[InstCombine] Move SSE2/AVX2 arithmetic vector shift folding to instcombiner

As discussed in D11760, this patch moves the (V)PSRA(WD) arithmetic shift-by-constant folding to InstCombine to match the logical shift implementations.

Differential Revision: http://reviews.llvm.org/D11886

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@244495 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 6dd4fb8d98eec40dc3e52eb6070b6b6f110f76de..aac4b1452c08dff13676e27216fce0de5203d57e 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -23464,50 +23464,6 @@ static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
 
     return SDValue();
   }
-
-  // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
-  case Intrinsic::x86_sse2_psrai_w:
-  case Intrinsic::x86_sse2_psrai_d:
-  case Intrinsic::x86_avx2_psrai_w:
-  case Intrinsic::x86_avx2_psrai_d:
-  case Intrinsic::x86_sse2_psra_w:
-  case Intrinsic::x86_sse2_psra_d:
-  case Intrinsic::x86_avx2_psra_w:
-  case Intrinsic::x86_avx2_psra_d: {
-    SDValue Op0 = N->getOperand(1);
-    SDValue Op1 = N->getOperand(2);
-    EVT VT = Op0.getValueType();
-    assert(VT.isVector() && "Expected a vector type!");
-
-    if (isa<BuildVectorSDNode>(Op1))
-      Op1 = Op1.getOperand(0);
-
-    if (!isa<ConstantSDNode>(Op1))
-      return SDValue();
-
-    EVT SVT = VT.getVectorElementType();
-    unsigned SVTBits = SVT.getSizeInBits();
-
-    ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
-    const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
-    uint64_t ShAmt = C.getZExtValue();
-
-    // Don't try to convert this shift into a ISD::SRA if the shift
-    // count is bigger than or equal to the element size.
-    if (ShAmt >= SVTBits)
-      return SDValue();
-
-    // Trivial case: if the shift count is zero, then fold this
-    // into the first operand.
-    if (ShAmt == 0)
-      return Op0;
-
-    // Replace this packed shift intrinsic with a target independent
-    // shift dag node.
-    SDLoc DL(N);
-    SDValue Splat = DAG.getConstant(C, DL, VT);
-    return DAG.getNode(ISD::SRA, DL, VT, Op0, Splat);
-  }
   }
 }
 

diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp
index de413c423481139f46a785f88df0dea098155327..600c8c3639297d094cd43701f46435304f87dbec 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -199,7 +199,9 @@ Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
 
 static Value *SimplifyX86immshift(const IntrinsicInst &II,
                                   InstCombiner::BuilderTy &Builder,
-                                  bool ShiftLeft) {
+                                  bool LogicalShift, bool ShiftLeft) {
+  assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
+
   // Simplify if count is constant.
   auto Arg1 = II.getArgOperand(1);
   auto CAZ = dyn_cast<ConstantAggregateZero>(Arg1);
@@ -238,9 +240,15 @@ static Value *SimplifyX86immshift(const IntrinsicInst &II,
   if (Count == 0)
     return Vec;
 
-  // Handle cases when Shift >= BitWidth - just return zero.
-  if (Count.uge(BitWidth))
-    return ConstantAggregateZero::get(VT);
+  // Handle cases when Shift >= BitWidth.
+  if (Count.uge(BitWidth)) {
+    // If LogicalShift - just return zero.
+    if (LogicalShift)
+      return ConstantAggregateZero::get(VT);
+
+    // If ArithmeticShift - clamp Shift to (BitWidth - 1).
+    Count = APInt(64, BitWidth - 1);
+  }
 
   // Get a constant vector of the same type as the first operand.
   auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth));
@@ -249,7 +257,10 @@ static Value *SimplifyX86immshift(const IntrinsicInst &II,
   if (ShiftLeft)
     return Builder.CreateShl(Vec, ShiftVec);
 
-  return Builder.CreateLShr(Vec, ShiftVec);
+  if (LogicalShift)
+    return Builder.CreateLShr(Vec, ShiftVec);
+
+  return Builder.CreateAShr(Vec, ShiftVec);
 }
 
 static Value *SimplifyX86extend(const IntrinsicInst &II,
@@ -776,6 +787,19 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
     break;
   }
 
