From 7377cff9e7641c75678fd5c80472942fd7ef869a Mon Sep 17 00:00:00 2001
From: Benjamin Kramer <benny.kra@googlemail.com>
Date: Wed, 23 Oct 2013 19:19:04 +0000
Subject: [PATCH] X86: Custom lower zext v16i8 to v16i16.

On sandy bridge (PR17654) we now get
	vpxor	%xmm1, %xmm1, %xmm1
	vpunpckhbw	%xmm1, %xmm0, %xmm2
	vpunpcklbw	%xmm1, %xmm0, %xmm0
	vinsertf128	$1, %xmm2, %ymm0, %ymm0

On haswell it's a simple
	vpmovzxbw	%xmm0, %ymm0

There is a maze of duplicated and dead transforms and patterns in this
area. Remove the dead custom lowering of zext v8i16 to v8i32, that's
already handled by LowerAVXExtend.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193262 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp   | 25 ++++++-------------------
 lib/Target/X86/X86InstrSSE.td        |  2 ++
 test/CodeGen/X86/avx-zext.ll         | 12 ++++++++++++
 test/CodeGen/X86/avx2-conversions.ll |  9 +++++++++
 4 files changed, 29 insertions(+), 19 deletions(-)

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 7dddb49beee..5f29f4f8ce3 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1160,6 +1160,7 @@ void X86TargetLowering::resetOperationActions() {
     setOperationAction(ISD::SINT_TO_FP,         MVT::v8i32, Legal);
     setOperationAction(ISD::FP_ROUND,           MVT::v4f32, Legal);
 
+    setOperationAction(ISD::ZERO_EXTEND,        MVT::v16i16, Custom);
     setOperationAction(ISD::ZERO_EXTEND,        MVT::v8i32, Custom);
     setOperationAction(ISD::UINT_TO_FP,         MVT::v8i8,  Custom);
     setOperationAction(ISD::UINT_TO_FP,         MVT::v8i16, Custom);
@@ -8864,7 +8865,8 @@ static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
   //   Concat upper and lower parts.
   //
 
-  if (((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
+  if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
+      ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
       ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
     return SDValue();
 
@@ -8944,24 +8946,9 @@ static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
       return Res;
   }
 
-  if (!VT.is256BitVector() || !SVT.is128BitVector() ||
-      VT.getVectorNumElements() != SVT.getVectorNumElements())
-    return SDValue();
-
-  assert(Subtarget->hasFp256() && "256-bit vector is observed without AVX!");
-
-  // AVX2 has better support of integer extending.
-  if (Subtarget->hasInt256())
-    return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
-
-  SDValue Lo = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32, In);
-  static const int Mask[] = {4, 5, 6, 7, -1, -1, -1, -1};
-  SDValue Hi = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32,
-                           DAG.getVectorShuffle(MVT::v8i16, DL, In,
-                                                DAG.getUNDEF(MVT::v8i16),
-                                                &Mask[0]));
-
-  return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i32, Lo, Hi);
+  assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
+         VT.getVectorNumElements() != SVT.getVectorNumElements());
+  return SDValue();
 }
 
 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index bf091919549..004710b67cc 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -5596,6 +5596,8 @@ let Predicates = [HasAVX2] in {
               (VPMOVZXDQYrr VR128:$src)>;
     def : Pat<(v8i32 (X86vzmovly (v8i16 VR128:$src))),
               (VPMOVZXWDYrr VR128:$src)>;
+    def : Pat<(v16i16 (X86vzmovly (v16i8 VR128:$src))),
+              (VPMOVZXBWYrr VR128:$src)>;
   }
 
   def : Pat<(v4i64 (X86vsmovl (v4i32 VR128:$src))), (VPMOVSXDQYrr VR128:$src)>;
diff --git a/test/CodeGen/X86/avx-zext.ll b/test/CodeGen/X86/avx-zext.ll
index e2b6c552da9..75117463bc3 100644
--- a/test/CodeGen/X86/avx-zext.ll
+++ b/test/CodeGen/X86/avx-zext.ll
@@ -27,3 +27,15 @@ define <8 x i32> @zext_8i8_to_8i32(<8 x i8> %z) {
   %t = zext <8 x i8> %z to <8 x i32>
   ret <8 x i32> %t
 }
+
+; PR17654
+define <16 x i16> @zext_16i8_to_16i16(<16 x i8> %z) {
+; CHECK-LABEL: zext_16i8_to_16i16:
+; CHECK: vpxor
+; CHECK: vpunpckhbw
+; CHECK: vpunpcklbw
+; CHECK: vinsertf128
+; CHECK: ret
+  %t = zext <16 x i8> %z to <16 x i16>
+  ret <16 x i16> %t
+}
diff --git a/test/CodeGen/X86/avx2-conversions.ll b/test/CodeGen/X86/avx2-conversions.ll
index 3ce08dcc737..0143f18fe24 100644
--- a/test/CodeGen/X86/avx2-conversions.ll
+++ b/test/CodeGen/X86/avx2-conversions.ll
@@ -63,6 +63,15 @@ define <8 x i32> @zext_8i8_8i32(<8 x i8> %A) nounwind {
   ret <8 x i32>%B
 }
 
+; CHECK-LABEL: zext_16i8_16i16:
+; CHECK: vpmovzxbw
+; CHECK-NOT: vinsert
+; CHECK: ret
+define <16 x i16> @zext_16i8_16i16(<16 x i8> %z) {
+  %t = zext <16 x i8> %z to <16 x i16>
+  ret <16 x i16> %t
+}
+
 ; CHECK: load_sext_test1
 ; CHECK: vpmovsxdq (%r{{[^,]*}}), %ymm{{.*}}
 ; CHECK: ret