From ccc85ac85569c55b6915180d8b722812b4a226b6 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Fri, 11 Oct 2019 04:16:49 +0000
Subject: [PATCH] [X86] Add a DAG combine to turn v16i16->v16i8 VTRUNCUS+store
 into a saturating truncating store.

llvm-svn: 374509
---
 llvm/lib/Target/X86/X86ISelLowering.cpp         | 13 +++++++++++++
 llvm/test/CodeGen/X86/min-legal-vector-width.ll |  3 +--
 2 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 275e876..0e11941 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/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.
diff --git a/llvm/test/CodeGen/X86/min-legal-vector-width.ll b/llvm/test/CodeGen/X86/min-legal-vector-width.ll
index 95e6d3e..eb90a2a 100644
--- a/llvm/test/CodeGen/X86/min-legal-vector-width.ll
+++ b/llvm/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
-- 
2.7.4