+  // Constant fold ashr( <A x Bi>, Ci ).
+  case Intrinsic::x86_sse2_psra_d:
+  case Intrinsic::x86_sse2_psra_w:
+  case Intrinsic::x86_sse2_psrai_d:
+  case Intrinsic::x86_sse2_psrai_w:
+  case Intrinsic::x86_avx2_psra_d:
+  case Intrinsic::x86_avx2_psra_w:
+  case Intrinsic::x86_avx2_psrai_d:
+  case Intrinsic::x86_avx2_psrai_w:
+    if (Value *V = SimplifyX86immshift(*II, *Builder, false, false))
+      return ReplaceInstUsesWith(*II, V);
+    break;
+
   // Constant fold lshr( <A x Bi>, Ci ).
   case Intrinsic::x86_sse2_psrl_d:
   case Intrinsic::x86_sse2_psrl_q:
@@ -789,7 +813,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
   case Intrinsic::x86_avx2_psrli_d:
   case Intrinsic::x86_avx2_psrli_q:
   case Intrinsic::x86_avx2_psrli_w:
-    if (Value *V = SimplifyX86immshift(*II, *Builder, false))
+    if (Value *V = SimplifyX86immshift(*II, *Builder, true, false))
       return ReplaceInstUsesWith(*II, V);
     break;
 
@@ -806,7 +830,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
   case Intrinsic::x86_avx2_pslli_d:
   case Intrinsic::x86_avx2_pslli_q:
   case Intrinsic::x86_avx2_pslli_w:
-    if (Value *V = SimplifyX86immshift(*II, *Builder, true))
+    if (Value *V = SimplifyX86immshift(*II, *Builder, true, true))
       return ReplaceInstUsesWith(*II, V);
     break;
 

diff --git a/test/CodeGen/X86/combine-avx2-intrinsics.ll b/test/CodeGen/X86/combine-avx2-intrinsics.ll
index 8794f8b868490418ec336ac364fcd2b40a4ff490..2951ec4b95d43de4a9c03b464f4b3dce78f8b864 100644 (file)
--- a/test/CodeGen/X86/combine-avx2-intrinsics.ll
+++ b/test/CodeGen/X86/combine-avx2-intrinsics.ll
@@ -3,47 +3,6 @@
 ; Verify that the backend correctly combines AVX2 builtin intrinsics.
 
 
-define <8 x i32> @test_psra_1(<8 x i32> %A) {
-  %1 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %A, i32 3)
-  %2 = tail call <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32> %1, <4 x i32> <i32 3, i32 0, i32 7, i32 0>)
-  %3 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %2, i32 2)
-  ret <8 x i32> %3
-}
-; CHECK-LABEL: test_psra_1
-; CHECK: vpsrad $8, %ymm0, %ymm0
-; CHECK-NEXT: ret
-
-define <16 x i16> @test_psra_2(<16 x i16> %A) {
-  %1 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %A, i32 3)
-  %2 = tail call <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16> %1, <8 x i16> <i16 3, i16 0, i16 0, i16 0, i16 7, i16 0, i16 0, i16 0>)
-  %3 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %2, i32 2)
-  ret <16 x i16> %3
-}
-; CHECK-LABEL: test_psra_2
-; CHECK: vpsraw $8, %ymm0, %ymm0
-; CHECK-NEXT: ret
-
-define <16 x i16> @test_psra_3(<16 x i16> %A) {
-  %1 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %A, i32 0)
-  %2 = tail call <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16> %1, <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 7, i16 0, i16 0, i16 0>)
-  %3 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %2, i32 0)
-  ret <16 x i16> %3
-}
-; CHECK-LABEL: test_psra_3
-; CHECK-NOT: vpsraw
-; CHECK: ret
-
-define <8 x i32> @test_psra_4(<8 x i32> %A) {
-  %1 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %A, i32 0)
-  %2 = tail call <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32> %1, <4 x i32> <i32 0, i32 0, i32 7, i32 0>)
-  %3 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %2, i32 0)
-  ret <8 x i32> %3
-}
-; CHECK-LABEL: test_psra_4
-; CHECK-NOT: vpsrad
-; CHECK: ret
-
-
 define <32 x i8> @test_x86_avx2_pblendvb(<32 x i8> %a0, <32 x i8> %a1) {
   %res = call <32 x i8> @llvm.x86.avx2.pblendvb(<32 x i8> %a0, <32 x i8> %a0, <32 x i8> %a1)
   ret <32 x i8> %res
@@ -157,8 +116,4 @@ declare <32 x i8> @llvm.x86.avx2.pblendvb(<32 x i8>, <32 x i8>, <32 x i8>)
 declare <16 x i16> @llvm.x86.avx2.pblendw(<16 x i16>, <16 x i16>, i32)
 declare <4 x i32> @llvm.x86.avx2.pblendd.128(<4 x i32>, <4 x i32>, i32)
 declare <8 x i32> @llvm.x86.avx2.pblendd.256(<8 x i32>, <8 x i32>, i32)
-declare <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16>, <8 x i16>)
-declare <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16>, i32)
-declare <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32>, <4 x i32>)
-declare <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32>, i32)
 

