[PowerPC] ELFv2 aggregate passing support

This patch adds infrastructure support for passing array types
directly.  These can be used by the front-end to pass aggregate
types (coerced to an appropriate array type).  The details of the
array type being used inform the back-end about ABI-relevant
properties.  Specifically, the array element type encodes:
- whether the parameter should be passed in FPRs, VRs, or just
  GPRs/stack slots  (for float / vector / integer element types,
  respectively)
- what the alignment requirements of the parameter are when passed in
  GPRs/stack slots  (8 for float / 16 for vector / the element type
  size for integer element types) -- this corresponds to the
  "byval align" field

Using the infrastructure provided by this patch, a companion patch
to clang will enable two features:
- In the ELFv2 ABI, pass (and return) "homogeneous" floating-point
  or vector aggregates in FPRs and VRs (this is similar to the ARM
  homogeneous aggregate ABI)
- As an optimization for both ELFv1 and ELFv2 ABIs, pass aggregates
  that fit fully in registers without using the "byval" mechanism

The patch uses the functionArgumentNeedsConsecutiveRegisters callback
to encode that special treatment is required for all directly-passed
array types.  The isInConsecutiveRegs / isInConsecutiveRegsLast bits set
as a results are then used to implement the required size and alignment
rules in CalculateStackSlotSize / CalculateStackSlotAlignment etc.

As a related change, the ABI routines have to be modified to support
passing floating-point types in GPRs.  This is necessary because with
homogeneous aggregates of 4-byte float type we can now run out of FPRs
*before* we run out of the 64-byte argument save area that is shadowed
by GPRs.  Any extra floating-point arguments that no longer fit in FPRs
must now be passed in GPRs until we run out of those too.

Note that there was already code to pass floating-point arguments in
GPRs used with vararg parameters, which was done by writing the argument
out to the argument save area first and then reloading into GPRs.  The
patch re-implements this, however, in favor of code packing float arguments
directly via extension/truncation, BITCAST, and BUILD_PAIR operations.

This is required to support the ELFv2 ABI, since we cannot unconditionally
write to the argument save area (which the caller might not have allocated).
The change does, however, affect ELFv1 varags routines too; but even here
the overall effect should be advantageous: Instead of loading the argument
into the FPR, then storing the argument to the stack slot, and finally
reloading the argument from the stack slot into a GPR, the new code now
just loads the argument into the FPR, and subsequently loads the argument
into the GPR (via BITCAST).  That BITCAST might imply a save/reload from
a stack temporary (in which case we're no worse than before); but it
might be implemented more efficiently in some cases.

The final part of the patch enables up to 8 FPRs and VRs for argument
return in PPCCallingConv.td; this is required to support returning
ELFv2 homogeneous aggregates.  (Note that this doesn't affect other ABIs
since LLVM wil only look for which register to use if the parameter is
marked as "direct" return anyway.)

Reviewed by Hal Finkel.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213493 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 502 insertions, 51 deletions

diff --git a/lib/Target/PowerPC/PPCCallingConv.td b/lib/Target/PowerPC/PPCCallingConv.td
index d48164d6216..222760a0cb9 100644
--- a/lib/Target/PowerPC/PPCCallingConv.td
+++ b/lib/Target/PowerPC/PPCCallingConv.td
@@ -31,13 +31,18 @@ def RetCC_PPC : CallingConv<[
   CCIfType<[i32], CCAssignToReg<[R3, R4, R5, R6, R7, R8, R9, R10]>>,
   CCIfType<[i64], CCAssignToReg<[X3, X4, X5, X6]>>,
   CCIfType<[i128], CCAssignToReg<[X3, X4, X5, X6]>>,
+
+  // Floating point types returned as "direct" go into F1 .. F8; note that
+  // only the ELFv2 ABI fully utilizes all these registers.
+  CCIfType<[f32], CCAssignToReg<[F1, F2, F3, F4, F5, F6, F7, F8]>>,
+  CCIfType<[f64], CCAssignToReg<[F1, F2, F3, F4, F5, F6, F7, F8]>>,
   
-  CCIfType<[f32], CCAssignToReg<[F1, F2]>>,
-  CCIfType<[f64], CCAssignToReg<[F1, F2, F3, F4]>>,
-  
-  // Vector types are always returned in V2.
-  CCIfType<[v16i8, v8i16, v4i32, v4f32], CCAssignToReg<[V2]>>,
-  CCIfType<[v2f64, v2i64], CCAssignToReg<[VSH2]>>
+  // Vector types returned as "direct" go into V2 .. V9; note that only the
+  // ELFv2 ABI fully utilizes all these registers.
+  CCIfType<[v16i8, v8i16, v4i32, v4f32],
+           CCAssignToReg<[V2, V3, V4, V5, V6, V7, V8, V9]>>,
+  CCIfType<[v2f64, v2i64],
+           CCAssignToReg<[VSH2, VSH3, VSH4, VSH5, VSH6, VSH7, VSH8, VSH9]>>
 ]>;
 
