With sse3 and when the source is a load or has multiple uses, favors movddup over shuffp*, pshufd, etc. Without sse3 or when the source is from a register, make use of movlhps

llvm-svn: 56620

4 files changed, 113 insertions, 17 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 55238eb863e..d0b726f13f2 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -2516,6 +2516,21 @@ bool X86::isSplatLoMask(SDNode *N) {
   return true;
 }
 
+/// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
+/// specifies a shuffle of elements that is suitable for input to MOVDDUP.
+bool X86::isMOVDDUPMask(SDNode *N) {
+  assert(N->getOpcode() == ISD::BUILD_VECTOR);
+
+  unsigned e = N->getNumOperands() / 2;
+  for (unsigned i = 0; i < e; ++i)
+    if (!isUndefOrEqual(N->getOperand(i), i))
+      return false;
+  for (unsigned i = 0; i < e; ++i)
+    if (!isUndefOrEqual(N->getOperand(e+i), i))
+      return false;
+  return true;
+}
+
 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
 /// instructions.
@@ -2683,15 +2698,14 @@ static bool ShouldXformToMOVHLPS(SDNode *Mask) {
 /// is promoted to a vector. It also returns the LoadSDNode by reference if
 /// required.
 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
-  if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) {
-    N = N->getOperand(0).getNode();
-    if (ISD::isNON_EXTLoad(N)) {
-      if (LD)
-        *LD = cast<LoadSDNode>(N);
-      return true;
-    }
-  }
-  return false;
+  if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
+    return false;
+  N = N->getOperand(0).getNode();
+  if (!ISD::isNON_EXTLoad(N))
+    return false;
+  if (LD)
+    *LD = cast<LoadSDNode>(N);
+  return true;
 }
 
 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
@@ -2943,6 +2957,46 @@ static SDValue PromoteSplat(SDValue Op, SelectionDAG &DAG, bool HasSSE2) {
   return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
 }
 
+/// isVectorLoad - Returns true if the node is a vector load, a scalar
+/// load that's promoted to vector, or a load bitcasted.
+static bool isVectorLoad(SDValue Op) {
+  assert(Op.getValueType().isVector() && "Expected a vector type");
+  if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR ||
+      Op.getOpcode() == ISD::BIT_CONVERT) {
+    return isa<LoadSDNode>(Op.getOperand(0));
+  }
+  return isa<LoadSDNode>(Op);
+}
+
+
+/// CanonicalizeMovddup - Cannonicalize movddup shuffle to v2f64.
+///
+static SDValue CanonicalizeMovddup(SDValue Op, SDValue V1, SDValue Mask,
+                                   SelectionDAG &DAG, bool HasSSE3) {
+  // If we have sse3 and shuffle has more than one use or input is a load, then
+  // use movddup. Otherwise, use movlhps.
+  bool UseMovddup = HasSSE3 && (!Op.hasOneUse() || isVectorLoad(V1));
+  MVT PVT = UseMovddup ? MVT::v2f64 : MVT::v4f32;
+  MVT VT = Op.getValueType();
+  if (VT == PVT)
+    return Op;
+  unsigned NumElems = PVT.getVectorNumElements();
+  if (NumElems == 2) {
+    SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
+    Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i32, Cst, Cst);
+  } else {
+    assert(NumElems == 4);
+    SDValue Cst0 = DAG.getTargetConstant(0, MVT::i32);
+    SDValue Cst1 = DAG.getTargetConstant(1, MVT::i32);
+    Mask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Cst0, Cst1, Cst0, Cst1);
+  }
+
+  V1 = DAG.getNode(ISD::BIT_CONVERT, PVT, V1);
+  SDValue Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, PVT, V1,
+                                DAG.getNode(ISD::UNDEF, PVT), Mask);
+  return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle);
+}
+
 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
 /// vector of zero or undef vector.  This produces a shuffle where the low
 /// element of V2 is swizzled into the zero/undef vector, landing at element
@@ -3894,6 +3948,11 @@ X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
     return PromoteSplat(Op, DAG, Subtarget->hasSSE2());
   }
 
