Pull in r199028 from upstream llvm trunk (by Jakob Stoklund Olesen):

  The SPARCv9 ABI returns a float in %f0.

  This is different from the argument passing convention which puts the
  first float argument in %f1.

  With this patch, all returned floats are treated as if the 'inreg' flag
  were set. This means multiple float return values get packed in %f0,
  %f1, %f2, ...

  Note that when returning a struct in registers, clang will set the
  'inreg' flag on the return value, so that behavior is unchanged. This
  also happens when returning a float _Complex.

Introduced here: http://svnweb.freebsd.org/changeset/base/262261

Index: test/CodeGen/SPARC/64abi.ll
===================================================================
--- test/CodeGen/SPARC/64abi.ll
+++ test/CodeGen/SPARC/64abi.ll
@@ -180,7 +180,7 @@ define void @call_inreg_fi(i32* %p, i32 %i1, float
 }
 
 ; CHECK: inreg_ff
-; CHECK: fsubs %f0, %f1, %f1
+; CHECK: fsubs %f0, %f1, %f0
 define float @inreg_ff(float inreg %a0,   ; %f0
                        float inreg %a1) { ; %f1
   %rv = fsub float %a0, %a1
@@ -262,10 +262,10 @@ define void @call_ret_i64_pair(i64* %i0) {
   ret void
 }
 
-; This is not a C struct, each member uses 8 bytes.
+; This is not a C struct, the i32 member uses 8 bytes, but the float only 4.
 ; CHECK: ret_i32_float_pair
 ; CHECK: ld [%i2], %i0
-; CHECK: ld [%i3], %f3
+; CHECK: ld [%i3], %f2
 define { i32, float } @ret_i32_float_pair(i32 %a0, i32 %a1,
                                           i32* %p, float* %q) {
   %r1 = load i32* %p
@@ -279,7 +279,7 @@ define { i32, float } @ret_i32_float_pair(i32 %a0,
 ; CHECK: call_ret_i32_float_pair
 ; CHECK: call ret_i32_float_pair
 ; CHECK: st %o0, [%i0]
-; CHECK: st %f3, [%i1]
+; CHECK: st %f2, [%i1]
 define void @call_ret_i32_float_pair(i32* %i0, float* %i1) {
   %rv = call { i32, float } @ret_i32_float_pair(i32 undef, i32 undef,
                                                 i32* undef, float* undef)
Index: test/CodeGen/SPARC/constpool.ll
===================================================================
--- test/CodeGen/SPARC/constpool.ll
+++ test/CodeGen/SPARC/constpool.ll
@@ -21,7 +21,7 @@ entry:
 ; abs44: add %[[R1]], %m44(.LCPI0_0), %[[R2:[gilo][0-7]]]
 ; abs44: sllx %[[R2]], 12, %[[R3:[gilo][0-7]]]
 ; abs44: retl
-; abs44: ld [%[[R3]]+%l44(.LCPI0_0)], %f1
+; abs44: ld [%[[R3]]+%l44(.LCPI0_0)], %f0
 
 
 ; abs64: floatCP
@@ -31,7 +31,7 @@ entry:
 ; abs64: add %[[R3]], %hm(.LCPI0_0), %[[R4:[gilo][0-7]]]
 ; abs64: sllx %[[R4]], 32, %[[R5:[gilo][0-7]]]
 ; abs64: retl
-; abs64: ld [%[[R5]]+%[[R2]]], %f1
+; abs64: ld [%[[R5]]+%[[R2]]], %f0
 
 
 ; v8pic32: floatCP
@@ -50,7 +50,7 @@ entry:
 ; v9pic32: sethi %hi(.LCPI0_0), %[[R1:[gilo][0-7]]]
 ; v9pic32: add %[[R1]], %lo(.LCPI0_0), %[[Goffs:[gilo][0-7]]]
 ; v9pic32: ldx [%[[GOT:[gilo][0-7]]]+%[[Goffs]]], %[[Gaddr:[gilo][0-7]]]
-; v9pic32: ld [%[[Gaddr]]], %f1
+; v9pic32: ld [%[[Gaddr]]], %f0
 ; v9pic32: ret
 ; v9pic32: restore
 
Index: test/CodeGen/SPARC/64cond.ll
===================================================================
--- test/CodeGen/SPARC/64cond.ll
+++ test/CodeGen/SPARC/64cond.ll
@@ -80,7 +80,7 @@ entry:
 ; CHECK: selectf32_xcc
 ; CHECK: cmp %i0, %i1
 ; CHECK: fmovsg %xcc, %f5, %f7
-; CHECK: fmovs %f7, %f1
+; CHECK: fmovs %f7, %f0
 define float @selectf32_xcc(i64 %x, i64 %y, float %a, float %b) {
 entry:
   %tobool = icmp sgt i64 %x, %y
Index: lib/Target/Sparc/SparcISelLowering.cpp
===================================================================
--- lib/Target/Sparc/SparcISelLowering.cpp
+++ lib/Target/Sparc/SparcISelLowering.cpp
@@ -254,7 +254,7 @@ SparcTargetLowering::LowerReturn_64(SDValue Chain,
                  DAG.getTarget(), RVLocs, *DAG.getContext());
 
   // Analyze return values.
-  CCInfo.AnalyzeReturn(Outs, CC_Sparc64);
+  CCInfo.AnalyzeReturn(Outs, RetCC_Sparc64);
 
   SDValue Flag;
   SmallVector<SDValue, 4> RetOps(1, Chain);
@@ -1258,7 +1258,7 @@ SparcTargetLowering::LowerCall_64(TargetLowering::
   if (CLI.Ins.size() == 1 && CLI.Ins[0].VT == MVT::f32 && CLI.CS == 0)
     CLI.Ins[0].Flags.setInReg();
 
-  RVInfo.AnalyzeCallResult(CLI.Ins, CC_Sparc64);
+  RVInfo.AnalyzeCallResult(CLI.Ins, RetCC_Sparc64);
 
   // Copy all of the result registers out of their specified physreg.
   for (unsigned i = 0; i != RVLocs.size(); ++i) {
Index: lib/Target/Sparc/SparcCallingConv.td
===================================================================
--- lib/Target/Sparc/SparcCallingConv.td
+++ lib/Target/Sparc/SparcCallingConv.td
@@ -103,7 +103,7 @@ def RetCC_Sparc32 : CallingConv<[
 // Function return values are passed exactly like function arguments, except a
 // struct up to 32 bytes in size can be returned in registers.
 
-// Function arguments AND return values.
+// Function arguments AND most return values.
 def CC_Sparc64 : CallingConv<[
   // The frontend uses the inreg flag to indicate i32 and float arguments from
   // structs. These arguments are not promoted to 64 bits, but they can still
@@ -118,6 +118,15 @@ def CC_Sparc64 : CallingConv<[
   CCCustom<"CC_Sparc64_Full">
 ]>;
 
+def RetCC_Sparc64 : CallingConv<[
+  // A single f32 return value always goes in %f0. The ABI doesn't specify what
+  // happens to multiple f32 return values outside a struct.
+  CCIfType<[f32], CCCustom<"CC_Sparc64_Half">>,
+
+  // Otherwise, return values are passed exactly like arguments.
+  CCDelegateTo<CC_Sparc64>
+]>;
+
 // Callee-saved registers are handled by the register window mechanism.
 def CSR : CalleeSavedRegs<(add)> {
   let OtherPreserved = (add (sequence "I%u", 0, 7),