 
@@ -69,10 +74,12 @@ def RetCC_PPC64_ELF_FIS : CallingConv<[
   CCIfType<[i32],  CCPromoteToType<i64>>,
   CCIfType<[i64],  CCAssignToReg<[X3, X4]>>,
   CCIfType<[i128], CCAssignToReg<[X3, X4, X5, X6]>>,
-  CCIfType<[f32],  CCAssignToReg<[F1, F2]>>,
-  CCIfType<[f64],  CCAssignToReg<[F1, F2, F3, F4]>>,
-  CCIfType<[v16i8, v8i16, v4i32, v4f32], CCAssignToReg<[V2]>>,
-  CCIfType<[v2f64, v2i64], CCAssignToReg<[VSH2]>>
+  CCIfType<[f32],  CCAssignToReg<[F1, F2, F3, F4, F5, F6, F7, F8]>>,
+  CCIfType<[f64],  CCAssignToReg<[F1, F2, F3, F4, F5, F6, F7, F8]>>,
+  CCIfType<[v16i8, v8i16, v4i32, v4f32],
+           CCAssignToReg<[V2, V3, V4, V5, V6, V7, V8, V9]>>,
+  CCIfType<[v2f64, v2i64],
+           CCAssignToReg<[VSH2, VSH3, VSH4, VSH5, VSH6, VSH7, VSH8, VSH9]>>
 ]>;
 
 //===----------------------------------------------------------------------===//
diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index fb40d5e204e..d699e0fdd82 100644
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -2158,14 +2158,19 @@ static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
   unsigned ArgSize = ArgVT.getStoreSize();
   if (Flags.isByVal())
     ArgSize = Flags.getByValSize();
-  ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
+
+  // Round up to multiples of the pointer size, except for array members,
+  // which are always packed.
+  if (!Flags.isInConsecutiveRegs())
+    ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
 
   return ArgSize;
 }
 
 /// CalculateStackSlotAlignment - Calculates the alignment of this argument
 /// on the stack.
-static unsigned CalculateStackSlotAlignment(EVT ArgVT, ISD::ArgFlagsTy Flags,
+static unsigned CalculateStackSlotAlignment(EVT ArgVT, EVT OrigVT,
+                                            ISD::ArgFlagsTy Flags,
                                             unsigned PtrByteSize) {
   unsigned Align = PtrByteSize;
 
@@ -2187,6 +2192,17 @@ static unsigned CalculateStackSlotAlignment(EVT ArgVT, ISD::ArgFlagsTy Flags,
     }
   }
 
+  // Array members are always packed to their original alignment.
+  if (Flags.isInConsecutiveRegs()) {
+    // If the array member was split into multiple registers, the first
+    // needs to be aligned to the size of the full type.  (Except for
+    // ppcf128, which is only aligned as its f64 components.)
+    if (Flags.isSplit() && OrigVT != MVT::ppcf128)
+      Align = OrigVT.getStoreSize();
+    else
+      Align = ArgVT.getStoreSize();
+  }
+
   return Align;
 }
 
@@ -2194,7 +2210,8 @@ static unsigned CalculateStackSlotAlignment(EVT ArgVT, ISD::ArgFlagsTy Flags,
 /// stack slot (instead of being passed in registers).  ArgOffset,
 /// AvailableFPRs, and AvailableVRs must hold the current argument
 /// position, and will be updated to account for this argument.
-static bool CalculateStackSlotUsed(EVT ArgVT, ISD::ArgFlagsTy Flags,
+static bool CalculateStackSlotUsed(EVT ArgVT, EVT OrigVT,
+                                   ISD::ArgFlagsTy Flags,
                                    unsigned PtrByteSize,
                                    unsigned LinkageSize,
                                    unsigned ParamAreaSize,
@@ -2204,7 +2221,8 @@ static bool CalculateStackSlotUsed(EVT ArgVT, ISD::ArgFlagsTy Flags,
   bool UseMemory = false;
 
