[SDAG] Handle non-integer preferred memset types for non-constant values

The existing code in getMemsetValue only handled integer-preferred types when
the fill value was not a constant. Make this more robust in two ways:

  1. If the preferred type is a floating-point value, do the mul-splat trick on
     the corresponding integer type and then bitcast.
  2. If the preferred type is a vector, do the mul-splat trick on one vector
     element, and then build a vector out of them.

Fixes PR22754 (although, we should also turn off use of vector types at -O0).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@233749 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -3791,12 +3791,27 @@ static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG,
     return DAG.getConstantFP(APFloat(DAG.EVTToAPFloatSemantics(VT), Val), VT);
   }
 
-  Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value);
+  assert(Value.getValueType() == MVT::i8 && "memset with non-byte fill value?");
+  EVT IntVT = VT.getScalarType();
+  if (!IntVT.isInteger())
+    IntVT = EVT::getIntegerVT(*DAG.getContext(), IntVT.getSizeInBits());
+
+  Value = DAG.getNode(ISD::ZERO_EXTEND, dl, IntVT, Value);
   if (NumBits > 8) {
     // Use a multiplication with 0x010101... to extend the input to the
     // required length.
     APInt Magic = APInt::getSplat(NumBits, APInt(8, 0x01));
-    Value = DAG.getNode(ISD::MUL, dl, VT, Value, DAG.getConstant(Magic, VT));
+    Value = DAG.getNode(ISD::MUL, dl, IntVT, Value,
+                        DAG.getConstant(Magic, IntVT));
+  }
+
+  if (VT != Value.getValueType() && !VT.isInteger())
+    Value = DAG.getNode(ISD::BITCAST, dl, VT.getScalarType(), Value);
+  if (VT != Value.getValueType()) {
+    assert(VT.getVectorElementType() == Value.getValueType() &&
+           "value type should be one vector element here");
+    SmallVector<SDValue, 8> BVOps(VT.getVectorNumElements(), Value);
+    Value = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, BVOps);
   }
 
   return Value;

test/CodeGen/PowerPC/memset-nc-le.ll

@@ -0,0 +1,24 @@
+; RUN: llc < %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-n32:64"
+target triple = "powerpc64le"
+
+; Function Attrs: nounwind
+define void @test_vsx() unnamed_addr #0 align 2 {
+entry:
+  %0 = load i32, i32* undef, align 4
+  %1 = trunc i32 %0 to i8
+  call void @llvm.memset.p0i8.i64(i8* null, i8 %1, i64 32, i32 1, i1 false)
+  ret void
+
+; CHECK-LABEL: @test_vsx
+; CHECK: stxvd2x
+; CHECK: stxvd2x
+; CHECK: blr
+}
+
+; Function Attrs: nounwind
+declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) #1
+
+attributes #0 = { nounwind "target-cpu"="pwr8" }
+attributes #1 = { nounwind }
+

test/CodeGen/PowerPC/memset-nc.ll

@@ -0,0 +1,39 @@
+; RUN: llc < %s | FileCheck %s
+target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f128:128:128-v128:128:128-n32:64"
+target triple = "powerpc64-bgq-linux"
+
+; Function Attrs: nounwind
+define void @test_qpx() unnamed_addr #0 align 2 {
+entry:
+  %0 = load i32, i32* undef, align 4
+  %1 = trunc i32 %0 to i8
+  call void @llvm.memset.p0i8.i64(i8* null, i8 %1, i64 64, i32 32, i1 false)
+  ret void
+
+; CHECK-LABEL: @test_qpx
+; CHECK: qvstfdx
+; CHECK: qvstfdx
+; CHECK: blr
+}
+
+; Function Attrs: nounwind
+declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) #1
+
+; Function Attrs: nounwind
+define void @test_vsx() unnamed_addr #2 align 2 {
+entry:
+  %0 = load i32, i32* undef, align 4
+  %1 = trunc i32 %0 to i8
+  call void @llvm.memset.p0i8.i64(i8* null, i8 %1, i64 32, i32 1, i1 false)
+  ret void
+
+; CHECK-LABEL: @test_vsx
+; CHECK: stxvw4x
+; CHECK: stxvw4x
+; CHECK: blr
+}
+
+attributes #0 = { nounwind "target-cpu"="a2q" }
+attributes #1 = { nounwind }
+attributes #2 = { nounwind "target-cpu"="pwr7" }
+