[SystemZ] Use "STC;MVC" for memset

Use "STC;MVC" for memsets that are too big for two STCs or MV...Is yet
small enough for a single MVC.  As with memcpy, I'm leaving longer cases
till later.

The number of tests might seem excessive, but f33 & f34 from memset-04.ll
failed the first cut because I'd not added the "?:" on the calculation
of Size1.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@185918 91177308-0d34-0410-b5e6-96231b3b80d8

7 files changed, 1171 insertions, 0 deletions

diff --git a/lib/Target/SystemZ/SystemZISelLowering.cpp b/lib/Target/SystemZ/SystemZISelLowering.cpp
index bf359468e6..b1abc2c3c1 100644
--- a/lib/Target/SystemZ/SystemZISelLowering.cpp
+++ b/lib/Target/SystemZ/SystemZISelLowering.cpp
@@ -245,6 +245,14 @@ SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm)
   // We want to use MVC in preference to even a single load/store pair.
   MaxStoresPerMemcpy = 0;
   MaxStoresPerMemcpyOptSize = 0;
+
+  // The main memset sequence is a byte store followed by an MVC.
+  // Two STC or MV..I stores win over that, but the kind of fused stores
+  // generated by target-independent code don't when the byte value is
+  // variable.  E.g.  "STC <reg>;MHI <reg>,257;STH <reg>" is not better
+  // than "STC;MVC".  Handle the choice in target-specific code instead.
+  MaxStoresPerMemset = 0;
+  MaxStoresPerMemsetOptSize = 0;
 }
 
 bool SystemZTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
diff --git a/lib/Target/SystemZ/SystemZSelectionDAGInfo.cpp b/lib/Target/SystemZ/SystemZSelectionDAGInfo.cpp
index d2da9d2a03..4ca9292092 100644
--- a/lib/Target/SystemZ/SystemZSelectionDAGInfo.cpp
+++ b/lib/Target/SystemZ/SystemZSelectionDAGInfo.cpp
@@ -44,3 +44,84 @@ EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
   }
   return SDValue();
 }
+
+// Handle a memset of 1, 2, 4 or 8 bytes with the operands given by
+// Chain, Dst, ByteVal and Size.  These cases are expected to use
+// MVI, MVHHI, MVHI and MVGHI respectively.
+static SDValue memsetStore(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
+                           SDValue Dst, uint64_t ByteVal, uint64_t Size,
+                           unsigned Align,
+                           MachinePointerInfo DstPtrInfo) {
+  uint64_t StoreVal = ByteVal;
+  for (unsigned I = 1; I < Size; ++I)
+    StoreVal |= ByteVal << (I * 8);
+  return DAG.getStore(Chain, DL,
+                      DAG.getConstant(StoreVal, MVT::getIntegerVT(Size * 8)),
+                      Dst, DstPtrInfo, false, false, Align);
+}
+
+SDValue SystemZSelectionDAGInfo::
+EmitTargetCodeForMemset(SelectionDAG &DAG, SDLoc DL, SDValue Chain,
+                        SDValue Dst, SDValue Byte, SDValue Size,
+                        unsigned Align, bool IsVolatile,
+                        MachinePointerInfo DstPtrInfo) const {
+  EVT DstVT = Dst.getValueType();
+
+  if (IsVolatile)
+    return SDValue();
+
+  if (ConstantSDNode *CSize = dyn_cast<ConstantSDNode>(Size)) {
+    uint64_t Bytes = CSize->getZExtValue();
+    if (Bytes == 0)
+      return SDValue();
+    if (ConstantSDNode *CByte = dyn_cast<ConstantSDNode>(Byte)) {
+      // Handle cases that can be done using at most two of
+      // MVI, MVHI, MVHHI and MVGHI.  The latter two can only be
+      // used if ByteVal is all zeros or all ones; in other casees,
+      // we can move at most 2 halfwords.
+      uint64_t ByteVal = CByte->getZExtValue();
+      if (ByteVal == 0 || ByteVal == 255 ?
+          Bytes <= 16 && CountPopulation_64(Bytes) <= 2 :
+          Bytes <= 4) {
+        unsigned Size1 = Bytes == 16 ? 8 : 1 << findLastSet(Bytes);
+        unsigned Size2 = Bytes - Size1;
+        SDValue Chain1 = memsetStore(DAG, DL, Chain, Dst, ByteVal, Size1,
+                                     Align, DstPtrInfo);
+        if (Size2 == 0)
+          return Chain1;
+        Dst = DAG.getNode(ISD::ADD, DL, DstVT, Dst,
+                          DAG.getConstant(Size1, DstVT));
+        DstPtrInfo = DstPtrInfo.getWithOffset(Size1);
+        SDValue Chain2 = memsetStore(DAG, DL, Chain, Dst, ByteVal, Size2,
+                                     std::min(Align, Size1), DstPtrInfo);
+        return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chain1, Chain2);
+      }
+    } else {
+      // Handle one and two bytes using STC.