   // Respect alignment of argument on the stack.
-  unsigned Align = CalculateStackSlotAlignment(ArgVT, Flags, PtrByteSize);
+  unsigned Align =
+    CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
   ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
   // If there's no space left in the argument save area, we must
   // use memory (this check also catches zero-sized arguments).
@@ -2213,6 +2231,8 @@ static bool CalculateStackSlotUsed(EVT ArgVT, ISD::ArgFlagsTy Flags,
 
   // Allocate argument on the stack.
   ArgOffset += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
+  if (Flags.isInConsecutiveRegsLast())
+    ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
   // If we overran the argument save area, we must use memory
   // (this check catches arguments passed partially in memory)
   if (ArgOffset > LinkageSize + ParamAreaSize)
@@ -2563,7 +2583,7 @@ PPCTargetLowering::LowerFormalArguments_64SVR4(
   unsigned AvailableFPRs = Num_FPR_Regs;
   unsigned AvailableVRs = Num_VR_Regs;
   for (unsigned i = 0, e = Ins.size(); i != e; ++i)
-    if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].Flags,
+    if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].ArgVT, Ins[i].Flags,
                                PtrByteSize, LinkageSize, ParamAreaSize,
                                NumBytes, AvailableFPRs, AvailableVRs))
       HasParameterArea = true;
@@ -2581,6 +2601,7 @@ PPCTargetLowering::LowerFormalArguments_64SVR4(
     SDValue ArgVal;
     bool needsLoad = false;
     EVT ObjectVT = Ins[ArgNo].VT;
+    EVT OrigVT = Ins[ArgNo].ArgVT;
     unsigned ObjSize = ObjectVT.getStoreSize();
     unsigned ArgSize = ObjSize;
     ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
@@ -2589,7 +2610,7 @@ PPCTargetLowering::LowerFormalArguments_64SVR4(
 
     /* Respect alignment of argument on the stack.  */
     unsigned Align =
-      CalculateStackSlotAlignment(ObjectVT, Flags, PtrByteSize);
+      CalculateStackSlotAlignment(ObjectVT, OrigVT, Flags, PtrByteSize);
     ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
     unsigned CurArgOffset = ArgOffset;
 
@@ -2701,6 +2722,9 @@ PPCTargetLowering::LowerFormalArguments_64SVR4(
     case MVT::i1:
     case MVT::i32:
     case MVT::i64:
+      // These can be scalar arguments or elements of an integer array type
+      // passed directly.  Clang may use those instead of "byval" aggregate
+      // types to avoid forcing arguments to memory unnecessarily.
       if (GPR_idx != Num_GPR_Regs) {
         unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
         ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
@@ -2718,6 +2742,9 @@ PPCTargetLowering::LowerFormalArguments_64SVR4(
 
     case MVT::f32:
     case MVT::f64:
+      // These can be scalar arguments or elements of a float array type
+      // passed directly.  The latter are used to implement ELFv2 homogenous
+      // float aggregates.
       if (FPR_idx != Num_FPR_Regs) {
         unsigned VReg;
 
@@ -2730,12 +2757,32 @@ PPCTargetLowering::LowerFormalArguments_64SVR4(
 
         ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
         ++FPR_idx;
+      } else if (GPR_idx != Num_GPR_Regs) {
+        // This can only ever happen in the presence of f32 array types,
+        // since otherwise we never run out of FPRs before running out
+        // of GPRs.
+        unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
+        ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
+
+        if (ObjectVT == MVT::f32) {
+          if ((ArgOffset % PtrByteSize) == (isLittleEndian ? 4 : 0))
+            ArgVal = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgVal,
+                                 DAG.getConstant(32, MVT::i32));
+          ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
+        }
+
+        ArgVal = DAG.getNode(ISD::BITCAST, dl, ObjectVT, ArgVal);
       } else {
         needsLoad = true;
-        ArgSize = PtrByteSize;
       }
 
-      ArgOffset += 8;
+      // When passing an array of floats, the array occupies consecutive
+      // space in the argument area; only round up to the next doubleword
+      // at the end of the array.  Otherwise, each float takes 8 bytes.
+      ArgSize = Flags.isInConsecutiveRegs() ? ObjSize : PtrByteSize;
+      ArgOffset += ArgSize;
+      if (Flags.isInConsecutiveRegsLast())
+        ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
       break;
     case MVT::v4f32:
     case MVT::v4i32:
@@ -2743,6 +2790,9 @@ PPCTargetLowering::LowerFormalArguments_64SVR4(
     case MVT::v16i8:
     case MVT::v2f64:
     case MVT::v2i64:
+      // These can be scalar arguments or elements of a vector array type
+      // passed directly.  The latter are used to implement ELFv2 homogenous
+      // vector aggregates.
       if (VR_idx != Num_VR_Regs) {
         unsigned VReg = (ObjectVT == MVT::v2f64 || ObjectVT == MVT::v2i64) ?
                         MF.addLiveIn(VSRH[VR_idx], &PPC::VSHRCRegClass) :
@@ -4105,12 +4155,16 @@ PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
   for (unsigned i = 0; i != NumOps; ++i) {
     ISD::ArgFlagsTy Flags = Outs[i].Flags;
     EVT ArgVT = Outs[i].VT;
+    EVT OrigVT = Outs[i].ArgVT;
 
