Reapply last harmless part of r112934, the pattern fragment to match X86Unpcklpd

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@113009 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index d9c119c43659b1663fa23dfb9d9d14f0570b197a..07cc0a7903fb3ea3434653b371e83bf9eed2afe9 100644 (file)
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -5819,6 +5819,13 @@ def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
 def : Pat<(v2f64 (X86Movlhpd VR128:$src1,
                     (scalar_to_vector (loadf64 addr:$src2)))),
           (MOVHPDrm VR128:$src1, addr:$src2)>;
+// FIXME: Instead of X86Unpcklpd, there should be a X86Movlhpd here, the problem
+// is during lowering, where it's not possible to recognize the load fold cause
+// it has two uses through a bitcast. One use disappears at isel time and the
+// fold opportunity reappears.
+def : Pat<(v2f64 (X86Unpcklpd VR128:$src1,
+                    (scalar_to_vector (loadf64 addr:$src2)))),
+          (MOVHPDrm VR128:$src1, addr:$src2)>;
 
 // Shuffle with MOVSS
 def : Pat<(v4f32 (X86Movss VR128:$src1, (scalar_to_vector FR32:$src2))),