+      if (Bytes <= 2) {
+        SDValue Chain1 = DAG.getStore(Chain, DL, Byte, Dst, DstPtrInfo,
+                                      false, false, Align);
+        if (Bytes == 1)
+          return Chain1;
+        SDValue Dst2 = DAG.getNode(ISD::ADD, DL, DstVT, Dst,
+                                   DAG.getConstant(1, DstVT));
+        SDValue Chain2 = DAG.getStore(Chain, DL, Byte, Dst2,
+                                      DstPtrInfo.getWithOffset(1),
+                                      false, false, 1);
+        return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chain1, Chain2);
+      }
+    }
+    assert(Bytes >= 2 && "Should have dealt with 0- and 1-byte cases already");
+    if (Bytes <= 0x101) {
+      // Copy the byte to the first location and then use MVC to copy
+      // it to the rest.
+      Chain = DAG.getStore(Chain, DL, Byte, Dst, DstPtrInfo,
+                           false, false, Align);
+      SDValue Dst2 = DAG.getNode(ISD::ADD, DL, DstVT, Dst,
+                                 DAG.getConstant(1, DstVT));
+      return DAG.getNode(SystemZISD::MVC, DL, MVT::Other, Chain, Dst2, Dst,
+                         DAG.getConstant(Bytes - 1, MVT::i32));
+    }
+  }
+  return SDValue();
+}
diff --git a/lib/Target/SystemZ/SystemZSelectionDAGInfo.h b/lib/Target/SystemZ/SystemZSelectionDAGInfo.h
index 39c149137c..9138a9cc08 100644
--- a/lib/Target/SystemZ/SystemZSelectionDAGInfo.h
+++ b/lib/Target/SystemZ/SystemZSelectionDAGInfo.h
@@ -33,6 +33,12 @@ public:
                                   MachinePointerInfo DstPtrInfo,
                                   MachinePointerInfo SrcPtrInfo) const
     LLVM_OVERRIDE;
+
+  virtual SDValue
+  EmitTargetCodeForMemset(SelectionDAG &DAG, SDLoc DL,
+                          SDValue Chain, SDValue Dst, SDValue Byte,
+                          SDValue Size, unsigned Align, bool IsVolatile,
+                          MachinePointerInfo DstPtrInfo) const;
 };
 
 }
diff --git a/test/CodeGen/SystemZ/memset-01.ll b/test/CodeGen/SystemZ/memset-01.ll
new file mode 100644
index 0000000000..1592318b8b
--- /dev/null
+++ b/test/CodeGen/SystemZ/memset-01.ll
@@ -0,0 +1,124 @@
+; Test memset in cases where the set value is variable.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
+
+declare void @llvm.memset.p0i8.i32(i8 *nocapture, i8, i32, i32, i1) nounwind
+declare void @llvm.memset.p0i8.i64(i8 *nocapture, i8, i64, i32, i1) nounwind
+
+; No bytes, i32 version.
+define void @f1(i8 *%dest, i8 %val) {
+; CHECK: f1:
+; CHECK-NOT: %r2
+; CHECK-NOT: %r3
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 %val, i32 0, i32 1, i1 false)
+  ret void
+}
+
+; No bytes, i64 version.
+define void @f2(i8 *%dest, i8 %val) {
+; CHECK: f2:
+; CHECK-NOT: %r2
+; CHECK-NOT: %r3
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 %val, i64 0, i32 1, i1 false)
+  ret void
+}
+
+; 1 byte, i32 version.
+define void @f3(i8 *%dest, i8 %val) {
+; CHECK: f3:
+; CHECK: stc %r3, 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 %val, i32 1, i32 1, i1 false)
+  ret void
+}
+
+; 1 byte, i64 version.
+define void @f4(i8 *%dest, i8 %val) {
+; CHECK: f4:
+; CHECK: stc %r3, 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 %val, i64 1, i32 1, i1 false)
+  ret void
+}
+
+; 2 bytes, i32 version.
