Use shuffles to implement insert_vector_elt for i32, i64, f32, and f64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@44929 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 58 insertions, 64 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 83751412e9..67217fd992 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -3922,53 +3922,35 @@ X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
 
 SDOperand
 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
-  // Transform it so it match pinsrw which expects a 16-bit value in a GR32
-  // as its second argument.
   MVT::ValueType VT = Op.getValueType();
-  MVT::ValueType BaseVT = MVT::getVectorElementType(VT);
+  MVT::ValueType EVT = MVT::getVectorElementType(VT);
+  if (EVT == MVT::i8)
+    return SDOperand();
+
   SDOperand N0 = Op.getOperand(0);
   SDOperand N1 = Op.getOperand(1);
   SDOperand N2 = Op.getOperand(2);
-  if (MVT::getSizeInBits(BaseVT) == 16) {
+
+  if (MVT::getSizeInBits(EVT) == 16) {
+    // Transform it so it match pinsrw which expects a 16-bit value in a GR32
+    // as its second argument.
     if (N1.getValueType() != MVT::i32)
       N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
     if (N2.getValueType() != MVT::i32)
       N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(),getPointerTy());
     return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
-  } else if (MVT::getSizeInBits(BaseVT) == 32) {
-    unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
-    if (Idx == 0) {
-      // Use a movss.
-      N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
-      MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
-      MVT::ValueType BaseVT = MVT::getVectorElementType(MaskVT);
-      SmallVector<SDOperand, 8> MaskVec;
-      MaskVec.push_back(DAG.getConstant(4, BaseVT));
-      for (unsigned i = 1; i <= 3; ++i)
-        MaskVec.push_back(DAG.getConstant(i, BaseVT));
-      return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
-                         DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
-                                     &MaskVec[0], MaskVec.size()));
-    } else {
-      // Use two pinsrw instructions to insert a 32 bit value.
-      Idx <<= 1;
-      if (MVT::isFloatingPoint(N1.getValueType())) {
-        N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1);
-        N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1);
-        N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1,
-                         DAG.getConstant(0, getPointerTy()));
-      }
-      N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0);
-      N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
-                       DAG.getConstant(Idx, getPointerTy()));
-      N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8));
-      N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1,
-                       DAG.getConstant(Idx+1, getPointerTy()));
-      return DAG.getNode(ISD::BIT_CONVERT, VT, N0);
-    }
   }
 
-  return SDOperand();
+  N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1);
+  unsigned Idx = cast<ConstantSDNode>(N2)->getValue();
+  MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4);
+  MVT::ValueType MaskEVT = MVT::getVectorElementType(MaskVT);
+  SmallVector<SDOperand, 4> MaskVec;
+  for (unsigned i = 0; i < 4; ++i)
+    MaskVec.push_back(DAG.getConstant((i == Idx) ? i+4 : i, MaskEVT));
+  return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1,
+                     DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
+                                 &MaskVec[0], MaskVec.size()));
 }
 