     /* Respect alignment of argument on the stack.  */
-    unsigned Align = CalculateStackSlotAlignment(ArgVT, Flags, PtrByteSize);
+    unsigned Align =
+      CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
     NumBytes = ((NumBytes + Align - 1) / Align) * Align;
 
     NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
+    if (Flags.isInConsecutiveRegsLast())
+      NumBytes = ((NumBytes + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
   }
 
   unsigned NumBytesActuallyUsed = NumBytes;
@@ -4187,10 +4241,12 @@ PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
   for (unsigned i = 0; i != NumOps; ++i) {
     SDValue Arg = OutVals[i];
     ISD::ArgFlagsTy Flags = Outs[i].Flags;
+    EVT ArgVT = Outs[i].VT;
+    EVT OrigVT = Outs[i].ArgVT;
 
     /* Respect alignment of argument on the stack.  */
     unsigned Align =
-      CalculateStackSlotAlignment(Outs[i].VT, Flags, PtrByteSize);
+      CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
     ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
 
     /* Compute GPR index associated with argument offset.  */
@@ -4330,6 +4386,9 @@ PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
     case MVT::i1:
     case MVT::i32:
     case MVT::i64:
+      // These can be scalar arguments or elements of an integer array type
+      // passed directly.  Clang may use those instead of "byval" aggregate
+      // types to avoid forcing arguments to memory unnecessarily.
       if (GPR_idx != NumGPRs) {
         RegsToPass.push_back(std::make_pair(GPR[GPR_idx], Arg));
       } else {
@@ -4340,39 +4399,70 @@ PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
       ArgOffset += PtrByteSize;
       break;
     case MVT::f32:
-    case MVT::f64:
-      if (FPR_idx != NumFPRs) {
+    case MVT::f64: {
+      // These can be scalar arguments or elements of a float array type
+      // passed directly.  The latter are used to implement ELFv2 homogenous
+      // float aggregates.
+
+      // Named arguments go into FPRs first, and once they overflow, the
+      // remaining arguments go into GPRs and then the parameter save area.
+      // Unnamed arguments for vararg functions always go to GPRs and
+      // then the parameter save area.  For now, put all arguments to vararg
+      // routines always in both locations (FPR *and* GPR or stack slot).
+      bool NeedGPROrStack = isVarArg || FPR_idx == NumFPRs;
+
+      // First load the argument into the next available FPR.
+      if (FPR_idx != NumFPRs)
         RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
 