+define void @f5(i8 *%dest, i8 %val) {
+; CHECK: f5:
+; CHECK-DAG: stc %r3, 0(%r2)
+; CHECK-DAG: stc %r3, 1(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 %val, i32 2, i32 1, i1 false)
+  ret void
+}
+
+; 2 bytes, i64 version.
+define void @f6(i8 *%dest, i8 %val) {
+; CHECK: f6:
+; CHECK-DAG: stc %r3, 0(%r2)
+; CHECK-DAG: stc %r3, 1(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 %val, i64 2, i32 1, i1 false)
+  ret void
+}
+
+; 3 bytes, i32 version.
+define void @f7(i8 *%dest, i8 %val) {
+; CHECK: f7:
+; CHECK: stc %r3, 0(%r2)
+; CHECK: mvc 1(2,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 %val, i32 3, i32 1, i1 false)
+  ret void
+}
+
+; 3 bytes, i64 version.
+define void @f8(i8 *%dest, i8 %val) {
+; CHECK: f8:
+; CHECK: stc %r3, 0(%r2)
+; CHECK: mvc 1(2,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 %val, i64 3, i32 1, i1 false)
+  ret void
+}
+
+; 257 bytes, i32 version.
+define void @f9(i8 *%dest, i8 %val) {
+; CHECK: f9:
+; CHECK: stc %r3, 0(%r2)
+; CHECK: mvc 1(256,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 %val, i32 257, i32 1, i1 false)
+  ret void
+}
+
+; 257 bytes, i64 version.
+define void @f10(i8 *%dest, i8 %val) {
+; CHECK: f10:
+; CHECK: stc %r3, 0(%r2)
+; CHECK: mvc 1(256,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 %val, i64 257, i32 1, i1 false)
+  ret void
+}
+
+; 258 bytes, i32 version.  258 bytes is too big for a single MVC.
+; For now expect none, so that the test fails and gets updated when
+; large copies are implemented.
+define void @f11(i8 *%dest, i8 %val) {
+; CHECK: f11:
+; CHECK-NOT: mvc
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 %val, i32 258, i32 1, i1 false)
+  ret void
+}
+
+; 258 bytes, i64 version, with the same comments as above.
+define void @f12(i8 *%dest, i8 %val) {
+; CHECK: f12:
+; CHECK-NOT: mvc
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 %val, i64 258, i32 1, i1 false)
+  ret void
+}
diff --git a/test/CodeGen/SystemZ/memset-02.ll b/test/CodeGen/SystemZ/memset-02.ll
new file mode 100644
index 0000000000..c2c45fb4ae
--- /dev/null
+++ b/test/CodeGen/SystemZ/memset-02.ll
@@ -0,0 +1,160 @@
+; Test memset in cases where the set value is a constant other than 0 and -1.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
+
+declare void @llvm.memset.p0i8.i32(i8 *nocapture, i8, i32, i32, i1) nounwind
+declare void @llvm.memset.p0i8.i64(i8 *nocapture, i8, i64, i32, i1) nounwind
+
+; No bytes, i32 version.
+define void @f1(i8 *%dest) {
+; CHECK: f1:
+; CHECK-NOT: %r2
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 128, i32 0, i32 1, i1 false)
+  ret void
+}
+
+; No bytes, i64 version.
+define void @f2(i8 *%dest) {
+; CHECK: f2:
+; CHECK-NOT: %r2
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 128, i64 0, i32 1, i1 false)
+  ret void
+}
+
+; 1 byte, i32 version.
+define void @f3(i8 *%dest) {
+; CHECK: f3:
+; CHECK: mvi 0(%r2), 128
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 128, i32 1, i32 1, i1 false)
+  ret void
+}
+
+; 1 byte, i64 version.
+define void @f4(i8 *%dest) {
+; CHECK: f4:
+; CHECK: mvi 0(%r2), 128
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 128, i64 1, i32 1, i1 false)
+  ret void
+}
+
+; 2 bytes, i32 version.
+define void @f5(i8 *%dest) {
+; CHECK: f5:
+; CHECK: mvhhi 0(%r2), -32640
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 128, i32 2, i32 1, i1 false)
+  ret void
+}
+
+; 2 bytes, i64 version.
+define void @f6(i8 *%dest) {
+; CHECK: f6:
+; CHECK: mvhhi 0(%r2), -32640
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 128, i64 2, i32 1, i1 false)
+  ret void
+}
+
+; 3 bytes, i32 version.
+define void @f7(i8 *%dest) {
+; CHECK: f7:
+; CHECK-DAG: mvhhi 0(%r2), -32640
+; CHECK-DAG: mvi 2(%r2), 128
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 128, i32 3, i32 1, i1 false)
+  ret void
+}
+
+; 3 bytes, i64 version.