+  // Canonicalize movddup shuffles.
+  if (V2IsUndef && Subtarget->hasSSE2() &&
+      X86::isMOVDDUPMask(PermMask.getNode()))
+    return CanonicalizeMovddup(Op, V1, PermMask, DAG, Subtarget->hasSSE3());
+
   // If the shuffle can be profitably rewritten as a narrower shuffle, then
   // do it!
   if (VT == MVT::v8i16 || VT == MVT::v16i8) {
diff --git a/llvm/lib/Target/X86/X86ISelLowering.h b/llvm/lib/Target/X86/X86ISelLowering.h
index a2d8a11dc3f..784069b4d20 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.h
+++ b/llvm/lib/Target/X86/X86ISelLowering.h
@@ -295,6 +295,10 @@ namespace llvm {
     /// specifies a splat of zero element.
     bool isSplatLoMask(SDNode *N);
 
+    /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
+    /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
+    bool isMOVDDUPMask(SDNode *N);
+
     /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
     /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
     /// instructions.
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index d67beac3e95..f9de8e74758 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -208,6 +208,10 @@ def SSE_splat_lo_mask : PatLeaf<(build_vector), [{
   return X86::isSplatLoMask(N);
 }]>;
 
+def MOVDDUP_shuffle_mask : PatLeaf<(build_vector), [{
+  return X86::isMOVDDUPMask(N);
+}]>;
+
 def MOVHLPS_shuffle_mask : PatLeaf<(build_vector), [{
   return X86::isMOVHLPSMask(N);
 }]>;
@@ -755,6 +759,11 @@ def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:
 } // AddedComplexity
 } // Constraints = "$src1 = $dst"
 
+let AddedComplexity = 15 in
+def : Pat<(v4f32 (vector_shuffle VR128:$src, (undef), MOVDDUP_shuffle_mask)),
+          (MOVLHPSrr VR128:$src, VR128:$src)>, Requires<[HasSSE1]>;
+
+
 
 
 // Arithmetic
@@ -2452,16 +2461,24 @@ def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
 
 def MOVDDUPrr  : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
                       "movddup\t{$src, $dst|$dst, $src}",
-                      [(set VR128:$dst, (v2f64 (vector_shuffle
-                                                VR128:$src, (undef),
-                                                SSE_splat_lo_mask)))]>;
+                      [(set VR128:$dst,
+                         (v2f64 (vector_shuffle VR128:$src, (undef),
+                                                MOVDDUP_shuffle_mask)))]>;
 def MOVDDUPrm  : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
                       "movddup\t{$src, $dst|$dst, $src}",
-                      [(set VR128:$dst,
-                        (v2f64 (vector_shuffle
-                                (scalar_to_vector (loadf64 addr:$src)),
-                                (undef),
-                                SSE_splat_lo_mask)))]>;
+                    [(set VR128:$dst,
+                      (v2f64 (vector_shuffle
+                              (scalar_to_vector (loadf64 addr:$src)),
+                              (undef), MOVDDUP_shuffle_mask)))]>;
+
+def : Pat<(vector_shuffle
+             (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
+             (undef), MOVDDUP_shuffle_mask),
+          (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
+def : Pat<(vector_shuffle
+            (memopv2f64 addr:$src), (undef), MOVDDUP_shuffle_mask),
+          (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
+
 
 // Arithmetic
 let Constraints = "$src1 = $dst" in {
diff --git a/llvm/test/CodeGen/X86/vec_shuffle-22.ll b/llvm/test/CodeGen/X86/vec_shuffle-22.ll
new file mode 100644
index 00000000000..5648356333b
--- /dev/null
+++ b/llvm/test/CodeGen/X86/vec_shuffle-22.ll
@@ -0,0 +1,16 @@
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2       | not grep shuf
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2,-sse3 | grep movlhps | count 2
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse3       | grep movddup | count 1
+
+define <4 x float> @t1(<4 x float> %a) nounwind  {
+entry:
+        %tmp1 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> < i32 0, i32 1, i32 0, i32 1 >       ; <<4 x float>> [#uses=1]
+        ret <4 x float> %tmp1
+}
+
+define <4 x i32> @t2(<4 x i32>* %a) nounwind {
+entry:
+        %tmp1 = load <4 x i32>* %a;
+	%tmp2 = shufflevector <4 x i32> %tmp1, <4 x i32> undef, <4 x i32> < i32 0, i32 1, i32 0, i32 1 >		; <<4 x i32>> [#uses=1]
+	ret <4 x i32> %tmp2
+}