-        if (isVarArg) {
-          // A single float or an aggregate containing only a single float
-          // must be passed right-justified in the stack doubleword, and
-          // in the GPR, if one is available.
-          SDValue StoreOff;
-          if (Arg.getSimpleValueType().SimpleTy == MVT::f32 &&
-              !isLittleEndian) {
-            SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
-            StoreOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
-          } else
-            StoreOff = PtrOff;
-
-          SDValue Store = DAG.getStore(Chain, dl, Arg, StoreOff,
-                                       MachinePointerInfo(), false, false, 0);
-          MemOpChains.push_back(Store);
-
-          // Float varargs are always shadowed in available integer registers
-          if (GPR_idx != NumGPRs) {
-            SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
-                                       MachinePointerInfo(), false, false,
-                                       false, 0);
-            MemOpChains.push_back(Load.getValue(1));
-            RegsToPass.push_back(std::make_pair(GPR[GPR_idx], Load));
-          }
-        }
+      // Next, load the argument into GPR or stack slot if needed.
+      if (!NeedGPROrStack)
+        ;
+      else if (GPR_idx != NumGPRs) {
+        // In the non-vararg case, this can only ever happen in the
+        // presence of f32 array types, since otherwise we never run
+        // out of FPRs before running out of GPRs.
+        SDValue ArgVal;
+
+        // Double values are always passed in a single GPR.
+        if (Arg.getValueType() != MVT::f32) {
+          ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
+
+        // Non-array float values are extended and passed in a GPR.
+        } else if (!Flags.isInConsecutiveRegs()) {
+          ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
+          ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
+
+        // If we have an array of floats, we collect every odd element
+        // together with its predecessor into one GPR.
+        } else if (ArgOffset % PtrByteSize != 0) {
+          SDValue Lo, Hi;
+          Lo = DAG.getNode(ISD::BITCAST, dl, MVT::i32, OutVals[i - 1]);
+          Hi = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
+          if (!isLittleEndian)
+            std::swap(Lo, Hi);
+          ArgVal = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
+
+        // The final element, if even, goes into the first half of a GPR.
+        } else if (Flags.isInConsecutiveRegsLast()) {
+          ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
+          ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
+          if (!isLittleEndian)
+            ArgVal = DAG.getNode(ISD::SHL, dl, MVT::i64, ArgVal,
+                                 DAG.getConstant(32, MVT::i32));
+
+        // Non-final even elements are skipped; they will be handled
+        // together the with subsequent argument on the next go-around.
+        } else
+          ArgVal = SDValue();
+
+        if (ArgVal.getNode())
+          RegsToPass.push_back(std::make_pair(GPR[GPR_idx], ArgVal));
       } else {
         // Single-precision floating-point values are mapped to the
         // second (rightmost) word of the stack doubleword.
-        if (Arg.getValueType() == MVT::f32 && !isLittleEndian) {
+        if (Arg.getValueType() == MVT::f32 &&
+            !isLittleEndian && !Flags.isInConsecutiveRegs()) {
           SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
           PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
         }
@@ -4381,14 +4471,25 @@ PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
                          true, isTailCall, false, MemOpChains,
                          TailCallArguments, dl);
       }
-      ArgOffset += 8;
+      // When passing an array of floats, the array occupies consecutive
+      // space in the argument area; only round up to the next doubleword
+      // at the end of the array.  Otherwise, each float takes 8 bytes.
+      ArgOffset += (Arg.getValueType() == MVT::f32 &&
+                    Flags.isInConsecutiveRegs()) ? 4 : 8;
+      if (Flags.isInConsecutiveRegsLast())
+        ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
       break;
+    }
     case MVT::v4f32:
     case MVT::v4i32:
     case MVT::v8i16:
     case MVT::v16i8:
     case MVT::v2f64:
     case MVT::v2i64:
+      // These can be scalar arguments or elements of a vector array type
+      // passed directly.  The latter are used to implement ELFv2 homogenous
+      // vector aggregates.
+
       // For a varargs call, named arguments go into VRs or on the stack as
       // usual; unnamed arguments always go to the stack or the corresponding
       // GPRs when within range.  For now, we always put the value in both
diff --git a/lib/Target/PowerPC/PPCISelLowering.h b/lib/Target/PowerPC/PPCISelLowering.h
index df05aa51ffa..ae8c300a4ff 100644
--- a/lib/Target/PowerPC/PPCISelLowering.h
+++ b/lib/Target/PowerPC/PPCISelLowering.h
@@ -510,6 +510,20 @@ namespace llvm {
     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
                              const TargetLibraryInfo *LibInfo) const override;
 