diff --git a/test/CodeGen/X86/combine-sse2-intrinsics.ll b/test/CodeGen/X86/combine-sse2-intrinsics.ll
deleted file mode 100644 (file)
index fa500e5..0000000
--- a/test/CodeGen/X86/combine-sse2-intrinsics.ll
+++ /dev/null
@@ -1,53 +0,0 @@
-; RUN: llc < %s -march=x86 -mcpu=core2 | FileCheck %s
-; RUN: llc < %s -march=x86-64 -mcpu=corei7 | FileCheck %s
-
-; Verify that the backend correctly combines SSE2 builtin intrinsics.
-
-
-define <4 x i32> @test_psra_1(<4 x i32> %A) {
-  %1 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %A, i32 3)
-  %2 = tail call <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32> %1, <4 x i32> <i32 3, i32 0, i32 7, i32 0>)
-  %3 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %2, i32 2)
-  ret <4 x i32> %3
-}
-; CHECK-LABEL: test_psra_1
-; CHECK: psrad $8, %xmm0
-; CHECK-NEXT: ret
-
-define <8 x i16> @test_psra_2(<8 x i16> %A) {
-  %1 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %A, i32 3)
-  %2 = tail call <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16> %1, <8 x i16> <i16 3, i16 0, i16 0, i16 0, i16 7, i16 0, i16 0, i16 0>)
-  %3 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %2, i32 2)
-  ret <8 x i16> %3
-}
-; CHECK-LABEL: test_psra_2
-; CHECK: psraw $8, %xmm0
-; CHECK-NEXT: ret
-
-define <4 x i32> @test_psra_3(<4 x i32> %A) {
-  %1 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %A, i32 0)
-  %2 = tail call <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32> %1, <4 x i32> <i32 0, i32 0, i32 7, i32 0>)
-  %3 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %2, i32 0)
-  ret <4 x i32> %3
-}
-; CHECK-LABEL: test_psra_3
-; CHECK-NOT: psrad
-; CHECK: ret
-
-
-define <8 x i16> @test_psra_4(<8 x i16> %A) {
-  %1 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %A, i32 0)
-  %2 = tail call <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16> %1, <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 7, i16 0, i16 0, i16 0>)
-  %3 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %2, i32 0)
-  ret <8 x i16> %3
-}
-; CHECK-LABEL: test_psra_4
-; CHECK-NOT: psraw
-; CHECK: ret
-
-
-declare <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16>, <8 x i16>)
-declare <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16>, i32)
-declare <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32>, <4 x i32>)
-declare <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32>, i32)
-