+define void @f8(i8 *%dest) {
+; CHECK: f8:
+; CHECK-DAG: mvhhi 0(%r2), -32640
+; CHECK-DAG: mvi 2(%r2), 128
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 128, i64 3, i32 1, i1 false)
+  ret void
+}
+
+; 4 bytes, i32 version.
+define void @f9(i8 *%dest) {
+; CHECK: f9:
+; CHECK: iilf [[REG:%r[0-5]]], 2155905152
+; CHECK: st [[REG]], 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 128, i32 4, i32 1, i1 false)
+  ret void
+}
+
+; 4 bytes, i64 version.
+define void @f10(i8 *%dest) {
+; CHECK: f10:
+; CHECK: iilf [[REG:%r[0-5]]], 2155905152
+; CHECK: st [[REG]], 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 128, i64 4, i32 1, i1 false)
+  ret void
+}
+
+; 5 bytes, i32 version.
+define void @f11(i8 *%dest) {
+; CHECK: f11:
+; CHECK: mvi 0(%r2), 128
+; CHECK: mvc 1(4,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 128, i32 5, i32 1, i1 false)
+  ret void
+}
+
+; 5 bytes, i64 version.
+define void @f12(i8 *%dest) {
+; CHECK: f12:
+; CHECK: mvi 0(%r2), 128
+; CHECK: mvc 1(4,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 128, i64 5, i32 1, i1 false)
+  ret void
+}
+
+; 257 bytes, i32 version.
+define void @f13(i8 *%dest) {
+; CHECK: f13:
+; CHECK: mvi 0(%r2), 128
+; CHECK: mvc 1(256,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 128, i32 257, i32 1, i1 false)
+  ret void
+}
+
+; 257 bytes, i64 version.
+define void @f14(i8 *%dest) {
+; CHECK: f14:
+; CHECK: mvi 0(%r2), 128
+; CHECK: mvc 1(256,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 128, i64 257, i32 1, i1 false)
+  ret void
+}
+
+; 258 bytes, i32 version.  258 bytes is too big for a single MVC.
+; For now expect none, so that the test fails and gets updated when
+; large copies are implemented.
+define void @f15(i8 *%dest) {
+; CHECK: f15:
+; CHECK-NOT: mvc
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 128, i32 258, i32 1, i1 false)
+  ret void
+}
+
+; 258 bytes, i64 version, with the same comments as above.
+define void @f16(i8 *%dest) {
+; CHECK: f16:
+; CHECK-NOT: mvc
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 128, i64 258, i32 1, i1 false)
+  ret void
+}
diff --git a/test/CodeGen/SystemZ/memset-03.ll b/test/CodeGen/SystemZ/memset-03.ll
new file mode 100644
index 0000000000..b18cca4ec4
--- /dev/null
+++ b/test/CodeGen/SystemZ/memset-03.ll
@@ -0,0 +1,396 @@
+; Test memsets that clear all bits.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
+
+declare void @llvm.memset.p0i8.i32(i8 *nocapture, i8, i32, i32, i1) nounwind
+declare void @llvm.memset.p0i8.i64(i8 *nocapture, i8, i64, i32, i1) nounwind
+
+; No bytes, i32 version.
+define void @f1(i8 *%dest) {
+; CHECK: f1:
+; CHECK-NOT: %r2
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 0, i32 1, i1 false)
+  ret void
+}
+
+; No bytes, i64 version.
+define void @f2(i8 *%dest) {
+; CHECK: f2:
+; CHECK-NOT: %r2
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 0, i32 1, i1 false)
+  ret void
+}
+
+; 1 byte, i32 version.
+define void @f3(i8 *%dest) {
+; CHECK: f3:
+; CHECK: mvi 0(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 1, i32 1, i1 false)
+  ret void
+}
+
+; 1 byte, i64 version.
+define void @f4(i8 *%dest) {
+; CHECK: f4:
+; CHECK: mvi 0(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 1, i32 1, i1 false)
+  ret void
+}
+
+; 2 bytes, i32 version.
+define void @f5(i8 *%dest) {
+; CHECK: f5:
+; CHECK: mvhhi 0(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 2, i32 1, i1 false)
+  ret void
+}
+
+; 2 bytes, i64 version.
+define void @f6(i8 *%dest) {
+; CHECK: f6:
+; CHECK: mvhhi 0(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 2, i32 1, i1 false)
+  ret void
+}
+
+; 3 bytes, i32 version.