+    /// \brief Returns true if an argument of type Ty needs to be passed in a
+    /// contiguous block of registers in calling convention CallConv.
+    bool functionArgumentNeedsConsecutiveRegisters(
+      Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override {
+      // We support any array type as "consecutive" block in the parameter
+      // save area.  The element type defines the alignment requirement and
+      // whether the argument should go in GPRs, FPRs, or VRs if available.
+      //
+      // Note that clang uses this capability both to implement the ELFv2
+      // homogeneous float/vector aggregate ABI, and to avoid having to use
+      // "byval" when passing aggregates that might fully fit in registers.
+      return Ty->isArrayTy();
+    }
+
   private:
     SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
     SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
diff --git a/test/CodeGen/PowerPC/ppc64le-aggregates.ll b/test/CodeGen/PowerPC/ppc64le-aggregates.ll
new file mode 100644
index 00000000000..9eed623baca
--- /dev/null
+++ b/test/CodeGen/PowerPC/ppc64le-aggregates.ll
@@ -0,0 +1,329 @@
+; RUN: llc < %s -march=ppc64le -mcpu=pwr8 -mattr=+altivec | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-n32:64"
+target triple = "powerpc64le-unknown-linux-gnu"
+
+;
+; Verify use of registers for float/vector aggregate return.
+;
+
+define [8 x float] @return_float([8 x float] %x) {
+entry:
+  ret [8 x float] %x
+}
+; CHECK-LABEL: @return_float
+; CHECK: %entry
+; CHECK-NEXT: blr
+
+define [8 x double] @return_double([8 x double] %x) {
+entry:
+  ret [8 x double] %x
+}
+; CHECK-LABEL: @return_double
+; CHECK: %entry
+; CHECK-NEXT: blr
+
+define [4 x ppc_fp128] @return_ppcf128([4 x ppc_fp128] %x) {
+entry:
+  ret [4 x ppc_fp128] %x
+}
+; CHECK-LABEL: @return_ppcf128
+; CHECK: %entry
+; CHECK-NEXT: blr
+
+define [8 x <4 x i32>] @return_v4i32([8 x <4 x i32>] %x) {
+entry:
+  ret [8 x <4 x i32>] %x
+}
+; CHECK-LABEL: @return_v4i32
+; CHECK: %entry
+; CHECK-NEXT: blr
+
+
+;
+; Verify amount of space taken up by aggregates in the parameter save area.
+;
+
+define i64 @callee_float([7 x float] %a, [7 x float] %b, i64 %c) {
+entry:
+  ret i64 %c
+}
+; CHECK-LABEL: @callee_float
+; CHECK: ld 3, 96(1)
+; CHECK: blr
+
+define void @caller_float(i64 %x, [7 x float] %y) {
+entry:
+  tail call void @test_float([7 x float] %y, [7 x float] %y, i64 %x)
+  ret void
+}
+; CHECK-LABEL: @caller_float
+; CHECK: std 3, 96(1)
+; CHECK: bl test_float
+
+declare void @test_float([7 x float], [7 x float], i64)
+
+define i64 @callee_double(i64 %a, [7 x double] %b, i64 %c) {
+entry:
+  ret i64 %c
+}
+; CHECK-LABEL: @callee_double
+; CHECK: ld 3, 96(1)
+; CHECK: blr
+
+define void @caller_double(i64 %x, [7 x double] %y) {
+entry:
+  tail call void @test_double(i64 %x, [7 x double] %y, i64 %x)
+  ret void
+}
+; CHECK-LABEL: @caller_double
+; CHECK: std 3, 96(1)
+; CHECK: bl test_double
+
+declare void @test_double(i64, [7 x double], i64)
+
+define i64 @callee_ppcf128(i64 %a, [4 x ppc_fp128] %b, i64 %c) {
+entry:
+  ret i64 %c
+}
+; CHECK-LABEL: @callee_ppcf128
+; CHECK: ld 3, 104(1)
+; CHECK: blr
+
+define void @caller_ppcf128(i64 %x, [4 x ppc_fp128] %y) {
+entry:
+  tail call void @test_ppcf128(i64 %x, [4 x ppc_fp128] %y, i64 %x)
+  ret void
+}
+; CHECK-LABEL: @caller_ppcf128
+; CHECK: std 3, 104(1)
+; CHECK: bl test_ppcf128
+