 SDOperand
diff --git a/test/CodeGen/X86/2007-07-31-VInsertBug.ll b/test/CodeGen/X86/2007-07-31-VInsertBug.ll
deleted file mode 100644
index 1b13b08045..0000000000
--- a/test/CodeGen/X86/2007-07-31-VInsertBug.ll
+++ /dev/null
@@ -1,16 +0,0 @@
-; RUN: llvm-as < %s | llc -mtriple=i686-apple-darwin -mattr=+sse2 | %prcontext {pinsrw \$2} 1 | grep "movl \$1"
-; RUN: llvm-as < %s | llc -mtriple=i686-apple-darwin -mattr=+sse2 | not grep movss
-
-@G = global <4 x float> zeroinitializer
-
-define void @test(i32 *%P1, i32* %P2, float *%FP) {
-        %T = load float* %FP
-        store i32 0, i32* %P1
-
-        %U = load <4 x float>* @G
-        store i32 1, i32* %P1
-        %V = insertelement <4 x float> %U, float %T, i32 1
-        store <4 x float> %V, <4 x float>* @G
-
-        ret void
-}
diff --git a/test/CodeGen/X86/vec_insert-2.ll b/test/CodeGen/X86/vec_insert-2.ll
new file mode 100644
index 0000000000..8fc97bfd21
--- /dev/null
+++ b/test/CodeGen/X86/vec_insert-2.ll
@@ -0,0 +1,23 @@
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep pinsrw | count 1
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep movhpd | count 1
+; RUN: llvm-as < %s | llc -march=x86-64 -mattr=+sse2 | grep unpcklpd | count 1
+
+define <4 x float> @t1(float %s, <4 x float> %tmp) {
+        %tmp1 = insertelement <4 x float> %tmp, float %s, i32 3
+        ret <4 x float> %tmp1
+}
+
+define <4 x i32> @t2(i32 %s, <4 x i32> %tmp) {
+        %tmp1 = insertelement <4 x i32> %tmp, i32 %s, i32 3
+        ret <4 x i32> %tmp1
+}
+
+define <2 x double> @t3(double %s, <2 x double> %tmp) {
+        %tmp1 = insertelement <2 x double> %tmp, double %s, i32 1
+        ret <2 x double> %tmp1
+}
+
+define <8 x i16> @t4(i16 %s, <8 x i16> %tmp) {
+        %tmp1 = insertelement <8 x i16> %tmp, i16 %s, i32 5
+        ret <8 x i16> %tmp1
+}
diff --git a/test/CodeGen/X86/vec_insert-3.ll b/test/CodeGen/X86/vec_insert-3.ll
new file mode 100644
index 0000000000..1d374b4b9c
--- /dev/null
+++ b/test/CodeGen/X86/vec_insert-3.ll
@@ -0,0 +1,6 @@
+; RUN: llvm-as < %s | llc -march=x86-64 -mattr=+sse2 | grep punpcklqdq | count 1
+
+define <2 x i64> @t1(i64 %s, <2 x i64> %tmp) {
+        %tmp1 = insertelement <2 x i64> %tmp, i64 %s, i32 1
+        ret <2 x i64> %tmp1
+}
diff --git a/test/CodeGen/X86/vec_insert.ll b/test/CodeGen/X86/vec_insert.ll
index 8ee0484fda..c8c9f141ec 100644
--- a/test/CodeGen/X86/vec_insert.ll
+++ b/test/CodeGen/X86/vec_insert.ll
@@ -1,20 +1,19 @@
-; RUN: llvm-upgrade < %s | llvm-as | llc -march=x86 -mattr=+sse2 -o %t -f
-; RUN: grep movss  %t | count 1
-; RUN: grep pinsrw %t | count 2
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | grep movss | count 1
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | not grep pinsrw
 
-void %test(<4 x float>* %F, int %I) {
-	%tmp = load <4 x float>* %F
-	%f = cast int %I to float
-	%tmp1 = insertelement <4 x float> %tmp, float %f, uint 0
-	%tmp18 = add <4 x float> %tmp1, %tmp1
+define void @test(<4 x float>* %F, i32 %I) {
+	%tmp = load <4 x float>* %F		; <<4 x float>> [#uses=1]
+	%f = sitofp i32 %I to float		; <float> [#uses=1]
+	%tmp1 = insertelement <4 x float> %tmp, float %f, i32 0		; <<4 x float>> [#uses=2]
+	%tmp18 = add <4 x float> %tmp1, %tmp1		; <<4 x float>> [#uses=1]
 	store <4 x float> %tmp18, <4 x float>* %F
 	ret void
 }
 
-void %test2(<4 x float>* %F, int %I, float %g) {
-	%tmp = load <4 x float>* %F
-	%f = cast int %I to float
-	%tmp1 = insertelement <4 x float> %tmp, float %f, uint 2
+define void @test2(<4 x float>* %F, i32 %I, float %g) {
+	%tmp = load <4 x float>* %F		; <<4 x float>> [#uses=1]
+	%f = sitofp i32 %I to float		; <float> [#uses=1]
+	%tmp1 = insertelement <4 x float> %tmp, float %f, i32 2		; <<4 x float>> [#uses=1]
 	store <4 x float> %tmp1, <4 x float>* %F
 	ret void
 }