::VR128BitRegClass) {
+        SDValue VecBC = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Src);
+        Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64, VecBC,
+                         DAG.getConstant(1, DL, MVT::i32));
+        Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64, VecBC,
+                         DAG.getConstant(0, DL, MVT::i32));
+      } else {
+        assert(getRepRegClassFor(MVT::f128) == &SystemZ::FP128BitRegClass &&
+               "Unrecognized register class for f128.");
+        SDValue LoFP = DAG.getTargetExtractSubreg(SystemZ::subreg_l64,
+                                                  DL, MVT::f64, Src);
+        SDValue HiFP = DAG.getTargetExtractSubreg(SystemZ::subreg_h64,
+                                                  DL, MVT::f64, Src);
+        Lo = DAG.getNode(ISD::BITCAST, DL, MVT::i64, LoFP);
+        Hi = DAG.getNode(ISD::BITCAST, DL, MVT::i64, HiFP);
+      }
+      Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128, Lo, Hi));
+    }
+    break;
+  }
   default:
     llvm_unreachable("Unexpected node to lower");
   }
 }
 
 void
 SystemZTargetLowering::ReplaceNodeResults(SDNode *N,
                                           SmallVectorImpl<SDValue> &Results,

llvm/test/CodeGen/SystemZ/fp-conv-19.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+
+; Test f128 to i128 bitcasts.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 \
+; RUN:   | FileCheck -check-prefix=Z10 %s
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 \
+; RUN:   | FileCheck -check-prefix=Z14 %s
+
+define i64 @extract_float_hi(fp128* %0, fp128* %1) {
+; Z10-LABEL: extract_float_hi:
+; Z10:       # %bb.0: # %entry
+; Z10-NEXT:    ld %f0, 0(%r2)
+; Z10-NEXT:    ld %f2, 8(%r2)
+; Z10-NEXT:    ld %f1, 0(%r3)
+; Z10-NEXT:    ld %f3, 8(%r3)
+; Z10-NEXT:    axbr %f1, %f0
+; Z10-NEXT:    lgdr %r2, %f1
+; Z10-NEXT:    br %r14
+;
+; Z14-LABEL: extract_float_hi:
+; Z14:       # %bb.0: # %entry
+; Z14-NEXT:    vl %v0, 0(%r2), 3
+; Z14-NEXT:    vl %v1, 0(%r3), 3
+; Z14-NEXT:    wfaxb %v0, %v0, %v1
+; Z14-NEXT:    vlgvg %r2, %v0, 0
+; Z14-NEXT:    br %r14
+entry:
+  %x = load fp128, fp128* %0
+  %y = load fp128, fp128* %1
+  %add = fadd fp128 %x, %y
+  %2 = bitcast fp128 %add to i128
+  %u.sroa.0.0.extract.shift = lshr i128 %2, 64
+  %u.sroa.0.0.extract.trunc = trunc i128 %u.sroa.0.0.extract.shift to i64
+  ret i64 %u.sroa.0.0.extract.trunc
+}
+
+define i64 @extract_float_lo(fp128* %0, fp128* %1) {
+; Z10-LABEL: extract_float_lo:
+; Z10:       # %bb.0: # %entry
+; Z10-NEXT:    ld %f0, 0(%r2)
+; Z10-NEXT:    ld %f2, 8(%r2)
+; Z10-NEXT:    ld %f1, 0(%r3)
+; Z10-NEXT:    ld %f3, 8(%r3)
+; Z10-NEXT:    axbr %f1, %f0
+; Z10-NEXT:    lgdr %r2, %f3
+; Z10-NEXT:    br %r14
+;
+; Z14-LABEL: extract_float_lo:
+; Z14:       # %bb.0: # %entry
+; Z14-NEXT:    vl %v0, 0(%r2), 3
+; Z14-NEXT:    vl %v1, 0(%r3), 3
+; Z14-NEXT:    wfaxb %v0, %v0, %v1
+; Z14-NEXT:    vlgvg %r2, %v0, 1
+; Z14-NEXT:    br %r14
+entry:
+  %x = load fp128, fp128* %0
+  %y = load fp128, fp128* %1
+  %add = fadd fp128 %x, %y
+  %2 = bitcast fp128 %add to i128
+  %u.sroa.0.0.extract.trunc = trunc i128 %2 to i64
+  ret i64 %u.sroa.0.0.extract.trunc
+}
+
+define i128 @bitcast_128(fp128* %0, fp128* %1) {
+; Z10-LABEL: bitcast_128:
+; Z10:       # %bb.0: # %entry
+; Z10-NEXT:    ld %f0, 0(%r3)
+; Z10-NEXT:    ld %f2, 8(%r3)
+; Z10-NEXT:    ld %f1, 0(%r4)
+; Z10-NEXT:    ld %f3, 8(%r4)
+; Z10-NEXT:    axbr %f1, %f0
+; Z10-NEXT:    lgdr %r0, %f3
+; Z10-NEXT:    lgdr %r1, %f1
+; Z10-NEXT:    oill %r1, 1
+; Z10-NEXT:    oill %r0, 3
+; Z10-NEXT:    stg %r0, 8(%r2)
+; Z10-NEXT:    stg %r1, 0(%r2)
+; Z10-NEXT:    br %r14
+;
+; Z14-LABEL: bitcast_128:
+; Z14:       # %bb.0: # %entry
+; Z14-NEXT:    vl %v0, 0(%r3), 3
+; Z14-NEXT:    vl %v1, 0(%r4), 3
+; Z14-NEXT:    wfaxb %v0, %v0, %v1
+; Z14-NEXT:    vlgvg %r0, %v0, 1
+; Z14-NEXT:    vlgvg %r1, %v0, 0
+; Z14-NEXT:    oill %r1, 1
+; Z14-NEXT:    oill %r0, 3
+; Z14-NEXT:    stg %r0, 8(%r2)
+; Z14-NEXT:    stg %r1, 0(%r2)
+; Z14-NEXT:    br %r14
+entry:
+  %x = load fp128, fp128* %0
+  %y = load fp128, fp128* %1
+  %add = fadd fp128 %x, %y
+  %i = bitcast fp128 %add to i128
+  %hibit = shl i128 1, 64
+  %i2 = or i128 %i, %hibit
+  %i3 = or i128 %i2, 3
+  ret i128 %i3
+}