[PATCH] [AVX512] Bring back vector-shuffle lowering support through broadcasts

[Chandler Carruth]

Sorry for the long delay...


================
Comment at: lib/Target/X86/X86ISelLowering.cpp:10235
@@ -10234,3 +10244,7 @@
   assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
-  assert(Subtarget->hasDQI() && "We can only lower v8i64 with AVX-512-DQI");
 
----------------
(ignore this)

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:10313-10319
@@ -10312,9 +10329,5 @@
   case MVT::v8i64:
-    if (Subtarget->hasDQI())
-      return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
-    break;
+    return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
   case MVT::v16i32:
-    if (Subtarget->hasDQI())
-      return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
-    break;
   case MVT::v32i16:
----------------
These changes don't make sense to me. My understanding was that DQI was needed to shuffle integers.

http://reviews.llvm.org/D5885

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/