diff --git a/test/Transforms/InstCombine/x86-vector-shifts.ll b/test/Transforms/InstCombine/x86-vector-shifts.ll
index fdfad4900f972fae9a26efcd47afb7593be29762..95700f08d810fe4733471a22d56a1c8fe6eefb26 100644 (file)
--- a/test/Transforms/InstCombine/x86-vector-shifts.ll
+++ b/test/Transforms/InstCombine/x86-vector-shifts.ll
@@ -1,6 +1,102 @@
 ; RUN: opt < %s -instcombine -S | FileCheck %s\r
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"\r
 \r
+;\r
+; ASHR - Immediate\r
+;\r
+\r
+define <8 x i16> @sse2_psrai_w_0(<8 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psrai_w_0\r
+; CHECK-NEXT: ret <8 x i16> %v\r
+  %1 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %v, i32 0)\r
+  ret <8 x i16> %1\r
+}\r
+\r
+define <8 x i16> @sse2_psrai_w_15(<8 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psrai_w_15\r
+; CHECK-NEXT: %1 = ashr <8 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <8 x i16> %1\r
+  %1 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %v, i32 15)\r
+  ret <8 x i16> %1\r
+}\r
+\r
+define <8 x i16> @sse2_psrai_w_64(<8 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psrai_w_64\r
+; CHECK-NEXT: %1 = ashr <8 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <8 x i16> %1\r
+  %1 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %v, i32 64)\r
+  ret <8 x i16> %1\r
+}\r
+\r
+define <4 x i32> @sse2_psrai_d_0(<4 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psrai_d_0\r
+; CHECK-NEXT: ret <4 x i32> %v\r
+  %1 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %v, i32 0)\r
+  ret <4 x i32> %1\r
+}\r
+\r
+define <4 x i32> @sse2_psrai_d_15(<4 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psrai_d_15\r
+; CHECK-NEXT: %1 = ashr <4 x i32> %v, <i32 15, i32 15, i32 15, i32 15>\r
+; CHECK-NEXT: ret <4 x i32> %1\r
+  %1 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %v, i32 15)\r
+  ret <4 x i32> %1\r
+}\r
+\r
+define <4 x i32> @sse2_psrai_d_64(<4 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psrai_d_64\r
+; CHECK-NEXT: %1 = ashr <4 x i32> %v, <i32 31, i32 31, i32 31, i32 31>\r
+; CHECK-NEXT: ret <4 x i32> %1\r
+  %1 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %v, i32 64)\r
+  ret <4 x i32> %1\r
+}\r
+\r
+define <16 x i16> @avx2_psrai_w_0(<16 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psrai_w_0\r
+; CHECK-NEXT: ret <16 x i16> %v\r
+  %1 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %v, i32 0)\r
+  ret <16 x i16> %1\r
+}\r
+\r
+define <16 x i16> @avx2_psrai_w_15(<16 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psrai_w_15\r
+; CHECK-NEXT: %1 = ashr <16 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <16 x i16> %1\r
+  %1 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %v, i32 15)\r
+  ret <16 x i16> %1\r
+}\r
+\r
+define <16 x i16> @avx2_psrai_w_64(<16 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psrai_w_64\r
+; CHECK-NEXT: %1 = ashr <16 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <16 x i16> %1\r
+  %1 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %v, i32 64)\r
+  ret <16 x i16> %1\r
+}\r
+\r
+define <8 x i32> @avx2_psrai_d_0(<8 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psrai_d_0\r
+; CHECK-NEXT: ret <8 x i32> %v\r
+  %1 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %v, i32 0)\r
+  ret <8 x i32> %1\r
+}\r
+\r
+define <8 x i32> @avx2_psrai_d_15(<8 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psrai_d_15\r
+; CHECK-NEXT: %1 = ashr <8 x i32> %v, <i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15>\r
+; CHECK-NEXT: ret <8 x i32> %1\r
+  %1 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %v, i32 15)\r
+  ret <8 x i32> %1\r
+}\r
+\r
+define <8 x i32> @avx2_psrai_d_64(<8 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psrai_d_64\r
+; CHECK-NEXT: %1 = ashr <8 x i32> %v, <i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31>\r
+; CHECK-NEXT: ret <8 x i32> %1\r