+define void @f7(i8 *%dest) {
+; CHECK: f7:
+; CHECK-DAG: mvhhi 0(%r2), 0
+; CHECK-DAG: mvi 2(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 3, i32 1, i1 false)
+  ret void
+}
+
+; 3 bytes, i64 version.
+define void @f8(i8 *%dest) {
+; CHECK: f8:
+; CHECK-DAG: mvhhi 0(%r2), 0
+; CHECK-DAG: mvi 2(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 3, i32 1, i1 false)
+  ret void
+}
+
+; 4 bytes, i32 version.
+define void @f9(i8 *%dest) {
+; CHECK: f9:
+; CHECK: mvhi 0(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 4, i32 1, i1 false)
+  ret void
+}
+
+; 4 bytes, i64 version.
+define void @f10(i8 *%dest) {
+; CHECK: f10:
+; CHECK: mvhi 0(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 4, i32 1, i1 false)
+  ret void
+}
+
+; 5 bytes, i32 version.
+define void @f11(i8 *%dest) {
+; CHECK: f11:
+; CHECK-DAG: mvhi 0(%r2), 0
+; CHECK-DAG: mvi 4(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 5, i32 1, i1 false)
+  ret void
+}
+
+; 5 bytes, i64 version.
+define void @f12(i8 *%dest) {
+; CHECK: f12:
+; CHECK-DAG: mvhi 0(%r2), 0
+; CHECK-DAG: mvi 4(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 5, i32 1, i1 false)
+  ret void
+}
+
+; 6 bytes, i32 version.
+define void @f13(i8 *%dest) {
+; CHECK: f13:
+; CHECK-DAG: mvhi 0(%r2), 0
+; CHECK-DAG: mvhhi 4(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 6, i32 1, i1 false)
+  ret void
+}
+
+; 6 bytes, i64 version.
+define void @f14(i8 *%dest) {
+; CHECK: f14:
+; CHECK-DAG: mvhi 0(%r2), 0
+; CHECK-DAG: mvhhi 4(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 6, i32 1, i1 false)
+  ret void
+}
+
+; 7 bytes, i32 version.
+define void @f15(i8 *%dest) {
+; CHECK: f15:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(6,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 7, i32 1, i1 false)
+  ret void
+}
+
+; 7 bytes, i64 version.
+define void @f16(i8 *%dest) {
+; CHECK: f16:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(6,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 7, i32 1, i1 false)
+  ret void
+}
+
+; 8 bytes, i32 version.
+define void @f17(i8 *%dest) {
+; CHECK: f17:
+; CHECK: mvghi 0(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 8, i32 1, i1 false)
+  ret void
+}
+
+; 8 bytes, i64 version.
+define void @f18(i8 *%dest) {
+; CHECK: f18:
+; CHECK: mvghi 0(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 8, i32 1, i1 false)
+  ret void
+}
+
+; 9 bytes, i32 version.
+define void @f19(i8 *%dest) {
+; CHECK: f19:
+; CHECK-DAG: mvghi 0(%r2), 0
+; CHECK-DAG: mvi 8(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 9, i32 1, i1 false)
+  ret void
+}
+
+; 9 bytes, i64 version.
+define void @f20(i8 *%dest) {
+; CHECK: f20:
+; CHECK-DAG: mvghi 0(%r2), 0
+; CHECK-DAG: mvi 8(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 9, i32 1, i1 false)
+  ret void
+}
+
+; 10 bytes, i32 version.
+define void @f21(i8 *%dest) {
+; CHECK: f21:
+; CHECK-DAG: mvghi 0(%r2), 0
+; CHECK-DAG: mvhhi 8(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 10, i32 1, i1 false)
+  ret void
+}
+
+; 10 bytes, i64 version.
+define void @f22(i8 *%dest) {
+; CHECK: f22:
+; CHECK-DAG: mvghi 0(%r2), 0
+; CHECK-DAG: mvhhi 8(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 10, i32 1, i1 false)
+  ret void
+}
+
+; 11 bytes, i32 version.
+define void @f23(i8 *%dest) {
+; CHECK: f23:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(10,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 11, i32 1, i1 false)
+  ret void
+}
+
+; 11 bytes, i64 version.
+define void @f24(i8 *%dest) {
+; CHECK: f24:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(10,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 11, i32 1, i1 false)
+  ret void
+}
+
+; 12 bytes, i32 version.
