Index: lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- lib/Target/X86/X86ISelLowering.cpp
+++ lib/Target/X86/X86ISelLowering.cpp
@@ -28627,10 +28627,11 @@
   EVT VT = N->getValueType(0);
   EVT SrcVT = N0.getValueType();
 
-  // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
-  // special and don't usually play with other vector types, it's better to
-  // handle them early to be sure we emit efficient code by avoiding
-  // store-load conversions.
+  // Since MMX types are special and don't usually play with other vector types,
+  // it's better to handle them early to be sure we emit efficient code by
+  // avoiding store-load conversions.
+
+  // Detect bitcasts between i32 to x86mmx low word.
   if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
       SrcVT == MVT::v2i32 && isNullConstant(N0.getOperand(1))) {
     SDValue N00 = N0->getOperand(0);
@@ -28638,6 +28639,14 @@
       return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
   }
 
+  // Detect bitcasts between v2i64/v2f64 extraction to x86mmx.
+  if (VT == MVT::x86mmx && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+      isNullConstant(N0.getOperand(1))) {
+    SDValue N00 = N0->getOperand(0);
+    if (N00.getValueType().is128BitVector())
+      return DAG.getNode(X86ISD::MOVDQ2Q, SDLoc(N00), VT, N00);
+  }
+
   // Convert a bitcasted integer logic operation that has one bitcasted
   // floating-point operand into a floating-point logic operation. This may
   // create a load of a constant, but that is cheaper than materializing the
Index: lib/Target/X86/X86InstrMMX.td
===================================================================
--- lib/Target/X86/X86InstrMMX.td
+++ lib/Target/X86/X86InstrMMX.td
@@ -670,6 +670,14 @@
           (MMX_MOVQ2FR64rr VR64:$src)>;
 def : Pat<(x86mmx (bitconvert (f64 FR64:$src))),
           (MMX_MOVFR642Qrr FR64:$src)>;
+def : Pat<(x86mmx (MMX_X86movdq2q (bc_v2i64 (v4i32 (int_x86_sse2_cvtps2dq VR128:$src))))),
+          (MMX_CVTPS2PIirr VR128:$src)>;
+def : Pat<(x86mmx (MMX_X86movdq2q (bc_v2i64 (v4i32 (fp_to_sint (v4f32 VR128:$src)))))),
+          (MMX_CVTTPS2PIirr VR128:$src)>;
+def : Pat<(x86mmx (MMX_X86movdq2q (bc_v2i64 (v4i32 (X86cvtp2Int (v2f64 VR128:$src)))))),
+          (MMX_CVTPD2PIirr VR128:$src)>;
+def : Pat<(x86mmx (MMX_X86movdq2q (bc_v2i64 (v4i32 (X86cvttp2si (v2f64 VR128:$src)))))),
+          (MMX_CVTTPD2PIirr VR128:$src)>;
 }
 
 
Index: test/CodeGen/X86/mmx-cvt.ll
===================================================================
--- test/CodeGen/X86/mmx-cvt.ll
+++ test/CodeGen/X86/mmx-cvt.ll
@@ -2,7 +2,7 @@
 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+mmx,+sse2 | FileCheck %s --check-prefix=X86
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+mmx,+sse2 | FileCheck %s --check-prefix=X64
 
-; FIXME: If we are transferring XMM conversion results to MMX registers we could use the MMX equivalents
+; If we are transferring XMM conversion results to MMX registers we could use the MMX equivalents
 ; (CVTPD2PI/CVTTPD2PI + CVTPS2PI/CVTTPS2PI) without affecting rounding/exceptions etc.
 