+  %1 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %v, i32 64)\r
+  ret <8 x i32> %1\r
+}\r
+\r
 ;\r
 ; LSHR - Immediate\r
 ;\r
@@ -273,6 +369,134 @@ define <4 x i64> @avx2_pslli_q_64(<4 x i64> %v) nounwind readnone uwtable {
   ret <4 x i64> %1\r
 }\r
 \r
+;\r
+; ASHR - Constant Vector\r
+;\r
+\r
+define <8 x i16> @sse2_psra_w_0(<8 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psra_w_0\r
+; CHECK-NEXT: ret <8 x i16> %v\r
+  %1 = tail call <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16> %v, <8 x i16> zeroinitializer)\r
+  ret <8 x i16> %1\r
+}\r
+\r
+define <8 x i16> @sse2_psra_w_15(<8 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psra_w_15\r
+; CHECK-NEXT: %1 = ashr <8 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <8 x i16> %1\r
+  %1 = tail call <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16> %v, <8 x i16> <i16 15, i16 0, i16 0, i16 0, i16 9999, i16 9999, i16 9999, i16 9999>)\r
+  ret <8 x i16> %1\r
+}\r
+\r
+define <8 x i16> @sse2_psra_w_15_splat(<8 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psra_w_15_splat\r
+; CHECK-NEXT: %1 = ashr <8 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <8 x i16> %1\r
+  %1 = tail call <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16> %v, <8 x i16> <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>)\r
+  ret <8 x i16> %1\r
+}\r
+\r
+define <8 x i16> @sse2_psra_w_64(<8 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psra_w_64\r
+; CHECK-NEXT: %1 = ashr <8 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <8 x i16> %1\r
+  %1 = tail call <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16> %v, <8 x i16> <i16 64, i16 0, i16 0, i16 0, i16 9999, i16 9999, i16 9999, i16 9999>)\r
+  ret <8 x i16> %1\r
+}\r
+\r
+define <4 x i32> @sse2_psra_d_0(<4 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psra_d_0\r
+; CHECK-NEXT: ret <4 x i32> %v\r
+  %1 = tail call <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32> %v, <4 x i32> zeroinitializer)\r
+  ret <4 x i32> %1\r
+}\r
+\r
+define <4 x i32> @sse2_psra_d_15(<4 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psra_d_15\r
+; CHECK-NEXT: %1 = ashr <4 x i32> %v, <i32 15, i32 15, i32 15, i32 15>\r
+; CHECK-NEXT: ret <4 x i32> %1\r
+  %1 = tail call <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32> %v, <4 x i32> <i32 15, i32 0, i32 9999, i32 9999>)\r
+  ret <4 x i32> %1\r
+}\r
+\r
+define <4 x i32> @sse2_psra_d_15_splat(<4 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psra_d_15_splat\r
+; CHECK-NEXT: %1 = ashr <4 x i32> %v, <i32 31, i32 31, i32 31, i32 31>\r
+; CHECK-NEXT: ret <4 x i32> %1\r
+  %1 = tail call <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32> %v, <4 x i32> <i32 15, i32 15, i32 15, i32 15>)\r
+  ret <4 x i32> %1\r
+}\r
+\r
+define <4 x i32> @sse2_psra_d_64(<4 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @sse2_psra_d_64\r
+; CHECK-NEXT: %1 = ashr <4 x i32> %v, <i32 31, i32 31, i32 31, i32 31>\r
+; CHECK-NEXT: ret <4 x i32> %1\r
+  %1 = tail call <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32> %v, <4 x i32> <i32 64, i32 0, i32 9999, i32 9999>)\r
+  ret <4 x i32> %1\r
+}\r
+\r
+define <16 x i16> @avx2_psra_w_0(<16 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psra_w_0\r
+; CHECK-NEXT: ret <16 x i16> %v\r
+  %1 = tail call <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16> %v, <8 x i16> zeroinitializer)\r
+  ret <16 x i16> %1\r
+}\r
+\r
+define <16 x i16> @avx2_psra_w_15(<16 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psra_w_15\r
+; CHECK-NEXT: %1 = ashr <16 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <16 x i16> %1\r
+  %1 = tail call <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16> %v, <8 x i16> <i16 15, i16 0, i16 0, i16 0, i16 9999, i16 9999, i16 9999, i16 9999>)\r
+  ret <16 x i16> %1\r
+}\r
+\r