+define void @f25(i8 *%dest) {
+; CHECK: f25:
+; CHECK-DAG: mvghi 0(%r2), 0
+; CHECK-DAG: mvhi 8(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 12, i32 1, i1 false)
+  ret void
+}
+
+; 12 bytes, i64 version.
+define void @f26(i8 *%dest) {
+; CHECK: f26:
+; CHECK-DAG: mvghi 0(%r2), 0
+; CHECK-DAG: mvhi 8(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 12, i32 1, i1 false)
+  ret void
+}
+
+; 13 bytes, i32 version.
+define void @f27(i8 *%dest) {
+; CHECK: f27:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(12,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 13, i32 1, i1 false)
+  ret void
+}
+
+; 13 bytes, i64 version.
+define void @f28(i8 *%dest) {
+; CHECK: f28:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(12,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 13, i32 1, i1 false)
+  ret void
+}
+
+; 14 bytes, i32 version.
+define void @f29(i8 *%dest) {
+; CHECK: f29:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(13,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 14, i32 1, i1 false)
+  ret void
+}
+
+; 14 bytes, i64 version.
+define void @f30(i8 *%dest) {
+; CHECK: f30:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(13,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 14, i32 1, i1 false)
+  ret void
+}
+
+; 15 bytes, i32 version.
+define void @f31(i8 *%dest) {
+; CHECK: f31:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(14,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 15, i32 1, i1 false)
+  ret void
+}
+
+; 15 bytes, i64 version.
+define void @f32(i8 *%dest) {
+; CHECK: f32:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(14,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 15, i32 1, i1 false)
+  ret void
+}
+
+; 16 bytes, i32 version.
+define void @f33(i8 *%dest) {
+; CHECK: f33:
+; CHECK-DAG: mvghi 0(%r2), 0
+; CHECK-DAG: mvghi 8(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 16, i32 1, i1 false)
+  ret void
+}
+
+; 16 bytes, i64 version.
+define void @f34(i8 *%dest) {
+; CHECK: f34:
+; CHECK-DAG: mvghi 0(%r2), 0
+; CHECK-DAG: mvghi 8(%r2), 0
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 16, i32 1, i1 false)
+  ret void
+}
+
+; 17 bytes, i32 version.
+define void @f35(i8 *%dest) {
+; CHECK: f35:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(16,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 17, i32 1, i1 false)
+  ret void
+}
+
+; 17 bytes, i64 version.
+define void @f36(i8 *%dest) {
+; CHECK: f36:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(16,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 17, i32 1, i1 false)
+  ret void
+}
+
+; 257 bytes, i32 version.
+define void @f37(i8 *%dest) {
+; CHECK: f37:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(256,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 257, i32 1, i1 false)
+  ret void
+}
+
+; 257 bytes, i64 version.
+define void @f38(i8 *%dest) {
+; CHECK: f38:
+; CHECK: mvi 0(%r2), 0
+; CHECK: mvc 1(256,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 257, i32 1, i1 false)
+  ret void
+}
+
+; 258 bytes, i32 version.  258 bytes is too big for a single MVC.
+; For now expect none, so that the test fails and gets updated when
+; large copies are implemented.
+define void @f39(i8 *%dest) {
+; CHECK: f39:
+; CHECK-NOT: mvc
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 0, i32 258, i32 1, i1 false)
+  ret void
+}
+
+; 258 bytes, i64 version, with the same comments as above.
+define void @f40(i8 *%dest) {
+; CHECK: f40:
+; CHECK-NOT: mvc
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 0, i64 258, i32 1, i1 false)
+  ret void
+}
diff --git a/test/CodeGen/SystemZ/memset-04.ll b/test/CodeGen/SystemZ/memset-04.ll
new file mode 100644
index 0000000000..679e21f8fc
--- /dev/null
+++ b/test/CodeGen/SystemZ/memset-04.ll
@@ -0,0 +1,396 @@
+; Test memsets that set all bits.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
+
+declare void @llvm.memset.p0i8.i32(i8 *nocapture, i8, i32, i32, i1) nounwind
+declare void @llvm.memset.p0i8.i64(i8 *nocapture, i8, i64, i32, i1) nounwind
+
+; No bytes, i32 version.
+define void @f1(i8 *%dest) {
+; CHECK: f1:
+; CHECK-NOT: %r2
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 0, i32 1, i1 false)
+  ret void
+}
+
+; No bytes, i64 version.
+define void @f2(i8 *%dest) {
+; CHECK: f2:
+; CHECK-NOT: %r2
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 0, i32 1, i1 false)
+  ret void
+}
+
+; 1 byte, i32 version.