 define void @cvt_v2f64_v2i32(<2 x double>, <1 x i64>*) nounwind {
@@ -11,13 +11,9 @@
 ; X86-NEXT:    pushl %ebp
 ; X86-NEXT:    movl %esp, %ebp
 ; X86-NEXT:    andl $-8, %esp
-; X86-NEXT:    subl $16, %esp
+; X86-NEXT:    subl $8, %esp
 ; X86-NEXT:    movl 8(%ebp), %eax
-; X86-NEXT:    cvtpd2dq %xmm0, %xmm0
-; X86-NEXT:    movd %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; X86-NEXT:    movd %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    movq {{[0-9]+}}(%esp), %mm0
+; X86-NEXT:    cvtpd2pi %xmm0, %mm0
 ; X86-NEXT:    paddd %mm0, %mm0
 ; X86-NEXT:    movq %mm0, (%esp)
 ; X86-NEXT:    movl (%esp), %ecx
@@ -30,8 +26,7 @@
 ;
 ; X64-LABEL: cvt_v2f64_v2i32:
 ; X64:       # BB#0:
-; X64-NEXT:    cvtpd2dq %xmm0, %xmm0
-; X64-NEXT:    movdq2q %xmm0, %mm0
+; X64-NEXT:    cvtpd2pi %xmm0, %mm0
 ; X64-NEXT:    paddd %mm0, %mm0
 ; X64-NEXT:    movq %mm0, (%rdi)
 ; X64-NEXT:    retq
@@ -52,13 +47,9 @@
 ; X86-NEXT:    pushl %ebp
 ; X86-NEXT:    movl %esp, %ebp
 ; X86-NEXT:    andl $-8, %esp
-; X86-NEXT:    subl $16, %esp
+; X86-NEXT:    subl $8, %esp
 ; X86-NEXT:    movl 8(%ebp), %eax
-; X86-NEXT:    cvttpd2dq %xmm0, %xmm0
-; X86-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; X86-NEXT:    movd %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    movq {{[0-9]+}}(%esp), %mm0
+; X86-NEXT:    cvttpd2pi %xmm0, %mm0
 ; X86-NEXT:    paddd %mm0, %mm0
 ; X86-NEXT:    movq %mm0, (%esp)
 ; X86-NEXT:    movl (%esp), %ecx
@@ -71,8 +62,7 @@
 ;
 ; X64-LABEL: cvtt_v2f64_v2i32:
 ; X64:       # BB#0:
-; X64-NEXT:    cvttpd2dq %xmm0, %xmm0
-; X64-NEXT:    movdq2q %xmm0, %mm0
+; X64-NEXT:    cvttpd2pi %xmm0, %mm0
 ; X64-NEXT:    paddd %mm0, %mm0
 ; X64-NEXT:    movq %mm0, (%rdi)
 ; X64-NEXT:    retq
@@ -131,13 +121,9 @@
 ; X86-NEXT:    pushl %ebp
 ; X86-NEXT:    movl %esp, %ebp
 ; X86-NEXT:    andl $-8, %esp
-; X86-NEXT:    subl $16, %esp
+; X86-NEXT:    subl $8, %esp
 ; X86-NEXT:    movl 8(%ebp), %eax
-; X86-NEXT:    cvtps2dq %xmm0, %xmm0
-; X86-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; X86-NEXT:    movd %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    movq {{[0-9]+}}(%esp), %mm0
+; X86-NEXT:    cvtps2pi %xmm0, %mm0
 ; X86-NEXT:    paddd %mm0, %mm0
 ; X86-NEXT:    movq %mm0, (%esp)
 ; X86-NEXT:    movl (%esp), %ecx
@@ -150,8 +136,7 @@
 ;
 ; X64-LABEL: cvt_v2f32_v2i32:
 ; X64:       # BB#0:
-; X64-NEXT:    cvtps2dq %xmm0, %xmm0
-; X64-NEXT:    movdq2q %xmm0, %mm0
+; X64-NEXT:    cvtps2pi %xmm0, %mm0
 ; X64-NEXT:    paddd %mm0, %mm0
 ; X64-NEXT:    movq %mm0, (%rdi)
 ; X64-NEXT:    retq
@@ -172,13 +157,9 @@
 ; X86-NEXT:    pushl %ebp
 ; X86-NEXT:    movl %esp, %ebp
 ; X86-NEXT:    andl $-8, %esp
-; X86-NEXT:    subl $16, %esp
+; X86-NEXT:    subl $8, %esp
 ; X86-NEXT:    movl 8(%ebp), %eax
-; X86-NEXT:    cvttps2dq %xmm0, %xmm0
-; X86-NEXT:    movd %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; X86-NEXT:    movd %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    movq {{[0-9]+}}(%esp), %mm0
+; X86-NEXT:    cvttps2pi %xmm0, %mm0
 ; X86-NEXT:    paddd %mm0, %mm0
 ; X86-NEXT:    movq %mm0, (%esp)
 ; X86-NEXT:    movl (%esp), %ecx
@@ -191,8 +172,7 @@
 ;
 ; X64-LABEL: cvtt_v2f32_v2i32:
 ; X64:       # BB#0:
-; X64-NEXT:    cvttps2dq %xmm0, %xmm0
-; X64-NEXT:    movdq2q %xmm0, %mm0
+; X64-NEXT:    cvttps2pi %xmm0, %mm0
 ; X64-NEXT:    paddd %mm0, %mm0
 ; X64-NEXT:    movq %mm0, (%rdi)
 ; X64-NEXT:    retq
@@ -213,13 +193,9 @@
 ; X86-NEXT:    pushl %ebp
 ; X86-NEXT:    movl %esp, %ebp
 ; X86-NEXT:    andl $-8, %esp
-; X86-NEXT:    subl $16, %esp
+; X86-NEXT:    subl $8, %esp
 ; X86-NEXT:    movl 8(%ebp), %eax
-; X86-NEXT:    cvttps2dq %xmm0, %xmm0
-; X86-NEXT:    movd %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; X86-NEXT:    movd %xmm0, {{[0-9]+}}(%esp)
-; X86-NEXT:    movq {{[0-9]+}}(%esp), %mm0
+; X86-NEXT:    cvttps2pi %xmm0, %mm0
 ; X86-NEXT:    paddd %mm0, %mm0
 ; X86-NEXT:    movq %mm0, (%esp)
 ; X86-NEXT:    movl (%esp), %ecx
@@ -232,8 +208,7 @@
 ;
 ; X64-LABEL: fptosi_v2f32_v2i32:
 ; X64:       # BB#0:
-; X64-NEXT:    cvttps2dq %xmm0, %xmm0
-; X64-NEXT:    movdq2q %xmm0, %mm0
+; X64-NEXT:    cvttps2pi %xmm0, %mm0
 ; X64-NEXT:    paddd %mm0, %mm0
 ; X64-NEXT:    movq %mm0, (%rdi)
 ; X64-NEXT:    retq