+define <16 x i16> @avx2_psra_w_15_splat(<16 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psra_w_15_splat\r
+; CHECK-NEXT: %1 = ashr <16 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <16 x i16> %1\r
+  %1 = tail call <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16> %v, <8 x i16> <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>)\r
+  ret <16 x i16> %1\r
+}\r
+\r
+define <16 x i16> @avx2_psra_w_64(<16 x i16> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psra_w_64\r
+; CHECK-NEXT: %1 = ashr <16 x i16> %v, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>\r
+; CHECK-NEXT: ret <16 x i16> %1\r
+  %1 = tail call <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16> %v, <8 x i16> <i16 64, i16 0, i16 0, i16 0, i16 9999, i16 9999, i16 9999, i16 9999>)\r
+  ret <16 x i16> %1\r
+}\r
+\r
+define <8 x i32> @avx2_psra_d_0(<8 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psra_d_0\r
+; CHECK-NEXT: ret <8 x i32> %v\r
+  %1 = tail call <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32> %v, <4 x i32> zeroinitializer)\r
+  ret <8 x i32> %1\r
+}\r
+\r
+define <8 x i32> @avx2_psra_d_15(<8 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psra_d_15\r
+; CHECK-NEXT: %1 = ashr <8 x i32> %v, <i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15>\r
+; CHECK-NEXT: ret <8 x i32> %1\r
+  %1 = tail call <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32> %v, <4 x i32> <i32 15, i32 0, i32 9999, i32 9999>)\r
+  ret <8 x i32> %1\r
+}\r
+\r
+define <8 x i32> @avx2_psra_d_15_splat(<8 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psra_d_15_splat\r
+; CHECK-NEXT: %1 = ashr <8 x i32> %v, <i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31>\r
+; CHECK-NEXT: ret <8 x i32> %1\r
+  %1 = tail call <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32> %v, <4 x i32> <i32 15, i32 15, i32 15, i32 15>)\r
+  ret <8 x i32> %1\r
+}\r
+\r
+define <8 x i32> @avx2_psra_d_64(<8 x i32> %v) nounwind readnone uwtable {\r
+; CHECK-LABEL: @avx2_psra_d_64\r
+; CHECK-NEXT: %1 = ashr <8 x i32> %v, <i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31>\r
+; CHECK-NEXT: ret <8 x i32> %1\r
+  %1 = tail call <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32> %v, <4 x i32> <i32 64, i32 0, i32 9999, i32 9999>)\r
+  ret <8 x i32> %1\r
+}\r
+\r
 ;\r
 ; LSHR - Constant Vector\r
 ;\r
@@ -605,6 +829,82 @@ define <4 x i64> @avx2_psll_q_64(<4 x i64> %v) nounwind readnone uwtable {
 ; Constant Folding\r
 ;\r
 \r
+define <8 x i16> @test_sse2_psra_w_0(<8 x i16> %A) {\r
+; CHECK-LABEL: @test_sse2_psra_w_0\r
+; CHECK-NEXT: ret <8 x i16> %A\r
+  %1 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %A, i32 0)\r
+  %2 = tail call <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16> %1, <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 7, i16 0, i16 0, i16 0>)\r
+  %3 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %2, i32 0)\r
+  ret <8 x i16> %3\r
+}\r
+\r
+define <8 x i16> @test_sse2_psra_w_8() {\r
+; CHECK-LABEL: @test_sse2_psra_w_8\r
+; CHECK-NEXT: ret <8 x i16> <i16 -128, i16 64, i16 32, i16 16, i16 -128, i16 64, i16 32, i16 16>\r
+  %1 = bitcast <2 x i64> <i64 1152956690052710400, i64 1152956690052710400> to <8 x i16>\r
+  %2 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %1, i32 3)\r
+  %3 = tail call <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16> %2, <8 x i16> <i16 3, i16 0, i16 0, i16 0, i16 7, i16 0, i16 0, i16 0>)\r
+  %4 = tail call <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16> %3, i32 2)\r
+  ret <8 x i16> %4\r
+}\r
+\r
+define <4 x i32> @test_sse2_psra_d_0(<4 x i32> %A) {\r
+; CHECK-LABEL: @test_sse2_psra_d_0\r
+; CHECK-NEXT: ret <4 x i32> %A\r
+  %1 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %A, i32 0)\r
+  %2 = tail call <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32> %1, <4 x i32> <i32 0, i32 0, i32 7, i32 0>)\r