+define void @f3(i8 *%dest) {
+; CHECK: f3:
+; CHECK: mvi 0(%r2), 255
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 1, i32 1, i1 false)
+  ret void
+}
+
+; 1 byte, i64 version.
+define void @f4(i8 *%dest) {
+; CHECK: f4:
+; CHECK: mvi 0(%r2), 255
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 1, i32 1, i1 false)
+  ret void
+}
+
+; 2 bytes, i32 version.
+define void @f5(i8 *%dest) {
+; CHECK: f5:
+; CHECK: mvhhi 0(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 2, i32 1, i1 false)
+  ret void
+}
+
+; 2 bytes, i64 version.
+define void @f6(i8 *%dest) {
+; CHECK: f6:
+; CHECK: mvhhi 0(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 2, i32 1, i1 false)
+  ret void
+}
+
+; 3 bytes, i32 version.
+define void @f7(i8 *%dest) {
+; CHECK: f7:
+; CHECK-DAG: mvhhi 0(%r2), -1
+; CHECK-DAG: mvi 2(%r2), 255
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 3, i32 1, i1 false)
+  ret void
+}
+
+; 3 bytes, i64 version.
+define void @f8(i8 *%dest) {
+; CHECK: f8:
+; CHECK-DAG: mvhhi 0(%r2), -1
+; CHECK-DAG: mvi 2(%r2), 255
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 3, i32 1, i1 false)
+  ret void
+}
+
+; 4 bytes, i32 version.
+define void @f9(i8 *%dest) {
+; CHECK: f9:
+; CHECK: mvhi 0(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 4, i32 1, i1 false)
+  ret void
+}
+
+; 4 bytes, i64 version.
+define void @f10(i8 *%dest) {
+; CHECK: f10:
+; CHECK: mvhi 0(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 4, i32 1, i1 false)
+  ret void
+}
+
+; 5 bytes, i32 version.
+define void @f11(i8 *%dest) {
+; CHECK: f11:
+; CHECK-DAG: mvhi 0(%r2), -1
+; CHECK-DAG: mvi 4(%r2), 255
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 5, i32 1, i1 false)
+  ret void
+}
+
+; 5 bytes, i64 version.
+define void @f12(i8 *%dest) {
+; CHECK: f12:
+; CHECK-DAG: mvhi 0(%r2), -1
+; CHECK-DAG: mvi 4(%r2), 255
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 5, i32 1, i1 false)
+  ret void
+}
+
+; 6 bytes, i32 version.
+define void @f13(i8 *%dest) {
+; CHECK: f13:
+; CHECK-DAG: mvhi 0(%r2), -1
+; CHECK-DAG: mvhhi 4(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 6, i32 1, i1 false)
+  ret void
+}
+
+; 6 bytes, i64 version.
+define void @f14(i8 *%dest) {
+; CHECK: f14:
+; CHECK-DAG: mvhi 0(%r2), -1
+; CHECK-DAG: mvhhi 4(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 6, i32 1, i1 false)
+  ret void
+}
+
+; 7 bytes, i32 version.
+define void @f15(i8 *%dest) {
+; CHECK: f15:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(6,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 7, i32 1, i1 false)
+  ret void
+}
+
+; 7 bytes, i64 version.
+define void @f16(i8 *%dest) {
+; CHECK: f16:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(6,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 7, i32 1, i1 false)
+  ret void
+}
+
+; 8 bytes, i32 version.
+define void @f17(i8 *%dest) {
+; CHECK: f17:
+; CHECK: mvghi 0(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 8, i32 1, i1 false)
+  ret void
+}
+
+; 8 bytes, i64 version.
+define void @f18(i8 *%dest) {
+; CHECK: f18:
+; CHECK: mvghi 0(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 8, i32 1, i1 false)
+  ret void
+}
+
+; 9 bytes, i32 version.
+define void @f19(i8 *%dest) {
+; CHECK: f19:
+; CHECK-DAG: mvghi 0(%r2), -1
+; CHECK-DAG: mvi 8(%r2), 255
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 9, i32 1, i1 false)
+  ret void
+}
+
+; 9 bytes, i64 version.
+define void @f20(i8 *%dest) {
+; CHECK: f20:
+; CHECK-DAG: mvghi 0(%r2), -1
+; CHECK-DAG: mvi 8(%r2), 255
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 9, i32 1, i1 false)
+  ret void
+}
+
+; 10 bytes, i32 version.
+define void @f21(i8 *%dest) {
+; CHECK: f21:
+; CHECK-DAG: mvghi 0(%r2), -1
+; CHECK-DAG: mvhhi 8(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 10, i32 1, i1 false)
+  ret void
+}
+
+; 10 bytes, i64 version.
