[X86] Add a DAG combine to turn v16i16->v16i8 VTRUNCUS+store into a saturating truncating store.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@374509 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/X86/X86ISelLowering.cpp

@@ -40448,6 +40448,19 @@ static SDValue combineStore(SDNode *N, SelectionDAG &DAG,
                              MVT::v16i8, St->getMemOperand());
   }
 
+  // Try to fold a vpmovuswb 256->128 into a truncating store.
+  // FIXME: Generalize this to other types.
+  // FIXME: Do the same for signed saturation.
+  if (!St->isTruncatingStore() && VT == MVT::v16i8 &&
+      St->getValue().getOpcode() == X86ISD::VTRUNCUS &&
+      St->getValue().getOperand(0).getValueType() == MVT::v16i16 &&
+      TLI.isTruncStoreLegal(MVT::v16i16, MVT::v16i8) &&
+      St->getValue().hasOneUse()) {
+    return EmitTruncSStore(false /* Unsigned saturation */, St->getChain(),
+                           dl, St->getValue().getOperand(0), St->getBasePtr(),
+                           MVT::v16i8, St->getMemOperand(), DAG);
+  }
+
   // Optimize trunc store (of multiple scalars) to shuffle and store.
   // First, pack all of the elements in one place. Next, store to memory
   // in fewer chunks.

test/CodeGen/X86/min-legal-vector-width.ll

@@ -1104,8 +1104,7 @@ define void @trunc_packus_v16i32_v16i8_store(<16 x i32>* %p, <16 x i8>* %q) "min
 ; CHECK-NEXT:    vmovdqa (%rdi), %ymm0
 ; CHECK-NEXT:    vpackusdw 32(%rdi), %ymm0, %ymm0
 ; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; CHECK-NEXT:    vpmovuswb %ymm0, %xmm0
-; CHECK-NEXT:    vmovdqa %xmm0, (%rsi)
+; CHECK-NEXT:    vpmovuswb %ymm0, (%rsi)
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %a = load <16 x i32>, <16 x i32>* %p