+declare void @test_ppcf128(i64, [4 x ppc_fp128], i64)
+
+define i64 @callee_i64(i64 %a, [7 x i64] %b, i64 %c) {
+entry:
+  ret i64 %c
+}
+; CHECK-LABEL: @callee_i64
+; CHECK: ld 3, 96(1)
+; CHECK: blr
+
+define void @caller_i64(i64 %x, [7 x i64] %y) {
+entry:
+  tail call void @test_i64(i64 %x, [7 x i64] %y, i64 %x)
+  ret void
+}
+; CHECK-LABEL: @caller_i64
+; CHECK: std 3, 96(1)
+; CHECK: bl test_i64
+
+declare void @test_i64(i64, [7 x i64], i64)
+
+define i64 @callee_i128(i64 %a, [4 x i128] %b, i64 %c) {
+entry:
+  ret i64 %c
+}
+; CHECK-LABEL: @callee_i128
+; CHECK: ld 3, 112(1)
+; CHECK: blr
+
+define void @caller_i128(i64 %x, [4 x i128] %y) {
+entry:
+  tail call void @test_i128(i64 %x, [4 x i128] %y, i64 %x)
+  ret void
+}
+; CHECK-LABEL: @caller_i128
+; CHECK: std 3, 112(1)
+; CHECK: bl test_i128
+
+declare void @test_i128(i64, [4 x i128], i64)
+
+define i64 @callee_v4i32(i64 %a, [4 x <4 x i32>] %b, i64 %c) {
+entry:
+  ret i64 %c
+}
+; CHECK-LABEL: @callee_v4i32
+; CHECK: ld 3, 112(1)
+; CHECK: blr
+
+define void @caller_v4i32(i64 %x, [4 x <4 x i32>] %y) {
+entry:
+  tail call void @test_v4i32(i64 %x, [4 x <4 x i32>] %y, i64 %x)
+  ret void
+}
+; CHECK-LABEL: @caller_v4i32
+; CHECK: std 3, 112(1)
+; CHECK: bl test_v4i32
+
+declare void @test_v4i32(i64, [4 x <4 x i32>], i64)
+
+
+;
+; Verify handling of floating point arguments in GPRs
+;
+
+%struct.float8 = type { [8 x float] }
+%struct.float5 = type { [5 x float] }
+%struct.float2 = type { [2 x float] }
+
+@g8 = common global %struct.float8 zeroinitializer, align 4
+@g5 = common global %struct.float5 zeroinitializer, align 4
+@g2 = common global %struct.float2 zeroinitializer, align 4
+
+define float @callee0([7 x float] %a, [7 x float] %b) {
+entry:
+  %b.extract = extractvalue [7 x float] %b, 6
+  ret float %b.extract
+}
+; CHECK-LABEL: @callee0
+; CHECK: stw 10, [[OFF:.*]](1)
+; CHECK: lfs 1, [[OFF]](1)
+; CHECK: blr
+
+define void @caller0([7 x float] %a) {
+entry:
+  tail call void @test0([7 x float] %a, [7 x float] %a)
+  ret void
+}
+; CHECK-LABEL: @caller0
+; CHECK-DAG: fmr 8, 1
+; CHECK-DAG: fmr 9, 2
+; CHECK-DAG: fmr 10, 3
+; CHECK-DAG: fmr 11, 4
+; CHECK-DAG: fmr 12, 5
+; CHECK-DAG: fmr 13, 6
+; CHECK-DAG: stfs 7, [[OFF:[0-9]+]](1)
+; CHECK-DAG: lwz 10, [[OFF]](1)
+; CHECK: bl test0
+
+declare void @test0([7 x float], [7 x float])
+
+define float @callee1([8 x float] %a, [8 x float] %b) {
+entry:
+  %b.extract = extractvalue [8 x float] %b, 7
+  ret float %b.extract
+}
+; CHECK-LABEL: @callee1
+; CHECK: rldicl [[REG:[0-9]+]], 10, 32, 32
+; CHECK: stw [[REG]], [[OFF:.*]](1)
+; CHECK: lfs 1, [[OFF]](1)
+; CHECK: blr
+
+define void @caller1([8 x float] %a) {
+entry:
+  tail call void @test1([8 x float] %a, [8 x float] %a)
+  ret void
+}
+; CHECK-LABEL: @caller1
+; CHECK-DAG: fmr 9, 1
+; CHECK-DAG: fmr 10, 2
+; CHECK-DAG: fmr 11, 3
+; CHECK-DAG: fmr 12, 4
+; CHECK-DAG: fmr 13, 5
+; CHECK-DAG: stfs 5, [[OFF0:[0-9]+]](1)
+; CHECK-DAG: stfs 6, [[OFF1:[0-9]+]](1)
+; CHECK-DAG: stfs 7, [[OFF2:[0-9]+]](1)
+; CHECK-DAG: stfs 8, [[OFF3:[0-9]+]](1)
+; CHECK-DAG: lwz [[REG0:[0-9]+]], [[OFF0]](1)
+; CHECK-DAG: lwz [[REG1:[0-9]+]], [[OFF1]](1)
+; CHECK-DAG: lwz [[REG2:[0-9]+]], [[OFF2]](1)
+; CHECK-DAG: lwz [[REG3:[0-9]+]], [[OFF3]](1)
+; CHECK-DAG: sldi [[REG1]], [[REG1]], 32
+; CHECK-DAG: sldi [[REG3]], [[REG3]], 32
+; CHECK-DAG: or 9, [[REG0]], [[REG1]]
+; CHECK-DAG: or 10, [[REG2]], [[REG3]]