+  %3 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %1, i32 0)\r
+  ret <4 x i32> %3\r
+}\r
+\r
+define <4 x i32> @sse2_psra_d_8() {\r
+; CHECK-LABEL: @sse2_psra_d_8\r
+; CHECK-NEXT: ret <4 x i32> <i32 4194432, i32 1048608, i32 4194432, i32 1048608>\r
+  %1 = bitcast <2 x i64> <i64 1152956690052710400, i64 1152956690052710400> to <4 x i32>\r
+  %2 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %1, i32 3)\r
+  %3 = tail call <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32> %2, <4 x i32> <i32 3, i32 0, i32 7, i32 0>)\r
+  %4 = tail call <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32> %3, i32 2)\r
+  ret <4 x i32> %4\r
+}\r
+\r
+define <16 x i16> @test_avx2_psra_w_0(<16 x i16> %A) {\r
+; CHECK-LABEL: @test_avx2_psra_w_0\r
+; CHECK-NEXT: ret <16 x i16> %A\r
+  %1 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %A, i32 0)\r
+  %2 = tail call <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16> %1, <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 7, i16 0, i16 0, i16 0>)\r
+  %3 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %2, i32 0)\r
+  ret <16 x i16> %3\r
+}\r
+\r
+define <16 x i16> @test_avx2_psra_w_8(<16 x i16> %A) {\r
+; CHECK-LABEL: @test_avx2_psra_w_8\r
+; CHECK-NEXT: ret <16 x i16> <i16 -128, i16 64, i16 32, i16 16, i16 -128, i16 64, i16 32, i16 16, i16 -128, i16 64, i16 32, i16 16, i16 -128, i16 64, i16 32, i16 16>\r
+  %1 = bitcast <4 x i64> <i64 1152956690052710400, i64 1152956690052710400, i64 1152956690052710400, i64 1152956690052710400> to <16 x i16>\r
+  %2 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %1, i32 3)\r
+  %3 = tail call <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16> %2, <8 x i16> <i16 3, i16 0, i16 0, i16 0, i16 7, i16 0, i16 0, i16 0>)\r
+  %4 = tail call <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16> %3, i32 2)\r
+  ret <16 x i16> %4\r
+}\r
+\r
+define <8 x i32> @test_avx2_psra_d_0(<8 x i32> %A) {\r
+; CHECK-LABEL: @test_avx2_psra_d_0\r
+; CHECK-NEXT: ret <8 x i32> %A\r
+  %1 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %A, i32 0)\r
+  %2 = tail call <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32> %1, <4 x i32> <i32 0, i32 0, i32 7, i32 0>)\r
+  %3 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %2, i32 0)\r
+  ret <8 x i32> %3\r
+}\r
+\r
+define <8 x i32> @test_avx2_psra_d_8() {\r
+; CHECK-LABEL: @test_avx2_psra_d_8\r
+; CHECK-NEXT: ret <8 x i32> <i32 4194432, i32 1048608, i32 4194432, i32 1048608, i32 4194432, i32 1048608, i32 4194432, i32 1048608>\r
+  %1 = bitcast <4 x i64> <i64 1152956690052710400, i64 1152956690052710400, i64 1152956690052710400, i64 1152956690052710400> to <8 x i32>\r
+  %2 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %1, i32 3)\r
+  %3 = tail call <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32> %2, <4 x i32> <i32 3, i32 0, i32 7, i32 0>)\r
+  %4 = tail call <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32> %3, i32 2)\r
+  ret <8 x i32> %4\r
+}\r
+\r
 define <2 x i64> @test_sse2_1() nounwind readnone uwtable {\r
   %S = bitcast i32 1 to i32\r
   %1 = zext i32 %S to i64\r
@@ -814,4 +1114,13 @@ declare <2 x i64> @llvm.x86.sse2.psrl.q(<2 x i64>, <2 x i64>) #1
 declare <4 x i32> @llvm.x86.sse2.psrl.d(<4 x i32>, <4 x i32>) #1\r
 declare <8 x i16> @llvm.x86.sse2.psrl.w(<8 x i16>, <8 x i16>) #1\r
 \r
+declare <8 x i32> @llvm.x86.avx2.psrai.d(<8 x i32>, i32) #1\r
+declare <16 x i16> @llvm.x86.avx2.psrai.w(<16 x i16>, i32) #1\r
+declare <8 x i32> @llvm.x86.avx2.psra.d(<8 x i32>, <4 x i32>) #1\r
+declare <16 x i16> @llvm.x86.avx2.psra.w(<16 x i16>, <8 x i16>) #1\r
+declare <4 x i32> @llvm.x86.sse2.psrai.d(<4 x i32>, i32) #1\r
+declare <8 x i16> @llvm.x86.sse2.psrai.w(<8 x i16>, i32) #1\r
+declare <4 x i32> @llvm.x86.sse2.psra.d(<4 x i32>, <4 x i32>) #1\r
+declare <8 x i16> @llvm.x86.sse2.psra.w(<8 x i16>, <8 x i16>) #1\r
+\r
 attributes #1 = { nounwind readnone }\r