Special handling for MMX values being passed in either GPR64 or lower 64-bits of...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50289 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 38949601f6abe910008fcaf3baaa358a71bae6cf..21178ef78d0c5d1c853284a13b473be3dfdaa1b2 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -41,6 +41,9 @@
 #include "llvm/ADT/StringExtras.h"
 using namespace llvm;
 
+// Forward declarations.
+static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG);
+
 X86TargetLowering::X86TargetLowering(TargetMachine &TM)
   : TargetLowering(TM) {
   Subtarget = &TM.getSubtarget<X86Subtarget>();
@@ -1547,8 +1550,7 @@ SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
   SDOperand StackPtr;
   bool containsTailCallByValArg = false;
   SmallVector<std::pair<unsigned, unsigned>, 8> TailCallByValClobberedVRegs;
-  SmallVector<MVT::ValueType, 8> TailCallByValClobberedVRegTypes;
-  
+  SmallVector<MVT::ValueType, 8> TailCallByValClobberedVRegTypes;  
 
   // Walk the register/memloc assignments, inserting copies/loads.  For tail
   // calls, remember all arguments for later special lowering.
@@ -1574,6 +1576,30 @@ SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
     }
     
     if (VA.isRegLoc()) {
+      if (Is64Bit) {
+        MVT::ValueType RegVT = VA.getLocVT();
+        if (MVT::isVector(RegVT) && MVT::getSizeInBits(RegVT) == 64)
+          switch (VA.getLocReg()) {
+          default:
+            break;
+          case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
+          case X86::R8: {
+            // Special case: passing MMX values in GPR registers.
+            Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
+            break;
+          }
+          case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
+          case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
+            // Special case: passing MMX values in XMM registers.
+            Arg = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Arg);
+            Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Arg);
+            Arg = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v2i64,
+                              DAG.getNode(ISD::UNDEF, MVT::v2i64), Arg,
+                              getMOVLMask(2, DAG));
+            break;
+          }
+          }
+      }
       RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
     } else {
       if (!IsTailCall || (IsTailCall && isByVal)) {

diff --git a/lib/Target/X86/X86InstrMMX.td b/lib/Target/X86/X86InstrMMX.td
index 6d9c02ba88963ecf125c8ae08f7be0f91130fdd8..600b1e6e76de25de6c371fc3ad90f3e4f422241b 100644 (file)
--- a/lib/Target/X86/X86InstrMMX.td
+++ b/lib/Target/X86/X86InstrMMX.td
@@ -637,3 +637,7 @@ def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v4i16 immAllOnesV_bc))),
 def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v8i8  immAllOnesV_bc))),
                   (load addr:$src2))),
           (MMX_PANDNrm VR64:$src1, addr:$src2)>;
+
+// Move MMX to lower 64-bit of XMM
+def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert VR64:$src)))),
+          (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;

diff --git a/test/CodeGen/X86/mmx-arg-passing2.ll b/test/CodeGen/X86/mmx-arg-passing2.ll
new file mode 100644 (file)
index 0000000..c30ca1b
--- /dev/null
+++ b/test/CodeGen/X86/mmx-arg-passing2.ll
@@ -0,0 +1,25 @@
+; RUN: llvm-as < %s | llc -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | grep movq2dq
+; RUN: llvm-as < %s | llc -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | grep movd | count 1
+; RUN: llvm-as < %s | llc -mtriple=x86_64-apple-darwin -mattr=+mmx,+sse2 | grep movq | count 4
+
+@g_v8qi = external global <8 x i8>
+
+define void @t1() nounwind  {
+       %tmp3 = load <8 x i8>* @g_v8qi, align 8
+       %tmp4 = tail call i32 (...)* @pass_v8qi( <8 x i8> %tmp3 ) nounwind
+       ret void
+}
+
+define void @t2(<8 x i8> %v1, <8 x i8> %v2) nounwind  {
+       %tmp3 = add <8 x i8> %v1, %v2
+       %tmp4 = tail call i32 (...)* @pass_v8qi( <8 x i8> %tmp3 ) nounwind
+       ret void
+}
+
+define void @t3() nounwind  {
+       call void @pass_v1di( <1 x i64> zeroinitializer )
+        ret void
+}
+
+declare i32 @pass_v8qi(...)
+declare void @pass_v1di(<1 x i64>)