+; CHECK: bl test1
+
+declare void @test1([8 x float], [8 x float])
+
+define float @callee2([8 x float] %a, [5 x float] %b, [2 x float] %c) {
+entry:
+  %c.extract = extractvalue [2 x float] %c, 1
+  ret float %c.extract
+}
+; CHECK-LABEL: @callee2
+; CHECK: rldicl [[REG:[0-9]+]], 10, 32, 32
+; CHECK: stw [[REG]], [[OFF:.*]](1)
+; CHECK: lfs 1, [[OFF]](1)
+; CHECK: blr
+
+define void @caller2() {
+entry:
+  %0 = load [8 x float]* getelementptr inbounds (%struct.float8* @g8, i64 0, i32 0), align 4
+  %1 = load [5 x float]* getelementptr inbounds (%struct.float5* @g5, i64 0, i32 0), align 4
+  %2 = load [2 x float]* getelementptr inbounds (%struct.float2* @g2, i64 0, i32 0), align 4
+  tail call void @test2([8 x float] %0, [5 x float] %1, [2 x float] %2)
+  ret void
+}
+; CHECK-LABEL: @caller2
+; CHECK: ld [[REG:[0-9]+]], .LC
+; CHECK-DAG: lfs 1, 0([[REG]])
+; CHECK-DAG: lfs 2, 4([[REG]])
+; CHECK-DAG: lfs 3, 8([[REG]])
+; CHECK-DAG: lfs 4, 12([[REG]])
+; CHECK-DAG: lfs 5, 16([[REG]])
+; CHECK-DAG: lfs 6, 20([[REG]])
+; CHECK-DAG: lfs 7, 24([[REG]])
+; CHECK-DAG: lfs 8, 28([[REG]])
+; CHECK: ld [[REG:[0-9]+]], .LC
+; CHECK-DAG: lfs 9, 0([[REG]])
+; CHECK-DAG: lfs 10, 4([[REG]])
+; CHECK-DAG: lfs 11, 8([[REG]])
+; CHECK-DAG: lfs 12, 12([[REG]])
+; CHECK-DAG: lfs 13, 16([[REG]])
+; CHECK: ld [[REG:[0-9]+]], .LC
+; CHECK-DAG: lwz [[REG0:[0-9]+]], 0([[REG]])
+; CHECK-DAG: lwz [[REG1:[0-9]+]], 4([[REG]])
+; CHECK-DAG: sldi [[REG1]], [[REG1]], 32
+; CHECK-DAG: or 10, [[REG0]], [[REG1]]
+; CHECK: bl test2
+
+declare void @test2([8 x float], [5 x float], [2 x float])
+
+define double @callee3([8 x float] %a, [5 x float] %b, double %c) {
+entry:
+  ret double %c
+}
+; CHECK-LABEL: @callee3
+; CHECK: std 10, [[OFF:.*]](1)
+; CHECK: lfd 1, [[OFF]](1)
+; CHECK: blr
+
+define void @caller3(double %d) {
+entry:
+  %0 = load [8 x float]* getelementptr inbounds (%struct.float8* @g8, i64 0, i32 0), align 4
+  %1 = load [5 x float]* getelementptr inbounds (%struct.float5* @g5, i64 0, i32 0), align 4
+  tail call void @test3([8 x float] %0, [5 x float] %1, double %d)
+  ret void
+}
+; CHECK-LABEL: @caller3
+; CHECK: stfd 1, [[OFF:.*]](1)
+; CHECK: ld 10, [[OFF]](1)
+; CHECK: bl test3
+
+declare void @test3([8 x float], [5 x float], double)
+
+define float @callee4([8 x float] %a, [5 x float] %b, float %c) {
+entry:
+  ret float %c
+}
+; CHECK-LABEL: @callee4
+; CHECK: stw 10, [[OFF:.*]](1)
+; CHECK: lfs 1, [[OFF]](1)
+; CHECK: blr
+
+define void @caller4(float %f) {
+entry:
+  %0 = load [8 x float]* getelementptr inbounds (%struct.float8* @g8, i64 0, i32 0), align 4
+  %1 = load [5 x float]* getelementptr inbounds (%struct.float5* @g5, i64 0, i32 0), align 4
+  tail call void @test4([8 x float] %0, [5 x float] %1, float %f)
+  ret void
+}
+; CHECK-LABEL: @caller4
+; CHECK: stfs 1, [[OFF:.*]](1)
+; CHECK: lwz 10, [[OFF]](1)
+; CHECK: bl test4
+
+declare void @test4([8 x float], [5 x float], float)
+
diff --git a/test/CodeGen/PowerPC/varargs-struct-float.ll b/test/CodeGen/PowerPC/varargs-struct-float.ll
index fb1835f580b..0fd9fc50892 100644
--- a/test/CodeGen/PowerPC/varargs-struct-float.ll
+++ b/test/CodeGen/PowerPC/varargs-struct-float.ll
@@ -16,8 +16,8 @@ entry:
   ret void
 }
 
-; CHECK: stfs {{[0-9]+}}, 60(1)
-; CHECK: ld 4, 56(1)
+; CHECK: stfs {{[0-9]+}}, 116(1)
+; CHECK: lwz 4, 116(1)
 ; CHECK: bl
 
 declare void @testvaSf1(i32, ...)