+define void @f22(i8 *%dest) {
+; CHECK: f22:
+; CHECK-DAG: mvghi 0(%r2), -1
+; CHECK-DAG: mvhhi 8(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 10, i32 1, i1 false)
+  ret void
+}
+
+; 11 bytes, i32 version.
+define void @f23(i8 *%dest) {
+; CHECK: f23:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(10,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 11, i32 1, i1 false)
+  ret void
+}
+
+; 11 bytes, i64 version.
+define void @f24(i8 *%dest) {
+; CHECK: f24:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(10,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 11, i32 1, i1 false)
+  ret void
+}
+
+; 12 bytes, i32 version.
+define void @f25(i8 *%dest) {
+; CHECK: f25:
+; CHECK-DAG: mvghi 0(%r2), -1
+; CHECK-DAG: mvhi 8(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 12, i32 1, i1 false)
+  ret void
+}
+
+; 12 bytes, i64 version.
+define void @f26(i8 *%dest) {
+; CHECK: f26:
+; CHECK-DAG: mvghi 0(%r2), -1
+; CHECK-DAG: mvhi 8(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 12, i32 1, i1 false)
+  ret void
+}
+
+; 13 bytes, i32 version.
+define void @f27(i8 *%dest) {
+; CHECK: f27:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(12,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 13, i32 1, i1 false)
+  ret void
+}
+
+; 13 bytes, i64 version.
+define void @f28(i8 *%dest) {
+; CHECK: f28:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(12,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 13, i32 1, i1 false)
+  ret void
+}
+
+; 14 bytes, i32 version.
+define void @f29(i8 *%dest) {
+; CHECK: f29:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(13,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 14, i32 1, i1 false)
+  ret void
+}
+
+; 14 bytes, i64 version.
+define void @f30(i8 *%dest) {
+; CHECK: f30:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(13,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 14, i32 1, i1 false)
+  ret void
+}
+
+; 15 bytes, i32 version.
+define void @f31(i8 *%dest) {
+; CHECK: f31:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(14,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 15, i32 1, i1 false)
+  ret void
+}
+
+; 15 bytes, i64 version.
+define void @f32(i8 *%dest) {
+; CHECK: f32:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(14,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 15, i32 1, i1 false)
+  ret void
+}
+
+; 16 bytes, i32 version.
+define void @f33(i8 *%dest) {
+; CHECK: f33:
+; CHECK-DAG: mvghi 0(%r2), -1
+; CHECK-DAG: mvghi 8(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 16, i32 1, i1 false)
+  ret void
+}
+
+; 16 bytes, i64 version.
+define void @f34(i8 *%dest) {
+; CHECK: f34:
+; CHECK-DAG: mvghi 0(%r2), -1
+; CHECK-DAG: mvghi 8(%r2), -1
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 16, i32 1, i1 false)
+  ret void
+}
+
+; 17 bytes, i32 version.
+define void @f35(i8 *%dest) {
+; CHECK: f35:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(16,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 17, i32 1, i1 false)
+  ret void
+}
+
+; 17 bytes, i64 version.
+define void @f36(i8 *%dest) {
+; CHECK: f36:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(16,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 17, i32 1, i1 false)
+  ret void
+}
+
+; 257 bytes, i32 version.
+define void @f37(i8 *%dest) {
+; CHECK: f37:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(256,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 257, i32 1, i1 false)
+  ret void
+}
+
+; 257 bytes, i64 version.
+define void @f38(i8 *%dest) {
+; CHECK: f38:
+; CHECK: mvi 0(%r2), 255
+; CHECK: mvc 1(256,%r2), 0(%r2)
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 257, i32 1, i1 false)
+  ret void
+}
+
+; 258 bytes, i32 version.  258 bytes is too big for a single MVC.
+; For now expect none, so that the test fails and gets updated when
+; large copies are implemented.
+define void @f39(i8 *%dest) {
+; CHECK: f39:
+; CHECK-NOT: mvc
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i32(i8 *%dest, i8 -1, i32 258, i32 1, i1 false)
+  ret void
+}
+
+; 258 bytes, i64 version, with the same comments as above.
+define void @f40(i8 *%dest) {
+; CHECK: f40:
+; CHECK-NOT: mvc
+; CHECK: br %r14
+  call void @llvm.memset.p0i8.i64(i8 *%dest, i8 -1, i64 258, i32 1, i1 false)
+  ret void
+}